diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_0/index_tensor([776, 776])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_0/index_tensor([776, 776])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1b81618f1b11bbebe818a3d33a310e310c4da85c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_0/index_tensor([776, 776])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "crane shot begins from a low angle, close to the ocean's surface, then steadily and powerfully advances, closely following an advanced unmanned patrol vessel as it races across a turbulent, inky ocean. The vessel boasts a streamlined, aerodynamic design, constructed from a matte gray carbon fiber composite. Its sharp, wave-piercing bow slices through the oncoming waves at breakneck speed. Each impact unleashes an explosion of white spray, sending countless droplets flying into the frigid air at high speed. These droplets instantly condense into a shimmering mist of ice crystals, perfectly capturing the explosive power and extreme speed characteristic of sports photography. Against a sky shrouded in thick, dark clouds, a diffuse, cold light accentuates the rigid contours of the vessel and the dynamic impact of the waves. The dominant tones of deep blue, cool gray, and dazzling white create a visual impact of tension, raw power, and unstoppable force as the camera zooms in and finally focuses on the vessel's bow cutting through the waves.", + "seed": 23861, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_0/index_776_seed_23861_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10230487454682588, + "vlm_logprob": -0.10230487454682588, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.10230487454682588, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an unmanned patrol vessel with a streamlined, aerodynamic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel constructed from a matte gray carbon fiber composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vessel racing at high speed through a turbulent, inky ocean with white spray and ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment set against a sky with thick dark clouds and diffuse cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a low angle near the ocean surface and steadily advance while zooming in on the vessel's bow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09283658862113953 + ], + "opposite_logprob_tokens": [ + -2.717836618423462 + ], + "logprob_reward": -0.09283658862113953, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.09283658862113953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2058849185705185, + 0.061501212418079376, + 0.03688957914710045, + 0.14146345853805542, + 0.08820245414972305, + 0.039756644517183304, + 0.049617934972047806, + 0.0760406032204628, + 0.14689676463603973, + 0.15374639630317688 + ], + "credit_map": [ + 0.01734965667128563, + 0.030312716960906982, + 0.02746509201824665, + 0.040022820234298706, + 0.023168105632066727, + 0.0181890781968832, + 0.03469021990895271, + 0.014687232673168182, + 0.0047491262666881084, + 0.008170143701136112, + 0.006928608287125826, + 0.010770880617201328, + 0.009334705770015717, + 0.008470478467643261, + 0.007358564995229244, + 0.005718708503991365, + 0.0034939709585160017, + 0.0038221983704715967, + 0.003996870014816523, + 0.0033325538970530033, + 0.004382382612675428, + 0.007976324297487736, + 0.002862268825992942, + 0.007023013196885586, + 0.004745177924633026, + 0.055513493716716766, + 0.004058042075484991, + 0.003646105993539095, + 0.0584699809551239, + 0.00761238019913435, + 0.00343232206068933, + 0.00398597028106451, + 0.008962106890976429, + 0.01064736396074295, + 0.015339965932071209, + 0.008284018374979496, + 0.009884598664939404, + 0.016080759465694427, + 0.005524465348571539, + 0.013479185290634632, + 0.0023276384454220533, + 0.006672696676105261, + 0.006064701825380325, + 0.00549581041559577, + 0.0045768991112709045, + 0.004079321399331093, + 0.006141216494143009, + 0.004398361314088106, + 0.0045509678311645985, + 0.004642487969249487, + 0.005947754252701998, + 0.007248387206345797, + 0.004801617935299873, + 0.007512387353926897, + 0.007311021443456411, + 0.007603310514241457, + 0.006936976220458746, + 0.009445136412978172, + 0.009307315573096275, + 0.011649878695607185, + 0.007867689244449139, + 0.009706194512546062, + 0.012201213277876377, + 0.00892619788646698, + 0.011272807605564594, + 0.018568815663456917, + 0.011577281169593334, + 0.03198494389653206, + 0.015397467650473118, + 0.019573748111724854, + 0.01860055886209011, + 0.01992115005850792, + 0.014741642400622368, + 0.030242683365941048, + 0.017925415188074112, + 0.015625718981027603, + 0.014543980360031128, + 0.02880161441862583, + 0.01200181245803833, + 0.019863540306687355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.09283658862113953 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an unmanned patrol vessel with a streamlined, aerodynamic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel constructed from a matte gray carbon fiber composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vessel racing at high speed through a turbulent, inky ocean with white spray and ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment set against a sky with thick dark clouds and diffuse cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a low angle near the ocean surface and steadily advance while zooming in on the vessel's bow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3488837480545044 + ], + "opposite_logprob_tokens": [ + -1.2238837480545044 + ], + "logprob_reward": -0.3488837480545044, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3488837480545044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3760862350463867, + 0.03229464963078499, + 0.028935516253113747, + 0.1680033951997757, + 0.07527641206979752, + 0.06353020668029785, + 0.07040389627218246, + 0.05788658186793327, + 0.09342993050813675, + 0.034153200685977936 + ], + "credit_map": [ + 0.03269924595952034, + 0.05770627409219742, + 0.0452953577041626, + 0.07559168338775635, + 0.03971779718995094, + 0.03473269194364548, + 0.0632186010479927, + 0.027124568819999695, + 0.0027725922409445047, + 0.005003197118639946, + 0.0036814981140196323, + 0.003837471827864647, + 0.005088692530989647, + 0.005085814278572798, + 0.0032980898395180702, + 0.0035272983368486166, + 0.002636814722791314, + 0.003001685021445155, + 0.005240600556135178, + 0.002791306935250759, + 0.0031505166552960873, + 0.004325304180383682, + 0.0030693376902490854, + 0.004719950258731842, + 0.00575798936188221, + 0.06486982852220535, + 0.004848745185881853, + 0.006051757372915745, + 0.06886473298072815, + 0.006453957408666611, + 0.005725418217480183, + 0.005430963821709156, + 0.006241061259061098, + 0.00993451476097107, + 0.011736181564629078, + 0.007831514813005924, + 0.007400440517812967, + 0.01752305030822754, + 0.004690468776971102, + 0.009919176809489727, + 0.003826821455731988, + 0.01742948219180107, + 0.012498832307755947, + 0.005142408888787031, + 0.006375598255544901, + 0.0061736502684652805, + 0.00677122687920928, + 0.005312190856784582, + 0.004807017743587494, + 0.025637367740273476, + 0.01791529357433319, + 0.004818878136575222, + 0.005457659717649221, + 0.005220997147262096, + 0.0030029495246708393, + 0.00354373618029058, + 0.004901216831058264, + 0.008958094753324986, + 0.007728209253400564, + 0.007532961666584015, + 0.006926403380930424, + 0.006018337793648243, + 0.008208269253373146, + 0.007613093592226505, + 0.006326024420559406, + 0.011148802936077118, + 0.006363300606608391, + 0.02683550864458084, + 0.007931907661259174, + 0.012552992440760136, + 0.010450966656208038, + 0.011820427142083645, + 0.00339309754781425, + 0.005858669988811016, + 0.002870711963623762, + 0.004257082007825375, + 0.0043181623332202435, + 0.0054768845438957214, + 0.003247322514653206, + 0.004731269087642431 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.3488837480545044 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an unmanned patrol vessel with a streamlined, aerodynamic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel constructed from a matte gray carbon fiber composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vessel racing at high speed through a turbulent, inky ocean with white spray and ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment set against a sky with thick dark clouds and diffuse cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a low angle near the ocean surface and steadily advance while zooming in on the vessel's bow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012538531795144081 + ], + "opposite_logprob_tokens": [ + -4.387538433074951 + ], + "logprob_reward": -0.012538531795144081, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012538531795144081, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3832571804523468, + 0.05917464941740036, + 0.04774392023682594, + 0.08351384848356247, + 0.06861330568790436, + 0.14398100972175598, + 0.02256191149353981, + 0.06876994669437408, + 0.05554819479584694, + 0.06683608889579773 + ], + "credit_map": [ + 0.03443881496787071, + 0.05810562148690224, + 0.04577343538403511, + 0.0771389901638031, + 0.03982052952051163, + 0.037108272314071655, + 0.0629267767071724, + 0.02794470824301243, + 0.0051983618177473545, + 0.007428309414535761, + 0.006262131500989199, + 0.011690998449921608, + 0.00791926309466362, + 0.006547118071466684, + 0.0066942982375621796, + 0.007434164173901081, + 0.004249838180840015, + 0.00658975588157773, + 0.005792270414531231, + 0.003996844869107008, + 0.007384015247225761, + 0.010284032672643661, + 0.003554512048140168, + 0.005892647895962, + 0.005169137381017208, + 0.026491766795516014, + 0.004630677867680788, + 0.0034944703802466393, + 0.02805814892053604, + 0.005536721553653479, + 0.0038875837344676256, + 0.0062453350983560085, + 0.005367071833461523, + 0.00849229283630848, + 0.011930882930755615, + 0.007044870872050524, + 0.007333353161811829, + 0.01359743531793356, + 0.004867190960794687, + 0.009980201721191406, + 0.009676549583673477, + 0.034470148384571075, + 0.027409253641963005, + 0.012684710323810577, + 0.012026038020849228, + 0.02146979607641697, + 0.011369539424777031, + 0.014874962158501148, + 0.0016880310140550137, + 0.002605399116873741, + 0.002567772287875414, + 0.0037407921627163887, + 0.0025604371912777424, + 0.004370654467493296, + 0.0022499116603285074, + 0.0027789101004600525, + 0.006612241733819246, + 0.009190691635012627, + 0.008537101559340954, + 0.010771794244647026, + 0.00707645108923316, + 0.009470967575907707, + 0.008110789582133293, + 0.008999908342957497, + 0.004751081578433514, + 0.007521375082433224, + 0.004230503924190998, + 0.01252637431025505, + 0.006876221392303705, + 0.005144543945789337, + 0.00884539820253849, + 0.005652694962918758, + 0.007916094735264778, + 0.01425855327397585, + 0.007113953586667776, + 0.005922282114624977, + 0.006379626225680113, + 0.016069116070866585, + 0.003594712121412158, + 0.005581751931458712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.012538531795144081 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an unmanned patrol vessel with a streamlined, aerodynamic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel constructed from a matte gray carbon fiber composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vessel racing at high speed through a turbulent, inky ocean with white spray and ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment set against a sky with thick dark clouds and diffuse cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a low angle near the ocean surface and steadily advance while zooming in on the vessel's bow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008633549325168133 + ], + "opposite_logprob_tokens": [ + -4.758633613586426 + ], + "logprob_reward": -0.008633549325168133, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008633549325168133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05860702320933342, + 0.06017271801829338, + 0.0795680582523346, + 0.17199662327766418, + 0.042837321758270264, + 0.13362650573253632, + 0.08443930000066757, + 0.25316399335861206, + 0.06166249141097069, + 0.053925950080156326 + ], + "credit_map": [ + 0.005160546861588955, + 0.010265078395605087, + 0.006607528310269117, + 0.007956045679748058, + 0.0077043138444423676, + 0.0060750083066523075, + 0.009084721095860004, + 0.005753780715167522, + 0.005038221832364798, + 0.007095064036548138, + 0.005418068263679743, + 0.01319094467908144, + 0.008502267301082611, + 0.00603253161534667, + 0.007202001288533211, + 0.007693621329963207, + 0.009892535395920277, + 0.010482551530003548, + 0.010283899493515491, + 0.006648453418165445, + 0.01105122733861208, + 0.015790345147252083, + 0.006046751979738474, + 0.00937229860574007, + 0.008237451314926147, + 0.06662199646234512, + 0.006825363729149103, + 0.005494378972798586, + 0.06767963618040085, + 0.008420147933065891, + 0.0036085708998143673, + 0.005109079647809267, + 0.0027766800485551357, + 0.007367927581071854, + 0.007593116257339716, + 0.005971812177449465, + 0.004526374861598015, + 0.004866117145866156, + 0.0038776760920882225, + 0.005857617128640413, + 0.009852089919149876, + 0.027685334905982018, + 0.02641451358795166, + 0.015808923169970512, + 0.011094401590526104, + 0.012288033962249756, + 0.015512305311858654, + 0.014970900490880013, + 0.006659759674221277, + 0.023017745465040207, + 0.014774066396057606, + 0.00613979110494256, + 0.007400353439152241, + 0.011504977010190487, + 0.007216115947812796, + 0.0077264937572181225, + 0.02775893732905388, + 0.05012398213148117, + 0.026857739314436913, + 0.018711067736148834, + 0.026583584025502205, + 0.053104206919670105, + 0.021906882524490356, + 0.028117604553699493, + 0.005411796737462282, + 0.007997937500476837, + 0.004573852755129337, + 0.013443294912576675, + 0.0070327529683709145, + 0.006487750913947821, + 0.01036600861698389, + 0.006349099334329367, + 0.006123894825577736, + 0.01200439129024744, + 0.007351930718868971, + 0.0038996615912765265, + 0.004348798654973507, + 0.0132262222468853, + 0.002580452710390091, + 0.004390595946460962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 5 + ], + "selected_reward": -0.008633549325168133 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an unmanned patrol vessel with a streamlined, aerodynamic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel constructed from a matte gray carbon fiber composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vessel racing at high speed through a turbulent, inky ocean with white spray and ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment set against a sky with thick dark clouds and diffuse cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a low angle near the ocean surface and steadily advance while zooming in on the vessel's bow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048631954938173294 + ], + "opposite_logprob_tokens": [ + -3.0486319065093994 + ], + "logprob_reward": -0.048631954938173294, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048631954938173294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2641828954219818, + 0.030750425532460213, + 0.03510010987520218, + 0.14771559834480286, + 0.03056485578417778, + 0.05320154130458832, + 0.20735375583171844, + 0.03274713456630707, + 0.033898789435625076, + 0.1644849330186844 + ], + "credit_map": [ + 0.024521159008145332, + 0.038579732179641724, + 0.02951701730489731, + 0.0520964115858078, + 0.026338282972574234, + 0.02867213264107704, + 0.0440506748855114, + 0.02040746435523033, + 0.0035604089498519897, + 0.004509024787694216, + 0.0032618597615510225, + 0.004355157725512981, + 0.004352165386080742, + 0.00437852181494236, + 0.0034368825145065784, + 0.0028964057564735413, + 0.002965150633826852, + 0.004220047499984503, + 0.0040968963876366615, + 0.0031023311894387007, + 0.004125071223825216, + 0.007874379865825176, + 0.0038138888776302338, + 0.004902344662696123, + 0.004739013966172934, + 0.05898306518793106, + 0.0037365963216871023, + 0.004199564456939697, + 0.06282214820384979, + 0.006417848169803619, + 0.0030737740453332663, + 0.0037435900885611773, + 0.0031327728647738695, + 0.004220778122544289, + 0.0040471055544912815, + 0.00351579743437469, + 0.0035263290628790855, + 0.005004228092730045, + 0.0029367338865995407, + 0.004181111231446266, + 0.0034251876641064882, + 0.006145205814391375, + 0.00552448770031333, + 0.0058427052572369576, + 0.007257386110723019, + 0.011706762947142124, + 0.007674252614378929, + 0.00562555156648159, + 0.013499735854566097, + 0.06197589263319969, + 0.042439453303813934, + 0.02056737244129181, + 0.01863657496869564, + 0.021727317944169044, + 0.012321939691901207, + 0.01618547923862934, + 0.002950231544673443, + 0.0035986988805234432, + 0.00483139231801033, + 0.005096904467791319, + 0.003597110975533724, + 0.0054007708095014095, + 0.003781104227527976, + 0.003490922274067998, + 0.0030843440908938646, + 0.004407576285302639, + 0.0031466884538531303, + 0.004705620463937521, + 0.004659260157495737, + 0.007199337240308523, + 0.003691374557092786, + 0.00300458911806345, + 0.018132204189896584, + 0.038094572722911835, + 0.020094918087124825, + 0.013060501776635647, + 0.013160903006792068, + 0.039260730147361755, + 0.008119458332657814, + 0.014561636373400688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.048631954938173294 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "crane shot begins from a low angle, close to the ocean's surface, then steadily and powerfully advances, closely following an advanced unmanned patrol vessel as it races across a turbulent, inky ocean. The vessel boasts a streamlined, aerodynamic design, constructed from a matte gray carbon fiber composite. Its sharp, wave-piercing bow slices through the oncoming waves at breakneck speed. Each impact unleashes an explosion of white spray, sending countless droplets flying into the frigid air at high speed. These droplets instantly condense into a shimmering mist of ice crystals, perfectly capturing the explosive power and extreme speed characteristic of sports photography. Against a sky shrouded in thick, dark clouds, a diffuse, cold light accentuates the rigid contours of the vessel and the dynamic impact of the waves. The dominant tones of deep blue, cool gray, and dazzling white create a visual impact of tension, raw power, and unstoppable force as the camera zooms in and finally focuses on the vessel's bow cutting through the waves.", + "seed": 23862, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_0/index_776_seed_23862_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05931188294198364, + "vlm_logprob": -0.05931188294198364, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.05931188294198364, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an unmanned patrol vessel with a streamlined, aerodynamic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel constructed from a matte gray carbon fiber composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vessel racing at high speed through a turbulent, inky ocean with white spray and ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment set against a sky with thick dark clouds and diffuse cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a low angle near the ocean surface and steadily advance while zooming in on the vessel's bow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0949779823422432 + ], + "opposite_logprob_tokens": [ + -2.844978094100952 + ], + "logprob_reward": -0.0949779823422432, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0949779823422432, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24107617139816284, + 0.17359139025211334, + 0.09268025308847427, + 0.03246514871716499, + 0.07579123973846436, + 0.07234760373830795, + 0.05045415088534355, + 0.12417680025100708, + 0.09872882813215256, + 0.03868845850229263 + ], + "credit_map": [ + 0.008080048486590385, + 0.09965667873620987, + 0.0086465859785676, + 0.006276387255638838, + 0.07620140165090561, + 0.029381239786744118, + 0.006676710210740566, + 0.006157122552394867, + 0.004726605024188757, + 0.06754366308450699, + 0.0062994747422635555, + 0.003524342318996787, + 0.07445834577083588, + 0.011305876076221466, + 0.0025229807943105698, + 0.0032101026736199856, + 0.006090128794312477, + 0.008909618481993675, + 0.005157696083188057, + 0.010038472712039948, + 0.020373210310935974, + 0.023168817162513733, + 0.012709252536296844, + 0.0062330542132258415, + 0.003577300813049078, + 0.005334839224815369, + 0.0029866250697523355, + 0.0026813631411641836, + 0.006750236265361309, + 0.006689728703349829, + 0.0022435891442000866, + 0.002201469149440527, + 0.006645678076893091, + 0.016947485506534576, + 0.004002199042588472, + 0.0074220565147697926, + 0.009105341508984566, + 0.021238714456558228, + 0.006102253217250109, + 0.004327510017901659, + 0.005545617546886206, + 0.02343132160604, + 0.0039016290102154016, + 0.007623973768204451, + 0.007961076684296131, + 0.016239825636148453, + 0.005023258272558451, + 0.0026209026109427214, + 0.005498123820871115, + 0.010121069848537445, + 0.005873959511518478, + 0.005428089294582605, + 0.007510608062148094, + 0.006896431557834148, + 0.00524530466645956, + 0.003880564123392105, + 0.012147649191319942, + 0.015601814724504948, + 0.020831864327192307, + 0.017374860122799873, + 0.020670758560299873, + 0.015602394938468933, + 0.012976014986634254, + 0.008971436880528927, + 0.00659187650308013, + 0.015109923668205738, + 0.00795299094170332, + 0.008319210261106491, + 0.018150804564356804, + 0.026567107066512108, + 0.010304768569767475, + 0.0057321470230817795, + 0.002180639887228608, + 0.0025051371194422245, + 0.002475474029779434, + 0.0031289253383874893, + 0.009752809070050716, + 0.009608741849660873, + 0.004538043402135372, + 0.004498687107115984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.0949779823422432 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an unmanned patrol vessel with a streamlined, aerodynamic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel constructed from a matte gray carbon fiber composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vessel racing at high speed through a turbulent, inky ocean with white spray and ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment set against a sky with thick dark clouds and diffuse cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a low angle near the ocean surface and steadily advance while zooming in on the vessel's bow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18002597987651825 + ], + "opposite_logprob_tokens": [ + -1.8050259351730347 + ], + "logprob_reward": -0.18002597987651825, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18002597987651825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26315543055534363, + 0.27338966727256775, + 0.1457938849925995, + 0.02684381976723671, + 0.03910108283162117, + 0.10024334490299225, + 0.017464542761445045, + 0.059467583894729614, + 0.05495239421725273, + 0.019588209688663483 + ], + "credit_map": [ + 0.016885079443454742, + 0.07118348032236099, + 0.01979093998670578, + 0.012697801925241947, + 0.023544080555438995, + 0.08445300161838531, + 0.013061887584626675, + 0.02153918705880642, + 0.007556060329079628, + 0.11259640753269196, + 0.0032602751161903143, + 0.0036991641391068697, + 0.12279403209686279, + 0.0174871776252985, + 0.0036712309811264277, + 0.0023253345862030983, + 0.007687153294682503, + 0.010940358974039555, + 0.006669893860816956, + 0.014081067405641079, + 0.0406036376953125, + 0.04073483869433403, + 0.016518015414476395, + 0.00855893176048994, + 0.002813120139762759, + 0.002865792019292712, + 0.001768101821653545, + 0.004319925792515278, + 0.006264626048505306, + 0.005414943210780621, + 0.0021518266294151545, + 0.0012454872485250235, + 0.003926401492208242, + 0.005708443466573954, + 0.0021031901706010103, + 0.004345031455159187, + 0.005453085992485285, + 0.007981162518262863, + 0.0049857087433338165, + 0.004598063882440329, + 0.0035419370979070663, + 0.043624453246593475, + 0.0017234607366845012, + 0.0035212805960327387, + 0.009835063479840755, + 0.03343132510781288, + 0.002795296488329768, + 0.001770547591149807, + 0.002162520308047533, + 0.002292673336341977, + 0.002141322009265423, + 0.0019038512837141752, + 0.003804319305345416, + 0.0023197450209409, + 0.0014326234813779593, + 0.0014074910432100296, + 0.006668408401310444, + 0.007292632013559341, + 0.010301287285983562, + 0.00958272349089384, + 0.008921511471271515, + 0.006321522407233715, + 0.006943952292203903, + 0.003435553517192602, + 0.004767030943185091, + 0.008161713369190693, + 0.004835954401642084, + 0.003721394343301654, + 0.010198241099715233, + 0.013354488648474216, + 0.00616592587903142, + 0.0037476527504622936, + 0.0019092155853286386, + 0.0019726650789380074, + 0.0019886030349880457, + 0.001653737504966557, + 0.0040015289559960365, + 0.003739838721230626, + 0.002262313384562731, + 0.00206031184643507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.18002597987651825 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an unmanned patrol vessel with a streamlined, aerodynamic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel constructed from a matte gray carbon fiber composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vessel racing at high speed through a turbulent, inky ocean with white spray and ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment set against a sky with thick dark clouds and diffuse cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a low angle near the ocean surface and steadily advance while zooming in on the vessel's bow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002820563269779086 + ], + "opposite_logprob_tokens": [ + -5.8778204917907715 + ], + "logprob_reward": -0.002820563269779086, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002820563269779086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2594224810600281, + 0.23845842480659485, + 0.0885055884718895, + 0.08332612365484238, + 0.11211013048887253, + 0.06316518038511276, + 0.03280258551239967, + 0.03274960815906525, + 0.021246079355478287, + 0.06821378320455551 + ], + "credit_map": [ + 0.01484721340239048, + 0.04091678932309151, + 0.01777266152203083, + 0.01589386910200119, + 0.09150484204292297, + 0.05040345340967178, + 0.013619634322822094, + 0.01446404680609703, + 0.006788536906242371, + 0.08785240352153778, + 0.0110934404656291, + 0.007292700465768576, + 0.0984010249376297, + 0.014351279474794865, + 0.006955699063837528, + 0.005723344627767801, + 0.0054865190759301186, + 0.00894032884389162, + 0.007212408818304539, + 0.006929379887878895, + 0.02141953818500042, + 0.025924744084477425, + 0.007385309785604477, + 0.005207360256463289, + 0.008078740909695625, + 0.009142285212874413, + 0.008607863448560238, + 0.0063103558495640755, + 0.021886086091399193, + 0.020936856046319008, + 0.004477479960769415, + 0.0038864577654749155, + 0.004731206689029932, + 0.043287910521030426, + 0.0026991695631295443, + 0.004832829348742962, + 0.007578226272016764, + 0.041544359177351, + 0.003983968403190374, + 0.0034524647053331137, + 0.0035746213980019093, + 0.023019788786768913, + 0.002337387762963772, + 0.0031414597760885954, + 0.004068329930305481, + 0.022839033976197243, + 0.002397633856162429, + 0.0017869309522211552, + 0.0031180381774902344, + 0.006083490792661905, + 0.0048431758768856525, + 0.0030895413365215063, + 0.005997588392347097, + 0.00337979756295681, + 0.0031307346653193235, + 0.0031602235976606607, + 0.003133812453597784, + 0.004334406927227974, + 0.005423242226243019, + 0.004122229292988777, + 0.004230483900755644, + 0.004441883414983749, + 0.003732338547706604, + 0.0033312172163277864, + 0.0016618616646155715, + 0.0032736093271523714, + 0.0024834685027599335, + 0.0023060718085616827, + 0.00317775527946651, + 0.0038541231770068407, + 0.003249588655307889, + 0.0012396018719300628, + 0.00398159958422184, + 0.0049344697035849094, + 0.005581432953476906, + 0.0062407758086919785, + 0.012562917545437813, + 0.014649524353444576, + 0.009958196431398392, + 0.010304870083928108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.002820563269779086 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an unmanned patrol vessel with a streamlined, aerodynamic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel constructed from a matte gray carbon fiber composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vessel racing at high speed through a turbulent, inky ocean with white spray and ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment set against a sky with thick dark clouds and diffuse cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a low angle near the ocean surface and steadily advance while zooming in on the vessel's bow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000559173640795052 + ], + "opposite_logprob_tokens": [ + -7.500559329986572 + ], + "logprob_reward": -0.000559173640795052, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.000559173640795052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36704567074775696, + 0.20780135691165924, + 0.06652680039405823, + 0.05381844565272331, + 0.07995405793190002, + 0.030003877356648445, + 0.0478912852704525, + 0.03421957418322563, + 0.09042699635028839, + 0.02231195569038391 + ], + "credit_map": [ + 0.01322736032307148, + 0.15359891951084137, + 0.012289956212043762, + 0.00896521843969822, + 0.08704090118408203, + 0.07514690607786179, + 0.0070846062153577805, + 0.009691793471574783, + 0.005752534605562687, + 0.08422156423330307, + 0.0055886018089950085, + 0.00393022270873189, + 0.09046395123004913, + 0.011410628445446491, + 0.0035851479042321444, + 0.002848685020580888, + 0.004729785490781069, + 0.0067613618448376656, + 0.0050384956412017345, + 0.0045778327621519566, + 0.0181596539914608, + 0.018538985401391983, + 0.004706535022705793, + 0.004014136269688606, + 0.006008434109389782, + 0.00771380215883255, + 0.004486958961933851, + 0.005309123545885086, + 0.009994998574256897, + 0.012697651982307434, + 0.0039947545155882835, + 0.0036127204075455666, + 0.00807089451700449, + 0.020363766700029373, + 0.0064871772192418575, + 0.006632021628320217, + 0.01713976263999939, + 0.008921192958950996, + 0.006541557144373655, + 0.005797677673399448, + 0.0031971249263733625, + 0.006761392578482628, + 0.0031970005948096514, + 0.0028799159917980433, + 0.003809926798567176, + 0.005502191372215748, + 0.002571307821199298, + 0.0020850144792348146, + 0.005076553672552109, + 0.010225821286439896, + 0.004548985045403242, + 0.004493604879826307, + 0.009793010540306568, + 0.00603637658059597, + 0.00404265895485878, + 0.003674273146316409, + 0.0040983292274177074, + 0.006194750312715769, + 0.0036991501692682505, + 0.003023199737071991, + 0.006420392543077469, + 0.003985912073403597, + 0.0038161573465913534, + 0.0029816790483891964, + 0.007202859967947006, + 0.014348906464874744, + 0.0062713236548006535, + 0.006349998060613871, + 0.016455598175525665, + 0.024199601262807846, + 0.009973579086363316, + 0.005625113844871521, + 0.0016853976994752884, + 0.0020214852411299944, + 0.0024236293975263834, + 0.0029036698397248983, + 0.0031508435495197773, + 0.003125053131952882, + 0.003067204961553216, + 0.003934667445719242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.000559173640795052 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an unmanned patrol vessel with a streamlined, aerodynamic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel constructed from a matte gray carbon fiber composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vessel racing at high speed through a turbulent, inky ocean with white spray and ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment set against a sky with thick dark clouds and diffuse cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a low angle near the ocean surface and steadily advance while zooming in on the vessel's bow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01817571558058262 + ], + "opposite_logprob_tokens": [ + -4.0181756019592285 + ], + "logprob_reward": -0.01817571558058262, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01817571558058262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2142525464296341, + 0.1526678502559662, + 0.05666482076048851, + 0.06979633867740631, + 0.03747088089585304, + 0.08567971736192703, + 0.06532519310712814, + 0.17210130393505096, + 0.09503129124641418, + 0.051010049879550934 + ], + "credit_map": [ + 0.008359436877071857, + 0.08286743611097336, + 0.007995445281267166, + 0.008791630156338215, + 0.05321485549211502, + 0.033667903393507004, + 0.010269594378769398, + 0.009086251258850098, + 0.004186737351119518, + 0.05980165675282478, + 0.002387464512139559, + 0.002938074991106987, + 0.06726283580064774, + 0.010670818388462067, + 0.0031262694392353296, + 0.0022939988411962986, + 0.004973218776285648, + 0.00913053099066019, + 0.00465925270691514, + 0.008736989460885525, + 0.007457842119038105, + 0.01232029777020216, + 0.005697418469935656, + 0.003689274424687028, + 0.010772190056741238, + 0.011794500052928925, + 0.007595549803227186, + 0.005822891369462013, + 0.009739646688103676, + 0.01229308545589447, + 0.006084304768592119, + 0.005694170482456684, + 0.004023827146738768, + 0.005863395985215902, + 0.0020917453803122044, + 0.00534335570409894, + 0.006301060318946838, + 0.005808540619909763, + 0.00482871662825346, + 0.0032102393452078104, + 0.0038931681774556637, + 0.03375096619129181, + 0.0028756617102771997, + 0.004746522754430771, + 0.010152899660170078, + 0.02378181740641594, + 0.0040434314869344234, + 0.0024352495092898607, + 0.006717205513268709, + 0.012072617188096046, + 0.009097842499613762, + 0.007187610026448965, + 0.010878118686378002, + 0.007267206907272339, + 0.006850985810160637, + 0.005253604147583246, + 0.009937481954693794, + 0.049926627427339554, + 0.01500007789582014, + 0.014188065193593502, + 0.0501912385225296, + 0.014548243954777718, + 0.010957380756735802, + 0.007352201268076897, + 0.00966511107981205, + 0.012006592936813831, + 0.014642242342233658, + 0.013123131357133389, + 0.013247950002551079, + 0.012185857631266117, + 0.013242019340395927, + 0.006918387021869421, + 0.002706725848838687, + 0.005014790687710047, + 0.0027190230321139097, + 0.004187355749309063, + 0.010162892751395702, + 0.011221058666706085, + 0.007431409787386656, + 0.007566794753074646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.01817571558058262 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "crane shot begins from a low angle, close to the ocean's surface, then steadily and powerfully advances, closely following an advanced unmanned patrol vessel as it races across a turbulent, inky ocean. The vessel boasts a streamlined, aerodynamic design, constructed from a matte gray carbon fiber composite. Its sharp, wave-piercing bow slices through the oncoming waves at breakneck speed. Each impact unleashes an explosion of white spray, sending countless droplets flying into the frigid air at high speed. These droplets instantly condense into a shimmering mist of ice crystals, perfectly capturing the explosive power and extreme speed characteristic of sports photography. Against a sky shrouded in thick, dark clouds, a diffuse, cold light accentuates the rigid contours of the vessel and the dynamic impact of the waves. The dominant tones of deep blue, cool gray, and dazzling white create a visual impact of tension, raw power, and unstoppable force as the camera zooms in and finally focuses on the vessel's bow cutting through the waves.", + "seed": 23863, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_0/index_776_seed_23863_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06075518066063523, + "vlm_logprob": -0.06075518066063523, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.06075518066063523, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an unmanned patrol vessel with a streamlined, aerodynamic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel constructed from a matte gray carbon fiber composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vessel racing at high speed through a turbulent, inky ocean with white spray and ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment set against a sky with thick dark clouds and diffuse cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a low angle near the ocean surface and steadily advance while zooming in on the vessel's bow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.032353851944208145 + ], + "opposite_logprob_tokens": [ + -4.03235387802124 + ], + "logprob_reward": -0.032353851944208145, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.032353851944208145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12744037806987762, + 0.26709145307540894, + 0.06964780390262604, + 0.05580165237188339, + 0.07960404455661774, + 0.18316906690597534, + 0.11056681722402573, + 0.037185411900281906, + 0.04254361242055893, + 0.02694975957274437 + ], + "credit_map": [ + 0.010851586237549782, + 0.01875516027212143, + 0.013510970398783684, + 0.017937811091542244, + 0.018659071996808052, + 0.01519135944545269, + 0.016825441271066666, + 0.01570897363126278, + 0.02354925125837326, + 0.062182746827602386, + 0.03760876506567001, + 0.024716198444366455, + 0.03692122921347618, + 0.02745123580098152, + 0.03039734996855259, + 0.024264663457870483, + 0.00399004714563489, + 0.011270985938608646, + 0.009168973192572594, + 0.010481084696948528, + 0.00916247721761465, + 0.0066363997757434845, + 0.011838354170322418, + 0.007099485956132412, + 0.00579456752166152, + 0.007998017594218254, + 0.009288551285862923, + 0.005161966197192669, + 0.01078295148909092, + 0.004800396505743265, + 0.005203264765441418, + 0.006771935150027275, + 0.009265470318496227, + 0.012857404537498951, + 0.007543455809354782, + 0.012692362070083618, + 0.00868169218301773, + 0.010155399329960346, + 0.00888466089963913, + 0.009523598477244377, + 0.011148213408887386, + 0.01601044274866581, + 0.0149171631783247, + 0.006966342218220234, + 0.06585895270109177, + 0.05731687694787979, + 0.005190407857298851, + 0.005760674364864826, + 0.01854407601058483, + 0.025617290288209915, + 0.01052339468151331, + 0.011213034391403198, + 0.015212301164865494, + 0.013660253956913948, + 0.00740621006116271, + 0.008390259929001331, + 0.00620087468996644, + 0.0034076953306794167, + 0.007833418436348438, + 0.003758421167731285, + 0.005218884442001581, + 0.00535347918048501, + 0.002679036697372794, + 0.0027336017228662968, + 0.006830502767115831, + 0.004468153230845928, + 0.007909003645181656, + 0.004607353359460831, + 0.005738925654441118, + 0.006163351237773895, + 0.0033063641749322414, + 0.003519961377605796, + 0.003378340508788824, + 0.00253357645124197, + 0.0042707850225269794, + 0.0038938424549996853, + 0.0038484560791403055, + 0.00394233176484704, + 0.002849180018529296, + 0.0022332470398396254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.032353851944208145 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an unmanned patrol vessel with a streamlined, aerodynamic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel constructed from a matte gray carbon fiber composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vessel racing at high speed through a turbulent, inky ocean with white spray and ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment set against a sky with thick dark clouds and diffuse cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a low angle near the ocean surface and steadily advance while zooming in on the vessel's bow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2523149847984314 + ], + "opposite_logprob_tokens": [ + -1.5023150444030762 + ], + "logprob_reward": -0.2523149847984314, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2523149847984314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18385708332061768, + 0.20285606384277344, + 0.05685698986053467, + 0.0440899021923542, + 0.08164563030004501, + 0.1068866029381752, + 0.1806306391954422, + 0.04397689551115036, + 0.058843813836574554, + 0.04035637900233269 + ], + "credit_map": [ + 0.014880228787660599, + 0.02148997038602829, + 0.024936463683843613, + 0.027553316205739975, + 0.024444781243801117, + 0.030727248638868332, + 0.021349957212805748, + 0.018475111573934555, + 0.022869668900966644, + 0.03960747271776199, + 0.03444204106926918, + 0.022417861968278885, + 0.027315974235534668, + 0.014108280651271343, + 0.023463204503059387, + 0.018631545826792717, + 0.004667987115681171, + 0.008265006355941296, + 0.005555969662964344, + 0.007622247561812401, + 0.00699883047491312, + 0.007129347883164883, + 0.007713135797530413, + 0.008904462680220604, + 0.003962160553783178, + 0.005503630265593529, + 0.01005854643881321, + 0.003964009694755077, + 0.005933989770710468, + 0.003925105091184378, + 0.004360463470220566, + 0.006381997838616371, + 0.005952524486929178, + 0.008342592045664787, + 0.0096126738935709, + 0.006120558828115463, + 0.018219372257590294, + 0.02099115587770939, + 0.004053866490721703, + 0.008352881297469139, + 0.007818492129445076, + 0.010641234926879406, + 0.010446948930621147, + 0.005218253470957279, + 0.033599987626075745, + 0.029362309724092484, + 0.003636149922385812, + 0.006163231562823057, + 0.03129763528704643, + 0.042682573199272156, + 0.022280359640717506, + 0.018917836248874664, + 0.020782126113772392, + 0.023249641060829163, + 0.010433129966259003, + 0.010987336747348309, + 0.0063938964158296585, + 0.003887671511620283, + 0.007128998637199402, + 0.005456832237541676, + 0.0050201984122395515, + 0.005124351941049099, + 0.0039723836816847324, + 0.006992561277002096, + 0.008523652330040932, + 0.0063663870096206665, + 0.00922801997512579, + 0.008291586302220821, + 0.006142886821180582, + 0.0075624678283929825, + 0.006381744518876076, + 0.006347073242068291, + 0.005396766122430563, + 0.006732156965881586, + 0.004873853642493486, + 0.006106640677899122, + 0.004259656183421612, + 0.006804493721574545, + 0.004148827865719795, + 0.0020339814946055412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.2523149847984314 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an unmanned patrol vessel with a streamlined, aerodynamic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel constructed from a matte gray carbon fiber composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vessel racing at high speed through a turbulent, inky ocean with white spray and ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment set against a sky with thick dark clouds and diffuse cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a low angle near the ocean surface and steadily advance while zooming in on the vessel's bow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004094788804650307 + ], + "opposite_logprob_tokens": [ + -5.50409460067749 + ], + "logprob_reward": -0.004094788804650307, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004094788804650307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23241427540779114, + 0.3305133581161499, + 0.04990208521485329, + 0.05986467003822327, + 0.0862918496131897, + 0.07368173450231552, + 0.08840810507535934, + 0.014906738884747028, + 0.025133607909083366, + 0.03888358548283577 + ], + "credit_map": [ + 0.01790875568985939, + 0.0268289465457201, + 0.032617513090372086, + 0.02763284556567669, + 0.03237278386950493, + 0.04007744416594505, + 0.02382260374724865, + 0.03115338459610939, + 0.031013263389468193, + 0.059652820229530334, + 0.04608364403247833, + 0.03760029375553131, + 0.038443222641944885, + 0.038092467933893204, + 0.039593055844306946, + 0.04003456234931946, + 0.004427924752235413, + 0.006033020094037056, + 0.005817800760269165, + 0.008492258377373219, + 0.007304707542061806, + 0.003016313537955284, + 0.008304507471621037, + 0.006505548022687435, + 0.005837000906467438, + 0.010384422726929188, + 0.010594591498374939, + 0.00649668462574482, + 0.007838008925318718, + 0.007218807004392147, + 0.005504693370312452, + 0.005990456789731979, + 0.011022977530956268, + 0.009644705802202225, + 0.01164519414305687, + 0.007811253424733877, + 0.01589384861290455, + 0.015332152135670185, + 0.005360598675906658, + 0.009581110440194607, + 0.006977285724133253, + 0.00944121927022934, + 0.008106968365609646, + 0.004747526720166206, + 0.020285682752728462, + 0.01608467288315296, + 0.003670436330139637, + 0.00436794338747859, + 0.01684178039431572, + 0.02179001271724701, + 0.008707436732947826, + 0.009812116622924805, + 0.010119563899934292, + 0.010026935487985611, + 0.006265895441174507, + 0.004844353999942541, + 0.0024303938262164593, + 0.0020503918640315533, + 0.0019839752931147814, + 0.0022666926961392164, + 0.001934386556968093, + 0.001958126202225685, + 0.0010512006701901555, + 0.0012315708445385098, + 0.0026998508255928755, + 0.002975843846797943, + 0.00329681602306664, + 0.004843983333557844, + 0.003314046887680888, + 0.00332043319940567, + 0.0024113729596138, + 0.002271258272230625, + 0.004565822426229715, + 0.004622114356607199, + 0.0048142545856535435, + 0.0046702162362635136, + 0.0063996328972280025, + 0.006156349554657936, + 0.004313070792704821, + 0.0033421192783862352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.004094788804650307 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an unmanned patrol vessel with a streamlined, aerodynamic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel constructed from a matte gray carbon fiber composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vessel racing at high speed through a turbulent, inky ocean with white spray and ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment set against a sky with thick dark clouds and diffuse cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a low angle near the ocean surface and steadily advance while zooming in on the vessel's bow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024851649068295956 + ], + "opposite_logprob_tokens": [ + -6.002485275268555 + ], + "logprob_reward": -0.0024851649068295956, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024851649068295956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19907605648040771, + 0.46721741557121277, + 0.0957321748137474, + 0.009337405674159527, + 0.04839755594730377, + 0.10719665884971619, + 0.03180526942014694, + 0.020538903772830963, + 0.011846914887428284, + 0.008851587772369385 + ], + "credit_map": [ + 0.01587543450295925, + 0.023640945553779602, + 0.02776437997817993, + 0.025113895535469055, + 0.026297099888324738, + 0.03592025861144066, + 0.021302709355950356, + 0.023161357268691063, + 0.04371827095746994, + 0.0867830291390419, + 0.06427992135286331, + 0.04536117985844612, + 0.05618763342499733, + 0.05888156592845917, + 0.05249739810824394, + 0.05950844660401344, + 0.006798108108341694, + 0.014921457506716251, + 0.02005809359252453, + 0.01059637125581503, + 0.009708423167467117, + 0.006960416678339243, + 0.010932750068604946, + 0.01575656048953533, + 0.0009565326618030667, + 0.001296206726692617, + 0.0013633378548547626, + 0.0011703209020197392, + 0.001023835502564907, + 0.0010542911477386951, + 0.0010664142901077867, + 0.0014064671704545617, + 0.005662592593580484, + 0.009680217131972313, + 0.006175778340548277, + 0.0033915529493242502, + 0.008776625618338585, + 0.00596804590895772, + 0.0024515914265066385, + 0.006291156634688377, + 0.007409499958157539, + 0.011274009943008423, + 0.009316020645201206, + 0.003768631722778082, + 0.037510525435209274, + 0.03114553913474083, + 0.0030306221451610327, + 0.003741807071492076, + 0.006620778702199459, + 0.008485116995871067, + 0.002845711773261428, + 0.004087132401764393, + 0.0033940509893000126, + 0.0035535821225494146, + 0.001317485817708075, + 0.0015014118980616331, + 0.00478935893625021, + 0.002494171727448702, + 0.0027571627870202065, + 0.0015374054200947285, + 0.0035922673996537924, + 0.0027668399270623922, + 0.0012060041772201657, + 0.0013956930488348007, + 0.0018084824550896883, + 0.0013320237630978227, + 0.0017942656995728612, + 0.0013904342195019126, + 0.0015705316327512264, + 0.0018079013098031282, + 0.0010764515027403831, + 0.0010668248869478703, + 0.001174621400423348, + 0.0009374092915095389, + 0.0015058951685205102, + 0.0009834786178544164, + 0.0007975687040016055, + 0.0016317440895363688, + 0.000962382648140192, + 0.0008584889001213014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.0024851649068295956 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an unmanned patrol vessel with a streamlined, aerodynamic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel constructed from a matte gray carbon fiber composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vessel racing at high speed through a turbulent, inky ocean with white spray and ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment set against a sky with thick dark clouds and diffuse cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a low angle near the ocean surface and steadily advance while zooming in on the vessel's bow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01252711284905672 + ], + "opposite_logprob_tokens": [ + -4.387526988983154 + ], + "logprob_reward": -0.01252711284905672, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01252711284905672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10547056049108505, + 0.1956305056810379, + 0.067251056432724, + 0.07784006744623184, + 0.07800634205341339, + 0.13577750325202942, + 0.06624460965394974, + 0.05806514248251915, + 0.15609171986579895, + 0.05962250381708145 + ], + "credit_map": [ + 0.00890368316322565, + 0.012812914326786995, + 0.014206956140697002, + 0.017096243798732758, + 0.010588114149868488, + 0.015414099209010601, + 0.0153542160987854, + 0.011094345711171627, + 0.017211170867085457, + 0.039128378033638, + 0.027475344017148018, + 0.01688466966152191, + 0.023419301956892014, + 0.02563246712088585, + 0.01988893933594227, + 0.025990255177021027, + 0.004747528117150068, + 0.0072931330651044846, + 0.015156681649386883, + 0.008594474755227566, + 0.005884284153580666, + 0.006447882391512394, + 0.009290464222431183, + 0.009836616925895214, + 0.006873132660984993, + 0.01063705887645483, + 0.012401792220771313, + 0.007055096793919802, + 0.012626484967768192, + 0.006363969296216965, + 0.010659269988536835, + 0.011223271489143372, + 0.006436883006244898, + 0.010983526706695557, + 0.01180101279169321, + 0.007255555130541325, + 0.013371454551815987, + 0.014732209034264088, + 0.005916429217904806, + 0.007509278599172831, + 0.014972631819546223, + 0.017036568373441696, + 0.01778816431760788, + 0.018393054604530334, + 0.02109231986105442, + 0.018092546612024307, + 0.01431109569966793, + 0.014091131277382374, + 0.012581830844283104, + 0.015620380640029907, + 0.0054325866512954235, + 0.008095874451100826, + 0.009160835295915604, + 0.007400015369057655, + 0.0043447865173220634, + 0.003608304774388671, + 0.008374958299100399, + 0.005039041396230459, + 0.005705353803932667, + 0.007579187862575054, + 0.007355531677603722, + 0.00783016812056303, + 0.008876488544046879, + 0.0073044179007411, + 0.017406223341822624, + 0.01337259542196989, + 0.026939505711197853, + 0.016394630074501038, + 0.02176625467836857, + 0.02302234247326851, + 0.018993770703673363, + 0.018196409568190575, + 0.00688045984134078, + 0.005965340416878462, + 0.009421262890100479, + 0.007223360240459442, + 0.008768345229327679, + 0.009610825218260288, + 0.007331212051212788, + 0.004421698860824108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 5 + ], + "selected_reward": -0.01252711284905672 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "crane shot begins from a low angle, close to the ocean's surface, then steadily and powerfully advances, closely following an advanced unmanned patrol vessel as it races across a turbulent, inky ocean. The vessel boasts a streamlined, aerodynamic design, constructed from a matte gray carbon fiber composite. Its sharp, wave-piercing bow slices through the oncoming waves at breakneck speed. Each impact unleashes an explosion of white spray, sending countless droplets flying into the frigid air at high speed. These droplets instantly condense into a shimmering mist of ice crystals, perfectly capturing the explosive power and extreme speed characteristic of sports photography. Against a sky shrouded in thick, dark clouds, a diffuse, cold light accentuates the rigid contours of the vessel and the dynamic impact of the waves. The dominant tones of deep blue, cool gray, and dazzling white create a visual impact of tension, raw power, and unstoppable force as the camera zooms in and finally focuses on the vessel's bow cutting through the waves.", + "seed": 23864, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_0/index_776_seed_23864_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07568175612250343, + "vlm_logprob": -0.07568175612250343, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.07568175612250343, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an unmanned patrol vessel with a streamlined, aerodynamic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel constructed from a matte gray carbon fiber composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vessel racing at high speed through a turbulent, inky ocean with white spray and ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment set against a sky with thick dark clouds and diffuse cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a low angle near the ocean surface and steadily advance while zooming in on the vessel's bow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.046765826642513275 + ], + "opposite_logprob_tokens": [ + -3.9217658042907715 + ], + "logprob_reward": -0.046765826642513275, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.046765826642513275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13356462121009827, + 0.055023208260536194, + 0.0509210079908371, + 0.10836134850978851, + 0.10782952606678009, + 0.07274755090475082, + 0.12865668535232544, + 0.1756351739168167, + 0.13529042899608612, + 0.03197041526436806 + ], + "credit_map": [ + 0.01678546704351902, + 0.024810105562210083, + 0.019704261794686317, + 0.008925271220505238, + 0.020496241748332977, + 0.014152360148727894, + 0.016142437234520912, + 0.012548471800982952, + 0.004858233500272036, + 0.01201816089451313, + 0.010077601298689842, + 0.004104708321392536, + 0.00454373937100172, + 0.009068495593965054, + 0.00642806151881814, + 0.003924207761883736, + 0.0047145322896540165, + 0.007662518881261349, + 0.004312165547162294, + 0.005017001647502184, + 0.006906155962496996, + 0.012213105335831642, + 0.005935684312134981, + 0.004159844946116209, + 0.008428913541138172, + 0.011921274475753307, + 0.01613754779100418, + 0.01357920840382576, + 0.010755560360848904, + 0.015310999937355518, + 0.019148366525769234, + 0.013079483062028885, + 0.009312348440289497, + 0.010917413048446178, + 0.014587875455617905, + 0.012050921097397804, + 0.013960311189293861, + 0.014186861924827099, + 0.016244027763605118, + 0.01656976342201233, + 0.00927409902215004, + 0.009155349805951118, + 0.006733796093612909, + 0.01650635525584221, + 0.007392617873847485, + 0.009351153858006, + 0.0068078599870204926, + 0.007526319939643145, + 0.01960677281022072, + 0.023484844714403152, + 0.01046549342572689, + 0.020805424079298973, + 0.02179672010242939, + 0.0099836066365242, + 0.009916865266859531, + 0.01259695179760456, + 0.0241279024630785, + 0.024871494621038437, + 0.021803151816129684, + 0.030206721276044846, + 0.01101962011307478, + 0.01853412389755249, + 0.02409030869603157, + 0.02098185010254383, + 0.011712998151779175, + 0.013404139317572117, + 0.009837291203439236, + 0.017784608528017998, + 0.013639872893691063, + 0.030147748067975044, + 0.022367069497704506, + 0.016396701335906982, + 0.0035116372164338827, + 0.002205870347097516, + 0.0025911559350788593, + 0.0036930381320416927, + 0.002731636632233858, + 0.004760542884469032, + 0.005568876396864653, + 0.006907659582793713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.046765826642513275 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an unmanned patrol vessel with a streamlined, aerodynamic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel constructed from a matte gray carbon fiber composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vessel racing at high speed through a turbulent, inky ocean with white spray and ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment set against a sky with thick dark clouds and diffuse cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a low angle near the ocean surface and steadily advance while zooming in on the vessel's bow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3136318027973175 + ], + "opposite_logprob_tokens": [ + -1.3136317729949951 + ], + "logprob_reward": -0.3136318027973175, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3136318027973175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21163833141326904, + 0.26850372552871704, + 0.05766721069812775, + 0.034937094897031784, + 0.0459110364317894, + 0.03560083359479904, + 0.0518500991165638, + 0.17504993081092834, + 0.08843593299388885, + 0.030405815690755844 + ], + "credit_map": [ + 0.02047063037753105, + 0.023976750671863556, + 0.05601585656404495, + 0.02242022566497326, + 0.01904740370810032, + 0.024515487253665924, + 0.018723392859101295, + 0.02646857500076294, + 0.010318884626030922, + 0.09709616005420685, + 0.012476734817028046, + 0.004630404990166426, + 0.10865296423435211, + 0.019238276407122612, + 0.008420548401772976, + 0.007669768296182156, + 0.006839406210929155, + 0.006800815463066101, + 0.006861458998173475, + 0.006594291888177395, + 0.0046597011387348175, + 0.00698441918939352, + 0.008968676440417767, + 0.009958439506590366, + 0.0033431323245167732, + 0.004987610038369894, + 0.00716413464397192, + 0.003354693530127406, + 0.004420072305947542, + 0.00504258181899786, + 0.003696945495903492, + 0.0029279221780598164, + 0.004896936472505331, + 0.0044373273849487305, + 0.009459993802011013, + 0.005031254608184099, + 0.003783252788707614, + 0.006383947096765041, + 0.0054488652385771275, + 0.006469462998211384, + 0.0051262653432786465, + 0.004646775312721729, + 0.00473098736256361, + 0.005074488464742899, + 0.00344048417173326, + 0.004414129536598921, + 0.0035122784320265055, + 0.004655423108488321, + 0.007959621958434582, + 0.010370586067438126, + 0.005105826538056135, + 0.0089102266356349, + 0.006524945143610239, + 0.003624862292781472, + 0.0040796366520226, + 0.005274390336126089, + 0.022408153861761093, + 0.027244148775935173, + 0.022074593231081963, + 0.026812352240085602, + 0.0117334658280015, + 0.01766960881650448, + 0.0224591251462698, + 0.02464849129319191, + 0.007462213281542063, + 0.00875791348516941, + 0.006323889829218388, + 0.010896673426032066, + 0.008665141649544239, + 0.022517945617437363, + 0.016223138198256493, + 0.007589027285575867, + 0.003682549810037017, + 0.0029889254365116358, + 0.00348395062610507, + 0.004086772911250591, + 0.0031475224532186985, + 0.002728808671236038, + 0.004919103812426329, + 0.005368180572986603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.3136318027973175 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an unmanned patrol vessel with a streamlined, aerodynamic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel constructed from a matte gray carbon fiber composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vessel racing at high speed through a turbulent, inky ocean with white spray and ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment set against a sky with thick dark clouds and diffuse cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a low angle near the ocean surface and steadily advance while zooming in on the vessel's bow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003192449454218149 + ], + "opposite_logprob_tokens": [ + -5.75319242477417 + ], + "logprob_reward": -0.003192449454218149, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003192449454218149, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23709796369075775, + 0.0628148689866066, + 0.03964227810502052, + 0.09629719704389572, + 0.047034069895744324, + 0.2330428659915924, + 0.07049225270748138, + 0.14588718116283417, + 0.04196133464574814, + 0.02572999708354473 + ], + "credit_map": [ + 0.025540612637996674, + 0.031069248914718628, + 0.05775156989693642, + 0.022369880229234695, + 0.02864597737789154, + 0.02455570362508297, + 0.018152540549635887, + 0.029012436047196388, + 0.0032743564806878567, + 0.020959120243787766, + 0.004957666154950857, + 0.002802494680508971, + 0.01385051291435957, + 0.008610280230641365, + 0.004732747562229633, + 0.003627694910392165, + 0.005436446517705917, + 0.006079133599996567, + 0.004606509581208229, + 0.004015249665826559, + 0.00628729397431016, + 0.003854739712551236, + 0.005547104403376579, + 0.003815804375335574, + 0.01479705236852169, + 0.018937818706035614, + 0.01574552431702614, + 0.008352137170732021, + 0.011352105997502804, + 0.0146913081407547, + 0.006774004548788071, + 0.0056472476571798325, + 0.004889561329036951, + 0.004239998292177916, + 0.007831819355487823, + 0.004983165301382542, + 0.0058903880417346954, + 0.007368065882474184, + 0.006880530156195164, + 0.004950542934238911, + 0.03690625727176666, + 0.030612491071224213, + 0.03423386812210083, + 0.023518089205026627, + 0.014807583764195442, + 0.03132116049528122, + 0.014063365757465363, + 0.0475800521671772, + 0.011265553534030914, + 0.012127634137868881, + 0.006521972361952066, + 0.011916235089302063, + 0.010847687721252441, + 0.005187345668673515, + 0.005657420959323645, + 0.006968400906771421, + 0.02018139511346817, + 0.021506639197468758, + 0.017933957278728485, + 0.02412226051092148, + 0.009781766682863235, + 0.01451803557574749, + 0.01984022743999958, + 0.01800290308892727, + 0.003337099449709058, + 0.004018051084131002, + 0.0033366349525749683, + 0.008451671339571476, + 0.004639876540750265, + 0.007799255196005106, + 0.005997893866151571, + 0.00438085850328207, + 0.0035891400184482336, + 0.0030960284639149904, + 0.0031439049635082483, + 0.00421015452593565, + 0.0020695661660283804, + 0.001599147799424827, + 0.004105123691260815, + 0.0039169322699308395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.003192449454218149 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an unmanned patrol vessel with a streamlined, aerodynamic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel constructed from a matte gray carbon fiber composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vessel racing at high speed through a turbulent, inky ocean with white spray and ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment set against a sky with thick dark clouds and diffuse cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a low angle near the ocean surface and steadily advance while zooming in on the vessel's bow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006330393371172249 + ], + "opposite_logprob_tokens": [ + -7.375633239746094 + ], + "logprob_reward": -0.0006330393371172249, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006330393371172249, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17214925587177277, + 0.30350369215011597, + 0.0796327218413353, + 0.028216641396284103, + 0.04527051001787186, + 0.0703447014093399, + 0.07643252611160278, + 0.14427180588245392, + 0.05724582448601723, + 0.02293233759701252 + ], + "credit_map": [ + 0.015587755478918552, + 0.02131148800253868, + 0.04562658816576004, + 0.016307802870869637, + 0.022346356883645058, + 0.017760267481207848, + 0.014738069847226143, + 0.01847092993557453, + 0.01622219756245613, + 0.10118807107210159, + 0.040006838738918304, + 0.010729845613241196, + 0.08911193907260895, + 0.02442949451506138, + 0.00878805574029684, + 0.013027246110141277, + 0.015358081087470055, + 0.010804234072566032, + 0.009820973500609398, + 0.0058106775395572186, + 0.01118420995771885, + 0.011104676872491837, + 0.006701595615595579, + 0.008848270401358604, + 0.0031752604991197586, + 0.004481459502130747, + 0.004184490069746971, + 0.00250988663174212, + 0.0031897432636469603, + 0.0029105672147125006, + 0.003182202111929655, + 0.0045830318704247475, + 0.004655537195503712, + 0.005099537782371044, + 0.007393022067844868, + 0.0043965596705675125, + 0.0052185929380357265, + 0.008060626685619354, + 0.005670040845870972, + 0.0047765919007360935, + 0.0061795348301529884, + 0.009429139085114002, + 0.006932688876986504, + 0.02341042086482048, + 0.004876676015555859, + 0.007685268763452768, + 0.005805813241750002, + 0.006025163922458887, + 0.008895078673958778, + 0.012842971831560135, + 0.0065713669173419476, + 0.014562251977622509, + 0.013323808088898659, + 0.00609457865357399, + 0.006096824072301388, + 0.008045644499361515, + 0.018611758947372437, + 0.02249218337237835, + 0.017669152468442917, + 0.02364608645439148, + 0.010585591197013855, + 0.015938838943839073, + 0.017735334113240242, + 0.017592865973711014, + 0.005361833609640598, + 0.00604882650077343, + 0.00481982808560133, + 0.006904117297381163, + 0.006388251669704914, + 0.013288559392094612, + 0.009317934513092041, + 0.005116472952067852, + 0.0024284468963742256, + 0.0022073816508054733, + 0.0019603190012276173, + 0.004094182047992945, + 0.0014884017873555422, + 0.0027584179770201445, + 0.004379275720566511, + 0.0036159104201942682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.0006330393371172249 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an unmanned patrol vessel with a streamlined, aerodynamic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel constructed from a matte gray carbon fiber composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vessel racing at high speed through a turbulent, inky ocean with white spray and ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment set against a sky with thick dark clouds and diffuse cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a low angle near the ocean surface and steadily advance while zooming in on the vessel's bow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014185662381350994 + ], + "opposite_logprob_tokens": [ + -4.264185428619385 + ], + "logprob_reward": -0.014185662381350994, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014185662381350994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3062094748020172, + 0.0634489506483078, + 0.05305056646466255, + 0.07100415974855423, + 0.020624471828341484, + 0.10209038853645325, + 0.1005372628569603, + 0.13438521325588226, + 0.109307661652565, + 0.039341893047094345 + ], + "credit_map": [ + 0.030433543026447296, + 0.04192708432674408, + 0.08212917298078537, + 0.027213146910071373, + 0.034287966787815094, + 0.03494597226381302, + 0.021672409027814865, + 0.033600177615880966, + 0.004725378472357988, + 0.013603762723505497, + 0.008870542980730534, + 0.004266043659299612, + 0.009982015937566757, + 0.006593982223421335, + 0.0068988981656730175, + 0.00850832462310791, + 0.006427302025258541, + 0.0076319314539432526, + 0.0059075686149299145, + 0.006401936989277601, + 0.005120718851685524, + 0.007555689197033644, + 0.009067046456038952, + 0.004938366822898388, + 0.007690581493079662, + 0.010311443358659744, + 0.009217065759003162, + 0.007176996674388647, + 0.0069095971994102, + 0.009521759115159512, + 0.009863141924142838, + 0.010313563980162144, + 0.002880625193938613, + 0.002939060563221574, + 0.0032159057445824146, + 0.002268635667860508, + 0.002535519190132618, + 0.0028243991546332836, + 0.001965669449418783, + 0.0019946543034166098, + 0.009643381461501122, + 0.013058701530098915, + 0.009715437889099121, + 0.028366385027766228, + 0.008103378117084503, + 0.012190042063593864, + 0.009407450444996357, + 0.01160560641437769, + 0.011352402158081532, + 0.016217071563005447, + 0.009792470373213291, + 0.020197561010718346, + 0.014640890061855316, + 0.008424355648458004, + 0.009013095870614052, + 0.01089941430836916, + 0.018560703843832016, + 0.017895182594656944, + 0.01771959476172924, + 0.02195456437766552, + 0.008694025687873363, + 0.014078235253691673, + 0.01717802695930004, + 0.018304873257875443, + 0.008153638802468777, + 0.010925541631877422, + 0.008406422100961208, + 0.022426124662160873, + 0.0115561718121171, + 0.019998794421553612, + 0.01481005921959877, + 0.013030906207859516, + 0.0049592237919569016, + 0.0036570702213793993, + 0.004987618420273066, + 0.00798700749874115, + 0.002618350787088275, + 0.0032911025919020176, + 0.005343291442841291, + 0.00649822223931551 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.014185662381350994 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_13/index_tensor([5339, 5339])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_13/index_tensor([5339, 5339])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..65494da6c7e4e1dc7b51eca9df11812875d397c8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_13/index_tensor([5339, 5339])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "fluid 360-degree dolly shot, held parallel to the subject, gracefully orbits a giant green sea turtle. It glides effortlessly through the crystal-clear, endless expanse of tropical water. Bright sunlight filters through the shimmering water, creating dancing, shimmering rays of light (the Tyndall effect) against the deep blue background. As the camera smoothly pans, these rays create a gorgeous, warm lens flare, with flecks of golden light softly blooming and shifting across the screen. The turtle is enormous, its massive shell, mottled with deep brown and amber, reflecting a soft luster, revealing the marks of age. Using its massive, powerful front flippers like wings, it makes slow, steady, and deliberate strokes. Each stroke is imbued with graceful power, propelling its massive body through the water like a ballerina. Beneath it, the pristine, sandy seabed is clearly visible, dotted with clumps of vibrant coral and swaying seagrass. The entire scene is dominated by a tranquil blue color scheme, illuminated by golden sunlight penetrating the water, creating a peaceful, magnificent and almost dreamy atmosphere, making people feel the peace of life and the awe-inspiring beauty of nature.", + "seed": 9494, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_13/index_5339_seed_9494_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.22884792583063246, + "vlm_logprob": -0.22884792583063246, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.22884792583063246, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a fluid 360-degree dolly shot orbiting the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle's shell mottled with deep brown and amber colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing rays of light (Tyndall effect) and golden lens flare visible in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the seabed visible beneath the turtle, showing sandy bottom, coral, and seagrass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03129817917943001 + ], + "opposite_logprob_tokens": [ + -3.5312981605529785 + ], + "logprob_reward": -0.03129817917943001, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03129817917943001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06728469580411911, + 0.05539018660783768, + 0.015070822089910507, + 0.19588741660118103, + 0.10905640572309494, + 0.0118842339143157, + 0.01270068809390068, + 0.4052667021751404, + 0.08589563518762589, + 0.041563160717487335 + ], + "credit_map": [ + 0.007620824500918388, + 0.006875071674585342, + 0.008766572922468185, + 0.005630456376820803, + 0.008164704777300358, + 0.010634568519890308, + 0.010809171013534069, + 0.008783326484262943, + 0.011592618189752102, + 0.004336005542427301, + 0.013554482720792294, + 0.004016439896076918, + 0.004857405088841915, + 0.003394660772755742, + 0.006027613300830126, + 0.007610961329191923, + 0.002429438754916191, + 0.0014748055255040526, + 0.0016711363568902016, + 0.0017354837618768215, + 0.0021585598587989807, + 0.0016805834602564573, + 0.0015284412074834108, + 0.002392374211922288, + 0.016904139891266823, + 0.01760702207684517, + 0.01344105415046215, + 0.015471511520445347, + 0.01844128966331482, + 0.042372871190309525, + 0.02329828217625618, + 0.04835125431418419, + 0.027893655002117157, + 0.010999109596014023, + 0.0100481528788805, + 0.010691358707845211, + 0.011201590299606323, + 0.01144392043352127, + 0.012294184416532516, + 0.01448444277048111, + 0.0016684341244399548, + 0.0020771685522049665, + 0.0009259262587875128, + 0.0014616455882787704, + 0.0017972958739846945, + 0.0011732516577467322, + 0.0014340667985379696, + 0.0013464450603350997, + 0.0013781150337308645, + 0.0019751950167119503, + 0.0015393744688481092, + 0.001375473802909255, + 0.00174798967782408, + 0.0014120228588581085, + 0.0016746360342949629, + 0.0015978813171386719, + 0.0792635977268219, + 0.027052663266658783, + 0.02998499758541584, + 0.0413966067135334, + 0.03621771186590195, + 0.038960956037044525, + 0.046351391822099686, + 0.10603879392147064, + 0.0067429435439407825, + 0.011972852982580662, + 0.010177782736718655, + 0.011762211099267006, + 0.009096389636397362, + 0.010110226459801197, + 0.007352957967668772, + 0.018680274486541748, + 0.0035021468065679073, + 0.004507889039814472, + 0.004394171759486198, + 0.0038409889675676823, + 0.0029975951183587313, + 0.005282765254378319, + 0.005260718520730734, + 0.011776884086430073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 4 + ], + "selected_reward": -0.03129817917943001 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a fluid 360-degree dolly shot orbiting the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle's shell mottled with deep brown and amber colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing rays of light (Tyndall effect) and golden lens flare visible in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the seabed visible beneath the turtle, showing sandy bottom, coral, and seagrass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9742255210876465 + ], + "opposite_logprob_tokens": [ + -0.4742255210876465 + ], + "logprob_reward": -0.9742255210876465, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9742255210876465, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.04402416571974754, + 0.13670264184474945, + 0.027034085243940353, + 0.06813216209411621, + 0.04458086937665939, + 0.019748279824852943, + 0.02354683354496956, + 0.5190156102180481, + 0.025070203468203545, + 0.09214519709348679 + ], + "credit_map": [ + 0.006209318991750479, + 0.004476112779229879, + 0.005154731683433056, + 0.00457372423261404, + 0.005066512618213892, + 0.005922798532992601, + 0.004083499312400818, + 0.008537466637790203, + 0.041154924780130386, + 0.008310782723128796, + 0.048540350049734116, + 0.005772394593805075, + 0.0062596797943115234, + 0.004817657172679901, + 0.004888596013188362, + 0.01695825904607773, + 0.003405703231692314, + 0.0030029641930013895, + 0.0030892565846443176, + 0.003938312176615, + 0.003840331919491291, + 0.0030742872040718794, + 0.0035080541856586933, + 0.0031751745846122503, + 0.0052343872375786304, + 0.005104766692966223, + 0.005993350874632597, + 0.005450270138680935, + 0.008513540960848331, + 0.018497614189982414, + 0.006125076673924923, + 0.013213160447776318, + 0.005247906781733036, + 0.004265978466719389, + 0.003968618810176849, + 0.006709935609251261, + 0.008729346096515656, + 0.004141350742429495, + 0.006613085977733135, + 0.004904645960777998, + 0.003352339146658778, + 0.0029411090072244406, + 0.0015523252077400684, + 0.0027222149074077606, + 0.002627135021612048, + 0.0020825720857828856, + 0.002470932900905609, + 0.0019996524788439274, + 0.0025801328010857105, + 0.0026172304060310125, + 0.00207670615054667, + 0.0025150650180876255, + 0.0042615290731191635, + 0.003049087943509221, + 0.0024961663875728846, + 0.003950916230678558, + 0.10760397464036942, + 0.03626956418156624, + 0.03493322432041168, + 0.04963947832584381, + 0.0398298017680645, + 0.05177193880081177, + 0.0577109158039093, + 0.14125670492649078, + 0.0020744376815855503, + 0.004514079075306654, + 0.0034153724554926157, + 0.004537533037364483, + 0.0026300677563995123, + 0.002419121330603957, + 0.0023878144565969706, + 0.003091778140515089, + 0.010190264321863651, + 0.018118206411600113, + 0.010459112003445625, + 0.013734360225498676, + 0.012294364161789417, + 0.009051238186657429, + 0.007339636795222759, + 0.010958009399473667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 9 + ], + "selected_reward": -0.9742255210876465 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a fluid 360-degree dolly shot orbiting the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle's shell mottled with deep brown and amber colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing rays of light (Tyndall effect) and golden lens flare visible in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the seabed visible beneath the turtle, showing sandy bottom, coral, and seagrass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018185431137681007 + ], + "opposite_logprob_tokens": [ + -4.018185615539551 + ], + "logprob_reward": -0.018185431137681007, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018185431137681007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06115260347723961, + 0.20274946093559265, + 0.057052332907915115, + 0.0754888653755188, + 0.034003835171461105, + 0.03309636190533638, + 0.028612002730369568, + 0.2919485867023468, + 0.09460828453302383, + 0.12128762900829315 + ], + "credit_map": [ + 0.007997095584869385, + 0.0055493018589913845, + 0.00701639661565423, + 0.00695125013589859, + 0.008076741360127926, + 0.009223232045769691, + 0.006913448218256235, + 0.009425139054656029, + 0.05160905420780182, + 0.014672607183456421, + 0.05958732217550278, + 0.01142904069274664, + 0.015286036767065525, + 0.01118474267423153, + 0.01857033185660839, + 0.02041032537817955, + 0.011548391543328762, + 0.0055678184144198895, + 0.0076856305822730064, + 0.003804646199569106, + 0.004964930471032858, + 0.003747437382116914, + 0.0055115707218647, + 0.014221902936697006, + 0.006468399427831173, + 0.004778922535479069, + 0.007691151462495327, + 0.00685065146535635, + 0.004187315236777067, + 0.0184260755777359, + 0.004968416877090931, + 0.02211793325841427, + 0.006279908120632172, + 0.002684987150132656, + 0.0023860728833824396, + 0.004269154742360115, + 0.006631352473050356, + 0.0037034722045063972, + 0.004447967745363712, + 0.0036009217146784067, + 0.003669889410957694, + 0.004874556325376034, + 0.0018047892954200506, + 0.0035743077751249075, + 0.004374807700514793, + 0.004634325858205557, + 0.003757732454687357, + 0.006405952852219343, + 0.0032651438377797604, + 0.0041006701067090034, + 0.0035492153838276863, + 0.002899884944781661, + 0.003567357547581196, + 0.00342803867533803, + 0.003256784053519368, + 0.004544910043478012, + 0.06005026027560234, + 0.01952608861029148, + 0.0203507449477911, + 0.02879323810338974, + 0.02376512996852398, + 0.028583068400621414, + 0.032132815569639206, + 0.0787472203373909, + 0.007464120164513588, + 0.011977220885455608, + 0.008822918869554996, + 0.009745607152581215, + 0.009432896040380001, + 0.01206748653203249, + 0.010423694737255573, + 0.02467433735728264, + 0.015420695766806602, + 0.008567332290112972, + 0.012482312507927418, + 0.012212598696351051, + 0.009558619931340218, + 0.013201942667365074, + 0.01581636257469654, + 0.034027766436338425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 9 + ], + "selected_reward": -0.018185431137681007 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a fluid 360-degree dolly shot orbiting the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle's shell mottled with deep brown and amber colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing rays of light (Tyndall effect) and golden lens flare visible in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the seabed visible beneath the turtle, showing sandy bottom, coral, and seagrass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007627647835761309 + ], + "opposite_logprob_tokens": [ + -4.882627487182617 + ], + "logprob_reward": -0.007627647835761309, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007627647835761309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04248351603746414, + 0.11335951089859009, + 0.05892086774110794, + 0.11006642132997513, + 0.058924462646245956, + 0.06397610157728195, + 0.07525487244129181, + 0.3336784243583679, + 0.05562061443924904, + 0.08771521598100662 + ], + "credit_map": [ + 0.0069264299236238, + 0.004031741991639137, + 0.005031059496104717, + 0.004813746083527803, + 0.00419800728559494, + 0.005540887359529734, + 0.004955000709742308, + 0.006986643187701702, + 0.03361130133271217, + 0.0073969392105937, + 0.03681453689932823, + 0.0050979722291231155, + 0.004641077946871519, + 0.004185888916254044, + 0.0074509019032120705, + 0.01416089478880167, + 0.010649238713085651, + 0.005326212849467993, + 0.008074276149272919, + 0.006835582200437784, + 0.004160206764936447, + 0.0035266010090708733, + 0.006271181628108025, + 0.0140775665640831, + 0.009511047974228859, + 0.005617699585855007, + 0.009148551151156425, + 0.009793519973754883, + 0.007288854569196701, + 0.02666427753865719, + 0.007231864612549543, + 0.03481060266494751, + 0.016914132982492447, + 0.005896907299757004, + 0.005971199832856655, + 0.004785148426890373, + 0.004331449046730995, + 0.006508643738925457, + 0.005406228359788656, + 0.009110754355788231, + 0.008632880635559559, + 0.005208198446780443, + 0.0038587015587836504, + 0.008113770745694637, + 0.005651283543556929, + 0.007485183887183666, + 0.009176407009363174, + 0.015849675983190536, + 0.009684713557362556, + 0.0067237624898552895, + 0.006990168243646622, + 0.007017078343778849, + 0.01042733620852232, + 0.01132166851311922, + 0.00886049959808588, + 0.014229641295969486, + 0.0699906125664711, + 0.027067165821790695, + 0.02343650348484516, + 0.03455366939306259, + 0.029517339542508125, + 0.03240567073225975, + 0.033421892672777176, + 0.08328557759523392, + 0.005324676167219877, + 0.0070536769926548, + 0.005996871739625931, + 0.010091567412018776, + 0.005704626441001892, + 0.0053766691125929356, + 0.006744908168911934, + 0.00932761374861002, + 0.013243301771581173, + 0.0061893584206700325, + 0.009071571752429008, + 0.008527061901986599, + 0.005332466680556536, + 0.007946341298520565, + 0.010127706453204155, + 0.02727741189301014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 3 + ], + "selected_reward": -0.007627647835761309 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a fluid 360-degree dolly shot orbiting the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle's shell mottled with deep brown and amber colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing rays of light (Tyndall effect) and golden lens flare visible in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the seabed visible beneath the turtle, showing sandy bottom, coral, and seagrass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11290284991264343 + ], + "opposite_logprob_tokens": [ + -2.237902879714966 + ], + "logprob_reward": -0.11290284991264343, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11290284991264343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03797655180096626, + 0.06154147535562515, + 0.034871917217969894, + 0.06897469609975815, + 0.08044757694005966, + 0.09035532176494598, + 0.033471498638391495, + 0.07664986699819565, + 0.1571113020181656, + 0.35859978199005127 + ], + "credit_map": [ + 0.005847501568496227, + 0.0035294517874717712, + 0.004287417978048325, + 0.0035936913918703794, + 0.0047069210559129715, + 0.004452614579349756, + 0.004498559050261974, + 0.0070603941567242146, + 0.011655133217573166, + 0.005284713581204414, + 0.013941287994384766, + 0.0033811021130532026, + 0.004954295698553324, + 0.004061967134475708, + 0.007306122686713934, + 0.010956855490803719, + 0.004335140343755484, + 0.003354737302288413, + 0.005188880488276482, + 0.005295468959957361, + 0.004819980822503567, + 0.002118889009580016, + 0.003509538946673274, + 0.00624928018078208, + 0.005802587606012821, + 0.005168502684682608, + 0.005131073296070099, + 0.006459125317633152, + 0.008278164081275463, + 0.01633082889020443, + 0.00759428646415472, + 0.014210129156708717, + 0.019862385466694832, + 0.007852691225707531, + 0.007162954658269882, + 0.008111642673611641, + 0.009938273578882217, + 0.007886834442615509, + 0.009034330025315285, + 0.010598459281027317, + 0.010648912750184536, + 0.0060765366069972515, + 0.005470011383295059, + 0.010030127130448818, + 0.007253685966134071, + 0.012595467269420624, + 0.014895196072757244, + 0.02338538132607937, + 0.007669349666684866, + 0.0044652619399130344, + 0.0033315797336399555, + 0.0042910282500088215, + 0.0032167842146009207, + 0.004071807023137808, + 0.002995623741298914, + 0.0034300663974136114, + 0.01605008915066719, + 0.007133806589990854, + 0.005149333272129297, + 0.006977303419262171, + 0.006057113409042358, + 0.007383004296571016, + 0.007971392013132572, + 0.019927825778722763, + 0.014760801568627357, + 0.018430719152092934, + 0.016058027744293213, + 0.01834314689040184, + 0.01956804469227791, + 0.01709975115954876, + 0.022052938118577003, + 0.030797874554991722, + 0.03784705698490143, + 0.041844692081213, + 0.032924920320510864, + 0.03898029774427414, + 0.027402153238654137, + 0.03468887135386467, + 0.047839075326919556, + 0.09707270562648773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.11290284991264343 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fluid 360-degree dolly shot, held parallel to the subject, gracefully orbits a giant green sea turtle. It glides effortlessly through the crystal-clear, endless expanse of tropical water. Bright sunlight filters through the shimmering water, creating dancing, shimmering rays of light (the Tyndall effect) against the deep blue background. As the camera smoothly pans, these rays create a gorgeous, warm lens flare, with flecks of golden light softly blooming and shifting across the screen. The turtle is enormous, its massive shell, mottled with deep brown and amber, reflecting a soft luster, revealing the marks of age. Using its massive, powerful front flippers like wings, it makes slow, steady, and deliberate strokes. Each stroke is imbued with graceful power, propelling its massive body through the water like a ballerina. Beneath it, the pristine, sandy seabed is clearly visible, dotted with clumps of vibrant coral and swaying seagrass. The entire scene is dominated by a tranquil blue color scheme, illuminated by golden sunlight penetrating the water, creating a peaceful, magnificent and almost dreamy atmosphere, making people feel the peace of life and the awe-inspiring beauty of nature.", + "seed": 9495, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_13/index_5339_seed_9495_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.21120871957391502, + "vlm_logprob": -0.21120871957391502, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.21120871957391502, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a fluid 360-degree dolly shot orbiting the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle's shell mottled with deep brown and amber colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing rays of light (Tyndall effect) and golden lens flare visible in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the seabed visible beneath the turtle, showing sandy bottom, coral, and seagrass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3159043490886688 + ], + "opposite_logprob_tokens": [ + -1.3159043788909912 + ], + "logprob_reward": -0.3159043490886688, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3159043490886688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1324378401041031, + 0.12567120790481567, + 0.2025725394487381, + 0.05154556781053543, + 0.11259066313505173, + 0.03358553349971771, + 0.0308561809360981, + 0.1199948713183403, + 0.11678849160671234, + 0.07395710796117783 + ], + "credit_map": [ + 0.012960191816091537, + 0.014529208652675152, + 0.017014339566230774, + 0.023364698514342308, + 0.017396776005625725, + 0.01868404820561409, + 0.0128428740426898, + 0.015645699575543404, + 0.013135641813278198, + 0.011234768666327, + 0.014347553253173828, + 0.029406728222966194, + 0.015160461887717247, + 0.014644615352153778, + 0.014947576448321342, + 0.012793859466910362, + 0.03240862488746643, + 0.007136359345167875, + 0.06889369338750839, + 0.01812819391489029, + 0.01141605619341135, + 0.043041981756687164, + 0.012388762086629868, + 0.009158861823379993, + 0.00575663847848773, + 0.004585426766425371, + 0.004913338925689459, + 0.00804912019520998, + 0.005957765970379114, + 0.005303387995809317, + 0.00710919639095664, + 0.00987069308757782, + 0.01281768549233675, + 0.017451567575335503, + 0.01629076525568962, + 0.01993948221206665, + 0.015531587414443493, + 0.010872002691030502, + 0.009665409103035927, + 0.010022163391113281, + 0.0034158260095864534, + 0.004039553459733725, + 0.003039316972717643, + 0.004598009865731001, + 0.006181958131492138, + 0.00478436890989542, + 0.003075262065976858, + 0.004451240412890911, + 0.0032390430569648743, + 0.0037917373701930046, + 0.0026256220880895853, + 0.004890633746981621, + 0.006701081991195679, + 0.003509806003421545, + 0.003419524058699608, + 0.002678732620552182, + 0.010847736150026321, + 0.014201493002474308, + 0.014273644424974918, + 0.020809350535273552, + 0.019652077928185463, + 0.014333355240523815, + 0.012192614376544952, + 0.01368460152298212, + 0.01598287560045719, + 0.010207955725491047, + 0.01220660749822855, + 0.020103611052036285, + 0.01894163154065609, + 0.015688804909586906, + 0.012911817990243435, + 0.01074519194662571, + 0.02030176855623722, + 0.0036565728951245546, + 0.005183677189052105, + 0.004584875423461199, + 0.003916765563189983, + 0.027394816279411316, + 0.002671355614438653, + 0.006247279699891806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.3159043490886688 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a fluid 360-degree dolly shot orbiting the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle's shell mottled with deep brown and amber colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing rays of light (Tyndall effect) and golden lens flare visible in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the seabed visible beneath the turtle, showing sandy bottom, coral, and seagrass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933250427246094 + ], + "opposite_logprob_tokens": [ + -0.6933250427246094 + ], + "logprob_reward": -0.6933250427246094, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933250427246094, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.059165798127651215, + 0.2744860351085663, + 0.11368592828512192, + 0.25699925422668457, + 0.04339331388473511, + 0.0327690951526165, + 0.03549479693174362, + 0.031324826180934906, + 0.10367874056100845, + 0.04900216683745384 + ], + "credit_map": [ + 0.0053528607822954655, + 0.005177023820579052, + 0.0071180141530931, + 0.009662888012826443, + 0.009260972030460835, + 0.007543931249529123, + 0.007662456948310137, + 0.007387647405266762, + 0.026214607059955597, + 0.02351212687790394, + 0.0323929563164711, + 0.06288453191518784, + 0.03596808761358261, + 0.029597340151667595, + 0.03498230502009392, + 0.028934085741639137, + 0.011713185347616673, + 0.004661396145820618, + 0.038552381098270416, + 0.007347733713686466, + 0.006520207040011883, + 0.030924567952752113, + 0.007976952008903027, + 0.005989506375044584, + 0.051390390843153, + 0.019762959331274033, + 0.029816657304763794, + 0.024407995864748955, + 0.017347782850265503, + 0.08116449415683746, + 0.014939782209694386, + 0.018169183284044266, + 0.004399579484015703, + 0.0067536551505327225, + 0.004016329534351826, + 0.006892164703458548, + 0.006309493910521269, + 0.005311992485076189, + 0.00432441895827651, + 0.005385678727179766, + 0.003832845017313957, + 0.003146585077047348, + 0.004098910838365555, + 0.004181091208010912, + 0.005656674038618803, + 0.003848266787827015, + 0.0032926350831985474, + 0.0047120871022343636, + 0.004036142956465483, + 0.005384006071835756, + 0.005125338677316904, + 0.0035949796438217163, + 0.006712088827043772, + 0.003964617848396301, + 0.003397807013243437, + 0.0032798158936202526, + 0.0028448523953557014, + 0.0036420351825654507, + 0.0036320756189525127, + 0.004983994178473949, + 0.004396532196551561, + 0.00550380302593112, + 0.003388984827324748, + 0.0029325487557798624, + 0.031770750880241394, + 0.006011021789163351, + 0.007439658511430025, + 0.007498275488615036, + 0.007249862886965275, + 0.03282109275460243, + 0.004617688246071339, + 0.006270392797887325, + 0.008522904478013515, + 0.0025887915398925543, + 0.006570813246071339, + 0.004520410671830177, + 0.004385072272270918, + 0.016261383891105652, + 0.0022749481722712517, + 0.003877844661474228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.6933250427246094 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a fluid 360-degree dolly shot orbiting the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle's shell mottled with deep brown and amber colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing rays of light (Tyndall effect) and golden lens flare visible in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the seabed visible beneath the turtle, showing sandy bottom, coral, and seagrass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011079339310526848 + ], + "opposite_logprob_tokens": [ + -4.51107931137085 + ], + "logprob_reward": -0.011079339310526848, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011079339310526848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07214828580617905, + 0.4595598876476288, + 0.21155592799186707, + 0.04908287897706032, + 0.05035235732793808, + 0.02372891455888748, + 0.015295792371034622, + 0.07030952721834183, + 0.02499317191541195, + 0.02297327294945717 + ], + "credit_map": [ + 0.007076093927025795, + 0.00751288840547204, + 0.00864168256521225, + 0.010595177300274372, + 0.011762289330363274, + 0.010903683491051197, + 0.0068590217269957066, + 0.008797438815236092, + 0.04601523280143738, + 0.04058212786912918, + 0.05027768015861511, + 0.10609474778175354, + 0.056149620562791824, + 0.05034785717725754, + 0.06065180152654648, + 0.049440816044807434, + 0.03350602462887764, + 0.006081919185817242, + 0.07631047815084457, + 0.015456399880349636, + 0.008784699253737926, + 0.05042608082294464, + 0.011427040211856365, + 0.009563258849084377, + 0.010599908418953419, + 0.0032272455282509327, + 0.004916419740766287, + 0.003970188554376364, + 0.0038252000231295824, + 0.014557122252881527, + 0.0031464528292417526, + 0.004840337671339512, + 0.007925442419946194, + 0.007699353154748678, + 0.004538381937891245, + 0.006423564162105322, + 0.005958754103630781, + 0.007504771463572979, + 0.004086330533027649, + 0.006215753499418497, + 0.0022196625359356403, + 0.002665521577000618, + 0.0020513141062110662, + 0.00325187249109149, + 0.004764670040458441, + 0.003164083929732442, + 0.002374592935666442, + 0.003237195312976837, + 0.0013461611233651638, + 0.0017440723022446036, + 0.001477482495829463, + 0.0027835348155349493, + 0.0030246837995946407, + 0.0017043594270944595, + 0.0014373656595125794, + 0.0017781310016289353, + 0.006224967539310455, + 0.007874804548919201, + 0.008054275996983051, + 0.012717480771243572, + 0.011459256522357464, + 0.007871948182582855, + 0.009228010661900043, + 0.006878770887851715, + 0.004901630338281393, + 0.0020001749508082867, + 0.0022712231148034334, + 0.0031581956427544355, + 0.002856574021279812, + 0.005498596932739019, + 0.0021744174882769585, + 0.0021323570981621742, + 0.003802893916144967, + 0.0027049353811889887, + 0.0026820579078048468, + 0.003039957955479622, + 0.0026141328271478415, + 0.004565757233649492, + 0.0016340112779289484, + 0.0019295254023745656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.011079339310526848 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a fluid 360-degree dolly shot orbiting the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle's shell mottled with deep brown and amber colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing rays of light (Tyndall effect) and golden lens flare visible in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the seabed visible beneath the turtle, showing sandy bottom, coral, and seagrass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005948578007519245 + ], + "opposite_logprob_tokens": [ + -5.130948543548584 + ], + "logprob_reward": -0.005948578007519245, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005948578007519245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08250148594379425, + 0.06676891446113586, + 0.13397225737571716, + 0.11973991990089417, + 0.0555424839258194, + 0.040062226355075836, + 0.052315156906843185, + 0.1847984790802002, + 0.07091117650270462, + 0.19338791072368622 + ], + "credit_map": [ + 0.007789517752826214, + 0.007287236396223307, + 0.009000604972243309, + 0.014854210428893566, + 0.010299106128513813, + 0.010627636685967445, + 0.01107398048043251, + 0.011569185182452202, + 0.006104214582592249, + 0.005208471789956093, + 0.008161145262420177, + 0.011649352498352528, + 0.006349015980958939, + 0.008153111673891544, + 0.010837809182703495, + 0.01030578464269638, + 0.022945059463381767, + 0.0070423465222120285, + 0.034909773617982864, + 0.017905961722135544, + 0.012268681079149246, + 0.017732784152030945, + 0.011420062743127346, + 0.009747576899826527, + 0.02148793637752533, + 0.007362367119640112, + 0.013905003666877747, + 0.009422151371836662, + 0.008510197512805462, + 0.03672453761100769, + 0.007939228788018227, + 0.014388482086360455, + 0.011684774421155453, + 0.006339583080261946, + 0.004199282731860876, + 0.004792710300534964, + 0.006336994469165802, + 0.012798583135008812, + 0.003881996963173151, + 0.005508549977093935, + 0.004776713903993368, + 0.003584275720641017, + 0.00613810308277607, + 0.00396467512473464, + 0.004784625489264727, + 0.005536521784961224, + 0.004260696470737457, + 0.007016613148152828, + 0.008539733476936817, + 0.005777058657258749, + 0.009428469464182854, + 0.004934651777148247, + 0.006245015654712915, + 0.006058134604245424, + 0.0064293984323740005, + 0.004902688320726156, + 0.0328088253736496, + 0.014496193267405033, + 0.02666105516254902, + 0.013743600808084011, + 0.0126494737342, + 0.052756015211343765, + 0.017793534323573112, + 0.01388976164162159, + 0.008648299612104893, + 0.009535448625683784, + 0.008644377812743187, + 0.010611322708427906, + 0.010517211630940437, + 0.00874024722725153, + 0.006929910276085138, + 0.007284349296241999, + 0.04642144963145256, + 0.01915307529270649, + 0.01702979765832424, + 0.011641590856015682, + 0.008879073895514011, + 0.06344663351774216, + 0.006095020100474358, + 0.02072126232087612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 2 + ], + "selected_reward": -0.005948578007519245 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a fluid 360-degree dolly shot orbiting the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle's shell mottled with deep brown and amber colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing rays of light (Tyndall effect) and golden lens flare visible in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the seabed visible beneath the turtle, showing sandy bottom, coral, and seagrass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029786288738250732 + ], + "opposite_logprob_tokens": [ + -3.5297863483428955 + ], + "logprob_reward": -0.029786288738250732, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029786288738250732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06500791758298874, + 0.14485378563404083, + 0.09051883965730667, + 0.14259763062000275, + 0.1812034696340561, + 0.02090018056333065, + 0.036677852272987366, + 0.08771108090877533, + 0.06623175740242004, + 0.1642974466085434 + ], + "credit_map": [ + 0.0065301586873829365, + 0.007186904549598694, + 0.00864927377551794, + 0.009260167367756367, + 0.00930420495569706, + 0.010205160826444626, + 0.00843427050858736, + 0.005437782499939203, + 0.013950194232165813, + 0.011981366202235222, + 0.01690695993602276, + 0.03064187988638878, + 0.01541424635797739, + 0.017209408804774284, + 0.02072368934750557, + 0.018026037141680717, + 0.012995610013604164, + 0.0030834281351417303, + 0.03408126160502434, + 0.004621491301804781, + 0.004062051884829998, + 0.023348193615674973, + 0.004861117340624332, + 0.003465684363618493, + 0.02326541766524315, + 0.008917740546166897, + 0.01693464070558548, + 0.009537547826766968, + 0.008371589705348015, + 0.03904101997613907, + 0.010631990619003773, + 0.0258976761251688, + 0.017412610352039337, + 0.018105369061231613, + 0.02054700441658497, + 0.02014322765171528, + 0.019360767677426338, + 0.018013320863246918, + 0.030242104083299637, + 0.03737905994057655, + 0.0032854799646884203, + 0.002348042791709304, + 0.0019424243364483118, + 0.0023165070451796055, + 0.0030083954334259033, + 0.0026679858565330505, + 0.0019529178971424699, + 0.0033784285187721252, + 0.0036541614681482315, + 0.004150322172790766, + 0.006144401617348194, + 0.0038592105265706778, + 0.006869638804346323, + 0.004339527804404497, + 0.0029812196735292673, + 0.004679369740188122, + 0.012520404532551765, + 0.008423310704529285, + 0.01737079583108425, + 0.008036301471292973, + 0.006717350333929062, + 0.018178125843405724, + 0.006280873902142048, + 0.010183921083807945, + 0.008899549022316933, + 0.008480419404804707, + 0.007612570654600859, + 0.010315080173313618, + 0.009303029626607895, + 0.008907310664653778, + 0.006297620479017496, + 0.00641617551445961, + 0.04368074983358383, + 0.010953055694699287, + 0.017166567966341972, + 0.009642946533858776, + 0.008210750296711922, + 0.057412851601839066, + 0.005683952011168003, + 0.011546563357114792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 1 + ], + "selected_reward": -0.029786288738250732 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fluid 360-degree dolly shot, held parallel to the subject, gracefully orbits a giant green sea turtle. It glides effortlessly through the crystal-clear, endless expanse of tropical water. Bright sunlight filters through the shimmering water, creating dancing, shimmering rays of light (the Tyndall effect) against the deep blue background. As the camera smoothly pans, these rays create a gorgeous, warm lens flare, with flecks of golden light softly blooming and shifting across the screen. The turtle is enormous, its massive shell, mottled with deep brown and amber, reflecting a soft luster, revealing the marks of age. Using its massive, powerful front flippers like wings, it makes slow, steady, and deliberate strokes. Each stroke is imbued with graceful power, propelling its massive body through the water like a ballerina. Beneath it, the pristine, sandy seabed is clearly visible, dotted with clumps of vibrant coral and swaying seagrass. The entire scene is dominated by a tranquil blue color scheme, illuminated by golden sunlight penetrating the water, creating a peaceful, magnificent and almost dreamy atmosphere, making people feel the peace of life and the awe-inspiring beauty of nature.", + "seed": 9496, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_13/index_5339_seed_9496_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.13157457821071147, + "vlm_logprob": -0.13157457821071147, + "vlm_yes_no_margin": 2.45, + "vlm_token_credit": -0.13157457821071147, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a fluid 360-degree dolly shot orbiting the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle's shell mottled with deep brown and amber colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing rays of light (Tyndall effect) and golden lens flare visible in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the seabed visible beneath the turtle, showing sandy bottom, coral, and seagrass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0925070270895958 + ], + "opposite_logprob_tokens": [ + -2.4675071239471436 + ], + "logprob_reward": -0.0925070270895958, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0925070270895958, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13682805001735687, + 0.045554738491773605, + 0.07956313341856003, + 0.048998720943927765, + 0.2357064038515091, + 0.12525789439678192, + 0.07951799035072327, + 0.13850046694278717, + 0.024087678641080856, + 0.08598491549491882 + ], + "credit_map": [ + 0.0186776090413332, + 0.020206470042467117, + 0.01204295177012682, + 0.018331307917833328, + 0.017174329608678818, + 0.019132304936647415, + 0.011587520129978657, + 0.01967555470764637, + 0.005404360126703978, + 0.005918304435908794, + 0.0032303279731422663, + 0.006040079984813929, + 0.007947946898639202, + 0.006573828402906656, + 0.005241992883384228, + 0.005197896156460047, + 0.008156433701515198, + 0.007834848016500473, + 0.007951725274324417, + 0.014218457974493504, + 0.008957513608038425, + 0.007897606119513512, + 0.012638225220143795, + 0.011908317916095257, + 0.006511823274195194, + 0.00463199894875288, + 0.004106602631509304, + 0.005244528409093618, + 0.007138554472476244, + 0.00557321822270751, + 0.005899867974221706, + 0.009892127476632595, + 0.03143412247300148, + 0.021141396835446358, + 0.021724242717027664, + 0.016040917485952377, + 0.025191405788064003, + 0.03978778049349785, + 0.020909860730171204, + 0.05947666987776756, + 0.01017654687166214, + 0.030114488676190376, + 0.008172323927283287, + 0.01582401990890503, + 0.019064247608184814, + 0.0211909469217062, + 0.011588978581130505, + 0.009126347489655018, + 0.006592697463929653, + 0.01659325882792473, + 0.003912261687219143, + 0.0032367759849876165, + 0.021176936104893684, + 0.00810568779706955, + 0.0057384660467505455, + 0.01416191179305315, + 0.011814371682703495, + 0.026990607380867004, + 0.006635467987507582, + 0.01105405855923891, + 0.0392400287091732, + 0.011985303834080696, + 0.016895996406674385, + 0.013884619809687138, + 0.003368403296917677, + 0.002798563102260232, + 0.0044021643698215485, + 0.0025514618027955294, + 0.002047577640041709, + 0.003025946207344532, + 0.0026497547514736652, + 0.0032438081689178944, + 0.01625980995595455, + 0.007501170504838228, + 0.008615517988801003, + 0.02871408872306347, + 0.0048797111958265305, + 0.005870289634913206, + 0.007712591905146837, + 0.0064317407086491585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.0925070270895958 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a fluid 360-degree dolly shot orbiting the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle's shell mottled with deep brown and amber colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing rays of light (Tyndall effect) and golden lens flare visible in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the seabed visible beneath the turtle, showing sandy bottom, coral, and seagrass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521040737628937 + ], + "opposite_logprob_tokens": [ + -1.5021040439605713 + ], + "logprob_reward": -0.2521040737628937, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521040737628937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1682998687028885, + 0.11948192119598389, + 0.05146820470690727, + 0.04567825049161911, + 0.07904685288667679, + 0.18618710339069366, + 0.12247931212186813, + 0.07087723910808563, + 0.01998976618051529, + 0.13649143278598785 + ], + "credit_map": [ + 0.021476078778505325, + 0.020690055564045906, + 0.017614899203181267, + 0.03640767186880112, + 0.01760462485253811, + 0.02215665951371193, + 0.01722772978246212, + 0.015122161246836185, + 0.01444209460169077, + 0.011907084845006466, + 0.01211217138916254, + 0.02208305522799492, + 0.016179924830794334, + 0.013198740780353546, + 0.019600367173552513, + 0.009958475828170776, + 0.0077047464437782764, + 0.004931304138153791, + 0.005212193820625544, + 0.005806250497698784, + 0.008350124582648277, + 0.006427022162824869, + 0.006859716027975082, + 0.006176843773573637, + 0.006604881025850773, + 0.005772262811660767, + 0.004675284959375858, + 0.0060701374895870686, + 0.0088431341573596, + 0.004648412112146616, + 0.0045915087684988976, + 0.00447262916713953, + 0.011378223076462746, + 0.007566236425191164, + 0.005812454968690872, + 0.0062373485416173935, + 0.009679132141172886, + 0.01166086457669735, + 0.007828899659216404, + 0.01888369210064411, + 0.019096974283456802, + 0.04702051356434822, + 0.016154300421476364, + 0.01396838016808033, + 0.027321338653564453, + 0.028675712645053864, + 0.016778510063886642, + 0.017171375453472137, + 0.009104236029088497, + 0.0156262069940567, + 0.0070194280706346035, + 0.014792018570005894, + 0.015243932604789734, + 0.02216998115181923, + 0.014076520688831806, + 0.024446990340948105, + 0.0049043940380215645, + 0.01566181145608425, + 0.0036319231148809195, + 0.006934440229088068, + 0.016026917845010757, + 0.008371377363801003, + 0.007719697430729866, + 0.007626680191606283, + 0.002790148137137294, + 0.0024740269873291254, + 0.002621775260195136, + 0.003137313760817051, + 0.0013955405447632074, + 0.0027404080610722303, + 0.0021075536496937275, + 0.0027229993138462305, + 0.024032991379499435, + 0.010595409199595451, + 0.01840175688266754, + 0.04739408567547798, + 0.007851868867874146, + 0.00941692665219307, + 0.010883957147598267, + 0.007914425805211067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.2521040737628937 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a fluid 360-degree dolly shot orbiting the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle's shell mottled with deep brown and amber colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing rays of light (Tyndall effect) and golden lens flare visible in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the seabed visible beneath the turtle, showing sandy bottom, coral, and seagrass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864751175045967 + ], + "opposite_logprob_tokens": [ + -3.048647403717041 + ], + "logprob_reward": -0.04864751175045967, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864751175045967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1545523852109909, + 0.06427759677171707, + 0.04097762703895569, + 0.042687106877565384, + 0.17798413336277008, + 0.058911651372909546, + 0.15084995329380035, + 0.13951599597930908, + 0.05893830209970474, + 0.11130522936582565 + ], + "credit_map": [ + 0.01849391683936119, + 0.019838087260723114, + 0.014376768842339516, + 0.029575634747743607, + 0.018163492903113365, + 0.01915329322218895, + 0.0217256061732769, + 0.013225582428276539, + 0.00761355459690094, + 0.006490130443125963, + 0.006295923609286547, + 0.013809069991111755, + 0.007412880193442106, + 0.006893886253237724, + 0.010212349705398083, + 0.005549799185246229, + 0.004986096173524857, + 0.006103727035224438, + 0.004006409086287022, + 0.00531885726377368, + 0.0051447199657559395, + 0.004547560587525368, + 0.0062633235938847065, + 0.004606933332979679, + 0.008505790494382381, + 0.004506330005824566, + 0.0037823531311005354, + 0.006238279398530722, + 0.004150807857513428, + 0.004466860089451075, + 0.005901702214032412, + 0.00513498717918992, + 0.022499805316329002, + 0.015598746947944164, + 0.01801430992782116, + 0.012469992972910404, + 0.0147952726110816, + 0.030514689162373543, + 0.016925886273384094, + 0.047165412455797195, + 0.004907997325062752, + 0.008038926869630814, + 0.0054077100940048695, + 0.00907246582210064, + 0.010969782248139381, + 0.005294729024171829, + 0.008743318729102612, + 0.006476723589003086, + 0.010436132550239563, + 0.03765835613012314, + 0.006160346791148186, + 0.012999491766095161, + 0.04648575186729431, + 0.016506381332874298, + 0.010582915507256985, + 0.010020575486123562, + 0.01169090811163187, + 0.027710793539881706, + 0.007869870401918888, + 0.012898939661681652, + 0.036661725491285324, + 0.015112287364900112, + 0.015554480254650116, + 0.012016995809972286, + 0.008960351347923279, + 0.008259986527264118, + 0.006577557884156704, + 0.011650267988443375, + 0.004446767270565033, + 0.00531360786408186, + 0.007634406443685293, + 0.006095357239246368, + 0.02173045463860035, + 0.008276759646832943, + 0.010538879781961441, + 0.041968293488025665, + 0.005984220188111067, + 0.007228064816445112, + 0.008898044005036354, + 0.006680514197796583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.04864751175045967 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a fluid 360-degree dolly shot orbiting the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle's shell mottled with deep brown and amber colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing rays of light (Tyndall effect) and golden lens flare visible in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the seabed visible beneath the turtle, showing sandy bottom, coral, and seagrass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537354603409767 + ], + "opposite_logprob_tokens": [ + -4.387537479400635 + ], + "logprob_reward": -0.012537354603409767, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537354603409767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13889841735363007, + 0.08563733100891113, + 0.03648524731397629, + 0.04826763644814491, + 0.09227077662944794, + 0.09834565222263336, + 0.1476828008890152, + 0.07474897801876068, + 0.11066650599241257, + 0.16699665784835815 + ], + "credit_map": [ + 0.023560233414173126, + 0.017454778775572777, + 0.01266347523778677, + 0.025485137477517128, + 0.013441754505038261, + 0.01841879077255726, + 0.015097582712769508, + 0.012776663526892662, + 0.013465593568980694, + 0.009379792958498001, + 0.008448004722595215, + 0.009813373908400536, + 0.011245248839259148, + 0.012480834499001503, + 0.008695253171026707, + 0.012109235860407352, + 0.004406825639307499, + 0.004095104988664389, + 0.0033898523543030024, + 0.0035097470972687006, + 0.006520428694784641, + 0.004967357963323593, + 0.0035810270346701145, + 0.006014903075993061, + 0.006236877758055925, + 0.004844051320105791, + 0.004463822580873966, + 0.01023513451218605, + 0.00615460891276598, + 0.005206127651035786, + 0.0071973358280956745, + 0.003929679282009602, + 0.00999657716602087, + 0.010099698789417744, + 0.008531961590051651, + 0.007972121238708496, + 0.01316707395017147, + 0.015485678799450397, + 0.010049840435385704, + 0.016967827454209328, + 0.010171248577535152, + 0.020346641540527344, + 0.006471435539424419, + 0.008890711702406406, + 0.017511717975139618, + 0.016025112941861153, + 0.009035808965563774, + 0.009892978705465794, + 0.014925310388207436, + 0.02983277663588524, + 0.005513708107173443, + 0.005664662923663855, + 0.03994941711425781, + 0.014007765799760818, + 0.009647025726735592, + 0.028142133727669716, + 0.005670811049640179, + 0.014451066963374615, + 0.0038836696185171604, + 0.005410539917647839, + 0.020837834104895592, + 0.008191453292965889, + 0.008766856975853443, + 0.007536750286817551, + 0.016401933506131172, + 0.016157755628228188, + 0.01395876333117485, + 0.013273470103740692, + 0.008553596213459969, + 0.018169919028878212, + 0.010947578586637974, + 0.013203496113419533, + 0.028830721974372864, + 0.020379001274704933, + 0.022727010771632195, + 0.030728472396731377, + 0.010432332754135132, + 0.02386508136987686, + 0.01542369369417429, + 0.014610359445214272 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.012537354603409767 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a fluid 360-degree dolly shot orbiting the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle's shell mottled with deep brown and amber colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing rays of light (Tyndall effect) and golden lens flare visible in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the seabed visible beneath the turtle, showing sandy bottom, coral, and seagrass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520769238471985 + ], + "opposite_logprob_tokens": [ + -1.5020768642425537 + ], + "logprob_reward": -0.2520769238471985, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520769238471985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06467710435390472, + 0.10051992535591125, + 0.02785501629114151, + 0.05779881030321121, + 0.11030012369155884, + 0.0521046444773674, + 0.18570728600025177, + 0.06249597296118736, + 0.048476263880729675, + 0.29006484150886536 + ], + "credit_map": [ + 0.01103993784636259, + 0.010318972170352936, + 0.004358504433184862, + 0.011041217483580112, + 0.006178097799420357, + 0.009103965014219284, + 0.0060852463357150555, + 0.006551163271069527, + 0.018062429502606392, + 0.011376133188605309, + 0.009782603941857815, + 0.011040139943361282, + 0.014861463569104671, + 0.013416928239166737, + 0.009091980755329132, + 0.012888239696621895, + 0.003244879888370633, + 0.0036016968078911304, + 0.0026110068429261446, + 0.0038700606673955917, + 0.004443338140845299, + 0.0034637225326150656, + 0.003379875561222434, + 0.0032404367811977863, + 0.011229036375880241, + 0.006772897206246853, + 0.005209372844547033, + 0.009167423471808434, + 0.0077941338531672955, + 0.006024939939379692, + 0.0067980638705194, + 0.004802937153726816, + 0.021792542189359665, + 0.011662907898426056, + 0.009134401567280293, + 0.011363361030817032, + 0.017126457765698433, + 0.013821426779031754, + 0.012996376492083073, + 0.012402651831507683, + 0.004619610961526632, + 0.011841800063848495, + 0.003201351733878255, + 0.00659447256475687, + 0.006821799091994762, + 0.00962590891867876, + 0.005207779351621866, + 0.00419192248955369, + 0.014138881117105484, + 0.04895791411399841, + 0.006835492327809334, + 0.0048094745725393295, + 0.06509502977132797, + 0.015118762850761414, + 0.010077587328851223, + 0.02067413553595543, + 0.0049741934053599834, + 0.014382346533238888, + 0.0029243764001876116, + 0.002885704394429922, + 0.021079860627651215, + 0.0057616750709712505, + 0.006444977596402168, + 0.004042843356728554, + 0.006313864141702652, + 0.006319880951195955, + 0.006795049179345369, + 0.005937421694397926, + 0.004650152754038572, + 0.008382231928408146, + 0.004721543285995722, + 0.005356119945645332, + 0.0534072145819664, + 0.034810859709978104, + 0.03644581884145737, + 0.0628809928894043, + 0.0169297456741333, + 0.03347368165850639, + 0.027209976688027382, + 0.024906568229198456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 4 + ], + "selected_reward": -0.2520769238471985 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fluid 360-degree dolly shot, held parallel to the subject, gracefully orbits a giant green sea turtle. It glides effortlessly through the crystal-clear, endless expanse of tropical water. Bright sunlight filters through the shimmering water, creating dancing, shimmering rays of light (the Tyndall effect) against the deep blue background. As the camera smoothly pans, these rays create a gorgeous, warm lens flare, with flecks of golden light softly blooming and shifting across the screen. The turtle is enormous, its massive shell, mottled with deep brown and amber, reflecting a soft luster, revealing the marks of age. Using its massive, powerful front flippers like wings, it makes slow, steady, and deliberate strokes. Each stroke is imbued with graceful power, propelling its massive body through the water like a ballerina. Beneath it, the pristine, sandy seabed is clearly visible, dotted with clumps of vibrant coral and swaying seagrass. The entire scene is dominated by a tranquil blue color scheme, illuminated by golden sunlight penetrating the water, creating a peaceful, magnificent and almost dreamy atmosphere, making people feel the peace of life and the awe-inspiring beauty of nature.", + "seed": 9497, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_13/index_5339_seed_9497_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.37495735315606, + "vlm_logprob": -0.37495735315606, + "vlm_yes_no_margin": 2.575, + "vlm_token_credit": -0.37495735315606, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a fluid 360-degree dolly shot orbiting the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle's shell mottled with deep brown and amber colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing rays of light (Tyndall effect) and golden lens flare visible in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the seabed visible beneath the turtle, showing sandy bottom, coral, and seagrass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02854011580348015 + ], + "opposite_logprob_tokens": [ + -3.6535401344299316 + ], + "logprob_reward": -0.02854011580348015, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02854011580348015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.085617795586586, + 0.05223263055086136, + 0.09935135394334793, + 0.43750572204589844, + 0.1019829511642456, + 0.02313283085823059, + 0.0744127407670021, + 0.051610734313726425, + 0.028945939615368843, + 0.04520731791853905 + ], + "credit_map": [ + 0.015246578492224216, + 0.012974441982805729, + 0.00562661699950695, + 0.007568432949483395, + 0.014435815624892712, + 0.009479355067014694, + 0.010442934930324554, + 0.009843617677688599, + 0.009149307385087013, + 0.007589834742248058, + 0.00467674620449543, + 0.005755054298788309, + 0.005985284689813852, + 0.005953997373580933, + 0.00573159521445632, + 0.0073908064514398575, + 0.015696844086050987, + 0.014861037023365498, + 0.011969988234341145, + 0.008441271260380745, + 0.011262457817792892, + 0.02048477903008461, + 0.007684782613068819, + 0.008950197137892246, + 0.02796514332294464, + 0.010906673036515713, + 0.1720074713230133, + 0.021095842123031616, + 0.01471630297601223, + 0.09227891266345978, + 0.027661839500069618, + 0.07087349146604538, + 0.009315277449786663, + 0.004517683293670416, + 0.028191911056637764, + 0.0062219914980232716, + 0.008004370145499706, + 0.03351980820298195, + 0.006194983143359423, + 0.006016929168254137, + 0.002851952100172639, + 0.003645743476226926, + 0.002804596908390522, + 0.0028290902264416218, + 0.0019937024917453527, + 0.00271798437461257, + 0.0028844145126640797, + 0.003405344672501087, + 0.009959190152585506, + 0.016266215592622757, + 0.008861619979143143, + 0.00518269557505846, + 0.01140095666050911, + 0.00759879918769002, + 0.006714787799865007, + 0.008428477682173252, + 0.0052838572300970554, + 0.006587791256606579, + 0.007764211855828762, + 0.006232427898794413, + 0.00886520091444254, + 0.006904971320182085, + 0.005318684969097376, + 0.00465358654037118, + 0.003154039615765214, + 0.003973307553678751, + 0.0041606430895626545, + 0.005464993417263031, + 0.00349810509942472, + 0.0026589520275592804, + 0.0025062214117497206, + 0.0035296787973493338, + 0.005234710872173309, + 0.007399597205221653, + 0.006370370276272297, + 0.005473403725773096, + 0.006239011883735657, + 0.00595466373488307, + 0.004002564586699009, + 0.004532997030764818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 2 + ], + "selected_reward": -0.02854011580348015 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a fluid 360-degree dolly shot orbiting the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle's shell mottled with deep brown and amber colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing rays of light (Tyndall effect) and golden lens flare visible in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the seabed visible beneath the turtle, showing sandy bottom, coral, and seagrass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.6005152463912964 + ], + "opposite_logprob_tokens": [ + -0.2255152016878128 + ], + "logprob_reward": -1.6005152463912964, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.6005152463912964, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08206123113632202, + 0.0505731925368309, + 0.11142705380916595, + 0.20305290818214417, + 0.058592118322849274, + 0.1183118149638176, + 0.01822182908654213, + 0.07605384290218353, + 0.24643723666667938, + 0.03526874631643295 + ], + "credit_map": [ + 0.005632045678794384, + 0.007368459831923246, + 0.00515340780839324, + 0.005907784681767225, + 0.018350642174482346, + 0.012204708531498909, + 0.011909057386219501, + 0.015535120852291584, + 0.007194361183792353, + 0.006281855516135693, + 0.005921392701566219, + 0.006834045052528381, + 0.005145520903170109, + 0.006812342908233404, + 0.006286471616476774, + 0.006097205448895693, + 0.01988585665822029, + 0.013985871337354183, + 0.013771014288067818, + 0.008012830279767513, + 0.013852237723767757, + 0.02445129118859768, + 0.007853199727833271, + 0.009614754468202591, + 0.015744134783744812, + 0.00799785926938057, + 0.07420073449611664, + 0.010543402284383774, + 0.007733864244073629, + 0.02752784453332424, + 0.012793012894690037, + 0.04651205986738205, + 0.012795533053576946, + 0.007362001575529575, + 0.007341734599322081, + 0.006934379693120718, + 0.0062364921905100346, + 0.006920081563293934, + 0.005701850168406963, + 0.005300044547766447, + 0.01995510794222355, + 0.018889961764216423, + 0.013756608590483665, + 0.016840394586324692, + 0.015730194747447968, + 0.011958474293351173, + 0.01115411426872015, + 0.010026964358985424, + 0.0027414364740252495, + 0.0025952591095119715, + 0.0028309612534940243, + 0.0019829212687909603, + 0.0017576668178662658, + 0.002265784190967679, + 0.0020686578936874866, + 0.0019791428931057453, + 0.009146803058683872, + 0.014543996192514896, + 0.011759670451283455, + 0.008146608248353004, + 0.007220238912850618, + 0.008991319686174393, + 0.008510219864547253, + 0.007734990678727627, + 0.03530113399028778, + 0.03237137943506241, + 0.045493099838495255, + 0.06672564893960953, + 0.014988955110311508, + 0.019717508926987648, + 0.01417060848325491, + 0.017668887972831726, + 0.004033999051898718, + 0.005088942125439644, + 0.005150034092366695, + 0.0038866179529577494, + 0.005005618557333946, + 0.004164012148976326, + 0.003676398890092969, + 0.00426312442868948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 5 + ], + "selected_reward": -1.6005152463912964 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a fluid 360-degree dolly shot orbiting the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle's shell mottled with deep brown and amber colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing rays of light (Tyndall effect) and golden lens flare visible in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the seabed visible beneath the turtle, showing sandy bottom, coral, and seagrass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012544183060526848 + ], + "opposite_logprob_tokens": [ + -4.38754415512085 + ], + "logprob_reward": -0.012544183060526848, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012544183060526848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058642592281103134, + 0.05947759374976158, + 0.10834100097417831, + 0.21819624304771423, + 0.08293112367391586, + 0.07058626413345337, + 0.0371103398501873, + 0.14899030327796936, + 0.17403578758239746, + 0.04168875142931938 + ], + "credit_map": [ + 0.00580806378275156, + 0.005122663453221321, + 0.004578791558742523, + 0.005183591041713953, + 0.009190537966787815, + 0.00813609641045332, + 0.009575905278325081, + 0.011046946980059147, + 0.006592989433556795, + 0.007620575372129679, + 0.004229977261275053, + 0.0062492843717336655, + 0.008059203624725342, + 0.006376984994858503, + 0.00785114150494337, + 0.0124974325299263, + 0.013970536179840565, + 0.01494604256004095, + 0.011976365931332111, + 0.011094149202108383, + 0.01103699766099453, + 0.016978956758975983, + 0.010556469671428204, + 0.017781484872102737, + 0.015550966374576092, + 0.007845854386687279, + 0.08154811710119247, + 0.011176944710314274, + 0.007150031626224518, + 0.032583627849817276, + 0.013991246931254864, + 0.04834945127367973, + 0.014092576690018177, + 0.007671097759157419, + 0.015084792859852314, + 0.007743823807686567, + 0.007556849159300327, + 0.016571378335356712, + 0.0073668695986270905, + 0.00684373639523983, + 0.009591563604772091, + 0.010644631460309029, + 0.009534922428429127, + 0.0066834017634391785, + 0.007286843843758106, + 0.006164225749671459, + 0.01221330463886261, + 0.008467367850244045, + 0.0057702804915606976, + 0.005313686560839415, + 0.004775770008563995, + 0.004017112776637077, + 0.0043761106207966805, + 0.0034554421436041594, + 0.00525765772908926, + 0.004144276957958937, + 0.014957006089389324, + 0.02831057645380497, + 0.016595063731074333, + 0.016479315236210823, + 0.025438586249947548, + 0.021141143515706062, + 0.013950187712907791, + 0.012118428014218807, + 0.026175212115049362, + 0.02171618677675724, + 0.03489373251795769, + 0.04869335889816284, + 0.00979525689035654, + 0.012372903525829315, + 0.009528187103569508, + 0.01086095068603754, + 0.008836072869598866, + 0.005230461247265339, + 0.006372862961143255, + 0.004221779759973288, + 0.005046452861279249, + 0.004747427999973297, + 0.0035372553393244743, + 0.0036964404862374067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 7 + ], + "selected_reward": -0.012544183060526848 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a fluid 360-degree dolly shot orbiting the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle's shell mottled with deep brown and amber colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing rays of light (Tyndall effect) and golden lens flare visible in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the seabed visible beneath the turtle, showing sandy bottom, coral, and seagrass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630250882357359 + ], + "opposite_logprob_tokens": [ + -4.882630348205566 + ], + "logprob_reward": -0.007630250882357359, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630250882357359, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1230166107416153, + 0.07197214663028717, + 0.11089281737804413, + 0.15801547467708588, + 0.0656203031539917, + 0.10939855128526688, + 0.05840982496738434, + 0.09262719005346298, + 0.134946808218956, + 0.07510033994913101 + ], + "credit_map": [ + 0.00968161504715681, + 0.009317845106124878, + 0.011288667097687721, + 0.010972077958285809, + 0.022767236456274986, + 0.019577834755182266, + 0.016712389886379242, + 0.022698935121297836, + 0.006616012658923864, + 0.008106289431452751, + 0.005992542020976543, + 0.007692377083003521, + 0.010845139622688293, + 0.008161342702805996, + 0.009064766578376293, + 0.015493674203753471, + 0.018668638542294502, + 0.014785829000175, + 0.010557777248322964, + 0.00886285025626421, + 0.013505428098142147, + 0.020159143954515457, + 0.008684213273227215, + 0.015668924897909164, + 0.014790212735533714, + 0.008440445177257061, + 0.0534694604575634, + 0.009661721996963024, + 0.009788264520466328, + 0.020068969577550888, + 0.009918677620589733, + 0.03187771141529083, + 0.012778894044458866, + 0.00791130866855383, + 0.009708006866276264, + 0.006678821984678507, + 0.006410414353013039, + 0.008399034850299358, + 0.008079731836915016, + 0.005654081702232361, + 0.015897192060947418, + 0.013416099362075329, + 0.013130205683410168, + 0.0101735545322299, + 0.009843122214078903, + 0.011859380640089512, + 0.017213156446814537, + 0.017865823581814766, + 0.009162677451968193, + 0.009885772131383419, + 0.0099702263250947, + 0.004005523398518562, + 0.006029958371073008, + 0.007883100770413876, + 0.004239481873810291, + 0.00723308278247714, + 0.009933211840689182, + 0.01708744838833809, + 0.01050337590277195, + 0.010594774968922138, + 0.015604226849973202, + 0.011003305204212666, + 0.00945968832820654, + 0.008441148325800896, + 0.0192749984562397, + 0.01697777397930622, + 0.022728150710463524, + 0.034833140671253204, + 0.009962761774659157, + 0.011616822332143784, + 0.008932114578783512, + 0.01062103547155857, + 0.015858937054872513, + 0.008886143565177917, + 0.011957266367971897, + 0.007017876021564007, + 0.009334942325949669, + 0.008772688917815685, + 0.007111856248229742, + 0.006160615012049675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.007630250882357359 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a fluid 360-degree dolly shot orbiting the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle's shell mottled with deep brown and amber colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing rays of light (Tyndall effect) and golden lens flare visible in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the seabed visible beneath the turtle, showing sandy bottom, coral, and seagrass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22555696964263916 + ], + "opposite_logprob_tokens": [ + -1.6005569696426392 + ], + "logprob_reward": -0.22555696964263916, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22555696964263916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062645323574543, + 0.1315164566040039, + 0.0320168100297451, + 0.4930697977542877, + 0.02447197586297989, + 0.0573769211769104, + 0.019134098663926125, + 0.01751154102385044, + 0.050271254032850266, + 0.11198575049638748 + ], + "credit_map": [ + 0.007320864591747522, + 0.006282540503889322, + 0.006184495519846678, + 0.004957007709890604, + 0.005845007952302694, + 0.007410796359181404, + 0.00819216575473547, + 0.016452448442578316, + 0.011185034178197384, + 0.015559366904199123, + 0.008089529350399971, + 0.012163023464381695, + 0.022923629730939865, + 0.015714731067419052, + 0.016146177425980568, + 0.029734976589679718, + 0.005474058445543051, + 0.003972701262682676, + 0.003841482102870941, + 0.002582930028438568, + 0.0029686575289815664, + 0.00646246736869216, + 0.0028506743256002665, + 0.0038638394325971603, + 0.03138912096619606, + 0.01260873582214117, + 0.1915232092142105, + 0.0237326268106699, + 0.017853153869509697, + 0.103643499314785, + 0.030361024662852287, + 0.08195845037698746, + 0.0031903046183288097, + 0.0022137626074254513, + 0.00449048075824976, + 0.0019575683400034904, + 0.001872256281785667, + 0.005431696772575378, + 0.003204621374607086, + 0.0021112882532179356, + 0.004790476057678461, + 0.005819334182888269, + 0.008308125659823418, + 0.006553163751959801, + 0.005564992316067219, + 0.004580201115459204, + 0.015163120813667774, + 0.00659751333296299, + 0.002163519151508808, + 0.002112660091370344, + 0.002759568626061082, + 0.002080450067296624, + 0.0019898132886737585, + 0.0021519784349948168, + 0.003136980114504695, + 0.0027391312178224325, + 0.002067459048703313, + 0.0031277616508305073, + 0.0026326309889554977, + 0.0020112250931560993, + 0.002604866400361061, + 0.0019797070417553186, + 0.0015110535314306617, + 0.0015768397133797407, + 0.00718621676787734, + 0.005630409810692072, + 0.010354915633797646, + 0.013726070523262024, + 0.003309165593236685, + 0.0033663297072052956, + 0.0029635890386998653, + 0.003734554862603545, + 0.02645283192396164, + 0.016255555674433708, + 0.018578477203845978, + 0.010548154823482037, + 0.01231144554913044, + 0.011806529015302658, + 0.008064956404268742, + 0.007967810146510601 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 9 + ], + "selected_reward": -0.22555696964263916 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_14/index_tensor([3923, 3923])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_14/index_tensor([3923, 3923])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ab6e9843a671a7cea1afb2c1f1e131461b598805 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_14/index_tensor([3923, 3923])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a single, unbroken shot, the camera begins with a static, top-down close-up of a sprawling blueprint spread across a dark, oil-stained wooden workbench. A maze of fine, cyan lines covers the crisp white construction paper, depicting the intricate schematics of a complex machine, complete with detailed gears, pistons, and wiring diagrams. Without any visible intervention, the blueprint begins to magically fold itself; sharp, precise creases form automatically, and the paper bends and folds with a life of its own, methodically transforming the complex 2D drawing into a sleek paper airplane. When the final fold is complete, the camera smoothly pulls back and begins tracking as the paper airplane, its wings now adorned with the fragmented schematics, gently lifts from the workbench and takes flight. It begins a graceful, soaring journey through a cluttered yet orderly inventor's workshop, the camera following seamlessly as the aircraft orbits a hanging industrial lamp, its warm glow illuminating the dancing dust motes in the air. It glides over half-assembled machinery, past pegboards neatly arranged with gleaming metal tools, swoops low over a metal lathe, and then rises again, navigating the complexities of space with effortless grace. The entire scene is bathed in the soft golden glow of afternoon sunlight, which filters through a large, grimy window, casting long shadows and creating an atmosphere of quiet wonder and dormant creative potential.", + "seed": 31655, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_14/index_3923_seed_31655_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.011560628376901149, + "vlm_logprob": -0.011560628376901149, + "vlm_yes_no_margin": 4.9, + "vlm_token_credit": -0.011560628376901149, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, top-down close-up of a blueprint on a dark, oil-stained wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the blueprint magically fold itself into a sleek paper airplane without visible human intervention?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back and track the paper airplane as it takes flight and orbits a hanging industrial lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an inventor's workshop containing half-assembled machinery, pegboards with metal tools, and a metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight filtering through a large, grimy window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007309716194868088 + ], + "opposite_logprob_tokens": [ + -5.757309913635254 + ], + "logprob_reward": -0.007309716194868088, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.007309716194868088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3355419933795929, + 0.05547000840306282, + 0.06137673184275627, + 0.07731590420007706, + 0.09488846361637115, + 0.06965915113687515, + 0.0479855053126812, + 0.06028164550662041, + 0.11665621399879456, + 0.0808243528008461 + ], + "credit_map": [ + 0.03289937227964401, + 0.047924965620040894, + 0.02687837742269039, + 0.0477704219520092, + 0.04230459779500961, + 0.025198129937052727, + 0.05772911384701729, + 0.05483702942728996, + 0.006940734572708607, + 0.009570545516908169, + 0.00620948476716876, + 0.0075440481305122375, + 0.0059159318916499615, + 0.00789379607886076, + 0.0040839556604623795, + 0.007311515975743532, + 0.010875129140913486, + 0.00546820554882288, + 0.014199340716004372, + 0.008912643417716026, + 0.0050172037445008755, + 0.0033412750344723463, + 0.0059033893048763275, + 0.007659544702619314, + 0.023067956790328026, + 0.003922515548765659, + 0.022476790472865105, + 0.0046892594546079636, + 0.002866422990337014, + 0.0077092149294912815, + 0.00471955630928278, + 0.007864189334213734, + 0.013339824043214321, + 0.013956652022898197, + 0.013824907131493092, + 0.012978377752006054, + 0.008403975516557693, + 0.011706535704433918, + 0.010887213982641697, + 0.009790981188416481, + 0.008920024149119854, + 0.005497321020811796, + 0.009617095813155174, + 0.009126763790845871, + 0.00630622822791338, + 0.009466045536100864, + 0.0060004256665706635, + 0.014725247398018837, + 0.009736530482769012, + 0.007576607167720795, + 0.007449499797075987, + 0.005361597053706646, + 0.005779299419373274, + 0.003936279099434614, + 0.004318050108850002, + 0.0038276431150734425, + 0.006155557930469513, + 0.006718122866004705, + 0.012049464508891106, + 0.006852957885712385, + 0.010127653367817402, + 0.006202575284987688, + 0.007558017503470182, + 0.004617294762283564, + 0.018616458401083946, + 0.015910856425762177, + 0.014677729457616806, + 0.011604244820773602, + 0.013780955225229263, + 0.013517364859580994, + 0.014984793029725552, + 0.013563811779022217, + 0.010028488002717495, + 0.00806923396885395, + 0.010021342895925045, + 0.008343778550624847, + 0.011613273061811924, + 0.008498686365783215, + 0.009417051449418068, + 0.01483250129967928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.007309716194868088 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, top-down close-up of a blueprint on a dark, oil-stained wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the blueprint magically fold itself into a sleek paper airplane without visible human intervention?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back and track the paper airplane as it takes flight and orbits a hanging industrial lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an inventor's workshop containing half-assembled machinery, pegboards with metal tools, and a metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight filtering through a large, grimy window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02983858808875084 + ], + "opposite_logprob_tokens": [ + -3.5298385620117188 + ], + "logprob_reward": -0.02983858808875084, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02983858808875084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1858234852552414, + 0.06877530366182327, + 0.09217046201229095, + 0.09161540120840073, + 0.17341729998588562, + 0.14614783227443695, + 0.052713192999362946, + 0.0740174800157547, + 0.042544424533843994, + 0.07277509570121765 + ], + "credit_map": [ + 0.026471182703971863, + 0.026844816282391548, + 0.016048705205321312, + 0.03194728493690491, + 0.027753416448831558, + 0.01416043657809496, + 0.020979544147849083, + 0.021618105471134186, + 0.008252923376858234, + 0.01158055942505598, + 0.006859154440462589, + 0.009808569215238094, + 0.00827640201896429, + 0.010520475916564465, + 0.004760855343192816, + 0.008716372773051262, + 0.018152566626667976, + 0.00790970679372549, + 0.013504335656762123, + 0.018040675669908524, + 0.0067216078750789165, + 0.0063235824927687645, + 0.01123922597616911, + 0.01027876790612936, + 0.020949924364686012, + 0.008631745353341103, + 0.015590665861964226, + 0.008113006129860878, + 0.006510936189442873, + 0.011366932652890682, + 0.008296501822769642, + 0.012155694887042046, + 0.028825052082538605, + 0.020019112154841423, + 0.02620808593928814, + 0.02595084346830845, + 0.013295143842697144, + 0.02328347973525524, + 0.0220663920044899, + 0.013769215904176235, + 0.01912517286837101, + 0.011530878022313118, + 0.01745653711259365, + 0.027751339599490166, + 0.01683482713997364, + 0.013446950353682041, + 0.01771799474954605, + 0.02228415384888649, + 0.010537618771195412, + 0.010293298400938511, + 0.007594163063913584, + 0.0048702312633395195, + 0.006907455623149872, + 0.005140970926731825, + 0.003675279440358281, + 0.0036941824946552515, + 0.007537607569247484, + 0.007499498315155506, + 0.016126099973917007, + 0.008613495156168938, + 0.010382344014942646, + 0.009472579695284367, + 0.009129269048571587, + 0.005256598349660635, + 0.007514298427850008, + 0.006931567098945379, + 0.004752784036099911, + 0.0045114727690815926, + 0.004502103663980961, + 0.004427017644047737, + 0.004771771840751171, + 0.005133415572345257, + 0.011901899240911007, + 0.009010988287627697, + 0.008572492748498917, + 0.006675065960735083, + 0.010249239392578602, + 0.007803644984960556, + 0.007754348684102297, + 0.010807420127093792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.02983858808875084 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, top-down close-up of a blueprint on a dark, oil-stained wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the blueprint magically fold itself into a sleek paper airplane without visible human intervention?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back and track the paper airplane as it takes flight and orbits a hanging industrial lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an inventor's workshop containing half-assembled machinery, pegboards with metal tools, and a metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight filtering through a large, grimy window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003624657168984413 + ], + "opposite_logprob_tokens": [ + -5.628624439239502 + ], + "logprob_reward": -0.003624657168984413, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003624657168984413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16234196722507477, + 0.05336839333176613, + 0.14495344460010529, + 0.08161959052085876, + 0.22032122313976288, + 0.05894697830080986, + 0.06335487216711044, + 0.12733320891857147, + 0.040308814495801926, + 0.04745153710246086 + ], + "credit_map": [ + 0.02073497511446476, + 0.024598045274615288, + 0.014025776647031307, + 0.02783229947090149, + 0.02318066544830799, + 0.011630063876509666, + 0.019191598519682884, + 0.02114853449165821, + 0.008500189520418644, + 0.010838055983185768, + 0.004927699454128742, + 0.00655505433678627, + 0.00602369848638773, + 0.006461388431489468, + 0.004532168619334698, + 0.005530142690986395, + 0.02702309750020504, + 0.008685479871928692, + 0.01818639039993286, + 0.03400179371237755, + 0.010939521715044975, + 0.008154494687914848, + 0.01871360093355179, + 0.01924905739724636, + 0.024805882945656776, + 0.004839593544602394, + 0.026928270235657692, + 0.005332372151315212, + 0.0032138023525476456, + 0.006868262775242329, + 0.0033684661611914635, + 0.006262934301048517, + 0.02880672924220562, + 0.015292706899344921, + 0.0329221710562706, + 0.024107009172439575, + 0.01459325011819601, + 0.05567478761076927, + 0.016437947750091553, + 0.03248663246631622, + 0.006936433259397745, + 0.005198195111006498, + 0.010210726410150528, + 0.006127059925347567, + 0.0071206362918019295, + 0.0072548119351267815, + 0.004069577902555466, + 0.012029536068439484, + 0.010974385775625706, + 0.011095128022134304, + 0.01051276084035635, + 0.006977390497922897, + 0.007644409779459238, + 0.004999002907425165, + 0.006050210911780596, + 0.005101583898067474, + 0.013238917104899883, + 0.009079722687602043, + 0.019151199609041214, + 0.012140152975916862, + 0.03017701767385006, + 0.02000737562775612, + 0.013579207472503185, + 0.009959620423614979, + 0.005308420397341251, + 0.00549322459846735, + 0.0054569849744439125, + 0.0037812101654708385, + 0.005207874346524477, + 0.005485002417117357, + 0.0049114348366856575, + 0.004664664622396231, + 0.0063817137852311134, + 0.0073219346813857555, + 0.005183184985071421, + 0.006229904014617205, + 0.006529515143483877, + 0.004061501938849688, + 0.005889774765819311, + 0.0058540101163089275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.003624657168984413 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, top-down close-up of a blueprint on a dark, oil-stained wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the blueprint magically fold itself into a sleek paper airplane without visible human intervention?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back and track the paper airplane as it takes flight and orbits a hanging industrial lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an inventor's workshop containing half-assembled machinery, pegboards with metal tools, and a metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight filtering through a large, grimy window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005953199230134487 + ], + "opposite_logprob_tokens": [ + -5.130953311920166 + ], + "logprob_reward": -0.005953199230134487, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005953199230134487, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2096892148256302, + 0.04296497255563736, + 0.07743111997842789, + 0.051231808960437775, + 0.11293701827526093, + 0.07450152933597565, + 0.05395165830850601, + 0.1403740495443344, + 0.08763550966978073, + 0.14928311109542847 + ], + "credit_map": [ + 0.02494201809167862, + 0.029656019061803818, + 0.018787678331136703, + 0.03382068872451782, + 0.02552870661020279, + 0.015282823704183102, + 0.028602786362171173, + 0.03306848928332329, + 0.007225220091640949, + 0.00818436685949564, + 0.003708293428644538, + 0.005383751355111599, + 0.00458156131207943, + 0.0053507667034864426, + 0.003988378681242466, + 0.004542634356766939, + 0.014094952493906021, + 0.004578389227390289, + 0.013581089675426483, + 0.014866611920297146, + 0.00639948109164834, + 0.004277900326997042, + 0.011229096911847591, + 0.008403608575463295, + 0.010739371180534363, + 0.004771007690578699, + 0.011570660397410393, + 0.005013700574636459, + 0.0032000357750803232, + 0.0060455272905528545, + 0.004899256397038698, + 0.004992251750081778, + 0.014596582390367985, + 0.01320794876664877, + 0.01686486043035984, + 0.015871267765760422, + 0.010248239152133465, + 0.01566903665661812, + 0.014776165597140789, + 0.011702913790941238, + 0.010439267382025719, + 0.005359064321964979, + 0.009343708865344524, + 0.012243294157087803, + 0.00752132385969162, + 0.00955931469798088, + 0.007786726579070091, + 0.012248835526406765, + 0.009890630841255188, + 0.009575794450938702, + 0.008236441761255264, + 0.004448062740266323, + 0.006249999161809683, + 0.00529105169698596, + 0.0053198449313640594, + 0.004939831327646971, + 0.014872942119836807, + 0.016139091923832893, + 0.02348240464925766, + 0.017505059018731117, + 0.03314594924449921, + 0.010729317553341389, + 0.01632361300289631, + 0.008175669237971306, + 0.013305896893143654, + 0.007624683436006308, + 0.014124655164778233, + 0.010145041160285473, + 0.009703502058982849, + 0.00790973287075758, + 0.011286905966699123, + 0.0135350888594985, + 0.015897627919912338, + 0.02079644985496998, + 0.01685374043881893, + 0.021702982485294342, + 0.018937887623906136, + 0.011441564187407494, + 0.02555050142109394, + 0.018102360889315605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.005953199230134487 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, top-down close-up of a blueprint on a dark, oil-stained wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the blueprint magically fold itself into a sleek paper airplane without visible human intervention?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back and track the paper airplane as it takes flight and orbits a hanging industrial lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an inventor's workshop containing half-assembled machinery, pegboards with metal tools, and a metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight filtering through a large, grimy window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076981201767921 + ], + "opposite_logprob_tokens": [ + -4.511076927185059 + ], + "logprob_reward": -0.011076981201767921, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076981201767921, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06666892766952515, + 0.03583786264061928, + 0.04729966074228287, + 0.09146510809659958, + 0.11498142778873444, + 0.11305202543735504, + 0.142752006649971, + 0.2751222252845764, + 0.03859977051615715, + 0.07422099262475967 + ], + "credit_map": [ + 0.009346667677164078, + 0.009930696338415146, + 0.005113192833960056, + 0.011687264777719975, + 0.009089785628020763, + 0.004170645494014025, + 0.013399209827184677, + 0.003931465093046427, + 0.004712226800620556, + 0.006434593349695206, + 0.0027266275137662888, + 0.0036845533177256584, + 0.003979808185249567, + 0.0040606120601296425, + 0.0038892098236829042, + 0.006350231356918812, + 0.008109554648399353, + 0.0025794419925659895, + 0.007153915707021952, + 0.009381059557199478, + 0.003942613489925861, + 0.002974415896460414, + 0.005923563148826361, + 0.0072351000271737576, + 0.031024932861328125, + 0.0035951840691268444, + 0.03276827931404114, + 0.00509825861081481, + 0.00267585227265954, + 0.007131180260330439, + 0.0031887455843389034, + 0.005982678383588791, + 0.012835203669965267, + 0.011750313453376293, + 0.016369475051760674, + 0.013606490567326546, + 0.01018241886049509, + 0.023542087525129318, + 0.013298111036419868, + 0.013397321105003357, + 0.013614187017083168, + 0.00812467746436596, + 0.015560812316834927, + 0.017632491886615753, + 0.011038959957659245, + 0.01651335507631302, + 0.011479911394417286, + 0.019087638705968857, + 0.03358624875545502, + 0.027286341413855553, + 0.015986597165465355, + 0.013111710548400879, + 0.018191838636994362, + 0.012919258326292038, + 0.014614141546189785, + 0.007055865600705147, + 0.030786627903580666, + 0.027708766981959343, + 0.06459934264421463, + 0.0351187102496624, + 0.04844873398542404, + 0.0154997194185853, + 0.03543828800320625, + 0.017522040754556656, + 0.006122033577412367, + 0.006479548756033182, + 0.004144641105085611, + 0.004452218767255545, + 0.00477266451343894, + 0.0044567459262907505, + 0.004560016095638275, + 0.003611903404816985, + 0.009762054309248924, + 0.006981611717492342, + 0.009057193994522095, + 0.008462054654955864, + 0.010104015469551086, + 0.009972253814339638, + 0.010983452200889587, + 0.00889835599809885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 4 + ], + "selected_reward": -0.011076981201767921 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a single, unbroken shot, the camera begins with a static, top-down close-up of a sprawling blueprint spread across a dark, oil-stained wooden workbench. A maze of fine, cyan lines covers the crisp white construction paper, depicting the intricate schematics of a complex machine, complete with detailed gears, pistons, and wiring diagrams. Without any visible intervention, the blueprint begins to magically fold itself; sharp, precise creases form automatically, and the paper bends and folds with a life of its own, methodically transforming the complex 2D drawing into a sleek paper airplane. When the final fold is complete, the camera smoothly pulls back and begins tracking as the paper airplane, its wings now adorned with the fragmented schematics, gently lifts from the workbench and takes flight. It begins a graceful, soaring journey through a cluttered yet orderly inventor's workshop, the camera following seamlessly as the aircraft orbits a hanging industrial lamp, its warm glow illuminating the dancing dust motes in the air. It glides over half-assembled machinery, past pegboards neatly arranged with gleaming metal tools, swoops low over a metal lathe, and then rises again, navigating the complexities of space with effortless grace. The entire scene is bathed in the soft golden glow of afternoon sunlight, which filters through a large, grimy window, casting long shadows and creating an atmosphere of quiet wonder and dormant creative potential.", + "seed": 31656, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_14/index_3923_seed_31656_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.032611006312072276, + "vlm_logprob": -0.032611006312072276, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.032611006312072276, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, top-down close-up of a blueprint on a dark, oil-stained wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the blueprint magically fold itself into a sleek paper airplane without visible human intervention?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back and track the paper airplane as it takes flight and orbits a hanging industrial lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an inventor's workshop containing half-assembled machinery, pegboards with metal tools, and a metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight filtering through a large, grimy window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012017820961773396 + ], + "opposite_logprob_tokens": [ + -5.012017726898193 + ], + "logprob_reward": -0.012017820961773396, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.012017820961773396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30832093954086304, + 0.023541321977972984, + 0.11931329220533371, + 0.37427040934562683, + 0.05894047021865845, + 0.016845636069774628, + 0.02419774979352951, + 0.013105507008731365, + 0.019876454025506973, + 0.04158812761306763 + ], + "credit_map": [ + 0.008885687217116356, + 0.009713011793792248, + 0.01901528611779213, + 0.1124734878540039, + 0.010377214290201664, + 0.009987172670662403, + 0.026423612609505653, + 0.11144550144672394, + 0.0020461000967770815, + 0.0017481918912380934, + 0.0038498644717037678, + 0.00445898761972785, + 0.0018435029778629541, + 0.003270597429946065, + 0.0028524675872176886, + 0.003471611998975277, + 0.005433006212115288, + 0.007327621802687645, + 0.0064493403770029545, + 0.020187318325042725, + 0.009443099610507488, + 0.008192005567252636, + 0.03982773423194885, + 0.02245318330824375, + 0.01373148150742054, + 0.016857562586665154, + 0.1647670418024063, + 0.016484351828694344, + 0.012002306059002876, + 0.011104780249297619, + 0.01562940888106823, + 0.12369350343942642, + 0.005728629883378744, + 0.006967110093683004, + 0.003121275920420885, + 0.007056047208607197, + 0.007969182915985584, + 0.02138812094926834, + 0.002963036997243762, + 0.00374707137234509, + 0.002015474485233426, + 0.003419096814468503, + 0.0014572030631825328, + 0.001551907742395997, + 0.001954088220372796, + 0.002882814733311534, + 0.001844750135205686, + 0.001720302039757371, + 0.003160918829962611, + 0.0020672171376645565, + 0.0030528383795171976, + 0.003071401035413146, + 0.0025695948861539364, + 0.0029763290658593178, + 0.0023682962637394667, + 0.0049311574548482895, + 0.0019537925254553556, + 0.0021971333771944046, + 0.0008298958418890834, + 0.0012447199551388621, + 0.001730691292323172, + 0.001517068361863494, + 0.0012796925147995353, + 0.002352513372898102, + 0.0014969822950661182, + 0.003590026404708624, + 0.0020662941969931126, + 0.0030460015404969454, + 0.0017281306209042668, + 0.0029042186215519905, + 0.0020914466585963964, + 0.002953355433419347, + 0.0027770595625042915, + 0.004019911400973797, + 0.0028593852184712887, + 0.0044812350533902645, + 0.004009145312011242, + 0.009590942412614822, + 0.006047207396477461, + 0.007803243584930897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.012017820961773396 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, top-down close-up of a blueprint on a dark, oil-stained wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the blueprint magically fold itself into a sleek paper airplane without visible human intervention?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back and track the paper airplane as it takes flight and orbits a hanging industrial lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an inventor's workshop containing half-assembled machinery, pegboards with metal tools, and a metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight filtering through a large, grimy window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04870496690273285 + ], + "opposite_logprob_tokens": [ + -3.0487048625946045 + ], + "logprob_reward": -0.04870496690273285, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04870496690273285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6577699184417725, + 0.017459604889154434, + 0.03304540738463402, + 0.06555716693401337, + 0.0640525370836258, + 0.01598057523369789, + 0.03297710418701172, + 0.05352236330509186, + 0.014047701843082905, + 0.04558764025568962 + ], + "credit_map": [ + 0.014822427183389664, + 0.0194118469953537, + 0.04136974364519119, + 0.2609226107597351, + 0.012192144058644772, + 0.01840278133749962, + 0.046034976840019226, + 0.24461334943771362, + 0.0014783501392230392, + 0.001145172631368041, + 0.0028599409852176905, + 0.0036460754927247763, + 0.0009695673361420631, + 0.0013642619596794248, + 0.0034054003190249205, + 0.002590833930298686, + 0.004359149839729071, + 0.003046429716050625, + 0.00336705450899899, + 0.004509988706558943, + 0.002736080437898636, + 0.0028461941983550787, + 0.005314357113093138, + 0.006866150535643101, + 0.0025346216280013323, + 0.0036776966881006956, + 0.02347666583955288, + 0.0029503030236810446, + 0.005758230108767748, + 0.0035838517360389233, + 0.003687449498102069, + 0.019888345152139664, + 0.005874889437109232, + 0.007903055287897587, + 0.003088325494900346, + 0.008467295207083225, + 0.009903891943395138, + 0.022160271182656288, + 0.0026763633359223604, + 0.003978439141064882, + 0.0021099357400089502, + 0.0026359595358371735, + 0.0013913162983953953, + 0.0018457856494933367, + 0.002089391928166151, + 0.0022079346235841513, + 0.001878602895885706, + 0.0018216456519439816, + 0.002764252247288823, + 0.004429001826792955, + 0.0025584427639842033, + 0.0031691868789494038, + 0.007738935761153698, + 0.0039571612142026424, + 0.004816147033125162, + 0.0035439711064100266, + 0.005032340530306101, + 0.009643971920013428, + 0.003094436600804329, + 0.0058013200759887695, + 0.011041892692446709, + 0.0062736570835113525, + 0.004633121658116579, + 0.008001618087291718, + 0.0014381264336407185, + 0.0020745941437780857, + 0.0010946341790258884, + 0.0019572575110942125, + 0.0016621624818071723, + 0.0023662233725190163, + 0.0012402760330587626, + 0.0022144257090985775, + 0.0032579207327216864, + 0.00433221273124218, + 0.002807315904647112, + 0.004190482664853334, + 0.0037216152995824814, + 0.011964588426053524, + 0.006001774687319994, + 0.00931172538548708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.04870496690273285 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, top-down close-up of a blueprint on a dark, oil-stained wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the blueprint magically fold itself into a sleek paper airplane without visible human intervention?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back and track the paper airplane as it takes flight and orbits a hanging industrial lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an inventor's workshop containing half-assembled machinery, pegboards with metal tools, and a metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight filtering through a large, grimy window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004648474045097828 + ], + "opposite_logprob_tokens": [ + -5.379648685455322 + ], + "logprob_reward": -0.004648474045097828, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004648474045097828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5454716682434082, + 0.028550073504447937, + 0.02644125185906887, + 0.11308252066373825, + 0.06833256036043167, + 0.022687925025820732, + 0.018987152725458145, + 0.08028676360845566, + 0.025154154747724533, + 0.0710059329867363 + ], + "credit_map": [ + 0.013855090364813805, + 0.01633678562939167, + 0.03463379666209221, + 0.21457456052303314, + 0.012375532649457455, + 0.016627583652734756, + 0.03736212104558945, + 0.19970621168613434, + 0.002158940304070711, + 0.0027349814772605896, + 0.004689456429332495, + 0.005950596183538437, + 0.002560695167630911, + 0.002625439316034317, + 0.0035791860427707434, + 0.004250779747962952, + 0.004647911060601473, + 0.0019833766855299473, + 0.0019136829068884254, + 0.0025649552699178457, + 0.0066458070650696754, + 0.001784072839654982, + 0.004134556278586388, + 0.0027668909169733524, + 0.004192199092358351, + 0.0049071162939071655, + 0.051145218312740326, + 0.00523760262876749, + 0.002920793602243066, + 0.003442921442911029, + 0.004799345973879099, + 0.03643731772899628, + 0.006496405228972435, + 0.008577450178563595, + 0.004623632412403822, + 0.006396682932972908, + 0.009313316084444523, + 0.023567650467157364, + 0.0035146684385836124, + 0.005842748563736677, + 0.0025570467114448547, + 0.005141021218150854, + 0.0017037963261827826, + 0.001882881042547524, + 0.002574075246229768, + 0.004032060503959656, + 0.002047932706773281, + 0.002749111270532012, + 0.0020645910408347845, + 0.0019689849577844143, + 0.002320104744285345, + 0.0022508688271045685, + 0.0024456714745610952, + 0.0025461744517087936, + 0.0022201240062713623, + 0.0031706334557384253, + 0.011269568465650082, + 0.014056989923119545, + 0.004537752363830805, + 0.0091960858553648, + 0.014285939745604992, + 0.01120456401258707, + 0.0066026742570102215, + 0.009133189916610718, + 0.0022781172301620245, + 0.002990228123962879, + 0.0025275538209825754, + 0.003175872378051281, + 0.002093010814860463, + 0.0048661008477211, + 0.002308306284248829, + 0.004914965480566025, + 0.0067209843546152115, + 0.007928726263344288, + 0.00534095149487257, + 0.0060616061091423035, + 0.009686016477644444, + 0.015367913991212845, + 0.008615250699222088, + 0.011284483596682549 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.004648474045097828 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, top-down close-up of a blueprint on a dark, oil-stained wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the blueprint magically fold itself into a sleek paper airplane without visible human intervention?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back and track the paper airplane as it takes flight and orbits a hanging industrial lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an inventor's workshop containing half-assembled machinery, pegboards with metal tools, and a metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight filtering through a large, grimy window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008647141046822071 + ], + "opposite_logprob_tokens": [ + -4.7586469650268555 + ], + "logprob_reward": -0.008647141046822071, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008647141046822071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22396759688854218, + 0.02906656637787819, + 0.05377979204058647, + 0.41608887910842896, + 0.044245846569538116, + 0.01695825718343258, + 0.0110096987336874, + 0.06064542755484581, + 0.03585809841752052, + 0.10837987810373306 + ], + "credit_map": [ + 0.007260079961270094, + 0.00735721830278635, + 0.01628889888525009, + 0.08284647762775421, + 0.0092622647061944, + 0.008048411458730698, + 0.017690660431981087, + 0.07521357387304306, + 0.00234330166131258, + 0.003020284930244088, + 0.005992131307721138, + 0.0050400616601109505, + 0.0025260711554437876, + 0.0030940216965973377, + 0.003009793348610401, + 0.004040899686515331, + 0.007157615851610899, + 0.006438858807086945, + 0.0023931709583848715, + 0.004579617641866207, + 0.01616843417286873, + 0.003033433808013797, + 0.008048330433666706, + 0.005960329435765743, + 0.015030018985271454, + 0.020145215094089508, + 0.18369156122207642, + 0.018399028107523918, + 0.012714991346001625, + 0.012040529400110245, + 0.01791517250239849, + 0.13615235686302185, + 0.004450873006135225, + 0.005340574309229851, + 0.003386274678632617, + 0.00410724151879549, + 0.006695940624922514, + 0.013076329603791237, + 0.002656014170497656, + 0.004532600287348032, + 0.0021876904647797346, + 0.0035457357298582792, + 0.001471631694585085, + 0.0013130210572853684, + 0.0016552546294406056, + 0.0037258039228618145, + 0.0015460425056517124, + 0.001513077411800623, + 0.00138825923204422, + 0.0011486477451398969, + 0.001658556517213583, + 0.001129783340729773, + 0.0014672060497105122, + 0.0014687696238979697, + 0.0012889750069007277, + 0.001459501450881362, + 0.012819724157452583, + 0.009299198165535927, + 0.00440928665921092, + 0.005815684329718351, + 0.007237395271658897, + 0.006584539078176022, + 0.004934942815452814, + 0.009544655680656433, + 0.0038369111716747284, + 0.0049577620811760426, + 0.004293536301702261, + 0.004081749822944403, + 0.0033446471206843853, + 0.00431753508746624, + 0.004372249823063612, + 0.0066537074744701385, + 0.010192759335041046, + 0.015157145448029041, + 0.008675089105963707, + 0.00988396443426609, + 0.018826253712177277, + 0.018796510994434357, + 0.013086043298244476, + 0.013762109912931919 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.008647141046822071 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, top-down close-up of a blueprint on a dark, oil-stained wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the blueprint magically fold itself into a sleek paper airplane without visible human intervention?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back and track the paper airplane as it takes flight and orbits a hanging industrial lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an inventor's workshop containing half-assembled machinery, pegboards with metal tools, and a metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight filtering through a large, grimy window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903662860393524 + ], + "opposite_logprob_tokens": [ + -2.464036703109741 + ], + "logprob_reward": -0.08903662860393524, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903662860393524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49951231479644775, + 0.06893839687108994, + 0.05143052712082863, + 0.049605101346969604, + 0.04503779113292694, + 0.042421773076057434, + 0.0252451840788126, + 0.03214982897043228, + 0.024571742862462997, + 0.16108731925487518 + ], + "credit_map": [ + 0.014945593662559986, + 0.01848497800529003, + 0.028873244300484657, + 0.18604739010334015, + 0.012051111087203026, + 0.020768487825989723, + 0.03854187950491905, + 0.1797996610403061, + 0.007044225465506315, + 0.004850448574870825, + 0.012936587445437908, + 0.010407223366200924, + 0.004643092397600412, + 0.007763429079204798, + 0.010996122844517231, + 0.01029727328568697, + 0.006565936375409365, + 0.00459839915856719, + 0.005730897653847933, + 0.005024399608373642, + 0.012032867409288883, + 0.005022337194532156, + 0.0049379970878362656, + 0.007517695426940918, + 0.003219062928110361, + 0.0056414236314594746, + 0.012606743723154068, + 0.005071575753390789, + 0.0065696449019014835, + 0.005390886217355728, + 0.0038628263864666224, + 0.0072429426945745945, + 0.003744652960449457, + 0.004820138681679964, + 0.002366408007219434, + 0.00631830794736743, + 0.006772326771169901, + 0.012536587193608284, + 0.0037927692756056786, + 0.0046866051852703094, + 0.007244577165693045, + 0.005872922949492931, + 0.0039167627692222595, + 0.00277938274666667, + 0.0038359772879630327, + 0.009446734562516212, + 0.0026974172797054052, + 0.006628002971410751, + 0.0021295591723173857, + 0.0026735803112387657, + 0.0018227072432637215, + 0.0029706719797104597, + 0.004278926644474268, + 0.003499929793179035, + 0.0035101473331451416, + 0.004359662998467684, + 0.0043097385205328465, + 0.004232752602547407, + 0.002514513209462166, + 0.0035499276127666235, + 0.00501210568472743, + 0.003953799605369568, + 0.00396908400580287, + 0.0046079121530056, + 0.0031910541001707315, + 0.0029135344084352255, + 0.0030887832399457693, + 0.0033719129860401154, + 0.0029463886748999357, + 0.0028849232476204634, + 0.002723676385357976, + 0.0034514705184847116, + 0.012752489186823368, + 0.023389771580696106, + 0.012583051808178425, + 0.014212971553206444, + 0.03216792643070221, + 0.024629879742860794, + 0.021559741348028183, + 0.01979149505496025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.08903662860393524 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a single, unbroken shot, the camera begins with a static, top-down close-up of a sprawling blueprint spread across a dark, oil-stained wooden workbench. A maze of fine, cyan lines covers the crisp white construction paper, depicting the intricate schematics of a complex machine, complete with detailed gears, pistons, and wiring diagrams. Without any visible intervention, the blueprint begins to magically fold itself; sharp, precise creases form automatically, and the paper bends and folds with a life of its own, methodically transforming the complex 2D drawing into a sleek paper airplane. When the final fold is complete, the camera smoothly pulls back and begins tracking as the paper airplane, its wings now adorned with the fragmented schematics, gently lifts from the workbench and takes flight. It begins a graceful, soaring journey through a cluttered yet orderly inventor's workshop, the camera following seamlessly as the aircraft orbits a hanging industrial lamp, its warm glow illuminating the dancing dust motes in the air. It glides over half-assembled machinery, past pegboards neatly arranged with gleaming metal tools, swoops low over a metal lathe, and then rises again, navigating the complexities of space with effortless grace. The entire scene is bathed in the soft golden glow of afternoon sunlight, which filters through a large, grimy window, casting long shadows and creating an atmosphere of quiet wonder and dormant creative potential.", + "seed": 31657, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_14/index_3923_seed_31657_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.015241984277963638, + "vlm_logprob": -0.015241984277963638, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.015241984277963638, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, top-down close-up of a blueprint on a dark, oil-stained wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the blueprint magically fold itself into a sleek paper airplane without visible human intervention?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back and track the paper airplane as it takes flight and orbits a hanging industrial lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an inventor's workshop containing half-assembled machinery, pegboards with metal tools, and a metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight filtering through a large, grimy window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0084653589874506 + ], + "opposite_logprob_tokens": [ + -5.75846529006958 + ], + "logprob_reward": -0.0084653589874506, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0084653589874506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20214807987213135, + 0.04882906749844551, + 0.22375203669071198, + 0.15691201388835907, + 0.04507417604327202, + 0.055343642830848694, + 0.033573295921087265, + 0.10140212625265121, + 0.016243951395154, + 0.11672161519527435 + ], + "credit_map": [ + 0.016560303047299385, + 0.03137961030006409, + 0.06680592149496078, + 0.030624592676758766, + 0.011839085258543491, + 0.007823607884347439, + 0.024483157321810722, + 0.012631790712475777, + 0.003931037150323391, + 0.004807952791452408, + 0.008230744861066341, + 0.006007175426930189, + 0.006893019191920757, + 0.002378864213824272, + 0.0125485360622406, + 0.0040317378006875515, + 0.023701924830675125, + 0.02903824672102928, + 0.026770642027258873, + 0.03792243078351021, + 0.020036114379763603, + 0.01643618382513523, + 0.04393100365996361, + 0.025915490463376045, + 0.015384429134428501, + 0.005778818391263485, + 0.048153288662433624, + 0.02069464884698391, + 0.0047082602977752686, + 0.020103270187973976, + 0.012536369264125824, + 0.029552923515439034, + 0.00588935986161232, + 0.004893335979431868, + 0.007991684600710869, + 0.0040625291876494884, + 0.003380788257345557, + 0.007198604289442301, + 0.005519730970263481, + 0.00613814452663064, + 0.0051180836744606495, + 0.006108903791755438, + 0.00851447694003582, + 0.00462013203650713, + 0.006858362816274166, + 0.008285890333354473, + 0.007951583713293076, + 0.007886209525167942, + 0.0033918770495802164, + 0.0036799469962716103, + 0.00551175232976675, + 0.004415062256157398, + 0.0035998444072902203, + 0.00446404330432415, + 0.003991099540144205, + 0.0045196679420769215, + 0.006941989529877901, + 0.008481545373797417, + 0.02971356362104416, + 0.003466063179075718, + 0.005583152174949646, + 0.008464307524263859, + 0.004326434340327978, + 0.03442506864666939, + 0.0024791620671749115, + 0.001608372083865106, + 0.002724286401644349, + 0.0022593012545257807, + 0.0014128353213891387, + 0.0022308111656457186, + 0.0014631518861278892, + 0.0020660303998738527, + 0.011272121220827103, + 0.008578015491366386, + 0.030001962557435036, + 0.00686363922432065, + 0.005238435696810484, + 0.007478875108063221, + 0.006400959100574255, + 0.040887605398893356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.0084653589874506 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, top-down close-up of a blueprint on a dark, oil-stained wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the blueprint magically fold itself into a sleek paper airplane without visible human intervention?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back and track the paper airplane as it takes flight and orbits a hanging industrial lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an inventor's workshop containing half-assembled machinery, pegboards with metal tools, and a metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight filtering through a large, grimy window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029840555042028427 + ], + "opposite_logprob_tokens": [ + -3.5298404693603516 + ], + "logprob_reward": -0.029840555042028427, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029840555042028427, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29010340571403503, + 0.038451291620731354, + 0.12167204916477203, + 0.09518542140722275, + 0.10420682281255722, + 0.07877089828252792, + 0.035015542060136795, + 0.17675569653511047, + 0.020573792979121208, + 0.039265040308237076 + ], + "credit_map": [ + 0.020818663761019707, + 0.042706601321697235, + 0.10554437339305878, + 0.042693909257650375, + 0.014517760835587978, + 0.009883090853691101, + 0.0303594209253788, + 0.023579586297273636, + 0.0036447567399591208, + 0.0035269036889076233, + 0.005274116061627865, + 0.007721845060586929, + 0.00329100270755589, + 0.002439586678519845, + 0.008963183499872684, + 0.0035898981150239706, + 0.013314837589859962, + 0.01302857045084238, + 0.014649062417447567, + 0.02042177878320217, + 0.011377600021660328, + 0.011969412676990032, + 0.024981770664453506, + 0.011929016560316086, + 0.016543420031666756, + 0.004840377252548933, + 0.03191874548792839, + 0.014627506025135517, + 0.0035521837417036295, + 0.008453709073364735, + 0.009282536804676056, + 0.005966942757368088, + 0.009595283307135105, + 0.014598220586776733, + 0.014774590730667114, + 0.011425468139350414, + 0.01396951638162136, + 0.016323547810316086, + 0.01175715122371912, + 0.011763041839003563, + 0.008447139523923397, + 0.00759241683408618, + 0.013697493821382523, + 0.005527597852051258, + 0.01128297857940197, + 0.01357491035014391, + 0.007283349055796862, + 0.011365015059709549, + 0.004666011780500412, + 0.004955590702593327, + 0.0038437226321548223, + 0.002780589507892728, + 0.005567890591919422, + 0.004850239027291536, + 0.004839566070586443, + 0.003511932445690036, + 0.01162843406200409, + 0.008085972629487514, + 0.060947686433792114, + 0.007175585255026817, + 0.008842733688652515, + 0.01594805158674717, + 0.009344593621790409, + 0.05478263646364212, + 0.002639633836224675, + 0.0037790306378155947, + 0.0021404230501502752, + 0.0018999342573806643, + 0.0019882104825228453, + 0.0024760658852756023, + 0.0034994115121662617, + 0.0021510838996618986, + 0.0033015308436006308, + 0.001923064817674458, + 0.016005417332053185, + 0.0033421660773456097, + 0.001804827363230288, + 0.0018392314668744802, + 0.001762819243595004, + 0.009285980835556984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.029840555042028427 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, top-down close-up of a blueprint on a dark, oil-stained wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the blueprint magically fold itself into a sleek paper airplane without visible human intervention?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back and track the paper airplane as it takes flight and orbits a hanging industrial lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an inventor's workshop containing half-assembled machinery, pegboards with metal tools, and a metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight filtering through a large, grimy window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005962442606687546 + ], + "opposite_logprob_tokens": [ + -5.130962371826172 + ], + "logprob_reward": -0.005962442606687546, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005962442606687546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2704726457595825, + 0.03429703414440155, + 0.0517515204846859, + 0.09100257605314255, + 0.07182090729475021, + 0.06191574037075043, + 0.03072175942361355, + 0.22834427654743195, + 0.06244918704032898, + 0.0972243994474411 + ], + "credit_map": [ + 0.01769384741783142, + 0.04442277178168297, + 0.10157756507396698, + 0.041598688811063766, + 0.009554860182106495, + 0.009629707783460617, + 0.02651897631585598, + 0.019476212561130524, + 0.004150712862610817, + 0.002580555621534586, + 0.004934846889227629, + 0.005737209692597389, + 0.0028342711739242077, + 0.002385720144957304, + 0.007222925778478384, + 0.0044507901184260845, + 0.00424980279058218, + 0.006796376314014196, + 0.007200242020189762, + 0.009094069711863995, + 0.004410557448863983, + 0.00422444473952055, + 0.011656222864985466, + 0.0041197980754077435, + 0.020808471366763115, + 0.0034168120473623276, + 0.031226007267832756, + 0.012735038995742798, + 0.0023253720719367266, + 0.009076209738850594, + 0.006520368624478579, + 0.00489429384469986, + 0.006054427940398455, + 0.00703250989317894, + 0.007113688159734011, + 0.005017005372792482, + 0.008751751855015755, + 0.012667141854763031, + 0.008140653371810913, + 0.017043720930814743, + 0.007874335162341595, + 0.0037942773196846247, + 0.010956562124192715, + 0.004619689658284187, + 0.00793546810746193, + 0.008965328335762024, + 0.009059689939022064, + 0.008710387162864208, + 0.003887631930410862, + 0.004291297402232885, + 0.003866247134283185, + 0.0037260542158037424, + 0.0024592343252152205, + 0.00455361558124423, + 0.004040619358420372, + 0.003897058079019189, + 0.013513040728867054, + 0.009072176180779934, + 0.0788915827870369, + 0.01130625419318676, + 0.01258066762238741, + 0.020505638793110847, + 0.012274005450308323, + 0.07020089775323868, + 0.006089471746236086, + 0.005273786839097738, + 0.006715143099427223, + 0.004852717276662588, + 0.005191111005842686, + 0.009953705593943596, + 0.011760607361793518, + 0.012612640857696533, + 0.013725443743169308, + 0.008594058454036713, + 0.026203703135252, + 0.0049839201383292675, + 0.004261536058038473, + 0.005574545823037624, + 0.005271830130368471, + 0.02860935777425766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.005962442606687546 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, top-down close-up of a blueprint on a dark, oil-stained wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the blueprint magically fold itself into a sleek paper airplane without visible human intervention?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back and track the paper airplane as it takes flight and orbits a hanging industrial lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an inventor's workshop containing half-assembled machinery, pegboards with metal tools, and a metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight filtering through a large, grimy window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008651277050375938 + ], + "opposite_logprob_tokens": [ + -4.758651256561279 + ], + "logprob_reward": -0.008651277050375938, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008651277050375938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2088034749031067, + 0.025771642103791237, + 0.07976023107767105, + 0.17702871561050415, + 0.028880102559924126, + 0.04210534319281578, + 0.020186632871627808, + 0.09989668428897858, + 0.15947072207927704, + 0.15809646248817444 + ], + "credit_map": [ + 0.014221461489796638, + 0.03311940282583237, + 0.0759311392903328, + 0.03183165565133095, + 0.009343208745121956, + 0.007384859025478363, + 0.020774714648723602, + 0.016197029501199722, + 0.002410909393802285, + 0.0023304661735892296, + 0.003732918994501233, + 0.004320045001804829, + 0.0030087220948189497, + 0.0015376775991171598, + 0.00591258192434907, + 0.0025183220859616995, + 0.008776171132922173, + 0.010633615776896477, + 0.00928647443652153, + 0.012622900307178497, + 0.007331663742661476, + 0.006889427080750465, + 0.014439023099839687, + 0.009780955500900745, + 0.04307083785533905, + 0.007207113783806562, + 0.05904509127140045, + 0.024446628987789154, + 0.003965379670262337, + 0.010895122773945332, + 0.01241710688918829, + 0.015981433913111687, + 0.003907942213118076, + 0.003083055140450597, + 0.0035738975275307894, + 0.0030988482758402824, + 0.003007377963513136, + 0.003423134796321392, + 0.005036480259150267, + 0.0037493647541850805, + 0.005889055319130421, + 0.004215369001030922, + 0.0071867480874061584, + 0.0040998742915689945, + 0.0042473869398236275, + 0.004441480152308941, + 0.00602701585739851, + 0.00599841121584177, + 0.0023510949686169624, + 0.0023193261586129665, + 0.0033559121657162905, + 0.001515734358690679, + 0.002013913821429014, + 0.0032543682027608156, + 0.0021878278348594904, + 0.003188454546034336, + 0.0097962012514472, + 0.012409967370331287, + 0.02112874574959278, + 0.008995382115244865, + 0.009091920219361782, + 0.008419112302362919, + 0.006774052511900663, + 0.023281311616301537, + 0.013815040700137615, + 0.032223910093307495, + 0.01451586652547121, + 0.010564520955085754, + 0.009808157570660114, + 0.016562294214963913, + 0.045383527874946594, + 0.016597405076026917, + 0.014242424629628658, + 0.013223479501903057, + 0.04523606225848198, + 0.008538366295397282, + 0.008134670555591583, + 0.009380998089909554, + 0.00707533722743392, + 0.05226512253284454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.008651277050375938 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, top-down close-up of a blueprint on a dark, oil-stained wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the blueprint magically fold itself into a sleek paper airplane without visible human intervention?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back and track the paper airplane as it takes flight and orbits a hanging industrial lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an inventor's workshop containing half-assembled machinery, pegboards with metal tools, and a metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight filtering through a large, grimy window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329028770327568 + ], + "opposite_logprob_tokens": [ + -3.7732903957366943 + ], + "logprob_reward": -0.02329028770327568, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329028770327568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12607337534427643, + 0.03581107035279274, + 0.08789627999067307, + 0.051454249769449234, + 0.05772765353322029, + 0.1135941594839096, + 0.15240757167339325, + 0.2182505577802658, + 0.07924814522266388, + 0.077536940574646 + ], + "credit_map": [ + 0.007961993105709553, + 0.016788937151432037, + 0.04233697056770325, + 0.01861030049622059, + 0.009499776177108288, + 0.004788034129887819, + 0.017152704298496246, + 0.008934647776186466, + 0.004442874807864428, + 0.002814504085108638, + 0.004880956374108791, + 0.0057952734641730785, + 0.0035120747052133083, + 0.0025232492480427027, + 0.005392671562731266, + 0.006449465174227953, + 0.00985558982938528, + 0.011292291805148125, + 0.01011250913143158, + 0.014628397300839424, + 0.007711281068623066, + 0.007036987692117691, + 0.017789281904697418, + 0.00946993287652731, + 0.004530229140073061, + 0.0031303397845476866, + 0.013113862834870815, + 0.004959017504006624, + 0.0024063242599368095, + 0.008670332841575146, + 0.003242506179958582, + 0.01140163466334343, + 0.004725289531052113, + 0.00488761393353343, + 0.005012919660657644, + 0.003960552625358105, + 0.00515389908105135, + 0.010548017919063568, + 0.007913305424153805, + 0.015526057220995426, + 0.013670082204043865, + 0.011709374375641346, + 0.02320272848010063, + 0.007302457932382822, + 0.015349102206528187, + 0.017180301249027252, + 0.009766696020960808, + 0.015413410030305386, + 0.013833953998982906, + 0.019904563203454018, + 0.022131890058517456, + 0.01644017919898033, + 0.01602388545870781, + 0.02759031392633915, + 0.017202386632561684, + 0.01928040385246277, + 0.013516966253519058, + 0.025830380618572235, + 0.05785712972283363, + 0.016025036573410034, + 0.024287935346364975, + 0.017039475962519646, + 0.014744923450052738, + 0.048948705196380615, + 0.00691975187510252, + 0.014445950277149677, + 0.0067469272762537, + 0.005403802264481783, + 0.004502051044255495, + 0.008638915605843067, + 0.02173530124127865, + 0.010855443775653839, + 0.014492582529783249, + 0.0060304622165858746, + 0.027184274047613144, + 0.005744528491050005, + 0.00604426022619009, + 0.0036658330354839563, + 0.006035549566149712, + 0.008339446038007736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.02329028770327568 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a single, unbroken shot, the camera begins with a static, top-down close-up of a sprawling blueprint spread across a dark, oil-stained wooden workbench. A maze of fine, cyan lines covers the crisp white construction paper, depicting the intricate schematics of a complex machine, complete with detailed gears, pistons, and wiring diagrams. Without any visible intervention, the blueprint begins to magically fold itself; sharp, precise creases form automatically, and the paper bends and folds with a life of its own, methodically transforming the complex 2D drawing into a sleek paper airplane. When the final fold is complete, the camera smoothly pulls back and begins tracking as the paper airplane, its wings now adorned with the fragmented schematics, gently lifts from the workbench and takes flight. It begins a graceful, soaring journey through a cluttered yet orderly inventor's workshop, the camera following seamlessly as the aircraft orbits a hanging industrial lamp, its warm glow illuminating the dancing dust motes in the air. It glides over half-assembled machinery, past pegboards neatly arranged with gleaming metal tools, swoops low over a metal lathe, and then rises again, navigating the complexities of space with effortless grace. The entire scene is bathed in the soft golden glow of afternoon sunlight, which filters through a large, grimy window, casting long shadows and creating an atmosphere of quiet wonder and dormant creative potential.", + "seed": 31658, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_14/index_3923_seed_31658_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03639850579202175, + "vlm_logprob": -0.03639850579202175, + "vlm_yes_no_margin": 4.375000023841858, + "vlm_token_credit": -0.03639850579202175, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, top-down close-up of a blueprint on a dark, oil-stained wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the blueprint magically fold itself into a sleek paper airplane without visible human intervention?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back and track the paper airplane as it takes flight and orbits a hanging industrial lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an inventor's workshop containing half-assembled machinery, pegboards with metal tools, and a metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight filtering through a large, grimy window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009572675451636314 + ], + "opposite_logprob_tokens": [ + -5.884572505950928 + ], + "logprob_reward": -0.009572675451636314, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.009572675451636314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15653595328330994, + 0.04720644652843475, + 0.05369236692786217, + 0.053344935178756714, + 0.07794564962387085, + 0.0723329409956932, + 0.271224707365036, + 0.09100847691297531, + 0.12434206902980804, + 0.05236642435193062 + ], + "credit_map": [ + 0.0173500869423151, + 0.020358433946967125, + 0.009331200271844864, + 0.01953882910311222, + 0.03397101163864136, + 0.01889783889055252, + 0.01792914606630802, + 0.019159402698278427, + 0.004389714449644089, + 0.00538713950663805, + 0.005281521938741207, + 0.010130161419510841, + 0.0046821157447993755, + 0.004487012047320604, + 0.006920062005519867, + 0.005928718019276857, + 0.0044971671886742115, + 0.005094216670840979, + 0.0035316001158207655, + 0.007368888705968857, + 0.011418703012168407, + 0.006857675034552813, + 0.004251488484442234, + 0.010672629810869694, + 0.007381231524050236, + 0.0052536725997924805, + 0.008417570032179356, + 0.008538085967302322, + 0.0054151760414242744, + 0.004746034275740385, + 0.006333951372653246, + 0.0072592138312757015, + 0.006795814260840416, + 0.00821849424391985, + 0.013058985583484173, + 0.006013830192387104, + 0.011307401582598686, + 0.018814636394381523, + 0.0073040989227592945, + 0.0064323837868869305, + 0.010214507579803467, + 0.008816332556307316, + 0.008221618831157684, + 0.009653509594500065, + 0.006618989631533623, + 0.007594976108521223, + 0.01076542865484953, + 0.010447578504681587, + 0.01798759400844574, + 0.021649707108736038, + 0.023419031873345375, + 0.030704181641340256, + 0.02673613838851452, + 0.07171253859996796, + 0.018985457718372345, + 0.06003005430102348, + 0.010084680281579494, + 0.008399647660553455, + 0.010614988394081593, + 0.013442674651741982, + 0.01205906830728054, + 0.008078091777861118, + 0.017536351457238197, + 0.010792980901896954, + 0.021947581321001053, + 0.017046095803380013, + 0.013404635712504387, + 0.015944380313158035, + 0.012552927248179913, + 0.006463248748332262, + 0.021780259907245636, + 0.015202953480184078, + 0.00399112980812788, + 0.006096755154430866, + 0.003935025539249182, + 0.010245542973279953, + 0.006757339928299189, + 0.0082645108923316, + 0.004721755161881447, + 0.008354362100362778 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.009572675451636314 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, top-down close-up of a blueprint on a dark, oil-stained wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the blueprint magically fold itself into a sleek paper airplane without visible human intervention?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back and track the paper airplane as it takes flight and orbits a hanging industrial lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an inventor's workshop containing half-assembled machinery, pegboards with metal tools, and a metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight filtering through a large, grimy window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14285625517368317 + ], + "opposite_logprob_tokens": [ + -2.0178563594818115 + ], + "logprob_reward": -0.14285625517368317, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14285625517368317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08223726600408554, + 0.018984923139214516, + 0.08382581919431686, + 0.07357851415872574, + 0.03982217237353325, + 0.07813737541437149, + 0.3351234495639801, + 0.1374625712633133, + 0.1351710706949234, + 0.015656854957342148 + ], + "credit_map": [ + 0.006693019066005945, + 0.0074373758397996426, + 0.004297665320336819, + 0.008425009436905384, + 0.018934519961476326, + 0.009616325609385967, + 0.018121350556612015, + 0.008711992762982845, + 0.002536613028496504, + 0.0023111889604479074, + 0.0017903695115819573, + 0.0032023394014686346, + 0.0014130916679278016, + 0.0027724369429051876, + 0.0025146433617919683, + 0.002444240264594555, + 0.0086686871945858, + 0.0063379062339663506, + 0.006194406189024448, + 0.013043122366070747, + 0.01769757643342018, + 0.011080357246100903, + 0.008633899502456188, + 0.012169855646789074, + 0.00888550654053688, + 0.007503222674131393, + 0.008033503778278828, + 0.011388140730559826, + 0.012662982568144798, + 0.007017090916633606, + 0.009054852649569511, + 0.009033209644258022, + 0.004235302563756704, + 0.004320470616221428, + 0.005815806332975626, + 0.004171828273683786, + 0.005455079022794962, + 0.00817173346877098, + 0.0040720137767493725, + 0.0035799338947981596, + 0.008061378262937069, + 0.008681681007146835, + 0.009462532587349415, + 0.012645184062421322, + 0.008076784200966358, + 0.008292133919894695, + 0.011632430367171764, + 0.01128524262458086, + 0.021869271993637085, + 0.024268116801977158, + 0.027624323964118958, + 0.035896606743335724, + 0.030829953029751778, + 0.09279979765415192, + 0.01969306543469429, + 0.08214228600263596, + 0.011171507649123669, + 0.012260195799171925, + 0.013255259953439236, + 0.02707298845052719, + 0.014940300956368446, + 0.008601607754826546, + 0.03312595933675766, + 0.0170347411185503, + 0.025420621037483215, + 0.01764090359210968, + 0.013103757984936237, + 0.01762060634791851, + 0.013563640415668488, + 0.007403003517538309, + 0.024337762966752052, + 0.016080766916275024, + 0.0010527168633416295, + 0.0023607106413692236, + 0.001159489736892283, + 0.002723930636420846, + 0.002099372213706374, + 0.0022501731291413307, + 0.001457549398764968, + 0.0025529111735522747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.14285625517368317 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, top-down close-up of a blueprint on a dark, oil-stained wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the blueprint magically fold itself into a sleek paper airplane without visible human intervention?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back and track the paper airplane as it takes flight and orbits a hanging industrial lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an inventor's workshop containing half-assembled machinery, pegboards with metal tools, and a metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight filtering through a large, grimy window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006751700770109892 + ], + "opposite_logprob_tokens": [ + -5.006751537322998 + ], + "logprob_reward": -0.006751700770109892, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006751700770109892, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08468254655599594, + 0.03799653798341751, + 0.07633782178163528, + 0.04983191937208176, + 0.052313752472400665, + 0.07351064682006836, + 0.2822236120700836, + 0.22358514368534088, + 0.08870532363653183, + 0.030812755227088928 + ], + "credit_map": [ + 0.008760135620832443, + 0.011325715109705925, + 0.005889498628675938, + 0.01365876104682684, + 0.015074796043336391, + 0.008553809486329556, + 0.01101676095277071, + 0.010403069667518139, + 0.004190176259726286, + 0.003928828984498978, + 0.0035203653387725353, + 0.005757051520049572, + 0.006083767861127853, + 0.003988341894000769, + 0.005512904841452837, + 0.00501509802415967, + 0.007549142464995384, + 0.006228411104530096, + 0.0056751081719994545, + 0.012769951485097408, + 0.01779177039861679, + 0.009265055879950523, + 0.006845579482614994, + 0.010212801396846771, + 0.007089480757713318, + 0.005231647286564112, + 0.007629156578332186, + 0.006918418221175671, + 0.005953674204647541, + 0.004646527115255594, + 0.005315935239195824, + 0.007047080900520086, + 0.004199957940727472, + 0.008944779634475708, + 0.005950942635536194, + 0.004598940256983042, + 0.008787820115685463, + 0.008684043772518635, + 0.007442865986377001, + 0.0037044049240648746, + 0.00967567227780819, + 0.009095791727304459, + 0.008094626478850842, + 0.009486997500061989, + 0.0092364102602005, + 0.007871181704103947, + 0.011390659026801586, + 0.008659305982291698, + 0.0158487930893898, + 0.01845000497996807, + 0.021920491009950638, + 0.029248761013150215, + 0.022222254425287247, + 0.07751325517892838, + 0.012417868711054325, + 0.08460218459367752, + 0.015747832134366035, + 0.02036668360233307, + 0.022822294384241104, + 0.04484299570322037, + 0.017986582592129707, + 0.015156988985836506, + 0.056628625839948654, + 0.030033135786652565, + 0.014993726275861263, + 0.010889141820371151, + 0.007928269915282726, + 0.010642833076417446, + 0.01091716717928648, + 0.004555250983685255, + 0.019610164687037468, + 0.009168775752186775, + 0.002473941771313548, + 0.004363717045634985, + 0.002877144142985344, + 0.005631622392684221, + 0.0036200894974172115, + 0.004535382147878408, + 0.0024602427147328854, + 0.004850614350289106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.006751700770109892 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, top-down close-up of a blueprint on a dark, oil-stained wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the blueprint magically fold itself into a sleek paper airplane without visible human intervention?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back and track the paper airplane as it takes flight and orbits a hanging industrial lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an inventor's workshop containing half-assembled machinery, pegboards with metal tools, and a metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight filtering through a large, grimy window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00674353027716279 + ], + "opposite_logprob_tokens": [ + -5.006743431091309 + ], + "logprob_reward": -0.00674353027716279, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00674353027716279, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1477554887533188, + 0.045128099620342255, + 0.04641414433717728, + 0.042227983474731445, + 0.07042322307825089, + 0.09992389380931854, + 0.25997844338417053, + 0.13541972637176514, + 0.056830115616321564, + 0.0958988294005394 + ], + "credit_map": [ + 0.017025358974933624, + 0.014157830737531185, + 0.011639833450317383, + 0.02778039500117302, + 0.01829713210463524, + 0.016450250521302223, + 0.02417466975748539, + 0.018230028450489044, + 0.005635017529129982, + 0.004921054933220148, + 0.004713221453130245, + 0.006565119605511427, + 0.0048398831859230995, + 0.004945153370499611, + 0.006799993105232716, + 0.006708658300340176, + 0.0047708190977573395, + 0.005670383106917143, + 0.004287905525416136, + 0.0068482523784041405, + 0.003972058184444904, + 0.007937035523355007, + 0.00479086535051465, + 0.008136826567351818, + 0.00478370999917388, + 0.003761938540264964, + 0.004373041447252035, + 0.004749387968331575, + 0.006165732629597187, + 0.0060952650383114815, + 0.004013857804238796, + 0.008285049349069595, + 0.006559150759130716, + 0.0070485565811395645, + 0.007330247201025486, + 0.010622936300933361, + 0.009753498248755932, + 0.00985583383589983, + 0.012205520644783974, + 0.007047475781291723, + 0.013507927767932415, + 0.009594837203621864, + 0.012430421076714993, + 0.014142804779112339, + 0.009688819758594036, + 0.013491819612681866, + 0.014576518908143044, + 0.01249073538929224, + 0.013233515433967113, + 0.019304322078824043, + 0.019090570509433746, + 0.03223760798573494, + 0.013175014406442642, + 0.05341748148202896, + 0.011656618677079678, + 0.0978633314371109, + 0.01911727897822857, + 0.013717405498027802, + 0.011303775943815708, + 0.01637270115315914, + 0.013025776483118534, + 0.012457135133445263, + 0.03331711143255234, + 0.016108538955450058, + 0.009153022430837154, + 0.007073753513395786, + 0.005150991957634687, + 0.007806809619069099, + 0.007738366723060608, + 0.003656670916825533, + 0.009953871369361877, + 0.006296626292169094, + 0.008365301415324211, + 0.012061599642038345, + 0.01025205384939909, + 0.022006187587976456, + 0.007788074668496847, + 0.011582013219594955, + 0.012461108155548573, + 0.011382492259144783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.00674353027716279 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, top-down close-up of a blueprint on a dark, oil-stained wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the blueprint magically fold itself into a sleek paper airplane without visible human intervention?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back and track the paper airplane as it takes flight and orbits a hanging industrial lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an inventor's workshop containing half-assembled machinery, pegboards with metal tools, and a metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight filtering through a large, grimy window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016068367287516594 + ], + "opposite_logprob_tokens": [ + -4.141068458557129 + ], + "logprob_reward": -0.016068367287516594, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016068367287516594, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11591602116823196, + 0.07209179550409317, + 0.11295562982559204, + 0.03699537366628647, + 0.06357789784669876, + 0.05100210756063461, + 0.13562314212322235, + 0.0888247936964035, + 0.1144292950630188, + 0.20858392119407654 + ], + "credit_map": [ + 0.010588061064481735, + 0.011188128963112831, + 0.010364577174186707, + 0.02266421541571617, + 0.013132939115166664, + 0.008070985786616802, + 0.022622045129537582, + 0.017285075038671494, + 0.008048580028116703, + 0.007310575805604458, + 0.008961303159594536, + 0.009467574767768383, + 0.010552817955613136, + 0.008738039992749691, + 0.009730031713843346, + 0.009282872080802917, + 0.010750861838459969, + 0.010940062813460827, + 0.006509921979159117, + 0.009217300452291965, + 0.02921786904335022, + 0.014547103084623814, + 0.009021524339914322, + 0.022750990465283394, + 0.005860522855073214, + 0.003274011891335249, + 0.004643463529646397, + 0.004050441086292267, + 0.003361284267157316, + 0.0038152558263391256, + 0.0035282133612781763, + 0.008462183177471161, + 0.003779591992497444, + 0.010562022216618061, + 0.006444230210036039, + 0.005639966577291489, + 0.012502631172537804, + 0.008367432281374931, + 0.008655653335154057, + 0.0076263705268502235, + 0.006294078193604946, + 0.007368174381554127, + 0.005529135465621948, + 0.006043429020792246, + 0.006768342107534409, + 0.005662944633513689, + 0.007628773804754019, + 0.005707231350243092, + 0.01138732023537159, + 0.015135714784264565, + 0.012659135274589062, + 0.018305199220776558, + 0.016570862382650375, + 0.032844822853803635, + 0.009397500194609165, + 0.019322583451867104, + 0.008276139385998249, + 0.010635579004883766, + 0.007212706841528416, + 0.01311364583671093, + 0.010675900615751743, + 0.005886657629162073, + 0.02237066812813282, + 0.010653488337993622, + 0.012578876689076424, + 0.016131503507494926, + 0.007019639480859041, + 0.011201683431863785, + 0.018247617408633232, + 0.0110554164275527, + 0.025677457451820374, + 0.012517101131379604, + 0.016273723915219307, + 0.02231048047542572, + 0.018793543800711632, + 0.03215998038649559, + 0.03235509619116783, + 0.02602708712220192, + 0.03619462996721268, + 0.024469370022416115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.016068367287516594 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_7/index_tensor([4110, 4110])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_7/index_tensor([4110, 4110])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..eb8efcb867d9800b0ca08b402285eacc7e3bd114 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_7/index_tensor([4110, 4110])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level close-up lingers mid-section of a lone birch tree, capturing its form in surreal, macroscopic detail. The camera remains perfectly still, focused on the tree's papery white bark, which peels away in thin, translucent strips, revealing a chalky, soft layer beneath, its surface marked by distinctive black, horizontal streaks. A gentle, almost imperceptible breeze sweeps across the landscape, causing the tree's delicate, heart-shaped leaves to tremble in a constant, subtle movement. Each leaf, a vibrant, dewy green, captures the soft, diffuse light of an overcast morning, highlighting its fine, intricate veins as it flutters. The background is a soft-focus blur of cool, misty fields, the gentle greens and grays enhancing the birch tree's stark, elegant loneliness. The atmosphere is one of profound tranquility and quiet observation, the cool palette and gentle, trembling movement creating a visually rich and emotionally contemplative scene.", + "seed": 23549, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_7/index_4110_seed_23549_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.032406279863789675, + "vlm_logprob": -0.032406279863789675, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.032406279863789675, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up focused on the mid-section of a birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the birch bark peel away in thin, translucent strips to reveal a chalky, soft layer beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves heart-shaped, vibrant dewy green, and marked by fine, intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves tremble in a constant, subtle movement caused by a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of cool, misty fields with greens and grays?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.046129804104566574 + ], + "opposite_logprob_tokens": [ + -3.4211297035217285 + ], + "logprob_reward": -0.046129804104566574, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.046129804104566574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11540739983320236, + 0.07207164913415909, + 0.03507866710424423, + 0.05047503113746643, + 0.11386770755052567, + 0.09202992916107178, + 0.12540559470653534, + 0.03783522546291351, + 0.050578776746988297, + 0.30725011229515076 + ], + "credit_map": [ + 0.012420185841619968, + 0.00957584660500288, + 0.01201510988175869, + 0.01360495574772358, + 0.008550821803510189, + 0.02082834206521511, + 0.01916358806192875, + 0.01924852468073368, + 0.008171399123966694, + 0.005627889186143875, + 0.008797748945653439, + 0.010527215898036957, + 0.007235560566186905, + 0.010410712100565434, + 0.008545607328414917, + 0.012755511328577995, + 0.004332092124968767, + 0.0023516404908150434, + 0.0077353534288704395, + 0.004821334965527058, + 0.0026296530850231647, + 0.004416138399392366, + 0.003959339577704668, + 0.0048331101424992085, + 0.005269370507448912, + 0.0049025691114366055, + 0.006249178200960159, + 0.004783672746270895, + 0.005549231544137001, + 0.009049772284924984, + 0.006151048466563225, + 0.008520182222127914, + 0.01716320961713791, + 0.008747955784201622, + 0.014436199329793453, + 0.011411740444600582, + 0.01202577818185091, + 0.018978195264935493, + 0.014906484633684158, + 0.016198135912418365, + 0.016386007890105247, + 0.00754733569920063, + 0.012774496339261532, + 0.01142208743840456, + 0.006268482189625502, + 0.01019059494137764, + 0.012452382594347, + 0.014988524839282036, + 0.014195363037288189, + 0.017528178170323372, + 0.012195215560495853, + 0.01264770794659853, + 0.018478401005268097, + 0.016186071559786797, + 0.01692691072821617, + 0.017247729003429413, + 0.0035451948642730713, + 0.0023060315288603306, + 0.0053116874769330025, + 0.005227094516158104, + 0.0023111270274966955, + 0.008034639991819859, + 0.004190986976027489, + 0.006908458657562733, + 0.00676956120878458, + 0.00577913923189044, + 0.0055771018378436565, + 0.007797648664563894, + 0.0047632623463869095, + 0.006623418536037207, + 0.0057228016667068005, + 0.007545835338532925, + 0.03690123185515404, + 0.010296722874045372, + 0.0915258452296257, + 0.033465128391981125, + 0.012631364166736603, + 0.08651963621377945, + 0.009569372981786728, + 0.026340769603848457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.046129804104566574 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up focused on the mid-section of a birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the birch bark peel away in thin, translucent strips to reveal a chalky, soft layer beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves heart-shaped, vibrant dewy green, and marked by fine, intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves tremble in a constant, subtle movement caused by a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of cool, misty fields with greens and grays?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07007338851690292 + ], + "opposite_logprob_tokens": [ + -2.695073366165161 + ], + "logprob_reward": -0.07007338851690292, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07007338851690292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2578331530094147, + 0.0751008689403534, + 0.09419521689414978, + 0.06251569092273712, + 0.09082414954900742, + 0.16482073068618774, + 0.048539839684963226, + 0.05854576453566551, + 0.10822214186191559, + 0.03940241411328316 + ], + "credit_map": [ + 0.030663898214697838, + 0.019743528217077255, + 0.03718588128685951, + 0.027637051418423653, + 0.019329434260725975, + 0.03965495154261589, + 0.04196405038237572, + 0.04165438190102577, + 0.008995144627988338, + 0.0067289755679667, + 0.008423741906881332, + 0.009485708549618721, + 0.006853899918496609, + 0.010790212079882622, + 0.010626718401908875, + 0.01319647766649723, + 0.009932131506502628, + 0.006000860594213009, + 0.01394287496805191, + 0.009365861304104328, + 0.007713942788541317, + 0.015658225864171982, + 0.013409081846475601, + 0.01817224733531475, + 0.008628303185105324, + 0.009028468281030655, + 0.006902780383825302, + 0.005671109072864056, + 0.011628110893070698, + 0.007407575845718384, + 0.0067793261259794235, + 0.0064700208604335785, + 0.011663948185741901, + 0.007011373993009329, + 0.011258640326559544, + 0.008003893308341503, + 0.008626529015600681, + 0.016745543107390404, + 0.012398432940244675, + 0.01511579193174839, + 0.015341246500611305, + 0.009858319535851479, + 0.019429804757237434, + 0.015129811130464077, + 0.014592697843909264, + 0.03782881796360016, + 0.025991899892687798, + 0.026648137718439102, + 0.006066623609513044, + 0.00461833830922842, + 0.009478413499891758, + 0.007102648261934519, + 0.0037132930010557175, + 0.00574145931750536, + 0.004632255528122187, + 0.00718681188300252, + 0.007918187417089939, + 0.004339076112955809, + 0.006834032479673624, + 0.0061153448186814785, + 0.004388304892927408, + 0.010326804593205452, + 0.00970712210983038, + 0.008916893042623997, + 0.013799989596009254, + 0.011247912421822548, + 0.01271663699299097, + 0.019283220171928406, + 0.008936836384236813, + 0.015532591380178928, + 0.01032113004475832, + 0.016383832320570946, + 0.004156691022217274, + 0.0036394421476870775, + 0.00756789231672883, + 0.003868972649797797, + 0.0024820994585752487, + 0.008963615633547306, + 0.003277260111644864, + 0.005446445196866989 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.07007338851690292 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up focused on the mid-section of a birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the birch bark peel away in thin, translucent strips to reveal a chalky, soft layer beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves heart-shaped, vibrant dewy green, and marked by fine, intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves tremble in a constant, subtle movement caused by a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of cool, misty fields with greens and grays?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003625488607212901 + ], + "opposite_logprob_tokens": [ + -5.628625392913818 + ], + "logprob_reward": -0.003625488607212901, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003625488607212901, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12043741345405579, + 0.05600249394774437, + 0.034917328506708145, + 0.03592732921242714, + 0.047450192272663116, + 0.042730480432510376, + 0.10368619114160538, + 0.13281337916851044, + 0.08316564559936523, + 0.3428695797920227 + ], + "credit_map": [ + 0.018575021997094154, + 0.010580267757177353, + 0.012741249054670334, + 0.020985804498195648, + 0.008427080698311329, + 0.014154856093227863, + 0.017358748242259026, + 0.017614388838410378, + 0.0058465818874537945, + 0.007174663245677948, + 0.011658361181616783, + 0.007886300794780254, + 0.0036519623827189207, + 0.00869500171393156, + 0.005428748205304146, + 0.005660870112478733, + 0.003974719438701868, + 0.005917532369494438, + 0.004669566173106432, + 0.006945919245481491, + 0.003048795973882079, + 0.003771676914766431, + 0.003172517754137516, + 0.003416598541662097, + 0.0046167075634002686, + 0.004702009726315737, + 0.004012983292341232, + 0.004340524319559336, + 0.00486166775226593, + 0.004505819175392389, + 0.0048894197680056095, + 0.003998194355517626, + 0.005351456813514233, + 0.0027518041897565126, + 0.005562867037951946, + 0.005691060330718756, + 0.00384539901278913, + 0.008963299915194511, + 0.00776850339025259, + 0.007515796925872564, + 0.006514586042612791, + 0.0032457965426146984, + 0.005899238400161266, + 0.005231834016740322, + 0.003148839110508561, + 0.005544982850551605, + 0.006497915834188461, + 0.0066472855396568775, + 0.011828289367258549, + 0.014557178132236004, + 0.011529970914125443, + 0.009262681007385254, + 0.015374786220490932, + 0.01273954939097166, + 0.014179532416164875, + 0.014214199036359787, + 0.015633370727300644, + 0.011893054470419884, + 0.019291197881102562, + 0.019871031865477562, + 0.01661829836666584, + 0.022198906168341637, + 0.009889214299619198, + 0.01741829700767994, + 0.007539215963333845, + 0.011242561042308807, + 0.010162433609366417, + 0.015381391160190105, + 0.008697333745658398, + 0.008572003804147243, + 0.009687845595180988, + 0.011882849968969822, + 0.04131828248500824, + 0.011787139810621738, + 0.1024772897362709, + 0.03742567449808121, + 0.013681621290743351, + 0.09697692096233368, + 0.01051743421703577, + 0.028685197234153748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.003625488607212901 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up focused on the mid-section of a birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the birch bark peel away in thin, translucent strips to reveal a chalky, soft layer beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves heart-shaped, vibrant dewy green, and marked by fine, intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves tremble in a constant, subtle movement caused by a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of cool, misty fields with greens and grays?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810923546552658 + ], + "opposite_logprob_tokens": [ + -3.288109302520752 + ], + "logprob_reward": -0.03810923546552658, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810923546552658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12447473406791687, + 0.1320827752351761, + 0.06888491660356522, + 0.04194248095154762, + 0.06726591289043427, + 0.08243164420127869, + 0.03786630928516388, + 0.056136056780815125, + 0.12427359819412231, + 0.2646416127681732 + ], + "credit_map": [ + 0.01290027517825365, + 0.012491662986576557, + 0.01423554215580225, + 0.019177185371518135, + 0.010182484053075314, + 0.01606062799692154, + 0.01906808651983738, + 0.020358864217996597, + 0.013372763060033321, + 0.0077463495545089245, + 0.03771939501166344, + 0.018447326496243477, + 0.008459648117423058, + 0.019088849425315857, + 0.0132147753611207, + 0.01403366681188345, + 0.007803198415786028, + 0.0075481366366147995, + 0.008030697703361511, + 0.008427334949374199, + 0.004966913256794214, + 0.01215375866740942, + 0.006921173073351383, + 0.013033704832196236, + 0.00544190825894475, + 0.003482317551970482, + 0.006061658728867769, + 0.005388143006712198, + 0.0035867108963429928, + 0.005200122017413378, + 0.006222576834261417, + 0.0065590450540184975, + 0.013907013460993767, + 0.00868278183043003, + 0.007869930006563663, + 0.006697612348943949, + 0.006995313335210085, + 0.006699381861835718, + 0.007799588609486818, + 0.00861428864300251, + 0.015561871230602264, + 0.00729124341160059, + 0.011817211285233498, + 0.012470322661101818, + 0.005579053424298763, + 0.007900591939687729, + 0.010591603815555573, + 0.011219745501875877, + 0.005860547069460154, + 0.004442940000444651, + 0.006061904598027468, + 0.004202534444630146, + 0.0038573523052036762, + 0.004594194702804089, + 0.0045717740431427956, + 0.004275059327483177, + 0.006301274988800287, + 0.004446523729711771, + 0.0073869312182068825, + 0.008785897865891457, + 0.006007840856909752, + 0.00859027449041605, + 0.006331395357847214, + 0.00828591175377369, + 0.0157010518014431, + 0.014373178593814373, + 0.014097064733505249, + 0.021057486534118652, + 0.010590333491563797, + 0.01853449083864689, + 0.011359640397131443, + 0.018560348078608513, + 0.031464479863643646, + 0.00918788556009531, + 0.07825542986392975, + 0.028833864256739616, + 0.011107348836958408, + 0.0738442987203598, + 0.008439662866294384, + 0.023508628830313683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.03810923546552658 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up focused on the mid-section of a birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the birch bark peel away in thin, translucent strips to reveal a chalky, soft layer beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves heart-shaped, vibrant dewy green, and marked by fine, intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves tremble in a constant, subtle movement caused by a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of cool, misty fields with greens and grays?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040934826247394085 + ], + "opposite_logprob_tokens": [ + -5.504093647003174 + ], + "logprob_reward": -0.0040934826247394085, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040934826247394085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09547685086727142, + 0.08655832707881927, + 0.03833633288741112, + 0.05095384642481804, + 0.11646430194377899, + 0.09203168749809265, + 0.283296674489975, + 0.024866575375199318, + 0.04070991650223732, + 0.17130550742149353 + ], + "credit_map": [ + 0.014179171063005924, + 0.010352123528718948, + 0.012737314216792583, + 0.014958872459828854, + 0.00838400237262249, + 0.010797531343996525, + 0.011639549396932125, + 0.012428288348019123, + 0.009122399613261223, + 0.007232328876852989, + 0.010799845680594444, + 0.012319098226726055, + 0.006194557528942823, + 0.013752860948443413, + 0.01258893497288227, + 0.014548302628099918, + 0.0037745656445622444, + 0.0025228585582226515, + 0.007983471266925335, + 0.003994546364992857, + 0.001840072451159358, + 0.009113581851124763, + 0.002691512228921056, + 0.006415727082639933, + 0.00552495289593935, + 0.006961374077945948, + 0.005467435345053673, + 0.004710078705102205, + 0.00857172068208456, + 0.008133066818118095, + 0.0050651878118515015, + 0.006520028691738844, + 0.017655283212661743, + 0.011042727157473564, + 0.015960693359375, + 0.01006999984383583, + 0.012869843281805515, + 0.017199471592903137, + 0.014022083953022957, + 0.017644193023443222, + 0.015488612465560436, + 0.006740884855389595, + 0.01836853288114071, + 0.012515760958194733, + 0.005989040248095989, + 0.010059445165097713, + 0.010268592275679111, + 0.012600814923644066, + 0.03437952697277069, + 0.03569955751299858, + 0.038539137691259384, + 0.025950197130441666, + 0.04107115790247917, + 0.03563091903924942, + 0.03583448752760887, + 0.036191683262586594, + 0.002463365439325571, + 0.0016290266066789627, + 0.002879535546526313, + 0.0034147563856095076, + 0.0016515548340976238, + 0.004524720832705498, + 0.003036676673218608, + 0.005266938358545303, + 0.004058851394802332, + 0.005179181694984436, + 0.005502660293132067, + 0.006921411491930485, + 0.0041597080416977406, + 0.0040339077822864056, + 0.004799521062523127, + 0.006054673343896866, + 0.020533664152026176, + 0.005791527219116688, + 0.04954594746232033, + 0.017744889482855797, + 0.007308573462069035, + 0.05050233006477356, + 0.005412336438894272, + 0.0144662382081151 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 4 + ], + "selected_reward": -0.0040934826247394085 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up lingers mid-section of a lone birch tree, capturing its form in surreal, macroscopic detail. The camera remains perfectly still, focused on the tree's papery white bark, which peels away in thin, translucent strips, revealing a chalky, soft layer beneath, its surface marked by distinctive black, horizontal streaks. A gentle, almost imperceptible breeze sweeps across the landscape, causing the tree's delicate, heart-shaped leaves to tremble in a constant, subtle movement. Each leaf, a vibrant, dewy green, captures the soft, diffuse light of an overcast morning, highlighting its fine, intricate veins as it flutters. The background is a soft-focus blur of cool, misty fields, the gentle greens and grays enhancing the birch tree's stark, elegant loneliness. The atmosphere is one of profound tranquility and quiet observation, the cool palette and gentle, trembling movement creating a visually rich and emotionally contemplative scene.", + "seed": 23550, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_7/index_4110_seed_23550_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09670383520424367, + "vlm_logprob": -0.09670383520424367, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.09670383520424367, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up focused on the mid-section of a birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the birch bark peel away in thin, translucent strips to reveal a chalky, soft layer beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves heart-shaped, vibrant dewy green, and marked by fine, intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves tremble in a constant, subtle movement caused by a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of cool, misty fields with greens and grays?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028616808354854584 + ], + "opposite_logprob_tokens": [ + -4.403616905212402 + ], + "logprob_reward": -0.028616808354854584, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.028616808354854584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10594544559717178, + 0.2607095241546631, + 0.032550375908613205, + 0.04786752909421921, + 0.049232449382543564, + 0.04196446016430855, + 0.059245407581329346, + 0.10113351792097092, + 0.07329311966896057, + 0.22805818915367126 + ], + "credit_map": [ + 0.007904994301497936, + 0.007140436675399542, + 0.016383685171604156, + 0.008377972058951855, + 0.015572227537631989, + 0.017082450911402702, + 0.01124524511396885, + 0.022238437086343765, + 0.021159101277589798, + 0.037314221262931824, + 0.04185618832707405, + 0.037760570645332336, + 0.016680721193552017, + 0.014121352694928646, + 0.03629802167415619, + 0.055519334971904755, + 0.0037226704880595207, + 0.003314957721158862, + 0.00669194245710969, + 0.003058966714888811, + 0.002144432859495282, + 0.0028321545105427504, + 0.004541333299130201, + 0.006243919022381306, + 0.0051431297324597836, + 0.0056549375876784325, + 0.010957433842122555, + 0.004724335856735706, + 0.003629679325968027, + 0.004410094115883112, + 0.004896823782473803, + 0.008451094850897789, + 0.0038487999700009823, + 0.004199075512588024, + 0.012286287732422352, + 0.004596240818500519, + 0.0037712117191404104, + 0.0030345129780471325, + 0.008193635381758213, + 0.00930268969386816, + 0.003037218004465103, + 0.003495030337944627, + 0.007397249806672335, + 0.0037225536070764065, + 0.003923467360436916, + 0.005548947956413031, + 0.004013583995401859, + 0.010826409794390202, + 0.005110828205943108, + 0.006000104360282421, + 0.014003920368850231, + 0.007758222986012697, + 0.002580586588010192, + 0.003548648441210389, + 0.003910458646714687, + 0.016332633793354034, + 0.0073188492096960545, + 0.00941784493625164, + 0.025325536727905273, + 0.0060126581229269505, + 0.0068269819021224976, + 0.008853653445839882, + 0.007539054844528437, + 0.029838936403393745, + 0.006413215305656195, + 0.010663011111319065, + 0.012528108432888985, + 0.00712301442399621, + 0.005951587110757828, + 0.007217444479465485, + 0.00905170664191246, + 0.01434502936899662, + 0.01796102337539196, + 0.030146678909659386, + 0.04698370024561882, + 0.0376773439347744, + 0.009663723409175873, + 0.01360626332461834, + 0.017597133293747902, + 0.05442233383655548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.028616808354854584 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up focused on the mid-section of a birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the birch bark peel away in thin, translucent strips to reveal a chalky, soft layer beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves heart-shaped, vibrant dewy green, and marked by fine, intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves tremble in a constant, subtle movement caused by a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of cool, misty fields with greens and grays?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10037167370319366 + ], + "opposite_logprob_tokens": [ + -2.3503715991973877 + ], + "logprob_reward": -0.10037167370319366, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10037167370319366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12930168211460114, + 0.3599916100502014, + 0.09918155521154404, + 0.05504123121500015, + 0.09458159655332565, + 0.06380564719438553, + 0.07137027382850647, + 0.04991847276687622, + 0.03368624672293663, + 0.04312170296907425 + ], + "credit_map": [ + 0.009456700645387173, + 0.014333745464682579, + 0.02755824662744999, + 0.013683585450053215, + 0.007168092764914036, + 0.011105922050774097, + 0.010847117751836777, + 0.03514828160405159, + 0.031591594219207764, + 0.04923124983906746, + 0.06353341042995453, + 0.04897201061248779, + 0.021480323746800423, + 0.01910979673266411, + 0.04483233019709587, + 0.08124089986085892, + 0.007890116423368454, + 0.015602229163050652, + 0.015439963899552822, + 0.015910232439637184, + 0.008662434294819832, + 0.010528367012739182, + 0.012337746098637581, + 0.01281046587973833, + 0.005781807471066713, + 0.006419952027499676, + 0.010067043825984001, + 0.004919433034956455, + 0.003580995136871934, + 0.004806352313607931, + 0.008204908110201359, + 0.011260739527642727, + 0.00677120964974165, + 0.012477253563702106, + 0.022361695766448975, + 0.009114965796470642, + 0.0046784463338553905, + 0.0049608987756073475, + 0.012770608067512512, + 0.021446513012051582, + 0.0057199797593057156, + 0.009145176038146019, + 0.009682973846793175, + 0.011259539984166622, + 0.005602715071290731, + 0.005280277691781521, + 0.007394128479063511, + 0.009720852598547935, + 0.00603690417483449, + 0.005502801388502121, + 0.01882214844226837, + 0.007087047677487135, + 0.0030143335461616516, + 0.0035831518471240997, + 0.004929225891828537, + 0.022394657135009766, + 0.003022127551957965, + 0.005134030245244503, + 0.007054279092699289, + 0.007199559360742569, + 0.006112896837294102, + 0.008037921041250229, + 0.005182789172977209, + 0.00817487109452486, + 0.00325897429138422, + 0.003180908504873514, + 0.004257781896740198, + 0.00296215177513659, + 0.004284118767827749, + 0.006252889521420002, + 0.003898242022842169, + 0.005591182503849268, + 0.002401008503511548, + 0.004768698941916227, + 0.010065248236060143, + 0.005306870676577091, + 0.0019357915734872222, + 0.0034493673592805862, + 0.0035493141040205956, + 0.011645402759313583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.10037167370319366 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up focused on the mid-section of a birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the birch bark peel away in thin, translucent strips to reveal a chalky, soft layer beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves heart-shaped, vibrant dewy green, and marked by fine, intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves tremble in a constant, subtle movement caused by a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of cool, misty fields with greens and grays?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018206385895609856 + ], + "opposite_logprob_tokens": [ + -4.018206596374512 + ], + "logprob_reward": -0.018206385895609856, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018206385895609856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12905368208885193, + 0.10750480741262436, + 0.11433351784944534, + 0.04679286479949951, + 0.0649808794260025, + 0.044901903718709946, + 0.0373033843934536, + 0.08961893618106842, + 0.22999151051044464, + 0.1355186253786087 + ], + "credit_map": [ + 0.008613735437393188, + 0.01287542562931776, + 0.02280549332499504, + 0.013681536540389061, + 0.013332322239875793, + 0.012043870985507965, + 0.02348828688263893, + 0.022212987765669823, + 0.010043373331427574, + 0.012326786294579506, + 0.01446597557514906, + 0.014673794619739056, + 0.008595423772931099, + 0.009069180116057396, + 0.019610101357102394, + 0.018720142543315887, + 0.008704353123903275, + 0.0082833431661129, + 0.02459978312253952, + 0.008708189241588116, + 0.0068248966708779335, + 0.012022719718515873, + 0.015260730870068073, + 0.02992948703467846, + 0.0027624471113085747, + 0.004319760482758284, + 0.008625410497188568, + 0.0045883250422775745, + 0.004692771006375551, + 0.005017849151045084, + 0.0043715788051486015, + 0.012414714321494102, + 0.0054172673262655735, + 0.0053770835511386395, + 0.014457848854362965, + 0.007461367640644312, + 0.004513460211455822, + 0.005776115693151951, + 0.007839825935661793, + 0.014137892052531242, + 0.004847159143537283, + 0.00538855604827404, + 0.008111497387290001, + 0.006027611438184977, + 0.0037790255155414343, + 0.004366043955087662, + 0.004810350015759468, + 0.007571651600301266, + 0.00298078591004014, + 0.002843226306140423, + 0.00901353731751442, + 0.002516798675060272, + 0.002246143762022257, + 0.004726221784949303, + 0.002274604281410575, + 0.010702061466872692, + 0.0068984609097242355, + 0.009643452242016792, + 0.02170262299478054, + 0.007650194689631462, + 0.004842758644372225, + 0.008221611380577087, + 0.006456622388213873, + 0.02420320175588131, + 0.015346402302384377, + 0.018556686118245125, + 0.04296259954571724, + 0.014153105206787586, + 0.02704186737537384, + 0.03582821413874626, + 0.0181876253336668, + 0.05791497230529785, + 0.008908095769584179, + 0.01522461324930191, + 0.02699725143611431, + 0.017710616812109947, + 0.008065289817750454, + 0.01573113724589348, + 0.01044989563524723, + 0.032431717962026596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.018206385895609856 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up focused on the mid-section of a birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the birch bark peel away in thin, translucent strips to reveal a chalky, soft layer beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves heart-shaped, vibrant dewy green, and marked by fine, intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves tremble in a constant, subtle movement caused by a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of cool, misty fields with greens and grays?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813657224178314 + ], + "opposite_logprob_tokens": [ + -1.4063657522201538 + ], + "logprob_reward": -0.2813657224178314, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813657224178314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33362579345703125, + 0.1359098255634308, + 0.059903476387262344, + 0.07521138340234756, + 0.09571653604507446, + 0.11053936928510666, + 0.09627018868923187, + 0.023715971037745476, + 0.03466033190488815, + 0.03444717824459076 + ], + "credit_map": [ + 0.02070281095802784, + 0.03290216997265816, + 0.05386226624250412, + 0.021168716251850128, + 0.0504370778799057, + 0.04820244759321213, + 0.03430738300085068, + 0.07204290479421616, + 0.011908495798707008, + 0.026197299361228943, + 0.0176339540630579, + 0.02190668322145939, + 0.00972729455679655, + 0.008091217838227749, + 0.018910858780145645, + 0.021534008905291557, + 0.004348299466073513, + 0.008246702142059803, + 0.010509551502764225, + 0.006752073764801025, + 0.0056815058924257755, + 0.0073873307555913925, + 0.0044598705135285854, + 0.012518137693405151, + 0.00714744022116065, + 0.008010550402104855, + 0.015141943469643593, + 0.0060518053360283375, + 0.007020251359790564, + 0.008400256745517254, + 0.008221371099352837, + 0.015217755921185017, + 0.007052168715745211, + 0.009350824169814587, + 0.020247073844075203, + 0.010076037608087063, + 0.007203972432762384, + 0.00804837141185999, + 0.011254974640905857, + 0.0224831011146307, + 0.010009834542870522, + 0.014041666872799397, + 0.02085389941930771, + 0.01075449213385582, + 0.007724956143647432, + 0.009345637634396553, + 0.015742691233754158, + 0.022066179662942886, + 0.006204524543136358, + 0.009915766306221485, + 0.020121997222304344, + 0.006887869909405708, + 0.008430102840065956, + 0.010007874108850956, + 0.00908877607434988, + 0.025613263249397278, + 0.002016079844906926, + 0.0031222845427691936, + 0.003702713642269373, + 0.003630824154242873, + 0.0021956583950668573, + 0.002883745124563575, + 0.002324364148080349, + 0.0038402993232011795, + 0.0032993415370583534, + 0.0034661933314055204, + 0.005038235802203417, + 0.0030149947851896286, + 0.0031323821749538183, + 0.005915561690926552, + 0.005404152441769838, + 0.00538946408778429, + 0.0026334708090871572, + 0.003483603708446026, + 0.005069108679890633, + 0.003287524450570345, + 0.0028731159400194883, + 0.007248324807733297, + 0.00341810192912817, + 0.006433924660086632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.2813657224178314 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up focused on the mid-section of a birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the birch bark peel away in thin, translucent strips to reveal a chalky, soft layer beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves heart-shaped, vibrant dewy green, and marked by fine, intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves tremble in a constant, subtle movement caused by a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of cool, misty fields with greens and grays?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054958585649728775 + ], + "opposite_logprob_tokens": [ + -2.9299585819244385 + ], + "logprob_reward": -0.054958585649728775, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054958585649728775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23240771889686584, + 0.060469187796115875, + 0.1327776461839676, + 0.06692898273468018, + 0.082663893699646, + 0.053290314972400665, + 0.043511614203453064, + 0.062314391136169434, + 0.14030954241752625, + 0.12532666325569153 + ], + "credit_map": [ + 0.015093579888343811, + 0.020596366375684738, + 0.03575534000992775, + 0.037025272846221924, + 0.024602431803941727, + 0.01536883506923914, + 0.04152276739478111, + 0.042443111538887024, + 0.0062498231418430805, + 0.006377884652465582, + 0.010342027060687542, + 0.009075293317437172, + 0.0067339567467570305, + 0.004744112957268953, + 0.0061081331223249435, + 0.010837957262992859, + 0.009834314696490765, + 0.010028212331235409, + 0.028679534792900085, + 0.01261322945356369, + 0.007570987101644278, + 0.013232131488621235, + 0.0141821950674057, + 0.03663703054189682, + 0.0035930227022618055, + 0.006489082705229521, + 0.011360207572579384, + 0.005163369234651327, + 0.00713411346077919, + 0.01114120613783598, + 0.00564924580976367, + 0.01639874093234539, + 0.007035992573946714, + 0.007898364216089249, + 0.019135139882564545, + 0.008801898919045925, + 0.006993704475462437, + 0.006216933950781822, + 0.011157573200762272, + 0.015424287877976894, + 0.004578747786581516, + 0.005302031058818102, + 0.007397205103188753, + 0.006088895723223686, + 0.0068696946837008, + 0.00833720713853836, + 0.00473643746227026, + 0.009980098344385624, + 0.0030147673096507788, + 0.0024162298068404198, + 0.00970382709056139, + 0.0026999320834875107, + 0.004645053762942553, + 0.004488422069698572, + 0.0038001853972673416, + 0.012743199244141579, + 0.005284387618303299, + 0.010331369936466217, + 0.007965261116623878, + 0.006738216616213322, + 0.007229336071759462, + 0.008396562188863754, + 0.007977228611707687, + 0.0083920294418931, + 0.009547763504087925, + 0.010915803723037243, + 0.027342839166522026, + 0.00893507432192564, + 0.016232511028647423, + 0.021441923454403877, + 0.009978445246815681, + 0.03591518849134445, + 0.009770954959094524, + 0.017089318484067917, + 0.025338279083371162, + 0.022257572039961815, + 0.005300483200699091, + 0.006615324877202511, + 0.010068356990814209, + 0.028886374086141586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.054958585649728775 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up lingers mid-section of a lone birch tree, capturing its form in surreal, macroscopic detail. The camera remains perfectly still, focused on the tree's papery white bark, which peels away in thin, translucent strips, revealing a chalky, soft layer beneath, its surface marked by distinctive black, horizontal streaks. A gentle, almost imperceptible breeze sweeps across the landscape, causing the tree's delicate, heart-shaped leaves to tremble in a constant, subtle movement. Each leaf, a vibrant, dewy green, captures the soft, diffuse light of an overcast morning, highlighting its fine, intricate veins as it flutters. The background is a soft-focus blur of cool, misty fields, the gentle greens and grays enhancing the birch tree's stark, elegant loneliness. The atmosphere is one of profound tranquility and quiet observation, the cool palette and gentle, trembling movement creating a visually rich and emotionally contemplative scene.", + "seed": 23551, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_7/index_4110_seed_23551_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.028653429355472326, + "vlm_logprob": -0.028653429355472326, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.028653429355472326, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up focused on the mid-section of a birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the birch bark peel away in thin, translucent strips to reveal a chalky, soft layer beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves heart-shaped, vibrant dewy green, and marked by fine, intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves tremble in a constant, subtle movement caused by a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of cool, misty fields with greens and grays?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023594098165631294 + ], + "opposite_logprob_tokens": [ + -4.398593902587891 + ], + "logprob_reward": -0.023594098165631294, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.023594098165631294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25709497928619385, + 0.040673013776540756, + 0.11844837665557861, + 0.0489073283970356, + 0.04640451818704605, + 0.08654338866472244, + 0.16082704067230225, + 0.08774162083864212, + 0.09330885112285614, + 0.06005091592669487 + ], + "credit_map": [ + 0.022944554686546326, + 0.013439878821372986, + 0.07121939212083817, + 0.054195404052734375, + 0.012530209496617317, + 0.023631785064935684, + 0.03184131532907486, + 0.02729240246117115, + 0.004360863473266363, + 0.0033502038568258286, + 0.007178331725299358, + 0.00648602657020092, + 0.002845659153535962, + 0.006128065753728151, + 0.0038740639574825764, + 0.006449792999774218, + 0.007995842024683952, + 0.007107178680598736, + 0.018236782401800156, + 0.016373734921216965, + 0.010261510498821735, + 0.021071448922157288, + 0.01628941483795643, + 0.02111244574189186, + 0.004645489156246185, + 0.003920730669051409, + 0.011840609833598137, + 0.010332615114748478, + 0.0033889953047037125, + 0.003495339769870043, + 0.0056433710269629955, + 0.005640173330903053, + 0.004680355079472065, + 0.0041742995381355286, + 0.010076464153826237, + 0.007766406983137131, + 0.0031996266916394234, + 0.004569887183606625, + 0.006152432411909103, + 0.005785040091723204, + 0.005759201943874359, + 0.005648489575833082, + 0.01646965742111206, + 0.015705743804574013, + 0.007749410346150398, + 0.014653583988547325, + 0.010712098330259323, + 0.009845195338129997, + 0.012896840460598469, + 0.02190604992210865, + 0.01950562372803688, + 0.01841120794415474, + 0.02991119772195816, + 0.017814531922340393, + 0.02015257626771927, + 0.020228993147611618, + 0.007400802336633205, + 0.00692442012950778, + 0.019870255142450333, + 0.018039843067526817, + 0.006578825414180756, + 0.01121317408978939, + 0.00788179226219654, + 0.00983249582350254, + 0.00748030561953783, + 0.014391687698662281, + 0.013723642565310001, + 0.013070203363895416, + 0.009220875799655914, + 0.011129835620522499, + 0.011971641331911087, + 0.012320640496909618, + 0.005104078911244869, + 0.0056312354281544685, + 0.012362033128738403, + 0.010036054067313671, + 0.005465081427246332, + 0.006712517701089382, + 0.007626226171851158, + 0.007113680709153414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.023594098165631294 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up focused on the mid-section of a birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the birch bark peel away in thin, translucent strips to reveal a chalky, soft layer beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves heart-shaped, vibrant dewy green, and marked by fine, intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves tremble in a constant, subtle movement caused by a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of cool, misty fields with greens and grays?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07903526723384857 + ], + "opposite_logprob_tokens": [ + -2.5790352821350098 + ], + "logprob_reward": -0.07903526723384857, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07903526723384857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06905461102724075, + 0.05490590259432793, + 0.04254840314388275, + 0.08781155198812485, + 0.17995589971542358, + 0.32697170972824097, + 0.10097122937440872, + 0.041606172919273376, + 0.04623241350054741, + 0.04994216561317444 + ], + "credit_map": [ + 0.008271023631095886, + 0.005655545741319656, + 0.012637649662792683, + 0.008072786964476109, + 0.004981942940503359, + 0.009292282164096832, + 0.01049446314573288, + 0.009648906998336315, + 0.0037853200919926167, + 0.002722970675677061, + 0.00738550117239356, + 0.007008622866123915, + 0.0036481013521552086, + 0.013372916728258133, + 0.006479905918240547, + 0.010502555407583714, + 0.0037954344879835844, + 0.0026171694044023752, + 0.006197282578796148, + 0.0055951448157429695, + 0.002851268742233515, + 0.0063659935258328915, + 0.006286130752414465, + 0.008839977905154228, + 0.009130530059337616, + 0.007924358360469341, + 0.015876810997724533, + 0.010874547995626926, + 0.0065310620702803135, + 0.010070289485156536, + 0.015267588198184967, + 0.012136359699070454, + 0.01304253377020359, + 0.008082030341029167, + 0.03552628308534622, + 0.03280990570783615, + 0.011678172275424004, + 0.030596869066357613, + 0.01923607848584652, + 0.028984010219573975, + 0.023814000189304352, + 0.021164461970329285, + 0.03980745002627373, + 0.03905338793992996, + 0.029989710077643394, + 0.0713791698217392, + 0.043383631855249405, + 0.05837986618280411, + 0.0069101732224226, + 0.013157445006072521, + 0.012698608450591564, + 0.013413116335868835, + 0.01694634184241295, + 0.012441190890967846, + 0.012251045554876328, + 0.01315329223871231, + 0.004123090300709009, + 0.002966236788779497, + 0.01045481488108635, + 0.007882057689130306, + 0.002675408497452736, + 0.003565769409760833, + 0.005668296478688717, + 0.004270493984222412, + 0.003000050550326705, + 0.002902938984334469, + 0.004962923005223274, + 0.007442553527653217, + 0.0031425366178154945, + 0.009557247161865234, + 0.005402577109634876, + 0.009821584448218346, + 0.0035318646114319563, + 0.0027338380459696054, + 0.00694021862000227, + 0.006943271029740572, + 0.0036427362356334925, + 0.009116043336689472, + 0.007243873551487923, + 0.00979031901806593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.07903526723384857 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up focused on the mid-section of a birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the birch bark peel away in thin, translucent strips to reveal a chalky, soft layer beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves heart-shaped, vibrant dewy green, and marked by fine, intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves tremble in a constant, subtle movement caused by a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of cool, misty fields with greens and grays?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011093014851212502 + ], + "opposite_logprob_tokens": [ + -4.5110931396484375 + ], + "logprob_reward": -0.011093014851212502, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011093014851212502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18944798409938812, + 0.10665430128574371, + 0.12579233944416046, + 0.0978507250547409, + 0.13144059479236603, + 0.12626181542873383, + 0.08420373499393463, + 0.03913130983710289, + 0.049258869141340256, + 0.04995832219719887 + ], + "credit_map": [ + 0.024198725819587708, + 0.014935058541595936, + 0.04411507397890091, + 0.02623618394136429, + 0.016058627516031265, + 0.01835525594651699, + 0.026063673198223114, + 0.01948537491261959, + 0.011898698285222054, + 0.009903945028781891, + 0.01847795397043228, + 0.019486835226416588, + 0.010863509960472584, + 0.011563990265130997, + 0.013236699625849724, + 0.011222665198147297, + 0.01037430390715599, + 0.013341868296265602, + 0.022889278829097748, + 0.020376984030008316, + 0.009878888726234436, + 0.01585724577307701, + 0.016887085512280464, + 0.016186676919460297, + 0.008583839051425457, + 0.01728774979710579, + 0.013640910387039185, + 0.013004408217966557, + 0.011156381107866764, + 0.010571696795523167, + 0.011778985150158405, + 0.011826750822365284, + 0.010453514754772186, + 0.013834815472364426, + 0.032476846128702164, + 0.030691273510456085, + 0.014054498635232449, + 0.006004902999848127, + 0.012910217046737671, + 0.011014514602720737, + 0.010145755484700203, + 0.009482711553573608, + 0.017988506704568863, + 0.013897056691348553, + 0.012678376398980618, + 0.024723147973418236, + 0.016494419425725937, + 0.020851844921708107, + 0.006577897351235151, + 0.009869307279586792, + 0.010899425484240055, + 0.013280639424920082, + 0.012782849371433258, + 0.008402140811085701, + 0.01220835279673338, + 0.010183123871684074, + 0.0036876010708510876, + 0.0034732434432953596, + 0.007895756512880325, + 0.00625739386305213, + 0.00464024068787694, + 0.004206002689898014, + 0.004157269839197397, + 0.004813795909285545, + 0.003975001163780689, + 0.00827324390411377, + 0.005876179318875074, + 0.006476245820522308, + 0.006410804111510515, + 0.007822650484740734, + 0.004590755794197321, + 0.005833983421325684, + 0.004697597119957209, + 0.0060640666633844376, + 0.0061323209665715694, + 0.00649406248703599, + 0.005098303779959679, + 0.008462398312985897, + 0.00580195477232337, + 0.007207615301012993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.011093014851212502 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up focused on the mid-section of a birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the birch bark peel away in thin, translucent strips to reveal a chalky, soft layer beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves heart-shaped, vibrant dewy green, and marked by fine, intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves tremble in a constant, subtle movement caused by a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of cool, misty fields with greens and grays?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635374292731285 + ], + "opposite_logprob_tokens": [ + -3.6513538360595703 + ], + "logprob_reward": -0.02635374292731285, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635374292731285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2789006531238556, + 0.07332592457532883, + 0.06291306763887405, + 0.04545847699046135, + 0.05310220643877983, + 0.14920559525489807, + 0.07460823655128479, + 0.1227322667837143, + 0.0781099870800972, + 0.061643581837415695 + ], + "credit_map": [ + 0.02644684538245201, + 0.013049628585577011, + 0.0791424810886383, + 0.058958813548088074, + 0.015735378488898277, + 0.020184867084026337, + 0.03682713955640793, + 0.028555497527122498, + 0.0072431364096701145, + 0.004293259233236313, + 0.01742018759250641, + 0.013938075862824917, + 0.004068202339112759, + 0.008363944478332996, + 0.008888241834938526, + 0.009110875427722931, + 0.005104264710098505, + 0.007716021500527859, + 0.014315702952444553, + 0.012484977021813393, + 0.004364565946161747, + 0.00576294818893075, + 0.006451400462538004, + 0.006713187322020531, + 0.004096547607332468, + 0.003341831499710679, + 0.008574760518968105, + 0.007247799541801214, + 0.0035512091126292944, + 0.005242346785962582, + 0.005452276673167944, + 0.007951704785227776, + 0.0033804241102188826, + 0.0032398556359112263, + 0.009019067510962486, + 0.010122050531208515, + 0.0036663420032709837, + 0.008381372317671776, + 0.006566216237843037, + 0.008726879954338074, + 0.011181864887475967, + 0.01262977160513401, + 0.025787072256207466, + 0.018224000930786133, + 0.011587838642299175, + 0.025741206482052803, + 0.019755342975258827, + 0.024298498407006264, + 0.0049901497550308704, + 0.005752030294388533, + 0.012888884171843529, + 0.012982530519366264, + 0.007935933768749237, + 0.005993953440338373, + 0.013618138618767262, + 0.010446617379784584, + 0.011087032034993172, + 0.009210837073624134, + 0.018515102565288544, + 0.017165839672088623, + 0.010516850277781487, + 0.02088327892124653, + 0.011758181266486645, + 0.023595135658979416, + 0.007037278264760971, + 0.012920036911964417, + 0.014610684476792812, + 0.015132890082895756, + 0.008607257157564163, + 0.0052931043319404125, + 0.006983910221606493, + 0.0075248233042657375, + 0.005115182604640722, + 0.004481921903789043, + 0.017211131751537323, + 0.011987682431936264, + 0.004788241349160671, + 0.004517857916653156, + 0.007269702851772308, + 0.0062718624249100685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.02635374292731285 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up focused on the mid-section of a birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the birch bark peel away in thin, translucent strips to reveal a chalky, soft layer beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves heart-shaped, vibrant dewy green, and marked by fine, intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves tremble in a constant, subtle movement caused by a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of cool, misty fields with greens and grays?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003191023599356413 + ], + "opposite_logprob_tokens": [ + -5.753190994262695 + ], + "logprob_reward": -0.003191023599356413, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003191023599356413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2039806842803955, + 0.06597890704870224, + 0.05375237762928009, + 0.05197952687740326, + 0.05156177282333374, + 0.08809377998113632, + 0.1530500203371048, + 0.08914875239133835, + 0.2130211442708969, + 0.029433010146021843 + ], + "credit_map": [ + 0.018885141238570213, + 0.01525808870792389, + 0.05049867182970047, + 0.03532979264855385, + 0.013337574899196625, + 0.021282296627759933, + 0.02664833702147007, + 0.022740785032510757, + 0.006480105221271515, + 0.004259171895682812, + 0.01829863339662552, + 0.011884658597409725, + 0.004037107340991497, + 0.006315161939710379, + 0.007106817327439785, + 0.0075972494669258595, + 0.005185410380363464, + 0.0035583502613008022, + 0.01028081402182579, + 0.009451963007450104, + 0.004302320536226034, + 0.0068055614829063416, + 0.007263528183102608, + 0.0069044288247823715, + 0.005081958137452602, + 0.005174091551452875, + 0.01208449061959982, + 0.009633570909500122, + 0.004333028569817543, + 0.003980109468102455, + 0.0064206053502857685, + 0.005271675065159798, + 0.004443005658686161, + 0.003405652241781354, + 0.014131966046988964, + 0.01327748503535986, + 0.00330794183537364, + 0.003794855671003461, + 0.00452555064111948, + 0.004675315227359533, + 0.0068358564749360085, + 0.010485244914889336, + 0.01404118537902832, + 0.009042112156748772, + 0.009830951690673828, + 0.012604529969394207, + 0.012034960091114044, + 0.01321894209831953, + 0.012494305148720741, + 0.018103189766407013, + 0.0388319231569767, + 0.025438331067562103, + 0.01304350420832634, + 0.012854079715907574, + 0.0162214245647192, + 0.016063274815678596, + 0.007860296405851841, + 0.006225016433745623, + 0.025949692353606224, + 0.020514121279120445, + 0.00499241566285491, + 0.006697498261928558, + 0.008932969532907009, + 0.007976745255291462, + 0.019378352910280228, + 0.04532643407583237, + 0.03820354491472244, + 0.03548261895775795, + 0.028120599687099457, + 0.01239261869341135, + 0.02072162926197052, + 0.01339534018188715, + 0.002581765642389655, + 0.0021266480907797813, + 0.004021076485514641, + 0.003908712882548571, + 0.0022052170243114233, + 0.004948574583977461, + 0.005196470767259598, + 0.004444546066224575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.003191023599356413 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up lingers mid-section of a lone birch tree, capturing its form in surreal, macroscopic detail. The camera remains perfectly still, focused on the tree's papery white bark, which peels away in thin, translucent strips, revealing a chalky, soft layer beneath, its surface marked by distinctive black, horizontal streaks. A gentle, almost imperceptible breeze sweeps across the landscape, causing the tree's delicate, heart-shaped leaves to tremble in a constant, subtle movement. Each leaf, a vibrant, dewy green, captures the soft, diffuse light of an overcast morning, highlighting its fine, intricate veins as it flutters. The background is a soft-focus blur of cool, misty fields, the gentle greens and grays enhancing the birch tree's stark, elegant loneliness. The atmosphere is one of profound tranquility and quiet observation, the cool palette and gentle, trembling movement creating a visually rich and emotionally contemplative scene.", + "seed": 23552, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_7/index_4110_seed_23552_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.028247063304297625, + "vlm_logprob": -0.028247063304297625, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.028247063304297625, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up focused on the mid-section of a birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the birch bark peel away in thin, translucent strips to reveal a chalky, soft layer beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves heart-shaped, vibrant dewy green, and marked by fine, intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves tremble in a constant, subtle movement caused by a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of cool, misty fields with greens and grays?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023691311478614807 + ], + "opposite_logprob_tokens": [ + -4.898691177368164 + ], + "logprob_reward": -0.023691311478614807, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.023691311478614807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18056128919124603, + 0.12501533329486847, + 0.0957699865102768, + 0.1600719541311264, + 0.17193910479545593, + 0.048875078558921814, + 0.07279539108276367, + 0.023353073745965958, + 0.06811150163412094, + 0.05350733920931816 + ], + "credit_map": [ + 0.0283916424959898, + 0.023915225639939308, + 0.02559705823659897, + 0.02544991485774517, + 0.014557992108166218, + 0.012604925781488419, + 0.02433415688574314, + 0.02571035549044609, + 0.009400167502462864, + 0.01639937423169613, + 0.017534179612994194, + 0.019167734310030937, + 0.010285643860697746, + 0.006153079215437174, + 0.029818255454301834, + 0.016256891191005707, + 0.009422681294381618, + 0.01226841937750578, + 0.015618555247783661, + 0.010934772901237011, + 0.011689338833093643, + 0.009053926914930344, + 0.013952160254120827, + 0.012830127961933613, + 0.018313979730010033, + 0.014816062524914742, + 0.034420453011989594, + 0.01795334368944168, + 0.011707735247910023, + 0.00957462377846241, + 0.01836487278342247, + 0.03492087125778198, + 0.016839610412716866, + 0.02481393702328205, + 0.02512086182832718, + 0.030618486925959587, + 0.010922799818217754, + 0.011563340201973915, + 0.029235299676656723, + 0.022824758663773537, + 0.005852715577930212, + 0.010436135344207287, + 0.006053329445421696, + 0.007693837396800518, + 0.003860409837216139, + 0.0027469564229249954, + 0.0063739256002008915, + 0.0058577656745910645, + 0.00792773999273777, + 0.012040399014949799, + 0.009430643171072006, + 0.012679936364293098, + 0.005358763970434666, + 0.004137990064918995, + 0.012355191633105278, + 0.008864729665219784, + 0.0028398570138961077, + 0.002362913219258189, + 0.004957287106662989, + 0.0025576953776180744, + 0.0018724100664258003, + 0.0011846194975078106, + 0.0028601016383618116, + 0.004718188662081957, + 0.0065211886540055275, + 0.0070329755544662476, + 0.01227445900440216, + 0.00882383156567812, + 0.006123345345258713, + 0.005283697508275509, + 0.009929501451551914, + 0.012122491374611855, + 0.005802407395094633, + 0.005213081371039152, + 0.010265999473631382, + 0.00972807127982378, + 0.004852405749261379, + 0.0024629856925457716, + 0.006826725322753191, + 0.008355660364031792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.023691311478614807 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up focused on the mid-section of a birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the birch bark peel away in thin, translucent strips to reveal a chalky, soft layer beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves heart-shaped, vibrant dewy green, and marked by fine, intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves tremble in a constant, subtle movement caused by a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of cool, misty fields with greens and grays?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054987695068120956 + ], + "opposite_logprob_tokens": [ + -2.929987668991089 + ], + "logprob_reward": -0.054987695068120956, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054987695068120956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15604163706302643, + 0.08782520890235901, + 0.056620899587869644, + 0.07989899814128876, + 0.14720101654529572, + 0.06936022639274597, + 0.1404610276222229, + 0.074783556163311, + 0.05587067827582359, + 0.13193678855895996 + ], + "credit_map": [ + 0.0230073444545269, + 0.017920490354299545, + 0.028036270290613174, + 0.019864320755004883, + 0.011350211687386036, + 0.009068096056580544, + 0.02153707854449749, + 0.02525782585144043, + 0.007894142530858517, + 0.010694749653339386, + 0.014237916097044945, + 0.013580208644270897, + 0.010168913751840591, + 0.005771817173808813, + 0.012874881736934185, + 0.012602587230503559, + 0.006965323351323605, + 0.005585303995758295, + 0.011160888709127903, + 0.006820073351264, + 0.007260750979185104, + 0.0046339016407728195, + 0.00545430788770318, + 0.008740350604057312, + 0.008931517601013184, + 0.009145967662334442, + 0.012103751301765442, + 0.009312039241194725, + 0.010184128768742085, + 0.009669264778494835, + 0.009803487919270992, + 0.010748839005827904, + 0.011801712214946747, + 0.02568642981350422, + 0.01851942390203476, + 0.021760722622275352, + 0.013861333020031452, + 0.012012413702905178, + 0.02772907167673111, + 0.015829913318157196, + 0.00807862263172865, + 0.008665989153087139, + 0.011217267252504826, + 0.009103440679609776, + 0.0063461982645094395, + 0.005572536028921604, + 0.011086822487413883, + 0.009289342910051346, + 0.027107395231723785, + 0.01125924102962017, + 0.03332548961043358, + 0.014328009448945522, + 0.007261789869517088, + 0.008850036188960075, + 0.010622707195580006, + 0.027706364169716835, + 0.006998057011514902, + 0.009961581788957119, + 0.012748205102980137, + 0.01570078916847706, + 0.0047616600058972836, + 0.0036471805069595575, + 0.008819280192255974, + 0.012146806344389915, + 0.008062394335865974, + 0.006977721117436886, + 0.009481022134423256, + 0.007531402632594109, + 0.0046710786409676075, + 0.0031692469492554665, + 0.008154667913913727, + 0.007823142223060131, + 0.013136387802660465, + 0.013230389915406704, + 0.02759905904531479, + 0.017003221437335014, + 0.009264693595468998, + 0.007689745165407658, + 0.015162026509642601, + 0.028851259499788284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.054987695068120956 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up focused on the mid-section of a birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the birch bark peel away in thin, translucent strips to reveal a chalky, soft layer beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves heart-shaped, vibrant dewy green, and marked by fine, intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves tremble in a constant, subtle movement caused by a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of cool, misty fields with greens and grays?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012558309361338615 + ], + "opposite_logprob_tokens": [ + -4.387558460235596 + ], + "logprob_reward": -0.012558309361338615, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012558309361338615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22079920768737793, + 0.08651169389486313, + 0.05676214024424553, + 0.17764995992183685, + 0.18710440397262573, + 0.031949739903211594, + 0.05564083531498909, + 0.03381521999835968, + 0.03101615235209465, + 0.11875063925981522 + ], + "credit_map": [ + 0.03458448871970177, + 0.019106026738882065, + 0.030405789613723755, + 0.031685106456279755, + 0.021790169179439545, + 0.021290382370352745, + 0.032239243388175964, + 0.029697980731725693, + 0.008556759916245937, + 0.012582704424858093, + 0.01304608304053545, + 0.011990486644208431, + 0.011597171425819397, + 0.0061965323984622955, + 0.01153420377522707, + 0.011007753200829029, + 0.009002345614135265, + 0.006419084500521421, + 0.010577136650681496, + 0.007069609593600035, + 0.005530102178454399, + 0.003427626332268119, + 0.006062332075089216, + 0.008673899807035923, + 0.014504064805805683, + 0.025538088753819466, + 0.022360332310199738, + 0.020612241700291634, + 0.02495708130300045, + 0.024609703570604324, + 0.02311643399298191, + 0.021952006965875626, + 0.014597075060009956, + 0.02541714534163475, + 0.026236452162265778, + 0.0327252522110939, + 0.015698831528425217, + 0.01536004152148962, + 0.0316288135945797, + 0.025440797209739685, + 0.0037953949067741632, + 0.005027064122259617, + 0.004301465582102537, + 0.004665847402065992, + 0.002992807887494564, + 0.002183894393965602, + 0.004334285389631987, + 0.0046489788219332695, + 0.005480943713337183, + 0.006986717227846384, + 0.008503913879394531, + 0.004935455974191427, + 0.006788244470953941, + 0.00647400040179491, + 0.006820566486567259, + 0.009650987572968006, + 0.002709675580263138, + 0.004115764982998371, + 0.0042979177087545395, + 0.005831107031553984, + 0.0031971598509699106, + 0.003687687451019883, + 0.005675069987773895, + 0.0043008397333323956, + 0.004399085883051157, + 0.0027868936304003, + 0.006906832568347454, + 0.0029565070290118456, + 0.0031871062237769365, + 0.00187857320997864, + 0.0034328128676861525, + 0.005468341056257486, + 0.012527944520115852, + 0.011133778840303421, + 0.0261270459741354, + 0.018855849280953407, + 0.009760294109582901, + 0.004252784885466099, + 0.015343839302659035, + 0.02074909768998623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.012558309361338615 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up focused on the mid-section of a birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the birch bark peel away in thin, translucent strips to reveal a chalky, soft layer beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves heart-shaped, vibrant dewy green, and marked by fine, intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves tremble in a constant, subtle movement caused by a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of cool, misty fields with greens and grays?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866136610507965 + ], + "opposite_logprob_tokens": [ + -3.048661470413208 + ], + "logprob_reward": -0.04866136610507965, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866136610507965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16121214628219604, + 0.08507904410362244, + 0.07033323496580124, + 0.1735355108976364, + 0.15512868762016296, + 0.10329761356115341, + 0.04904460534453392, + 0.023765822872519493, + 0.08473551273345947, + 0.09386786818504333 + ], + "credit_map": [ + 0.01966811716556549, + 0.019873348996043205, + 0.02595318853855133, + 0.019401906058192253, + 0.016737671568989754, + 0.012875033542513847, + 0.020113466307520866, + 0.026589401066303253, + 0.010789032094180584, + 0.009819092229008675, + 0.014777874574065208, + 0.010468866676092148, + 0.008809332735836506, + 0.005566118750721216, + 0.01153479516506195, + 0.013313932344317436, + 0.0060848044231534, + 0.009901512414216995, + 0.010285289026796818, + 0.008751961402595043, + 0.007428592070937157, + 0.007145567797124386, + 0.011465173214673996, + 0.009270334616303444, + 0.01613483391702175, + 0.016819577664136887, + 0.029155516996979713, + 0.021065015345811844, + 0.01649332232773304, + 0.021062033250927925, + 0.02931065484881401, + 0.02349456399679184, + 0.013150515966117382, + 0.02304990403354168, + 0.024344487115740776, + 0.026460282504558563, + 0.011369128711521626, + 0.010579201392829418, + 0.024504322558641434, + 0.02167084999382496, + 0.014628242701292038, + 0.017901839688420296, + 0.015444070100784302, + 0.015327656641602516, + 0.005653651896864176, + 0.004151422064751387, + 0.013353781774640083, + 0.016836950555443764, + 0.005916685331612825, + 0.005730900447815657, + 0.00953041110187769, + 0.0065529788844287395, + 0.004543270915746689, + 0.0043443478643894196, + 0.004623290151357651, + 0.007802722044289112, + 0.002948984270915389, + 0.003071094397455454, + 0.004291577730327845, + 0.003246897831559181, + 0.0014387615956366062, + 0.0014338071923702955, + 0.003477912163361907, + 0.0038567876908928156, + 0.008998801931738853, + 0.010455355048179626, + 0.010812388733029366, + 0.012886936776340008, + 0.007883722893893719, + 0.0064428080804646015, + 0.015523014590144157, + 0.011732491664588451, + 0.008149098604917526, + 0.009803706780076027, + 0.017917949706315994, + 0.016983360052108765, + 0.008146103471517563, + 0.0034684576094150543, + 0.014329311437904835, + 0.015069879591464996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.04866136610507965 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up focused on the mid-section of a birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the birch bark peel away in thin, translucent strips to reveal a chalky, soft layer beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves heart-shaped, vibrant dewy green, and marked by fine, intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves tremble in a constant, subtle movement caused by a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of cool, misty fields with greens and grays?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013366345083341002 + ], + "opposite_logprob_tokens": [ + -6.626336574554443 + ], + "logprob_reward": -0.0013366345083341002, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013366345083341002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21766023337841034, + 0.07686333358287811, + 0.1246776208281517, + 0.13834433257579803, + 0.09552352875471115, + 0.06502608954906464, + 0.03606574982404709, + 0.07871178537607193, + 0.08630084246397018, + 0.08082649856805801 + ], + "credit_map": [ + 0.020692115649580956, + 0.017294269055128098, + 0.051833223551511765, + 0.02204575575888157, + 0.016786780208349228, + 0.009565047919750214, + 0.016003886237740517, + 0.06343916058540344, + 0.01021316833794117, + 0.008456851355731487, + 0.016199367120862007, + 0.010367942042648792, + 0.006608658004552126, + 0.0031350369099527597, + 0.010722691193223, + 0.011159618385136127, + 0.01370176300406456, + 0.015965063124895096, + 0.01989116705954075, + 0.018202537670731544, + 0.011905936524271965, + 0.008150672540068626, + 0.019208217039704323, + 0.017652258276939392, + 0.02303648367524147, + 0.01363409124314785, + 0.03208627551794052, + 0.014028227888047695, + 0.013816149905323982, + 0.008588427677750587, + 0.010029693134129047, + 0.023124994710087776, + 0.013614436611533165, + 0.01264223549515009, + 0.013618342578411102, + 0.015689503401517868, + 0.006996810436248779, + 0.005975118838250637, + 0.013980958610773087, + 0.013006128370761871, + 0.014229646883904934, + 0.006102710496634245, + 0.014484573155641556, + 0.007361797150224447, + 0.003591586835682392, + 0.0033617650624364614, + 0.0054543716832995415, + 0.01043963897973299, + 0.00554997893050313, + 0.0028733627405017614, + 0.007873096503317356, + 0.0036184177733957767, + 0.002408392261713743, + 0.001746890600770712, + 0.002686944557353854, + 0.009308669716119766, + 0.016270458698272705, + 0.007267671171575785, + 0.01655547507107258, + 0.010885501280426979, + 0.0037113509606570005, + 0.0050326865166425705, + 0.006558925379067659, + 0.012429716996848583, + 0.015168571844696999, + 0.00823156163096428, + 0.020201891660690308, + 0.009969797916710377, + 0.004748695995658636, + 0.004794855136424303, + 0.00783677026629448, + 0.015348697081208229, + 0.007410041056573391, + 0.008508879691362381, + 0.011642947793006897, + 0.01606626808643341, + 0.007010183297097683, + 0.003986269235610962, + 0.01423563715070486, + 0.011966271325945854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.0013366345083341002 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_0/index_tensor([656, 656])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_0/index_tensor([656, 656])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ec6f66d3d8a9049dec09976c2286781f3813bc1a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_0/index_tensor([656, 656])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, smooth, low-angle tracking shot moves along a frozen riverbank, frost-covered and gleaming coldly in the dim moonlight. Broken ice fragments appear and disappear in the nearly frozen black water. As the camera slowly stops, it reveals its subject, hidden in the deep shadow cast by a withered tree on the bank: a massive Rough-legged Fishing Owl. Its thick, fluffy feathers, a mottled palette of dark brown and light brown, almost blend into the darkness. Its body leans slightly forward, crouching in a poised posture, motionless. In the near-total darkness, its unsettlingly piercing yellow eyes, like two burning amber lamps, pierce the deep blue and black night, becoming the sole source of light and focal point in the entire frame, its gaze fixed firmly on a point beyond the camera. The texture of the entire scene is like a Rembrandt-style thick oil painting, with extremely strong contrast between light and shadow. Deep Prussian blue and pure black constitute the main colors of the picture, creating a breathtaking and silent atmosphere full of dramatic tension.", + "seed": 24741, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_0/index_656_seed_24741_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02764750234782696, + "vlm_logprob": -0.02764750234782696, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.02764750234782696, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a frozen riverbank with broken ice fragments in nearly frozen black water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive Rough-legged Fishing Owl crouching in a poised, motionless posture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl have thick fluffy feathers with a mottled palette of dark brown and light brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the owl's eyes piercing yellow, acting as the sole source of light in the deep blue and black night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow, smooth, low-angle tracking shot that stops to reveal the owl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04715709388256073 + ], + "opposite_logprob_tokens": [ + -3.172157049179077 + ], + "logprob_reward": -0.04715709388256073, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04715709388256073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23296357691287994, + 0.1421443670988083, + 0.0819346234202385, + 0.35163363814353943, + 0.07198445498943329, + 0.013598980382084846, + 0.027746044099330902, + 0.01657225377857685, + 0.03844087943434715, + 0.02298116497695446 + ], + "credit_map": [ + 0.04720325767993927, + 0.031045865267515182, + 0.024823471903800964, + 0.02835577167570591, + 0.026016784831881523, + 0.026912709698081017, + 0.025979993864893913, + 0.022625738754868507, + 0.01550016738474369, + 0.017131399363279343, + 0.028583314269781113, + 0.01949993334710598, + 0.011458825320005417, + 0.010056234896183014, + 0.018395744264125824, + 0.02151876501739025, + 0.01100677065551281, + 0.013617527671158314, + 0.013913213275372982, + 0.008471229113638401, + 0.010727625340223312, + 0.009326319210231304, + 0.007175239268690348, + 0.007696706801652908, + 0.03467206284403801, + 0.038912542164325714, + 0.05278407782316208, + 0.035913821309804916, + 0.06951172649860382, + 0.04944567382335663, + 0.03206101804971695, + 0.038332726806402206, + 0.006870830897241831, + 0.005798663944005966, + 0.017211927101016045, + 0.00748309725895524, + 0.006904893554747105, + 0.01808731071650982, + 0.003699597669765353, + 0.00592813640832901, + 0.0014039662200957537, + 0.0013301429571583867, + 0.0022641774266958237, + 0.0016628665616735816, + 0.0018744633998721838, + 0.0019692787900567055, + 0.001464229659177363, + 0.0016298566479235888, + 0.0015857921680435538, + 0.0034039681777358055, + 0.0038202106952667236, + 0.00537006463855505, + 0.005045637022703886, + 0.0037689392920583487, + 0.0026615073438733816, + 0.0020899279043078423, + 0.0010246863821521401, + 0.002382048638537526, + 0.002185429912060499, + 0.0026271967217326164, + 0.0024988981895148754, + 0.002683025086298585, + 0.001538169919513166, + 0.0016327996272593737, + 0.0031671947799623013, + 0.004866302944719791, + 0.005640687420964241, + 0.005433152429759502, + 0.006292314734309912, + 0.005706810392439365, + 0.003423459129408002, + 0.003910961095243692, + 0.0015765295829623938, + 0.002947932342067361, + 0.00419986667111516, + 0.002701070625334978, + 0.0036539179272949696, + 0.004419285804033279, + 0.0018002029974013567, + 0.00168236147146672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.04715709388256073 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a frozen riverbank with broken ice fragments in nearly frozen black water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive Rough-legged Fishing Owl crouching in a poised, motionless posture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl have thick fluffy feathers with a mottled palette of dark brown and light brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the owl's eyes piercing yellow, acting as the sole source of light in the deep blue and black night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow, smooth, low-angle tracking shot that stops to reveal the owl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254771463572979 + ], + "opposite_logprob_tokens": [ + -4.387547492980957 + ], + "logprob_reward": -0.01254771463572979, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254771463572979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09068341553211212, + 0.10797929763793945, + 0.12890644371509552, + 0.3586339056491852, + 0.08557061105966568, + 0.037605732679367065, + 0.03861989453434944, + 0.0942184329032898, + 0.038470447063446045, + 0.01931180991232395 + ], + "credit_map": [ + 0.012260681949555874, + 0.013667717576026917, + 0.008726476691663265, + 0.010258160531520844, + 0.010641074739396572, + 0.012558314949274063, + 0.01510796882212162, + 0.007463015615940094, + 0.01139579527080059, + 0.03173310309648514, + 0.01451521273702383, + 0.017618436366319656, + 0.007493684068322182, + 0.007095710840076208, + 0.009456457570195198, + 0.00867090281099081, + 0.01515887025743723, + 0.019506268203258514, + 0.014241728000342846, + 0.018008850514888763, + 0.01759178377687931, + 0.025970345363020897, + 0.007676875684410334, + 0.010751712135970592, + 0.031317055225372314, + 0.04328085109591484, + 0.06041349098086357, + 0.04375549033284187, + 0.06409303843975067, + 0.04656641185283661, + 0.032871901988983154, + 0.03633566200733185, + 0.005624765995889902, + 0.013780016452074051, + 0.015930455178022385, + 0.012912393547594547, + 0.008594566956162453, + 0.016985198482871056, + 0.006071623880416155, + 0.005671591032296419, + 0.002222399227321148, + 0.004715979099273682, + 0.005526380613446236, + 0.0075136893428862095, + 0.00564246391877532, + 0.0053613497875630856, + 0.003881835611537099, + 0.0027416362427175045, + 0.003362028393894434, + 0.004911789204925299, + 0.005546989850699902, + 0.00551050016656518, + 0.007692032027989626, + 0.0050353859551250935, + 0.003155704587697983, + 0.0034054648131132126, + 0.005592893343418837, + 0.011671793647110462, + 0.012658927589654922, + 0.01736532151699066, + 0.01775270514190197, + 0.013466146774590015, + 0.008645393885672092, + 0.007065252400934696, + 0.002539278706535697, + 0.005611979868263006, + 0.00542051438242197, + 0.006239698268473148, + 0.005394480656832457, + 0.006274989806115627, + 0.0035153476055711508, + 0.0034741568379104137, + 0.0014381660148501396, + 0.0022163039539009333, + 0.003097713692113757, + 0.0031257830560207367, + 0.003217739751562476, + 0.0025939319748431444, + 0.0019999241922050714, + 0.0016222468111664057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 1 + ], + "selected_reward": -0.01254771463572979 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a frozen riverbank with broken ice fragments in nearly frozen black water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive Rough-legged Fishing Owl crouching in a poised, motionless posture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl have thick fluffy feathers with a mottled palette of dark brown and light brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the owl's eyes piercing yellow, acting as the sole source of light in the deep blue and black night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow, smooth, low-angle tracking shot that stops to reveal the owl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006740570068359375 + ], + "opposite_logprob_tokens": [ + -5.006740570068359 + ], + "logprob_reward": -0.006740570068359375, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006740570068359375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16085894405841827, + 0.10509888827800751, + 0.054595138877630234, + 0.19910266995429993, + 0.13322713971138, + 0.02907184138894081, + 0.09686332195997238, + 0.14699053764343262, + 0.0489056259393692, + 0.025285951793193817 + ], + "credit_map": [ + 0.018067585304379463, + 0.01639457978308201, + 0.015935268253087997, + 0.019938519224524498, + 0.017844073474407196, + 0.027886848896741867, + 0.020603254437446594, + 0.024188809096813202, + 0.011475145816802979, + 0.024031467735767365, + 0.016966920346021652, + 0.014739601872861385, + 0.0073068588972091675, + 0.006608885712921619, + 0.010527892969548702, + 0.013442101888358593, + 0.006731933914124966, + 0.008541932329535484, + 0.008403636515140533, + 0.006271039601415396, + 0.0075454991310834885, + 0.006920315325260162, + 0.00538995023816824, + 0.004790823440998793, + 0.014012613333761692, + 0.026712926104664803, + 0.043955691158771515, + 0.03782968968153, + 0.024745166301727295, + 0.027184894308447838, + 0.012505882419645786, + 0.012155769392848015, + 0.008268200792372227, + 0.017796874046325684, + 0.0218084417283535, + 0.022447574883699417, + 0.019790831953287125, + 0.022994056344032288, + 0.011902492493391037, + 0.00821866188198328, + 0.0027512528467923403, + 0.004155104514211416, + 0.004338288679718971, + 0.004145285580307245, + 0.004488800186663866, + 0.0041083949618041515, + 0.0025866341311484575, + 0.002498079091310501, + 0.005494910292327404, + 0.011391686275601387, + 0.01345922239124775, + 0.017599642276763916, + 0.01956130936741829, + 0.012293334119021893, + 0.009067565202713013, + 0.007995645515620708, + 0.008487381041049957, + 0.01913287304341793, + 0.01922191120684147, + 0.027897216379642487, + 0.026863515377044678, + 0.020671317353844643, + 0.013484567403793335, + 0.011231747455894947, + 0.003431454300880432, + 0.006142860744148493, + 0.006762634497135878, + 0.007449156139045954, + 0.008523846045136452, + 0.007366596255451441, + 0.0046823713928461075, + 0.004546700976788998, + 0.001619514892809093, + 0.003294244408607483, + 0.004061452578753233, + 0.004675417672842741, + 0.0031017151195555925, + 0.0037403376772999763, + 0.0025419951416552067, + 0.00225127418525517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.006740570068359375 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a frozen riverbank with broken ice fragments in nearly frozen black water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive Rough-legged Fishing Owl crouching in a poised, motionless posture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl have thick fluffy feathers with a mottled palette of dark brown and light brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the owl's eyes piercing yellow, acting as the sole source of light in the deep blue and black night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow, smooth, low-angle tracking shot that stops to reveal the owl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038105905055999756 + ], + "opposite_logprob_tokens": [ + -3.2881059646606445 + ], + "logprob_reward": -0.038105905055999756, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038105905055999756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07454927265644073, + 0.10821203887462616, + 0.07352803647518158, + 0.3697773516178131, + 0.0879916101694107, + 0.024126386269927025, + 0.08369075506925583, + 0.10177023708820343, + 0.051929205656051636, + 0.024425044655799866 + ], + "credit_map": [ + 0.007638893090188503, + 0.008328450843691826, + 0.007292279973626137, + 0.0068901884369552135, + 0.007928107865154743, + 0.01091759279370308, + 0.015412663109600544, + 0.01014111191034317, + 0.01040791254490614, + 0.014323599636554718, + 0.016548287123441696, + 0.014431498013436794, + 0.014581807889044285, + 0.014427507296204567, + 0.011164533905684948, + 0.012326901778578758, + 0.008895568549633026, + 0.013095305301249027, + 0.007734974846243858, + 0.01020498014986515, + 0.009815921075642109, + 0.009484422393143177, + 0.006969524081796408, + 0.007327347993850708, + 0.03540128469467163, + 0.03788970038294792, + 0.0577484667301178, + 0.03896640986204147, + 0.06967616826295853, + 0.05060587450861931, + 0.03389200568199158, + 0.04559749364852905, + 0.005591295659542084, + 0.014387345872819424, + 0.014609555713832378, + 0.012586478143930435, + 0.012023244984447956, + 0.015414418652653694, + 0.007919185794889927, + 0.005460094660520554, + 0.002089816378429532, + 0.00304163689725101, + 0.00360071356408298, + 0.0033097679261118174, + 0.0029447015840560198, + 0.003751032054424286, + 0.0022846146021038294, + 0.0031041069887578487, + 0.005058889277279377, + 0.010754579678177834, + 0.011114035733044147, + 0.015614093281328678, + 0.014096472412347794, + 0.011756469495594501, + 0.00825649406760931, + 0.0070397332310676575, + 0.005752466153353453, + 0.011918495409190655, + 0.014306078664958477, + 0.019323833286762238, + 0.019988972693681717, + 0.01413634978234768, + 0.008613597601652145, + 0.007730462122708559, + 0.0032563279382884502, + 0.00747672887519002, + 0.006788074504584074, + 0.007955044507980347, + 0.008118418045341969, + 0.008084848523139954, + 0.005397172644734383, + 0.004852598067373037, + 0.001665621530264616, + 0.0036580886226147413, + 0.0036860816180706024, + 0.0034945064689964056, + 0.003995809238404036, + 0.003959053661674261, + 0.0020396397449076176, + 0.0019262455170974135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 7 + ], + "selected_reward": -0.038105905055999756 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a frozen riverbank with broken ice fragments in nearly frozen black water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive Rough-legged Fishing Owl crouching in a poised, motionless posture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl have thick fluffy feathers with a mottled palette of dark brown and light brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the owl's eyes piercing yellow, acting as the sole source of light in the deep blue and black night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow, smooth, low-angle tracking shot that stops to reveal the owl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368622809648514 + ], + "opposite_logprob_tokens": [ + -3.4086861610412598 + ], + "logprob_reward": -0.03368622809648514, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368622809648514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15937437117099762, + 0.1796034872531891, + 0.13243216276168823, + 0.0753713846206665, + 0.045668844133615494, + 0.04280901327729225, + 0.045190148055553436, + 0.18236833810806274, + 0.12149423360824585, + 0.015688035637140274 + ], + "credit_map": [ + 0.019287047907710075, + 0.017299069091677666, + 0.020580600947141647, + 0.024384882301092148, + 0.018976125866174698, + 0.025012539699673653, + 0.02188095822930336, + 0.011953159235417843, + 0.0241702850908041, + 0.02297702245414257, + 0.03626149892807007, + 0.027092549949884415, + 0.014097125269472599, + 0.015518835745751858, + 0.01904047280550003, + 0.020445693284273148, + 0.014051726087927818, + 0.027262458577752113, + 0.015665780752897263, + 0.013967175036668777, + 0.020571626722812653, + 0.020122237503528595, + 0.009586666710674763, + 0.011204494163393974, + 0.00710895424708724, + 0.008529813028872013, + 0.0110080037266016, + 0.007593309506773949, + 0.014400362968444824, + 0.011471239849925041, + 0.007054722867906094, + 0.008204985409975052, + 0.003306310623884201, + 0.00380636821500957, + 0.010574139654636383, + 0.004315339494496584, + 0.006099878344684839, + 0.009747345000505447, + 0.0038131107576191425, + 0.0040063560009002686, + 0.0033937019761651754, + 0.0045170909725129604, + 0.007005500141531229, + 0.007012330461293459, + 0.006344215478748083, + 0.005901409313082695, + 0.004683210048824549, + 0.003951555117964745, + 0.003768522758036852, + 0.005081205628812313, + 0.00662121269851923, + 0.006692466326057911, + 0.00779295340180397, + 0.0064082639291882515, + 0.005605925340205431, + 0.0032195996027439833, + 0.01070738397538662, + 0.021505141630768776, + 0.025336703285574913, + 0.033754535019397736, + 0.033351458609104156, + 0.02540445514023304, + 0.01795915514230728, + 0.014349511824548244, + 0.007760825101286173, + 0.01493888907134533, + 0.016540126875042915, + 0.020653506740927696, + 0.02055925689637661, + 0.018025873228907585, + 0.012457567267119884, + 0.01055818423628807, + 0.0010951996082440019, + 0.001845383900217712, + 0.0026551370974630117, + 0.0025359170977026224, + 0.002237281994894147, + 0.002218478126451373, + 0.0013751143123954535, + 0.0017255244310945272 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.03368622809648514 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, smooth, low-angle tracking shot moves along a frozen riverbank, frost-covered and gleaming coldly in the dim moonlight. Broken ice fragments appear and disappear in the nearly frozen black water. As the camera slowly stops, it reveals its subject, hidden in the deep shadow cast by a withered tree on the bank: a massive Rough-legged Fishing Owl. Its thick, fluffy feathers, a mottled palette of dark brown and light brown, almost blend into the darkness. Its body leans slightly forward, crouching in a poised posture, motionless. In the near-total darkness, its unsettlingly piercing yellow eyes, like two burning amber lamps, pierce the deep blue and black night, becoming the sole source of light and focal point in the entire frame, its gaze fixed firmly on a point beyond the camera. The texture of the entire scene is like a Rembrandt-style thick oil painting, with extremely strong contrast between light and shadow. Deep Prussian blue and pure black constitute the main colors of the picture, creating a breathtaking and silent atmosphere full of dramatic tension.", + "seed": 24742, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_0/index_656_seed_24742_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.026158180832862855, + "vlm_logprob": -0.026158180832862855, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.026158180832862855, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a frozen riverbank with broken ice fragments in nearly frozen black water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive Rough-legged Fishing Owl crouching in a poised, motionless posture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl have thick fluffy feathers with a mottled palette of dark brown and light brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the owl's eyes piercing yellow, acting as the sole source of light in the deep blue and black night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow, smooth, low-angle tracking shot that stops to reveal the owl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017827512696385384 + ], + "opposite_logprob_tokens": [ + -4.26782751083374 + ], + "logprob_reward": -0.017827512696385384, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.017827512696385384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16482363641262054, + 0.022410890087485313, + 0.032513368874788284, + 0.02644830197095871, + 0.10377687215805054, + 0.019005734473466873, + 0.01494675874710083, + 0.2916354238986969, + 0.29313117265701294, + 0.03130783513188362 + ], + "credit_map": [ + 0.009408288635313511, + 0.0182656217366457, + 0.009464780800044537, + 0.017458366230130196, + 0.043650560081005096, + 0.04440369829535484, + 0.014810031279921532, + 0.007362302392721176, + 0.0026067106518894434, + 0.004091572016477585, + 0.0025070547126233578, + 0.0022463055793195963, + 0.003833342809230089, + 0.003054617205634713, + 0.00218561221845448, + 0.0018856775714084506, + 0.002542852656915784, + 0.004097375553101301, + 0.004243312869220972, + 0.0065513900481164455, + 0.004341434221714735, + 0.003018063260242343, + 0.003525173058733344, + 0.00419376976788044, + 0.002899592975154519, + 0.0026698799338191748, + 0.0038150593172758818, + 0.0030122154857963324, + 0.004128413740545511, + 0.0033119628205895424, + 0.0038771890103816986, + 0.0027339907828718424, + 0.010828075930476189, + 0.011264562606811523, + 0.01761375181376934, + 0.011753222905099392, + 0.016841210424900055, + 0.011880884878337383, + 0.01064700074493885, + 0.012948175892233849, + 0.0015194620937108994, + 0.002322157844901085, + 0.003787693567574024, + 0.0018550478853285313, + 0.0019245067378506064, + 0.0025050598196685314, + 0.0021786668803542852, + 0.002913141855970025, + 0.0012343950802460313, + 0.0021790964528918266, + 0.0027886119205504656, + 0.0022388058714568615, + 0.0018498832359910011, + 0.001431078533641994, + 0.0017620804719626904, + 0.0014628083445131779, + 0.020686518400907516, + 0.061237309128046036, + 0.0371151939034462, + 0.042764876037836075, + 0.050257984548807144, + 0.025640372186899185, + 0.031209519132971764, + 0.022723659873008728, + 0.013385321013629436, + 0.009591867215931416, + 0.11053615063428879, + 0.014726677909493446, + 0.009283080697059631, + 0.11110502481460571, + 0.011445079930126667, + 0.013057999312877655, + 0.002987713785842061, + 0.0028964008670300245, + 0.004601103253662586, + 0.004225919023156166, + 0.0035070779267698526, + 0.004008801653981209, + 0.00436068419367075, + 0.004720134660601616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.017827512696385384 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a frozen riverbank with broken ice fragments in nearly frozen black water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive Rough-legged Fishing Owl crouching in a poised, motionless posture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl have thick fluffy feathers with a mottled palette of dark brown and light brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the owl's eyes piercing yellow, acting as the sole source of light in the deep blue and black night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow, smooth, low-angle tracking shot that stops to reveal the owl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026353394612669945 + ], + "opposite_logprob_tokens": [ + -3.651353359222412 + ], + "logprob_reward": -0.026353394612669945, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026353394612669945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06691400706768036, + 0.016915490850806236, + 0.013302193023264408, + 0.012219112366437912, + 0.017118636518716812, + 0.011406327597796917, + 0.03282930701971054, + 0.3900229036808014, + 0.37109383940696716, + 0.06817825138568878 + ], + "credit_map": [ + 0.0034420702140778303, + 0.005577520001679659, + 0.004674370400607586, + 0.005285653285682201, + 0.01833566464483738, + 0.019942987710237503, + 0.004878009669482708, + 0.0047777313739061356, + 0.002724661258980632, + 0.004016696475446224, + 0.001580181997269392, + 0.0010601818794384599, + 0.0031067028176039457, + 0.0025475886650383472, + 0.0010543606476858258, + 0.0008251167018897831, + 0.0010304577881470323, + 0.0017086906591430306, + 0.001766011118888855, + 0.0025197574868798256, + 0.0018125901697203517, + 0.0012847891775891185, + 0.0017314334399998188, + 0.0014484627172350883, + 0.001522717415355146, + 0.001329478109255433, + 0.0016517704352736473, + 0.0014050619211047888, + 0.0018124999478459358, + 0.0015421571442857385, + 0.0017501418478786945, + 0.0012052853126078844, + 0.0019405940547585487, + 0.003180423518642783, + 0.0020342343486845493, + 0.0024291377048939466, + 0.002358783269301057, + 0.0012773718917742372, + 0.0016871743137016892, + 0.002210917416960001, + 0.001107197953388095, + 0.001331191509962082, + 0.0020803671795874834, + 0.001509205438196659, + 0.0011675654677674174, + 0.0015591431874781847, + 0.0012630122946575284, + 0.0013886444503441453, + 0.0026807019021362066, + 0.004685555584728718, + 0.005420855712145567, + 0.004346967209130526, + 0.004318225663155317, + 0.0033184527419507504, + 0.004229316953569651, + 0.0038292300887405872, + 0.029072850942611694, + 0.08341395854949951, + 0.04617675021290779, + 0.056866355240345, + 0.06907697021961212, + 0.035774748772382736, + 0.03991571441292763, + 0.0297255776822567, + 0.01689319871366024, + 0.012937726452946663, + 0.13841187953948975, + 0.019723612815141678, + 0.012470723129808903, + 0.1383538544178009, + 0.015243416652083397, + 0.01705942489206791, + 0.005997847765684128, + 0.007405488286167383, + 0.012519460171461105, + 0.009087146259844303, + 0.004635388031601906, + 0.01441846415400505, + 0.006252320948988199, + 0.007862133905291557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.026353394612669945 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a frozen riverbank with broken ice fragments in nearly frozen black water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive Rough-legged Fishing Owl crouching in a poised, motionless posture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl have thick fluffy feathers with a mottled palette of dark brown and light brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the owl's eyes piercing yellow, acting as the sole source of light in the deep blue and black night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow, smooth, low-angle tracking shot that stops to reveal the owl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107780635356903 + ], + "opposite_logprob_tokens": [ + -4.511077880859375 + ], + "logprob_reward": -0.01107780635356903, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107780635356903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05338704586029053, + 0.010076365433633327, + 0.027838105335831642, + 0.009414003230631351, + 0.030389774590730667, + 0.026680374518036842, + 0.035946231335401535, + 0.45088881254196167, + 0.3109976649284363, + 0.04438153654336929 + ], + "credit_map": [ + 0.003039318835362792, + 0.005853759124875069, + 0.0029969012830406427, + 0.004876371473073959, + 0.013246641494333744, + 0.01580788381397724, + 0.0036516522523015738, + 0.003914519678801298, + 0.0012974153505638242, + 0.0015970650129020214, + 0.0016787729691714048, + 0.0008887505391612649, + 0.0014617317356169224, + 0.0014662351459264755, + 0.0007112524472177029, + 0.0009751427569426596, + 0.002275388455018401, + 0.004388477187603712, + 0.003526060376316309, + 0.004537044558674097, + 0.0037473791744560003, + 0.002387195359915495, + 0.0037341092247515917, + 0.0032424512319266796, + 0.001312884734943509, + 0.0010453822324052453, + 0.0012730094604194164, + 0.0011217412538826466, + 0.0013410281389951706, + 0.0010293014347553253, + 0.0012104602064937353, + 0.0010801967000588775, + 0.002697694581001997, + 0.004652621224522591, + 0.004081449005752802, + 0.004151109606027603, + 0.0036479055415838957, + 0.003615232650190592, + 0.0034560086205601692, + 0.004087754990905523, + 0.0022746133618056774, + 0.004313228651881218, + 0.004991387948393822, + 0.004018062260001898, + 0.002550492761656642, + 0.0026263222098350525, + 0.0032743425108492374, + 0.0026319262105971575, + 0.0028428537771105766, + 0.005582308396697044, + 0.005433039739727974, + 0.005039970390498638, + 0.005402211099863052, + 0.003971540369093418, + 0.00404354277998209, + 0.0036307654809206724, + 0.03296523541212082, + 0.09341787546873093, + 0.05813853070139885, + 0.06725817918777466, + 0.07792368531227112, + 0.04050923138856888, + 0.04731910303235054, + 0.03335701301693916, + 0.01430522371083498, + 0.009711114689707756, + 0.11578331142663956, + 0.015343266539275646, + 0.009321088902652264, + 0.12105357646942139, + 0.012383734807372093, + 0.013096385635435581, + 0.0031376637052744627, + 0.0031938478350639343, + 0.00981783028692007, + 0.005935227498412132, + 0.0029106978327035904, + 0.011291254311800003, + 0.003593273926526308, + 0.004501740448176861 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.01107780635356903 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a frozen riverbank with broken ice fragments in nearly frozen black water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive Rough-legged Fishing Owl crouching in a poised, motionless posture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl have thick fluffy feathers with a mottled palette of dark brown and light brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the owl's eyes piercing yellow, acting as the sole source of light in the deep blue and black night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow, smooth, low-angle tracking shot that stops to reveal the owl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496027693152428 + ], + "opposite_logprob_tokens": [ + -2.929960250854492 + ], + "logprob_reward": -0.05496027693152428, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496027693152428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04947181046009064, + 0.031941529363393784, + 0.018803589046001434, + 0.02184867300093174, + 0.030591752380132675, + 0.04218602925539017, + 0.036241788417100906, + 0.44519683718681335, + 0.13764260709285736, + 0.1860753744840622 + ], + "credit_map": [ + 0.002706116298213601, + 0.0036311049479991198, + 0.0037975956220179796, + 0.0035152628552168608, + 0.013264601118862629, + 0.015134228393435478, + 0.0033481987193226814, + 0.004074702039361, + 0.0037560458295047283, + 0.009018576703965664, + 0.002717680996283889, + 0.0027819934766739607, + 0.005142603535205126, + 0.0038749477826058865, + 0.0022742156870663166, + 0.0023754609283059835, + 0.0017300951294600964, + 0.00270581035874784, + 0.0023222891613841057, + 0.0031006440985947847, + 0.002911002142354846, + 0.001946349861100316, + 0.0022897750604897738, + 0.0017976216040551662, + 0.003723844652995467, + 0.002420929027721286, + 0.0021335172932595015, + 0.002514908090233803, + 0.0032623005099594593, + 0.002671573543921113, + 0.0029165924061089754, + 0.0022050056140869856, + 0.002865817863494158, + 0.0035961938556283712, + 0.0037918537855148315, + 0.0037615967448800802, + 0.0036501851864159107, + 0.0031486726365983486, + 0.004200011026114225, + 0.0055774194188416, + 0.004103081300854683, + 0.0054395669139921665, + 0.005987850949168205, + 0.0045316945761442184, + 0.003762293141335249, + 0.006468593142926693, + 0.0044960821978747845, + 0.007396864239126444, + 0.0027592675760388374, + 0.005644531454890966, + 0.005625363904982805, + 0.005369060672819614, + 0.0050327228382229805, + 0.004144002683460712, + 0.004397012293338776, + 0.003269822569563985, + 0.03154454007744789, + 0.08746375888586044, + 0.05814189091324806, + 0.06473524868488312, + 0.07648077607154846, + 0.04093706235289574, + 0.04792095348238945, + 0.037972573190927505, + 0.007678569294512272, + 0.005972207989543676, + 0.04369557648897171, + 0.008707677945494652, + 0.006367100402712822, + 0.04971819743514061, + 0.006362492218613625, + 0.00914077553898096, + 0.016657754778862, + 0.022567976266145706, + 0.02461938001215458, + 0.02227015048265457, + 0.022634072229266167, + 0.03812681511044502, + 0.017295129597187042, + 0.02190408855676651 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.05496027693152428 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a frozen riverbank with broken ice fragments in nearly frozen black water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive Rough-legged Fishing Owl crouching in a poised, motionless posture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl have thick fluffy feathers with a mottled palette of dark brown and light brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the owl's eyes piercing yellow, acting as the sole source of light in the deep blue and black night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow, smooth, low-angle tracking shot that stops to reveal the owl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020571913570165634 + ], + "opposite_logprob_tokens": [ + -3.8955719470977783 + ], + "logprob_reward": -0.020571913570165634, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020571913570165634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11049148440361023, + 0.04002263396978378, + 0.11875845491886139, + 0.0771188735961914, + 0.03735601529479027, + 0.12828995287418365, + 0.09935516119003296, + 0.19752487540245056, + 0.1106564998626709, + 0.08042607456445694 + ], + "credit_map": [ + 0.00835437048226595, + 0.014473773539066315, + 0.01231492217630148, + 0.012166362255811691, + 0.02173851802945137, + 0.01947362907230854, + 0.008612656965851784, + 0.01335724163800478, + 0.0039019568357616663, + 0.01066508051007986, + 0.002523588016629219, + 0.0036776969209313393, + 0.006768953055143356, + 0.006342156324535608, + 0.0027353940531611443, + 0.0034078063908964396, + 0.009812542237341404, + 0.02031480148434639, + 0.015552163124084473, + 0.02265966683626175, + 0.01727699115872383, + 0.00934507790952921, + 0.013058892451226711, + 0.010738306678831577, + 0.015107516199350357, + 0.008147800341248512, + 0.008379310369491577, + 0.008965590968728065, + 0.00937309954315424, + 0.00946290697902441, + 0.01081073097884655, + 0.0068719093687832355, + 0.003909106366336346, + 0.0036047794856131077, + 0.005351108964532614, + 0.004194419831037521, + 0.004913078621029854, + 0.004976005293428898, + 0.0045862519182264805, + 0.005821261089295149, + 0.011830151081085205, + 0.016054678708314896, + 0.028820976614952087, + 0.017468243837356567, + 0.011487023904919624, + 0.013819877989590168, + 0.016360532492399216, + 0.01244845800101757, + 0.008208390325307846, + 0.013708563521504402, + 0.017742883414030075, + 0.013476978056132793, + 0.011481071822345257, + 0.012488736771047115, + 0.011595649644732475, + 0.010652881115674973, + 0.015822168439626694, + 0.024487944319844246, + 0.04900757223367691, + 0.027795735746622086, + 0.020680934190750122, + 0.018377525731921196, + 0.022231686860322952, + 0.019121292978525162, + 0.0051740845665335655, + 0.003959547262638807, + 0.03842112049460411, + 0.005994553677737713, + 0.0036302406806498766, + 0.04294402524828911, + 0.004647713154554367, + 0.005885212682187557, + 0.00751178152859211, + 0.00899434369057417, + 0.010535310953855515, + 0.015525715425610542, + 0.008291371166706085, + 0.009300793521106243, + 0.009011146612465382, + 0.01125560887157917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 2 + ], + "selected_reward": -0.020571913570165634 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, smooth, low-angle tracking shot moves along a frozen riverbank, frost-covered and gleaming coldly in the dim moonlight. Broken ice fragments appear and disappear in the nearly frozen black water. As the camera slowly stops, it reveals its subject, hidden in the deep shadow cast by a withered tree on the bank: a massive Rough-legged Fishing Owl. Its thick, fluffy feathers, a mottled palette of dark brown and light brown, almost blend into the darkness. Its body leans slightly forward, crouching in a poised posture, motionless. In the near-total darkness, its unsettlingly piercing yellow eyes, like two burning amber lamps, pierce the deep blue and black night, becoming the sole source of light and focal point in the entire frame, its gaze fixed firmly on a point beyond the camera. The texture of the entire scene is like a Rembrandt-style thick oil painting, with extremely strong contrast between light and shadow. Deep Prussian blue and pure black constitute the main colors of the picture, creating a breathtaking and silent atmosphere full of dramatic tension.", + "seed": 24743, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_0/index_656_seed_24743_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.022971011698246002, + "vlm_logprob": -0.022971011698246002, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.022971011698246002, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a frozen riverbank with broken ice fragments in nearly frozen black water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive Rough-legged Fishing Owl crouching in a poised, motionless posture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl have thick fluffy feathers with a mottled palette of dark brown and light brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the owl's eyes piercing yellow, acting as the sole source of light in the deep blue and black night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow, smooth, low-angle tracking shot that stops to reveal the owl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018515722826123238 + ], + "opposite_logprob_tokens": [ + -4.393515586853027 + ], + "logprob_reward": -0.018515722826123238, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.018515722826123238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6085467338562012, + 0.050970930606126785, + 0.07923384755849838, + 0.04225147143006325, + 0.0777692198753357, + 0.027758484706282616, + 0.028054790571331978, + 0.030657896772027016, + 0.028449879959225655, + 0.026306714862585068 + ], + "credit_map": [ + 0.03316005691885948, + 0.10061180591583252, + 0.043121472001075745, + 0.13135860860347748, + 0.13610613346099854, + 0.08498644083738327, + 0.05493602529168129, + 0.024266181513667107, + 0.004053711891174316, + 0.006876444909721613, + 0.005007856525480747, + 0.0058122738264501095, + 0.009294012561440468, + 0.006425437517464161, + 0.007856931537389755, + 0.005644260440021753, + 0.007516443729400635, + 0.00819770060479641, + 0.007151692174375057, + 0.009688043966889381, + 0.013774649240076542, + 0.013598262332379818, + 0.012387076392769814, + 0.006919975858181715, + 0.004553896840661764, + 0.004375682678073645, + 0.002739317249506712, + 0.008378783240914345, + 0.00742323650047183, + 0.006599742453545332, + 0.005581737961620092, + 0.0025990766007453203, + 0.0063289012759923935, + 0.0071999249048531055, + 0.00957709550857544, + 0.010539105162024498, + 0.013595078140497208, + 0.012688085436820984, + 0.012887868098914623, + 0.0049531604163348675, + 0.003499874146655202, + 0.004260579589754343, + 0.001962590031325817, + 0.0030588002409785986, + 0.00520524475723505, + 0.0037882328033447266, + 0.0037582633085548878, + 0.0022249009925872087, + 0.00448994804173708, + 0.0036982898600399494, + 0.0017253340920433402, + 0.0035890734288841486, + 0.0037869319785386324, + 0.004317782819271088, + 0.004192299675196409, + 0.002255131723359227, + 0.0048930528573691845, + 0.004875113256275654, + 0.0025695625226944685, + 0.003595756134018302, + 0.003851033514365554, + 0.004249595105648041, + 0.003805645043030381, + 0.002818136941641569, + 0.003962193615734577, + 0.004326654598116875, + 0.0024625214282423258, + 0.004014140460640192, + 0.0034495287109166384, + 0.004501728340983391, + 0.0038041884545236826, + 0.0019289246993139386, + 0.003967077471315861, + 0.003974234685301781, + 0.002292242366820574, + 0.0035985885187983513, + 0.0033598891459405422, + 0.0038101414684206247, + 0.0029700309969484806, + 0.00233450997620821 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.018515722826123238 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a frozen riverbank with broken ice fragments in nearly frozen black water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive Rough-legged Fishing Owl crouching in a poised, motionless posture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl have thick fluffy feathers with a mottled palette of dark brown and light brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the owl's eyes piercing yellow, acting as the sole source of light in the deep blue and black night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow, smooth, low-angle tracking shot that stops to reveal the owl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014202586375176907 + ], + "opposite_logprob_tokens": [ + -4.26420259475708 + ], + "logprob_reward": -0.014202586375176907, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202586375176907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4616096019744873, + 0.07276634871959686, + 0.08935212343931198, + 0.055052097886800766, + 0.08717042952775955, + 0.06949325650930405, + 0.04148834943771362, + 0.040503352880477905, + 0.04680933430790901, + 0.03575514256954193 + ], + "credit_map": [ + 0.02258971892297268, + 0.08199805021286011, + 0.025868374854326248, + 0.09347798675298691, + 0.10803617537021637, + 0.07269758731126785, + 0.03971048817038536, + 0.017231177538633347, + 0.006093122530728579, + 0.007489200681447983, + 0.005873567424714565, + 0.007127195131033659, + 0.018004195764660835, + 0.014633222483098507, + 0.007468608673661947, + 0.006077232304960489, + 0.0075666047632694244, + 0.01089623011648655, + 0.007163982838392258, + 0.010948184877634048, + 0.016292374581098557, + 0.01722903922200203, + 0.014009758830070496, + 0.005245935171842575, + 0.00486713508144021, + 0.007504344452172518, + 0.0029975350480526686, + 0.010240347124636173, + 0.011007516644895077, + 0.007032437715679407, + 0.007802458014339209, + 0.0036003233399242163, + 0.00947620626538992, + 0.008858827874064445, + 0.006590661592781544, + 0.013404936529695988, + 0.014598994515836239, + 0.012863488867878914, + 0.014481180347502232, + 0.0068961381912231445, + 0.009000583551824093, + 0.00754341296851635, + 0.006793555337935686, + 0.007285524159669876, + 0.011281456798315048, + 0.009310959838330746, + 0.01188689935952425, + 0.006390859372913837, + 0.006278371438384056, + 0.00505194952711463, + 0.0035292021930217743, + 0.004383002407848835, + 0.0056253657676279545, + 0.006327948532998562, + 0.005966232158243656, + 0.004326277878135443, + 0.00784549955278635, + 0.003984704148024321, + 0.002614708384498954, + 0.005022265948355198, + 0.006564912851899862, + 0.005305523984134197, + 0.005992546211928129, + 0.003173193661496043, + 0.010510282590985298, + 0.005846601910889149, + 0.0024186691734939814, + 0.00503226974979043, + 0.0074108317494392395, + 0.006185520440340042, + 0.0058824606239795685, + 0.0035227001644670963, + 0.007859187200665474, + 0.004614117555320263, + 0.001848593121394515, + 0.004674085881561041, + 0.004983467049896717, + 0.004522200673818588, + 0.004592068027704954, + 0.0026614207308739424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.014202586375176907 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a frozen riverbank with broken ice fragments in nearly frozen black water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive Rough-legged Fishing Owl crouching in a poised, motionless posture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl have thick fluffy feathers with a mottled palette of dark brown and light brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the owl's eyes piercing yellow, acting as the sole source of light in the deep blue and black night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow, smooth, low-angle tracking shot that stops to reveal the owl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005952014587819576 + ], + "opposite_logprob_tokens": [ + -5.130951881408691 + ], + "logprob_reward": -0.005952014587819576, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005952014587819576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5434899926185608, + 0.08380415290594101, + 0.0700433999300003, + 0.05354967713356018, + 0.052172839641571045, + 0.03987798094749451, + 0.04837753251194954, + 0.03699791058897972, + 0.032920267432928085, + 0.0387662872672081 + ], + "credit_map": [ + 0.028392553329467773, + 0.10108290612697601, + 0.033877868205308914, + 0.11477311700582504, + 0.12125343084335327, + 0.0731133446097374, + 0.050665948539972305, + 0.020330797880887985, + 0.008664288558065891, + 0.007986688986420631, + 0.005847972352057695, + 0.010150037705898285, + 0.016093695536255836, + 0.018208879977464676, + 0.010138272307813168, + 0.006714313756674528, + 0.0073552364483475685, + 0.00803054217249155, + 0.00769384577870369, + 0.0102621391415596, + 0.012887946330010891, + 0.008313325233757496, + 0.009302309714257717, + 0.0061980560421943665, + 0.00746798375621438, + 0.0051056696102023125, + 0.003980261739343405, + 0.009420177899301052, + 0.00755737628787756, + 0.00856101419776678, + 0.007267755456268787, + 0.004189437720924616, + 0.004294524434953928, + 0.004567905329167843, + 0.0038579455576837063, + 0.008669174276292324, + 0.008623580448329449, + 0.009750178083777428, + 0.008022185415029526, + 0.004387347027659416, + 0.005855372175574303, + 0.004669781308621168, + 0.0025622451212257147, + 0.006323592271655798, + 0.00555272726342082, + 0.005762611050158739, + 0.006256537977606058, + 0.002895114943385124, + 0.007197936996817589, + 0.006769164931029081, + 0.003543917089700699, + 0.006418358534574509, + 0.007323079742491245, + 0.007498828694224358, + 0.006579701788723469, + 0.003046545898541808, + 0.008114158175885677, + 0.004249121528118849, + 0.002195446053519845, + 0.003690569195896387, + 0.005086997058242559, + 0.005394599866122007, + 0.0038437992334365845, + 0.004423217382282019, + 0.004933634772896767, + 0.004645852372050285, + 0.0018753970507532358, + 0.004356572404503822, + 0.004907307215034962, + 0.005363028030842543, + 0.00456973584368825, + 0.0022687409073114395, + 0.008374065160751343, + 0.005247573833912611, + 0.0022058577742427588, + 0.004982820712029934, + 0.0048246970400214195, + 0.005376987159252167, + 0.004943896550685167, + 0.002810389269143343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.005952014587819576 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a frozen riverbank with broken ice fragments in nearly frozen black water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive Rough-legged Fishing Owl crouching in a poised, motionless posture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl have thick fluffy feathers with a mottled palette of dark brown and light brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the owl's eyes piercing yellow, acting as the sole source of light in the deep blue and black night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow, smooth, low-angle tracking shot that stops to reveal the owl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809994086623192 + ], + "opposite_logprob_tokens": [ + -3.288100004196167 + ], + "logprob_reward": -0.03809994086623192, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809994086623192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5510704517364502, + 0.08325079828500748, + 0.0904090628027916, + 0.07819981127977371, + 0.07636068761348724, + 0.04720788449048996, + 0.02289123274385929, + 0.02121661975979805, + 0.013873347081243992, + 0.015520053915679455 + ], + "credit_map": [ + 0.03727448359131813, + 0.08728539198637009, + 0.05380270630121231, + 0.11986622959375381, + 0.10808888077735901, + 0.07711037248373032, + 0.04523593559861183, + 0.022406425327062607, + 0.007437259890139103, + 0.008976117707788944, + 0.007767803966999054, + 0.009091559797525406, + 0.01635914109647274, + 0.010308931581676006, + 0.0118859326466918, + 0.0114240488037467, + 0.008031909354031086, + 0.010017668828368187, + 0.006403177510946989, + 0.0099648991599679, + 0.017994282767176628, + 0.01565033197402954, + 0.016712825745344162, + 0.005633966065943241, + 0.00600461196154356, + 0.00807015411555767, + 0.00629596458747983, + 0.010960891842842102, + 0.01323750615119934, + 0.01299695111811161, + 0.012538532726466656, + 0.008095193654298782, + 0.005935104563832283, + 0.007627703249454498, + 0.010477431118488312, + 0.009082824923098087, + 0.012199779972434044, + 0.012000536546111107, + 0.013200660236179829, + 0.005836648400872946, + 0.008035536855459213, + 0.006398375146090984, + 0.004976235330104828, + 0.005426689982414246, + 0.006494984030723572, + 0.0060034277848899364, + 0.006816002074629068, + 0.0030566328205168247, + 0.002928937552496791, + 0.00347452936694026, + 0.0016056078020483255, + 0.0029230692889541388, + 0.0032445667311549187, + 0.0038368285167962313, + 0.0028819956351071596, + 0.0019956976175308228, + 0.002054387703537941, + 0.0033738825004547834, + 0.0013488206313923001, + 0.0024737699422985315, + 0.003774778451770544, + 0.003108387580141425, + 0.0032192100770771503, + 0.001863381708972156, + 0.0015027213376015425, + 0.0013434351421892643, + 0.0011735456064343452, + 0.0019260613480582833, + 0.0021965729538351297, + 0.0021570874378085136, + 0.0022685539443045855, + 0.0013053689617663622, + 0.0019024465000256896, + 0.0019718604162335396, + 0.0010170828318223357, + 0.0018850249471142888, + 0.0024476381950080395, + 0.002500728238373995, + 0.002315846038982272, + 0.0014794274466112256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.03809994086623192 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a frozen riverbank with broken ice fragments in nearly frozen black water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive Rough-legged Fishing Owl crouching in a poised, motionless posture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl have thick fluffy feathers with a mottled palette of dark brown and light brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the owl's eyes piercing yellow, acting as the sole source of light in the deep blue and black night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow, smooth, low-angle tracking shot that stops to reveal the owl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808479383587837 + ], + "opposite_logprob_tokens": [ + -3.2880847454071045 + ], + "logprob_reward": -0.03808479383587837, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808479383587837, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2339904010295868, + 0.08475388586521149, + 0.1884055882692337, + 0.1133822351694107, + 0.0847090557217598, + 0.1442069411277771, + 0.04645325616002083, + 0.02091466635465622, + 0.060075871646404266, + 0.02310808375477791 + ], + "credit_map": [ + 0.016066361218690872, + 0.03502988815307617, + 0.03697527199983597, + 0.054579444229602814, + 0.035306066274642944, + 0.023317553102970123, + 0.01985299400985241, + 0.01286283414810896, + 0.006220471113920212, + 0.011426812969148159, + 0.008658302016556263, + 0.009571135975420475, + 0.017195148393511772, + 0.01583843305706978, + 0.00969394389539957, + 0.006149632856249809, + 0.01666029915213585, + 0.019127435982227325, + 0.008850280195474625, + 0.015577136538922787, + 0.028208201751112938, + 0.04760707914829254, + 0.04030263423919678, + 0.01207253523170948, + 0.010661734268069267, + 0.014981799758970737, + 0.007483960594981909, + 0.014490606263279915, + 0.020230431109666824, + 0.018610404804348946, + 0.017930060625076294, + 0.008993243798613548, + 0.010934904217720032, + 0.008864183910191059, + 0.0076355901546776295, + 0.01012243889272213, + 0.011859070509672165, + 0.012696080841124058, + 0.011175015941262245, + 0.011421765200793743, + 0.016708772629499435, + 0.01706121675670147, + 0.0177169032394886, + 0.016508538275957108, + 0.021806227043271065, + 0.022021763026714325, + 0.022488709539175034, + 0.009894811548292637, + 0.006312916986644268, + 0.006648350041359663, + 0.003446754300966859, + 0.0053632063791155815, + 0.006769207306206226, + 0.006918453611433506, + 0.007612019777297974, + 0.00338234705850482, + 0.002610183320939541, + 0.003852501977235079, + 0.0022186748683452606, + 0.0025601608213037252, + 0.0025292846839874983, + 0.0029526553116738796, + 0.0026399402413517237, + 0.0015512657118961215, + 0.007995941676199436, + 0.006596169900149107, + 0.00467959837988019, + 0.005855697672814131, + 0.01166938804090023, + 0.008720386773347855, + 0.008141552098095417, + 0.006417140830308199, + 0.0031266312580555677, + 0.0026879343204200268, + 0.0016899823676794767, + 0.002817136235535145, + 0.0037582169752568007, + 0.0036236895248293877, + 0.002982007572427392, + 0.0024224845692515373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.03808479383587837 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, smooth, low-angle tracking shot moves along a frozen riverbank, frost-covered and gleaming coldly in the dim moonlight. Broken ice fragments appear and disappear in the nearly frozen black water. As the camera slowly stops, it reveals its subject, hidden in the deep shadow cast by a withered tree on the bank: a massive Rough-legged Fishing Owl. Its thick, fluffy feathers, a mottled palette of dark brown and light brown, almost blend into the darkness. Its body leans slightly forward, crouching in a poised posture, motionless. In the near-total darkness, its unsettlingly piercing yellow eyes, like two burning amber lamps, pierce the deep blue and black night, becoming the sole source of light and focal point in the entire frame, its gaze fixed firmly on a point beyond the camera. The texture of the entire scene is like a Rembrandt-style thick oil painting, with extremely strong contrast between light and shadow. Deep Prussian blue and pure black constitute the main colors of the picture, creating a breathtaking and silent atmosphere full of dramatic tension.", + "seed": 24744, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_0/index_656_seed_24744_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.018897120282053946, + "vlm_logprob": -0.018897120282053946, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.018897120282053946, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a frozen riverbank with broken ice fragments in nearly frozen black water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive Rough-legged Fishing Owl crouching in a poised, motionless posture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl have thick fluffy feathers with a mottled palette of dark brown and light brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the owl's eyes piercing yellow, acting as the sole source of light in the deep blue and black night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow, smooth, low-angle tracking shot that stops to reveal the owl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021810319274663925 + ], + "opposite_logprob_tokens": [ + -4.021810531616211 + ], + "logprob_reward": -0.021810319274663925, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.021810319274663925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07188234478235245, + 0.14955249428749084, + 0.023378880694508553, + 0.12246676534414291, + 0.31463322043418884, + 0.03673221170902252, + 0.09352261573076248, + 0.06529559195041656, + 0.07793447375297546, + 0.04460136964917183 + ], + "credit_map": [ + 0.010279921814799309, + 0.007262909784913063, + 0.011100544594228268, + 0.007049704436212778, + 0.007910484448075294, + 0.008279111236333847, + 0.009476130828261375, + 0.010523535311222076, + 0.05215699225664139, + 0.054599180817604065, + 0.006129586603492498, + 0.00474261911585927, + 0.0113221425563097, + 0.006458207033574581, + 0.006389595568180084, + 0.007754181977361441, + 0.001755431992933154, + 0.0032629589550197124, + 0.0025576327461749315, + 0.002663402119651437, + 0.003241358557716012, + 0.0024963037576526403, + 0.004300134256482124, + 0.0031016585417091846, + 0.004930790979415178, + 0.0363059863448143, + 0.01017558854073286, + 0.04894975572824478, + 0.005169028416275978, + 0.005057784728705883, + 0.007076377049088478, + 0.0048014502972364426, + 0.025775868445634842, + 0.03702038154006004, + 0.13893036544322968, + 0.054149068892002106, + 0.00922780204564333, + 0.0278474111109972, + 0.009771911427378654, + 0.01191042922437191, + 0.006034539081156254, + 0.0038594496436417103, + 0.005052079446613789, + 0.0027370890602469444, + 0.002371669514104724, + 0.009448881261050701, + 0.003448282368481159, + 0.003780222497880459, + 0.014635637402534485, + 0.016134491190314293, + 0.00936836190521717, + 0.006737637333571911, + 0.006112930830568075, + 0.02404385432600975, + 0.008298427797853947, + 0.008191274479031563, + 0.006007048767060041, + 0.01150679774582386, + 0.010395095683634281, + 0.006402140948921442, + 0.006580794230103493, + 0.008421330712735653, + 0.00968237780034542, + 0.006300002336502075, + 0.011212213896214962, + 0.008779480122029781, + 0.013823112472891808, + 0.0062677301466465, + 0.004143235739320517, + 0.01823001727461815, + 0.005891157314181328, + 0.009587525390088558, + 0.007638402283191681, + 0.004985876381397247, + 0.007133249193429947, + 0.003119455184787512, + 0.0022045813966542482, + 0.012026011012494564, + 0.002933215582743287, + 0.004560576286166906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.021810319274663925 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a frozen riverbank with broken ice fragments in nearly frozen black water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive Rough-legged Fishing Owl crouching in a poised, motionless posture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl have thick fluffy feathers with a mottled palette of dark brown and light brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the owl's eyes piercing yellow, acting as the sole source of light in the deep blue and black night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow, smooth, low-angle tracking shot that stops to reveal the owl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016071535646915436 + ], + "opposite_logprob_tokens": [ + -4.141071319580078 + ], + "logprob_reward": -0.016071535646915436, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016071535646915436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.031282730400562286, + 0.05224134027957916, + 0.024841854348778725, + 0.033147912472486496, + 0.7127270698547363, + 0.04962495341897011, + 0.01713496260344982, + 0.0374705009162426, + 0.027292491868138313, + 0.014236219227313995 + ], + "credit_map": [ + 0.004135523922741413, + 0.0037317893002182245, + 0.004821464419364929, + 0.003305569291114807, + 0.0033188220113515854, + 0.0037064540665596724, + 0.004414011258631945, + 0.0038490963634103537, + 0.017677534371614456, + 0.01876537874341011, + 0.0023319462779909372, + 0.002184711629524827, + 0.00391915999352932, + 0.0026305452920496464, + 0.002376992255449295, + 0.002355070086196065, + 0.0015732941683381796, + 0.0022380573209375143, + 0.0028435320127755404, + 0.0020891884341835976, + 0.005390728358179331, + 0.0031533273868262768, + 0.0037947744131088257, + 0.003758952021598816, + 0.003720645559951663, + 0.007441399153321981, + 0.0033954482059925795, + 0.005198194645345211, + 0.004804105963557959, + 0.003753956640139222, + 0.0025495539885014296, + 0.002284607384353876, + 0.059273529797792435, + 0.08520251512527466, + 0.32011809945106506, + 0.12447689473628998, + 0.01764701120555401, + 0.05978558957576752, + 0.020550858229398727, + 0.025672554969787598, + 0.006874314043670893, + 0.006741690449416637, + 0.00603780010715127, + 0.005032130517065525, + 0.003680726746097207, + 0.009721110574901104, + 0.006534590385854244, + 0.0050025926902890205, + 0.001876586931757629, + 0.0022014924325048923, + 0.0021622825879603624, + 0.0015914709074422717, + 0.0020832500886172056, + 0.0037533340509980917, + 0.0018619258189573884, + 0.0016046210657805204, + 0.004269002936780453, + 0.00582703435793519, + 0.0042966739274561405, + 0.0027983051259070635, + 0.003519329708069563, + 0.007236376870423555, + 0.0058372425846755505, + 0.003686535404995084, + 0.0022073988802731037, + 0.0036147180944681168, + 0.006403257604688406, + 0.003025928046554327, + 0.002181036164984107, + 0.003691187361255288, + 0.002713964320719242, + 0.0034549999982118607, + 0.0010624384740367532, + 0.0016593129839748144, + 0.002581487875431776, + 0.001601537223905325, + 0.0012400769628584385, + 0.0019338969141244888, + 0.002687894506379962, + 0.001469574635848403 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 5 + ], + "selected_reward": -0.016071535646915436 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a frozen riverbank with broken ice fragments in nearly frozen black water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive Rough-legged Fishing Owl crouching in a poised, motionless posture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl have thick fluffy feathers with a mottled palette of dark brown and light brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the owl's eyes piercing yellow, acting as the sole source of light in the deep blue and black night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow, smooth, low-angle tracking shot that stops to reveal the owl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014196710661053658 + ], + "opposite_logprob_tokens": [ + -4.264196872711182 + ], + "logprob_reward": -0.014196710661053658, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014196710661053658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046124186366796494, + 0.05734214186668396, + 0.025601567700505257, + 0.15089920163154602, + 0.33295828104019165, + 0.10968395322561264, + 0.06933516263961792, + 0.08483248949050903, + 0.08163468539714813, + 0.041588328778743744 + ], + "credit_map": [ + 0.0062913293950259686, + 0.006851749960333109, + 0.005502033047378063, + 0.005980789661407471, + 0.004904153291136026, + 0.0035932809114456177, + 0.007467217743396759, + 0.005533632356673479, + 0.02142782136797905, + 0.02282503992319107, + 0.002225850010290742, + 0.0019417802104726434, + 0.004051486030220985, + 0.002056017518043518, + 0.0011830686125904322, + 0.0016310797072947025, + 0.002026796108111739, + 0.0031762022990733385, + 0.0046355533413589, + 0.0027912622317671776, + 0.003467756789177656, + 0.004194395150989294, + 0.003274037502706051, + 0.0020355647429823875, + 0.0065791914239525795, + 0.044041454792022705, + 0.010555695742368698, + 0.055635299533605576, + 0.010451641865074635, + 0.007368731778115034, + 0.009895656257867813, + 0.006371534429490566, + 0.02696697972714901, + 0.03903336822986603, + 0.14811936020851135, + 0.05727056786417961, + 0.010588129051029682, + 0.02724517695605755, + 0.010973654687404633, + 0.012761038728058338, + 0.015496083535254002, + 0.013431806117296219, + 0.018749084323644638, + 0.011224337853491306, + 0.006030084565281868, + 0.026558799669146538, + 0.008394841104745865, + 0.009798910468816757, + 0.0110728545114398, + 0.008903220295906067, + 0.0066470373421907425, + 0.007376853842288256, + 0.00499030202627182, + 0.01636560633778572, + 0.006472348235547543, + 0.007506945636123419, + 0.010026982985436916, + 0.013706842437386513, + 0.008900064043700695, + 0.006907441187649965, + 0.00790153257548809, + 0.016806812956929207, + 0.01239203754812479, + 0.008190777152776718, + 0.010080743581056595, + 0.008009122684597969, + 0.018471919000148773, + 0.006115629803389311, + 0.005522129591554403, + 0.01715957187116146, + 0.006156417075544596, + 0.010119151324033737, + 0.0051325284875929356, + 0.004783530719578266, + 0.00898919627070427, + 0.0027706283144652843, + 0.0029732033144682646, + 0.00878700241446495, + 0.00407577445730567, + 0.004076466429978609 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.014196710661053658 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a frozen riverbank with broken ice fragments in nearly frozen black water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive Rough-legged Fishing Owl crouching in a poised, motionless posture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl have thick fluffy feathers with a mottled palette of dark brown and light brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the owl's eyes piercing yellow, acting as the sole source of light in the deep blue and black night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow, smooth, low-angle tracking shot that stops to reveal the owl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026348169893026352 + ], + "opposite_logprob_tokens": [ + -3.651348114013672 + ], + "logprob_reward": -0.026348169893026352, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026348169893026352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045924216508865356, + 0.08134663850069046, + 0.05362796410918236, + 0.1739620715379715, + 0.08698099851608276, + 0.22767989337444305, + 0.050191521644592285, + 0.04024853929877281, + 0.10862220078706741, + 0.1314159631729126 + ], + "credit_map": [ + 0.0071395644918084145, + 0.004357367753982544, + 0.007220914121717215, + 0.004309538751840591, + 0.004593235440552235, + 0.0043104263022542, + 0.007841416634619236, + 0.0061517516151070595, + 0.029438724741339684, + 0.031671468168497086, + 0.002603928092867136, + 0.0023189042694866657, + 0.005839167628437281, + 0.003655664622783661, + 0.002560579450801015, + 0.003258204786106944, + 0.003906110767275095, + 0.006667270325124264, + 0.008559092879295349, + 0.00585665600374341, + 0.007188313640654087, + 0.0061447215266525745, + 0.009109430946409702, + 0.006196363363415003, + 0.0074739302508533, + 0.03381390497088432, + 0.014985628426074982, + 0.05538473278284073, + 0.022743195295333862, + 0.010330529883503914, + 0.016844626516103745, + 0.012385528534650803, + 0.006580382119864225, + 0.009630491025745869, + 0.032428231090307236, + 0.012940955348312855, + 0.005419102497398853, + 0.008163625374436378, + 0.006399880163371563, + 0.005418334621936083, + 0.041161295026540756, + 0.03386123105883598, + 0.04777953028678894, + 0.021706456318497658, + 0.009085442870855331, + 0.04486532881855965, + 0.009705676697194576, + 0.01951494626700878, + 0.007972377352416515, + 0.006209813989698887, + 0.004756272304803133, + 0.003522421233355999, + 0.0037549284752458334, + 0.01364219468086958, + 0.00523495813831687, + 0.005098552908748388, + 0.003279595635831356, + 0.007951457053422928, + 0.0062880804762244225, + 0.004584208130836487, + 0.0035410576965659857, + 0.003955758176743984, + 0.006894679740071297, + 0.0037537061143666506, + 0.021009299904108047, + 0.014633659273386002, + 0.01223586592823267, + 0.008435815572738647, + 0.006480770651251078, + 0.0288986973464489, + 0.006434289272874594, + 0.010493798181414604, + 0.024852413684129715, + 0.015053221955895424, + 0.017063958570361137, + 0.00892509613186121, + 0.006200444418936968, + 0.037694014608860016, + 0.008838264271616936, + 0.012788549065589905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 9 + ], + "selected_reward": -0.026348169893026352 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a frozen riverbank with broken ice fragments in nearly frozen black water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive Rough-legged Fishing Owl crouching in a poised, motionless posture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl have thick fluffy feathers with a mottled palette of dark brown and light brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the owl's eyes piercing yellow, acting as the sole source of light in the deep blue and black night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow, smooth, low-angle tracking shot that stops to reveal the owl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016058865934610367 + ], + "opposite_logprob_tokens": [ + -4.141058921813965 + ], + "logprob_reward": -0.016058865934610367, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016058865934610367, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.020296217873692513, + 0.04244057834148407, + 0.024000031873583794, + 0.181594118475914, + 0.5786731243133545, + 0.03025156632065773, + 0.0176469087600708, + 0.02644079551100731, + 0.06441055238246918, + 0.014246135950088501 + ], + "credit_map": [ + 0.0027613395359367132, + 0.002014443278312683, + 0.0026465514674782753, + 0.001587175764143467, + 0.0025768806226551533, + 0.0028340360149741173, + 0.001905170502141118, + 0.00397062161937356, + 0.013746753334999084, + 0.013580729253590107, + 0.001897216192446649, + 0.00221778335981071, + 0.0036459199618548155, + 0.002521042013540864, + 0.0025135886389762163, + 0.0023175491951406, + 0.001664743642322719, + 0.002678937977179885, + 0.002812352031469345, + 0.002526976400986314, + 0.0038157179951667786, + 0.0027421782724559307, + 0.003701168345287442, + 0.004057957790791988, + 0.008852725848555565, + 0.05717700719833374, + 0.013284225016832352, + 0.0663284882903099, + 0.009587756358087063, + 0.009879572317004204, + 0.009533992037177086, + 0.006950362119823694, + 0.04810044914484024, + 0.06980468332767487, + 0.2588023543357849, + 0.10079815983772278, + 0.014941620640456676, + 0.0482291579246521, + 0.01658923365175724, + 0.0214074719697237, + 0.002178856171667576, + 0.0027395880315452814, + 0.011262375861406326, + 0.0033915613312274218, + 0.001699867774732411, + 0.005267017986625433, + 0.0019235125510022044, + 0.001788785681128502, + 0.002212694613263011, + 0.0030281839426606894, + 0.0020261851605027914, + 0.002215947723016143, + 0.0012185866944491863, + 0.003431157674640417, + 0.001813902286812663, + 0.001700250431895256, + 0.001961767440661788, + 0.00475298473611474, + 0.005599341820925474, + 0.002982419217005372, + 0.0027461061254143715, + 0.0018914854153990746, + 0.004141154233366251, + 0.0023655355907976627, + 0.01173858717083931, + 0.007122897543013096, + 0.00954213086515665, + 0.004947330337017775, + 0.0028448239900171757, + 0.01702488213777542, + 0.004379409831017256, + 0.006810490507632494, + 0.002337798709049821, + 0.0014266489306464791, + 0.0027581353206187487, + 0.0011391087900847197, + 0.000766973418649286, + 0.003573940135538578, + 0.0009846637258306146, + 0.0012588661629706621 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 8 + ], + "selected_reward": -0.016058865934610367 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_1/index_tensor([5954, 5954])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_1/index_tensor([5954, 5954])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8c2c05f356ca2e9aea066781bafb98f496b3fc33 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_1/index_tensor([5954, 5954])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, shot from a level angle, slowly and steadily moves forward, focusing on an elderly monk. His shiny head gleams faintly in the morning light, his forehead and eyes crisscrossed with deep, furrowed lines. Clad in a faded, ochre-yellow, coarse robe, he sits cross-legged on a dark cushion in the quiet courtyard of an ancient monastery. Behind him, the temple's pristine wooden eaves, their outlines gradually becoming clear in the morning light, and beneath his feet, a lightly frosted bluestone floor. His back is straight, his hands clasped in his lap in meditation mudra, his eyes closed, his face radiating utter serenity and concentration. With each slow, deep, and rhythmic breath, a small, dense cloud of white mist gently emanates from his nostrils, briefly condensing and swirling in the cold air before curling upwards and silently dissipating into the chill dawn light. The dominant tones of the entire painting are calm blue-gray and deep wood tones. Only the first glimmer of light at dawn gives the monk's outline a faint golden edge. The overall atmosphere is solemn and tranquil, filled with Zen and a deep sense of tranquility that transcends time.", + "seed": 23954, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_1/index_5954_seed_23954_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08385998643934726, + "vlm_logprob": -0.08385998643934726, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.08385998643934726, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly monk wearing a faded, ochre-yellow, coarse robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk sitting cross-legged on a dark cushion with his back straight and hands in meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a quiet courtyard of an ancient monastery with pristine wooden eaves and a lightly frosted bluestone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk emit a small, dense cloud of white mist from his nostrils that swirls and dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot from a level angle that slowly and steadily moves forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016178162768483162 + ], + "opposite_logprob_tokens": [ + -4.391178131103516 + ], + "logprob_reward": -0.016178162768483162, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.016178162768483162, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30885085463523865, + 0.1075175404548645, + 0.03417498990893364, + 0.03345414251089096, + 0.057385120540857315, + 0.19640915095806122, + 0.07257149368524551, + 0.05297243222594261, + 0.06439074873924255, + 0.0722736194729805 + ], + "credit_map": [ + 0.037066154181957245, + 0.03127267584204674, + 0.045170124620199203, + 0.022476786747574806, + 0.05175362899899483, + 0.039096519351005554, + 0.04749701917171478, + 0.03451790660619736, + 0.013111842796206474, + 0.01072213239967823, + 0.014270706102252007, + 0.006588084623217583, + 0.01815057173371315, + 0.017238065600395203, + 0.015842149034142494, + 0.011593978852033615, + 0.0035864096134901047, + 0.0027313234750181437, + 0.004289750475436449, + 0.0028401988092809916, + 0.006969331298023462, + 0.00499247619882226, + 0.005232140421867371, + 0.0035333556588739157, + 0.002857389161363244, + 0.0030281306244432926, + 0.0037424247711896896, + 0.002807125449180603, + 0.007401526905596256, + 0.003797652432695031, + 0.006951455492526293, + 0.00286843697540462, + 0.00398833304643631, + 0.004138440825045109, + 0.01193256862461567, + 0.005023186095058918, + 0.007732694502919912, + 0.005122947506606579, + 0.009217813611030579, + 0.01022912934422493, + 0.015745509415864944, + 0.010058984160423279, + 0.059249330312013626, + 0.012447640299797058, + 0.01691642589867115, + 0.014026395045220852, + 0.016061771661043167, + 0.051903072744607925, + 0.004064988344907761, + 0.0025910523254424334, + 0.011210381984710693, + 0.006865359377115965, + 0.010049758478999138, + 0.0059660556726157665, + 0.008785435929894447, + 0.023038456216454506, + 0.00828542560338974, + 0.005491207353770733, + 0.005720656830817461, + 0.0038827331736683846, + 0.008692672476172447, + 0.007753031328320503, + 0.005435909144580364, + 0.007710792124271393, + 0.007162203546613455, + 0.005090474151074886, + 0.005579669028520584, + 0.006698907352983952, + 0.011156279593706131, + 0.008816185407340527, + 0.007269372697919607, + 0.012617645785212517, + 0.011580429039895535, + 0.006631880532950163, + 0.006176788359880447, + 0.008513578213751316, + 0.014592203311622143, + 0.008100551553070545, + 0.010429954156279564, + 0.00624822685495019 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.016178162768483162 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly monk wearing a faded, ochre-yellow, coarse robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk sitting cross-legged on a dark cushion with his back straight and hands in meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a quiet courtyard of an ancient monastery with pristine wooden eaves and a lightly frosted bluestone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk emit a small, dense cloud of white mist from his nostrils that swirls and dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot from a level angle that slowly and steadily moves forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048684071749448776 + ], + "opposite_logprob_tokens": [ + -3.0486841201782227 + ], + "logprob_reward": -0.048684071749448776, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048684071749448776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41458287835121155, + 0.08540798723697662, + 0.026258844882249832, + 0.026209790259599686, + 0.04678452014923096, + 0.1412515938282013, + 0.09547722339630127, + 0.08633460104465485, + 0.03206063434481621, + 0.04563196748495102 + ], + "credit_map": [ + 0.04466231167316437, + 0.038878146559000015, + 0.0526898019015789, + 0.03811054676771164, + 0.09213171899318695, + 0.03937356546521187, + 0.07352284342050552, + 0.035213958472013474, + 0.00782283116132021, + 0.007089270278811455, + 0.011110367253422737, + 0.007503557950258255, + 0.016325397416949272, + 0.01169308740645647, + 0.014256106689572334, + 0.009607361629605293, + 0.0019261291017755866, + 0.0019164071418344975, + 0.0031209641601890326, + 0.0029756631702184677, + 0.006008078344166279, + 0.0027838400565087795, + 0.0042465305887162685, + 0.003281232202425599, + 0.0023812996223568916, + 0.0024150959216058254, + 0.0026824395172297955, + 0.0031139126513153315, + 0.006052167154848576, + 0.0026661655865609646, + 0.0043227095156908035, + 0.0025760005228221416, + 0.0035684253089129925, + 0.003140827175229788, + 0.006466270424425602, + 0.004300602711737156, + 0.009539135731756687, + 0.0055351960472762585, + 0.007957211695611477, + 0.006276850588619709, + 0.01298338733613491, + 0.006924496032297611, + 0.04643775150179863, + 0.008496193215250969, + 0.00605539046227932, + 0.009907860308885574, + 0.010227877646684647, + 0.04021862894296646, + 0.005262976046651602, + 0.0034700208343565464, + 0.013485071249306202, + 0.011380988173186779, + 0.012498470954596996, + 0.008658573031425476, + 0.012543892487883568, + 0.028177225962281227, + 0.010171808302402496, + 0.010505231097340584, + 0.007916748523712158, + 0.008113507181406021, + 0.016757136210799217, + 0.012400970794260502, + 0.009666253812611103, + 0.010802938602864742, + 0.003987123724073172, + 0.0030291967559605837, + 0.003516250057145953, + 0.002679517725482583, + 0.006360935512930155, + 0.004067610949277878, + 0.0030832744669169188, + 0.005336724687367678, + 0.009710765443742275, + 0.005787233356386423, + 0.004812783561646938, + 0.0027716809418052435, + 0.008229363709688187, + 0.005257823038846254, + 0.00482245022431016, + 0.0042398665100336075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.048684071749448776 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly monk wearing a faded, ochre-yellow, coarse robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk sitting cross-legged on a dark cushion with his back straight and hands in meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a quiet courtyard of an ancient monastery with pristine wooden eaves and a lightly frosted bluestone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk emit a small, dense cloud of white mist from his nostrils that swirls and dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot from a level angle that slowly and steadily moves forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16036953032016754 + ], + "opposite_logprob_tokens": [ + -1.9103695154190063 + ], + "logprob_reward": -0.16036953032016754, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16036953032016754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3067292869091034, + 0.09369740635156631, + 0.03116113878786564, + 0.03741997107863426, + 0.10321860015392303, + 0.08793359249830246, + 0.06672972440719604, + 0.10473021119832993, + 0.06329341977834702, + 0.10508666187524796 + ], + "credit_map": [ + 0.035219937562942505, + 0.04265043884515762, + 0.04490683972835541, + 0.024050507694482803, + 0.03186677396297455, + 0.05107954517006874, + 0.03690829873085022, + 0.04004695639014244, + 0.011179848574101925, + 0.013626042753458023, + 0.009793381206691265, + 0.006750207394361496, + 0.012327928096055984, + 0.017141766846179962, + 0.011841337196528912, + 0.011036896146833897, + 0.0038486430421471596, + 0.004678203724324703, + 0.00356360268779099, + 0.003074652748182416, + 0.004120065830647945, + 0.004010417032986879, + 0.0047804792411625385, + 0.0030850721523165703, + 0.004572080913931131, + 0.004590611904859543, + 0.005163827911019325, + 0.0033750850707292557, + 0.0060563962906599045, + 0.003735383041203022, + 0.006406840402632952, + 0.0035197422839701176, + 0.012231549248099327, + 0.01398328598588705, + 0.01873384602367878, + 0.008243906311690807, + 0.010202601552009583, + 0.011150560341775417, + 0.008961166255176067, + 0.019711686298251152, + 0.012047729454934597, + 0.006926706992089748, + 0.010683471336960793, + 0.00893932394683361, + 0.009094606153666973, + 0.013585172593593597, + 0.010433212853968143, + 0.016223371028900146, + 0.003759062848985195, + 0.0034189382568001747, + 0.008689237758517265, + 0.005492203403264284, + 0.011756028048694134, + 0.0064288838766515255, + 0.008579554967582226, + 0.01860581338405609, + 0.01700463704764843, + 0.013794448226690292, + 0.011738895438611507, + 0.009840368293225765, + 0.014115173369646072, + 0.013134755194187164, + 0.011541912332177162, + 0.013560025952756405, + 0.00851256400346756, + 0.0069327326491475105, + 0.0057432930916547775, + 0.005403426010161638, + 0.00945628434419632, + 0.010458866134285927, + 0.005616454873234034, + 0.011169792152941227, + 0.01956668496131897, + 0.008192867040634155, + 0.01722608506679535, + 0.007433115504682064, + 0.010706296190619469, + 0.01833176612854004, + 0.008346742019057274, + 0.01528310589492321 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.16036953032016754 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly monk wearing a faded, ochre-yellow, coarse robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk sitting cross-legged on a dark cushion with his back straight and hands in meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a quiet courtyard of an ancient monastery with pristine wooden eaves and a lightly frosted bluestone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk emit a small, dense cloud of white mist from his nostrils that swirls and dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot from a level angle that slowly and steadily moves forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603783667087555 + ], + "opposite_logprob_tokens": [ + -1.910378336906433 + ], + "logprob_reward": -0.1603783667087555, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603783667087555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23529507219791412, + 0.0716749057173729, + 0.03194510191679001, + 0.05506907403469086, + 0.09372976422309875, + 0.30181941390037537, + 0.08411464095115662, + 0.06914078444242477, + 0.029914522543549538, + 0.027296708896756172 + ], + "credit_map": [ + 0.022501928731799126, + 0.020374802872538567, + 0.0438384935259819, + 0.022596843540668488, + 0.03237781301140785, + 0.026109283789992332, + 0.036856599152088165, + 0.030639341101050377, + 0.007617919240146875, + 0.00677600409835577, + 0.007975170388817787, + 0.007492379751056433, + 0.012484977953135967, + 0.009938460774719715, + 0.009760533459484577, + 0.009629461914300919, + 0.0031680737156420946, + 0.003532860428094864, + 0.003827338106930256, + 0.004706128966063261, + 0.004347707144916058, + 0.0043824645690619946, + 0.004882202949374914, + 0.003098327200859785, + 0.004815972410142422, + 0.005158269312232733, + 0.007330150343477726, + 0.00599157577380538, + 0.011519013904035091, + 0.006371648050844669, + 0.0085776811465621, + 0.005304770078510046, + 0.009121046401560307, + 0.011098752729594707, + 0.008638466708362103, + 0.010253874585032463, + 0.01517285592854023, + 0.009421680122613907, + 0.01911461167037487, + 0.010908480733633041, + 0.023604949936270714, + 0.015357817523181438, + 0.09736186265945435, + 0.018557051196694374, + 0.013751555234193802, + 0.021734653040766716, + 0.022386664524674416, + 0.08906486630439758, + 0.004537314176559448, + 0.004396942909806967, + 0.012815282680094242, + 0.004664963111281395, + 0.010867403820157051, + 0.00867484975606203, + 0.00956776924431324, + 0.028590118512511253, + 0.00429950375109911, + 0.004634598735719919, + 0.0055590057745575905, + 0.007068636827170849, + 0.01585381291806698, + 0.0080431392416358, + 0.012286531738936901, + 0.01139556523412466, + 0.0024969857186079025, + 0.003081766190007329, + 0.002459887880831957, + 0.00364408059976995, + 0.005291464272886515, + 0.004862673580646515, + 0.00321830902248621, + 0.004859356675297022, + 0.004413461312651634, + 0.0023861180525273085, + 0.0032875805627554655, + 0.0026264474727213383, + 0.004040542058646679, + 0.004359046928584576, + 0.0032234778627753258, + 0.0029600393027067184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.1603783667087555 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly monk wearing a faded, ochre-yellow, coarse robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk sitting cross-legged on a dark cushion with his back straight and hands in meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a quiet courtyard of an ancient monastery with pristine wooden eaves and a lightly frosted bluestone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk emit a small, dense cloud of white mist from his nostrils that swirls and dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot from a level angle that slowly and steadily moves forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368980064988136 + ], + "opposite_logprob_tokens": [ + -3.4086897373199463 + ], + "logprob_reward": -0.03368980064988136, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368980064988136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3700096905231476, + 0.09070317447185516, + 0.04449507221579552, + 0.04728155955672264, + 0.05508887767791748, + 0.15971730649471283, + 0.07627417892217636, + 0.04425054043531418, + 0.03821517527103424, + 0.07396438717842102 + ], + "credit_map": [ + 0.04153024032711983, + 0.03311513736844063, + 0.051882676780223846, + 0.0364941842854023, + 0.053695254027843475, + 0.05498073250055313, + 0.05504198744893074, + 0.04326951876282692, + 0.009071883745491505, + 0.009135919623076916, + 0.007951767183840275, + 0.00811499822884798, + 0.016717366874217987, + 0.01443874929100275, + 0.013871658593416214, + 0.011400839313864708, + 0.003896886482834816, + 0.00451698387041688, + 0.0041113137267529964, + 0.004790500272065401, + 0.011002296581864357, + 0.0051984479650855064, + 0.006124552339315414, + 0.004854095634073019, + 0.003455959027633071, + 0.004508711863309145, + 0.0041747987270355225, + 0.005614861845970154, + 0.0120980991050601, + 0.005030592437833548, + 0.007748667616397142, + 0.004649870563298464, + 0.00409576715901494, + 0.004682171158492565, + 0.005942284129559994, + 0.0049594976007938385, + 0.012044718489050865, + 0.005723761394619942, + 0.009948542341589928, + 0.007692134007811546, + 0.014766433276236057, + 0.0076979887671768665, + 0.050411712378263474, + 0.009299371391534805, + 0.00889524631202221, + 0.011677213944494724, + 0.012051444500684738, + 0.04491790011525154, + 0.007876652292907238, + 0.005479156505316496, + 0.00864351261407137, + 0.007028290070593357, + 0.012328995391726494, + 0.013789554126560688, + 0.008109960705041885, + 0.013018065132200718, + 0.007906845770776272, + 0.0048413584008812904, + 0.0037052836269140244, + 0.004428551532328129, + 0.007873579859733582, + 0.003864406840875745, + 0.0059571112506091595, + 0.005673408042639494, + 0.002719785552471876, + 0.0035343938507139683, + 0.0035343405324965715, + 0.0031550659332424402, + 0.006584944203495979, + 0.0061393314972519875, + 0.002902451902627945, + 0.009644865058362484, + 0.0105538759380579, + 0.006329445634037256, + 0.008711462840437889, + 0.005568657536059618, + 0.00968991406261921, + 0.012520097196102142, + 0.006990229710936546, + 0.013600708916783333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.03368980064988136 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a level angle, slowly and steadily moves forward, focusing on an elderly monk. His shiny head gleams faintly in the morning light, his forehead and eyes crisscrossed with deep, furrowed lines. Clad in a faded, ochre-yellow, coarse robe, he sits cross-legged on a dark cushion in the quiet courtyard of an ancient monastery. Behind him, the temple's pristine wooden eaves, their outlines gradually becoming clear in the morning light, and beneath his feet, a lightly frosted bluestone floor. His back is straight, his hands clasped in his lap in meditation mudra, his eyes closed, his face radiating utter serenity and concentration. With each slow, deep, and rhythmic breath, a small, dense cloud of white mist gently emanates from his nostrils, briefly condensing and swirling in the cold air before curling upwards and silently dissipating into the chill dawn light. The dominant tones of the entire painting are calm blue-gray and deep wood tones. Only the first glimmer of light at dawn gives the monk's outline a faint golden edge. The overall atmosphere is solemn and tranquil, filled with Zen and a deep sense of tranquility that transcends time.", + "seed": 23955, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_1/index_5954_seed_23955_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04666650053113699, + "vlm_logprob": -0.04666650053113699, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.04666650053113699, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly monk wearing a faded, ochre-yellow, coarse robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk sitting cross-legged on a dark cushion with his back straight and hands in meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a quiet courtyard of an ancient monastery with pristine wooden eaves and a lightly frosted bluestone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk emit a small, dense cloud of white mist from his nostrils that swirls and dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot from a level angle that slowly and steadily moves forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010473879054188728 + ], + "opposite_logprob_tokens": [ + -4.885473728179932 + ], + "logprob_reward": -0.010473879054188728, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.010473879054188728, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04438352957367897, + 0.050617530941963196, + 0.027438567951321602, + 0.01915578544139862, + 0.07593797892332077, + 0.07324854284524918, + 0.045889534056186676, + 0.05314633622765541, + 0.09321357309818268, + 0.5169685482978821 + ], + "credit_map": [ + 0.006953248288482428, + 0.006398661062121391, + 0.005699862260371447, + 0.004211733117699623, + 0.007937537506222725, + 0.004913635551929474, + 0.0049509210512042046, + 0.003317933762446046, + 0.006345237605273724, + 0.005308008287101984, + 0.0037833808455616236, + 0.005545444320887327, + 0.015127100050449371, + 0.006778548937290907, + 0.0048917667008936405, + 0.0028380497824400663, + 0.003446781076490879, + 0.0030466162133961916, + 0.002057315781712532, + 0.0023877567145973444, + 0.007137760054320097, + 0.0031555097084492445, + 0.004142186604440212, + 0.0020646434277296066, + 0.0030349174048751593, + 0.0021948812063783407, + 0.0018361901165917516, + 0.0016060523921623826, + 0.0041771396063268185, + 0.0026903939433395863, + 0.0022157945204526186, + 0.0014004171825945377, + 0.009395240806043148, + 0.009826396591961384, + 0.004720697645097971, + 0.007519193459302187, + 0.020501308143138885, + 0.008738572709262371, + 0.00900863204151392, + 0.006227945908904076, + 0.012106419540941715, + 0.0068691186606884, + 0.00488703977316618, + 0.01049472764134407, + 0.010881816036999226, + 0.010180717334151268, + 0.012399419210851192, + 0.005429295357316732, + 0.0052823289297521114, + 0.006484176963567734, + 0.003838181961327791, + 0.0057147350162267685, + 0.010200371034443378, + 0.005550680682063103, + 0.005344204604625702, + 0.00347485882230103, + 0.011108482256531715, + 0.0035025919787585735, + 0.004704308230429888, + 0.005658342968672514, + 0.007770082447677851, + 0.011246941983699799, + 0.006382447201758623, + 0.002773144980892539, + 0.012763699516654015, + 0.006956455763429403, + 0.02065829373896122, + 0.00692772539332509, + 0.013690979219973087, + 0.012283029034733772, + 0.009766479954123497, + 0.010166928172111511, + 0.15898685157299042, + 0.026829442009329796, + 0.13951429724693298, + 0.025539208203554153, + 0.028209762647747993, + 0.04912567883729935, + 0.032280273735523224, + 0.05648307502269745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.010473879054188728 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly monk wearing a faded, ochre-yellow, coarse robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk sitting cross-legged on a dark cushion with his back straight and hands in meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a quiet courtyard of an ancient monastery with pristine wooden eaves and a lightly frosted bluestone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk emit a small, dense cloud of white mist from his nostrils that swirls and dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot from a level angle that slowly and steadily moves forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012558192014694214 + ], + "opposite_logprob_tokens": [ + -4.3875579833984375 + ], + "logprob_reward": -0.012558192014694214, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012558192014694214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13317441940307617, + 0.0703003853559494, + 0.05823436379432678, + 0.10143108665943146, + 0.035504210740327835, + 0.0671861544251442, + 0.08915246278047562, + 0.04358009994029999, + 0.1478998064994812, + 0.25353702902793884 + ], + "credit_map": [ + 0.024120772257447243, + 0.0205165334045887, + 0.011528080329298973, + 0.013423603028059006, + 0.022603396326303482, + 0.017096633091568947, + 0.01689234934747219, + 0.006993028335273266, + 0.008199009113013744, + 0.00788759347051382, + 0.005932895466685295, + 0.0058555686846375465, + 0.02216641791164875, + 0.00762549601495266, + 0.0077992877922952175, + 0.004834114573895931, + 0.0071869236417114735, + 0.00729399686679244, + 0.005093132611364126, + 0.006352666299790144, + 0.013385091908276081, + 0.007126301992684603, + 0.006810344289988279, + 0.004985904321074486, + 0.014793135225772858, + 0.009917622432112694, + 0.007471925113350153, + 0.008439041674137115, + 0.022262444719672203, + 0.018688375130295753, + 0.012609746307134628, + 0.007248790934681892, + 0.006357284262776375, + 0.003570203436538577, + 0.005650798324495554, + 0.003524098312482238, + 0.004866293631494045, + 0.003610013984143734, + 0.004118077922612429, + 0.0038074387703090906, + 0.009731121361255646, + 0.007840766571462154, + 0.006224766839295626, + 0.00646591791883111, + 0.017595821991562843, + 0.006878903601318598, + 0.007420077919960022, + 0.005028781481087208, + 0.009602370671927929, + 0.010070809163153172, + 0.008498897776007652, + 0.01371368020772934, + 0.015173140913248062, + 0.011005363427102566, + 0.014247105456888676, + 0.006841087248176336, + 0.009144216775894165, + 0.0038944424595683813, + 0.003740705084055662, + 0.004220335744321346, + 0.005512726027518511, + 0.008715590462088585, + 0.0054902490228414536, + 0.0028618338983505964, + 0.018009107559919357, + 0.01263438444584608, + 0.02975788153707981, + 0.013874624855816364, + 0.02180328778922558, + 0.015975089743733406, + 0.02146550826728344, + 0.014379900880157948, + 0.07978760451078415, + 0.012685243971645832, + 0.07344260066747665, + 0.012688538990914822, + 0.007422726135700941, + 0.021479735150933266, + 0.01641569659113884, + 0.029614876955747604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.012558192014694214 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly monk wearing a faded, ochre-yellow, coarse robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk sitting cross-legged on a dark cushion with his back straight and hands in meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a quiet courtyard of an ancient monastery with pristine wooden eaves and a lightly frosted bluestone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk emit a small, dense cloud of white mist from his nostrils that swirls and dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot from a level angle that slowly and steadily moves forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496016517281532 + ], + "opposite_logprob_tokens": [ + -2.929960250854492 + ], + "logprob_reward": -0.05496016517281532, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496016517281532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23430894315242767, + 0.07826872915029526, + 0.07777377963066101, + 0.050586484372615814, + 0.14734847843647003, + 0.041145775467157364, + 0.0793667808175087, + 0.06230941042304039, + 0.07832194119691849, + 0.15056967735290527 + ], + "credit_map": [ + 0.0493340827524662, + 0.028335247188806534, + 0.034603361040353775, + 0.016526665538549423, + 0.027992263436317444, + 0.040065065026283264, + 0.01860749162733555, + 0.018844766542315483, + 0.016136907041072845, + 0.006463938858360052, + 0.010948861949145794, + 0.004847703035920858, + 0.01324803289026022, + 0.010116621851921082, + 0.007766088470816612, + 0.008740569464862347, + 0.01928427815437317, + 0.008514972403645515, + 0.007231556810438633, + 0.006484708283096552, + 0.0072372364811599255, + 0.016319740563631058, + 0.005788798909634352, + 0.006912493612617254, + 0.007998472079634666, + 0.006478656083345413, + 0.00613026600331068, + 0.004850059747695923, + 0.008431320078670979, + 0.00569521076977253, + 0.004921505693346262, + 0.00608099065721035, + 0.019698262214660645, + 0.018634995445609093, + 0.029140226542949677, + 0.013339935801923275, + 0.018056131899356842, + 0.018497003242373466, + 0.011453366838395596, + 0.018528549000620842, + 0.005978941917419434, + 0.004126026760786772, + 0.004207331221550703, + 0.006191677879542112, + 0.004928450100123882, + 0.005107305012643337, + 0.007169435266405344, + 0.0034366059117019176, + 0.011461624875664711, + 0.009359081275761127, + 0.009435546584427357, + 0.008770568296313286, + 0.01477448083460331, + 0.009052559733390808, + 0.008979094214737415, + 0.007533824071288109, + 0.015147816389799118, + 0.004597015213221312, + 0.005233443807810545, + 0.004765457008033991, + 0.007395809981971979, + 0.01618446409702301, + 0.005672214087098837, + 0.0033131965901702642, + 0.01168797817081213, + 0.0073951901867985725, + 0.010060855187475681, + 0.008240148425102234, + 0.011471363715827465, + 0.011139003559947014, + 0.01167032029479742, + 0.0066570863127708435, + 0.03441271558403969, + 0.009115869179368019, + 0.04192829132080078, + 0.012162194587290287, + 0.01159342285245657, + 0.011729276739060879, + 0.011179851368069649, + 0.018448056653141975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.05496016517281532 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly monk wearing a faded, ochre-yellow, coarse robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk sitting cross-legged on a dark cushion with his back straight and hands in meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a quiet courtyard of an ancient monastery with pristine wooden eaves and a lightly frosted bluestone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk emit a small, dense cloud of white mist from his nostrils that swirls and dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot from a level angle that slowly and steadily moves forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428031474351883 + ], + "opposite_logprob_tokens": [ + -2.017803192138672 + ], + "logprob_reward": -0.1428031474351883, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1428031474351883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07394225150346756, + 0.0709686204791069, + 0.06341306865215302, + 0.06129363551735878, + 0.10820619016885757, + 0.22019904851913452, + 0.03379248455166817, + 0.020169993862509727, + 0.06897581368684769, + 0.2790389358997345 + ], + "credit_map": [ + 0.014874570071697235, + 0.008749444968998432, + 0.006208071485161781, + 0.007466612849384546, + 0.00885731726884842, + 0.01239580474793911, + 0.010100753046572208, + 0.005289673805236816, + 0.011102333664894104, + 0.006367525085806847, + 0.006088061258196831, + 0.007003668695688248, + 0.014727451838552952, + 0.012707909569144249, + 0.008595045655965805, + 0.004376625642180443, + 0.008372141048312187, + 0.006824067328125238, + 0.003669018391519785, + 0.004916413687169552, + 0.017569320276379585, + 0.010698298923671246, + 0.007342022377997637, + 0.004021784290671349, + 0.00905704777687788, + 0.005575880408287048, + 0.0076090800575912, + 0.006486076395958662, + 0.009302929043769836, + 0.01111435517668724, + 0.006888206582516432, + 0.005260058678686619, + 0.014359808526933193, + 0.011271758936345577, + 0.008858243934810162, + 0.013328073546290398, + 0.021298306062817574, + 0.01634822227060795, + 0.015932120382785797, + 0.006809653248637915, + 0.03251330927014351, + 0.021321389824151993, + 0.014769170433282852, + 0.030688155442476273, + 0.037778254598379135, + 0.030833963304758072, + 0.0341942124068737, + 0.018100587651133537, + 0.005866220220923424, + 0.004024188965559006, + 0.0028153189923614264, + 0.0038312808610498905, + 0.00407473836094141, + 0.004413997754454613, + 0.006386141758412123, + 0.0023805969394743443, + 0.0030196323059499264, + 0.002368688816204667, + 0.0027326324488967657, + 0.0021693799644708633, + 0.002291502198204398, + 0.0027388266753405333, + 0.002856853185221553, + 0.0019924775697290897, + 0.011003602296113968, + 0.005986079573631287, + 0.01248889695852995, + 0.007426484487950802, + 0.010546884499490261, + 0.006012252531945705, + 0.009703424759209156, + 0.005808187182992697, + 0.09275469183921814, + 0.01310106460005045, + 0.07176905870437622, + 0.013829702511429787, + 0.00785916019231081, + 0.03140517324209213, + 0.019803741946816444, + 0.02851634845137596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 4 + ], + "selected_reward": -0.1428031474351883 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly monk wearing a faded, ochre-yellow, coarse robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk sitting cross-legged on a dark cushion with his back straight and hands in meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a quiet courtyard of an ancient monastery with pristine wooden eaves and a lightly frosted bluestone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk emit a small, dense cloud of white mist from his nostrils that swirls and dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot from a level angle that slowly and steadily moves forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253711897879839 + ], + "opposite_logprob_tokens": [ + -4.387537002563477 + ], + "logprob_reward": -0.01253711897879839, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253711897879839, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05802517384290695, + 0.025331659242510796, + 0.01833159849047661, + 0.028941169381141663, + 0.04303882643580437, + 0.07805410772562027, + 0.06955261528491974, + 0.03782710060477257, + 0.13952389359474182, + 0.501373827457428 + ], + "credit_map": [ + 0.011128556914627552, + 0.005987042095512152, + 0.004813802894204855, + 0.006380817852914333, + 0.007743173278868198, + 0.006844314280897379, + 0.009242772124707699, + 0.00588469160720706, + 0.004065656568855047, + 0.0025183989200741053, + 0.0020201685838401318, + 0.0022428766824305058, + 0.005878905765712261, + 0.004001113120466471, + 0.002577585633844137, + 0.0020269551314413548, + 0.002720276825129986, + 0.002107032807543874, + 0.0016911919228732586, + 0.0024831173941493034, + 0.0031851017847657204, + 0.0024905442260205746, + 0.002439499832689762, + 0.001214833464473486, + 0.005531031638383865, + 0.0025136310141533613, + 0.0028360772412270308, + 0.0032613968942314386, + 0.0037197645287960768, + 0.006219121627509594, + 0.002452095737680793, + 0.0024080509319901466, + 0.005832785740494728, + 0.00485253008082509, + 0.003627413185313344, + 0.004225356969982386, + 0.011166559532284737, + 0.0061133126728236675, + 0.004364802502095699, + 0.0028560664504766464, + 0.00975709967315197, + 0.008807871490716934, + 0.004807453136891127, + 0.005789300426840782, + 0.02572464942932129, + 0.009152323938906193, + 0.008890156634151936, + 0.00512525113299489, + 0.009409814141690731, + 0.006859986577183008, + 0.006696261465549469, + 0.008653481490910053, + 0.011451244354248047, + 0.008233762346208096, + 0.009860302321612835, + 0.008387763053178787, + 0.00681470287963748, + 0.002469278872013092, + 0.0034151908475905657, + 0.00354608497582376, + 0.005438258405774832, + 0.008832474239170551, + 0.004557157400995493, + 0.0027539501897990704, + 0.020110005512833595, + 0.012104163877665997, + 0.02406233735382557, + 0.01399125624448061, + 0.019217176362872124, + 0.01845330372452736, + 0.018376832827925682, + 0.013208813033998013, + 0.14263293147087097, + 0.02498171292245388, + 0.131201833486557, + 0.03302320837974548, + 0.029745586216449738, + 0.036182332783937454, + 0.04961853474378586, + 0.05398768186569214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.01253711897879839 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a level angle, slowly and steadily moves forward, focusing on an elderly monk. His shiny head gleams faintly in the morning light, his forehead and eyes crisscrossed with deep, furrowed lines. Clad in a faded, ochre-yellow, coarse robe, he sits cross-legged on a dark cushion in the quiet courtyard of an ancient monastery. Behind him, the temple's pristine wooden eaves, their outlines gradually becoming clear in the morning light, and beneath his feet, a lightly frosted bluestone floor. His back is straight, his hands clasped in his lap in meditation mudra, his eyes closed, his face radiating utter serenity and concentration. With each slow, deep, and rhythmic breath, a small, dense cloud of white mist gently emanates from his nostrils, briefly condensing and swirling in the cold air before curling upwards and silently dissipating into the chill dawn light. The dominant tones of the entire painting are calm blue-gray and deep wood tones. Only the first glimmer of light at dawn gives the monk's outline a faint golden edge. The overall atmosphere is solemn and tranquil, filled with Zen and a deep sense of tranquility that transcends time.", + "seed": 23956, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_1/index_5954_seed_23956_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06788666471838951, + "vlm_logprob": -0.06788666471838951, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.06788666471838951, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly monk wearing a faded, ochre-yellow, coarse robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk sitting cross-legged on a dark cushion with his back straight and hands in meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a quiet courtyard of an ancient monastery with pristine wooden eaves and a lightly frosted bluestone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk emit a small, dense cloud of white mist from his nostrils that swirls and dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot from a level angle that slowly and steadily moves forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00994996726512909 + ], + "opposite_logprob_tokens": [ + -5.009950160980225 + ], + "logprob_reward": -0.00994996726512909, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00994996726512909, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2592760920524597, + 0.0750453770160675, + 0.07248751819133759, + 0.09060466289520264, + 0.04520988091826439, + 0.051408831030130386, + 0.08671120554208755, + 0.12050429731607437, + 0.12605929374694824, + 0.07269281893968582 + ], + "credit_map": [ + 0.05655611306428909, + 0.035524141043424606, + 0.036002714186906815, + 0.018378913402557373, + 0.02321436256170273, + 0.043395232409238815, + 0.028110811486840248, + 0.018093815073370934, + 0.017677098512649536, + 0.011102017015218735, + 0.008261062204837799, + 0.004829109646379948, + 0.006656465120613575, + 0.01341144647449255, + 0.006854652427136898, + 0.006253537256270647, + 0.009115209802985191, + 0.0226645078510046, + 0.010605359449982643, + 0.0037992140278220177, + 0.007069411687552929, + 0.008560528047382832, + 0.005957658402621746, + 0.004715632647275925, + 0.015522345900535583, + 0.022361543029546738, + 0.014451945200562477, + 0.006041340995579958, + 0.007230428513139486, + 0.013270758092403412, + 0.0071283676661551, + 0.004597949329763651, + 0.005266166757792234, + 0.008564947172999382, + 0.004316394682973623, + 0.005662319716066122, + 0.005711392033845186, + 0.0039693936705589294, + 0.008401533588767052, + 0.0033177356235682964, + 0.006777503993362188, + 0.012439172714948654, + 0.006864754483103752, + 0.005350873339921236, + 0.005734201520681381, + 0.0068421559408307076, + 0.004869488067924976, + 0.0025306849274784327, + 0.009730805642902851, + 0.02034825086593628, + 0.012111467309296131, + 0.006719884928315878, + 0.009984482079744339, + 0.011652090586721897, + 0.00949924811720848, + 0.006664982531219721, + 0.016515394672751427, + 0.02962101809680462, + 0.014095858670771122, + 0.010600008070468903, + 0.01158701442182064, + 0.014140717685222626, + 0.014864521101117134, + 0.009079758077859879, + 0.008984825573861599, + 0.04760448634624481, + 0.027296755462884903, + 0.00489733787253499, + 0.012192398309707642, + 0.00643889931961894, + 0.014054309576749802, + 0.0045902966521680355, + 0.005953536368906498, + 0.013222814537584782, + 0.008094096556305885, + 0.006045411806553602, + 0.009997962974011898, + 0.007021793629974127, + 0.012663722969591618, + 0.00969348382204771 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.00994996726512909 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly monk wearing a faded, ochre-yellow, coarse robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk sitting cross-legged on a dark cushion with his back straight and hands in meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a quiet courtyard of an ancient monastery with pristine wooden eaves and a lightly frosted bluestone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk emit a small, dense cloud of white mist from his nostrils that swirls and dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot from a level angle that slowly and steadily moves forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029816025868058205 + ], + "opposite_logprob_tokens": [ + -3.529815912246704 + ], + "logprob_reward": -0.029816025868058205, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029816025868058205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2327955663204193, + 0.03840223699808121, + 0.16534405946731567, + 0.09559560567140579, + 0.2131568342447281, + 0.06310946494340897, + 0.03144674375653267, + 0.04533310607075691, + 0.06532609462738037, + 0.049490343779325485 + ], + "credit_map": [ + 0.048555485904216766, + 0.03379106521606445, + 0.03574639558792114, + 0.016187302768230438, + 0.02505525015294552, + 0.035060834139585495, + 0.02353276126086712, + 0.014866424724459648, + 0.006331646349281073, + 0.005286112893372774, + 0.003922862932085991, + 0.002979515353217721, + 0.004213598091155291, + 0.006697296164929867, + 0.005019026808440685, + 0.0039521693252027035, + 0.019594088196754456, + 0.05971476063132286, + 0.030062824487686157, + 0.007621565833687782, + 0.012598138302564621, + 0.014357013627886772, + 0.013699661009013653, + 0.0076959640718996525, + 0.01566779799759388, + 0.025668516755104065, + 0.013965792022645473, + 0.00611737510189414, + 0.009347120299935341, + 0.011519033461809158, + 0.00837240181863308, + 0.004937547259032726, + 0.033519379794597626, + 0.06293417513370514, + 0.038997069001197815, + 0.012526019476354122, + 0.01696198061108589, + 0.019880663603544235, + 0.018176427111029625, + 0.010161090642213821, + 0.011066669598221779, + 0.007701528258621693, + 0.005005722865462303, + 0.006501294672489166, + 0.00881496723741293, + 0.010995370335876942, + 0.008640299551188946, + 0.004383602645248175, + 0.006025268230587244, + 0.005349336192011833, + 0.0027942892629653215, + 0.002185551216825843, + 0.0031442265026271343, + 0.007216593716293573, + 0.0028145085088908672, + 0.0019169648876413703, + 0.007878494448959827, + 0.007629744242876768, + 0.0047923303209245205, + 0.004998908843845129, + 0.004746236838400364, + 0.006104969885200262, + 0.005744146648794413, + 0.003438265062868595, + 0.004448938649147749, + 0.022514425218105316, + 0.015117142349481583, + 0.004461286589503288, + 0.006166716571897268, + 0.003934427630156279, + 0.0065397960133850574, + 0.0021433497313410044, + 0.004276860039681196, + 0.014284956268966198, + 0.009491214528679848, + 0.0031149701680988073, + 0.00513277156278491, + 0.00419164402410388, + 0.005485437344759703, + 0.0035124800633639097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.029816025868058205 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly monk wearing a faded, ochre-yellow, coarse robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk sitting cross-legged on a dark cushion with his back straight and hands in meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a quiet courtyard of an ancient monastery with pristine wooden eaves and a lightly frosted bluestone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk emit a small, dense cloud of white mist from his nostrils that swirls and dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot from a level angle that slowly and steadily moves forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129402294754982 + ], + "opposite_logprob_tokens": [ + -2.2379403114318848 + ], + "logprob_reward": -0.1129402294754982, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129402294754982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1548536717891693, + 0.04687368869781494, + 0.12285028398036957, + 0.06623046845197678, + 0.23867681622505188, + 0.02604222670197487, + 0.04115339368581772, + 0.1960490196943283, + 0.07728660851716995, + 0.029983816668391228 + ], + "credit_map": [ + 0.03578063100576401, + 0.029549527913331985, + 0.01763094961643219, + 0.010193884372711182, + 0.013383101671934128, + 0.024960484355688095, + 0.01435840129852295, + 0.008996681310236454, + 0.009980573318898678, + 0.01005629450082779, + 0.005305456928908825, + 0.0030825240537524223, + 0.00330139696598053, + 0.00811717938631773, + 0.003953092265874147, + 0.003077171044424176, + 0.015437177382409573, + 0.042165521532297134, + 0.022470630705356598, + 0.004774928092956543, + 0.006782748270779848, + 0.017094988375902176, + 0.007890310138463974, + 0.0062339757569134235, + 0.009214308112859726, + 0.010872039012610912, + 0.012592515908181667, + 0.006998368538916111, + 0.007597954012453556, + 0.008937029168009758, + 0.005226690787822008, + 0.004791559185832739, + 0.03723743557929993, + 0.07472141087055206, + 0.044945620000362396, + 0.012468241155147552, + 0.01564684882760048, + 0.02671895921230316, + 0.01446431316435337, + 0.012473990209400654, + 0.003836961230263114, + 0.004632533527910709, + 0.002847768599167466, + 0.002864362206310034, + 0.003571559675037861, + 0.0035484889522194862, + 0.002997852861881256, + 0.0017426982522010803, + 0.009159286506474018, + 0.007542390376329422, + 0.005486840847879648, + 0.0030006736051291227, + 0.0031131121795624495, + 0.006967766676098108, + 0.0038556403014808893, + 0.002027678769081831, + 0.0189448781311512, + 0.06298933178186417, + 0.02064107358455658, + 0.020922493189573288, + 0.02078983373939991, + 0.025994786992669106, + 0.018411722034215927, + 0.007354895118623972, + 0.006299589294940233, + 0.028884269297122955, + 0.017972080036997795, + 0.0029724177438765764, + 0.0069799418561160564, + 0.003652276238426566, + 0.007887950167059898, + 0.002638083416968584, + 0.00225063762627542, + 0.00881381705403328, + 0.00570374121889472, + 0.0016458968166261911, + 0.003168137278407812, + 0.002022368600592017, + 0.003803055500611663, + 0.0025761642027646303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.1129402294754982 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly monk wearing a faded, ochre-yellow, coarse robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk sitting cross-legged on a dark cushion with his back straight and hands in meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a quiet courtyard of an ancient monastery with pristine wooden eaves and a lightly frosted bluestone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk emit a small, dense cloud of white mist from his nostrils that swirls and dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot from a level angle that slowly and steadily moves forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16038891673088074 + ], + "opposite_logprob_tokens": [ + -1.9103889465332031 + ], + "logprob_reward": -0.16038891673088074, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16038891673088074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11984380334615707, + 0.031823255121707916, + 0.10788005590438843, + 0.18901227414608002, + 0.2220233827829361, + 0.054081156849861145, + 0.028799235820770264, + 0.14680272340774536, + 0.03396480530500412, + 0.0657692477107048 + ], + "credit_map": [ + 0.02066073939204216, + 0.01996118575334549, + 0.014843443408608437, + 0.009818017482757568, + 0.01261144783347845, + 0.017250673845410347, + 0.016212301328778267, + 0.008486003614962101, + 0.004047923255711794, + 0.005065779201686382, + 0.003563743783161044, + 0.002600695937871933, + 0.0040440340526402, + 0.00602557510137558, + 0.003606001613661647, + 0.002869503339752555, + 0.013865262269973755, + 0.03838128224015236, + 0.021438932046294212, + 0.004936580080538988, + 0.00682986481115222, + 0.009939311072230339, + 0.006801106501370668, + 0.005687721539288759, + 0.0304112546145916, + 0.06123367324471474, + 0.03322041034698486, + 0.009338506497442722, + 0.013523964211344719, + 0.020545141771435738, + 0.010837281122803688, + 0.009902063757181168, + 0.03674793988466263, + 0.05514143779873848, + 0.037993960082530975, + 0.012914649210870266, + 0.017633352428674698, + 0.03191922605037689, + 0.01829702779650688, + 0.011375811882317066, + 0.007013852708041668, + 0.007744791451841593, + 0.004730803892016411, + 0.007040195167064667, + 0.008262553252279758, + 0.0068616983480751514, + 0.008501512929797173, + 0.003925755154341459, + 0.0035223299637436867, + 0.0031689561437815428, + 0.002861281391233206, + 0.002942606806755066, + 0.005595730617642403, + 0.0030082878656685352, + 0.005311348009854555, + 0.002388694789260626, + 0.024461708962917328, + 0.03696243464946747, + 0.01737361028790474, + 0.014922967180609703, + 0.012118916027247906, + 0.019325660541653633, + 0.014253636822104454, + 0.007383797783404589, + 0.0028116947505623102, + 0.01053509023040533, + 0.00651375949382782, + 0.0019213918130844831, + 0.0038629863411188126, + 0.0028678770177066326, + 0.0038511028978973627, + 0.0016009029932320118, + 0.00517989369109273, + 0.01976373977959156, + 0.011794453486800194, + 0.003488061483949423, + 0.006870832294225693, + 0.004497436340898275, + 0.008474057540297508, + 0.0057007791474461555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 7 + ], + "selected_reward": -0.16038891673088074 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly monk wearing a faded, ochre-yellow, coarse robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk sitting cross-legged on a dark cushion with his back straight and hands in meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a quiet courtyard of an ancient monastery with pristine wooden eaves and a lightly frosted bluestone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk emit a small, dense cloud of white mist from his nostrils that swirls and dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot from a level angle that slowly and steadily moves forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026338184252381325 + ], + "opposite_logprob_tokens": [ + -3.6513381004333496 + ], + "logprob_reward": -0.026338184252381325, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026338184252381325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10508023947477341, + 0.03063041903078556, + 0.09434917569160461, + 0.07805455476045609, + 0.19352726638317108, + 0.07247712463140488, + 0.030962694436311722, + 0.060481905937194824, + 0.2355465590953827, + 0.09888999909162521 + ], + "credit_map": [ + 0.01842261478304863, + 0.014383070170879364, + 0.013530748896300793, + 0.007262664381414652, + 0.01572584919631481, + 0.014824609272181988, + 0.012492879293859005, + 0.0084378095343709, + 0.0061400108970701694, + 0.005225552245974541, + 0.00323447585105896, + 0.002206824952736497, + 0.003040274837985635, + 0.003965068142861128, + 0.003085346193984151, + 0.0037328675389289856, + 0.011936837807297707, + 0.029634015634655952, + 0.013442398980259895, + 0.00400659441947937, + 0.0074845063500106335, + 0.011155039072036743, + 0.009957559406757355, + 0.006732237059623003, + 0.011639602482318878, + 0.022764792665839195, + 0.011510366573929787, + 0.0044168210588395596, + 0.008006450720131397, + 0.009743952192366123, + 0.006446742452681065, + 0.0035258366260677576, + 0.03199158236384392, + 0.05611443519592285, + 0.03410877659916878, + 0.012510985136032104, + 0.014140564948320389, + 0.01825285330414772, + 0.016740921884775162, + 0.0096671711653471, + 0.012256328947842121, + 0.01058027520775795, + 0.005307548679411411, + 0.006332946475595236, + 0.008763199672102928, + 0.01442132331430912, + 0.009169762954115868, + 0.0056457435712218285, + 0.00368239707313478, + 0.006798068061470985, + 0.0032180179841816425, + 0.002827306278049946, + 0.003568444401025772, + 0.005280275829136372, + 0.003270503133535385, + 0.002317683305591345, + 0.013805338181555271, + 0.010276886634528637, + 0.007014244794845581, + 0.005338778253644705, + 0.005152870435267687, + 0.010365409776568413, + 0.005394643172621727, + 0.0031337470281869173, + 0.01756579801440239, + 0.09121086448431015, + 0.05540842190384865, + 0.008691838942468166, + 0.019929716363549232, + 0.009904923848807812, + 0.025310734286904335, + 0.007524295710027218, + 0.007633992936462164, + 0.029127001762390137, + 0.01721196435391903, + 0.004781542345881462, + 0.010660773143172264, + 0.005495007615536451, + 0.013215283863246441, + 0.010764443315565586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.026338184252381325 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a level angle, slowly and steadily moves forward, focusing on an elderly monk. His shiny head gleams faintly in the morning light, his forehead and eyes crisscrossed with deep, furrowed lines. Clad in a faded, ochre-yellow, coarse robe, he sits cross-legged on a dark cushion in the quiet courtyard of an ancient monastery. Behind him, the temple's pristine wooden eaves, their outlines gradually becoming clear in the morning light, and beneath his feet, a lightly frosted bluestone floor. His back is straight, his hands clasped in his lap in meditation mudra, his eyes closed, his face radiating utter serenity and concentration. With each slow, deep, and rhythmic breath, a small, dense cloud of white mist gently emanates from his nostrils, briefly condensing and swirling in the cold air before curling upwards and silently dissipating into the chill dawn light. The dominant tones of the entire painting are calm blue-gray and deep wood tones. Only the first glimmer of light at dawn gives the monk's outline a faint golden edge. The overall atmosphere is solemn and tranquil, filled with Zen and a deep sense of tranquility that transcends time.", + "seed": 23957, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_1/index_5954_seed_23957_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.14988996647298336, + "vlm_logprob": -0.14988996647298336, + "vlm_yes_no_margin": 2.425, + "vlm_token_credit": -0.14988996647298336, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly monk wearing a faded, ochre-yellow, coarse robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk sitting cross-legged on a dark cushion with his back straight and hands in meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a quiet courtyard of an ancient monastery with pristine wooden eaves and a lightly frosted bluestone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk emit a small, dense cloud of white mist from his nostrils that swirls and dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot from a level angle that slowly and steadily moves forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02135463058948517 + ], + "opposite_logprob_tokens": [ + -4.146354675292969 + ], + "logprob_reward": -0.02135463058948517, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.02135463058948517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13169878721237183, + 0.09504829347133636, + 0.17244750261306763, + 0.03641097992658615, + 0.08966971933841705, + 0.0852607861161232, + 0.0876486599445343, + 0.0867030918598175, + 0.1539674997329712, + 0.061144743114709854 + ], + "credit_map": [ + 0.014617822133004665, + 0.028198717162013054, + 0.020968539640307426, + 0.016869792714715004, + 0.01016923412680626, + 0.013544072397053242, + 0.012989734299480915, + 0.01434086263179779, + 0.009662960655987263, + 0.018214304000139236, + 0.01145726628601551, + 0.015882296487689018, + 0.00926839467138052, + 0.010720370337367058, + 0.008554347790777683, + 0.011288348585367203, + 0.020462166517972946, + 0.04165922850370407, + 0.028705833479762077, + 0.028375810012221336, + 0.012461029924452305, + 0.015343078412115574, + 0.012003068812191486, + 0.013437265530228615, + 0.004441477824002504, + 0.008012481965124607, + 0.003345595672726631, + 0.003743763780221343, + 0.0037689704913645983, + 0.005587258376181126, + 0.0034833289682865143, + 0.004028101917356253, + 0.006545851938426495, + 0.010758223943412304, + 0.013702531345188618, + 0.014834623783826828, + 0.009201363660395145, + 0.010046234354376793, + 0.013851838186383247, + 0.010729051195085049, + 0.00835242960602045, + 0.010264383628964424, + 0.0062838466838002205, + 0.011799261905252934, + 0.017544718459248543, + 0.011340745724737644, + 0.009000885300338268, + 0.010674512013792992, + 0.006154848262667656, + 0.01909313164651394, + 0.013064471073448658, + 0.01107111107558012, + 0.009260828606784344, + 0.012909002602100372, + 0.00783148780465126, + 0.008263764902949333, + 0.007148146629333496, + 0.022189315408468246, + 0.01034869160503149, + 0.011154024861752987, + 0.011488386429846287, + 0.007603674195706844, + 0.009786712005734444, + 0.006984134204685688, + 0.01319024246186018, + 0.021953260526061058, + 0.017461102455854416, + 0.015273325145244598, + 0.021539412438869476, + 0.01860184408724308, + 0.028119612485170364, + 0.017828689888119698, + 0.007472148630768061, + 0.010625248774886131, + 0.007209157571196556, + 0.006810465361922979, + 0.006710410118103027, + 0.007310526445508003, + 0.007295259740203619, + 0.007711525075137615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.02135463058948517 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly monk wearing a faded, ochre-yellow, coarse robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk sitting cross-legged on a dark cushion with his back straight and hands in meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a quiet courtyard of an ancient monastery with pristine wooden eaves and a lightly frosted bluestone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk emit a small, dense cloud of white mist from his nostrils that swirls and dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot from a level angle that slowly and steadily moves forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11298143118619919 + ], + "opposite_logprob_tokens": [ + -2.2379813194274902 + ], + "logprob_reward": -0.11298143118619919, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11298143118619919, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13941611349582672, + 0.07878807187080383, + 0.17301273345947266, + 0.09167265892028809, + 0.10030418634414673, + 0.06500911712646484, + 0.0831889733672142, + 0.07931643724441528, + 0.12842988967895508, + 0.060861874371767044 + ], + "credit_map": [ + 0.020630283281207085, + 0.02704687975347042, + 0.018917758017778397, + 0.01834830269217491, + 0.011495274491608143, + 0.011075710877776146, + 0.017907196655869484, + 0.013994717039167881, + 0.010914420709013939, + 0.010301226750016212, + 0.00830881018191576, + 0.010281873866915703, + 0.010218805633485317, + 0.010806835256516933, + 0.008360530249774456, + 0.009595564566552639, + 0.017344357445836067, + 0.04219042509794235, + 0.023529227823019028, + 0.02855408750474453, + 0.013349286280572414, + 0.016338791698217392, + 0.015521240420639515, + 0.016185319051146507, + 0.007886596024036407, + 0.018673449754714966, + 0.014991872012615204, + 0.013774343766272068, + 0.009958709590137005, + 0.007767536677420139, + 0.009759939275681973, + 0.008860217407345772, + 0.008944683708250523, + 0.010568490251898766, + 0.008549428544938564, + 0.020422279834747314, + 0.011921697296202183, + 0.00897697452455759, + 0.017665816470980644, + 0.01325481291860342, + 0.005118712782859802, + 0.012440108694136143, + 0.006042788736522198, + 0.007379818707704544, + 0.007301541045308113, + 0.007561680860817432, + 0.009140419773757458, + 0.010024046525359154, + 0.0063234674744307995, + 0.01302539836615324, + 0.007442330941557884, + 0.008753260597586632, + 0.011671379208564758, + 0.012510797008872032, + 0.013590344227850437, + 0.009871994145214558, + 0.007079367060214281, + 0.014861045405268669, + 0.008748102001845837, + 0.009632863104343414, + 0.012726265005767345, + 0.007690491620451212, + 0.010581712238490582, + 0.007996591739356518, + 0.013201246038079262, + 0.013519641011953354, + 0.012849012389779091, + 0.010687788017094135, + 0.018751846626400948, + 0.023748528212308884, + 0.02153981477022171, + 0.014132009819149971, + 0.008543171919882298, + 0.006779499817639589, + 0.006580329034477472, + 0.006053668446838856, + 0.00798507034778595, + 0.008250078186392784, + 0.008021600544452667, + 0.008648456074297428 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.11298143118619919 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly monk wearing a faded, ochre-yellow, coarse robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk sitting cross-legged on a dark cushion with his back straight and hands in meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a quiet courtyard of an ancient monastery with pristine wooden eaves and a lightly frosted bluestone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk emit a small, dense cloud of white mist from his nostrils that swirls and dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot from a level angle that slowly and steadily moves forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3871579170227051 + ], + "opposite_logprob_tokens": [ + -1.137157917022705 + ], + "logprob_reward": -0.3871579170227051, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3871579170227051, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1763160526752472, + 0.07016599178314209, + 0.1562444567680359, + 0.07442674040794373, + 0.04597241058945656, + 0.062462519854307175, + 0.07106713205575943, + 0.15560029447078705, + 0.1112508773803711, + 0.07649350166320801 + ], + "credit_map": [ + 0.017775759100914, + 0.022896993905305862, + 0.018261563032865524, + 0.01753975637257099, + 0.03018794395029545, + 0.027679214254021645, + 0.021589670330286026, + 0.02038515917956829, + 0.007506738416850567, + 0.011733266524970531, + 0.006030571646988392, + 0.006143365986645222, + 0.01422509178519249, + 0.011073033325374126, + 0.006939623039215803, + 0.006514308974146843, + 0.020748773589730263, + 0.032672569155693054, + 0.02199520915746689, + 0.024591833353042603, + 0.01314591895788908, + 0.01565336063504219, + 0.01472396682947874, + 0.01271284930408001, + 0.00625253701582551, + 0.013409013859927654, + 0.013404437340795994, + 0.010513005778193474, + 0.006408404558897018, + 0.008095921948552132, + 0.007406431715935469, + 0.00893698912113905, + 0.005752177909016609, + 0.0037570230197161436, + 0.005669500213116407, + 0.006774562411010265, + 0.005692505743354559, + 0.006022179964929819, + 0.00749601423740387, + 0.004808447323739529, + 0.00606673676520586, + 0.005287335719913244, + 0.0047792112454771996, + 0.006410665810108185, + 0.009493733756244183, + 0.015124652534723282, + 0.005830604117363691, + 0.00946958176791668, + 0.005624144803732634, + 0.0066575296223163605, + 0.005308876279741526, + 0.008180812932550907, + 0.010846746154129505, + 0.008734062314033508, + 0.01532046403735876, + 0.010394501499831676, + 0.01349784154444933, + 0.01884419098496437, + 0.012050646357238293, + 0.016761625185608864, + 0.031199704855680466, + 0.015831071883440018, + 0.025493305176496506, + 0.021921930834650993, + 0.009406475350260735, + 0.009159580804407597, + 0.012625654228031635, + 0.00953695084899664, + 0.01691298373043537, + 0.015966802835464478, + 0.02268218621611595, + 0.014960243366658688, + 0.009190823882818222, + 0.004909167066216469, + 0.0065037463791668415, + 0.005153140053153038, + 0.010480581782758236, + 0.010845865122973919, + 0.013015244156122208, + 0.01639493741095066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.3871579170227051 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly monk wearing a faded, ochre-yellow, coarse robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk sitting cross-legged on a dark cushion with his back straight and hands in meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a quiet courtyard of an ancient monastery with pristine wooden eaves and a lightly frosted bluestone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk emit a small, dense cloud of white mist from his nostrils that swirls and dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot from a level angle that slowly and steadily moves forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20161418616771698 + ], + "opposite_logprob_tokens": [ + -1.7016141414642334 + ], + "logprob_reward": -0.20161418616771698, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20161418616771698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23376663029193878, + 0.12569591403007507, + 0.08357555419206619, + 0.03878533095121384, + 0.0854049101471901, + 0.08838692307472229, + 0.0830375999212265, + 0.08269711583852768, + 0.12225018441677094, + 0.05639982968568802 + ], + "credit_map": [ + 0.03400777280330658, + 0.03412681818008423, + 0.024723751470446587, + 0.0290695670992136, + 0.0199254360049963, + 0.024862242862582207, + 0.03794259577989578, + 0.029108457267284393, + 0.016194909811019897, + 0.01346353068947792, + 0.014651531353592873, + 0.01435747742652893, + 0.02174687571823597, + 0.01475661899894476, + 0.014078578911721706, + 0.016446391120553017, + 0.007357223425060511, + 0.011431512422859669, + 0.015144645236432552, + 0.015019171871244907, + 0.008808090351521969, + 0.00788249634206295, + 0.008345971815288067, + 0.009586445055902004, + 0.004387775901705027, + 0.004952374380081892, + 0.005386814009398222, + 0.005508446134626865, + 0.0046320161782205105, + 0.005887527950108051, + 0.0036354088224470615, + 0.004394968971610069, + 0.007297699339687824, + 0.010137570090591908, + 0.010139171034097672, + 0.015353302471339703, + 0.008833768777549267, + 0.005789693910628557, + 0.014371098950505257, + 0.01348260510712862, + 0.007312926463782787, + 0.012029617093503475, + 0.0078004151582717896, + 0.01681448146700859, + 0.01190345548093319, + 0.004356213845312595, + 0.015732184052467346, + 0.012437630444765091, + 0.005561488214880228, + 0.011410273611545563, + 0.007186288945376873, + 0.009137550368905067, + 0.013205497525632381, + 0.006827664561569691, + 0.017830973491072655, + 0.01187786553055048, + 0.005921503063291311, + 0.01862155832350254, + 0.011267421767115593, + 0.013789653778076172, + 0.009567901492118835, + 0.0068035670556128025, + 0.007951253093779087, + 0.00877425353974104, + 0.010883490554988384, + 0.010214133188128471, + 0.01260458305478096, + 0.011316176503896713, + 0.01721869222819805, + 0.015284746885299683, + 0.02717612311244011, + 0.017552228644490242, + 0.006427232176065445, + 0.005518297199159861, + 0.007774610538035631, + 0.007081638555973768, + 0.009213918820023537, + 0.00993288867175579, + 0.00470740906894207, + 0.005743832793086767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.20161418616771698 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly monk wearing a faded, ochre-yellow, coarse robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk sitting cross-legged on a dark cushion with his back straight and hands in meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a quiet courtyard of an ancient monastery with pristine wooden eaves and a lightly frosted bluestone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk emit a small, dense cloud of white mist from his nostrils that swirls and dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot from a level angle that slowly and steadily moves forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026341667398810387 + ], + "opposite_logprob_tokens": [ + -3.651341676712036 + ], + "logprob_reward": -0.026341667398810387, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026341667398810387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11616520583629608, + 0.05769685283303261, + 0.18793991208076477, + 0.12900525331497192, + 0.07171163707971573, + 0.07654480636119843, + 0.06578216701745987, + 0.09570367634296417, + 0.1275942325592041, + 0.07185621559619904 + ], + "credit_map": [ + 0.017878985032439232, + 0.018203284591436386, + 0.011443240568041801, + 0.012453543953597546, + 0.012883943505585194, + 0.01643303781747818, + 0.012990916147828102, + 0.013878255151212215, + 0.010971715673804283, + 0.007473635487258434, + 0.005158222746104002, + 0.008875701576471329, + 0.005901059601455927, + 0.007839434780180454, + 0.006192944943904877, + 0.0052841342985630035, + 0.02053670398890972, + 0.0448954813182354, + 0.028672581538558006, + 0.03132592514157295, + 0.014132854528725147, + 0.017770079895853996, + 0.0147990258410573, + 0.01580725610256195, + 0.011449001729488373, + 0.02808716520667076, + 0.01839362643659115, + 0.016720252111554146, + 0.01349618285894394, + 0.014155286364257336, + 0.013143899850547314, + 0.013559842482209206, + 0.007722329348325729, + 0.009476928971707821, + 0.010127559304237366, + 0.01123854611068964, + 0.010533890686929226, + 0.0068510486744344234, + 0.008940665982663631, + 0.006820666138082743, + 0.007044016383588314, + 0.013858458027243614, + 0.006785412784665823, + 0.009941632859408855, + 0.008736665360629559, + 0.01106543280184269, + 0.007947342470288277, + 0.011165841482579708, + 0.005339034833014011, + 0.005910107400268316, + 0.005904944613575935, + 0.00709130335599184, + 0.00910408515483141, + 0.009986959397792816, + 0.01338308583945036, + 0.009062645956873894, + 0.0077027189545333385, + 0.011993872001767159, + 0.010234312154352665, + 0.013020401820540428, + 0.015292548574507236, + 0.010326916351914406, + 0.013868819922208786, + 0.013264093548059464, + 0.009936929680407047, + 0.012300998903810978, + 0.01666804775595665, + 0.010951691307127476, + 0.02057337574660778, + 0.015633579343557358, + 0.026529086753726006, + 0.015000523999333382, + 0.007481176406145096, + 0.007745197508484125, + 0.009144655428826809, + 0.007512597367167473, + 0.009785099886357784, + 0.010746694169938564, + 0.008259708993136883, + 0.011181087233126163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 8 + ], + "selected_reward": -0.026341667398810387 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_10/index_tensor([96, 96])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_10/index_tensor([96, 96])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..560a5b662b79c8dd28001be7751c736beed2fccd --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_10/index_tensor([96, 96])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Utilizing high-speed camera technology, a single, sweeping, 360-degree sweep captures the breathtaking spectacle of the mythical Xuanwu, a tortoise, majestically situated at the center of a dazzling galaxy. This mountain-like tortoise's shell is no ordinary object, but rather composed of deep cosmic obsidian. Inscribed on its surface are intricate, ancient golden patterns resembling star patterns. Reflected by the nebula, these patterns appear to come alive, flowing with energy in perceptible slow motion. An equally enormous black snake, its scales shimmering like amethyst, coils tightly and powerfully around its body. Its head rises high, its scarlet eyes ablaze with primordial fire, and its forked tongue, captured in extreme slow motion, juts out like a dark red lightning bolt, a gaze surveying the universe with majesty. Against the backdrop lies the boundless cosmos, a swirling nebula of deep purple, fiery orange, and icy blue gas. Billions of stars, transformed into blurred streaks of light, cast a sacred glow around the Xuanwu's silhouette. The entire scene is shot in extreme slow motion, magnifying every tiny muscle movement and energy flow of Xuanwu to the extreme. There is no fierce fighting, but it is full of the tense and unparalleled sense of power before the start of a cosmic-level competitive game, as if the entire galaxy is holding its breath, witnessing this silent and magnificent confrontation.", + "seed": 27838, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_10/index_96_seed_27838_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.13712587924674152, + "vlm_logprob": -0.13712587924674152, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.13712587924674152, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large tortoise resembling the mythical Xuanwu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tortoise's shell made of deep cosmic obsidian with intricate golden patterns resembling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large black snake with amethyst-shimmering scales coiled tightly around the tortoise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a swirling nebula of deep purple, fiery orange, and icy blue gas with blurred star streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004814935382455587 + ], + "opposite_logprob_tokens": [ + -5.504815101623535 + ], + "logprob_reward": -0.004814935382455587, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004814935382455587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21101532876491547, + 0.10648136585950851, + 0.13263404369354248, + 0.04420913755893707, + 0.1005314365029335, + 0.06432611495256424, + 0.046596843749284744, + 0.08348090201616287, + 0.07366371154785156, + 0.137061208486557 + ], + "credit_map": [ + 0.028508923947811127, + 0.06750144064426422, + 0.011741383001208305, + 0.021724458783864975, + 0.019832247868180275, + 0.019213629886507988, + 0.017379123717546463, + 0.025114096701145172, + 0.015038501471281052, + 0.023815149441361427, + 0.009969707578420639, + 0.011318612843751907, + 0.015614225529134274, + 0.015268513932824135, + 0.007830129005014896, + 0.007626511622220278, + 0.025818318128585815, + 0.035184022039175034, + 0.005399944260716438, + 0.007294314913451672, + 0.035801857709884644, + 0.009606104344129562, + 0.007097868714481592, + 0.006431618705391884, + 0.007084065116941929, + 0.0057955593802034855, + 0.003913541790097952, + 0.004057881888002157, + 0.008627369999885559, + 0.007621550466865301, + 0.0033434720244258642, + 0.0037656917702406645, + 0.014550352469086647, + 0.02034733258187771, + 0.005548923742026091, + 0.013604371808469296, + 0.02036115899682045, + 0.012711336836218834, + 0.006706412881612778, + 0.006701542995870113, + 0.011705361306667328, + 0.007422429975122213, + 0.008922204375267029, + 0.007290109992027283, + 0.008462858386337757, + 0.007089145947247744, + 0.00440779235213995, + 0.00902620330452919, + 0.006039137952029705, + 0.005360986106097698, + 0.003918289206922054, + 0.005279707256704569, + 0.0048815226182341576, + 0.008297797292470932, + 0.005041873082518578, + 0.007777524180710316, + 0.012315473519265652, + 0.008009390905499458, + 0.006291660945862532, + 0.007875761017203331, + 0.007560123689472675, + 0.018489329144358635, + 0.008504131808876991, + 0.01443502213805914, + 0.014773566275835037, + 0.014596237801015377, + 0.007737409323453903, + 0.005898178089410067, + 0.0120425159111619, + 0.006167812738567591, + 0.005738742649555206, + 0.0067092375829815865, + 0.025388170033693314, + 0.03448401018977165, + 0.016474738717079163, + 0.011548809707164764, + 0.012995557859539986, + 0.011189633049070835, + 0.008697090670466423, + 0.016283182427287102 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.004814935382455587 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large tortoise resembling the mythical Xuanwu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tortoise's shell made of deep cosmic obsidian with intricate golden patterns resembling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large black snake with amethyst-shimmering scales coiled tightly around the tortoise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a swirling nebula of deep purple, fiery orange, and icy blue gas with blurred star streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329203300178051 + ], + "opposite_logprob_tokens": [ + -3.773292064666748 + ], + "logprob_reward": -0.02329203300178051, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329203300178051, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17184919118881226, + 0.06627143174409866, + 0.2490113377571106, + 0.030434297397732735, + 0.06533164530992508, + 0.028268825262784958, + 0.045780111104249954, + 0.08382640779018402, + 0.11370095610618591, + 0.14552585780620575 + ], + "credit_map": [ + 0.02308378368616104, + 0.056671805679798126, + 0.00958485808223486, + 0.01450319867581129, + 0.018358910456299782, + 0.01884172298014164, + 0.012992043979465961, + 0.017812857404351234, + 0.008868986740708351, + 0.011028971523046494, + 0.004877015016973019, + 0.006718858145177364, + 0.010756377130746841, + 0.00871253665536642, + 0.007104668766260147, + 0.00820401031523943, + 0.05363018810749054, + 0.06619095057249069, + 0.011413857340812683, + 0.014579404145479202, + 0.06011863052845001, + 0.016465863212943077, + 0.011986086145043373, + 0.01462634652853012, + 0.0062000686302781105, + 0.004351879935711622, + 0.0020606964826583862, + 0.003921231254935265, + 0.005848109722137451, + 0.0028802910819649696, + 0.0027900247368961573, + 0.002381994854658842, + 0.008150517009198666, + 0.017923200502991676, + 0.0033266476821154356, + 0.0055320593528449535, + 0.01675422489643097, + 0.005455506034195423, + 0.0037311867345124483, + 0.0044583044946193695, + 0.004260567482560873, + 0.004909987095743418, + 0.002496573608368635, + 0.003615678520873189, + 0.004386990796774626, + 0.0037981143686920404, + 0.00216922746039927, + 0.0026316873263567686, + 0.0068633356131613255, + 0.0053934091702103615, + 0.004489363171160221, + 0.004775800742208958, + 0.003964731469750404, + 0.005051571410149336, + 0.0035872880835086107, + 0.011654609814286232, + 0.012471701018512249, + 0.008824032731354237, + 0.005990908946841955, + 0.007224775850772858, + 0.008210219442844391, + 0.01785162463784218, + 0.008722069673240185, + 0.014531073160469532, + 0.02557053044438362, + 0.01986050419509411, + 0.012063358910381794, + 0.009768120013177395, + 0.012751410715281963, + 0.013964845798909664, + 0.008358878083527088, + 0.011363317258656025, + 0.024534733965992928, + 0.03160262107849121, + 0.018939225003123283, + 0.0138817448168993, + 0.013484181836247444, + 0.01473866868764162, + 0.009363613091409206, + 0.01898106187582016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.02329203300178051 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large tortoise resembling the mythical Xuanwu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tortoise's shell made of deep cosmic obsidian with intricate golden patterns resembling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large black snake with amethyst-shimmering scales coiled tightly around the tortoise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a swirling nebula of deep purple, fiery orange, and icy blue gas with blurred star streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07894394546747208 + ], + "opposite_logprob_tokens": [ + -2.578943967819214 + ], + "logprob_reward": -0.07894394546747208, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07894394546747208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0729944035410881, + 0.14808614552021027, + 0.05315662547945976, + 0.08686777204275131, + 0.18407689034938812, + 0.07742465287446976, + 0.0974530354142189, + 0.0556897334754467, + 0.09504707902669907, + 0.1292036473751068 + ], + "credit_map": [ + 0.008329343050718307, + 0.011068855412304401, + 0.005771538242697716, + 0.008646488189697266, + 0.005608026869595051, + 0.00827780831605196, + 0.011289767920970917, + 0.014002573676407337, + 0.017413977533578873, + 0.019582504406571388, + 0.014609064906835556, + 0.03040233813226223, + 0.018061289563775063, + 0.01908404380083084, + 0.017406631261110306, + 0.01152628380805254, + 0.008514163084328175, + 0.009498346596956253, + 0.0044844369404017925, + 0.005073599983006716, + 0.009653487242758274, + 0.007872527465224266, + 0.004325099755078554, + 0.0037349627818912268, + 0.016324317082762718, + 0.009666511788964272, + 0.007955288514494896, + 0.010752563364803791, + 0.011649567633867264, + 0.014509943313896656, + 0.007597721181809902, + 0.008411857299506664, + 0.02254643850028515, + 0.04890626668930054, + 0.008516347967088223, + 0.015432864427566528, + 0.047310713678598404, + 0.018229642882943153, + 0.01112357433885336, + 0.01201104186475277, + 0.01079715508967638, + 0.010686250403523445, + 0.009058634750545025, + 0.0104924775660038, + 0.008548890240490437, + 0.010508814826607704, + 0.0077261910773813725, + 0.009606244042515755, + 0.013973950408399105, + 0.019735459238290787, + 0.006576710846275091, + 0.009350772015750408, + 0.020909158512949944, + 0.012000672519207, + 0.00562812527641654, + 0.009278183802962303, + 0.009233768098056316, + 0.006009592209011316, + 0.004602300934493542, + 0.005393072031438351, + 0.004956863354891539, + 0.011015237309038639, + 0.0061106812208890915, + 0.008368217386305332, + 0.011265883222222328, + 0.013941032811999321, + 0.008743911050260067, + 0.008936206810176373, + 0.012571527622640133, + 0.02126225456595421, + 0.007197953760623932, + 0.011128307320177555, + 0.018416525796055794, + 0.024282511323690414, + 0.009450985118746758, + 0.0172563586384058, + 0.013124688528478146, + 0.019135601818561554, + 0.012067582458257675, + 0.015469389967620373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 9 + ], + "selected_reward": -0.07894394546747208 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large tortoise resembling the mythical Xuanwu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tortoise's shell made of deep cosmic obsidian with intricate golden patterns resembling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large black snake with amethyst-shimmering scales coiled tightly around the tortoise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a swirling nebula of deep purple, fiery orange, and icy blue gas with blurred star streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024845702573657036 + ], + "opposite_logprob_tokens": [ + -6.0024847984313965 + ], + "logprob_reward": -0.0024845702573657036, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024845702573657036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17056813836097717, + 0.11068683862686157, + 0.15443642437458038, + 0.04836402088403702, + 0.21113425493240356, + 0.09774654358625412, + 0.03085009567439556, + 0.034957561641931534, + 0.040273476392030716, + 0.10098256915807724 + ], + "credit_map": [ + 0.02019972912967205, + 0.05391065776348114, + 0.010006405413150787, + 0.01466768141835928, + 0.017301715910434723, + 0.024127691984176636, + 0.015217598527669907, + 0.015136663801968098, + 0.013005114160478115, + 0.01780449040234089, + 0.010242944583296776, + 0.01647649146616459, + 0.016403377056121826, + 0.014659265987575054, + 0.012363165616989136, + 0.009731994941830635, + 0.024458449333906174, + 0.032793931663036346, + 0.006302697584033012, + 0.011138618923723698, + 0.04504147544503212, + 0.013681384734809399, + 0.009898020885884762, + 0.011121846735477448, + 0.007751876022666693, + 0.006401974242180586, + 0.0029111907351762056, + 0.004684275481849909, + 0.01401247177273035, + 0.005751710385084152, + 0.0032661650329828262, + 0.003584357211366296, + 0.0278584323823452, + 0.05690881609916687, + 0.009509856812655926, + 0.01738651469349861, + 0.05424235388636589, + 0.019521547481417656, + 0.012066452763974667, + 0.013640292920172215, + 0.018135782331228256, + 0.021854551509022713, + 0.007333321496844292, + 0.007590491324663162, + 0.02518930286169052, + 0.005977306980639696, + 0.005854064133018255, + 0.005811723414808512, + 0.004440440330654383, + 0.003643132047727704, + 0.0029950125608593225, + 0.0036036078818142414, + 0.0033116452395915985, + 0.004558822605758905, + 0.002339377300813794, + 0.005958057474344969, + 0.005112507846206427, + 0.0042063589207828045, + 0.002619487466290593, + 0.002848594216629863, + 0.004387015476822853, + 0.006776537746191025, + 0.0034072333946824074, + 0.005599827505648136, + 0.009083847515285015, + 0.008477828465402126, + 0.004445626866072416, + 0.0034629476722329855, + 0.004908638074994087, + 0.004145022481679916, + 0.0024909882340580225, + 0.003258576150983572, + 0.016942571848630905, + 0.020909743383526802, + 0.009813710115849972, + 0.010313793085515499, + 0.010633065365254879, + 0.013025447726249695, + 0.009969085454940796, + 0.009375148452818394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.0024845702573657036 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large tortoise resembling the mythical Xuanwu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tortoise's shell made of deep cosmic obsidian with intricate golden patterns resembling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large black snake with amethyst-shimmering scales coiled tightly around the tortoise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a swirling nebula of deep purple, fiery orange, and icy blue gas with blurred star streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5760939121246338 + ], + "opposite_logprob_tokens": [ + -0.8260939121246338 + ], + "logprob_reward": -0.5760939121246338, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5760939121246338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10447452962398529, + 0.1378200799226761, + 0.08478149026632309, + 0.05306345969438553, + 0.18979045748710632, + 0.027110643684864044, + 0.03225959837436676, + 0.026249924674630165, + 0.08477486670017242, + 0.25967487692832947 + ], + "credit_map": [ + 0.012405689805746078, + 0.020173536613583565, + 0.009024719707667828, + 0.00813277903944254, + 0.00976169016212225, + 0.01542049553245306, + 0.012277422472834587, + 0.017278196290135384, + 0.01660452038049698, + 0.02033291384577751, + 0.014931637793779373, + 0.02882189303636551, + 0.015816444531083107, + 0.01569688878953457, + 0.015634121373295784, + 0.00998167134821415, + 0.01295938715338707, + 0.019189096987247467, + 0.004813372157514095, + 0.007715275045484304, + 0.01842200942337513, + 0.009477718733251095, + 0.006654184777289629, + 0.0055504534393548965, + 0.009431933984160423, + 0.005945821292698383, + 0.00418497622013092, + 0.008748382329940796, + 0.007854164578020573, + 0.00715893879532814, + 0.0048963879235088825, + 0.0048428550362586975, + 0.024810491129755974, + 0.051771119236946106, + 0.0077612269669771194, + 0.016812901943922043, + 0.049973856657743454, + 0.01653706654906273, + 0.01008809544146061, + 0.012035708874464035, + 0.005299608688801527, + 0.0032218496780842543, + 0.002816456602886319, + 0.0026492541655898094, + 0.003442011307924986, + 0.003334557870402932, + 0.002544896909967065, + 0.0038020103238523006, + 0.0040330663323402405, + 0.006359525956213474, + 0.0018727399874478579, + 0.0031324634328484535, + 0.005789344664663076, + 0.005384816788136959, + 0.002531516831368208, + 0.003156127408146858, + 0.0037316849920898676, + 0.0031817685812711716, + 0.002723564160987735, + 0.0030642629135400057, + 0.003414698177948594, + 0.003666140139102936, + 0.0028712551575154066, + 0.0035965493880212307, + 0.029981788247823715, + 0.027845434844493866, + 0.004441000521183014, + 0.003370286663994193, + 0.005356216337531805, + 0.006076504476368427, + 0.003165244357660413, + 0.00453840009868145, + 0.043857015669345856, + 0.05803607776761055, + 0.027604756876826286, + 0.02136683277785778, + 0.023257626220583916, + 0.02510030008852482, + 0.018290873616933823, + 0.04216143861413002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 1 + ], + "selected_reward": -0.5760939121246338 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Utilizing high-speed camera technology, a single, sweeping, 360-degree sweep captures the breathtaking spectacle of the mythical Xuanwu, a tortoise, majestically situated at the center of a dazzling galaxy. This mountain-like tortoise's shell is no ordinary object, but rather composed of deep cosmic obsidian. Inscribed on its surface are intricate, ancient golden patterns resembling star patterns. Reflected by the nebula, these patterns appear to come alive, flowing with energy in perceptible slow motion. An equally enormous black snake, its scales shimmering like amethyst, coils tightly and powerfully around its body. Its head rises high, its scarlet eyes ablaze with primordial fire, and its forked tongue, captured in extreme slow motion, juts out like a dark red lightning bolt, a gaze surveying the universe with majesty. Against the backdrop lies the boundless cosmos, a swirling nebula of deep purple, fiery orange, and icy blue gas. Billions of stars, transformed into blurred streaks of light, cast a sacred glow around the Xuanwu's silhouette. The entire scene is shot in extreme slow motion, magnifying every tiny muscle movement and energy flow of Xuanwu to the extreme. There is no fierce fighting, but it is full of the tense and unparalleled sense of power before the start of a cosmic-level competitive game, as if the entire galaxy is holding its breath, witnessing this silent and magnificent confrontation.", + "seed": 27839, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_10/index_96_seed_27839_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11141917062923312, + "vlm_logprob": -0.11141917062923312, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.11141917062923312, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large tortoise resembling the mythical Xuanwu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tortoise's shell made of deep cosmic obsidian with intricate golden patterns resembling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large black snake with amethyst-shimmering scales coiled tightly around the tortoise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a swirling nebula of deep purple, fiery orange, and icy blue gas with blurred star streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006769106257706881 + ], + "opposite_logprob_tokens": [ + -5.131769180297852 + ], + "logprob_reward": -0.006769106257706881, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.006769106257706881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08686517179012299, + 0.04935454577207565, + 0.11200886219739914, + 0.10429266840219498, + 0.20217902958393097, + 0.22814500331878662, + 0.033158257603645325, + 0.041970979422330856, + 0.050387583673000336, + 0.09163790941238403 + ], + "credit_map": [ + 0.013678805902600288, + 0.010191734880208969, + 0.0066924141719937325, + 0.010850119404494762, + 0.013638319447636604, + 0.010076082311570644, + 0.014187859371304512, + 0.007549833040684462, + 0.008796107955276966, + 0.004253405146300793, + 0.005230688489973545, + 0.005506743676960468, + 0.003645743941888213, + 0.010618877597153187, + 0.006959468591958284, + 0.004343509208410978, + 0.01760202832520008, + 0.018072782084345818, + 0.005618375260382891, + 0.008208483457565308, + 0.025777291506528854, + 0.011463722214102745, + 0.01630939543247223, + 0.008956794627010822, + 0.015092691406607628, + 0.01054976787418127, + 0.010505836457014084, + 0.019000347703695297, + 0.016712920740246773, + 0.012154573574662209, + 0.01158875972032547, + 0.008687768131494522, + 0.05950925499200821, + 0.010356701910495758, + 0.010737930424511433, + 0.006906751077622175, + 0.014543653465807438, + 0.0789012536406517, + 0.011644178070127964, + 0.0095793092623353, + 0.07552365958690643, + 0.00840587168931961, + 0.006053274497389793, + 0.011488037183880806, + 0.013294012285768986, + 0.09657967835664749, + 0.009509594179689884, + 0.007290862035006285, + 0.004541352856904268, + 0.0036300974898040295, + 0.003471768693998456, + 0.00522330217063427, + 0.0031277795787900686, + 0.002526830183342099, + 0.007956109009683132, + 0.002681018551811576, + 0.0067872353829443455, + 0.003752754768356681, + 0.00511843990534544, + 0.005412991624325514, + 0.004297778941690922, + 0.005594121757894754, + 0.0056921932846307755, + 0.00531546026468277, + 0.007616108283400536, + 0.00780984852463007, + 0.005029576364904642, + 0.007387660909444094, + 0.0050399769097566605, + 0.007264082320034504, + 0.006652018520981073, + 0.0035883127711713314, + 0.010588768869638443, + 0.007911127991974354, + 0.009559721685945988, + 0.010290907695889473, + 0.011411633342504501, + 0.021188518032431602, + 0.01123465783894062, + 0.009452568367123604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 2 + ], + "selected_reward": -0.006769106257706881 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large tortoise resembling the mythical Xuanwu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tortoise's shell made of deep cosmic obsidian with intricate golden patterns resembling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large black snake with amethyst-shimmering scales coiled tightly around the tortoise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a swirling nebula of deep purple, fiery orange, and icy blue gas with blurred star streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012547596357762814 + ], + "opposite_logprob_tokens": [ + -4.387547492980957 + ], + "logprob_reward": -0.012547596357762814, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012547596357762814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11607439070940018, + 0.03870224952697754, + 0.0690387561917305, + 0.10094328969717026, + 0.24413327872753143, + 0.29038873314857483, + 0.03516837954521179, + 0.01492055132985115, + 0.036850448697805405, + 0.053779978305101395 + ], + "credit_map": [ + 0.01818474754691124, + 0.01611366495490074, + 0.010437902063131332, + 0.013653979636728764, + 0.014774533919990063, + 0.014027664437890053, + 0.016455790027976036, + 0.012426100671291351, + 0.009313753806054592, + 0.0036145232152193785, + 0.00339067867025733, + 0.0022619995288550854, + 0.003512399969622493, + 0.008994187228381634, + 0.0035547856241464615, + 0.004059919621795416, + 0.011156552471220493, + 0.011200263164937496, + 0.003546132706105709, + 0.004974399693310261, + 0.016910945996642113, + 0.006989928428083658, + 0.008393651805818081, + 0.005866875872015953, + 0.014385762624442577, + 0.009461633861064911, + 0.010106186382472515, + 0.020466655492782593, + 0.014050873927772045, + 0.012063112109899521, + 0.012253005988895893, + 0.008156058378517628, + 0.07252620160579681, + 0.012188602238893509, + 0.013781859539449215, + 0.007672702893614769, + 0.01735164038836956, + 0.09464945644140244, + 0.014454049058258533, + 0.011508761905133724, + 0.09812420606613159, + 0.010734033770859241, + 0.0074915154837071896, + 0.012167808599770069, + 0.014822994358837605, + 0.12650224566459656, + 0.011248011142015457, + 0.009297916665673256, + 0.00539798429235816, + 0.002705291146412492, + 0.0035924233961850405, + 0.004825539421290159, + 0.0035351463593542576, + 0.0028598522767424583, + 0.008845936506986618, + 0.003406203119084239, + 0.002643082058057189, + 0.0015871275682002306, + 0.0019128118874505162, + 0.0016004313947632909, + 0.0016324889147654176, + 0.001720781670883298, + 0.002443137811496854, + 0.0013806879287585616, + 0.005868364125490189, + 0.004806618206202984, + 0.0038266240153461695, + 0.004410047549754381, + 0.005101313814520836, + 0.004245012067258358, + 0.00442354055121541, + 0.004168927203863859, + 0.007642220240086317, + 0.0035435159225016832, + 0.0053848824463784695, + 0.005564912687987089, + 0.007469399832189083, + 0.014890668913722038, + 0.004474925342947245, + 0.004809449426829815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.012547596357762814 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large tortoise resembling the mythical Xuanwu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tortoise's shell made of deep cosmic obsidian with intricate golden patterns resembling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large black snake with amethyst-shimmering scales coiled tightly around the tortoise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a swirling nebula of deep purple, fiery orange, and icy blue gas with blurred star streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012534528970718384 + ], + "opposite_logprob_tokens": [ + -4.3875346183776855 + ], + "logprob_reward": -0.012534528970718384, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012534528970718384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1127651259303093, + 0.04366327077150345, + 0.17081375420093536, + 0.13454236090183258, + 0.2441428303718567, + 0.09437857568264008, + 0.06275084614753723, + 0.04417233541607857, + 0.04642924293875694, + 0.04634157568216324 + ], + "credit_map": [ + 0.016672471538186073, + 0.016711033880710602, + 0.008709298446774483, + 0.013137866742908955, + 0.013069681823253632, + 0.013874935917556286, + 0.019411345943808556, + 0.011178510263562202, + 0.006253897212445736, + 0.004466371610760689, + 0.0044410196132957935, + 0.005516470409929752, + 0.0038809559773653746, + 0.009247020818293095, + 0.006082989741116762, + 0.003774551209062338, + 0.026614516973495483, + 0.027360666543245316, + 0.008817859925329685, + 0.013103500008583069, + 0.04086678475141525, + 0.016513969749212265, + 0.023186996579170227, + 0.014349478296935558, + 0.02123262919485569, + 0.01285144966095686, + 0.014154116623103619, + 0.023758213967084885, + 0.019654404371976852, + 0.01562800072133541, + 0.01597844809293747, + 0.011285102926194668, + 0.07130635529756546, + 0.013015210628509521, + 0.010150945745408535, + 0.008520104922354221, + 0.01779172569513321, + 0.09675281494855881, + 0.014531932771205902, + 0.012073751538991928, + 0.026437681168317795, + 0.0052681113593280315, + 0.0034420171286910772, + 0.007247950416058302, + 0.006629118230193853, + 0.036018285900354385, + 0.005002948455512524, + 0.004332471173256636, + 0.00975760817527771, + 0.0047178189270198345, + 0.006949663627892733, + 0.008721019141376019, + 0.004869182128459215, + 0.004460021387785673, + 0.01707679219543934, + 0.006198747083544731, + 0.007094002794474363, + 0.005678281653672457, + 0.006165564060211182, + 0.0052227117121219635, + 0.0036869796458631754, + 0.005740677006542683, + 0.004954115021973848, + 0.005630007945001125, + 0.007934531196951866, + 0.007304694969207048, + 0.004996111616492271, + 0.005803402047604322, + 0.004416367970407009, + 0.005986189469695091, + 0.004829687532037497, + 0.0051582674495875835, + 0.006705062463879585, + 0.004599715117365122, + 0.005277332384139299, + 0.005382336676120758, + 0.005680341273546219, + 0.008085353299975395, + 0.0049665444530546665, + 0.005644892808049917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 3 + ], + "selected_reward": -0.012534528970718384 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large tortoise resembling the mythical Xuanwu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tortoise's shell made of deep cosmic obsidian with intricate golden patterns resembling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large black snake with amethyst-shimmering scales coiled tightly around the tortoise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a swirling nebula of deep purple, fiery orange, and icy blue gas with blurred star streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00193813256919384 + ], + "opposite_logprob_tokens": [ + -6.251938343048096 + ], + "logprob_reward": -0.00193813256919384, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.00193813256919384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05009455978870392, + 0.033286359161138535, + 0.06355482339859009, + 0.0734780952334404, + 0.27423518896102905, + 0.412507563829422, + 0.022392811253666878, + 0.026746751740574837, + 0.02014625258743763, + 0.023557614535093307 + ], + "credit_map": [ + 0.008470476604998112, + 0.008253354579210281, + 0.004519338719546795, + 0.005417859647423029, + 0.005690868943929672, + 0.006454820279031992, + 0.005791108123958111, + 0.005496733356267214, + 0.006411116104573011, + 0.002358237514272332, + 0.004041033796966076, + 0.002219254383817315, + 0.0032888713758438826, + 0.006906271446496248, + 0.005055839661508799, + 0.0030057369731366634, + 0.010334974154829979, + 0.010334400460124016, + 0.0030698413029313087, + 0.004422543570399284, + 0.016092782840132713, + 0.006302305497229099, + 0.007676103617995977, + 0.005321873817592859, + 0.010075360536575317, + 0.007567510008811951, + 0.008081948384642601, + 0.012141117826104164, + 0.01216745376586914, + 0.00898236595094204, + 0.008071825839579105, + 0.006390515714883804, + 0.08157963305711746, + 0.013932613655924797, + 0.012794372625648975, + 0.008427012711763382, + 0.01979173719882965, + 0.10859813541173935, + 0.016364743933081627, + 0.01274694874882698, + 0.14228428900241852, + 0.014236140064895153, + 0.009820332750678062, + 0.016381172463297844, + 0.020323695614933968, + 0.18154428899288177, + 0.015034851618111134, + 0.01288277842104435, + 0.003193689975887537, + 0.0015900582075119019, + 0.002307708840817213, + 0.0032189565245062113, + 0.0018739254446700215, + 0.001981924520805478, + 0.00602308101952076, + 0.0022034680005162954, + 0.00425987783819437, + 0.0018437746912240982, + 0.0034591639414429665, + 0.003279323223978281, + 0.0027914317324757576, + 0.0040869638323783875, + 0.004246940836310387, + 0.0027792740147560835, + 0.003224695101380348, + 0.0029021254740655422, + 0.0021301191300153732, + 0.0024431634228676558, + 0.001849851687438786, + 0.0026134257204830647, + 0.0024229770060628653, + 0.00255989539436996, + 0.0030410645995289087, + 0.0020846270490437746, + 0.0031720343977212906, + 0.0031073528807610273, + 0.0030779088847339153, + 0.0036071862559765577, + 0.003082399722188711, + 0.0023850398138165474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.00193813256919384 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large tortoise resembling the mythical Xuanwu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tortoise's shell made of deep cosmic obsidian with intricate golden patterns resembling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large black snake with amethyst-shimmering scales coiled tightly around the tortoise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a swirling nebula of deep purple, fiery orange, and icy blue gas with blurred star streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233064889907837 + ], + "opposite_logprob_tokens": [ + -0.8983064889907837 + ], + "logprob_reward": -0.5233064889907837, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233064889907837, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03472286835312843, + 0.01808295212686062, + 0.08112816512584686, + 0.047714050859212875, + 0.096631720662117, + 0.5379916429519653, + 0.03583722189068794, + 0.09663044661283493, + 0.024490611627697945, + 0.026770267635583878 + ], + "credit_map": [ + 0.004257523454725742, + 0.005115929991006851, + 0.004557010252028704, + 0.004559954162687063, + 0.002896946156397462, + 0.003988988697528839, + 0.004587971139699221, + 0.004758544731885195, + 0.0030083763413131237, + 0.0017468675505369902, + 0.0020166472531855106, + 0.0016329531790688634, + 0.001307232421822846, + 0.0038025486283004284, + 0.001987470081076026, + 0.0025808552745729685, + 0.012162087485194206, + 0.013102812692523003, + 0.004306260962039232, + 0.005756201688200235, + 0.019160788506269455, + 0.008099699392914772, + 0.011271117255091667, + 0.007269198074936867, + 0.008007052354514599, + 0.004495546221733093, + 0.005130172241479158, + 0.008086549118161201, + 0.006379543337970972, + 0.005597834940999746, + 0.005902301054447889, + 0.004115050658583641, + 0.027448726817965508, + 0.005017162766307592, + 0.005239578895270824, + 0.003801263403147459, + 0.007462301291525364, + 0.037015367299318314, + 0.005729570519179106, + 0.004917748272418976, + 0.18413445353507996, + 0.019036991521716118, + 0.012812006287276745, + 0.022900547832250595, + 0.027518609538674355, + 0.23475830256938934, + 0.01975902169942856, + 0.017071738839149475, + 0.00477500120177865, + 0.002691595582291484, + 0.0039796833880245686, + 0.0050901263020932674, + 0.0033488678745925426, + 0.003286962630227208, + 0.009006382897496223, + 0.0036586010828614235, + 0.015391537919640541, + 0.0071837203577160835, + 0.011034601368010044, + 0.011095652356743813, + 0.010416064411401749, + 0.019077898934483528, + 0.013273145072162151, + 0.009157826192677021, + 0.004712667316198349, + 0.003310213331133127, + 0.0023173207882791758, + 0.0024020180571824312, + 0.002913889940828085, + 0.003096298547461629, + 0.0027032834477722645, + 0.003034919733181596, + 0.0033781961537897587, + 0.0014077952364459634, + 0.002685071201995015, + 0.0022567319683730602, + 0.004018042702227831, + 0.0077290344052016735, + 0.002917948178946972, + 0.002377446973696351 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 7 + ], + "selected_reward": -0.5233064889907837 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Utilizing high-speed camera technology, a single, sweeping, 360-degree sweep captures the breathtaking spectacle of the mythical Xuanwu, a tortoise, majestically situated at the center of a dazzling galaxy. This mountain-like tortoise's shell is no ordinary object, but rather composed of deep cosmic obsidian. Inscribed on its surface are intricate, ancient golden patterns resembling star patterns. Reflected by the nebula, these patterns appear to come alive, flowing with energy in perceptible slow motion. An equally enormous black snake, its scales shimmering like amethyst, coils tightly and powerfully around its body. Its head rises high, its scarlet eyes ablaze with primordial fire, and its forked tongue, captured in extreme slow motion, juts out like a dark red lightning bolt, a gaze surveying the universe with majesty. Against the backdrop lies the boundless cosmos, a swirling nebula of deep purple, fiery orange, and icy blue gas. Billions of stars, transformed into blurred streaks of light, cast a sacred glow around the Xuanwu's silhouette. The entire scene is shot in extreme slow motion, magnifying every tiny muscle movement and energy flow of Xuanwu to the extreme. There is no fierce fighting, but it is full of the tense and unparalleled sense of power before the start of a cosmic-level competitive game, as if the entire galaxy is holding its breath, witnessing this silent and magnificent confrontation.", + "seed": 27840, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_10/index_96_seed_27840_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.054965903307311235, + "vlm_logprob": -0.054965903307311235, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.054965903307311235, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large tortoise resembling the mythical Xuanwu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tortoise's shell made of deep cosmic obsidian with intricate golden patterns resembling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large black snake with amethyst-shimmering scales coiled tightly around the tortoise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a swirling nebula of deep purple, fiery orange, and icy blue gas with blurred star streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0037547105457633734 + ], + "opposite_logprob_tokens": [ + -5.753754615783691 + ], + "logprob_reward": -0.0037547105457633734, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0037547105457633734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12572279572486877, + 0.1635957956314087, + 0.03868217393755913, + 0.08064310252666473, + 0.045867834240198135, + 0.15193282067775726, + 0.03034653142094612, + 0.22347581386566162, + 0.06306225061416626, + 0.07667086273431778 + ], + "credit_map": [ + 0.026714177802205086, + 0.014694915153086185, + 0.01830318011343479, + 0.011000663973391056, + 0.009281538426876068, + 0.01829420030117035, + 0.01362994872033596, + 0.013804162852466106, + 0.018942469730973244, + 0.027717409655451775, + 0.012622523121535778, + 0.010481012985110283, + 0.03649336099624634, + 0.023308703675866127, + 0.01550319790840149, + 0.018527114763855934, + 0.004963921383023262, + 0.0038122457917779684, + 0.004549937788397074, + 0.004405690357089043, + 0.0038366832304745913, + 0.005675074178725481, + 0.0063583762384951115, + 0.005080245900899172, + 0.010844369418919086, + 0.006236734334379435, + 0.011161751113831997, + 0.004801838658750057, + 0.011248166672885418, + 0.020084811374545097, + 0.007327885366976261, + 0.008937547914683819, + 0.006176050752401352, + 0.007498558145016432, + 0.0077044907957315445, + 0.004471711814403534, + 0.004424637649208307, + 0.005642153788357973, + 0.004456595983356237, + 0.00549363624304533, + 0.021282240748405457, + 0.01750933565199375, + 0.020572716370224953, + 0.011452564969658852, + 0.012363904155790806, + 0.032343532890081406, + 0.01363700907677412, + 0.022771524265408516, + 0.0034160795621573925, + 0.004850445780903101, + 0.004471424501389265, + 0.0038123975973576307, + 0.00248982897028327, + 0.004979778081178665, + 0.00325054582208395, + 0.0030760313384234905, + 0.025021828711032867, + 0.02349736914038658, + 0.02321179397404194, + 0.028459567576646805, + 0.04013634845614433, + 0.031041650101542473, + 0.028705939650535583, + 0.023401319980621338, + 0.008517146110534668, + 0.003765727858990431, + 0.009285670705139637, + 0.0073382859118282795, + 0.005544485989958048, + 0.00813430454581976, + 0.00898135919123888, + 0.011495264247059822, + 0.007496785372495651, + 0.015099618583917618, + 0.012376922182738781, + 0.006135262548923492, + 0.01046588271856308, + 0.007827132940292358, + 0.010173007845878601, + 0.0070962486788630486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 5 + ], + "selected_reward": -0.0037547105457633734 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large tortoise resembling the mythical Xuanwu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tortoise's shell made of deep cosmic obsidian with intricate golden patterns resembling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large black snake with amethyst-shimmering scales coiled tightly around the tortoise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a swirling nebula of deep purple, fiery orange, and icy blue gas with blurred star streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005958887282758951 + ], + "opposite_logprob_tokens": [ + -5.1309590339660645 + ], + "logprob_reward": -0.005958887282758951, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005958887282758951, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05989179015159607, + 0.39931145310401917, + 0.020971430465579033, + 0.020334932953119278, + 0.0649949237704277, + 0.2354641854763031, + 0.02002999372780323, + 0.038286350667476654, + 0.06367818266153336, + 0.0770367830991745 + ], + "credit_map": [ + 0.01061838585883379, + 0.006417073309421539, + 0.007950058206915855, + 0.00713750347495079, + 0.006389874964952469, + 0.00873049907386303, + 0.006828376557677984, + 0.005820017773658037, + 0.0442538857460022, + 0.08554603904485703, + 0.030302351340651512, + 0.02882748655974865, + 0.10305741429328918, + 0.03590429201722145, + 0.04197683930397034, + 0.029443129897117615, + 0.002930881455540657, + 0.0030913257505744696, + 0.002500674920156598, + 0.0023260912857949734, + 0.002184472279623151, + 0.003131367266178131, + 0.002500721486285329, + 0.0023058955557644367, + 0.002528951270505786, + 0.00255011604167521, + 0.002558782696723938, + 0.002012121258303523, + 0.0030922393780201674, + 0.003186854301020503, + 0.002410904271528125, + 0.0019949637353420258, + 0.008723601698875427, + 0.009191379882395267, + 0.008557530120015144, + 0.007085309363901615, + 0.005884620826691389, + 0.011194915510714054, + 0.006880621891468763, + 0.007476941216737032, + 0.041782259941101074, + 0.04314698278903961, + 0.03360062837600708, + 0.01645919494330883, + 0.018875418230891228, + 0.048360373824834824, + 0.013338993303477764, + 0.01990034244954586, + 0.002355649834498763, + 0.0025992542505264282, + 0.003979602362960577, + 0.0029543107375502586, + 0.0013965561520308256, + 0.00207977881655097, + 0.0024581439793109894, + 0.0022066982928663492, + 0.00432470953091979, + 0.004272385500371456, + 0.004395191092044115, + 0.004821399226784706, + 0.005025734659284353, + 0.005210514180362225, + 0.0059052761644124985, + 0.004331141244620085, + 0.008839153684675694, + 0.003798280842602253, + 0.005903030280023813, + 0.009248240850865841, + 0.006959196180105209, + 0.008368817158043385, + 0.0112337376922369, + 0.009327719919383526, + 0.006722874008119106, + 0.013407672755420208, + 0.009748848155140877, + 0.006871250923722982, + 0.011774078942835331, + 0.009049586951732635, + 0.013112761080265045, + 0.0063497102819383144 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 9 + ], + "selected_reward": -0.005958887282758951 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large tortoise resembling the mythical Xuanwu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tortoise's shell made of deep cosmic obsidian with intricate golden patterns resembling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large black snake with amethyst-shimmering scales coiled tightly around the tortoise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a swirling nebula of deep purple, fiery orange, and icy blue gas with blurred star streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03807767853140831 + ], + "opposite_logprob_tokens": [ + -3.2880775928497314 + ], + "logprob_reward": -0.03807767853140831, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03807767853140831, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14303171634674072, + 0.14109335839748383, + 0.03361504524946213, + 0.04168730229139328, + 0.08250822126865387, + 0.1984763890504837, + 0.04478642717003822, + 0.10049928724765778, + 0.0902905985713005, + 0.12401165813207626 + ], + "credit_map": [ + 0.026842303574085236, + 0.014605165459215641, + 0.017228346318006516, + 0.015376055613160133, + 0.016103455796837807, + 0.021793503314256668, + 0.016364427283406258, + 0.014718452468514442, + 0.0166773721575737, + 0.019073888659477234, + 0.013573718257248402, + 0.01501027587801218, + 0.025487158447504044, + 0.0208871029317379, + 0.014560718089342117, + 0.015823127701878548, + 0.004958209581673145, + 0.004280353430658579, + 0.003137294203042984, + 0.0031939963810145855, + 0.0034088706597685814, + 0.007301089819520712, + 0.0042279851622879505, + 0.003107244148850441, + 0.004943191539496183, + 0.004175497684627771, + 0.004277847241610289, + 0.0036327207926660776, + 0.00735477963462472, + 0.008061770349740982, + 0.004635177552700043, + 0.004606316331773996, + 0.012132365256547928, + 0.013566014356911182, + 0.009476622566580772, + 0.009310089983046055, + 0.006324039306491613, + 0.013618840835988522, + 0.008068353869020939, + 0.010011897422373295, + 0.03406967222690582, + 0.030557839199900627, + 0.03319387137889862, + 0.01716557890176773, + 0.015596417710185051, + 0.03688971698284149, + 0.012523937039077282, + 0.018479349091649055, + 0.005684922449290752, + 0.0060257199220359325, + 0.006029548589140177, + 0.004166698083281517, + 0.0034656282514333725, + 0.011379767209291458, + 0.003433424048125744, + 0.004600718151777983, + 0.01112182904034853, + 0.011171572841703892, + 0.010552065446972847, + 0.012245483696460724, + 0.017841363325715065, + 0.014679173938930035, + 0.01280609518289566, + 0.010081703774631023, + 0.009785359725356102, + 0.004726480692625046, + 0.010720106773078442, + 0.01396801695227623, + 0.00879494659602642, + 0.01124162320047617, + 0.019855080172419548, + 0.011198990046977997, + 0.014177910052239895, + 0.02483302727341652, + 0.01970396563410759, + 0.011842343956232071, + 0.013737224042415619, + 0.015906667336821556, + 0.01435172650963068, + 0.009458787739276886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.03807767853140831 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large tortoise resembling the mythical Xuanwu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tortoise's shell made of deep cosmic obsidian with intricate golden patterns resembling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large black snake with amethyst-shimmering scales coiled tightly around the tortoise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a swirling nebula of deep purple, fiery orange, and icy blue gas with blurred star streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001510670525021851 + ], + "opposite_logprob_tokens": [ + -6.5015106201171875 + ], + "logprob_reward": -0.001510670525021851, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001510670525021851, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05289188027381897, + 0.09405118972063065, + 0.015928369015455246, + 0.15340788662433624, + 0.03649666905403137, + 0.47876399755477905, + 0.01645558327436447, + 0.03725335747003555, + 0.045221131294965744, + 0.06952989846467972 + ], + "credit_map": [ + 0.007997891865670681, + 0.006186429411172867, + 0.008022477850317955, + 0.00558497803285718, + 0.005199192091822624, + 0.00619832519441843, + 0.006553587503731251, + 0.0071489992551505566, + 0.012223813682794571, + 0.014471138827502728, + 0.01098764967173338, + 0.006582642439752817, + 0.01962227001786232, + 0.011344337835907936, + 0.009355374611914158, + 0.00946397427469492, + 0.0025164533872157335, + 0.0013504140079021454, + 0.0025574418250471354, + 0.0014913359191268682, + 0.0014431193703785539, + 0.003186901332810521, + 0.002065303735435009, + 0.0013174006016924977, + 0.0179672222584486, + 0.014134261757135391, + 0.017807310447096825, + 0.010469378903508186, + 0.027719058096408844, + 0.03387904167175293, + 0.014913564547896385, + 0.016518043354153633, + 0.004985998850315809, + 0.004726987797766924, + 0.004651044495403767, + 0.004217176232486963, + 0.0027301383670419455, + 0.005702327936887741, + 0.003494855249300599, + 0.0059881447814404964, + 0.08460026234388351, + 0.08340282738208771, + 0.06647717207670212, + 0.03348865732550621, + 0.03864859789609909, + 0.10240747034549713, + 0.02883058413863182, + 0.040908440947532654, + 0.0020467874128371477, + 0.0024542436003684998, + 0.0022799314465373755, + 0.0015456032706424594, + 0.0012660461943596601, + 0.002806641161441803, + 0.001365436939522624, + 0.0026908956933766603, + 0.006625653710216284, + 0.012172254733741283, + 0.004582558758556843, + 0.00307572353631258, + 0.002853469690307975, + 0.004261448048055172, + 0.0017193184467032552, + 0.0019629329908639193, + 0.007011084817349911, + 0.006233420222997665, + 0.006173532921820879, + 0.004569411277770996, + 0.004004497081041336, + 0.005963448900729418, + 0.002971435897052288, + 0.00829430390149355, + 0.01130073331296444, + 0.016395363956689835, + 0.010153470560908318, + 0.005584897939115763, + 0.007924944162368774, + 0.007468965835869312, + 0.00520293740555644, + 0.005498588550835848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 1 + ], + "selected_reward": -0.001510670525021851 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large tortoise resembling the mythical Xuanwu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tortoise's shell made of deep cosmic obsidian with intricate golden patterns resembling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large black snake with amethyst-shimmering scales coiled tightly around the tortoise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a swirling nebula of deep purple, fiery orange, and icy blue gas with blurred star streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2255275696516037 + ], + "opposite_logprob_tokens": [ + -1.6005275249481201 + ], + "logprob_reward": -0.2255275696516037, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255275696516037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13615237176418304, + 0.3071193993091583, + 0.05939498543739319, + 0.13937169313430786, + 0.03921779617667198, + 0.11770844459533691, + 0.019585568457841873, + 0.09836193919181824, + 0.054085832089185715, + 0.02900194190442562 + ], + "credit_map": [ + 0.02790691889822483, + 0.010477459989488125, + 0.016876379027962685, + 0.012202548794448376, + 0.011597229167819023, + 0.021098725497722626, + 0.021434864029288292, + 0.014558251947164536, + 0.030894123017787933, + 0.0502646379172802, + 0.03233208507299423, + 0.024311769753694534, + 0.0656423568725586, + 0.03296726197004318, + 0.02968648076057434, + 0.04102069139480591, + 0.008777623064815998, + 0.006359043065458536, + 0.005971146281808615, + 0.007008515764027834, + 0.007604993414133787, + 0.008551967330276966, + 0.00884535163640976, + 0.006276347208768129, + 0.01725667342543602, + 0.013343560509383678, + 0.020051145926117897, + 0.011145641095936298, + 0.020102480426430702, + 0.029222823679447174, + 0.011140391230583191, + 0.01710897497832775, + 0.005445405840873718, + 0.005101755261421204, + 0.006002032198011875, + 0.004133840091526508, + 0.0028254890348762274, + 0.005611251574009657, + 0.003651346080005169, + 0.006446676794439554, + 0.02103523164987564, + 0.015183312818408012, + 0.015353662893176079, + 0.008262841030955315, + 0.009837731719017029, + 0.03090800903737545, + 0.007342588622123003, + 0.009785063564777374, + 0.0023718117736279964, + 0.003448971314355731, + 0.002519327914342284, + 0.0028259980026632547, + 0.0017139724222943187, + 0.0021984197665005922, + 0.0021108565852046013, + 0.002396209631115198, + 0.010911275632679462, + 0.012767097912728786, + 0.010595835745334625, + 0.011636180803179741, + 0.015167140401899815, + 0.01551573071628809, + 0.012263256125152111, + 0.00950542651116848, + 0.006946774665266275, + 0.0034218786749988794, + 0.00521302642300725, + 0.007144398055970669, + 0.00522435549646616, + 0.00668542692437768, + 0.009563828818500042, + 0.00988614372909069, + 0.0038240416906774044, + 0.006001620553433895, + 0.00354410195723176, + 0.002307718386873603, + 0.003869517706334591, + 0.0037771319039165974, + 0.0033425840083509684, + 0.0023352261632680893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.2255275696516037 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Utilizing high-speed camera technology, a single, sweeping, 360-degree sweep captures the breathtaking spectacle of the mythical Xuanwu, a tortoise, majestically situated at the center of a dazzling galaxy. This mountain-like tortoise's shell is no ordinary object, but rather composed of deep cosmic obsidian. Inscribed on its surface are intricate, ancient golden patterns resembling star patterns. Reflected by the nebula, these patterns appear to come alive, flowing with energy in perceptible slow motion. An equally enormous black snake, its scales shimmering like amethyst, coils tightly and powerfully around its body. Its head rises high, its scarlet eyes ablaze with primordial fire, and its forked tongue, captured in extreme slow motion, juts out like a dark red lightning bolt, a gaze surveying the universe with majesty. Against the backdrop lies the boundless cosmos, a swirling nebula of deep purple, fiery orange, and icy blue gas. Billions of stars, transformed into blurred streaks of light, cast a sacred glow around the Xuanwu's silhouette. The entire scene is shot in extreme slow motion, magnifying every tiny muscle movement and energy flow of Xuanwu to the extreme. There is no fierce fighting, but it is full of the tense and unparalleled sense of power before the start of a cosmic-level competitive game, as if the entire galaxy is holding its breath, witnessing this silent and magnificent confrontation.", + "seed": 27841, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_10/index_96_seed_27841_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.15044980430975557, + "vlm_logprob": -0.15044980430975557, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.15044980430975557, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large tortoise resembling the mythical Xuanwu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tortoise's shell made of deep cosmic obsidian with intricate golden patterns resembling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large black snake with amethyst-shimmering scales coiled tightly around the tortoise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a swirling nebula of deep purple, fiery orange, and icy blue gas with blurred star streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004531709477305412 + ], + "opposite_logprob_tokens": [ + -5.7545318603515625 + ], + "logprob_reward": -0.004531709477305412, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.004531709477305412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1757323443889618, + 0.12271678447723389, + 0.12134318798780441, + 0.08667758852243423, + 0.09554947912693024, + 0.10257428884506226, + 0.04339500144124031, + 0.08293524384498596, + 0.12218806892633438, + 0.046887971460819244 + ], + "credit_map": [ + 0.02959110587835312, + 0.01640809141099453, + 0.019295882433652878, + 0.03201857954263687, + 0.024078546091914177, + 0.015847153961658478, + 0.02061813324689865, + 0.017874853685498238, + 0.017128989100456238, + 0.00959757249802351, + 0.013608966954052448, + 0.013295094482600689, + 0.01746428571641445, + 0.01928367093205452, + 0.020077167078852654, + 0.012261046096682549, + 0.014325952157378197, + 0.022949617356061935, + 0.011000851169228554, + 0.010852338746190071, + 0.016607675701379776, + 0.02254386618733406, + 0.008432365022599697, + 0.014630523510277271, + 0.01118413731455803, + 0.013253075070679188, + 0.007367311976850033, + 0.015046292915940285, + 0.01151816826313734, + 0.011889752000570297, + 0.0070263054221868515, + 0.009392544627189636, + 0.015180199407041073, + 0.01214465033262968, + 0.006821758113801479, + 0.011402733623981476, + 0.009051558561623096, + 0.019352857023477554, + 0.007765724323689938, + 0.013829993084073067, + 0.019872890785336494, + 0.010414098389446735, + 0.009808475151658058, + 0.014539510011672974, + 0.011555374599993229, + 0.012357877567410469, + 0.007936070673167706, + 0.016089988872408867, + 0.0068263397552073, + 0.0030535077676177025, + 0.004121513105928898, + 0.005637319292873144, + 0.0040433392859995365, + 0.0074079460464417934, + 0.004169047344475985, + 0.008135986514389515, + 0.014800566248595715, + 0.010749335400760174, + 0.007433238439261913, + 0.008669758215546608, + 0.007805971894413233, + 0.0102078877389431, + 0.007782995235174894, + 0.015485493466258049, + 0.017478158697485924, + 0.023277660831809044, + 0.015665020793676376, + 0.02004927583038807, + 0.011083732359111309, + 0.009684900753200054, + 0.008539806120097637, + 0.016409510746598244, + 0.006359134800732136, + 0.005560396239161491, + 0.006767595652490854, + 0.004666838329285383, + 0.00480629550293088, + 0.006169623229652643, + 0.0035372471902519464, + 0.00902083795517683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.004531709477305412 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large tortoise resembling the mythical Xuanwu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tortoise's shell made of deep cosmic obsidian with intricate golden patterns resembling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large black snake with amethyst-shimmering scales coiled tightly around the tortoise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a swirling nebula of deep purple, fiery orange, and icy blue gas with blurred star streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009790733456611633 + ], + "opposite_logprob_tokens": [ + -4.634790897369385 + ], + "logprob_reward": -0.009790733456611633, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009790733456611633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14576959609985352, + 0.2022106647491455, + 0.13132774829864502, + 0.08980287611484528, + 0.06579922884702682, + 0.05780711770057678, + 0.13163962960243225, + 0.060290075838565826, + 0.07408404350280762, + 0.041269052773714066 + ], + "credit_map": [ + 0.027551036328077316, + 0.02332584746181965, + 0.016007209196686745, + 0.020762521773576736, + 0.014317830093204975, + 0.013239556923508644, + 0.016983555629849434, + 0.01358203124254942, + 0.027162861078977585, + 0.037517789751291275, + 0.02389369159936905, + 0.02853521704673767, + 0.02277461439371109, + 0.023532913997769356, + 0.020922958850860596, + 0.01787061057984829, + 0.01718352735042572, + 0.017307819798588753, + 0.012547335587441921, + 0.017717771232128143, + 0.013103348203003407, + 0.026428520679473877, + 0.011624939739704132, + 0.015414468944072723, + 0.01403732132166624, + 0.015620003454387188, + 0.007847691886126995, + 0.015033570118248463, + 0.010468341410160065, + 0.011877545155584812, + 0.008118311874568462, + 0.006800089962780476, + 0.009474080055952072, + 0.010283042676746845, + 0.005904131568968296, + 0.011541534215211868, + 0.00722713116556406, + 0.007092950865626335, + 0.006931012496352196, + 0.007345337420701981, + 0.0093202143907547, + 0.006892079021781683, + 0.0072688087821006775, + 0.007467550225555897, + 0.00655771279707551, + 0.007189157884567976, + 0.005467327311635017, + 0.00764426589012146, + 0.01968192681670189, + 0.012243342585861683, + 0.017743676900863647, + 0.02356692962348461, + 0.011913877911865711, + 0.014529550448060036, + 0.012330497615039349, + 0.019629817456007004, + 0.009495362639427185, + 0.004519735928624868, + 0.005570948589593172, + 0.006698377430438995, + 0.0077375248074531555, + 0.007938913069665432, + 0.007477029226720333, + 0.010852179490029812, + 0.01411495078355074, + 0.0055799325928092, + 0.013791118748486042, + 0.005263542290776968, + 0.00787443108856678, + 0.009181465022265911, + 0.005930753890424967, + 0.012347839772701263, + 0.005438587162643671, + 0.0059982286766171455, + 0.007658379152417183, + 0.006572152022272348, + 0.0050102295354008675, + 0.003558218013495207, + 0.003236268414184451, + 0.003796986071392894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.009790733456611633 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large tortoise resembling the mythical Xuanwu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tortoise's shell made of deep cosmic obsidian with intricate golden patterns resembling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large black snake with amethyst-shimmering scales coiled tightly around the tortoise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a swirling nebula of deep purple, fiery orange, and icy blue gas with blurred star streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16031448543071747 + ], + "opposite_logprob_tokens": [ + -1.9103144407272339 + ], + "logprob_reward": -0.16031448543071747, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16031448543071747, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09707972407341003, + 0.5756100416183472, + 0.09953560680150986, + 0.03194804862141609, + 0.03687359765172005, + 0.035702601075172424, + 0.02969192899763584, + 0.02055596560239792, + 0.049397509545087814, + 0.023605002090334892 + ], + "credit_map": [ + 0.014005047269165516, + 0.0088308434933424, + 0.017373526468873024, + 0.014472538605332375, + 0.008326368406414986, + 0.011196908541023731, + 0.011993679217994213, + 0.01088081393390894, + 0.06915590912103653, + 0.06169285997748375, + 0.05135634168982506, + 0.06741567701101303, + 0.07536263763904572, + 0.10976302623748779, + 0.0832744836807251, + 0.057589091360569, + 0.012282568030059338, + 0.013745984062552452, + 0.009476744569838047, + 0.01320614293217659, + 0.010970712639391422, + 0.01685326360166073, + 0.011467347852885723, + 0.011532842181622982, + 0.004678991157561541, + 0.0041855573654174805, + 0.0030592449475079775, + 0.004205173812806606, + 0.003709714859724045, + 0.005039828829467297, + 0.0025698160752654076, + 0.0044997199438512325, + 0.006849413737654686, + 0.0037776348181068897, + 0.003979600500315428, + 0.00432941410690546, + 0.003510932670906186, + 0.006068743299692869, + 0.0038581909611821175, + 0.0044996654614806175, + 0.007584191858768463, + 0.00396064342930913, + 0.003229212947189808, + 0.005605217069387436, + 0.0031453962437808514, + 0.0054138596169650555, + 0.0027375200297683477, + 0.004026560578495264, + 0.0036284178495407104, + 0.0024540910962969065, + 0.0038210172206163406, + 0.0064381384290754795, + 0.00241716462187469, + 0.0035424476955085993, + 0.00298857968300581, + 0.004402074031531811, + 0.0034979309421032667, + 0.001682671019807458, + 0.0025961529463529587, + 0.0020822470542043447, + 0.0024580087047070265, + 0.0030358764342963696, + 0.001980275847017765, + 0.0032228021882474422, + 0.007202720269560814, + 0.0066786883398890495, + 0.01230167131870985, + 0.006867614109069109, + 0.004280797205865383, + 0.00379665894433856, + 0.003236742690205574, + 0.005032617598772049, + 0.0036999296862632036, + 0.001869492931291461, + 0.0046774279326200485, + 0.002517205663025379, + 0.0024749524891376495, + 0.0029315545689314604, + 0.002103998325765133, + 0.003330440493300557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.16031448543071747 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large tortoise resembling the mythical Xuanwu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tortoise's shell made of deep cosmic obsidian with intricate golden patterns resembling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large black snake with amethyst-shimmering scales coiled tightly around the tortoise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a swirling nebula of deep purple, fiery orange, and icy blue gas with blurred star streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001510432455688715 + ], + "opposite_logprob_tokens": [ + -6.5015106201171875 + ], + "logprob_reward": -0.001510432455688715, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001510432455688715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1929018497467041, + 0.17674891650676727, + 0.11346037685871124, + 0.10076197236776352, + 0.056036628782749176, + 0.08320333808660507, + 0.08647222071886063, + 0.08656027168035507, + 0.05482187867164612, + 0.049032628536224365 + ], + "credit_map": [ + 0.03606146201491356, + 0.026754243299365044, + 0.014401447027921677, + 0.025505077093839645, + 0.043756045401096344, + 0.013068688102066517, + 0.018508290871977806, + 0.014846578240394592, + 0.023003188893198967, + 0.024722952395677567, + 0.016709018498659134, + 0.024690426886081696, + 0.03188583627343178, + 0.02590244635939598, + 0.017264634370803833, + 0.012570377439260483, + 0.013445345684885979, + 0.02115514874458313, + 0.010894582606852055, + 0.012688321061432362, + 0.013346116989850998, + 0.016904866322875023, + 0.012237186543643475, + 0.012788808904588223, + 0.01363534014672041, + 0.01650630123913288, + 0.007936579175293446, + 0.013845824636518955, + 0.01512162946164608, + 0.015289573930203915, + 0.007925114594399929, + 0.010501599870622158, + 0.011485371738672256, + 0.00510537251830101, + 0.004448631312698126, + 0.006046989932656288, + 0.006287794094532728, + 0.010792089626193047, + 0.005772930569946766, + 0.00609744293615222, + 0.01679387502372265, + 0.005604460835456848, + 0.010595510713756084, + 0.012252166867256165, + 0.008445600979030132, + 0.012313120998442173, + 0.006650944706052542, + 0.010547655634582043, + 0.014059960842132568, + 0.00920314621180296, + 0.010646562092006207, + 0.015924010425806046, + 0.00784386321902275, + 0.009671332314610481, + 0.008330201730132103, + 0.01079313363879919, + 0.01170246209949255, + 0.008626078255474567, + 0.013718794099986553, + 0.007221114821732044, + 0.00838450901210308, + 0.009890914894640446, + 0.006694707088172436, + 0.020321683958172798, + 0.006979040335863829, + 0.006898177787661552, + 0.009105444885790348, + 0.005580149590969086, + 0.00460964348167181, + 0.007320859003812075, + 0.004681318998336792, + 0.009647242724895477, + 0.005962572526186705, + 0.004796735011041164, + 0.00924853514879942, + 0.004874115809798241, + 0.0040865130722522736, + 0.007769062649458647, + 0.00417490117251873, + 0.008120188489556313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.001510432455688715 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large tortoise resembling the mythical Xuanwu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tortoise's shell made of deep cosmic obsidian with intricate golden patterns resembling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large black snake with amethyst-shimmering scales coiled tightly around the tortoise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a swirling nebula of deep purple, fiery orange, and icy blue gas with blurred star streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761016607284546 + ], + "opposite_logprob_tokens": [ + -0.8261016607284546 + ], + "logprob_reward": -0.5761016607284546, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761016607284546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08811376243829727, + 0.34986990690231323, + 0.055541858077049255, + 0.03594814985990524, + 0.06645885854959488, + 0.13083355128765106, + 0.055585939437150955, + 0.0690988302230835, + 0.08996199071407318, + 0.05858714506030083 + ], + "credit_map": [ + 0.01412980817258358, + 0.011022630147635937, + 0.008866989985108376, + 0.012636703439056873, + 0.01189429685473442, + 0.0073012723587453365, + 0.012093894183635712, + 0.010168167762458324, + 0.04101837798953056, + 0.03702692687511444, + 0.030907614156603813, + 0.045779258012771606, + 0.041845157742500305, + 0.06852687895298004, + 0.04980666935443878, + 0.03495901823043823, + 0.007256876211613417, + 0.007528369780629873, + 0.004146123304963112, + 0.00920600350946188, + 0.005376089364290237, + 0.008180375210940838, + 0.005841230042278767, + 0.008006793446838856, + 0.0060883439145982265, + 0.004113693255931139, + 0.005926688201725483, + 0.0032931307796388865, + 0.004728865809738636, + 0.004926206078380346, + 0.002766582416370511, + 0.004104640334844589, + 0.014167670160531998, + 0.006305539980530739, + 0.006127676926553249, + 0.008245340548455715, + 0.006954718381166458, + 0.010193818248808384, + 0.006239832378923893, + 0.008224266581237316, + 0.024912603199481964, + 0.008675680495798588, + 0.018867628648877144, + 0.013926852494478226, + 0.011236526072025299, + 0.016568180173635483, + 0.011318180710077286, + 0.025327907875180244, + 0.008473639376461506, + 0.004653044044971466, + 0.006105768959969282, + 0.009654724039137363, + 0.004925158806145191, + 0.007935774512588978, + 0.005229358095675707, + 0.008608466945588589, + 0.011431518010795116, + 0.0162097979336977, + 0.006704968865960836, + 0.009957910515367985, + 0.006397991441190243, + 0.005830808542668819, + 0.004585165996104479, + 0.007980668917298317, + 0.013715561479330063, + 0.0063131400384008884, + 0.014811885543167591, + 0.009641882963478565, + 0.007435184437781572, + 0.009921988472342491, + 0.008680962957441807, + 0.019441386684775352, + 0.008458802476525307, + 0.007267232984304428, + 0.006532424129545689, + 0.00905988086014986, + 0.005392852704972029, + 0.007163965608924627, + 0.0047728801146149635, + 0.009939108043909073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 8 + ], + "selected_reward": -0.5761016607284546 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_11/index_tensor([3980, 3980])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_11/index_tensor([3980, 3980])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f4a6f7c3abed21c8d92676667b1211ceb4fd1b32 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_11/index_tensor([3980, 3980])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "With surreal, documentary-like clarity, a static, low-angle close-up held in the foreground captures the resilient Sorolepidium glacier fern. Its dark, almost black, scaly fronds coil tightly, clinging tenaciously to sharp fragments of gray rock in the crisp, cold light of the high-altitude sun. Each scale on the fronds is rendered with microscopic precision, revealing a rough, protective texture, while tiny, hair-like roots visibly grasp the cracks of the icy stone. In the background, vast and imposing, a massive glacier wall fills the frame, its ancient ice a compressed tapestry of white and deep, translucent blue within its cracks. A subtle, almost imperceptible forward creep animates the glacier, and as small chunks of ice and rock are steadily and deliberately ground and pushed forward at the glacier's base, the glacier's immense geological force becomes apparent, creating a stark and powerful contrast between the fern's delicate, motionless stillness and the awe-inspiring, indifferent movement of the immense geological force behind it.", + "seed": 32062, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_11/index_3980_seed_32062_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.18823759565129877, + "vlm_logprob": -0.18823759565129877, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.18823759565129877, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Sorolepidium glacier fern with dark, almost black, scaly fronds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fern shown in a static, low-angle close-up held in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern clinging to sharp fragments of gray rock with visible hair-like roots grasping the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive glacier wall with white and deep translucent blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glacier exhibit a subtle forward creep with small chunks of ice and rock being pushed forward at its base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5846896767616272 + ], + "opposite_logprob_tokens": [ + -0.8346896767616272 + ], + "logprob_reward": -0.5846896767616272, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5846896767616272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13370315730571747, + 0.08680308610200882, + 0.09581884741783142, + 0.06977874785661697, + 0.05508960783481598, + 0.23929902911186218, + 0.024759182706475258, + 0.0171873327344656, + 0.039382487535476685, + 0.2381785809993744 + ], + "credit_map": [ + 0.015165395103394985, + 0.015013102442026138, + 0.01151213701814413, + 0.009930265136063099, + 0.022879676893353462, + 0.036904748529195786, + 0.0098827900364995, + 0.012415025383234024, + 0.010509409941732883, + 0.009101541712880135, + 0.014523496851325035, + 0.010889164172112942, + 0.013251345604658127, + 0.0139564024284482, + 0.007798104081302881, + 0.006773625500500202, + 0.013366201892495155, + 0.006640211679041386, + 0.011049470864236355, + 0.0069690668024122715, + 0.012444812804460526, + 0.017168324440717697, + 0.01197906769812107, + 0.016201695427298546, + 0.007202698849141598, + 0.004923806991428137, + 0.009423649869859219, + 0.010334644466638565, + 0.009979801252484322, + 0.013361833058297634, + 0.007093395106494427, + 0.007458919659256935, + 0.006178572773933411, + 0.005213126074522734, + 0.008631248958408833, + 0.0062417821027338505, + 0.008181532844901085, + 0.009908715263009071, + 0.006760954856872559, + 0.003973673097789288, + 0.02465253695845604, + 0.026314985007047653, + 0.03434443473815918, + 0.026586933061480522, + 0.026501046493649483, + 0.03300820663571358, + 0.02957558073103428, + 0.0383153036236763, + 0.0024832868948578835, + 0.002223186194896698, + 0.0030735479667782784, + 0.003677476430311799, + 0.0024423024151474237, + 0.0030719551723450422, + 0.0041946591809391975, + 0.003592767985537648, + 0.0012062207097187638, + 0.002230350160971284, + 0.002143431454896927, + 0.002470544306561351, + 0.0017407359555363655, + 0.0018819653196260333, + 0.002811665181070566, + 0.0027024184819310904, + 0.0032149844337254763, + 0.00395171158015728, + 0.00415155990049243, + 0.004871216136962175, + 0.004300630651414394, + 0.004320296924561262, + 0.005041901953518391, + 0.009530183859169483, + 0.02509477734565735, + 0.03078552521765232, + 0.029544277116656303, + 0.02354566380381584, + 0.02502298727631569, + 0.045413658022880554, + 0.025206269696354866, + 0.03356543555855751 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.5846896767616272 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Sorolepidium glacier fern with dark, almost black, scaly fronds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fern shown in a static, low-angle close-up held in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern clinging to sharp fragments of gray rock with visible hair-like roots grasping the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive glacier wall with white and deep translucent blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glacier exhibit a subtle forward creep with small chunks of ice and rock being pushed forward at its base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014224211685359478 + ], + "opposite_logprob_tokens": [ + -4.264224052429199 + ], + "logprob_reward": -0.014224211685359478, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014224211685359478, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17751537263393402, + 0.08544639497995377, + 0.18603327870368958, + 0.09312416613101959, + 0.033516187220811844, + 0.17623202502727509, + 0.0284569188952446, + 0.019200464710593224, + 0.028005335479974747, + 0.17246989905834198 + ], + "credit_map": [ + 0.025976281613111496, + 0.018804728984832764, + 0.027797192335128784, + 0.013839228078722954, + 0.026986517012119293, + 0.028696438297629356, + 0.014461807906627655, + 0.020953182131052017, + 0.013509487733244896, + 0.00851798988878727, + 0.010957583785057068, + 0.00781730655580759, + 0.012952599674463272, + 0.015429753810167313, + 0.008174456655979156, + 0.008087224327027798, + 0.020121388137340546, + 0.0146707808598876, + 0.029511457309126854, + 0.02487991936504841, + 0.024005744606256485, + 0.026637010276317596, + 0.018783940002322197, + 0.02742304839193821, + 0.009627350606024265, + 0.006810226943343878, + 0.01345253549516201, + 0.01146882027387619, + 0.014013240113854408, + 0.0198934655636549, + 0.007494133897125721, + 0.010364393703639507, + 0.0028715585358440876, + 0.0025431043468415737, + 0.006602122914046049, + 0.003582824720069766, + 0.005135469138622284, + 0.005983668379485607, + 0.003745214082300663, + 0.003052223939448595, + 0.017887720838189125, + 0.019176730886101723, + 0.026045100763440132, + 0.02005651965737343, + 0.01932733878493309, + 0.024368517100811005, + 0.021321147680282593, + 0.028048936277627945, + 0.002566237235441804, + 0.00343732419423759, + 0.004294306971132755, + 0.004180603194981813, + 0.0029498471412807703, + 0.002747908001765609, + 0.004432744812220335, + 0.003847947111353278, + 0.0018109978409484029, + 0.002065577544271946, + 0.0024858498945832253, + 0.0025464375503361225, + 0.0021660064812749624, + 0.00275879492983222, + 0.002825631992891431, + 0.0025411671958863735, + 0.0024695375468581915, + 0.004279629793018103, + 0.002822969574481249, + 0.003674929030239582, + 0.003983302041888237, + 0.003750562435016036, + 0.0034945642109960318, + 0.003529840614646673, + 0.019766010344028473, + 0.02318563126027584, + 0.023119274526834488, + 0.015488484874367714, + 0.01770579256117344, + 0.033470578491687775, + 0.018360426649451256, + 0.021373696625232697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.014224211685359478 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Sorolepidium glacier fern with dark, almost black, scaly fronds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fern shown in a static, low-angle close-up held in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern clinging to sharp fragments of gray rock with visible hair-like roots grasping the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive glacier wall with white and deep translucent blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glacier exhibit a subtle forward creep with small chunks of ice and rock being pushed forward at its base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026363030076026917 + ], + "opposite_logprob_tokens": [ + -3.6513631343841553 + ], + "logprob_reward": -0.026363030076026917, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026363030076026917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1927431970834732, + 0.07194754481315613, + 0.10540139675140381, + 0.0781901627779007, + 0.051328085362911224, + 0.16158242523670197, + 0.03787502646446228, + 0.018814248964190483, + 0.09135843813419342, + 0.19075949490070343 + ], + "credit_map": [ + 0.03090071864426136, + 0.01892189495265484, + 0.022642143070697784, + 0.011717627756297588, + 0.031142650172114372, + 0.03784053400158882, + 0.017136648297309875, + 0.022440973669290543, + 0.007886909879744053, + 0.007872858084738255, + 0.012713158503174782, + 0.010356973856687546, + 0.009367830120027065, + 0.009725946933031082, + 0.00804806500673294, + 0.005975805688649416, + 0.009789851494133472, + 0.009060407057404518, + 0.016878848895430565, + 0.01225109864026308, + 0.012485229410231113, + 0.012860341928899288, + 0.01550440862774849, + 0.016571197658777237, + 0.008127279579639435, + 0.006940836552530527, + 0.013225104659795761, + 0.008750755339860916, + 0.011743182316422462, + 0.015060282312333584, + 0.00593522097915411, + 0.008407503366470337, + 0.00651213712990284, + 0.005036803428083658, + 0.008034825325012207, + 0.005224044434726238, + 0.007417432963848114, + 0.01003037579357624, + 0.004995347931981087, + 0.004077119752764702, + 0.017298534512519836, + 0.018199261277914047, + 0.024292515590786934, + 0.018267592415213585, + 0.0177596565335989, + 0.021734319627285004, + 0.01862356625497341, + 0.025406965985894203, + 0.00395995145663619, + 0.0035934653133153915, + 0.004388040397316217, + 0.004739397671073675, + 0.0045149801298975945, + 0.006045900750905275, + 0.004907243885099888, + 0.005726046860218048, + 0.0014935730723664165, + 0.0026489843148738146, + 0.0021564147900789976, + 0.0018302984535694122, + 0.00161217141430825, + 0.00245523639023304, + 0.00240154261700809, + 0.004216027446091175, + 0.007780191954225302, + 0.009853191673755646, + 0.008121965453028679, + 0.007857603952288628, + 0.018333060666918755, + 0.014551029540598392, + 0.012734915129840374, + 0.012126478366553783, + 0.02172820083796978, + 0.025518346577882767, + 0.0240873284637928, + 0.017416885122656822, + 0.020296720787882805, + 0.0358852744102478, + 0.021009190008044243, + 0.024817537516355515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.026363030076026917 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Sorolepidium glacier fern with dark, almost black, scaly fronds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fern shown in a static, low-angle close-up held in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern clinging to sharp fragments of gray rock with visible hair-like roots grasping the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive glacier wall with white and deep translucent blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glacier exhibit a subtle forward creep with small chunks of ice and rock being pushed forward at its base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024867109023034573 + ], + "opposite_logprob_tokens": [ + -6.002486705780029 + ], + "logprob_reward": -0.0024867109023034573, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024867109023034573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4013131558895111, + 0.05502380430698395, + 0.08686131238937378, + 0.04442905634641647, + 0.06257425248622894, + 0.1401137411594391, + 0.029040446504950523, + 0.023449290543794632, + 0.030832352116703987, + 0.12636253237724304 + ], + "credit_map": [ + 0.05718458071351051, + 0.036688607186079025, + 0.04485148936510086, + 0.024244800209999084, + 0.06913580000400543, + 0.09408850222826004, + 0.03054523468017578, + 0.04457416757941246, + 0.009143863804638386, + 0.0044369823299348354, + 0.008153073489665985, + 0.004439594689756632, + 0.008430751971900463, + 0.010933468118309975, + 0.003967823460698128, + 0.0055182562209665775, + 0.013072230853140354, + 0.007846349850296974, + 0.01225605420768261, + 0.005561967380344868, + 0.013194790109992027, + 0.01725260354578495, + 0.007048529572784901, + 0.010628807358443737, + 0.004632928408682346, + 0.0035739545710384846, + 0.006108726840466261, + 0.0075173634104430676, + 0.0062031131237745285, + 0.007984927855432034, + 0.003779103048145771, + 0.004628944676369429, + 0.007838012650609016, + 0.005681388080120087, + 0.00747156934812665, + 0.006110177841037512, + 0.009286157786846161, + 0.013830679468810558, + 0.00725262425839901, + 0.00510364631190896, + 0.014845028519630432, + 0.015169029124081135, + 0.02417427860200405, + 0.015265842899680138, + 0.01610965095460415, + 0.022407229989767075, + 0.015172731131315231, + 0.016969969496130943, + 0.0029297461733222008, + 0.002828432247042656, + 0.0033363699913024902, + 0.003447264200076461, + 0.0032234233804047108, + 0.004405257757753134, + 0.0037105828523635864, + 0.005159373395144939, + 0.0020908310543745756, + 0.0024009502958506346, + 0.003749188967049122, + 0.002351816277951002, + 0.003051302395761013, + 0.00497756153345108, + 0.002222072798758745, + 0.0026055718772113323, + 0.00249020685441792, + 0.0033655136357992887, + 0.0025829218793660402, + 0.002799778012558818, + 0.0074052163399755955, + 0.006767424289137125, + 0.0028421550523489714, + 0.0025791404768824577, + 0.014471984468400478, + 0.015508589334785938, + 0.016403622925281525, + 0.012076133862137794, + 0.013838890939950943, + 0.022246796637773514, + 0.014276587404310703, + 0.017539940774440765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.0024867109023034573 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Sorolepidium glacier fern with dark, almost black, scaly fronds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fern shown in a static, low-angle close-up held in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern clinging to sharp fragments of gray rock with visible hair-like roots grasping the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive glacier wall with white and deep translucent blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glacier exhibit a subtle forward creep with small chunks of ice and rock being pushed forward at its base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31342434883117676 + ], + "opposite_logprob_tokens": [ + -1.3134243488311768 + ], + "logprob_reward": -0.31342434883117676, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31342434883117676, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22189544141292572, + 0.06481758505105972, + 0.14989975094795227, + 0.1312931776046753, + 0.11424406617879868, + 0.1020130068063736, + 0.036224715411663055, + 0.03906645253300667, + 0.0806313082575798, + 0.059914588928222656 + ], + "credit_map": [ + 0.023760536685585976, + 0.018695401027798653, + 0.030675148591399193, + 0.01596851460635662, + 0.03552360460162163, + 0.051321689039468765, + 0.01705761067569256, + 0.028892934322357178, + 0.006917349062860012, + 0.010687916539609432, + 0.010655361227691174, + 0.007679949980229139, + 0.00762689346447587, + 0.009177515283226967, + 0.008259650319814682, + 0.003812952898442745, + 0.018564051017165184, + 0.011500435881316662, + 0.019754933193325996, + 0.02166234888136387, + 0.0197610966861248, + 0.0284590944647789, + 0.018185846507549286, + 0.01201193779706955, + 0.018110085278749466, + 0.012036067433655262, + 0.018780425190925598, + 0.013516571372747421, + 0.016944613307714462, + 0.023679733276367188, + 0.012099489569664001, + 0.016126194968819618, + 0.013967407867312431, + 0.008827787823975086, + 0.019183168187737465, + 0.012568364851176739, + 0.016297249123454094, + 0.02189221978187561, + 0.011797224171459675, + 0.009710649959743023, + 0.013362515717744827, + 0.008908585645258427, + 0.0157565176486969, + 0.008973718620836735, + 0.014323252253234386, + 0.020528145134449005, + 0.008865054696798325, + 0.011295212432742119, + 0.0038928634021431208, + 0.0034369355998933315, + 0.004735973663628101, + 0.00415256479755044, + 0.004309764131903648, + 0.006421115715056658, + 0.004389197565615177, + 0.004886297509074211, + 0.003032747656106949, + 0.004933562595397234, + 0.003929801285266876, + 0.0037186264526098967, + 0.0034651157911866903, + 0.0041411519050598145, + 0.005228782072663307, + 0.010616665706038475, + 0.007471188902854919, + 0.007436790969222784, + 0.007537619210779667, + 0.00857472699135542, + 0.012073257938027382, + 0.015241491608321667, + 0.00916128046810627, + 0.013134951703250408, + 0.007060710806399584, + 0.007881851866841316, + 0.008275783620774746, + 0.006124685984104872, + 0.0059782108291983604, + 0.008563786745071411, + 0.007612473331391811, + 0.008417084813117981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.31342434883117676 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "With surreal, documentary-like clarity, a static, low-angle close-up held in the foreground captures the resilient Sorolepidium glacier fern. Its dark, almost black, scaly fronds coil tightly, clinging tenaciously to sharp fragments of gray rock in the crisp, cold light of the high-altitude sun. Each scale on the fronds is rendered with microscopic precision, revealing a rough, protective texture, while tiny, hair-like roots visibly grasp the cracks of the icy stone. In the background, vast and imposing, a massive glacier wall fills the frame, its ancient ice a compressed tapestry of white and deep, translucent blue within its cracks. A subtle, almost imperceptible forward creep animates the glacier, and as small chunks of ice and rock are steadily and deliberately ground and pushed forward at the glacier's base, the glacier's immense geological force becomes apparent, creating a stark and powerful contrast between the fern's delicate, motionless stillness and the awe-inspiring, indifferent movement of the immense geological force behind it.", + "seed": 32063, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_11/index_3980_seed_32063_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.4369738640263677, + "vlm_logprob": -0.4369738640263677, + "vlm_yes_no_margin": 2.125, + "vlm_token_credit": -0.4369738640263677, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Sorolepidium glacier fern with dark, almost black, scaly fronds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fern shown in a static, low-angle close-up held in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern clinging to sharp fragments of gray rock with visible hair-like roots grasping the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive glacier wall with white and deep translucent blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glacier exhibit a subtle forward creep with small chunks of ice and rock being pushed forward at its base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.6026837825775146 + ], + "opposite_logprob_tokens": [ + -0.22768378257751465 + ], + "logprob_reward": -1.6026837825775146, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.6026837825775146, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11603852361440659, + 0.08257146179676056, + 0.03314419090747833, + 0.07716314494609833, + 0.09943889081478119, + 0.04599108546972275, + 0.44959843158721924, + 0.017761962488293648, + 0.029484761878848076, + 0.048807542771101 + ], + "credit_map": [ + 0.012116187252104282, + 0.01643289066851139, + 0.014531344175338745, + 0.01780414581298828, + 0.011701374314725399, + 0.018460996448993683, + 0.011361837387084961, + 0.013629746623337269, + 0.006833676714450121, + 0.004459704272449017, + 0.005921421572566032, + 0.007305362727493048, + 0.0046919239684939384, + 0.024715712293982506, + 0.004822391550987959, + 0.02382127195596695, + 0.003065682016313076, + 0.004763841163367033, + 0.004719225224107504, + 0.004619814455509186, + 0.002832449274137616, + 0.0037148864939808846, + 0.004257130902260542, + 0.005171163007616997, + 0.009479687549173832, + 0.008986135944724083, + 0.009538519196212292, + 0.010637247003614902, + 0.007371705956757069, + 0.01054629497230053, + 0.008683336898684502, + 0.011920222081243992, + 0.009220809675753117, + 0.009822545573115349, + 0.01284308172762394, + 0.013623692095279694, + 0.012979038059711456, + 0.012705509550869465, + 0.015690414234995842, + 0.012553798966109753, + 0.003726612776517868, + 0.0042456965893507, + 0.0037686973810195923, + 0.005383728072047234, + 0.008660433813929558, + 0.008046700619161129, + 0.0063651129603385925, + 0.005794108845293522, + 0.041534289717674255, + 0.05961013212800026, + 0.04197932779788971, + 0.07668015360832214, + 0.0555574931204319, + 0.045341480523347855, + 0.0726853460073471, + 0.05621020495891571, + 0.0031400874722748995, + 0.0011351702269166708, + 0.002250346587970853, + 0.0017431575106456876, + 0.001757679507136345, + 0.0024328904692083597, + 0.0024850841145962477, + 0.002817547181621194, + 0.004176258575171232, + 0.0028996048495173454, + 0.003971696365624666, + 0.0035393647849559784, + 0.0036312839947640896, + 0.0037714738864451647, + 0.0032423667144030333, + 0.004252712707966566, + 0.00533196609467268, + 0.0046188971027731895, + 0.00752851041033864, + 0.005942909047007561, + 0.006615841761231422, + 0.00468098372220993, + 0.007139022462069988, + 0.006949411705136299 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -1.6026837825775146 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Sorolepidium glacier fern with dark, almost black, scaly fronds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fern shown in a static, low-angle close-up held in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern clinging to sharp fragments of gray rock with visible hair-like roots grasping the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive glacier wall with white and deep translucent blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glacier exhibit a subtle forward creep with small chunks of ice and rock being pushed forward at its base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0337466225028038 + ], + "opposite_logprob_tokens": [ + -3.4087467193603516 + ], + "logprob_reward": -0.0337466225028038, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0337466225028038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1175675168633461, + 0.08206728845834732, + 0.05245542898774147, + 0.11345972120761871, + 0.07303544878959656, + 0.061502642929553986, + 0.40391093492507935, + 0.032554417848587036, + 0.02633264847099781, + 0.037114039063453674 + ], + "credit_map": [ + 0.012471317313611507, + 0.015037273056805134, + 0.013476424850523472, + 0.01680072583258152, + 0.01129116304218769, + 0.02224710024893284, + 0.011563313193619251, + 0.01468019187450409, + 0.00844008382409811, + 0.005833142902702093, + 0.007527453359216452, + 0.008541904389858246, + 0.00621079932898283, + 0.019793622195720673, + 0.005348889622837305, + 0.02037138305604458, + 0.007923632860183716, + 0.0049203126691281796, + 0.004657791927456856, + 0.008180839940905571, + 0.003890405409038067, + 0.006991899572312832, + 0.006725047715008259, + 0.009165490977466106, + 0.013517423532903194, + 0.013148069381713867, + 0.013359280303120613, + 0.01531840581446886, + 0.011702570132911205, + 0.01425519585609436, + 0.013014854863286018, + 0.019143911078572273, + 0.006870936136692762, + 0.007286897394806147, + 0.009164762683212757, + 0.010016928426921368, + 0.009730220772325993, + 0.009282596409320831, + 0.011451183818280697, + 0.009231919422745705, + 0.005470005329698324, + 0.00518460851162672, + 0.006875205785036087, + 0.007400975096970797, + 0.010645615868270397, + 0.010555220767855644, + 0.0081363245844841, + 0.007234685122966766, + 0.037761494517326355, + 0.05191553384065628, + 0.039369627833366394, + 0.06802665442228317, + 0.04941176623106003, + 0.040186282247304916, + 0.0668296292424202, + 0.050409894436597824, + 0.006595767103135586, + 0.002077613491564989, + 0.004625534173101187, + 0.002962246537208557, + 0.002846464281901717, + 0.004622329957783222, + 0.0029933154582977295, + 0.0058311475440859795, + 0.005557849537581205, + 0.00232313247397542, + 0.0030911557842046022, + 0.0021872185170650482, + 0.003140711225569248, + 0.0037368619814515114, + 0.0030744001269340515, + 0.0032213155645877123, + 0.003990532830357552, + 0.004021608270704746, + 0.005987502168864012, + 0.0047279540449380875, + 0.004717949777841568, + 0.003478456987068057, + 0.00476852199062705, + 0.005421509966254234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.0337466225028038 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Sorolepidium glacier fern with dark, almost black, scaly fronds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fern shown in a static, low-angle close-up held in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern clinging to sharp fragments of gray rock with visible hair-like roots grasping the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive glacier wall with white and deep translucent blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glacier exhibit a subtle forward creep with small chunks of ice and rock being pushed forward at its base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005660235881805 + ], + "opposite_logprob_tokens": [ + -2.695056676864624 + ], + "logprob_reward": -0.07005660235881805, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005660235881805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0942406877875328, + 0.16084657609462738, + 0.05937087535858154, + 0.1242118775844574, + 0.026101233437657356, + 0.0856383666396141, + 0.3796934187412262, + 0.024753093719482422, + 0.0128232566639781, + 0.03232058882713318 + ], + "credit_map": [ + 0.00985212903469801, + 0.008134476840496063, + 0.008453317917883396, + 0.00955015979707241, + 0.02109149657189846, + 0.019896125420928, + 0.00684936111792922, + 0.010413626208901405, + 0.013101154007017612, + 0.009190990589559078, + 0.01107919029891491, + 0.011660444550216198, + 0.007864264771342278, + 0.051190782338380814, + 0.007155216298997402, + 0.04960453510284424, + 0.006509768310934305, + 0.008699861355125904, + 0.007853344082832336, + 0.008101217448711395, + 0.005089037120342255, + 0.007416741922497749, + 0.006163308396935463, + 0.009537591598927975, + 0.013370734639465809, + 0.013780017383396626, + 0.01470057014375925, + 0.01632288098335266, + 0.01373254880309105, + 0.014575609005987644, + 0.013713058084249496, + 0.02401646040380001, + 0.00348080531693995, + 0.002630603965371847, + 0.0029358090832829475, + 0.002727226121351123, + 0.0028037717565894127, + 0.0032529677264392376, + 0.0033457796089351177, + 0.004924269393086433, + 0.00980356801301241, + 0.006366632413119078, + 0.01708373613655567, + 0.008736168034374714, + 0.009118998423218727, + 0.012237363494932652, + 0.0071760485880076885, + 0.01511585433036089, + 0.034499749541282654, + 0.0520036518573761, + 0.035065438598394394, + 0.06595176458358765, + 0.04710531607270241, + 0.03978634253144264, + 0.0592605285346508, + 0.046020641922950745, + 0.004316776990890503, + 0.0018098660511896014, + 0.0038039744831621647, + 0.002316102385520935, + 0.0023850242141634226, + 0.0032844541128724813, + 0.00243445485830307, + 0.004402440972626209, + 0.0022340642753988504, + 0.0012150024995207787, + 0.0016566606936976314, + 0.0014133746735751629, + 0.0014055602950975299, + 0.0017291916301473975, + 0.0013194206403568387, + 0.0018499837024137378, + 0.003035465255379677, + 0.0034317723475396633, + 0.004493312444537878, + 0.003967261407524347, + 0.004112051799893379, + 0.003226981032639742, + 0.005053086206316948, + 0.0050006587989628315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 3 + ], + "selected_reward": -0.07005660235881805 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Sorolepidium glacier fern with dark, almost black, scaly fronds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fern shown in a static, low-angle close-up held in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern clinging to sharp fragments of gray rock with visible hair-like roots grasping the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive glacier wall with white and deep translucent blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glacier exhibit a subtle forward creep with small chunks of ice and rock being pushed forward at its base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004091345705091953 + ], + "opposite_logprob_tokens": [ + -5.504091262817383 + ], + "logprob_reward": -0.004091345705091953, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004091345705091953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18159665167331696, + 0.06899621337652206, + 0.035624291747808456, + 0.04951469600200653, + 0.05928546562790871, + 0.03603798523545265, + 0.44782277941703796, + 0.03590977191925049, + 0.014582761563360691, + 0.07062935084104538 + ], + "credit_map": [ + 0.01604677177965641, + 0.0194070003926754, + 0.019736357033252716, + 0.020199909806251526, + 0.02257096953690052, + 0.02934201806783676, + 0.013933081179857254, + 0.04036055877804756, + 0.011188991367816925, + 0.005584868602454662, + 0.00811503455042839, + 0.005781063809990883, + 0.006009240634739399, + 0.013501230627298355, + 0.006352358032017946, + 0.012463427148759365, + 0.00449901819229126, + 0.003813663963228464, + 0.0030328091233968735, + 0.006787812802940607, + 0.002448332030326128, + 0.004412203095853329, + 0.0052248151041567326, + 0.005405638832598925, + 0.005184451583772898, + 0.006349816452711821, + 0.007512067444622517, + 0.006742624565958977, + 0.005029201507568359, + 0.006086424458771944, + 0.004768680315464735, + 0.00784143153578043, + 0.006940463557839394, + 0.005629098974168301, + 0.006513429805636406, + 0.007728820201009512, + 0.006408803164958954, + 0.007907181046903133, + 0.00890650786459446, + 0.009251164272427559, + 0.004096169024705887, + 0.0035688895732164383, + 0.008946135640144348, + 0.0034223173279315233, + 0.002999793039634824, + 0.0041273292154073715, + 0.003266462590545416, + 0.0056108892895281315, + 0.04172388091683388, + 0.05944649502635002, + 0.042173199355602264, + 0.07603216171264648, + 0.0548911951482296, + 0.04622567817568779, + 0.07159397751092911, + 0.05573620647192001, + 0.007276269607245922, + 0.00212425971403718, + 0.005097307730466127, + 0.0040601217187941074, + 0.0032142261043190956, + 0.004776641260832548, + 0.0034967323299497366, + 0.005864213686436415, + 0.0016643328126519918, + 0.0012175820302218199, + 0.001710090902633965, + 0.0019278699764981866, + 0.001956009306013584, + 0.0018838901305571198, + 0.001708940602838993, + 0.00251404638402164, + 0.013755560852587223, + 0.005547427572309971, + 0.009691755287349224, + 0.00760389631614089, + 0.007327554747462273, + 0.00951377023011446, + 0.008274470455944538, + 0.008914914913475513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.004091345705091953 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Sorolepidium glacier fern with dark, almost black, scaly fronds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fern shown in a static, low-angle close-up held in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern clinging to sharp fragments of gray rock with visible hair-like roots grasping the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive glacier wall with white and deep translucent blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glacier exhibit a subtle forward creep with small chunks of ice and rock being pushed forward at its base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47429096698760986 + ], + "opposite_logprob_tokens": [ + -0.9742909669876099 + ], + "logprob_reward": -0.47429096698760986, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47429096698760986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1038670539855957, + 0.3534896671772003, + 0.02776157669723034, + 0.03901161998510361, + 0.05335674434900284, + 0.03431041166186333, + 0.22489671409130096, + 0.049788568168878555, + 0.056757569313049316, + 0.056760162115097046 + ], + "credit_map": [ + 0.012106143869459629, + 0.007291368208825588, + 0.007098458707332611, + 0.006785204168409109, + 0.02490737847983837, + 0.018976809456944466, + 0.006313217803835869, + 0.020388463512063026, + 0.023327169939875603, + 0.008517441339790821, + 0.009355809539556503, + 0.024881480261683464, + 0.01115875318646431, + 0.1303766816854477, + 0.01051961537450552, + 0.13535268604755402, + 0.002973817987367511, + 0.00315264449454844, + 0.0031318662222474813, + 0.0034331774804741144, + 0.0028554187156260014, + 0.003671449376270175, + 0.002883338835090399, + 0.005659859627485275, + 0.006474373862147331, + 0.0033553882967680693, + 0.004361168947070837, + 0.004071839153766632, + 0.004173685796558857, + 0.004986822605133057, + 0.003674285253509879, + 0.007914053276181221, + 0.0042842659167945385, + 0.005573438014835119, + 0.0067093889228999615, + 0.007477232255041599, + 0.008496282622218132, + 0.006589195225387812, + 0.007938615046441555, + 0.006288321688771248, + 0.00437958212569356, + 0.002830451587215066, + 0.009636398404836655, + 0.0038019935600459576, + 0.0031792065128684044, + 0.004186197184026241, + 0.0032178331166505814, + 0.00307874521240592, + 0.019905410706996918, + 0.03162035718560219, + 0.022021183744072914, + 0.038911595940589905, + 0.02817855216562748, + 0.02312387153506279, + 0.03647062927484512, + 0.024665100499987602, + 0.005848389118909836, + 0.004163818433880806, + 0.007425478659570217, + 0.007163211237639189, + 0.005359342787414789, + 0.005166810937225819, + 0.006891640834510326, + 0.0077698687091469765, + 0.012374063022434711, + 0.0048142001032829285, + 0.007544018793851137, + 0.004701557103544474, + 0.005338672082871199, + 0.007864509709179401, + 0.005795816890895367, + 0.00832472462207079, + 0.008685605600476265, + 0.004456628579646349, + 0.006013702601194382, + 0.009826211258769035, + 0.0044898721389472485, + 0.006845084484666586, + 0.008391142822802067, + 0.00805190671235323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.47429096698760986 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "With surreal, documentary-like clarity, a static, low-angle close-up held in the foreground captures the resilient Sorolepidium glacier fern. Its dark, almost black, scaly fronds coil tightly, clinging tenaciously to sharp fragments of gray rock in the crisp, cold light of the high-altitude sun. Each scale on the fronds is rendered with microscopic precision, revealing a rough, protective texture, while tiny, hair-like roots visibly grasp the cracks of the icy stone. In the background, vast and imposing, a massive glacier wall fills the frame, its ancient ice a compressed tapestry of white and deep, translucent blue within its cracks. A subtle, almost imperceptible forward creep animates the glacier, and as small chunks of ice and rock are steadily and deliberately ground and pushed forward at the glacier's base, the glacier's immense geological force becomes apparent, creating a stark and powerful contrast between the fern's delicate, motionless stillness and the awe-inspiring, indifferent movement of the immense geological force behind it.", + "seed": 32064, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_11/index_3980_seed_32064_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09212671942077577, + "vlm_logprob": -0.09212671942077577, + "vlm_yes_no_margin": 3.900000011920929, + "vlm_token_credit": -0.09212671942077577, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Sorolepidium glacier fern with dark, almost black, scaly fronds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fern shown in a static, low-angle close-up held in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern clinging to sharp fragments of gray rock with visible hair-like roots grasping the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive glacier wall with white and deep translucent blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glacier exhibit a subtle forward creep with small chunks of ice and rock being pushed forward at its base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04588901624083519 + ], + "opposite_logprob_tokens": [ + -3.170888900756836 + ], + "logprob_reward": -0.04588901624083519, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04588901624083519, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11794612556695938, + 0.06001952663064003, + 0.03718818724155426, + 0.027735276147723198, + 0.07051602751016617, + 0.07309897243976593, + 0.13829950988292694, + 0.16199716925621033, + 0.13720262050628662, + 0.1759965717792511 + ], + "credit_map": [ + 0.014887510798871517, + 0.010620697401463985, + 0.018394188955426216, + 0.015062456950545311, + 0.01388515718281269, + 0.014691960997879505, + 0.017624642699956894, + 0.012779506854712963, + 0.008476775139570236, + 0.004890763200819492, + 0.007900560274720192, + 0.006944379769265652, + 0.005591332446783781, + 0.008351067081093788, + 0.0062663499265909195, + 0.011598298326134682, + 0.003302702447399497, + 0.0038777031004428864, + 0.006308209616690874, + 0.007264961488544941, + 0.003789891256019473, + 0.004153101705014706, + 0.004728284664452076, + 0.0037633334286510944, + 0.0038874433375895023, + 0.0032744337804615498, + 0.0038970704190433025, + 0.002701230114325881, + 0.0034102986101061106, + 0.004100239370018244, + 0.002513038692995906, + 0.003951523452997208, + 0.010406291112303734, + 0.008499938063323498, + 0.010146289132535458, + 0.008632426150143147, + 0.008098219521343708, + 0.008696001023054123, + 0.006876010447740555, + 0.00916085485368967, + 0.007941084913909435, + 0.00960626918822527, + 0.012292886152863503, + 0.009705541655421257, + 0.00876761693507433, + 0.00755134504288435, + 0.009172522462904453, + 0.008061708882451057, + 0.017231211066246033, + 0.009047472849488258, + 0.009750171564519405, + 0.011328550986945629, + 0.036387745290994644, + 0.03894002363085747, + 0.005331599619239569, + 0.01028272695839405, + 0.011789442040026188, + 0.012967973947525024, + 0.010739204473793507, + 0.012312954291701317, + 0.048063378781080246, + 0.0504554845392704, + 0.006034285761415958, + 0.009634441696107388, + 0.017170891165733337, + 0.01661655306816101, + 0.017074689269065857, + 0.0155112873762846, + 0.016664724797010422, + 0.01849452592432499, + 0.01153709925711155, + 0.024132857099175453, + 0.026500286534428596, + 0.024139633402228355, + 0.012841636314988136, + 0.011636282317340374, + 0.023864561691880226, + 0.03156888112425804, + 0.011140481568872929, + 0.03430480509996414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.04588901624083519 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Sorolepidium glacier fern with dark, almost black, scaly fronds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fern shown in a static, low-angle close-up held in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern clinging to sharp fragments of gray rock with visible hair-like roots grasping the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive glacier wall with white and deep translucent blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glacier exhibit a subtle forward creep with small chunks of ice and rock being pushed forward at its base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014217747375369072 + ], + "opposite_logprob_tokens": [ + -4.264217853546143 + ], + "logprob_reward": -0.014217747375369072, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014217747375369072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04791758209466934, + 0.02174159325659275, + 0.01859455369412899, + 0.01877826824784279, + 0.013272827491164207, + 0.030962809920310974, + 0.11088671535253525, + 0.6394850611686707, + 0.037582989782094955, + 0.06077757850289345 + ], + "credit_map": [ + 0.007032679859548807, + 0.0032511402387171984, + 0.006261710077524185, + 0.005262632388621569, + 0.005821633618324995, + 0.008006902411580086, + 0.005524362437427044, + 0.006756526883691549, + 0.0017097785603255033, + 0.001499795587733388, + 0.0039273882284760475, + 0.002735479036346078, + 0.0020062867552042007, + 0.002775142202153802, + 0.002854186575859785, + 0.004233537707477808, + 0.0020042036194354296, + 0.0018816195661202073, + 0.0028255744837224483, + 0.0023060739040374756, + 0.0018025196623057127, + 0.003012734930962324, + 0.001853979891166091, + 0.0029078491497784853, + 0.002732552122324705, + 0.0027022049762308598, + 0.0026586148887872696, + 0.0020856002811342478, + 0.002123685320839286, + 0.0023477121721953154, + 0.0019989032298326492, + 0.002128995954990387, + 0.0015873312950134277, + 0.001117441337555647, + 0.0018402013229206204, + 0.0013855553697794676, + 0.0013552741147577763, + 0.002271524164825678, + 0.0012111879186704755, + 0.002504314063116908, + 0.0030734024476259947, + 0.003947633318603039, + 0.003183981403708458, + 0.003505644155666232, + 0.004557370208203793, + 0.004354099277406931, + 0.0039961026050150394, + 0.004344582092016935, + 0.00523015670478344, + 0.006452555768191814, + 0.005125387571752071, + 0.004398298915475607, + 0.04173525422811508, + 0.03575940057635307, + 0.0035007107071578503, + 0.008684962056577206, + 0.04550400376319885, + 0.0357159748673439, + 0.09682637453079224, + 0.05225361883640289, + 0.15220020711421967, + 0.21844284236431122, + 0.01367848739027977, + 0.024863576516509056, + 0.003926623612642288, + 0.004864671733230352, + 0.004544395487755537, + 0.004883123561739922, + 0.004072657786309719, + 0.00725041376426816, + 0.0031264671124517918, + 0.004914641845971346, + 0.007110076490789652, + 0.006400228012353182, + 0.006327860057353973, + 0.006668825633823872, + 0.007621635217219591, + 0.008702022023499012, + 0.007898906245827675, + 0.010048029012978077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.014217747375369072 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Sorolepidium glacier fern with dark, almost black, scaly fronds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fern shown in a static, low-angle close-up held in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern clinging to sharp fragments of gray rock with visible hair-like roots grasping the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive glacier wall with white and deep translucent blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glacier exhibit a subtle forward creep with small chunks of ice and rock being pushed forward at its base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012549715116620064 + ], + "opposite_logprob_tokens": [ + -4.387549877166748 + ], + "logprob_reward": -0.012549715116620064, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012549715116620064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12091169506311417, + 0.081868015229702, + 0.04301641881465912, + 0.020651234313845634, + 0.04102079197764397, + 0.05558587610721588, + 0.16620349884033203, + 0.1624692976474762, + 0.12688934803009033, + 0.1813838630914688 + ], + "credit_map": [ + 0.013318322598934174, + 0.008403502404689789, + 0.019327064976096153, + 0.013651713728904724, + 0.014259985648095608, + 0.01848169043660164, + 0.012674679048359394, + 0.02079472318291664, + 0.005797096528112888, + 0.005555412732064724, + 0.013569273985922337, + 0.008695614524185658, + 0.008772020228207111, + 0.013658088631927967, + 0.00844644382596016, + 0.017374062910676003, + 0.004057526588439941, + 0.003911672160029411, + 0.005686921998858452, + 0.0066180783323943615, + 0.0036322057712823153, + 0.006622149609029293, + 0.0044147358275949955, + 0.008073125965893269, + 0.0013549855211749673, + 0.0021218296606093645, + 0.0029804869554936886, + 0.0019194349879398942, + 0.002706760074943304, + 0.0035984283313155174, + 0.0016283126315101981, + 0.004340995103120804, + 0.004620621912181377, + 0.0031692979391664267, + 0.00574828777462244, + 0.00446762889623642, + 0.004634582903236151, + 0.006206890102475882, + 0.004565624985843897, + 0.007607852574437857, + 0.006615615915507078, + 0.005737680941820145, + 0.008772407658398151, + 0.00629194313660264, + 0.007586618419736624, + 0.006730417720973492, + 0.006042385008186102, + 0.007808805909007788, + 0.014392583630979061, + 0.00941284280270338, + 0.008755745366215706, + 0.011028924025595188, + 0.05159115418791771, + 0.049408067017793655, + 0.006771798711270094, + 0.014842375181615353, + 0.011397953145205975, + 0.010729326866567135, + 0.018030069768428802, + 0.010580167174339294, + 0.039560914039611816, + 0.0501580648124218, + 0.013212849386036396, + 0.008799947798252106, + 0.012943423353135586, + 0.014940939843654633, + 0.014585829339921474, + 0.010508378967642784, + 0.013374559581279755, + 0.019878311082720757, + 0.008609101176261902, + 0.03204880654811859, + 0.017814340069890022, + 0.019121944904327393, + 0.028181200847029686, + 0.019339682534337044, + 0.020944884046912193, + 0.028731079772114754, + 0.014245318248867989, + 0.03300539404153824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.012549715116620064 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Sorolepidium glacier fern with dark, almost black, scaly fronds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fern shown in a static, low-angle close-up held in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern clinging to sharp fragments of gray rock with visible hair-like roots grasping the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive glacier wall with white and deep translucent blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glacier exhibit a subtle forward creep with small chunks of ice and rock being pushed forward at its base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000918681500479579 + ], + "opposite_logprob_tokens": [ + -7.000918865203857 + ], + "logprob_reward": -0.000918681500479579, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.000918681500479579, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07498060166835785, + 0.06395052373409271, + 0.036558426916599274, + 0.034052468836307526, + 0.05087723955512047, + 0.06532066315412521, + 0.05690307170152664, + 0.3676712214946747, + 0.1115773618221283, + 0.13810837268829346 + ], + "credit_map": [ + 0.007147088646888733, + 0.005553026683628559, + 0.010216382332146168, + 0.007877904921770096, + 0.00966668501496315, + 0.01331424992531538, + 0.00887498538941145, + 0.012330276891589165, + 0.00536584435030818, + 0.006263346876949072, + 0.009583460167050362, + 0.007205497939139605, + 0.0074552251026034355, + 0.009585878811776638, + 0.007390142418444157, + 0.011101128533482552, + 0.0052293697372078896, + 0.0028608920983970165, + 0.0048500788398087025, + 0.004106923937797546, + 0.003666080767288804, + 0.005830710753798485, + 0.0036606681533157825, + 0.006353703327476978, + 0.002401233883574605, + 0.002514852676540613, + 0.006164184305816889, + 0.003990476951003075, + 0.0031939828768372536, + 0.005510917864739895, + 0.003904595272615552, + 0.006372223608195782, + 0.005036108661442995, + 0.006092770490795374, + 0.005748525727540255, + 0.004183449782431126, + 0.006750747561454773, + 0.008338332176208496, + 0.005190487951040268, + 0.009536818601191044, + 0.007417827844619751, + 0.009006318636238575, + 0.009315709583461285, + 0.007440940011292696, + 0.007713641971349716, + 0.007966139353811741, + 0.0076131741516292095, + 0.008846906945109367, + 0.01005240622907877, + 0.005349511280655861, + 0.0061510708183050156, + 0.007545751519501209, + 0.005706764291971922, + 0.009613282047212124, + 0.003538739401847124, + 0.008945546112954617, + 0.028120234608650208, + 0.021637627854943275, + 0.052351538091897964, + 0.030426625162363052, + 0.08265241980552673, + 0.11914098262786865, + 0.011861489154398441, + 0.02148030512034893, + 0.010658804327249527, + 0.012695803306996822, + 0.012209607288241386, + 0.00889741349965334, + 0.010894114151597023, + 0.019420523196458817, + 0.006985609419643879, + 0.029815485700964928, + 0.022963156923651695, + 0.017311882227659225, + 0.010706806555390358, + 0.009282213635742664, + 0.025963759049773216, + 0.024994676932692528, + 0.008126477710902691, + 0.018759410828351974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.000918681500479579 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Sorolepidium glacier fern with dark, almost black, scaly fronds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fern shown in a static, low-angle close-up held in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern clinging to sharp fragments of gray rock with visible hair-like roots grasping the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive glacier wall with white and deep translucent blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glacier exhibit a subtle forward creep with small chunks of ice and rock being pushed forward at its base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38705843687057495 + ], + "opposite_logprob_tokens": [ + -1.1370584964752197 + ], + "logprob_reward": -0.38705843687057495, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.38705843687057495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11041679233312607, + 0.038795262575149536, + 0.03415326774120331, + 0.04254567623138428, + 0.03191021829843521, + 0.0430513359606266, + 0.23103472590446472, + 0.23367905616760254, + 0.09147096425294876, + 0.14294274151325226 + ], + "credit_map": [ + 0.011265807785093784, + 0.006850054487586021, + 0.015015204437077045, + 0.010799022391438484, + 0.013597971759736538, + 0.02279208041727543, + 0.0097047695890069, + 0.020391885191202164, + 0.005032380577176809, + 0.0030761791858822107, + 0.005373826250433922, + 0.004088941961526871, + 0.004027077462524176, + 0.0066819521598517895, + 0.003571653738617897, + 0.006943251937627792, + 0.0036013044882565737, + 0.002581845037639141, + 0.005989943165332079, + 0.0034657663200050592, + 0.0037308658938854933, + 0.005622610449790955, + 0.0026729698292911053, + 0.006487964652478695, + 0.0029197391122579575, + 0.0030989476945251226, + 0.006745289079844952, + 0.004551065154373646, + 0.004120418336242437, + 0.006646191701292992, + 0.002957239979878068, + 0.011506783775985241, + 0.002400157041847706, + 0.0035934364423155785, + 0.003994753584265709, + 0.003907278645783663, + 0.004131074994802475, + 0.0046177455224096775, + 0.004203656688332558, + 0.00506211631000042, + 0.00258022453635931, + 0.003309514606371522, + 0.005824776366353035, + 0.0038037714548408985, + 0.005587996449321508, + 0.008658822625875473, + 0.003993029240518808, + 0.009293201379477978, + 0.01909475028514862, + 0.013448563404381275, + 0.011567777954041958, + 0.014191001653671265, + 0.075394406914711, + 0.07275678217411041, + 0.007816042751073837, + 0.016765402629971504, + 0.022577255964279175, + 0.019797323271632195, + 0.044844795018434525, + 0.027369752526283264, + 0.027081402018666267, + 0.060520414263010025, + 0.010515504516661167, + 0.020972607657313347, + 0.01103468332439661, + 0.011495373211801052, + 0.009979944676160812, + 0.007525577209889889, + 0.01365551445633173, + 0.010881959460675716, + 0.007785872556269169, + 0.019112039357423782, + 0.0164735596626997, + 0.018136441707611084, + 0.014462091028690338, + 0.011419227346777916, + 0.016161058098077774, + 0.027266014367341995, + 0.010008569806814194, + 0.02901577204465866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.38705843687057495 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "With surreal, documentary-like clarity, a static, low-angle close-up held in the foreground captures the resilient Sorolepidium glacier fern. Its dark, almost black, scaly fronds coil tightly, clinging tenaciously to sharp fragments of gray rock in the crisp, cold light of the high-altitude sun. Each scale on the fronds is rendered with microscopic precision, revealing a rough, protective texture, while tiny, hair-like roots visibly grasp the cracks of the icy stone. In the background, vast and imposing, a massive glacier wall fills the frame, its ancient ice a compressed tapestry of white and deep, translucent blue within its cracks. A subtle, almost imperceptible forward creep animates the glacier, and as small chunks of ice and rock are steadily and deliberately ground and pushed forward at the glacier's base, the glacier's immense geological force becomes apparent, creating a stark and powerful contrast between the fern's delicate, motionless stillness and the awe-inspiring, indifferent movement of the immense geological force behind it.", + "seed": 32065, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_11/index_3980_seed_32065_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.20480516809038818, + "vlm_logprob": -0.20480516809038818, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.20480516809038818, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Sorolepidium glacier fern with dark, almost black, scaly fronds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fern shown in a static, low-angle close-up held in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern clinging to sharp fragments of gray rock with visible hair-like roots grasping the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive glacier wall with white and deep translucent blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glacier exhibit a subtle forward creep with small chunks of ice and rock being pushed forward at its base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.35252341628074646 + ], + "opposite_logprob_tokens": [ + -1.2275234460830688 + ], + "logprob_reward": -0.35252341628074646, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.35252341628074646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08621715009212494, + 0.054198455065488815, + 0.20796777307987213, + 0.13523323833942413, + 0.18328875303268433, + 0.07263597100973129, + 0.0330418162047863, + 0.11098767817020416, + 0.058789242058992386, + 0.05763988569378853 + ], + "credit_map": [ + 0.01214208547025919, + 0.015142450109124184, + 0.008386064320802689, + 0.006216662470251322, + 0.02365606278181076, + 0.009470243938267231, + 0.005014191847294569, + 0.006189392413944006, + 0.006925026886165142, + 0.010683099739253521, + 0.004543669521808624, + 0.004803946707397699, + 0.013866623863577843, + 0.004996946081519127, + 0.004654822871088982, + 0.00372431892901659, + 0.040368419140577316, + 0.05951518937945366, + 0.006713994313031435, + 0.01023103017359972, + 0.06433431059122086, + 0.01385799515992403, + 0.008035961538553238, + 0.004910873249173164, + 0.0314008854329586, + 0.04491276293992996, + 0.0036297764163464308, + 0.0030273518059402704, + 0.036593060940504074, + 0.009531006217002869, + 0.0037009590305387974, + 0.0024374297354370356, + 0.06409037113189697, + 0.07153254002332687, + 0.007429522927850485, + 0.0038724360056221485, + 0.018049556761980057, + 0.0066554127261042595, + 0.006176282186061144, + 0.005482621490955353, + 0.010029944591224194, + 0.011880382895469666, + 0.005071843508630991, + 0.007768542505800724, + 0.017757022753357887, + 0.006350055802613497, + 0.008523860014975071, + 0.0052543212659657, + 0.005484834313392639, + 0.0049809422343969345, + 0.0029662686865776777, + 0.0031239455565810204, + 0.004876439925283194, + 0.003671378130093217, + 0.004399792291224003, + 0.0035382153000682592, + 0.016947440803050995, + 0.029904888942837715, + 0.005472023505717516, + 0.004685395862907171, + 0.03643364459276199, + 0.009073792956769466, + 0.004121249075978994, + 0.004349248483777046, + 0.008934248238801956, + 0.015267045237123966, + 0.0024056618567556143, + 0.0027277173940092325, + 0.019566865637898445, + 0.004917924292385578, + 0.0029928910080343485, + 0.0019768914207816124, + 0.007497664075344801, + 0.008049051277339458, + 0.007680280599743128, + 0.0057793185114860535, + 0.008880894631147385, + 0.005845533683896065, + 0.006182814948260784, + 0.007724322844296694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 3 + ], + "selected_reward": -0.35252341628074646 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Sorolepidium glacier fern with dark, almost black, scaly fronds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fern shown in a static, low-angle close-up held in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern clinging to sharp fragments of gray rock with visible hair-like roots grasping the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive glacier wall with white and deep translucent blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glacier exhibit a subtle forward creep with small chunks of ice and rock being pushed forward at its base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0125654898583889 + ], + "opposite_logprob_tokens": [ + -4.387565612792969 + ], + "logprob_reward": -0.0125654898583889, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.0125654898583889, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14412665367126465, + 0.07333819568157196, + 0.030148399993777275, + 0.07011517137289047, + 0.1731380671262741, + 0.042203497141599655, + 0.04823683202266693, + 0.2200644612312317, + 0.07882031053304672, + 0.11980840563774109 + ], + "credit_map": [ + 0.029471632093191147, + 0.027983132749795914, + 0.013553624972701073, + 0.007794657256454229, + 0.02741166390478611, + 0.020435364916920662, + 0.006212598644196987, + 0.011263980530202389, + 0.010806716047227383, + 0.01691172458231449, + 0.003260683501139283, + 0.003934723790735006, + 0.025831516832113266, + 0.007649040315300226, + 0.0022642002440989017, + 0.0026795919984579086, + 0.0036566846538335085, + 0.0037904689088463783, + 0.004651953931897879, + 0.004530925769358873, + 0.005660704802721739, + 0.0027776912320405245, + 0.0026706974022090435, + 0.002409272128716111, + 0.009139406494796276, + 0.014442089945077896, + 0.006292642094194889, + 0.0061674718745052814, + 0.01921970397233963, + 0.005766409914940596, + 0.005339765455573797, + 0.003747685346752405, + 0.058336012065410614, + 0.06578627973794937, + 0.008417157456278801, + 0.0046574813313782215, + 0.017596229910850525, + 0.007803128566592932, + 0.004989432170987129, + 0.005552351009100676, + 0.006806491874158382, + 0.005811117123812437, + 0.004772982560098171, + 0.004911270923912525, + 0.0069083571434021, + 0.004566410090774298, + 0.004889339674264193, + 0.0035375263541936874, + 0.008032212033867836, + 0.010343768633902073, + 0.003782173153012991, + 0.00375457014888525, + 0.010997001081705093, + 0.004259619861841202, + 0.003575528971850872, + 0.003491952084004879, + 0.03778389096260071, + 0.06506434082984924, + 0.010436196811497211, + 0.010623451322317123, + 0.06186613813042641, + 0.016554366797208786, + 0.009011123329401016, + 0.008724959567189217, + 0.010982982814311981, + 0.015229200944304466, + 0.004100203514099121, + 0.004603929817676544, + 0.02298874780535698, + 0.006642311345785856, + 0.0056810262612998486, + 0.008591908030211926, + 0.015748482197523117, + 0.014552227221429348, + 0.015316320583224297, + 0.012077824212610722, + 0.018702447414398193, + 0.012303516268730164, + 0.013261158019304276, + 0.017846425995230675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.0125654898583889 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Sorolepidium glacier fern with dark, almost black, scaly fronds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fern shown in a static, low-angle close-up held in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern clinging to sharp fragments of gray rock with visible hair-like roots grasping the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive glacier wall with white and deep translucent blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glacier exhibit a subtle forward creep with small chunks of ice and rock being pushed forward at its base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023303797468543053 + ], + "opposite_logprob_tokens": [ + -3.773303747177124 + ], + "logprob_reward": -0.023303797468543053, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023303797468543053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19135712087154388, + 0.054878830909729004, + 0.1621299684047699, + 0.06561574339866638, + 0.2200731784105301, + 0.033171940594911575, + 0.036051858216524124, + 0.059629056602716446, + 0.09478126466274261, + 0.08231101185083389 + ], + "credit_map": [ + 0.030783625319600105, + 0.0353868193924427, + 0.016936801373958588, + 0.011309796944260597, + 0.05354432761669159, + 0.022146761417388916, + 0.008723239414393902, + 0.012525760568678379, + 0.008466294035315514, + 0.011557473801076412, + 0.004164298065006733, + 0.0037778550758957863, + 0.017389247193932533, + 0.0053635952062904835, + 0.0017926843138411641, + 0.002367382636293769, + 0.03262396156787872, + 0.048145223408937454, + 0.005426859483122826, + 0.0062601263634860516, + 0.05003973841667175, + 0.010908534750342369, + 0.00507139740511775, + 0.0036541263107210398, + 0.013747815042734146, + 0.020169002935290337, + 0.002434941940009594, + 0.0018700513755902648, + 0.018375789746642113, + 0.005390113685280085, + 0.0018735213670879602, + 0.0017545059090480208, + 0.08064120262861252, + 0.08880363404750824, + 0.006469959393143654, + 0.0033781148958951235, + 0.025086810812354088, + 0.0067826793529093266, + 0.004677827004343271, + 0.004232941661030054, + 0.004507685545831919, + 0.004164799582213163, + 0.003997566644102335, + 0.004339292645454407, + 0.005084102042019367, + 0.0034237478394061327, + 0.004299698863178492, + 0.003355046734213829, + 0.006031777709722519, + 0.004461252596229315, + 0.0035835665185004473, + 0.0033259717747569084, + 0.006147581152617931, + 0.004361853934824467, + 0.004020194057375193, + 0.004119659308344126, + 0.008100733160972595, + 0.014452029950916767, + 0.004946314264088869, + 0.003026396268978715, + 0.01910271681845188, + 0.004408292938023806, + 0.002230279380455613, + 0.003362296847626567, + 0.015190164558589458, + 0.024700263515114784, + 0.005335135385394096, + 0.009701745584607124, + 0.01738726906478405, + 0.0065171001479029655, + 0.008361803367733955, + 0.007587777450680733, + 0.011092948727309704, + 0.01126371044665575, + 0.00974990427494049, + 0.00763667980208993, + 0.013849940150976181, + 0.00891174003481865, + 0.009112648665904999, + 0.010693436488509178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.023303797468543053 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Sorolepidium glacier fern with dark, almost black, scaly fronds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fern shown in a static, low-angle close-up held in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern clinging to sharp fragments of gray rock with visible hair-like roots grasping the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive glacier wall with white and deep translucent blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glacier exhibit a subtle forward creep with small chunks of ice and rock being pushed forward at its base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028156896587461233 + ], + "opposite_logprob_tokens": [ + -5.8778157234191895 + ], + "logprob_reward": -0.0028156896587461233, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028156896587461233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25239327549934387, + 0.06803886592388153, + 0.019780173897743225, + 0.043526940047740936, + 0.322332501411438, + 0.03286603093147278, + 0.02004668302834034, + 0.07976648211479187, + 0.06836617738008499, + 0.09288276731967926 + ], + "credit_map": [ + 0.04482635110616684, + 0.053275853395462036, + 0.019270138815045357, + 0.01200337428599596, + 0.06936681270599365, + 0.028426088392734528, + 0.009143033996224403, + 0.01608164794743061, + 0.009714130312204361, + 0.015825442969799042, + 0.002744804136455059, + 0.003528362140059471, + 0.02362874522805214, + 0.006842422299087048, + 0.0025152843445539474, + 0.003239678218960762, + 0.0023364743683487177, + 0.0023335793521255255, + 0.0018464430468156934, + 0.003294967347756028, + 0.0033640405163168907, + 0.0016929006669670343, + 0.0034414641559123993, + 0.0014703072374686599, + 0.010411285795271397, + 0.013872071169316769, + 0.00224614841863513, + 0.0020399324130266905, + 0.009110692888498306, + 0.0028052462730556726, + 0.0019547578413039446, + 0.0010868087410926819, + 0.1110924780368805, + 0.12364624440670013, + 0.009440098889172077, + 0.004948049783706665, + 0.04648425802588463, + 0.012521643191576004, + 0.007473960984498262, + 0.006725798361003399, + 0.005337317939847708, + 0.005346075166016817, + 0.0034688774030655622, + 0.00424965750426054, + 0.004311987664550543, + 0.0035147902090102434, + 0.003820634912699461, + 0.0028166910633444786, + 0.003616887144744396, + 0.003498073434457183, + 0.0016614225460216403, + 0.0017586249159649014, + 0.002482139505445957, + 0.0014841676456853747, + 0.0029425350949168205, + 0.002602834952995181, + 0.012979918159544468, + 0.021837415173649788, + 0.006814470048993826, + 0.00501422630622983, + 0.016466006636619568, + 0.0069084917195141315, + 0.004155290778726339, + 0.005590680055320263, + 0.010767491534352303, + 0.017659923061728477, + 0.003329090541228652, + 0.004195240791887045, + 0.01845509000122547, + 0.005871433299034834, + 0.004202581476420164, + 0.0038853308651596308, + 0.016201699152588844, + 0.016330908983945847, + 0.005973236169666052, + 0.0072214058600366116, + 0.02003244310617447, + 0.009681209921836853, + 0.009061086922883987, + 0.008380780927836895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.0028156896587461233 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Sorolepidium glacier fern with dark, almost black, scaly fronds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fern shown in a static, low-angle close-up held in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fern clinging to sharp fragments of gray rock with visible hair-like roots grasping the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive glacier wall with white and deep translucent blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glacier exhibit a subtle forward creep with small chunks of ice and rock being pushed forward at its base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328174471855164 + ], + "opposite_logprob_tokens": [ + -0.7578174471855164 + ], + "logprob_reward": -0.6328174471855164, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328174471855164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1190720647573471, + 0.07219540327787399, + 0.15231303870677948, + 0.07858230918645859, + 0.35157471895217896, + 0.0362839438021183, + 0.02982020378112793, + 0.04511698707938194, + 0.07891328632831573, + 0.036128029227256775 + ], + "credit_map": [ + 0.01753339357674122, + 0.02555709145963192, + 0.006223506759852171, + 0.006170878652483225, + 0.04054781049489975, + 0.011630536057054996, + 0.005622698459774256, + 0.005786164663732052, + 0.01084006018936634, + 0.01512681134045124, + 0.005414779298007488, + 0.005686433985829353, + 0.020464133471250534, + 0.006001603789627552, + 0.004617596976459026, + 0.0040439884178340435, + 0.028375376015901566, + 0.04203547537326813, + 0.005900165997445583, + 0.009453159756958485, + 0.0457705482840538, + 0.010064668022096157, + 0.0065924557857215405, + 0.004121205769479275, + 0.009745423682034016, + 0.017654461786150932, + 0.005161856766790152, + 0.006759474985301495, + 0.0256551131606102, + 0.0065391696989536285, + 0.003973328974097967, + 0.0030934892129153013, + 0.1165640726685524, + 0.13076308369636536, + 0.00807729922235012, + 0.0042330557480454445, + 0.06331838667392731, + 0.015737077221274376, + 0.007287552114576101, + 0.005594208370894194, + 0.005957977380603552, + 0.006926150061190128, + 0.0034389158245176077, + 0.0025694607757031918, + 0.008403574116528034, + 0.003385064657777548, + 0.002753442618995905, + 0.0028493618592619896, + 0.00388390407897532, + 0.0064352466724812984, + 0.0025748307816684246, + 0.0025551908183842897, + 0.0074933432042598724, + 0.00237542693503201, + 0.0020655933767557144, + 0.0024366697762161493, + 0.006245753727853298, + 0.008398220874369144, + 0.0021120538003742695, + 0.0020113785285502672, + 0.014168702997267246, + 0.005660873372107744, + 0.002746781799942255, + 0.0037732271011918783, + 0.01230680476874113, + 0.02008596621453762, + 0.004676234442740679, + 0.006220738869160414, + 0.019682463258504868, + 0.005800348706543446, + 0.005308964289724827, + 0.0048317676410079, + 0.005320901051163673, + 0.0048986212350428104, + 0.0032604995649307966, + 0.0038666187319904566, + 0.00573087390512228, + 0.004549145698547363, + 0.004666437860578299, + 0.003834931179881096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.6328174471855164 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_12/index_tensor([6485, 6485])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_12/index_tensor([6485, 6485])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1985f4c47e804291f5bde1938a13a6eac66c167f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_12/index_tensor([6485, 6485])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An epic wide shot, the camera moving in slow, graceful arcs, captures Nezha, a young man in red, riding on a pair of fire wheels, slowly circling above a boundless, majestic sea of clouds. He wears his signature twin black buns, bound with red ropes. His face is handsome, his brows radiating a determination belying his age and a hint of rebelliousness. He wears a sleeveless, crimson armor fashioned from the Hun Tian Ling, intricately embroidered with golden flames, gleaming with a divine sheen in the light. Steadily in his right hand, he grips a fire-tipped spear taller than himself. The gleaming metallic gleam of the spear's shaft contrasts sharply with the ever-burning, blazing flames of true fire atop its tip. His feet rest firmly on a pair of flaming fire wheels, which rotate with a seemingly slow yet powerful rhythm, leaving a long trail of interwoven golden and red flames behind him. Below him are thick, rolling clouds like cotton, forming continuous mountains in the clouds and bottomless canyons. Brilliant golden sunlight pours down from the clear blue sky, penetrating the gaps in the clouds and casting several magnificent rays of light. It not only outlines a circle of dazzling golden edges for the outline of his red armor, but also makes the whole scene full of a sacred, majestic and tranquil atmosphere, presenting a grand cinematic feeling like an epic oriental mythology.", + "seed": 15758, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_12/index_6485_seed_15758_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.024545255256816743, + "vlm_logprob": -0.024545255256816743, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.024545255256816743, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide shot moving in slow, graceful arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha, a young man with twin black buns bound by red ropes, visible riding fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha wearing sleeveless crimson armor embroidered with golden flames and holding a fire-tipped spear taller than himself in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of clouds with thick rolling formations and brilliant golden sunlight casting rays from a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fire wheels rotating and leaving a long trail of interwoven golden and red flames behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06563979387283325 + ], + "opposite_logprob_tokens": [ + -2.8156397342681885 + ], + "logprob_reward": -0.06563979387283325, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06563979387283325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07984554767608643, + 0.08098320662975311, + 0.058371029794216156, + 0.019099093973636627, + 0.11657846719026566, + 0.4116564393043518, + 0.025475872680544853, + 0.03662732243537903, + 0.12427354604005814, + 0.04708945378661156 + ], + "credit_map": [ + 0.007370676379650831, + 0.006487880367785692, + 0.012452230788767338, + 0.010823243297636509, + 0.008577846921980381, + 0.012166106142103672, + 0.009659023024141788, + 0.012308544479310513, + 0.009462345391511917, + 0.009068907238543034, + 0.010124234482645988, + 0.007473487872630358, + 0.007030025124549866, + 0.009868581779301167, + 0.011088741943240166, + 0.016866886988282204, + 0.008647732436656952, + 0.004861680790781975, + 0.007476259022951126, + 0.005675220862030983, + 0.007393233012408018, + 0.006053389050066471, + 0.007468478288501501, + 0.010795041918754578, + 0.003155405167490244, + 0.0018663934897631407, + 0.0025470934342592955, + 0.002055563498288393, + 0.0019514894811436534, + 0.0024651819840073586, + 0.0022202960681170225, + 0.0028376728296279907, + 0.014883168041706085, + 0.03527795895934105, + 0.031189097091555595, + 0.0054506296291947365, + 0.003908331505954266, + 0.007053028792142868, + 0.00431700237095356, + 0.014499247074127197, + 0.05867718905210495, + 0.11644347757101059, + 0.10234384983778, + 0.01814775913953781, + 0.01373250037431717, + 0.03246397152543068, + 0.011658512987196445, + 0.058189183473587036, + 0.003777802223339677, + 0.0028241267427802086, + 0.0037336559034883976, + 0.003476210404187441, + 0.002444461453706026, + 0.0025023650377988815, + 0.0028241530526429415, + 0.00389309786260128, + 0.004919054452329874, + 0.00597485201433301, + 0.005596486385911703, + 0.002837427658960223, + 0.005023851525038481, + 0.0033948528580367565, + 0.003612611209973693, + 0.005268185865134001, + 0.013877457939088345, + 0.013008533976972103, + 0.02385532669723034, + 0.015426740981638432, + 0.013921653851866722, + 0.010686329565942287, + 0.015028228051960468, + 0.018469277769327164, + 0.004369817674160004, + 0.005884256213903427, + 0.007411639671772718, + 0.00576774962246418, + 0.0035367426462471485, + 0.003661598777398467, + 0.005247773602604866, + 0.011209875345230103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 4 + ], + "selected_reward": -0.06563979387283325 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide shot moving in slow, graceful arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha, a young man with twin black buns bound by red ropes, visible riding fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha wearing sleeveless crimson armor embroidered with golden flames and holding a fire-tipped spear taller than himself in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of clouds with thick rolling formations and brilliant golden sunlight casting rays from a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fire wheels rotating and leaving a long trail of interwoven golden and red flames behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014208933338522911 + ], + "opposite_logprob_tokens": [ + -4.264208793640137 + ], + "logprob_reward": -0.014208933338522911, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014208933338522911, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08864333480596542, + 0.05093229562044144, + 0.04515698924660683, + 0.02489781752228737, + 0.4217628836631775, + 0.09268490225076675, + 0.0712192952632904, + 0.04478664696216583, + 0.14094330370426178, + 0.01897253468632698 + ], + "credit_map": [ + 0.011521266773343086, + 0.007476368919014931, + 0.013150595128536224, + 0.01291971281170845, + 0.008698100224137306, + 0.009771517477929592, + 0.010936610400676727, + 0.014169160276651382, + 0.006671394687145948, + 0.004497169051319361, + 0.008074774406850338, + 0.005595176015049219, + 0.004064307548105717, + 0.007127333898097277, + 0.007480486761778593, + 0.0074216583743691444, + 0.003558892523869872, + 0.0030589464586228132, + 0.005301229190081358, + 0.005523243919014931, + 0.007480526342988014, + 0.005032812710851431, + 0.007805742789059877, + 0.007395595777779818, + 0.0028531134594231844, + 0.0018669948913156986, + 0.004758602939546108, + 0.0018142411718145013, + 0.002728919265791774, + 0.0024252827279269695, + 0.0029847011901438236, + 0.005465960130095482, + 0.059026703238487244, + 0.1500207483768463, + 0.11484625935554504, + 0.015871265903115273, + 0.011524977162480354, + 0.02414955012500286, + 0.009831718169152737, + 0.03649168089032173, + 0.0069871325977146626, + 0.006968917325139046, + 0.026471450924873352, + 0.00591403478756547, + 0.006036990322172642, + 0.006668837275356054, + 0.006991861388087273, + 0.02664567530155182, + 0.00881819985806942, + 0.006087173707783222, + 0.0147929098457098, + 0.013312717899680138, + 0.00618921360000968, + 0.004598116502165794, + 0.008278090506792068, + 0.009142874740064144, + 0.006507356185466051, + 0.007307335268706083, + 0.006435347720980644, + 0.004298186395317316, + 0.005257371347397566, + 0.0033057695254683495, + 0.006271404679864645, + 0.005403878632932901, + 0.014081521891057491, + 0.013964900746941566, + 0.024703023955225945, + 0.018644506111741066, + 0.014934007078409195, + 0.011871203780174255, + 0.018953682854771614, + 0.02379046007990837, + 0.001728016883134842, + 0.002096863929182291, + 0.0029337049927562475, + 0.0031402953900396824, + 0.0019461798947304487, + 0.0017635842086747289, + 0.002231243997812271, + 0.003132645972073078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 5 + ], + "selected_reward": -0.014208933338522911 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide shot moving in slow, graceful arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha, a young man with twin black buns bound by red ropes, visible riding fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha wearing sleeveless crimson armor embroidered with golden flames and holding a fire-tipped spear taller than himself in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of clouds with thick rolling formations and brilliant golden sunlight casting rays from a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fire wheels rotating and leaving a long trail of interwoven golden and red flames behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038119446486234665 + ], + "opposite_logprob_tokens": [ + -3.2881195545196533 + ], + "logprob_reward": -0.038119446486234665, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038119446486234665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05772690102458, + 0.055773571133613586, + 0.054918620735406876, + 0.04656926542520523, + 0.5744747519493103, + 0.0645841732621193, + 0.027292683720588684, + 0.03757641837000847, + 0.04245642200112343, + 0.038627129048109055 + ], + "credit_map": [ + 0.0053452858701348305, + 0.005200985819101334, + 0.007911880500614643, + 0.0079589132219553, + 0.00637463852763176, + 0.007987468503415585, + 0.00985976867377758, + 0.007087963167577982, + 0.006856562104076147, + 0.0046820794232189655, + 0.007173636462539434, + 0.005797546356916428, + 0.0061508724465966225, + 0.005063982214778662, + 0.012125461362302303, + 0.007923434488475323, + 0.005418663844466209, + 0.004129502456635237, + 0.006725847255438566, + 0.006062970496714115, + 0.0077818515710532665, + 0.0067049176432192326, + 0.008131169714033604, + 0.009963705204427242, + 0.003444344736635685, + 0.003483137348666787, + 0.009352783672511578, + 0.003843597834929824, + 0.00405204389244318, + 0.004507283214479685, + 0.005254468414932489, + 0.012631609104573727, + 0.08140693604946136, + 0.2141818404197693, + 0.1519738733768463, + 0.023113276809453964, + 0.017347780987620354, + 0.03385157510638237, + 0.012723555788397789, + 0.03987596184015274, + 0.009620849043130875, + 0.019338686019182205, + 0.005546052008867264, + 0.004011173266917467, + 0.002449692925438285, + 0.006654697470366955, + 0.0022431910037994385, + 0.014719841070473194, + 0.004529993049800396, + 0.003140025306493044, + 0.006591166835278273, + 0.0033383630216121674, + 0.002402920974418521, + 0.0019151230808347464, + 0.002295093610882759, + 0.00308000179938972, + 0.005111070349812508, + 0.004158367868512869, + 0.005390611942857504, + 0.004746715538203716, + 0.004253423772752285, + 0.002837055129930377, + 0.006010615732520819, + 0.005068561062216759, + 0.0062252492643892765, + 0.004591228906065226, + 0.007184232119470835, + 0.005102232564240694, + 0.0034743209835141897, + 0.004644987639039755, + 0.004955466371029615, + 0.0062787095084786415, + 0.0043522994965314865, + 0.004954068455845118, + 0.006102187559008598, + 0.006810253486037254, + 0.0037059884052723646, + 0.002149831736460328, + 0.004297120030969381, + 0.006255383137613535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.038119446486234665 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide shot moving in slow, graceful arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha, a young man with twin black buns bound by red ropes, visible riding fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha wearing sleeveless crimson armor embroidered with golden flames and holding a fire-tipped spear taller than himself in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of clouds with thick rolling formations and brilliant golden sunlight casting rays from a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fire wheels rotating and leaving a long trail of interwoven golden and red flames behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019393223337829113 + ], + "opposite_logprob_tokens": [ + -6.251939296722412 + ], + "logprob_reward": -0.0019393223337829113, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019393223337829113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05159195885062218, + 0.029091330245137215, + 0.038815561681985855, + 0.027046287432312965, + 0.24509523808956146, + 0.33471253514289856, + 0.0893392562866211, + 0.038292694836854935, + 0.05302263796329498, + 0.09299248456954956 + ], + "credit_map": [ + 0.007685734890401363, + 0.005439793225377798, + 0.007871747016906738, + 0.007388690952211618, + 0.0028933680150657892, + 0.008303267881274223, + 0.004083680454641581, + 0.007925678044557571, + 0.0027234486769884825, + 0.002910724375396967, + 0.0036229323595762253, + 0.003087507328018546, + 0.002750623505562544, + 0.0035043067764490843, + 0.005781802348792553, + 0.004709984175860882, + 0.008193541318178177, + 0.003044873010367155, + 0.005353450775146484, + 0.004464435391128063, + 0.0036566508933901787, + 0.0036698514595627785, + 0.004104962106794119, + 0.006327797658741474, + 0.0051558236591517925, + 0.0035206691827625036, + 0.003667631419375539, + 0.0018346018623560667, + 0.0023949211463332176, + 0.003641771851107478, + 0.001812645117752254, + 0.005018222611397505, + 0.03380627930164337, + 0.0825418010354042, + 0.06592580676078796, + 0.010665793903172016, + 0.008912903256714344, + 0.013022483326494694, + 0.009140039794147015, + 0.021080126985907555, + 0.0507708415389061, + 0.12231843918561935, + 0.026629047468304634, + 0.018893849104642868, + 0.012850954197347164, + 0.02703683450818062, + 0.009145190939307213, + 0.06706736981868744, + 0.010486681014299393, + 0.015045449137687683, + 0.01691744476556778, + 0.010704811662435532, + 0.006656625773757696, + 0.009121299721300602, + 0.006922704633325338, + 0.013484242372214794, + 0.005627949256449938, + 0.00552272517234087, + 0.005529685877263546, + 0.0047078244388103485, + 0.0034991044085472822, + 0.004313833080232143, + 0.004163420293480158, + 0.004928152076900005, + 0.0053874836303293705, + 0.00591130368411541, + 0.007808256428688765, + 0.007096724584698677, + 0.005587269552052021, + 0.004822050221264362, + 0.007199135608971119, + 0.009210415184497833, + 0.011535382829606533, + 0.012402638792991638, + 0.013896281830966473, + 0.010432912036776543, + 0.007972552441060543, + 0.011062822304666042, + 0.007947600446641445, + 0.01774229109287262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 9 + ], + "selected_reward": -0.0019393223337829113 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide shot moving in slow, graceful arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha, a young man with twin black buns bound by red ropes, visible riding fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha wearing sleeveless crimson armor embroidered with golden flames and holding a fire-tipped spear taller than himself in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of clouds with thick rolling formations and brilliant golden sunlight casting rays from a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fire wheels rotating and leaving a long trail of interwoven golden and red flames behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028187802527099848 + ], + "opposite_logprob_tokens": [ + -5.877818584442139 + ], + "logprob_reward": -0.0028187802527099848, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028187802527099848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09648788720369339, + 0.06189147010445595, + 0.02808026783168316, + 0.020371560007333755, + 0.412761390209198, + 0.23334798216819763, + 0.07720605283975601, + 0.020426373928785324, + 0.03378431499004364, + 0.015642784535884857 + ], + "credit_map": [ + 0.014412601478397846, + 0.00822706799954176, + 0.013752485625445843, + 0.018310366198420525, + 0.009215489029884338, + 0.010338127613067627, + 0.011408013291656971, + 0.010823721066117287, + 0.008401288650929928, + 0.005323045887053013, + 0.007759055122733116, + 0.007860100828111172, + 0.005598093848675489, + 0.006501337047666311, + 0.009516989812254906, + 0.010931548662483692, + 0.0037559685297310352, + 0.0025146258994936943, + 0.0038386289961636066, + 0.003955645952373743, + 0.002453824505209923, + 0.002958565717563033, + 0.0030738194473087788, + 0.00552918529137969, + 0.002355180913582444, + 0.0020854033064097166, + 0.0038142260164022446, + 0.0016815760172903538, + 0.0017140752170234919, + 0.0026624982710927725, + 0.0013811506796628237, + 0.0046774460934102535, + 0.055667661130428314, + 0.14583422243595123, + 0.1138121634721756, + 0.015419011004269123, + 0.010161717422306538, + 0.02258770354092121, + 0.008523081429302692, + 0.04075581207871437, + 0.03000076301395893, + 0.06426054984331131, + 0.06361488252878189, + 0.009915359318256378, + 0.007074013352394104, + 0.012554152868688107, + 0.006259334739297628, + 0.03966889902949333, + 0.0071407826617360115, + 0.007875153794884682, + 0.012731591239571571, + 0.014632203616201878, + 0.007716786116361618, + 0.005358218215405941, + 0.011920738965272903, + 0.009830567054450512, + 0.003761047264561057, + 0.0031898654997348785, + 0.0029512071050703526, + 0.00165712705347687, + 0.0019084564410150051, + 0.0013692063512280583, + 0.00272993347607553, + 0.002859526313841343, + 0.005765583831816912, + 0.0037220458034425974, + 0.006900816224515438, + 0.0035648224875330925, + 0.0029779146425426006, + 0.0035114639904350042, + 0.0032012234441936016, + 0.004140438977628946, + 0.0017151772044599056, + 0.0014937528176233172, + 0.0024942737072706223, + 0.002628996269777417, + 0.0010991909075528383, + 0.0016611336031928658, + 0.0020818132907152176, + 0.002468444872647524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.0028187802527099848 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic wide shot, the camera moving in slow, graceful arcs, captures Nezha, a young man in red, riding on a pair of fire wheels, slowly circling above a boundless, majestic sea of clouds. He wears his signature twin black buns, bound with red ropes. His face is handsome, his brows radiating a determination belying his age and a hint of rebelliousness. He wears a sleeveless, crimson armor fashioned from the Hun Tian Ling, intricately embroidered with golden flames, gleaming with a divine sheen in the light. Steadily in his right hand, he grips a fire-tipped spear taller than himself. The gleaming metallic gleam of the spear's shaft contrasts sharply with the ever-burning, blazing flames of true fire atop its tip. His feet rest firmly on a pair of flaming fire wheels, which rotate with a seemingly slow yet powerful rhythm, leaving a long trail of interwoven golden and red flames behind him. Below him are thick, rolling clouds like cotton, forming continuous mountains in the clouds and bottomless canyons. Brilliant golden sunlight pours down from the clear blue sky, penetrating the gaps in the clouds and casting several magnificent rays of light. It not only outlines a circle of dazzling golden edges for the outline of his red armor, but also makes the whole scene full of a sacred, majestic and tranquil atmosphere, presenting a grand cinematic feeling like an epic oriental mythology.", + "seed": 15759, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_12/index_6485_seed_15759_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.011881885351613164, + "vlm_logprob": -0.011881885351613164, + "vlm_yes_no_margin": 4.749999952316284, + "vlm_token_credit": -0.011881885351613164, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide shot moving in slow, graceful arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha, a young man with twin black buns bound by red ropes, visible riding fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha wearing sleeveless crimson armor embroidered with golden flames and holding a fire-tipped spear taller than himself in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of clouds with thick rolling formations and brilliant golden sunlight casting rays from a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fire wheels rotating and leaving a long trail of interwoven golden and red flames behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020182136446237564 + ], + "opposite_logprob_tokens": [ + -4.145182132720947 + ], + "logprob_reward": -0.020182136446237564, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.020182136446237564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10170269757509232, + 0.06306828558444977, + 0.15150219202041626, + 0.1257658749818802, + 0.15250647068023682, + 0.1683814376592636, + 0.09745625406503677, + 0.034388862550258636, + 0.0693286880850792, + 0.035899288952350616 + ], + "credit_map": [ + 0.007458585314452648, + 0.021010693162679672, + 0.009958373382687569, + 0.016655460000038147, + 0.0170118510723114, + 0.007191941607743502, + 0.009843535721302032, + 0.012572248466312885, + 0.005425835959613323, + 0.007376684341579676, + 0.00644743861630559, + 0.012448834255337715, + 0.008295344188809395, + 0.005078874994069338, + 0.009110623970627785, + 0.008884639479219913, + 0.023486359044909477, + 0.012024023570120335, + 0.010078558698296547, + 0.012282037176191807, + 0.03047187067568302, + 0.01706547476351261, + 0.019065456464886665, + 0.027028393000364304, + 0.025539135560393333, + 0.012270160019397736, + 0.021346021443605423, + 0.006157278548926115, + 0.029160132631659508, + 0.008867508731782436, + 0.008479030802845955, + 0.013946587219834328, + 0.014897880144417286, + 0.012287999503314495, + 0.02163928747177124, + 0.008632801473140717, + 0.03496123477816582, + 0.03763646259903908, + 0.011116514913737774, + 0.011334271170198917, + 0.03170352429151535, + 0.0261115413159132, + 0.0340951532125473, + 0.010472442954778671, + 0.01567646861076355, + 0.015885159373283386, + 0.016235865652561188, + 0.018201256170868874, + 0.013480284251272678, + 0.009895830415189266, + 0.01704784296452999, + 0.00833381712436676, + 0.00813183095306158, + 0.01430224347859621, + 0.011126041412353516, + 0.015138350427150726, + 0.0033367550931870937, + 0.003780071157962084, + 0.005102020688354969, + 0.0026521100662648678, + 0.0038444080855697393, + 0.005558260716497898, + 0.006218440365046263, + 0.0038967912551015615, + 0.0064049819484353065, + 0.007222830317914486, + 0.010026001371443272, + 0.006821972783654928, + 0.010195689275860786, + 0.01084521971642971, + 0.006577460095286369, + 0.011234522797167301, + 0.003889513900503516, + 0.0026672822423279285, + 0.006581893190741539, + 0.004791993647813797, + 0.005269624292850494, + 0.003597593167796731, + 0.005284395068883896, + 0.003816989017650485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 2 + ], + "selected_reward": -0.020182136446237564 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide shot moving in slow, graceful arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha, a young man with twin black buns bound by red ropes, visible riding fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha wearing sleeveless crimson armor embroidered with golden flames and holding a fire-tipped spear taller than himself in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of clouds with thick rolling formations and brilliant golden sunlight casting rays from a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fire wheels rotating and leaving a long trail of interwoven golden and red flames behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014216807670891285 + ], + "opposite_logprob_tokens": [ + -4.264216899871826 + ], + "logprob_reward": -0.014216807670891285, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014216807670891285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14580312371253967, + 0.06586010754108429, + 0.059872936457395554, + 0.18972578644752502, + 0.18066945672035217, + 0.058853860944509506, + 0.0718706026673317, + 0.04358018934726715, + 0.14162740111351013, + 0.04213664308190346 + ], + "credit_map": [ + 0.010719947516918182, + 0.02155471406877041, + 0.017520777881145477, + 0.021225955337285995, + 0.023442434147000313, + 0.014567852020263672, + 0.015551969408988953, + 0.021219460293650627, + 0.005903136916458607, + 0.011163000948727131, + 0.007794593460857868, + 0.01401897519826889, + 0.008650439791381359, + 0.005401094909757376, + 0.0076491148211061954, + 0.005279740318655968, + 0.0062444438226521015, + 0.009865500964224339, + 0.005545738618820906, + 0.007438203785568476, + 0.008460285142064095, + 0.007081786170601845, + 0.007042346056550741, + 0.00819462165236473, + 0.06018455699086189, + 0.007866431958973408, + 0.008055216632783413, + 0.004827563650906086, + 0.07832884788513184, + 0.018144700676202774, + 0.006519908085465431, + 0.0057985433377325535, + 0.01653132401406765, + 0.007935300469398499, + 0.016475100070238113, + 0.005220730789005756, + 0.044713277369737625, + 0.06665016710758209, + 0.007361228112131357, + 0.015782300382852554, + 0.005335611756891012, + 0.005374700762331486, + 0.014373160898685455, + 0.006803678814321756, + 0.0060411724261939526, + 0.009239356964826584, + 0.005737896542996168, + 0.005948275793343782, + 0.0076039317063987255, + 0.006075764540582895, + 0.00950908288359642, + 0.005422318819910288, + 0.007152467500418425, + 0.010166221298277378, + 0.019186925143003464, + 0.006753882393240929, + 0.00459402846172452, + 0.004987840075045824, + 0.004687365144491196, + 0.003554732771590352, + 0.008756818249821663, + 0.004769286140799522, + 0.007103501353412867, + 0.005126611795276403, + 0.012516778893768787, + 0.021308815106749535, + 0.014073043130338192, + 0.007804682943969965, + 0.022372497245669365, + 0.029782257974147797, + 0.01405632309615612, + 0.019712982699275017, + 0.0057868012227118015, + 0.005498709622770548, + 0.004468681290745735, + 0.005767884198576212, + 0.005416031926870346, + 0.003940424881875515, + 0.005525012966245413, + 0.005733093246817589 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.014216807670891285 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide shot moving in slow, graceful arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha, a young man with twin black buns bound by red ropes, visible riding fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha wearing sleeveless crimson armor embroidered with golden flames and holding a fire-tipped spear taller than himself in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of clouds with thick rolling formations and brilliant golden sunlight casting rays from a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fire wheels rotating and leaving a long trail of interwoven golden and red flames behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820252276957035 + ], + "opposite_logprob_tokens": [ + -4.018202304840088 + ], + "logprob_reward": -0.01820252276957035, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01820252276957035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06134774163365364, + 0.03786129876971245, + 0.06758324801921844, + 0.14080725610256195, + 0.2698570787906647, + 0.11666740477085114, + 0.14587558805942535, + 0.028413334861397743, + 0.08421879261732101, + 0.04736824706196785 + ], + "credit_map": [ + 0.0040653664618730545, + 0.011484059505164623, + 0.008377316407859325, + 0.010531737469136715, + 0.010771403089165688, + 0.00399786839261651, + 0.005978750064969063, + 0.006141241639852524, + 0.003432595869526267, + 0.006351585034281015, + 0.005174483638256788, + 0.007937089540064335, + 0.004964903462678194, + 0.0032343927305191755, + 0.0035421671345829964, + 0.0032240853179246187, + 0.011211149394512177, + 0.008918730542063713, + 0.0040395320393145084, + 0.010094291530549526, + 0.016221685335040092, + 0.00580963771790266, + 0.005149334669113159, + 0.006138884462416172, + 0.04267365485429764, + 0.008428205735981464, + 0.007088447920978069, + 0.005514693912118673, + 0.05358750373125076, + 0.012301145121455193, + 0.006422246340662241, + 0.004791364073753357, + 0.030305851250886917, + 0.011212448589503765, + 0.026120668277144432, + 0.00608424236997962, + 0.05761164799332619, + 0.100639209151268, + 0.011321216821670532, + 0.026561785489320755, + 0.025518666952848434, + 0.01834413781762123, + 0.028211547061800957, + 0.0077235461212694645, + 0.010204430669546127, + 0.00843009352684021, + 0.008670764043927193, + 0.009564215317368507, + 0.013937595300376415, + 0.010692626237869263, + 0.016525674611330032, + 0.011627738364040852, + 0.015573297627270222, + 0.01804683730006218, + 0.0439833328127861, + 0.01548848208039999, + 0.0034114529844373465, + 0.004481105599552393, + 0.003936638589948416, + 0.001961314585059881, + 0.004455024842172861, + 0.002814463572576642, + 0.0038344936911016703, + 0.003518841229379177, + 0.0057335542514920235, + 0.007151636760681868, + 0.009318863041698933, + 0.004335416480898857, + 0.004059679806232452, + 0.025967730209231377, + 0.004935912322252989, + 0.02271600440144539, + 0.004565750248730183, + 0.004321433138102293, + 0.005498836748301983, + 0.005236292723566294, + 0.008635137230157852, + 0.004511004779487848, + 0.010610598139464855, + 0.0039891935884952545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 3 + ], + "selected_reward": -0.01820252276957035 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide shot moving in slow, graceful arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha, a young man with twin black buns bound by red ropes, visible riding fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha wearing sleeveless crimson armor embroidered with golden flames and holding a fire-tipped spear taller than himself in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of clouds with thick rolling formations and brilliant golden sunlight casting rays from a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fire wheels rotating and leaving a long trail of interwoven golden and red flames behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003190429415553808 + ], + "opposite_logprob_tokens": [ + -5.753190517425537 + ], + "logprob_reward": -0.003190429415553808, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003190429415553808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22845998406410217, + 0.04341649264097214, + 0.12557674944400787, + 0.12689252197742462, + 0.05563003197312355, + 0.16481897234916687, + 0.10128720104694366, + 0.0906195193529129, + 0.04231887683272362, + 0.02097966894507408 + ], + "credit_map": [ + 0.01928221434354782, + 0.02764434926211834, + 0.035142142325639725, + 0.024914970621466637, + 0.034007858484983444, + 0.021822750568389893, + 0.026974396780133247, + 0.038671307265758514, + 0.0040096319280564785, + 0.004731322173029184, + 0.005419999361038208, + 0.003744569607079029, + 0.005720203276723623, + 0.0060553839430212975, + 0.00594813097268343, + 0.007787254638969898, + 0.01706051640212536, + 0.010246888734400272, + 0.011739322915673256, + 0.011072173714637756, + 0.01688779704272747, + 0.013758777640759945, + 0.02004219777882099, + 0.02476908080279827, + 0.03176659345626831, + 0.00619430560618639, + 0.012472118251025677, + 0.003462117165327072, + 0.042231567203998566, + 0.01092465315014124, + 0.006033156532794237, + 0.013808008283376694, + 0.006405106745660305, + 0.0056691416539251804, + 0.010810057632625103, + 0.0035803918726742268, + 0.007208002731204033, + 0.007277766242623329, + 0.00564607186242938, + 0.009033488109707832, + 0.018020952120423317, + 0.014566460624337196, + 0.05020618438720703, + 0.006525037344545126, + 0.014964891597628593, + 0.04354714974761009, + 0.008675361052155495, + 0.008312935009598732, + 0.01329000573605299, + 0.007543168030679226, + 0.018934151157736778, + 0.007263575680553913, + 0.009721608832478523, + 0.013713572174310684, + 0.013223857618868351, + 0.0175972580909729, + 0.011089665815234184, + 0.009899740107357502, + 0.019565505906939507, + 0.006727572064846754, + 0.011335444636642933, + 0.014591723680496216, + 0.00523220282047987, + 0.012177668511867523, + 0.004106574691832066, + 0.003941001370549202, + 0.0054339757189154625, + 0.004565907642245293, + 0.005736636463552713, + 0.005779746919870377, + 0.006002083420753479, + 0.006752952467650175, + 0.0023750783875584602, + 0.0019800784066319466, + 0.0043851756490767, + 0.002812889637425542, + 0.0023892573080956936, + 0.0020424844697117805, + 0.0025201672688126564, + 0.0024745375849306583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.003190429415553808 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide shot moving in slow, graceful arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha, a young man with twin black buns bound by red ropes, visible riding fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha wearing sleeveless crimson armor embroidered with golden flames and holding a fire-tipped spear taller than himself in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of clouds with thick rolling formations and brilliant golden sunlight casting rays from a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fire wheels rotating and leaving a long trail of interwoven golden and red flames behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003617530455812812 + ], + "opposite_logprob_tokens": [ + -5.628617763519287 + ], + "logprob_reward": -0.003617530455812812, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003617530455812812, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04366191104054451, + 0.05803028494119644, + 0.11042307317256927, + 0.07387906312942505, + 0.2958047091960907, + 0.23798929154872894, + 0.0802675187587738, + 0.04543023928999901, + 0.027957472950220108, + 0.026556437835097313 + ], + "credit_map": [ + 0.0035823241341859102, + 0.006546265911310911, + 0.006560690235346556, + 0.008663702756166458, + 0.006043681409209967, + 0.0032558797392994165, + 0.004419328644871712, + 0.0045900363475084305, + 0.004876266699284315, + 0.0056623234413564205, + 0.008256621658802032, + 0.007266873959451914, + 0.006433148868381977, + 0.004019667394459248, + 0.01190374419093132, + 0.009611640125513077, + 0.017254533246159554, + 0.0084052300080657, + 0.006674603093415499, + 0.009928533807396889, + 0.023544196039438248, + 0.012248129583895206, + 0.012007691897451878, + 0.020360155031085014, + 0.01871331036090851, + 0.009031660854816437, + 0.00384938414208591, + 0.0033686337992548943, + 0.02507702447474003, + 0.006488359067589045, + 0.003989061340689659, + 0.00336162350140512, + 0.02064778283238411, + 0.019003691151738167, + 0.02533199079334736, + 0.00849999487400055, + 0.08692711591720581, + 0.10025263577699661, + 0.011102324351668358, + 0.02403918467462063, + 0.015910429880023003, + 0.006785659585148096, + 0.07271600514650345, + 0.00917243305593729, + 0.02078169584274292, + 0.08081455528736115, + 0.01754249446094036, + 0.014266024343669415, + 0.008635398931801319, + 0.006301811896264553, + 0.009094913490116596, + 0.006930986884981394, + 0.00942925363779068, + 0.008974605239927769, + 0.024372069165110588, + 0.006528480444103479, + 0.004205208737403154, + 0.005700356792658567, + 0.00410083681344986, + 0.004428897053003311, + 0.011141848750412464, + 0.00413252878934145, + 0.006718181539326906, + 0.005002379883080721, + 0.0030171393882483244, + 0.004474146291613579, + 0.004585205111652613, + 0.002846712712198496, + 0.002831234596669674, + 0.002828261349350214, + 0.003117271000519395, + 0.004257501102983952, + 0.0026140143163502216, + 0.0026172748766839504, + 0.0028448523953557014, + 0.0027939036954194307, + 0.004595338366925716, + 0.0031513264402747154, + 0.004952854476869106, + 0.0029868728015571833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.003617530455812812 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic wide shot, the camera moving in slow, graceful arcs, captures Nezha, a young man in red, riding on a pair of fire wheels, slowly circling above a boundless, majestic sea of clouds. He wears his signature twin black buns, bound with red ropes. His face is handsome, his brows radiating a determination belying his age and a hint of rebelliousness. He wears a sleeveless, crimson armor fashioned from the Hun Tian Ling, intricately embroidered with golden flames, gleaming with a divine sheen in the light. Steadily in his right hand, he grips a fire-tipped spear taller than himself. The gleaming metallic gleam of the spear's shaft contrasts sharply with the ever-burning, blazing flames of true fire atop its tip. His feet rest firmly on a pair of flaming fire wheels, which rotate with a seemingly slow yet powerful rhythm, leaving a long trail of interwoven golden and red flames behind him. Below him are thick, rolling clouds like cotton, forming continuous mountains in the clouds and bottomless canyons. Brilliant golden sunlight pours down from the clear blue sky, penetrating the gaps in the clouds and casting several magnificent rays of light. It not only outlines a circle of dazzling golden edges for the outline of his red armor, but also makes the whole scene full of a sacred, majestic and tranquil atmosphere, presenting a grand cinematic feeling like an epic oriental mythology.", + "seed": 15760, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_12/index_6485_seed_15760_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.012065955461002887, + "vlm_logprob": -0.012065955461002887, + "vlm_yes_no_margin": 5.124999952316284, + "vlm_token_credit": -0.012065955461002887, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide shot moving in slow, graceful arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha, a young man with twin black buns bound by red ropes, visible riding fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha wearing sleeveless crimson armor embroidered with golden flames and holding a fire-tipped spear taller than himself in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of clouds with thick rolling formations and brilliant golden sunlight casting rays from a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fire wheels rotating and leaving a long trail of interwoven golden and red flames behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021401071920990944 + ], + "opposite_logprob_tokens": [ + -4.0214009284973145 + ], + "logprob_reward": -0.021401071920990944, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.021401071920990944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0706271231174469, + 0.09469296038150787, + 0.0501948818564415, + 0.2918711006641388, + 0.08071133494377136, + 0.128562331199646, + 0.04478590935468674, + 0.10124699026346207, + 0.07134933769702911, + 0.06595800071954727 + ], + "credit_map": [ + 0.008502611890435219, + 0.008679630234837532, + 0.010862695053219795, + 0.0072920010425150394, + 0.007000379264354706, + 0.010195756331086159, + 0.006221451330929995, + 0.01187259703874588, + 0.01214482169598341, + 0.009774294681847095, + 0.015376884490251541, + 0.009899832308292389, + 0.015668001025915146, + 0.012292534112930298, + 0.010839407332241535, + 0.008697194047272205, + 0.005300105549395084, + 0.007204770110547543, + 0.006699864286929369, + 0.008762269280850887, + 0.005614376626908779, + 0.005564448889344931, + 0.005442877300083637, + 0.00560616934671998, + 0.0445525124669075, + 0.04799553006887436, + 0.04884208366274834, + 0.021700125187635422, + 0.025910912081599236, + 0.03478506952524185, + 0.034109435975551605, + 0.03397543355822563, + 0.011229964904487133, + 0.008459176868200302, + 0.017531508579850197, + 0.010258840397000313, + 0.005778565537184477, + 0.013182498514652252, + 0.007239829748868942, + 0.00703095318749547, + 0.01886807568371296, + 0.02544061839580536, + 0.015026072971522808, + 0.012284623458981514, + 0.007122923620045185, + 0.021229790523648262, + 0.0100529994815588, + 0.01853722147643566, + 0.0059669725596904755, + 0.004479784984141588, + 0.011885548010468483, + 0.006683830171823502, + 0.003293311223387718, + 0.006072673946619034, + 0.00326111470349133, + 0.0031426744535565376, + 0.012976517900824547, + 0.013556241057813168, + 0.012249980121850967, + 0.014379139989614487, + 0.014705201610922813, + 0.011542191728949547, + 0.012959536164999008, + 0.008878187276422977, + 0.007619123440235853, + 0.008463350124657154, + 0.00907641276717186, + 0.010561163537204266, + 0.006773574743419886, + 0.010367280803620815, + 0.011166293174028397, + 0.007322145160287619, + 0.005384355317801237, + 0.004466786049306393, + 0.00953393243253231, + 0.007015608716756105, + 0.0059522222727537155, + 0.008807939477264881, + 0.017384858801960945, + 0.007412299048155546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 7 + ], + "selected_reward": -0.021401071920990944 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide shot moving in slow, graceful arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha, a young man with twin black buns bound by red ropes, visible riding fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha wearing sleeveless crimson armor embroidered with golden flames and holding a fire-tipped spear taller than himself in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of clouds with thick rolling formations and brilliant golden sunlight casting rays from a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fire wheels rotating and leaving a long trail of interwoven golden and red flames behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608091965317726 + ], + "opposite_logprob_tokens": [ + -4.141080856323242 + ], + "logprob_reward": -0.01608091965317726, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608091965317726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13601936399936676, + 0.06732650846242905, + 0.10668560862541199, + 0.21989749372005463, + 0.040555790066719055, + 0.12828794121742249, + 0.05324879288673401, + 0.06121758371591568, + 0.12957796454429626, + 0.05718301236629486 + ], + "credit_map": [ + 0.015613995492458344, + 0.012933016754686832, + 0.019366620108485222, + 0.014899235218763351, + 0.017022928223013878, + 0.020193034783005714, + 0.015379142016172409, + 0.020611384883522987, + 0.007922316901385784, + 0.008279828354716301, + 0.013586585409939289, + 0.005113847553730011, + 0.006429876666516066, + 0.01590370200574398, + 0.004706413019448519, + 0.005383939947932959, + 0.010117664933204651, + 0.013821284286677837, + 0.010823741555213928, + 0.01673121005296707, + 0.01306881383061409, + 0.010308989323675632, + 0.01492173969745636, + 0.01689215935766697, + 0.025448361411690712, + 0.031228939071297646, + 0.023716770112514496, + 0.024159206077456474, + 0.025708787143230438, + 0.02730383537709713, + 0.026758404448628426, + 0.03557318076491356, + 0.005063941236585379, + 0.004169445019215345, + 0.005971534177660942, + 0.005501278676092625, + 0.0044268774800002575, + 0.005691183730959892, + 0.0035110691096633673, + 0.006220459938049316, + 0.013892993330955505, + 0.012691168114542961, + 0.018020233139395714, + 0.013933517970144749, + 0.010162229649722576, + 0.025823764503002167, + 0.011610145680606365, + 0.022153889760375023, + 0.006226943340152502, + 0.004482937511056662, + 0.010016110725700855, + 0.006842569913715124, + 0.006165019236505032, + 0.010261919349431992, + 0.004335560370236635, + 0.004917732905596495, + 0.010028297081589699, + 0.008447358384728432, + 0.008449233137071133, + 0.008580164983868599, + 0.005627467297017574, + 0.007149967364966869, + 0.004816609900444746, + 0.008118491619825363, + 0.019516361877322197, + 0.014218101277947426, + 0.017239795997738838, + 0.015080024488270283, + 0.010991659015417099, + 0.018939625471830368, + 0.020892105996608734, + 0.012700289487838745, + 0.004545155446976423, + 0.004628073889762163, + 0.008742508478462696, + 0.005636792629957199, + 0.0053650918416678905, + 0.007894940674304962, + 0.012984237633645535, + 0.007386215031147003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.01608091965317726 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide shot moving in slow, graceful arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha, a young man with twin black buns bound by red ropes, visible riding fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha wearing sleeveless crimson armor embroidered with golden flames and holding a fire-tipped spear taller than himself in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of clouds with thick rolling formations and brilliant golden sunlight casting rays from a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fire wheels rotating and leaving a long trail of interwoven golden and red flames behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020592350512742996 + ], + "opposite_logprob_tokens": [ + -3.895592451095581 + ], + "logprob_reward": -0.020592350512742996, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020592350512742996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16643396019935608, + 0.09921474754810333, + 0.06985845416784286, + 0.08040677011013031, + 0.09250640869140625, + 0.20931711792945862, + 0.0669848620891571, + 0.06084614619612694, + 0.040542419999837875, + 0.11388909071683884 + ], + "credit_map": [ + 0.012476777657866478, + 0.024363316595554352, + 0.02611786685883999, + 0.016720885410904884, + 0.01946771703660488, + 0.024480028077960014, + 0.017017370089888573, + 0.02578999474644661, + 0.010456816293299198, + 0.012878816574811935, + 0.01469776313751936, + 0.011332444846630096, + 0.010304633527994156, + 0.020984400063753128, + 0.009389979764819145, + 0.00916989054530859, + 0.0068088253028690815, + 0.009398211725056171, + 0.008287983946502209, + 0.009732985869050026, + 0.006207497324794531, + 0.006876684259623289, + 0.00877353921532631, + 0.013772728852927685, + 0.01042227167636156, + 0.011325040832161903, + 0.01050957664847374, + 0.010072351433336735, + 0.009807427413761616, + 0.010174508206546307, + 0.00912413652986288, + 0.008971461094915867, + 0.010278391651809216, + 0.010877076536417007, + 0.012708913534879684, + 0.01166393794119358, + 0.009721498005092144, + 0.013857124373316765, + 0.009250296279788017, + 0.014149180613458157, + 0.02273738756775856, + 0.05094233527779579, + 0.02860371395945549, + 0.021602677181363106, + 0.012497957795858383, + 0.03229507803916931, + 0.014111151918768883, + 0.026526829227805138, + 0.009197555482387543, + 0.006810054183006287, + 0.012668975628912449, + 0.00739672826603055, + 0.006850478705018759, + 0.009324372746050358, + 0.008693481795489788, + 0.006043218541890383, + 0.00978017970919609, + 0.009072094224393368, + 0.00786745548248291, + 0.00807031150907278, + 0.005872038193047047, + 0.006412820424884558, + 0.005736913066357374, + 0.008034337311983109, + 0.0037106398958712816, + 0.005086719524115324, + 0.005928829777985811, + 0.0048693339340388775, + 0.0037764341104775667, + 0.006721429526805878, + 0.00604331074282527, + 0.004405725747346878, + 0.007647643331438303, + 0.007718520704656839, + 0.017068523913621902, + 0.016189077869057655, + 0.00949736125767231, + 0.009409544989466667, + 0.036710578948259354, + 0.009647850878536701 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.020592350512742996 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide shot moving in slow, graceful arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha, a young man with twin black buns bound by red ropes, visible riding fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha wearing sleeveless crimson armor embroidered with golden flames and holding a fire-tipped spear taller than himself in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of clouds with thick rolling formations and brilliant golden sunlight casting rays from a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fire wheels rotating and leaving a long trail of interwoven golden and red flames behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009190387791022658 + ], + "opposite_logprob_tokens": [ + -7.000918865203857 + ], + "logprob_reward": -0.0009190387791022658, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009190387791022658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07893145829439163, + 0.08713751286268234, + 0.047682616859674454, + 0.07583891600370407, + 0.08291658014059067, + 0.18037281930446625, + 0.33835041522979736, + 0.056838348507881165, + 0.02689320035278797, + 0.025038154795765877 + ], + "credit_map": [ + 0.00931583158671856, + 0.006331751588732004, + 0.013027439825236797, + 0.008389610797166824, + 0.008223000913858414, + 0.011007457040250301, + 0.00876928772777319, + 0.013867078348994255, + 0.014379083178937435, + 0.006719878409057856, + 0.015105937607586384, + 0.010421360842883587, + 0.009202297776937485, + 0.013575549237430096, + 0.008146172389388084, + 0.009587240405380726, + 0.006768819876015186, + 0.004125362262129784, + 0.009778447449207306, + 0.0075012631714344025, + 0.005092621315270662, + 0.005144361872226, + 0.005403525196015835, + 0.003868212690576911, + 0.010111198760569096, + 0.009303498081862926, + 0.012028872966766357, + 0.006026852410286665, + 0.009360428899526596, + 0.01032733079046011, + 0.011076105758547783, + 0.007604626473039389, + 0.009310215711593628, + 0.008336007595062256, + 0.016347452998161316, + 0.009506836533546448, + 0.005612158682197332, + 0.01548679731786251, + 0.0074220821261405945, + 0.010895024985074997, + 0.015404739417135715, + 0.018221620470285416, + 0.04569048807024956, + 0.019243435934185982, + 0.006638334598392248, + 0.047422487288713455, + 0.008219098672270775, + 0.019532619044184685, + 0.051415473222732544, + 0.04022592306137085, + 0.05985016003251076, + 0.04777579382061958, + 0.021253619343042374, + 0.06346283107995987, + 0.02205684781074524, + 0.032309774309396744, + 0.007396970875561237, + 0.008500803261995316, + 0.007249961607158184, + 0.0074372673407197, + 0.0077064246870577335, + 0.006912728305906057, + 0.006190442945808172, + 0.005443749018013477, + 0.0021436254028230906, + 0.003593545174226165, + 0.004246245604008436, + 0.0035815443843603134, + 0.003539441153407097, + 0.0037534604780375957, + 0.0033937219996005297, + 0.0026416159234941006, + 0.0025139562785625458, + 0.005020318552851677, + 0.0039331912994384766, + 0.002186578931286931, + 0.0024428812321275473, + 0.0026589285116642714, + 0.0028989503625780344, + 0.003383351257070899 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 1 + ], + "selected_reward": -0.0009190387791022658 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide shot moving in slow, graceful arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha, a young man with twin black buns bound by red ropes, visible riding fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha wearing sleeveless crimson armor embroidered with golden flames and holding a fire-tipped spear taller than himself in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of clouds with thick rolling formations and brilliant golden sunlight casting rays from a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fire wheels rotating and leaving a long trail of interwoven golden and red flames behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013363964390009642 + ], + "opposite_logprob_tokens": [ + -6.626336574554443 + ], + "logprob_reward": -0.0013363964390009642, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013363964390009642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19010892510414124, + 0.06933004409074783, + 0.07589246332645416, + 0.10339028388261795, + 0.07647781074047089, + 0.04917484521865845, + 0.12199416756629944, + 0.0886896625161171, + 0.09460166841745377, + 0.13034018874168396 + ], + "credit_map": [ + 0.019435768947005272, + 0.01600278541445732, + 0.02963174134492874, + 0.016252286732196808, + 0.017542215064167976, + 0.0264563150703907, + 0.021794121712446213, + 0.0429936908185482, + 0.007424703799188137, + 0.00860882643610239, + 0.011454018764197826, + 0.00879975501447916, + 0.008123739622533321, + 0.009182434529066086, + 0.00832950696349144, + 0.007407063618302345, + 0.00840976182371378, + 0.007675496395677328, + 0.009967045858502388, + 0.009992463514208794, + 0.010115526616573334, + 0.00817778892815113, + 0.011793461628258228, + 0.009760918095707893, + 0.012450896203517914, + 0.0103062205016613, + 0.01367021631449461, + 0.010865040123462677, + 0.011991358362138271, + 0.0145279411226511, + 0.011966163292527199, + 0.017612449824810028, + 0.00747200520709157, + 0.00737022003158927, + 0.014252175576984882, + 0.011961069889366627, + 0.005957470275461674, + 0.010426998138427734, + 0.006818908732384443, + 0.012218961492180824, + 0.007137394975870848, + 0.0058516571298241615, + 0.008563501760363579, + 0.006760282441973686, + 0.002987704472616315, + 0.0076356795616447926, + 0.0033391902688890696, + 0.006899434141814709, + 0.01822465844452381, + 0.013249660842120647, + 0.01727086678147316, + 0.01731226220726967, + 0.008955069817602634, + 0.025179395452141762, + 0.008043796755373478, + 0.013758452609181404, + 0.00969536043703556, + 0.012801792472600937, + 0.009979437105357647, + 0.008507017977535725, + 0.01389108132570982, + 0.009771429933607578, + 0.01212057564407587, + 0.011922965757548809, + 0.012548308819532394, + 0.009795662946999073, + 0.013460281305015087, + 0.010857236571609974, + 0.007387983612716198, + 0.014741617254912853, + 0.01605420559644699, + 0.009756378829479218, + 0.01263363752514124, + 0.013334628194570541, + 0.0163599643856287, + 0.01625549979507923, + 0.010187270119786263, + 0.01673063263297081, + 0.029755130410194397, + 0.015083418227732182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.0013363964390009642 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic wide shot, the camera moving in slow, graceful arcs, captures Nezha, a young man in red, riding on a pair of fire wheels, slowly circling above a boundless, majestic sea of clouds. He wears his signature twin black buns, bound with red ropes. His face is handsome, his brows radiating a determination belying his age and a hint of rebelliousness. He wears a sleeveless, crimson armor fashioned from the Hun Tian Ling, intricately embroidered with golden flames, gleaming with a divine sheen in the light. Steadily in his right hand, he grips a fire-tipped spear taller than himself. The gleaming metallic gleam of the spear's shaft contrasts sharply with the ever-burning, blazing flames of true fire atop its tip. His feet rest firmly on a pair of flaming fire wheels, which rotate with a seemingly slow yet powerful rhythm, leaving a long trail of interwoven golden and red flames behind him. Below him are thick, rolling clouds like cotton, forming continuous mountains in the clouds and bottomless canyons. Brilliant golden sunlight pours down from the clear blue sky, penetrating the gaps in the clouds and casting several magnificent rays of light. It not only outlines a circle of dazzling golden edges for the outline of his red armor, but also makes the whole scene full of a sacred, majestic and tranquil atmosphere, presenting a grand cinematic feeling like an epic oriental mythology.", + "seed": 15761, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_12/index_6485_seed_15761_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.012441706808749587, + "vlm_logprob": -0.012441706808749587, + "vlm_yes_no_margin": 5.025, + "vlm_token_credit": -0.012441706808749587, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide shot moving in slow, graceful arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha, a young man with twin black buns bound by red ropes, visible riding fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha wearing sleeveless crimson armor embroidered with golden flames and holding a fire-tipped spear taller than himself in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of clouds with thick rolling formations and brilliant golden sunlight casting rays from a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fire wheels rotating and leaving a long trail of interwoven golden and red flames behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017200814560055733 + ], + "opposite_logprob_tokens": [ + -4.392200946807861 + ], + "logprob_reward": -0.017200814560055733, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.017200814560055733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20262864232063293, + 0.04831841588020325, + 0.07607182115316391, + 0.08855701237916946, + 0.03759715333580971, + 0.22285588085651398, + 0.06693343818187714, + 0.06979672610759735, + 0.12245257943868637, + 0.064788319170475 + ], + "credit_map": [ + 0.016776148229837418, + 0.028660286217927933, + 0.037403758615255356, + 0.019364308565855026, + 0.03458532318472862, + 0.010114934295415878, + 0.025955911725759506, + 0.02976798638701439, + 0.004613845143467188, + 0.0067740194499492645, + 0.007931122556328773, + 0.005792807321995497, + 0.00639802822843194, + 0.0031907297670841217, + 0.006566022057086229, + 0.007051845081150532, + 0.008904934860765934, + 0.012773172929883003, + 0.014532564207911491, + 0.005311314482241869, + 0.009920021519064903, + 0.010949480347335339, + 0.0063121123239398, + 0.0073682209476828575, + 0.010072099976241589, + 0.008532088249921799, + 0.020914925262331963, + 0.006307029165327549, + 0.007979395799338818, + 0.010491836816072464, + 0.007973812520503998, + 0.01628582738339901, + 0.0035300767049193382, + 0.008099549449980259, + 0.004899215418845415, + 0.003400934860110283, + 0.004012908786535263, + 0.005967222154140472, + 0.00341496872715652, + 0.0042722770012915134, + 0.019817551597952843, + 0.04465766251087189, + 0.03296573832631111, + 0.017000705003738403, + 0.020066499710083008, + 0.034354325383901596, + 0.02408764325082302, + 0.029905755072832108, + 0.011773881502449512, + 0.0072931344620883465, + 0.00968822930008173, + 0.005611122120171785, + 0.007500890176743269, + 0.008483657613396645, + 0.00890466570854187, + 0.00767785357311368, + 0.007760022766888142, + 0.011349938809871674, + 0.007243324536830187, + 0.006730073131620884, + 0.010580153204500675, + 0.009452885948121548, + 0.006499325856566429, + 0.010180996730923653, + 0.013306694105267525, + 0.014594039879739285, + 0.014940411783754826, + 0.014780129306018353, + 0.016087794676423073, + 0.012774311937391758, + 0.013267342001199722, + 0.02270185574889183, + 0.0071861702017486095, + 0.011671745218336582, + 0.007464115042239428, + 0.007312542758882046, + 0.006745950784534216, + 0.00590535206720233, + 0.010238217189908028, + 0.008264227770268917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.017200814560055733 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide shot moving in slow, graceful arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha, a young man with twin black buns bound by red ropes, visible riding fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha wearing sleeveless crimson armor embroidered with golden flames and holding a fire-tipped spear taller than himself in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of clouds with thick rolling formations and brilliant golden sunlight casting rays from a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fire wheels rotating and leaving a long trail of interwoven golden and red flames behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011087710037827492 + ], + "opposite_logprob_tokens": [ + -4.511087894439697 + ], + "logprob_reward": -0.011087710037827492, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011087710037827492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35834476351737976, + 0.05081312358379364, + 0.04472304880619049, + 0.08238600194454193, + 0.1445915848016739, + 0.048317816108465195, + 0.07675477117300034, + 0.11265453696250916, + 0.025016192346811295, + 0.05639820173382759 + ], + "credit_map": [ + 0.029889727011322975, + 0.05712776258587837, + 0.05550234764814377, + 0.05983864516019821, + 0.05463555082678795, + 0.01595049723982811, + 0.04812568426132202, + 0.03727453947067261, + 0.00329135381616652, + 0.006314569152891636, + 0.005240175407379866, + 0.00949858408421278, + 0.008664567954838276, + 0.003355487482622266, + 0.006012344732880592, + 0.008436041884124279, + 0.0034503010101616383, + 0.005821275524795055, + 0.008558699861168861, + 0.004231009166687727, + 0.008312826976180077, + 0.0038888126146048307, + 0.005502068903297186, + 0.0049580540508031845, + 0.0069250985980033875, + 0.008498787879943848, + 0.014964401721954346, + 0.007420454639941454, + 0.011853951960802078, + 0.011323650367558002, + 0.009416729211807251, + 0.011982924304902554, + 0.01166786439716816, + 0.023821888491511345, + 0.015209524892270565, + 0.012385002337396145, + 0.01574956811964512, + 0.02584642358124256, + 0.01803637482225895, + 0.02187493070960045, + 0.003962499089539051, + 0.00751124694943428, + 0.004798130597919226, + 0.00704935984686017, + 0.007055070251226425, + 0.0062009068205952644, + 0.007448552642017603, + 0.004292048513889313, + 0.01081746257841587, + 0.008962387219071388, + 0.008139915764331818, + 0.011068606749176979, + 0.007420872803777456, + 0.008361953310668468, + 0.014259879477322102, + 0.00772369047626853, + 0.012479598633944988, + 0.01879790425300598, + 0.01104770228266716, + 0.019835416227579117, + 0.013042500242590904, + 0.012651260010898113, + 0.01734105311334133, + 0.00745909521356225, + 0.0025751800276339054, + 0.00239876052364707, + 0.004386340267956257, + 0.0038553462363779545, + 0.0038834207225590944, + 0.0029059455264359713, + 0.002309000352397561, + 0.002702197292819619, + 0.006571044679731131, + 0.011221409775316715, + 0.004721454810351133, + 0.007135747000575066, + 0.007647775579243898, + 0.006063599605113268, + 0.006400919519364834, + 0.006636252161115408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.011087710037827492 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide shot moving in slow, graceful arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha, a young man with twin black buns bound by red ropes, visible riding fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha wearing sleeveless crimson armor embroidered with golden flames and holding a fire-tipped spear taller than himself in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of clouds with thick rolling formations and brilliant golden sunlight casting rays from a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fire wheels rotating and leaving a long trail of interwoven golden and red flames behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0298079252243042 + ], + "opposite_logprob_tokens": [ + -3.5298080444335938 + ], + "logprob_reward": -0.0298079252243042, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0298079252243042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1483115255832672, + 0.08584152162075043, + 0.1001347005367279, + 0.09907660633325577, + 0.1886117160320282, + 0.17252163589000702, + 0.056904878467321396, + 0.055597081780433655, + 0.032806262373924255, + 0.06019403040409088 + ], + "credit_map": [ + 0.014829366467893124, + 0.020937787368893623, + 0.013662895187735558, + 0.017852509394288063, + 0.029777396470308304, + 0.010609937831759453, + 0.030163725838065147, + 0.01047790888696909, + 0.00552960392087698, + 0.011337238363921642, + 0.010746575891971588, + 0.014409616589546204, + 0.01545721385627985, + 0.005698447581380606, + 0.01255458127707243, + 0.010108244605362415, + 0.0104240532964468, + 0.011306088417768478, + 0.029440339654684067, + 0.0063955639488995075, + 0.009946907870471478, + 0.009057020768523216, + 0.005578436888754368, + 0.01798628829419613, + 0.01089891791343689, + 0.00821941252797842, + 0.021621357649564743, + 0.008114892058074474, + 0.011414289474487305, + 0.011564583517611027, + 0.010938775725662708, + 0.01630437932908535, + 0.016372838988900185, + 0.03876131400465965, + 0.024189576506614685, + 0.016203001141548157, + 0.015847623348236084, + 0.02981688268482685, + 0.025254487991333008, + 0.02216600999236107, + 0.0183321051299572, + 0.030470222234725952, + 0.01370120607316494, + 0.018889743834733963, + 0.02648155577480793, + 0.02382335066795349, + 0.022730812430381775, + 0.018092650920152664, + 0.0106191486120224, + 0.0068019963800907135, + 0.006643334869295359, + 0.007387987803667784, + 0.005750870332121849, + 0.006972871255129576, + 0.005768466740846634, + 0.006960203405469656, + 0.007009296212345362, + 0.006423956714570522, + 0.0056281681172549725, + 0.009521705098450184, + 0.00882893055677414, + 0.0077330851927399635, + 0.006325645837932825, + 0.004126290790736675, + 0.0035859656054526567, + 0.006954572629183531, + 0.0031816575210541487, + 0.00413314625620842, + 0.004279032349586487, + 0.003375453408807516, + 0.003378259716555476, + 0.003918174654245377, + 0.007204558700323105, + 0.009417747147381306, + 0.006701551377773285, + 0.006072418764233589, + 0.009381514973938465, + 0.006060832180082798, + 0.0058771888725459576, + 0.009478219784796238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.0298079252243042 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide shot moving in slow, graceful arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha, a young man with twin black buns bound by red ropes, visible riding fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha wearing sleeveless crimson armor embroidered with golden flames and holding a fire-tipped spear taller than himself in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of clouds with thick rolling formations and brilliant golden sunlight casting rays from a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fire wheels rotating and leaving a long trail of interwoven golden and red flames behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009198724874295294 + ], + "opposite_logprob_tokens": [ + -7.000919818878174 + ], + "logprob_reward": -0.0009198724874295294, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009198724874295294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24743084609508514, + 0.09134384989738464, + 0.1310637891292572, + 0.07320854067802429, + 0.04882621020078659, + 0.08228809386491776, + 0.09716803580522537, + 0.041276294738054276, + 0.0611991360783577, + 0.1261952966451645 + ], + "credit_map": [ + 0.027531277388334274, + 0.034337498247623444, + 0.04309310019016266, + 0.03283163532614708, + 0.03020954690873623, + 0.01593371480703354, + 0.030209792777895927, + 0.033284250646829605, + 0.009786192327737808, + 0.013673653826117516, + 0.014140589162707329, + 0.010333654470741749, + 0.010747388936579227, + 0.01235862448811531, + 0.006599828600883484, + 0.01370390597730875, + 0.013764118775725365, + 0.015011915937066078, + 0.03428787365555763, + 0.010287233628332615, + 0.01778411492705345, + 0.015027245506644249, + 0.008157053962349892, + 0.016744211316108704, + 0.008699973113834858, + 0.007529789581894875, + 0.016447734087705612, + 0.004392670933157206, + 0.006734068505465984, + 0.00935504212975502, + 0.005149219650775194, + 0.014900034293532372, + 0.004775661043822765, + 0.010084765963256359, + 0.006963379215449095, + 0.003949671518057585, + 0.004337710328400135, + 0.007558838464319706, + 0.0043101063929498196, + 0.006846071686595678, + 0.011081668548285961, + 0.013953270390629768, + 0.015715524554252625, + 0.007130755577236414, + 0.006525418721139431, + 0.01119947899132967, + 0.009073213674128056, + 0.007608755491673946, + 0.014690118841826916, + 0.009814845398068428, + 0.014984202571213245, + 0.008572302758693695, + 0.010039106011390686, + 0.012852250598371029, + 0.01191449724137783, + 0.014300706796348095, + 0.004816045053303242, + 0.007140748202800751, + 0.004602132830768824, + 0.005100712180137634, + 0.0060186381451785564, + 0.005049695260822773, + 0.004370519891381264, + 0.004177798982709646, + 0.004799677524715662, + 0.0056617907248437405, + 0.00927846971899271, + 0.005940788891166449, + 0.007956601679325104, + 0.0067019108682870865, + 0.00826748926192522, + 0.012592402286827564, + 0.013339967466890812, + 0.01485973410308361, + 0.013322531245648861, + 0.010811645537614822, + 0.015197654254734516, + 0.014831268228590488, + 0.017681239172816277, + 0.02615123800933361 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.0009198724874295294 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide shot moving in slow, graceful arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha, a young man with twin black buns bound by red ropes, visible riding fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha wearing sleeveless crimson armor embroidered with golden flames and holding a fire-tipped spear taller than himself in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of clouds with thick rolling formations and brilliant golden sunlight casting rays from a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fire wheels rotating and leaving a long trail of interwoven golden and red flames behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031922117341309786 + ], + "opposite_logprob_tokens": [ + -5.75319242477417 + ], + "logprob_reward": -0.0031922117341309786, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031922117341309786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10751769691705704, + 0.03283246234059334, + 0.04232724383473396, + 0.04393859580159187, + 0.044694170355796814, + 0.07268150895833969, + 0.033042728900909424, + 0.045857660472393036, + 0.21182973682880402, + 0.3652781844139099 + ], + "credit_map": [ + 0.00970784854143858, + 0.015189778991043568, + 0.012995772063732147, + 0.014891456812620163, + 0.016460765153169632, + 0.008628004230558872, + 0.014941423200070858, + 0.014702649787068367, + 0.0025374561082571745, + 0.004257216118276119, + 0.004974098410457373, + 0.004102686885744333, + 0.005832691211253405, + 0.0023580824490636587, + 0.005039603915065527, + 0.003730627242475748, + 0.004723409190773964, + 0.0057393223978579044, + 0.010971812531352043, + 0.0032401084899902344, + 0.004185554571449757, + 0.004948163405060768, + 0.0029223731253296137, + 0.005596499890089035, + 0.004300979897379875, + 0.0049571264535188675, + 0.004478379152715206, + 0.00481942854821682, + 0.005725509487092495, + 0.005894012749195099, + 0.006359375547617674, + 0.007403783034533262, + 0.004101309925317764, + 0.005457484629005194, + 0.006009462755173445, + 0.003703016322106123, + 0.003370481077581644, + 0.0068010855466127396, + 0.005973971914499998, + 0.009277354925870895, + 0.006708506494760513, + 0.012737910263240337, + 0.010879837907850742, + 0.005187067203223705, + 0.0051515125669538975, + 0.011469908989965916, + 0.009105001576244831, + 0.011441761627793312, + 0.0044200001284480095, + 0.0034560083877295256, + 0.004206191282719374, + 0.004174851346760988, + 0.0028975491877645254, + 0.004424381535500288, + 0.005268623121082783, + 0.004195123910903931, + 0.007754630409181118, + 0.004300714936107397, + 0.005873347632586956, + 0.004039380233734846, + 0.007008166518062353, + 0.0065905326046049595, + 0.0038642045110464096, + 0.006426684558391571, + 0.02807989902794361, + 0.012081260792911053, + 0.03596116602420807, + 0.01759110577404499, + 0.03009624406695366, + 0.023364413529634476, + 0.018482651561498642, + 0.04617299139499664, + 0.036446645855903625, + 0.03073137253522873, + 0.041065700352191925, + 0.028956027701497078, + 0.040467847138643265, + 0.04406305029988289, + 0.062061842530965805, + 0.08148570358753204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.0031922117341309786 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_13/index_tensor([3937, 3937])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_13/index_tensor([3937, 3937])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a070ef8cf0e1cb97553d7dc3cab3284ec7fef507 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_13/index_tensor([3937, 3937])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "steady medium shot, shot from a low-to-medium angle, captured under the bright, soft overhead lighting of the stadium, perfectly captures the ecstatic moment of a player celebrating a goal during a fierce Ultimate Frisbee tournament. The young player, his dark brown curls draped against his forehead, his face radiating pure excitement and triumph, his eyes wide with ecstasy, his lips stretched into a brilliant grin. He wears a bright red, short-sleeved jersey made of moisture-wicking polyester that gleams subtly in the light. The jersey features an abstract white team crest pattern on the chest, paired with matching white shorts and knee-high red socks. Having just scored, he lowers his weight, swings his right leg back, and with a powerful, powerful kick, he unleashes a powerful force towards a ball that rolls right at his feet. Grass dust flies into the air, perfectly capturing the climax of his unleashed passion. Beneath his feet is a well-maintained, emerald green natural grass field. In the background, the silhouettes of other players in different jerseys are blurred by the depth of field. The entire picture is saturated in color and full of the clear texture of professional TV news broadcasts, accurately conveying the high-energy and contagious celebratory atmosphere in competitive sports.", + "seed": 22910, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_13/index_3937_seed_22910_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07558774249628186, + "vlm_logprob": -0.07558774249628186, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.07558774249628186, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a steady medium shot from a low-to-medium angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young player wearing a bright red short-sleeved jersey with an abstract white team crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player perform a powerful kick with his right leg towards a ball rolling at his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a well-maintained emerald green natural grass field with blurred silhouettes of other players in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the clear texture and color saturation typical of professional TV news broadcasts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09566451609134674 + ], + "opposite_logprob_tokens": [ + -3.7206645011901855 + ], + "logprob_reward": -0.09566451609134674, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.09566451609134674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10607089847326279, + 0.13810047507286072, + 0.09924058616161346, + 0.02656875178217888, + 0.1593044251203537, + 0.13832980394363403, + 0.035910509526729584, + 0.07770224660634995, + 0.17509783804416656, + 0.04367450997233391 + ], + "credit_map": [ + 0.016824349761009216, + 0.009976422414183617, + 0.015885911881923676, + 0.012119607999920845, + 0.014464707113802433, + 0.01558239758014679, + 0.010844068601727486, + 0.010373425669968128, + 0.015715131536126137, + 0.011824622750282288, + 0.016659781336784363, + 0.015222282148897648, + 0.012169839814305305, + 0.03484617546200752, + 0.017546020448207855, + 0.014116610400378704, + 0.010500166565179825, + 0.007448098622262478, + 0.013747318647801876, + 0.010360533371567726, + 0.01000992115586996, + 0.019377082586288452, + 0.011322482489049435, + 0.016474977135658264, + 0.0034364121966063976, + 0.003199707716703415, + 0.0029605787713080645, + 0.0031260123942047358, + 0.004982327576726675, + 0.0033520220313221216, + 0.003255915129557252, + 0.002255773637443781, + 0.012760146521031857, + 0.00789942778646946, + 0.03174750134348869, + 0.01639075204730034, + 0.0123593769967556, + 0.013815787620842457, + 0.015605519525706768, + 0.04872589558362961, + 0.01890387199819088, + 0.010430015623569489, + 0.022013060748577118, + 0.022511815652251244, + 0.010405711829662323, + 0.019392449408769608, + 0.019279180094599724, + 0.015393687412142754, + 0.002868387848138809, + 0.0024369042366743088, + 0.004609858151525259, + 0.004543770104646683, + 0.003710311371833086, + 0.0041471607983112335, + 0.004254313185811043, + 0.009339799173176289, + 0.007075930014252663, + 0.00682798819616437, + 0.013017338700592518, + 0.008946293964982033, + 0.007800557184964418, + 0.007751974742859602, + 0.013596856035292149, + 0.012685307301580906, + 0.009275821037590504, + 0.007839501835405827, + 0.053351178765296936, + 0.014540066942572594, + 0.0104310791939497, + 0.009453997947275639, + 0.014623040333390236, + 0.055583156645298004, + 0.004093654919415712, + 0.00647283811122179, + 0.00561134796589613, + 0.00410194369032979, + 0.006534174084663391, + 0.005333655048161745, + 0.006641027983278036, + 0.0048858653753995895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 5 + ], + "selected_reward": -0.09566451609134674 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a steady medium shot from a low-to-medium angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young player wearing a bright red short-sleeved jersey with an abstract white team crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player perform a powerful kick with his right leg towards a ball rolling at his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a well-maintained emerald green natural grass field with blurred silhouettes of other players in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the clear texture and color saturation typical of professional TV news broadcasts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0097922682762146 + ], + "opposite_logprob_tokens": [ + -4.634792327880859 + ], + "logprob_reward": -0.0097922682762146, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0097922682762146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0628693476319313, + 0.09245910495519638, + 0.07119900733232498, + 0.03169674798846245, + 0.1349060982465744, + 0.13822631537914276, + 0.05655565485358238, + 0.20836986601352692, + 0.17119242250919342, + 0.03252548724412918 + ], + "credit_map": [ + 0.007869796827435493, + 0.005573374219238758, + 0.006618701387196779, + 0.010157203301787376, + 0.011523067019879818, + 0.007619598880410194, + 0.007853764109313488, + 0.005653843749314547, + 0.01079566590487957, + 0.010916624218225479, + 0.012599322944879532, + 0.007459798362106085, + 0.00803324393928051, + 0.021414294838905334, + 0.005930398125201464, + 0.01530975941568613, + 0.006325042340904474, + 0.004259615670889616, + 0.009754486382007599, + 0.009432271122932434, + 0.008819730952382088, + 0.010453226044774055, + 0.008098646067082882, + 0.014055994339287281, + 0.0038288957439363003, + 0.005720372311770916, + 0.003686201525852084, + 0.0028097722679376602, + 0.004775803070515394, + 0.0034621383529156446, + 0.0038605513982474804, + 0.0035530156455934048, + 0.009526882320642471, + 0.007629802450537682, + 0.025617338716983795, + 0.016334837302565575, + 0.012662035413086414, + 0.012964333407580853, + 0.012361031025648117, + 0.037809841334819794, + 0.016881098970770836, + 0.009484473615884781, + 0.023111963644623756, + 0.022191720083355904, + 0.010266653262078762, + 0.01976771093904972, + 0.019527947530150414, + 0.016994759440422058, + 0.003803964937105775, + 0.00469640688970685, + 0.008956380188465118, + 0.010172040201723576, + 0.009243056178092957, + 0.004028683062642813, + 0.007487236522138119, + 0.008167886175215244, + 0.012680507265031338, + 0.015962403267621994, + 0.04221590608358383, + 0.028115693479776382, + 0.0207330621778965, + 0.024885859340429306, + 0.031041279435157776, + 0.03273514285683632, + 0.009543872438371181, + 0.009076680056750774, + 0.05596598982810974, + 0.006806615274399519, + 0.0113713713362813, + 0.011299104429781437, + 0.008716526441276073, + 0.058412257581949234, + 0.0030404881108552217, + 0.004096070770174265, + 0.004665278363972902, + 0.0037059830501675606, + 0.004423831589519978, + 0.0035524924751371145, + 0.004602521657943726, + 0.004438820760697126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.0097922682762146 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a steady medium shot from a low-to-medium angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young player wearing a bright red short-sleeved jersey with an abstract white team crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player perform a powerful kick with his right leg towards a ball rolling at his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a well-maintained emerald green natural grass field with blurred silhouettes of other players in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the clear texture and color saturation typical of professional TV news broadcasts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798742264509201 + ], + "opposite_logprob_tokens": [ + -1.8048741817474365 + ], + "logprob_reward": -0.1798742264509201, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798742264509201, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03253282606601715, + 0.10961399227380753, + 0.07629591226577759, + 0.042429205030202866, + 0.13585083186626434, + 0.055246371775865555, + 0.06655044853687286, + 0.312811940908432, + 0.10290303826332092, + 0.06576542556285858 + ], + "credit_map": [ + 0.003905162215232849, + 0.003125803079456091, + 0.004389840178191662, + 0.004668809939175844, + 0.004898048471659422, + 0.0037955136504024267, + 0.00436774967238307, + 0.0033819028176367283, + 0.013109586201608181, + 0.008071658201515675, + 0.014366986230015755, + 0.011375855654478073, + 0.009006110019981861, + 0.03111150488257408, + 0.010864236392080784, + 0.011708066798746586, + 0.007582176011055708, + 0.0052757770754396915, + 0.010636013932526112, + 0.008969646878540516, + 0.007758803199976683, + 0.01400677114725113, + 0.008663296699523926, + 0.013403436169028282, + 0.0033735826145857573, + 0.007072432432323694, + 0.004666877444833517, + 0.004775825422257185, + 0.00929583702236414, + 0.004033816047012806, + 0.004485853016376495, + 0.004724983591586351, + 0.010249936021864414, + 0.007012476213276386, + 0.026069438084959984, + 0.015043424442410469, + 0.012830809690058231, + 0.01287118997424841, + 0.01323250774294138, + 0.038541048765182495, + 0.00815147627145052, + 0.008425608277320862, + 0.007958024740219116, + 0.0062799532897770405, + 0.003791084047406912, + 0.006804817821830511, + 0.006364673841744661, + 0.007470732554793358, + 0.003978794440627098, + 0.005099420435726643, + 0.009958025999367237, + 0.01417762041091919, + 0.00997480470687151, + 0.005739249754697084, + 0.009615226648747921, + 0.008007310330867767, + 0.01677200384438038, + 0.02004060707986355, + 0.06386607885360718, + 0.04557756334543228, + 0.027941694483160973, + 0.043653663247823715, + 0.04549221694469452, + 0.04946814104914665, + 0.010560118593275547, + 0.01076498907059431, + 0.019820211455225945, + 0.007421831134706736, + 0.015150725841522217, + 0.010437977500259876, + 0.010221467353403568, + 0.018525725230574608, + 0.010432438924908638, + 0.0068124812096357346, + 0.007700627204030752, + 0.005649925209581852, + 0.008898871019482613, + 0.007445964962244034, + 0.012060048058629036, + 0.006765074096620083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 1 + ], + "selected_reward": -0.1798742264509201 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a steady medium shot from a low-to-medium angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young player wearing a bright red short-sleeved jersey with an abstract white team crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player perform a powerful kick with his right leg towards a ball rolling at his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a well-maintained emerald green natural grass field with blurred silhouettes of other players in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the clear texture and color saturation typical of professional TV news broadcasts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00361218536272645 + ], + "opposite_logprob_tokens": [ + -5.628612041473389 + ], + "logprob_reward": -0.00361218536272645, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00361218536272645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0667310282588005, + 0.04437176138162613, + 0.05627674609422684, + 0.015727365389466286, + 0.04767373949289322, + 0.5599986910820007, + 0.024591444060206413, + 0.05916221812367439, + 0.11129071563482285, + 0.014176316559314728 + ], + "credit_map": [ + 0.01019323617219925, + 0.007343894802033901, + 0.010081502608954906, + 0.009368081577122211, + 0.008385653607547283, + 0.0071854908019304276, + 0.006870466750115156, + 0.007302704267203808, + 0.005183655768632889, + 0.004353310912847519, + 0.00721947755664587, + 0.003535559168085456, + 0.003903194097802043, + 0.00664706714451313, + 0.0039209467358887196, + 0.009608550928533077, + 0.006083086133003235, + 0.0075989533215761185, + 0.006106970831751823, + 0.008618206717073917, + 0.007146218791604042, + 0.007295184303075075, + 0.005590622778981924, + 0.007837505079805851, + 0.0014144503511488438, + 0.0020418085623532534, + 0.0019062620121985674, + 0.0014194693649187684, + 0.0032128237653523684, + 0.0017818429041653872, + 0.0015137172304093838, + 0.002436990849673748, + 0.004015990998595953, + 0.004912363365292549, + 0.007708659395575523, + 0.0067365760914981365, + 0.004370633512735367, + 0.0037225347477942705, + 0.004671999718993902, + 0.011534979566931725, + 0.07398396730422974, + 0.033535294234752655, + 0.09625296294689178, + 0.08729489892721176, + 0.043371789157390594, + 0.08706407994031906, + 0.07320598512887955, + 0.06528971344232559, + 0.001465442474000156, + 0.0015595639124512672, + 0.003000442637130618, + 0.003913432359695435, + 0.0035026001278311014, + 0.002495052758604288, + 0.003292204812169075, + 0.005362705793231726, + 0.0063487165607512, + 0.007971270941197872, + 0.012897065840661526, + 0.0049964957870543, + 0.007020384538918734, + 0.005258215591311455, + 0.0075496165081858635, + 0.007120456080883741, + 0.004706912208348513, + 0.003958598710596561, + 0.04145926609635353, + 0.0047714486718177795, + 0.0052503361366689205, + 0.005389492493122816, + 0.0034272430930286646, + 0.04232742637395859, + 0.001444838591851294, + 0.0015760472742840648, + 0.0017418559873476624, + 0.001476429053582251, + 0.0016257275128737092, + 0.0020108625758439302, + 0.0022006200160831213, + 0.00209993589669466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.00361218536272645 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a steady medium shot from a low-to-medium angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young player wearing a bright red short-sleeved jersey with an abstract white team crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player perform a powerful kick with his right leg towards a ball rolling at his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a well-maintained emerald green natural grass field with blurred silhouettes of other players in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the clear texture and color saturation typical of professional TV news broadcasts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08899551630020142 + ], + "opposite_logprob_tokens": [ + -2.4639954566955566 + ], + "logprob_reward": -0.08899551630020142, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08899551630020142, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11346407234668732, + 0.11411965638399124, + 0.07535884529352188, + 0.025354545563459396, + 0.032822832465171814, + 0.15484972298145294, + 0.04661690071225166, + 0.27820855379104614, + 0.08211175352334976, + 0.07709311693906784 + ], + "credit_map": [ + 0.015018416568636894, + 0.014089321717619896, + 0.01609094999730587, + 0.015240165404975414, + 0.014584993943572044, + 0.013501868583261967, + 0.013341731391847134, + 0.011596627533435822, + 0.010548202320933342, + 0.010415008291602135, + 0.017534684389829636, + 0.007380490656942129, + 0.010680069215595722, + 0.01984640769660473, + 0.008900979533791542, + 0.028813820332288742, + 0.008726012893021107, + 0.006811067461967468, + 0.010671451687812805, + 0.00788967963308096, + 0.007422936148941517, + 0.01590094156563282, + 0.006085861008614302, + 0.011850889772176743, + 0.0036023783031851053, + 0.0026652372907847166, + 0.0027600193861871958, + 0.001866714097559452, + 0.004577333573251963, + 0.004440523684024811, + 0.002334986347705126, + 0.0031073547434061766, + 0.003768100868910551, + 0.0027281560469418764, + 0.0037252353504300117, + 0.00459875725209713, + 0.0051964037120342255, + 0.0024869150947779417, + 0.006500608753412962, + 0.0038186556193977594, + 0.01742466539144516, + 0.0310040432959795, + 0.01887468248605728, + 0.014224409125745296, + 0.014139886945486069, + 0.023402106016874313, + 0.015107301063835621, + 0.020672641694545746, + 0.004993993788957596, + 0.004436661023646593, + 0.0076267169788479805, + 0.004509237129241228, + 0.00721185514703393, + 0.006234401371330023, + 0.005175113677978516, + 0.00642892112955451, + 0.021285774186253548, + 0.026717932894825935, + 0.0493696928024292, + 0.028396980836987495, + 0.02251538634300232, + 0.02704671211540699, + 0.058906733989715576, + 0.04396933689713478, + 0.007333481218665838, + 0.006140234414488077, + 0.02367199771106243, + 0.004524086136370897, + 0.006044362671673298, + 0.005074275191873312, + 0.004768967162817717, + 0.02455434575676918, + 0.007215116173028946, + 0.008999690413475037, + 0.012343455106019974, + 0.007612748071551323, + 0.010104166343808174, + 0.010118622332811356, + 0.007799230515956879, + 0.012900091707706451 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 1 + ], + "selected_reward": -0.08899551630020142 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady medium shot, shot from a low-to-medium angle, captured under the bright, soft overhead lighting of the stadium, perfectly captures the ecstatic moment of a player celebrating a goal during a fierce Ultimate Frisbee tournament. The young player, his dark brown curls draped against his forehead, his face radiating pure excitement and triumph, his eyes wide with ecstasy, his lips stretched into a brilliant grin. He wears a bright red, short-sleeved jersey made of moisture-wicking polyester that gleams subtly in the light. The jersey features an abstract white team crest pattern on the chest, paired with matching white shorts and knee-high red socks. Having just scored, he lowers his weight, swings his right leg back, and with a powerful, powerful kick, he unleashes a powerful force towards a ball that rolls right at his feet. Grass dust flies into the air, perfectly capturing the climax of his unleashed passion. Beneath his feet is a well-maintained, emerald green natural grass field. In the background, the silhouettes of other players in different jerseys are blurred by the depth of field. The entire picture is saturated in color and full of the clear texture of professional TV news broadcasts, accurately conveying the high-energy and contagious celebratory atmosphere in competitive sports.", + "seed": 22911, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_13/index_3937_seed_22911_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05879441751167178, + "vlm_logprob": -0.05879441751167178, + "vlm_yes_no_margin": 3.524999976158142, + "vlm_token_credit": -0.05879441751167178, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a steady medium shot from a low-to-medium angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young player wearing a bright red short-sleeved jersey with an abstract white team crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player perform a powerful kick with his right leg towards a ball rolling at his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a well-maintained emerald green natural grass field with blurred silhouettes of other players in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the clear texture and color saturation typical of professional TV news broadcasts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08607218414545059 + ], + "opposite_logprob_tokens": [ + -3.3360722064971924 + ], + "logprob_reward": -0.08607218414545059, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.08607218414545059, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11719504743814468, + 0.07145361602306366, + 0.047490499913692474, + 0.11538245528936386, + 0.048573922365903854, + 0.11526870727539062, + 0.14959481358528137, + 0.05750071629881859, + 0.1069524884223938, + 0.1705876737833023 + ], + "credit_map": [ + 0.017466984689235687, + 0.013703329488635063, + 0.015487244352698326, + 0.010744065977633, + 0.011035202071070671, + 0.020715150982141495, + 0.012367460876703262, + 0.015675611793994904, + 0.005929703824222088, + 0.009039824828505516, + 0.006830576341599226, + 0.005832262337207794, + 0.007183058653026819, + 0.015233160927891731, + 0.0064118653535842896, + 0.014993168413639069, + 0.004145733080804348, + 0.006664748769253492, + 0.005709407385438681, + 0.0070577687583863735, + 0.004799667280167341, + 0.0055345455184578896, + 0.007434744853526354, + 0.006143881473690271, + 0.007308760192245245, + 0.026190217584371567, + 0.011812089011073112, + 0.019723841920495033, + 0.010267557576298714, + 0.013564377091825008, + 0.012456805445253849, + 0.01405880507081747, + 0.0048540555872023106, + 0.00923331268131733, + 0.005309332627803087, + 0.005576999392360449, + 0.005813842173665762, + 0.006202130578458309, + 0.005446204449981451, + 0.006138044409453869, + 0.014049499295651913, + 0.01539897732436657, + 0.010934790596365929, + 0.011996989138424397, + 0.013170575723052025, + 0.0225494634360075, + 0.013382324948906898, + 0.013786082156002522, + 0.01293653342872858, + 0.019115453585982323, + 0.03075534850358963, + 0.019203169271349907, + 0.012283711694180965, + 0.015601333230733871, + 0.017780767753720284, + 0.021918511018157005, + 0.008090648800134659, + 0.006417268421500921, + 0.009223808534443378, + 0.005251584108918905, + 0.006251803599298, + 0.008425703272223473, + 0.006225030403584242, + 0.007614871487021446, + 0.01761269010603428, + 0.00885002501308918, + 0.01877344399690628, + 0.012620417401194572, + 0.006965031381696463, + 0.012823735363781452, + 0.013897272758185863, + 0.015409869141876698, + 0.028866196051239967, + 0.031243033707141876, + 0.026399800553917885, + 0.01732914336025715, + 0.01541360467672348, + 0.014556250534951687, + 0.021002354100346565, + 0.01577729731798172 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.08607218414545059 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a steady medium shot from a low-to-medium angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young player wearing a bright red short-sleeved jersey with an abstract white team crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player perform a powerful kick with his right leg towards a ball rolling at his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a well-maintained emerald green natural grass field with blurred silhouettes of other players in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the clear texture and color saturation typical of professional TV news broadcasts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014215397648513317 + ], + "opposite_logprob_tokens": [ + -4.264215469360352 + ], + "logprob_reward": -0.014215397648513317, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014215397648513317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13119295239448547, + 0.06630833446979523, + 0.07486357539892197, + 0.19032169878482819, + 0.02850649505853653, + 0.16911455988883972, + 0.11545947939157486, + 0.05043870210647583, + 0.09229280054569244, + 0.08150141686201096 + ], + "credit_map": [ + 0.0204033050686121, + 0.010252624750137329, + 0.008066699840128422, + 0.006995626259595156, + 0.010860546492040157, + 0.03199326619505882, + 0.014725884422659874, + 0.027895012870430946, + 0.006554451771080494, + 0.006621084176003933, + 0.0076194568537175655, + 0.006479216273874044, + 0.007645773235708475, + 0.009011058136820793, + 0.010070156306028366, + 0.012307138182222843, + 0.007018425967544317, + 0.01882976107299328, + 0.0051020351238548756, + 0.010940435342490673, + 0.009867384098470211, + 0.007064482197165489, + 0.008734320290386677, + 0.00730673735961318, + 0.015359452925622463, + 0.04824960604310036, + 0.026019280776381493, + 0.03306794539093971, + 0.019372018054127693, + 0.012755894102156162, + 0.02486874721944332, + 0.010628758929669857, + 0.0026183377485722303, + 0.004874898586422205, + 0.0027364278212189674, + 0.005476511549204588, + 0.004058330785483122, + 0.0024298217613250017, + 0.0038618806283921003, + 0.0024502866435796022, + 0.02795652113854885, + 0.019592750817537308, + 0.016086921095848083, + 0.02216041460633278, + 0.014351815916597843, + 0.029480671510100365, + 0.028394874185323715, + 0.011090584099292755, + 0.009570054709911346, + 0.013202470727264881, + 0.01903594844043255, + 0.01198185607790947, + 0.012336698360741138, + 0.017276274040341377, + 0.01819438673555851, + 0.013861793093383312, + 0.006228596903383732, + 0.00747696915641427, + 0.0068532247096300125, + 0.004661259241402149, + 0.00575241120532155, + 0.009631980210542679, + 0.004475175403058529, + 0.005359083414077759, + 0.013006316497921944, + 0.008241142146289349, + 0.013279355131089687, + 0.013541518710553646, + 0.007622695527970791, + 0.012479721568524837, + 0.01418539322912693, + 0.009936656802892685, + 0.014489540830254555, + 0.012234977446496487, + 0.010448204353451729, + 0.012094205245375633, + 0.008381525985896587, + 0.00580802746117115, + 0.010558633133769035, + 0.007486298680305481 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.014215397648513317 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a steady medium shot from a low-to-medium angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young player wearing a bright red short-sleeved jersey with an abstract white team crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player perform a powerful kick with his right leg towards a ball rolling at his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a well-maintained emerald green natural grass field with blurred silhouettes of other players in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the clear texture and color saturation typical of professional TV news broadcasts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203853711485863 + ], + "opposite_logprob_tokens": [ + -2.8120384216308594 + ], + "logprob_reward": -0.06203853711485863, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203853711485863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04688828065991402, + 0.08812510967254639, + 0.05063150078058243, + 0.039858780801296234, + 0.03743980452418327, + 0.14510701596736908, + 0.2551095485687256, + 0.1413906365633011, + 0.07401224225759506, + 0.12143706530332565 + ], + "credit_map": [ + 0.008745647966861725, + 0.004565372131764889, + 0.004769134800881147, + 0.003643242409452796, + 0.004654159769415855, + 0.012683349661529064, + 0.002921303967013955, + 0.00490606902167201, + 0.006124700885266066, + 0.006738627329468727, + 0.00593674648553133, + 0.0057701170444488525, + 0.009816809557378292, + 0.02339126728475094, + 0.007263075094670057, + 0.02308376133441925, + 0.004161387216299772, + 0.008793304674327374, + 0.005886188708245754, + 0.006888978648930788, + 0.007646452635526657, + 0.0034188181161880493, + 0.008539843373000622, + 0.00529652601107955, + 0.0034797638654708862, + 0.007260067388415337, + 0.004859594162553549, + 0.006165683269500732, + 0.0038532817270606756, + 0.005365252494812012, + 0.004082066472619772, + 0.004793071188032627, + 0.004042191430926323, + 0.007937259040772915, + 0.0033542837481945753, + 0.0055648633278906345, + 0.005317392759025097, + 0.003680732101202011, + 0.0033954204991459846, + 0.004147661384195089, + 0.012591942213475704, + 0.01788572035729885, + 0.015454502776265144, + 0.03034101240336895, + 0.015027366578578949, + 0.016028912737965584, + 0.02241620048880577, + 0.015361352823674679, + 0.026973580941557884, + 0.044178809970617294, + 0.031236251816153526, + 0.03008653223514557, + 0.0244465172290802, + 0.041904326528310776, + 0.0307301077991724, + 0.02555340714752674, + 0.018632667139172554, + 0.016033539548516273, + 0.021424125880002975, + 0.01321073342114687, + 0.01642710156738758, + 0.023368772119283676, + 0.014862213283777237, + 0.01743147522211075, + 0.012292562983930111, + 0.005006115883588791, + 0.010327063500881195, + 0.009346372447907925, + 0.0062868837267160416, + 0.007589560933411121, + 0.011510848999023438, + 0.011652832850813866, + 0.01724105328321457, + 0.01612519659101963, + 0.015673214569687843, + 0.017069891095161438, + 0.01224767230451107, + 0.009476747363805771, + 0.018739594146609306, + 0.014863701537251472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.06203853711485863 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a steady medium shot from a low-to-medium angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young player wearing a bright red short-sleeved jersey with an abstract white team crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player perform a powerful kick with his right leg towards a ball rolling at his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a well-maintained emerald green natural grass field with blurred silhouettes of other players in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the clear texture and color saturation typical of professional TV news broadcasts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046369642950594425 + ], + "opposite_logprob_tokens": [ + -5.379636764526367 + ], + "logprob_reward": -0.0046369642950594425, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046369642950594425, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18047605454921722, + 0.0882914662361145, + 0.09740961343050003, + 0.15291063487529755, + 0.060518767684698105, + 0.0851820632815361, + 0.04451671242713928, + 0.07524140924215317, + 0.13649551570415497, + 0.07895779609680176 + ], + "credit_map": [ + 0.03125198557972908, + 0.014056231826543808, + 0.009213726036250591, + 0.010881662368774414, + 0.013563627377152443, + 0.04465257376432419, + 0.017472708597779274, + 0.03938354551792145, + 0.006150755099952221, + 0.00549369677901268, + 0.004385596141219139, + 0.004331492353230715, + 0.006548229139298201, + 0.02515619806945324, + 0.00649257143959403, + 0.029732923954725266, + 0.010067805647850037, + 0.014678088016808033, + 0.014253110624849796, + 0.014414033852517605, + 0.008787100203335285, + 0.00890545453876257, + 0.016621896997094154, + 0.009682120755314827, + 0.00897905696183443, + 0.03154236078262329, + 0.01870192028582096, + 0.026111219078302383, + 0.01364121213555336, + 0.01817558892071247, + 0.017126811668276787, + 0.01863245852291584, + 0.005599470343440771, + 0.011567305773496628, + 0.007069949060678482, + 0.008884319104254246, + 0.005768504925072193, + 0.006083737593144178, + 0.006601128727197647, + 0.008944353088736534, + 0.013122934848070145, + 0.00968312006443739, + 0.009165472351014614, + 0.007443847134709358, + 0.008081386797130108, + 0.016203608363866806, + 0.014815562404692173, + 0.006666127126663923, + 0.0056795417331159115, + 0.0053246202878654, + 0.004734167829155922, + 0.005527706351131201, + 0.0041292826645076275, + 0.008545110002160072, + 0.005901794880628586, + 0.004674485884606838, + 0.013339737430214882, + 0.014590069651603699, + 0.009233134798705578, + 0.006862741429358721, + 0.007710791192948818, + 0.008686440996825695, + 0.006038299296051264, + 0.008780195377767086, + 0.018068868666887283, + 0.009274138137698174, + 0.019773216918110847, + 0.013984957709908485, + 0.013463669456541538, + 0.013379473239183426, + 0.031249625608325005, + 0.01730157621204853, + 0.014410718344151974, + 0.014795433729887009, + 0.010816034860908985, + 0.010456479154527187, + 0.006832177750766277, + 0.004247178323566914, + 0.01069904025644064, + 0.006700736470520496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.0046369642950594425 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a steady medium shot from a low-to-medium angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young player wearing a bright red short-sleeved jersey with an abstract white team crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player perform a powerful kick with his right leg towards a ball rolling at his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a well-maintained emerald green natural grass field with blurred silhouettes of other players in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the clear texture and color saturation typical of professional TV news broadcasts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12700900435447693 + ], + "opposite_logprob_tokens": [ + -2.1270089149475098 + ], + "logprob_reward": -0.12700900435447693, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12700900435447693, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0817272886633873, + 0.03829493746161461, + 0.07296472787857056, + 0.22263000905513763, + 0.11345779895782471, + 0.055636510252952576, + 0.06840267032384872, + 0.12278597801923752, + 0.18122056126594543, + 0.04287956282496452 + ], + "credit_map": [ + 0.01197097823023796, + 0.009692706167697906, + 0.008857715874910355, + 0.006089637987315655, + 0.010449090041220188, + 0.011084613390266895, + 0.010188176296651363, + 0.0133943697437644, + 0.003490453353151679, + 0.0047977459616959095, + 0.003490301314741373, + 0.004058612044900656, + 0.005814529489725828, + 0.0054910932667553425, + 0.003992515150457621, + 0.007159683387726545, + 0.006979044061154127, + 0.012733523733913898, + 0.006275566760450602, + 0.009505780413746834, + 0.00942220352590084, + 0.008446349762380123, + 0.010464366525411606, + 0.009137887507677078, + 0.013681557960808277, + 0.03971220180392265, + 0.03168551251292229, + 0.03721548616886139, + 0.021265137940645218, + 0.027878589928150177, + 0.025591813027858734, + 0.025599701330065727, + 0.012056160718202591, + 0.01688244380056858, + 0.012288704514503479, + 0.018577763810753822, + 0.013797793537378311, + 0.014460870064795017, + 0.012986370362341404, + 0.01240767352283001, + 0.005034720990806818, + 0.006724619772285223, + 0.004212670028209686, + 0.00888556893914938, + 0.006334242410957813, + 0.010866772383451462, + 0.009439628571271896, + 0.004138281103223562, + 0.006629147566854954, + 0.011045074090361595, + 0.007583482190966606, + 0.007221910171210766, + 0.007827834226191044, + 0.012747852131724358, + 0.008451525121927261, + 0.006895833183079958, + 0.023404520004987717, + 0.021803060546517372, + 0.016547098755836487, + 0.010346065275371075, + 0.010556335560977459, + 0.016469640657305717, + 0.00808769278228283, + 0.015571551397442818, + 0.020954694598913193, + 0.01336679421365261, + 0.025287257507443428, + 0.019815882667899132, + 0.019190585240721703, + 0.016686197370290756, + 0.04543359950184822, + 0.020485537126660347, + 0.0076153152622282505, + 0.007217620965093374, + 0.005719131790101528, + 0.005124811083078384, + 0.005303496029227972, + 0.0032173749059438705, + 0.004437547642737627, + 0.00424426281824708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 7 + ], + "selected_reward": -0.12700900435447693 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady medium shot, shot from a low-to-medium angle, captured under the bright, soft overhead lighting of the stadium, perfectly captures the ecstatic moment of a player celebrating a goal during a fierce Ultimate Frisbee tournament. The young player, his dark brown curls draped against his forehead, his face radiating pure excitement and triumph, his eyes wide with ecstasy, his lips stretched into a brilliant grin. He wears a bright red, short-sleeved jersey made of moisture-wicking polyester that gleams subtly in the light. The jersey features an abstract white team crest pattern on the chest, paired with matching white shorts and knee-high red socks. Having just scored, he lowers his weight, swings his right leg back, and with a powerful, powerful kick, he unleashes a powerful force towards a ball that rolls right at his feet. Grass dust flies into the air, perfectly capturing the climax of his unleashed passion. Beneath his feet is a well-maintained, emerald green natural grass field. In the background, the silhouettes of other players in different jerseys are blurred by the depth of field. The entire picture is saturated in color and full of the clear texture of professional TV news broadcasts, accurately conveying the high-energy and contagious celebratory atmosphere in competitive sports.", + "seed": 22912, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_13/index_3937_seed_22912_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08851412059739232, + "vlm_logprob": -0.08851412059739232, + "vlm_yes_no_margin": 3.300000023841858, + "vlm_token_credit": -0.08851412059739232, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a steady medium shot from a low-to-medium angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young player wearing a bright red short-sleeved jersey with an abstract white team crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player perform a powerful kick with his right leg towards a ball rolling at his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a well-maintained emerald green natural grass field with blurred silhouettes of other players in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the clear texture and color saturation typical of professional TV news broadcasts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0568295456469059 + ], + "opposite_logprob_tokens": [ + -3.8068294525146484 + ], + "logprob_reward": -0.0568295456469059, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0568295456469059, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19485025107860565, + 0.10524143278598785, + 0.07829929143190384, + 0.1303061544895172, + 0.08801250159740448, + 0.07235937565565109, + 0.07623650133609772, + 0.09444421529769897, + 0.0834948793053627, + 0.07675538212060928 + ], + "credit_map": [ + 0.03326965495944023, + 0.03204004839062691, + 0.030301978811621666, + 0.01700534112751484, + 0.025928279384970665, + 0.02610841952264309, + 0.01619558222591877, + 0.014000948518514633, + 0.016900915652513504, + 0.011362244375050068, + 0.011108757928013802, + 0.019378796219825745, + 0.01582178846001625, + 0.009282151237130165, + 0.011954146437346935, + 0.009432632476091385, + 0.010170462541282177, + 0.010333272628486156, + 0.011286048218607903, + 0.0059938132762908936, + 0.011016466654837132, + 0.007451517973095179, + 0.008630075491964817, + 0.013417634181678295, + 0.009567614644765854, + 0.012537951581180096, + 0.020653972402215004, + 0.01906576007604599, + 0.01475473865866661, + 0.016882240772247314, + 0.017805200070142746, + 0.019038675352931023, + 0.018529972061514854, + 0.006147869396954775, + 0.013204659335315228, + 0.004625992383807898, + 0.007405142765492201, + 0.016490167006850243, + 0.0061828456819057465, + 0.015425852499902248, + 0.008853538893163204, + 0.009156658314168453, + 0.011679383926093578, + 0.008563942275941372, + 0.008784779347479343, + 0.008481799624860287, + 0.009539676830172539, + 0.007299595046788454, + 0.012440472841262817, + 0.013764884322881699, + 0.014039723202586174, + 0.007417225278913975, + 0.006626021582633257, + 0.008040208369493484, + 0.008498091250658035, + 0.005409872159361839, + 0.019080566242337227, + 0.011309786699712276, + 0.01582634262740612, + 0.009577056393027306, + 0.010001564398407936, + 0.01624300517141819, + 0.006550907157361507, + 0.005854989401996136, + 0.00855198409408331, + 0.009308917447924614, + 0.009926860220730305, + 0.019902203232049942, + 0.008751996792852879, + 0.008645103313028812, + 0.012466001324355602, + 0.0059418086893856525, + 0.01148985605686903, + 0.008315766230225563, + 0.010207355953752995, + 0.01068747229874134, + 0.01250176690518856, + 0.00785159319639206, + 0.00822727382183075, + 0.007474295794963837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.0568295456469059 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a steady medium shot from a low-to-medium angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young player wearing a bright red short-sleeved jersey with an abstract white team crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player perform a powerful kick with his right leg towards a ball rolling at his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a well-maintained emerald green natural grass field with blurred silhouettes of other players in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the clear texture and color saturation typical of professional TV news broadcasts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009800649248063564 + ], + "opposite_logprob_tokens": [ + -4.634800434112549 + ], + "logprob_reward": -0.009800649248063564, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009800649248063564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18812872469425201, + 0.1164337620139122, + 0.11013897508382797, + 0.11396609991788864, + 0.06289853155612946, + 0.07931632548570633, + 0.11235206574201584, + 0.09277213364839554, + 0.05983784422278404, + 0.06415556371212006 + ], + "credit_map": [ + 0.01813633367419243, + 0.02499585784971714, + 0.02972913719713688, + 0.02066408470273018, + 0.036253590136766434, + 0.022933626547455788, + 0.018126709386706352, + 0.017289385199546814, + 0.014248346909880638, + 0.013679970055818558, + 0.010944006964564323, + 0.021217796951532364, + 0.01867973618209362, + 0.01207676064223051, + 0.016377124935388565, + 0.009210008196532726, + 0.008915559388697147, + 0.012592709623277187, + 0.017439013347029686, + 0.013301475904881954, + 0.015540005639195442, + 0.012039415538311005, + 0.013566127046942711, + 0.01674466021358967, + 0.007020315155386925, + 0.007630286738276482, + 0.011159609071910381, + 0.008964497596025467, + 0.011108018457889557, + 0.029079733416438103, + 0.013288918882608414, + 0.02571471594274044, + 0.011163393035531044, + 0.0036630702670663595, + 0.005802880506962538, + 0.002880630549043417, + 0.00482656667008996, + 0.016097726300358772, + 0.003907306119799614, + 0.014556952752172947, + 0.013056334108114243, + 0.010025796480476856, + 0.010776937939226627, + 0.009219574742019176, + 0.009581970050930977, + 0.008458951488137245, + 0.009746463969349861, + 0.008450295776128769, + 0.017651667818427086, + 0.014685857109725475, + 0.01610884815454483, + 0.010296429507434368, + 0.0168426726013422, + 0.012264694087207317, + 0.013896510936319828, + 0.010605378076434135, + 0.019277526065707207, + 0.009782595559954643, + 0.012180095538496971, + 0.012182345613837242, + 0.0089389868080616, + 0.016005894169211388, + 0.0072157722897827625, + 0.007188916206359863, + 0.007186318282037973, + 0.0077772303484380245, + 0.007253141142427921, + 0.010188836604356766, + 0.006027160212397575, + 0.008851466700434685, + 0.006579522509127855, + 0.005974162369966507, + 0.006085822358727455, + 0.00865932647138834, + 0.009929451160132885, + 0.012341773137450218, + 0.008334661833941936, + 0.006185219623148441, + 0.00681960629299283, + 0.005799697246402502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.009800649248063564 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a steady medium shot from a low-to-medium angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young player wearing a bright red short-sleeved jersey with an abstract white team crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player perform a powerful kick with his right leg towards a ball rolling at his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a well-maintained emerald green natural grass field with blurred silhouettes of other players in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the clear texture and color saturation typical of professional TV news broadcasts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22555458545684814 + ], + "opposite_logprob_tokens": [ + -1.6005545854568481 + ], + "logprob_reward": -0.22555458545684814, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22555458545684814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1513620913028717, + 0.12613125145435333, + 0.1184120699763298, + 0.08419893682003021, + 0.08448740094900131, + 0.03577079623937607, + 0.07939103990793228, + 0.19521883130073547, + 0.0683765560388565, + 0.05665097013115883 + ], + "credit_map": [ + 0.01572834886610508, + 0.027479469776153564, + 0.019781602546572685, + 0.01321815513074398, + 0.025020182132720947, + 0.0165794026106596, + 0.0209704227745533, + 0.01258451584726572, + 0.017704015597701073, + 0.016461489722132683, + 0.014135674573481083, + 0.01848820596933365, + 0.01964125595986843, + 0.013094504363834858, + 0.017428355291485786, + 0.009177758358418941, + 0.008793354034423828, + 0.014347193762660027, + 0.014536464586853981, + 0.009698240086436272, + 0.017119476571679115, + 0.01507830061018467, + 0.012636170722544193, + 0.026202870532870293, + 0.003691301681101322, + 0.0035537509247660637, + 0.007500263396650553, + 0.005332200322300196, + 0.0064612082205712795, + 0.025316864252090454, + 0.007618263829499483, + 0.024725085124373436, + 0.01716078259050846, + 0.004780096467584372, + 0.00884808786213398, + 0.004784596152603626, + 0.0059067439287900925, + 0.02101912721991539, + 0.004638300277292728, + 0.017349667847156525, + 0.006056483834981918, + 0.004200206138193607, + 0.005089231301099062, + 0.0031809103675186634, + 0.004231998231261969, + 0.0042746164835989475, + 0.004658474586904049, + 0.0040788729675114155, + 0.012127860449254513, + 0.010953863151371479, + 0.011811231262981892, + 0.00951145775616169, + 0.008196528069674969, + 0.007739327847957611, + 0.012337876483798027, + 0.0067128986120224, + 0.028868522495031357, + 0.021775636821985245, + 0.027605842798948288, + 0.02344958670437336, + 0.02353745698928833, + 0.0418110191822052, + 0.012995949015021324, + 0.015174824744462967, + 0.013116179034113884, + 0.010322623886168003, + 0.009234236553311348, + 0.005947829224169254, + 0.00615478353574872, + 0.011931967921555042, + 0.006076708436012268, + 0.005592228379100561, + 0.0054020662792027, + 0.008350121788680553, + 0.0070009902119636536, + 0.011582559905946255, + 0.006803636904805899, + 0.0048563843593001366, + 0.008204187266528606, + 0.004451022017747164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.22555458545684814 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a steady medium shot from a low-to-medium angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young player wearing a bright red short-sleeved jersey with an abstract white team crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player perform a powerful kick with his right leg towards a ball rolling at his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a well-maintained emerald green natural grass field with blurred silhouettes of other players in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the clear texture and color saturation typical of professional TV news broadcasts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007625873666256666 + ], + "opposite_logprob_tokens": [ + -4.882626056671143 + ], + "logprob_reward": -0.007625873666256666, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625873666256666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18108628690242767, + 0.126571387052536, + 0.12496262043714523, + 0.07926598936319351, + 0.0648263618350029, + 0.08553601801395416, + 0.09701066464185715, + 0.1205027848482132, + 0.06680811941623688, + 0.0534297339618206 + ], + "credit_map": [ + 0.0302482470870018, + 0.030439600348472595, + 0.03875225409865379, + 0.014826125465333462, + 0.015855124220252037, + 0.021940156817436218, + 0.012106127105653286, + 0.016918662935495377, + 0.03229966387152672, + 0.01770390197634697, + 0.013478679582476616, + 0.016823498532176018, + 0.015151728875935078, + 0.013219944201409817, + 0.010623466223478317, + 0.007270511239767075, + 0.014546067453920841, + 0.01548840757459402, + 0.02217702753841877, + 0.013192059472203255, + 0.014108750969171524, + 0.013912208378314972, + 0.014105787500739098, + 0.01743232272565365, + 0.00489063048735261, + 0.0075785075314342976, + 0.008245685137808323, + 0.008753527887165546, + 0.006461557000875473, + 0.01709875464439392, + 0.00920759979635477, + 0.01702973060309887, + 0.014020102098584175, + 0.004861216060817242, + 0.00572775537148118, + 0.00513296015560627, + 0.005934187676757574, + 0.013247482478618622, + 0.004635035991668701, + 0.01126762293279171, + 0.02019478753209114, + 0.009598464705049992, + 0.01810307614505291, + 0.005201044492423534, + 0.006659512408077717, + 0.01655811257660389, + 0.00420580618083477, + 0.005015222355723381, + 0.01627650111913681, + 0.015553530305624008, + 0.012559598311781883, + 0.009192480705678463, + 0.009447651915252209, + 0.013282613828778267, + 0.012103973887860775, + 0.008594325743615627, + 0.030876455828547478, + 0.009986198507249355, + 0.019453231245279312, + 0.00823843665421009, + 0.014530106447637081, + 0.021197890862822533, + 0.005671486724168062, + 0.010548993945121765, + 0.007410749793052673, + 0.01235189475119114, + 0.008098888210952282, + 0.01190823782235384, + 0.005676476750522852, + 0.007061653770506382, + 0.010692806914448738, + 0.003607418853789568, + 0.006863868795335293, + 0.005327839404344559, + 0.006926208268851042, + 0.0058310809545218945, + 0.009310026653110981, + 0.007012302055954933, + 0.004801018629223108, + 0.007357393857091665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.007625873666256666 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a steady medium shot from a low-to-medium angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young player wearing a bright red short-sleeved jersey with an abstract white team crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player perform a powerful kick with his right leg towards a ball rolling at his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a well-maintained emerald green natural grass field with blurred silhouettes of other players in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the clear texture and color saturation typical of professional TV news broadcasts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14275994896888733 + ], + "opposite_logprob_tokens": [ + -2.0177600383758545 + ], + "logprob_reward": -0.14275994896888733, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14275994896888733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29365038871765137, + 0.047088373452425, + 0.08550836145877838, + 0.05750438570976257, + 0.07739313691854477, + 0.07010024785995483, + 0.2032715380191803, + 0.04685189947485924, + 0.048664748668670654, + 0.06996696442365646 + ], + "credit_map": [ + 0.024760423228144646, + 0.055854085832834244, + 0.06785690784454346, + 0.024532800540328026, + 0.03303159773349762, + 0.04051687940955162, + 0.021701643243432045, + 0.025396045297384262, + 0.00768629927188158, + 0.005306710489094257, + 0.004511899780482054, + 0.0059067318215966225, + 0.007290243171155453, + 0.0053129252046346664, + 0.006167836487293243, + 0.004905726294964552, + 0.01080912072211504, + 0.009505826979875565, + 0.00859556533396244, + 0.010030854493379593, + 0.009506859816610813, + 0.009946837089955807, + 0.01159325335174799, + 0.015520037151873112, + 0.003575787879526615, + 0.005111475940793753, + 0.007248264737427235, + 0.008370071649551392, + 0.005584369879215956, + 0.009154052473604679, + 0.00874878466129303, + 0.009711572900414467, + 0.011793697252869606, + 0.007923251949250698, + 0.009589075110852718, + 0.0061278920620679855, + 0.00983133353292942, + 0.011739995330572128, + 0.0072607845067977905, + 0.013127101585268974, + 0.01338622160255909, + 0.006286982912570238, + 0.013084063306450844, + 0.006764544174075127, + 0.007140824105590582, + 0.007106001954525709, + 0.0090689891949296, + 0.0072626168839633465, + 0.03785504773259163, + 0.022948643192648888, + 0.03151547908782959, + 0.02235504612326622, + 0.020368020981550217, + 0.02803330309689045, + 0.023621128872036934, + 0.01657485030591488, + 0.007160812150686979, + 0.006677642464637756, + 0.005953971296548843, + 0.006037464831024408, + 0.0048643662594258785, + 0.007343512028455734, + 0.004527753219008446, + 0.004286373965442181, + 0.005848235916346312, + 0.008170862682163715, + 0.005044679157435894, + 0.008274869993329048, + 0.005391934420913458, + 0.005666996818035841, + 0.007671336643397808, + 0.002595824422314763, + 0.010383677668869495, + 0.005721413064748049, + 0.010477256961166859, + 0.008341862820088863, + 0.011390012688934803, + 0.00871338602155447, + 0.007445949595421553, + 0.007493398617953062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.14275994896888733 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady medium shot, shot from a low-to-medium angle, captured under the bright, soft overhead lighting of the stadium, perfectly captures the ecstatic moment of a player celebrating a goal during a fierce Ultimate Frisbee tournament. The young player, his dark brown curls draped against his forehead, his face radiating pure excitement and triumph, his eyes wide with ecstasy, his lips stretched into a brilliant grin. He wears a bright red, short-sleeved jersey made of moisture-wicking polyester that gleams subtly in the light. The jersey features an abstract white team crest pattern on the chest, paired with matching white shorts and knee-high red socks. Having just scored, he lowers his weight, swings his right leg back, and with a powerful, powerful kick, he unleashes a powerful force towards a ball that rolls right at his feet. Grass dust flies into the air, perfectly capturing the climax of his unleashed passion. Beneath his feet is a well-maintained, emerald green natural grass field. In the background, the silhouettes of other players in different jerseys are blurred by the depth of field. The entire picture is saturated in color and full of the clear texture of professional TV news broadcasts, accurately conveying the high-energy and contagious celebratory atmosphere in competitive sports.", + "seed": 22913, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_13/index_3937_seed_22913_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0810714865103364, + "vlm_logprob": -0.0810714865103364, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.0810714865103364, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a steady medium shot from a low-to-medium angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young player wearing a bright red short-sleeved jersey with an abstract white team crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player perform a powerful kick with his right leg towards a ball rolling at his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a well-maintained emerald green natural grass field with blurred silhouettes of other players in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the clear texture and color saturation typical of professional TV news broadcasts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07151862978935242 + ], + "opposite_logprob_tokens": [ + -3.821518659591675 + ], + "logprob_reward": -0.07151862978935242, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.07151862978935242, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10266541689634323, + 0.05951966345310211, + 0.052946917712688446, + 0.12497220933437347, + 0.20652933418750763, + 0.07483112066984177, + 0.19051219522953033, + 0.04960351437330246, + 0.057117778807878494, + 0.08130192011594772 + ], + "credit_map": [ + 0.01967943087220192, + 0.010852274484932423, + 0.0133442971855402, + 0.010682706721127033, + 0.010903487913310528, + 0.012697113677859306, + 0.012131169438362122, + 0.012374920770525932, + 0.0071868556551635265, + 0.011825037188827991, + 0.008952239528298378, + 0.006161823403090239, + 0.007456577382981777, + 0.006270244251936674, + 0.0057083070278167725, + 0.005958572030067444, + 0.005909518804401159, + 0.007147420197725296, + 0.0052935900166630745, + 0.008299215696752071, + 0.006509358063340187, + 0.00743047846481204, + 0.007605828810483217, + 0.004751496948301792, + 0.01675432361662388, + 0.017674008384346962, + 0.013464514166116714, + 0.026470065116882324, + 0.013008426874876022, + 0.00983086321502924, + 0.0177791565656662, + 0.00999082624912262, + 0.025352006778120995, + 0.038454972207546234, + 0.031179318204522133, + 0.03254283219575882, + 0.015507594682276249, + 0.02050502598285675, + 0.028375515714287758, + 0.014612039551138878, + 0.00974384043365717, + 0.013974306173622608, + 0.009941201657056808, + 0.0077828895300626755, + 0.009055855683982372, + 0.009005745872855186, + 0.008886427618563175, + 0.006440846249461174, + 0.03363511338829994, + 0.02904004603624344, + 0.018038330599665642, + 0.023111887276172638, + 0.027157258242368698, + 0.02089785598218441, + 0.024111149832606316, + 0.014520535245537758, + 0.005131513345986605, + 0.005340766627341509, + 0.007908659987151623, + 0.008199633099138737, + 0.006192030385136604, + 0.0038331043906509876, + 0.007046019192785025, + 0.00595178222283721, + 0.007993624545633793, + 0.005687136203050613, + 0.006779870949685574, + 0.007898119278252125, + 0.0077756354585289955, + 0.005130326841026545, + 0.009442189708352089, + 0.0064108651131391525, + 0.01004057563841343, + 0.008523765951395035, + 0.01413126289844513, + 0.007447527255862951, + 0.010595811530947685, + 0.012704703025519848, + 0.008690038695931435, + 0.009168227203190327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 3 + ], + "selected_reward": -0.07151862978935242 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a steady medium shot from a low-to-medium angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young player wearing a bright red short-sleeved jersey with an abstract white team crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player perform a powerful kick with his right leg towards a ball rolling at his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a well-maintained emerald green natural grass field with blurred silhouettes of other players in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the clear texture and color saturation typical of professional TV news broadcasts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012556779198348522 + ], + "opposite_logprob_tokens": [ + -4.387556552886963 + ], + "logprob_reward": -0.012556779198348522, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012556779198348522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08839984983205795, + 0.05755815654993057, + 0.07580293715000153, + 0.14653798937797546, + 0.11359920352697372, + 0.10225261002779007, + 0.15030574798583984, + 0.11453510075807571, + 0.05458313599228859, + 0.09642525762319565 + ], + "credit_map": [ + 0.01347836758941412, + 0.009854979813098907, + 0.01262255385518074, + 0.009499640204012394, + 0.010515792295336723, + 0.01181202381849289, + 0.010834215208888054, + 0.009782275184988976, + 0.005772503558546305, + 0.008252185769379139, + 0.009864681400358677, + 0.00978297833353281, + 0.005761958658695221, + 0.004858146421611309, + 0.0071229576133191586, + 0.0061427452601492405, + 0.008698689751327038, + 0.00963024515658617, + 0.010005874559283257, + 0.012712739408016205, + 0.00604691868647933, + 0.009334193542599678, + 0.012579179368913174, + 0.006795097142457962, + 0.01287499163299799, + 0.026076778769493103, + 0.027832413092255592, + 0.02551264874637127, + 0.014341053552925587, + 0.010861294344067574, + 0.017112450674176216, + 0.011926361359655857, + 0.010501161217689514, + 0.0233203936368227, + 0.017201926559209824, + 0.018160035833716393, + 0.009031764231622219, + 0.012292783707380295, + 0.014717469923198223, + 0.008373663760721684, + 0.013012077659368515, + 0.01910911127924919, + 0.011141511611640453, + 0.01147605199366808, + 0.011020811274647713, + 0.018917866051197052, + 0.008918351493775845, + 0.008656825870275497, + 0.024119434878230095, + 0.021811988204717636, + 0.011782755143940449, + 0.01726558804512024, + 0.02407151460647583, + 0.019618645310401917, + 0.01942361891269684, + 0.012212199158966541, + 0.014873848296701908, + 0.014630584977567196, + 0.01828371360898018, + 0.02359590493142605, + 0.007932085543870926, + 0.006621555425226688, + 0.014421751722693443, + 0.01417565904557705, + 0.007007339969277382, + 0.005762831307947636, + 0.007510967552661896, + 0.006465660408139229, + 0.008959393948316574, + 0.00428170058876276, + 0.008762967772781849, + 0.005832275841385126, + 0.012139846570789814, + 0.011189878918230534, + 0.008701424114406109, + 0.013981151394546032, + 0.015322026796638966, + 0.013021493330597878, + 0.010329707525670528, + 0.011739730834960938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 7 + ], + "selected_reward": -0.012556779198348522 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a steady medium shot from a low-to-medium angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young player wearing a bright red short-sleeved jersey with an abstract white team crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player perform a powerful kick with his right leg towards a ball rolling at his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a well-maintained emerald green natural grass field with blurred silhouettes of other players in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the clear texture and color saturation typical of professional TV news broadcasts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2255527824163437 + ], + "opposite_logprob_tokens": [ + -1.6005527973175049 + ], + "logprob_reward": -0.2255527824163437, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255527824163437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07748033851385117, + 0.053614355623722076, + 0.05141512304544449, + 0.05491911247372627, + 0.048549603670835495, + 0.12479902058839798, + 0.25684109330177307, + 0.13328857719898224, + 0.1006418764591217, + 0.09845084697008133 + ], + "credit_map": [ + 0.0085506746545434, + 0.00940923672169447, + 0.012462826445698738, + 0.011573286727070808, + 0.00987772922962904, + 0.007754762656986713, + 0.008770042099058628, + 0.009081784635782242, + 0.007483664434403181, + 0.007138291839510202, + 0.007967697456479073, + 0.005536717362701893, + 0.005471186246722937, + 0.007915140129625797, + 0.00675971107557416, + 0.005341947544366121, + 0.006554819643497467, + 0.008874326944351196, + 0.006841578055173159, + 0.006964625790715218, + 0.005153720732778311, + 0.005363918840885162, + 0.007371750194579363, + 0.004290382843464613, + 0.006723434664309025, + 0.008718149736523628, + 0.007333791349083185, + 0.00816118810325861, + 0.006047687027603388, + 0.006448295898735523, + 0.006907543167471886, + 0.004579024389386177, + 0.005180904641747475, + 0.008714298717677593, + 0.011108982376754284, + 0.005246651824563742, + 0.0051860809326171875, + 0.003574984846636653, + 0.005520306061953306, + 0.004017395433038473, + 0.01622319407761097, + 0.026604922488331795, + 0.014052500016987324, + 0.01341177523136139, + 0.012082807719707489, + 0.019271375611424446, + 0.011900067329406738, + 0.011252372525632381, + 0.033440858125686646, + 0.04230010136961937, + 0.030418576672673225, + 0.035243920981884, + 0.025089100003242493, + 0.02146448753774166, + 0.04347991943359375, + 0.025404153391718864, + 0.011918481439352036, + 0.013444102369248867, + 0.021772000938653946, + 0.022259937599301338, + 0.011327619664371014, + 0.007158943917602301, + 0.026387104764580727, + 0.01902039907872677, + 0.013706809841096401, + 0.01015140675008297, + 0.01667582057416439, + 0.013520277105271816, + 0.013911706395447254, + 0.007346927188336849, + 0.013656138442456722, + 0.011672792956233025, + 0.010255230590701103, + 0.01029515266418457, + 0.006309944670647383, + 0.015092603862285614, + 0.019569717347621918, + 0.012627333402633667, + 0.01108876895159483, + 0.01321209967136383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.2255527824163437 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a steady medium shot from a low-to-medium angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young player wearing a bright red short-sleeved jersey with an abstract white team crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player perform a powerful kick with his right leg towards a ball rolling at his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a well-maintained emerald green natural grass field with blurred silhouettes of other players in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the clear texture and color saturation typical of professional TV news broadcasts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067326370626688 + ], + "opposite_logprob_tokens": [ + -5.00673246383667 + ], + "logprob_reward": -0.0067326370626688, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067326370626688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09862873703241348, + 0.09036143869161606, + 0.12666501104831696, + 0.11098019033670425, + 0.13772805035114288, + 0.11133618652820587, + 0.11918164789676666, + 0.09386701881885529, + 0.06361976265907288, + 0.04763198271393776 + ], + "credit_map": [ + 0.013122307136654854, + 0.009227077476680279, + 0.0150977848097682, + 0.011962881311774254, + 0.00863889791071415, + 0.017338251695036888, + 0.009812958538532257, + 0.01342857163399458, + 0.009092372842133045, + 0.011362803168594837, + 0.01878780499100685, + 0.010076554492115974, + 0.00880828034132719, + 0.01591379940509796, + 0.00795187707990408, + 0.00836794450879097, + 0.010852813720703125, + 0.015509814955294132, + 0.012421704828739166, + 0.017542116343975067, + 0.018472449854016304, + 0.02719428390264511, + 0.013646237552165985, + 0.011025596410036087, + 0.016955703496932983, + 0.014076459221541882, + 0.011993315070867538, + 0.01828686334192753, + 0.0132706044241786, + 0.013651612214744091, + 0.014126932248473167, + 0.008618700318038464, + 0.013686086051166058, + 0.011982104741036892, + 0.04071632772684097, + 0.017696676775813103, + 0.010236863046884537, + 0.01493464782834053, + 0.01605258695781231, + 0.01242275070399046, + 0.009095191955566406, + 0.012626124545931816, + 0.010936143808066845, + 0.008959031663835049, + 0.012258102186024189, + 0.036492008715867996, + 0.010471315123140812, + 0.01049826480448246, + 0.01641400344669819, + 0.012339712120592594, + 0.010779980570077896, + 0.014328019693493843, + 0.02122356742620468, + 0.014864985831081867, + 0.019280191510915756, + 0.009951191022992134, + 0.009596304036676884, + 0.011081521399319172, + 0.013820084743201733, + 0.016768507659435272, + 0.007139556109905243, + 0.006520973052829504, + 0.016470840200781822, + 0.012469232082366943, + 0.0068139610812067986, + 0.00547824427485466, + 0.007359279319643974, + 0.00961767602711916, + 0.012921749614179134, + 0.003918062429875135, + 0.010579594410955906, + 0.006931192707270384, + 0.00654546357691288, + 0.00540029164403677, + 0.005552253685891628, + 0.007535744924098253, + 0.005867534317076206, + 0.007375702727586031, + 0.004637011792510748, + 0.004717981442809105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 6 + ], + "selected_reward": -0.0067326370626688 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a steady medium shot from a low-to-medium angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young player wearing a bright red short-sleeved jersey with an abstract white team crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player perform a powerful kick with his right leg towards a ball rolling at his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a well-maintained emerald green natural grass field with blurred silhouettes of other players in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the clear texture and color saturation typical of professional TV news broadcasts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08899660408496857 + ], + "opposite_logprob_tokens": [ + -2.463996648788452 + ], + "logprob_reward": -0.08899660408496857, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08899660408496857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20357826352119446, + 0.12347758561372757, + 0.14966432750225067, + 0.04441078379750252, + 0.0593285895884037, + 0.1250317245721817, + 0.06716327369213104, + 0.13653817772865295, + 0.055693257600069046, + 0.035114090889692307 + ], + "credit_map": [ + 0.017957469448447227, + 0.0320298932492733, + 0.026118818670511246, + 0.025665132328867912, + 0.03801187127828598, + 0.018859822303056717, + 0.022666536271572113, + 0.02226869761943817, + 0.011324158869683743, + 0.022390905767679214, + 0.01876438781619072, + 0.01521794218569994, + 0.016605857759714127, + 0.012279248796403408, + 0.015179001726210117, + 0.011716071516275406, + 0.014479061588644981, + 0.024802125990390778, + 0.019154516980051994, + 0.026256969198584557, + 0.010837048292160034, + 0.019495144486427307, + 0.021669059991836548, + 0.012970393523573875, + 0.00765650300309062, + 0.005875854287296534, + 0.005164023954421282, + 0.006872051861137152, + 0.005236227065324783, + 0.005462064407765865, + 0.005107402335852385, + 0.0030366512946784496, + 0.005192921031266451, + 0.012345795519649982, + 0.010225933976471424, + 0.008999736048281193, + 0.0042425477877259254, + 0.007055572234094143, + 0.006880829576402903, + 0.004385247826576233, + 0.019126802682876587, + 0.020266752690076828, + 0.016337214037775993, + 0.014632084406912327, + 0.014495345763862133, + 0.016509387642145157, + 0.013504245318472385, + 0.010159879922866821, + 0.009815207682549953, + 0.007650367915630341, + 0.0055175283923745155, + 0.008660485967993736, + 0.010840307921171188, + 0.00822429172694683, + 0.01083555817604065, + 0.005619521252810955, + 0.01382923312485218, + 0.01521796453744173, + 0.02373228408396244, + 0.023210112005472183, + 0.00987526960670948, + 0.006540124304592609, + 0.02478465624153614, + 0.019348526373505592, + 0.007458470296114683, + 0.0051640779711306095, + 0.007813955657184124, + 0.010190194472670555, + 0.006253920495510101, + 0.0031980276107788086, + 0.009407103061676025, + 0.0062075029127299786, + 0.0033908679615706205, + 0.0030858370009809732, + 0.004496684763580561, + 0.003340585622936487, + 0.00385658023878932, + 0.008949756622314453, + 0.003673752536997199, + 0.004320025444030762 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.08899660408496857 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_14/index_tensor([908, 908])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_14/index_tensor([908, 908])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0fcd39ff6f72042ddbc1fc8cae71be7718bf3b62 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_14/index_tensor([908, 908])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium shot captures a young Southeast Asian man in his early twenties, filled with infectious energy, enthusiastically explaining a complex idea to an unseen person off-camera. He has short, tousled black hair, and his expressive dark eyes are wide and excited. Wearing a simple, fitted olive green cotton T-shirt, he leans slightly forward at the waist, his body language focused. His hands are in constant, dynamic motion; one moment they dart through the air to emphasize a critical point, the next, his fingers carefully trace an invisible diagram, shaping abstract concepts with precise, clear gestures. The scene is set in a bright, modern creative studio, with a large whiteboard covered with a complex flowchart, and colorful sticky notes visible against a softly blurred background. The space is illuminated by clean, diffuse light from large, unseen windows, creating a vibrant, high-key look with minimal shadows. The atmosphere is filled with intellectual curiosity and creative enthusiasm, completely focused on the man's vibrant and engaging expression of his ideas.", + "seed": 9815, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_14/index_908_seed_9815_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.012634895183146, + "vlm_logprob": -0.012634895183146, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.012634895183146, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with short, tousled black hair and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a fitted olive green cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning slightly forward at the waist with hands in dynamic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright modern studio with a large whiteboard covered in a flowchart and colorful sticky notes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with clean, diffuse lighting and minimal shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009569015353918076 + ], + "opposite_logprob_tokens": [ + -4.75956916809082 + ], + "logprob_reward": -0.009569015353918076, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.009569015353918076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3660648465156555, + 0.03836991265416145, + 0.07431318610906601, + 0.3516695499420166, + 0.041735097765922546, + 0.05861731991171837, + 0.014654294587671757, + 0.0135547686368227, + 0.010424240492284298, + 0.030596816912293434 + ], + "credit_map": [ + 0.03755055367946625, + 0.02169586904346943, + 0.11436427384614944, + 0.043257392942905426, + 0.016709398478269577, + 0.08774520456790924, + 0.021931860595941544, + 0.022810306400060654, + 0.003766287351027131, + 0.0028459387831389904, + 0.009386658668518066, + 0.005629493854939938, + 0.002289633499458432, + 0.007194594480097294, + 0.0034936578013002872, + 0.00376364728435874, + 0.006360458675771952, + 0.008650831878185272, + 0.022499429062008858, + 0.005141434259712696, + 0.002359258709475398, + 0.02230261079967022, + 0.0025313235819339752, + 0.004467833787202835, + 0.019878380000591278, + 0.005737427156418562, + 0.15463687479496002, + 0.011062443256378174, + 0.005058960989117622, + 0.1372642070055008, + 0.01034259982407093, + 0.00768864993005991, + 0.006755701266229153, + 0.005588769447058439, + 0.006426359061151743, + 0.004202749114483595, + 0.0045796106569468975, + 0.004722646903246641, + 0.0030100455041974783, + 0.006449214182794094, + 0.004971116315573454, + 0.007209744770079851, + 0.008091400377452374, + 0.0056845275685191154, + 0.007309610955417156, + 0.008117482997477055, + 0.0060149189084768295, + 0.011218524537980556, + 0.0018076871056109667, + 0.0019506757380440831, + 0.0020574363879859447, + 0.0015518513973802328, + 0.0008024905691854656, + 0.00218620034866035, + 0.0015162222553044558, + 0.00278173154219985, + 0.0013269344344735146, + 0.0011139900889247656, + 0.0022766494657844305, + 0.002568020485341549, + 0.0013275790261104703, + 0.001769705326296389, + 0.0010803609620779753, + 0.0020915295463055372, + 0.0016348888166248798, + 0.0011009757872670889, + 0.0021732081659138203, + 0.00109678425360471, + 0.0007953650783747435, + 0.001826556632295251, + 0.0007162953843362629, + 0.0010801659664139152, + 0.0027988972142338753, + 0.0038663705345243216, + 0.008616812527179718, + 0.002363501349464059, + 0.0011288865935057402, + 0.007798703387379646, + 0.0013446062803268433, + 0.0026790392585098743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.009569015353918076 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with short, tousled black hair and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a fitted olive green cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning slightly forward at the waist with hands in dynamic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright modern studio with a large whiteboard covered in a flowchart and colorful sticky notes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with clean, diffuse lighting and minimal shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025032395496964455 + ], + "opposite_logprob_tokens": [ + -6.002503395080566 + ], + "logprob_reward": -0.0025032395496964455, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025032395496964455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2306305468082428, + 0.048276904970407486, + 0.09715761244297028, + 0.41858723759651184, + 0.026204289868474007, + 0.03574681282043457, + 0.04285481572151184, + 0.012688107788562775, + 0.047759439796209335, + 0.04009423404932022 + ], + "credit_map": [ + 0.01877181977033615, + 0.011218373663723469, + 0.07762722671031952, + 0.02584819123148918, + 0.009743897244334221, + 0.06086266040802002, + 0.01385408639907837, + 0.012704281136393547, + 0.003457710612565279, + 0.0047376626171171665, + 0.011293678544461727, + 0.006188804283738136, + 0.0029258897993713617, + 0.009483943693339825, + 0.00332343066111207, + 0.006865780800580978, + 0.009776974096894264, + 0.016276389360427856, + 0.0256477203220129, + 0.010080463252961636, + 0.0038574913050979376, + 0.024043360725045204, + 0.003230280941352248, + 0.004244924057275057, + 0.02362896502017975, + 0.006686374545097351, + 0.1839759200811386, + 0.01324037741869688, + 0.006035896949470043, + 0.16297435760498047, + 0.012341412715613842, + 0.009703896939754486, + 0.0029705213382840157, + 0.002741969656199217, + 0.0035718055441975594, + 0.004168298561125994, + 0.002409650245681405, + 0.002394802635535598, + 0.00208395067602396, + 0.005863288417458534, + 0.0033851610496640205, + 0.003552403999492526, + 0.00559667544439435, + 0.0051004826091229916, + 0.004986083135008812, + 0.005038144998252392, + 0.004522600211203098, + 0.0035652562510222197, + 0.0059946742840111256, + 0.00277923047542572, + 0.009987120516598225, + 0.003943721763789654, + 0.002074148738756776, + 0.006757892202585936, + 0.0020481969695538282, + 0.009269827976822853, + 0.0015762075781822205, + 0.002935819560661912, + 0.0016008198726922274, + 0.0019295569509267807, + 0.0012479015858843923, + 0.001517460448667407, + 0.0008414885378442705, + 0.001038852846249938, + 0.005163924768567085, + 0.002902982523664832, + 0.016362424939870834, + 0.002676088362932205, + 0.0020343533251434565, + 0.012643559835851192, + 0.002188988495618105, + 0.003787111258134246, + 0.0039986418560147285, + 0.00674749119207263, + 0.008677673526108265, + 0.003907338250428438, + 0.0019373685354366899, + 0.00755638163536787, + 0.002032822696492076, + 0.005236513912677765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.0025032395496964455 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with short, tousled black hair and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a fitted olive green cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning slightly forward at the waist with hands in dynamic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright modern studio with a large whiteboard covered in a flowchart and colorful sticky notes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with clean, diffuse lighting and minimal shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369694948196411 + ], + "opposite_logprob_tokens": [ + -3.4086968898773193 + ], + "logprob_reward": -0.03369694948196411, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369694948196411, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2580071985721588, + 0.05386838689446449, + 0.09584468603134155, + 0.22128309309482574, + 0.07644671201705933, + 0.02005663700401783, + 0.1861613541841507, + 0.04923254996538162, + 0.01351858489215374, + 0.025580808520317078 + ], + "credit_map": [ + 0.025464581325650215, + 0.011528081260621548, + 0.08824142813682556, + 0.02619139663875103, + 0.009327389299869537, + 0.07008450478315353, + 0.011600502766668797, + 0.01556929387152195, + 0.0044332509860396385, + 0.0036316984333097935, + 0.016564149409532547, + 0.004132658708840609, + 0.002302769338712096, + 0.014650216326117516, + 0.0031528875697404146, + 0.005000753328204155, + 0.010380890220403671, + 0.01845708116889, + 0.02233557030558586, + 0.011409194208681583, + 0.003826049156486988, + 0.02100328356027603, + 0.003307949285954237, + 0.005124668590724468, + 0.012184139341115952, + 0.004526617471128702, + 0.09626144170761108, + 0.006992610637098551, + 0.0034008987713605165, + 0.08526923507452011, + 0.006572359707206488, + 0.006075803656131029, + 0.008850190788507462, + 0.009510033763945103, + 0.009092666208744049, + 0.0050367615185678005, + 0.008358906954526901, + 0.007830544374883175, + 0.004995774943381548, + 0.02277182601392269, + 0.0019086896209046245, + 0.0028191276360303164, + 0.0024791874457150698, + 0.002632120857015252, + 0.002261138753965497, + 0.002420030767098069, + 0.0017913037445396185, + 0.0037450373638421297, + 0.026638921350240707, + 0.022312993183732033, + 0.037344034761190414, + 0.020773788914084435, + 0.009408639743924141, + 0.027671528980135918, + 0.010145909152925014, + 0.031865548342466354, + 0.004431419540196657, + 0.004004405811429024, + 0.00797751173377037, + 0.009049240499734879, + 0.0052068899385631084, + 0.006430318579077721, + 0.004621336702257395, + 0.007511429954320192, + 0.0026070973835885525, + 0.0015763501869514585, + 0.002229332458227873, + 0.0012512741377577186, + 0.0011002675164490938, + 0.001959925750270486, + 0.0008046007715165615, + 0.0019897366873919964, + 0.002568712690845132, + 0.004248267505317926, + 0.00597304617986083, + 0.0022544057574123144, + 0.0010298402048647404, + 0.005826765205711126, + 0.0009951782412827015, + 0.0026845920365303755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.03369694948196411 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with short, tousled black hair and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a fitted olive green cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning slightly forward at the waist with hands in dynamic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright modern studio with a large whiteboard covered in a flowchart and colorful sticky notes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with clean, diffuse lighting and minimal shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978058110922575 + ], + "opposite_logprob_tokens": [ + -4.634780406951904 + ], + "logprob_reward": -0.00978058110922575, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978058110922575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11751987040042877, + 0.08144344389438629, + 0.028828714042901993, + 0.37458494305610657, + 0.0281854048371315, + 0.03460121154785156, + 0.12875071167945862, + 0.020208489149808884, + 0.051332853734493256, + 0.1345442682504654 + ], + "credit_map": [ + 0.011246826499700546, + 0.01005993690341711, + 0.03122914396226406, + 0.014491686597466469, + 0.005607023369520903, + 0.027727549895644188, + 0.0068541946820914745, + 0.010303527116775513, + 0.006763458251953125, + 0.0036491011269390583, + 0.03156861290335655, + 0.004214631859213114, + 0.0023582533467561007, + 0.02602166123688221, + 0.0024397193919867277, + 0.004428007174283266, + 0.0037022328469902277, + 0.0019792162347584963, + 0.006251767743378878, + 0.002836039289832115, + 0.0015986151993274689, + 0.005127453710883856, + 0.002170213731005788, + 0.005163178779184818, + 0.020968381315469742, + 0.006781811360269785, + 0.16446250677108765, + 0.011794311925768852, + 0.005289044231176376, + 0.14524033665657043, + 0.01105741411447525, + 0.008991189301013947, + 0.003807787550613284, + 0.002987945219501853, + 0.0035945537965744734, + 0.002283765934407711, + 0.003736726241186261, + 0.0036537128034979105, + 0.001957837026566267, + 0.006163078360259533, + 0.00796293281018734, + 0.005545319523662329, + 0.006524485070258379, + 0.0023917066864669323, + 0.0015533702680841088, + 0.003355900989845395, + 0.0016576661728322506, + 0.005609835498034954, + 0.019332917407155037, + 0.007627324666827917, + 0.03214031457901001, + 0.011088610626757145, + 0.004996661562472582, + 0.02013031393289566, + 0.006624996662139893, + 0.026809580624103546, + 0.002931424882262945, + 0.002813252154737711, + 0.004290812183171511, + 0.0028098805341869593, + 0.001166504924185574, + 0.002552007557824254, + 0.001194360083900392, + 0.002450248459354043, + 0.004776190035045147, + 0.0026533419732004404, + 0.0196473840624094, + 0.0018148033414036036, + 0.0014921448891982436, + 0.015601695515215397, + 0.0014574433444067836, + 0.0038898559287190437, + 0.011293409392237663, + 0.01902219094336033, + 0.03361601382493973, + 0.010611157864332199, + 0.0045225732028484344, + 0.03268969804048538, + 0.004820025060325861, + 0.01796921342611313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 6 + ], + "selected_reward": -0.00978058110922575 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with short, tousled black hair and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a fitted olive green cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning slightly forward at the waist with hands in dynamic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright modern studio with a large whiteboard covered in a flowchart and colorful sticky notes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with clean, diffuse lighting and minimal shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007624690420925617 + ], + "opposite_logprob_tokens": [ + -4.882624626159668 + ], + "logprob_reward": -0.007624690420925617, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007624690420925617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3261999189853668, + 0.0495033785700798, + 0.11552560329437256, + 0.26792266964912415, + 0.06282912939786911, + 0.02632666565477848, + 0.09700468182563782, + 0.018274564296007156, + 0.01970246061682701, + 0.01671086624264717 + ], + "credit_map": [ + 0.031536493450403214, + 0.014375757426023483, + 0.11046546697616577, + 0.03244095668196678, + 0.011641758494079113, + 0.08773447573184967, + 0.016111696138978004, + 0.02189333550632, + 0.005865834187716246, + 0.004933551419526339, + 0.012303785420954227, + 0.005402257665991783, + 0.002826901152729988, + 0.010645413771271706, + 0.002684761770069599, + 0.004840872250497341, + 0.009992117993533611, + 0.01522032916545868, + 0.03423216566443443, + 0.010566048324108124, + 0.003861445700749755, + 0.03229397535324097, + 0.0038959842640906572, + 0.005463543813675642, + 0.01547746080905199, + 0.006274063140153885, + 0.11449914425611496, + 0.008751156739890575, + 0.004748665727674961, + 0.10232419520616531, + 0.00792940054088831, + 0.007918569259345531, + 0.013023940846323967, + 0.009280893951654434, + 0.007663148455321789, + 0.006092321593314409, + 0.005198780447244644, + 0.0062082200311124325, + 0.004645772743970156, + 0.010716056451201439, + 0.0032803004141896963, + 0.004536972381174564, + 0.0038610275369137526, + 0.0030306035187095404, + 0.0029450673609972, + 0.0027407188899815083, + 0.0020956837106496096, + 0.003836290678009391, + 0.013694962486624718, + 0.014110765419900417, + 0.02030562236905098, + 0.01045124139636755, + 0.0038528365548700094, + 0.013557963073253632, + 0.005539844743907452, + 0.015491453930735588, + 0.002093763556331396, + 0.0016346571501344442, + 0.0035442819353193045, + 0.002348890993744135, + 0.0015636367024853826, + 0.0024094937834888697, + 0.0016235668445006013, + 0.003056274726986885, + 0.0036976775154471397, + 0.003350241342559457, + 0.0026330375112593174, + 0.0020046497229486704, + 0.0010611288016662002, + 0.0033856064546853304, + 0.0010513461893424392, + 0.0025187721475958824, + 0.0027922755107283592, + 0.0013605807907879353, + 0.0034490518737584352, + 0.0012541444739326835, + 0.0008168910280801356, + 0.0026542923878878355, + 0.000773642968852073, + 0.003609985811635852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.007624690420925617 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures a young Southeast Asian man in his early twenties, filled with infectious energy, enthusiastically explaining a complex idea to an unseen person off-camera. He has short, tousled black hair, and his expressive dark eyes are wide and excited. Wearing a simple, fitted olive green cotton T-shirt, he leans slightly forward at the waist, his body language focused. His hands are in constant, dynamic motion; one moment they dart through the air to emphasize a critical point, the next, his fingers carefully trace an invisible diagram, shaping abstract concepts with precise, clear gestures. The scene is set in a bright, modern creative studio, with a large whiteboard covered with a complex flowchart, and colorful sticky notes visible against a softly blurred background. The space is illuminated by clean, diffuse light from large, unseen windows, creating a vibrant, high-key look with minimal shadows. The atmosphere is filled with intellectual curiosity and creative enthusiasm, completely focused on the man's vibrant and engaging expression of his ideas.", + "seed": 9816, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_14/index_908_seed_9816_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.010197416553273797, + "vlm_logprob": -0.010197416553273797, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.010197416553273797, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with short, tousled black hair and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a fitted olive green cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning slightly forward at the waist with hands in dynamic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright modern studio with a large whiteboard covered in a flowchart and colorful sticky notes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with clean, diffuse lighting and minimal shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010135963559150696 + ], + "opposite_logprob_tokens": [ + -4.885136127471924 + ], + "logprob_reward": -0.010135963559150696, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.010135963559150696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22970722615718842, + 0.07304371893405914, + 0.051089320331811905, + 0.21243897080421448, + 0.08314686268568039, + 0.05829343572258949, + 0.07437402009963989, + 0.06049622595310211, + 0.12418848276138306, + 0.03322174400091171 + ], + "credit_map": [ + 0.0646810457110405, + 0.014972064644098282, + 0.02003948949277401, + 0.015931759029626846, + 0.012697727419435978, + 0.04553981497883797, + 0.013801032677292824, + 0.042044274508953094, + 0.006829108577221632, + 0.004096399527043104, + 0.008765317499637604, + 0.007557213772088289, + 0.005091104190796614, + 0.01252810750156641, + 0.007943869568407536, + 0.020232589915394783, + 0.006859055254608393, + 0.003897076239809394, + 0.005318529438227415, + 0.007641524076461792, + 0.0033120017033070326, + 0.007755477447062731, + 0.005101029295474291, + 0.01120462641119957, + 0.03843580558896065, + 0.01909882016479969, + 0.018007494509220123, + 0.010657493956387043, + 0.006748370360583067, + 0.060454435646533966, + 0.011351373977959156, + 0.04768514633178711, + 0.009316920302808285, + 0.007075537461787462, + 0.014032833278179169, + 0.01130001526325941, + 0.008528822101652622, + 0.011278209276497364, + 0.007142895832657814, + 0.0144716277718544, + 0.0062720575369894505, + 0.003406299976631999, + 0.005074681714177132, + 0.004816844593733549, + 0.0028283407445997, + 0.011339112184941769, + 0.00487964041531086, + 0.0196764525026083, + 0.013568500988185406, + 0.006088310852646828, + 0.00915178470313549, + 0.005668588913977146, + 0.0056706261821091175, + 0.010952385142445564, + 0.008870873600244522, + 0.014402946457266808, + 0.007359842304140329, + 0.005836278200149536, + 0.007390299811959267, + 0.006576182320713997, + 0.005334035027772188, + 0.008983933366835117, + 0.008010165765881538, + 0.01100548543035984, + 0.029828384518623352, + 0.0065176524221897125, + 0.01120650302618742, + 0.006363588385283947, + 0.007056100759655237, + 0.023347588256001472, + 0.014282072894275188, + 0.02558659389615059, + 0.0029543659184128046, + 0.0035151459742337465, + 0.005445542745292187, + 0.003629641141742468, + 0.0021410081535577774, + 0.004910297691822052, + 0.0036347515415400267, + 0.006990986410528421 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.010135963559150696 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with short, tousled black hair and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a fitted olive green cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning slightly forward at the waist with hands in dynamic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright modern studio with a large whiteboard covered in a flowchart and colorful sticky notes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with clean, diffuse lighting and minimal shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003638672875240445 + ], + "opposite_logprob_tokens": [ + -5.628638744354248 + ], + "logprob_reward": -0.003638672875240445, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003638672875240445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4198034703731537, + 0.06314089894294739, + 0.03897077590227127, + 0.1426001340150833, + 0.048355866223573685, + 0.05708374083042145, + 0.025641711428761482, + 0.08659320324659348, + 0.05591704323887825, + 0.061893172562122345 + ], + "credit_map": [ + 0.10487331449985504, + 0.016592495143413544, + 0.032470814883708954, + 0.024222305044531822, + 0.021164216101169586, + 0.08095064759254456, + 0.036788005381822586, + 0.10274170339107513, + 0.008882544003427029, + 0.005330202169716358, + 0.006196037400513887, + 0.00592048792168498, + 0.0052124978974461555, + 0.01033481489866972, + 0.006798335816711187, + 0.014465986751019955, + 0.005679793190211058, + 0.003538019023835659, + 0.005358767695724964, + 0.004955397453159094, + 0.002717914991080761, + 0.005350172985345125, + 0.006109568756073713, + 0.005261139944195747, + 0.034972190856933594, + 0.011891227215528488, + 0.010440140962600708, + 0.005329375620931387, + 0.003925700671970844, + 0.054250456392765045, + 0.005951297003775835, + 0.015839751809835434, + 0.006697272881865501, + 0.005094680469483137, + 0.005060045514255762, + 0.004813502077013254, + 0.004049353301525116, + 0.00814901851117611, + 0.004664539359509945, + 0.00982745736837387, + 0.0058201891370117664, + 0.003809332149103284, + 0.011097247712314129, + 0.01136886328458786, + 0.004224606789648533, + 0.007810758892446756, + 0.005918969865888357, + 0.007033779751509428, + 0.004971515852957964, + 0.0021246473770588636, + 0.0029386100359261036, + 0.002107140375301242, + 0.0021657319739460945, + 0.0037574833258986473, + 0.003003568621352315, + 0.004573017358779907, + 0.009255128912627697, + 0.011256578378379345, + 0.011307706125080585, + 0.01848127320408821, + 0.008232717402279377, + 0.011233281344175339, + 0.00873066857457161, + 0.008095855824649334, + 0.005929768085479736, + 0.0028692851774394512, + 0.006468322593718767, + 0.0035943035036325455, + 0.004162709694355726, + 0.009323539212346077, + 0.009167062118649483, + 0.014402058906853199, + 0.005634250119328499, + 0.004910048097372055, + 0.005351974628865719, + 0.009781694039702415, + 0.003670063801109791, + 0.008589677512645721, + 0.007876154966652393, + 0.016079312190413475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.003638672875240445 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with short, tousled black hair and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a fitted olive green cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning slightly forward at the waist with hands in dynamic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright modern studio with a large whiteboard covered in a flowchart and colorful sticky notes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with clean, diffuse lighting and minimal shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020582539960741997 + ], + "opposite_logprob_tokens": [ + -3.895582437515259 + ], + "logprob_reward": -0.020582539960741997, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020582539960741997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1346977949142456, + 0.2140687108039856, + 0.10749145597219467, + 0.10994754731655121, + 0.11585884541273117, + 0.06228887662291527, + 0.036317434161901474, + 0.05829264968633652, + 0.10068517178297043, + 0.06035156548023224 + ], + "credit_map": [ + 0.016702571883797646, + 0.010958798229694366, + 0.013410390354692936, + 0.016738587990403175, + 0.010402072221040726, + 0.020155780017375946, + 0.016921410337090492, + 0.029408186674118042, + 0.02002352476119995, + 0.014444135129451752, + 0.022969167679548264, + 0.019175654277205467, + 0.0156811885535717, + 0.03732549026608467, + 0.02398124523460865, + 0.060468290001153946, + 0.015086035244166851, + 0.0077883535996079445, + 0.020844614133238792, + 0.012299531139433384, + 0.00466854777187109, + 0.015665307641029358, + 0.008142905309796333, + 0.022996149957180023, + 0.015625447034835815, + 0.009437370114028454, + 0.008832192048430443, + 0.008498854003846645, + 0.005995120853185654, + 0.03228387609124184, + 0.010567538440227509, + 0.018707139417529106, + 0.01102227158844471, + 0.011991414241492748, + 0.02047022059559822, + 0.012767639011144638, + 0.014383472502231598, + 0.017689330503344536, + 0.00985626969486475, + 0.017678216099739075, + 0.007616232614964247, + 0.004033232107758522, + 0.00634809909388423, + 0.005165856797248125, + 0.003470076248049736, + 0.01019780058413744, + 0.008371872827410698, + 0.01708570122718811, + 0.0031372886151075363, + 0.0036904970183968544, + 0.0040990374982357025, + 0.004503398668020964, + 0.0032542147673666477, + 0.0049026599153876305, + 0.004058899823576212, + 0.008671434596180916, + 0.010625900700688362, + 0.004086674191057682, + 0.008060458116233349, + 0.005006169900298119, + 0.0033120624721050262, + 0.008797751739621162, + 0.006834974978119135, + 0.011568649671971798, + 0.014258194714784622, + 0.005076440051198006, + 0.008350299671292305, + 0.007259019184857607, + 0.006126953288912773, + 0.018134552985429764, + 0.015624059364199638, + 0.02585565485060215, + 0.00716026034206152, + 0.004930149298161268, + 0.008004263043403625, + 0.005724573973566294, + 0.0022712883073836565, + 0.00971118826419115, + 0.00439325300976634, + 0.018156588077545166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.020582539960741997 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with short, tousled black hair and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a fitted olive green cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning slightly forward at the waist with hands in dynamic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright modern studio with a large whiteboard covered in a flowchart and colorful sticky notes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with clean, diffuse lighting and minimal shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004094788804650307 + ], + "opposite_logprob_tokens": [ + -5.50409460067749 + ], + "logprob_reward": -0.004094788804650307, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004094788804650307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12457658350467682, + 0.09117686003446579, + 0.04952964931726456, + 0.27860429883003235, + 0.15156495571136475, + 0.0392480194568634, + 0.051840197294950485, + 0.07993447035551071, + 0.07507026195526123, + 0.058454617857933044 + ], + "credit_map": [ + 0.02261456288397312, + 0.009369204752147198, + 0.013406085781753063, + 0.012169621884822845, + 0.008509805426001549, + 0.01913347654044628, + 0.012626453302800655, + 0.02674739994108677, + 0.012722665444016457, + 0.008233332075178623, + 0.011267967522144318, + 0.007297128904610872, + 0.005427721422165632, + 0.01592194102704525, + 0.008638367056846619, + 0.02166775055229664, + 0.0057639628648757935, + 0.0041519287042319775, + 0.004719972610473633, + 0.007185221184045076, + 0.003154811216518283, + 0.007891959510743618, + 0.004791433457285166, + 0.011870364658534527, + 0.06282621622085571, + 0.022959547117352486, + 0.02552105113863945, + 0.012465433217585087, + 0.008783616125583649, + 0.09717687964439392, + 0.01563829369843006, + 0.03323331102728844, + 0.015633193776011467, + 0.00784143153578043, + 0.03008946031332016, + 0.011976411566138268, + 0.006234539672732353, + 0.0280381478369236, + 0.009730357676744461, + 0.04202142730355263, + 0.004188832826912403, + 0.0021760889794677496, + 0.0029063280671834946, + 0.0027095128316432238, + 0.0020132435020059347, + 0.007318704389035702, + 0.005148398689925671, + 0.012786910869181156, + 0.008357702754437923, + 0.0053227925673127174, + 0.007228075992316008, + 0.004998028744012117, + 0.005214740056544542, + 0.00752324378117919, + 0.005397471133619547, + 0.007798147853463888, + 0.009261500090360641, + 0.004472609609365463, + 0.005601163022220135, + 0.006035827100276947, + 0.0041745491325855255, + 0.015444409102201462, + 0.006921746302396059, + 0.028022674843668938, + 0.008900994434952736, + 0.00428799819201231, + 0.010427010245621204, + 0.006402797531336546, + 0.005246727727353573, + 0.012646757997572422, + 0.00897543877363205, + 0.01818254217505455, + 0.0058588930405676365, + 0.004621338564902544, + 0.006414037197828293, + 0.004969651810824871, + 0.0030244626104831696, + 0.0099021689966321, + 0.006839034613221884, + 0.016825033351778984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.004094788804650307 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with short, tousled black hair and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a fitted olive green cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning slightly forward at the waist with hands in dynamic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright modern studio with a large whiteboard covered in a flowchart and colorful sticky notes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with clean, diffuse lighting and minimal shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253511756658554 + ], + "opposite_logprob_tokens": [ + -4.387535095214844 + ], + "logprob_reward": -0.01253511756658554, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253511756658554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12822110950946808, + 0.12539546191692352, + 0.04965744540095329, + 0.22219082713127136, + 0.1443277895450592, + 0.06305905431509018, + 0.10722999274730682, + 0.08742783218622208, + 0.03132765740156174, + 0.041162870824337006 + ], + "credit_map": [ + 0.020119378343224525, + 0.007418313529342413, + 0.012474693357944489, + 0.01147074531763792, + 0.009343616664409637, + 0.02129819430410862, + 0.014117654412984848, + 0.031978506594896317, + 0.022944245487451553, + 0.009002753533422947, + 0.022396819666028023, + 0.014981143176555634, + 0.007830716669559479, + 0.021972687914967537, + 0.009029582142829895, + 0.01723751425743103, + 0.006563160568475723, + 0.004713668487966061, + 0.005416533909738064, + 0.0073058814741671085, + 0.0030177896842360497, + 0.006993252784013748, + 0.005241253413259983, + 0.010405906476080418, + 0.04120093956589699, + 0.019834691658616066, + 0.014997984282672405, + 0.01131534855812788, + 0.006480547599494457, + 0.0636943131685257, + 0.011005566455423832, + 0.05366142839193344, + 0.013889068737626076, + 0.020333316177129745, + 0.01371148694306612, + 0.015587053261697292, + 0.022560274228453636, + 0.021405963227152824, + 0.015416121110320091, + 0.021424509584903717, + 0.006675005424767733, + 0.0031079286709427834, + 0.007084970362484455, + 0.0075118462555110455, + 0.00372943258844316, + 0.010241836309432983, + 0.008226659148931503, + 0.01648137904703617, + 0.02181445062160492, + 0.00884766411036253, + 0.011256465688347816, + 0.009288419969379902, + 0.00754077872261405, + 0.016962219029664993, + 0.015504499897360802, + 0.016015490517020226, + 0.014332566410303116, + 0.005508471745997667, + 0.0066519915126264095, + 0.006365310866385698, + 0.006002576556056738, + 0.015212326310575008, + 0.009725505486130714, + 0.02362908236682415, + 0.006437927018851042, + 0.0024380988907068968, + 0.004963902290910482, + 0.002654139418154955, + 0.0020821832586079836, + 0.004893408622592688, + 0.003406136529520154, + 0.00445185974240303, + 0.0052305664867162704, + 0.0037484741769731045, + 0.005380107555538416, + 0.0027414713986217976, + 0.002502200659364462, + 0.0062772976234555244, + 0.0045366729609668255, + 0.010746079497039318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.01253511756658554 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures a young Southeast Asian man in his early twenties, filled with infectious energy, enthusiastically explaining a complex idea to an unseen person off-camera. He has short, tousled black hair, and his expressive dark eyes are wide and excited. Wearing a simple, fitted olive green cotton T-shirt, he leans slightly forward at the waist, his body language focused. His hands are in constant, dynamic motion; one moment they dart through the air to emphasize a critical point, the next, his fingers carefully trace an invisible diagram, shaping abstract concepts with precise, clear gestures. The scene is set in a bright, modern creative studio, with a large whiteboard covered with a complex flowchart, and colorful sticky notes visible against a softly blurred background. The space is illuminated by clean, diffuse light from large, unseen windows, creating a vibrant, high-key look with minimal shadows. The atmosphere is filled with intellectual curiosity and creative enthusiasm, completely focused on the man's vibrant and engaging expression of his ideas.", + "seed": 9817, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_14/index_908_seed_9817_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.016740206629037857, + "vlm_logprob": -0.016740206629037857, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.016740206629037857, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with short, tousled black hair and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a fitted olive green cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning slightly forward at the waist with hands in dynamic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright modern studio with a large whiteboard covered in a flowchart and colorful sticky notes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with clean, diffuse lighting and minimal shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0070028104819357395 + ], + "opposite_logprob_tokens": [ + -5.132002830505371 + ], + "logprob_reward": -0.0070028104819357395, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0070028104819357395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22614262998104095, + 0.08657971769571304, + 0.06336532533168793, + 0.05634366720914841, + 0.12879517674446106, + 0.0844857320189476, + 0.1280270367860794, + 0.0706644281744957, + 0.05391565337777138, + 0.10168063640594482 + ], + "credit_map": [ + 0.02615303359925747, + 0.028705665841698647, + 0.028275471180677414, + 0.035702381283044815, + 0.03197833150625229, + 0.023746205493807793, + 0.023962227627635002, + 0.027619313448667526, + 0.011660421267151833, + 0.014420179650187492, + 0.011259140446782112, + 0.015235576778650284, + 0.008863769471645355, + 0.009645150043070316, + 0.007871096022427082, + 0.007624389138072729, + 0.0077882008627057076, + 0.008477164432406425, + 0.007774088066071272, + 0.012785178609192371, + 0.00605797627940774, + 0.010399149730801582, + 0.006342838518321514, + 0.003740723943337798, + 0.007854187861084938, + 0.006277678534388542, + 0.011237787082791328, + 0.006290891673415899, + 0.005730052944272757, + 0.008874774910509586, + 0.005074742250144482, + 0.0050035510212183, + 0.014611820690333843, + 0.010184595361351967, + 0.013096055015921593, + 0.021601824089884758, + 0.017697202041745186, + 0.017591197043657303, + 0.018314825370907784, + 0.015697654336690903, + 0.011129673570394516, + 0.007429893594235182, + 0.012529319152235985, + 0.010916393250226974, + 0.009739922359585762, + 0.012019189074635506, + 0.013162501156330109, + 0.007558836601674557, + 0.017321592196822166, + 0.00905551202595234, + 0.017471881583333015, + 0.039318036288022995, + 0.0064224740490317345, + 0.022539915516972542, + 0.009711256250739098, + 0.006186367478221655, + 0.008067399263381958, + 0.006892894394695759, + 0.008597585372626781, + 0.017948471009731293, + 0.004685797728598118, + 0.010402512736618519, + 0.008327823132276535, + 0.0057419463992118835, + 0.004840028937906027, + 0.00601027999073267, + 0.0054038334637880325, + 0.008863515220582485, + 0.006634625606238842, + 0.007300335913896561, + 0.00885030161589384, + 0.006012735888361931, + 0.010879026725888252, + 0.013433719985187054, + 0.00880106445401907, + 0.016176963225007057, + 0.013513050973415375, + 0.013392241671681404, + 0.013866905122995377, + 0.011617668904364109 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.0070028104819357395 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with short, tousled black hair and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a fitted olive green cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning slightly forward at the waist with hands in dynamic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright modern studio with a large whiteboard covered in a flowchart and colorful sticky notes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with clean, diffuse lighting and minimal shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004653813783079386 + ], + "opposite_logprob_tokens": [ + -5.3796539306640625 + ], + "logprob_reward": -0.004653813783079386, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004653813783079386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1798933893442154, + 0.16253016889095306, + 0.09022244065999985, + 0.09596067667007446, + 0.1296219676733017, + 0.0807315930724144, + 0.05815700814127922, + 0.10225392878055573, + 0.02954668179154396, + 0.07108219712972641 + ], + "credit_map": [ + 0.02136586420238018, + 0.023848481476306915, + 0.025350255891680717, + 0.022357115522027016, + 0.023528799414634705, + 0.020284496247768402, + 0.021348165348172188, + 0.02181021124124527, + 0.020335432142019272, + 0.024504665285348892, + 0.020902229472994804, + 0.022695088759064674, + 0.019187649711966515, + 0.015880392864346504, + 0.019575931131839752, + 0.01944877579808235, + 0.013355810195207596, + 0.011522985063493252, + 0.009762734174728394, + 0.016533637419342995, + 0.009159130975604057, + 0.01429474912583828, + 0.00885759387165308, + 0.006735802162438631, + 0.012771557085216045, + 0.009896841831505299, + 0.020616210997104645, + 0.010947403497993946, + 0.00824511144310236, + 0.013530165888369083, + 0.011747379787266254, + 0.00820600613951683, + 0.013341892510652542, + 0.00905617792159319, + 0.013444271869957447, + 0.02439388819038868, + 0.015470392070710659, + 0.020565085113048553, + 0.017924487590789795, + 0.015425760298967361, + 0.009014880284667015, + 0.007322108838707209, + 0.008426951244473457, + 0.0105087049305439, + 0.013634911738336086, + 0.010238991118967533, + 0.01111948024481535, + 0.010465568862855434, + 0.007816166616976261, + 0.004443690180778503, + 0.007652038708329201, + 0.011324428953230381, + 0.005334919784218073, + 0.008874237537384033, + 0.008443266153335571, + 0.004268262069672346, + 0.009915156289935112, + 0.009807121008634567, + 0.012633788399398327, + 0.021397801116108894, + 0.010845518670976162, + 0.013970527797937393, + 0.011882510967552662, + 0.01180150918662548, + 0.0037371432408690453, + 0.002831228543072939, + 0.0037788900081068277, + 0.004670682828873396, + 0.0029295834247022867, + 0.004161630291491747, + 0.004339027218520641, + 0.003098494838923216, + 0.007701862137764692, + 0.0070068626664578915, + 0.006907268892973661, + 0.00961995217949152, + 0.010378259234130383, + 0.009083949960768223, + 0.011242346838116646, + 0.009141700342297554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.004653813783079386 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with short, tousled black hair and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a fitted olive green cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning slightly forward at the waist with hands in dynamic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright modern studio with a large whiteboard covered in a flowchart and colorful sticky notes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with clean, diffuse lighting and minimal shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048650916665792465 + ], + "opposite_logprob_tokens": [ + -3.0486509799957275 + ], + "logprob_reward": -0.048650916665792465, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048650916665792465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23857498168945312, + 0.08326715975999832, + 0.06840486824512482, + 0.22984011471271515, + 0.0723043903708458, + 0.0905541479587555, + 0.056750692427158356, + 0.07181858271360397, + 0.048029541969299316, + 0.04045560210943222 + ], + "credit_map": [ + 0.023113783448934555, + 0.03288266435265541, + 0.044298671185970306, + 0.04384420067071915, + 0.028567500412464142, + 0.021661484614014626, + 0.020132850855588913, + 0.024073798209428787, + 0.0122908316552639, + 0.011693979613482952, + 0.01428705733269453, + 0.015198671258985996, + 0.006199447903782129, + 0.007070583291351795, + 0.007742660120129585, + 0.008783920668065548, + 0.009236453101038933, + 0.008468586020171642, + 0.014110222458839417, + 0.008941015228629112, + 0.005790411029011011, + 0.01245197281241417, + 0.005353779532015324, + 0.004052416887134314, + 0.030651509761810303, + 0.0249540526419878, + 0.042159125208854675, + 0.02140110358595848, + 0.019092023372650146, + 0.0324140302836895, + 0.03465111553668976, + 0.024517124518752098, + 0.008511488325893879, + 0.004142906051129103, + 0.011068380437791348, + 0.015505551360547543, + 0.0034776723477989435, + 0.010918560437858105, + 0.010495660826563835, + 0.00818416103720665, + 0.011093180626630783, + 0.010073908604681492, + 0.015927273780107498, + 0.02068585343658924, + 0.006963537074625492, + 0.012470654211938381, + 0.007450764998793602, + 0.005888964980840683, + 0.005796564277261496, + 0.004298645071685314, + 0.008159193210303783, + 0.010403851047158241, + 0.005802782252430916, + 0.007666202262043953, + 0.006378551945090294, + 0.008244896307587624, + 0.009082825854420662, + 0.006580383516848087, + 0.009386022575199604, + 0.017672831192612648, + 0.0067821079865098, + 0.009778684936463833, + 0.007142945658415556, + 0.005392778664827347, + 0.005482868757098913, + 0.004649405367672443, + 0.007812490686774254, + 0.008789315819740295, + 0.0032253770623356104, + 0.006291818339377642, + 0.005647051148116589, + 0.006131209898740053, + 0.005221635568886995, + 0.004293578211218119, + 0.004902965389192104, + 0.007630349136888981, + 0.004376158583909273, + 0.005900359712541103, + 0.0038079479709267616, + 0.004322603344917297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.048650916665792465 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with short, tousled black hair and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a fitted olive green cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning slightly forward at the waist with hands in dynamic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright modern studio with a large whiteboard covered in a flowchart and colorful sticky notes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with clean, diffuse lighting and minimal shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002818542532622814 + ], + "opposite_logprob_tokens": [ + -5.877818584442139 + ], + "logprob_reward": -0.002818542532622814, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002818542532622814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18813279271125793, + 0.07954776287078857, + 0.10830675065517426, + 0.06615210324525833, + 0.16043518483638763, + 0.05910969898104668, + 0.17359225451946259, + 0.059033963829278946, + 0.046657342463731766, + 0.059032097458839417 + ], + "credit_map": [ + 0.026017161086201668, + 0.026659797877073288, + 0.025775544345378876, + 0.04484906047582626, + 0.017867103219032288, + 0.014255211688578129, + 0.015253996476531029, + 0.017454922199249268, + 0.010737570002675056, + 0.014043379575014114, + 0.009382841177284718, + 0.0200786292552948, + 0.0054350183345377445, + 0.006189266219735146, + 0.007044818252325058, + 0.006636237259954214, + 0.01398009154945612, + 0.013687641359865665, + 0.015382155776023865, + 0.02339884266257286, + 0.007665618788450956, + 0.016249237582087517, + 0.009945498779416084, + 0.007997668348252773, + 0.008387139067053795, + 0.006415190175175667, + 0.011776844039559364, + 0.01439896784722805, + 0.005101399030536413, + 0.009673588909208775, + 0.005523929838091135, + 0.0048750462010502815, + 0.016545843333005905, + 0.009543915279209614, + 0.024477915838360786, + 0.04356073960661888, + 0.007928213104605675, + 0.03177042677998543, + 0.012928372249007225, + 0.013679766096174717, + 0.005850356537848711, + 0.004584112670272589, + 0.007331882603466511, + 0.008740931749343872, + 0.004864479880779982, + 0.007719411049038172, + 0.012403967790305614, + 0.007614553440362215, + 0.021470513194799423, + 0.011573848314583302, + 0.02767479605972767, + 0.047727614641189575, + 0.009877788834273815, + 0.02881620265543461, + 0.014871293678879738, + 0.01158020831644535, + 0.007531909272074699, + 0.007357741706073284, + 0.007985004223883152, + 0.013305454514920712, + 0.005040979944169521, + 0.009135307744145393, + 0.005334429908543825, + 0.0033431355841457844, + 0.005320743191987276, + 0.003755836747586727, + 0.004234148655086756, + 0.012900376692414284, + 0.0038325602654367685, + 0.006601748522371054, + 0.005302023142576218, + 0.004709908273071051, + 0.007498112972825766, + 0.00526842288672924, + 0.009679908864200115, + 0.013021597638726234, + 0.004593152552843094, + 0.008544412441551685, + 0.006467426661401987, + 0.003959061577916145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.002818542532622814 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with short, tousled black hair and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a fitted olive green cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning slightly forward at the waist with hands in dynamic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright modern studio with a large whiteboard covered in a flowchart and colorful sticky notes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with clean, diffuse lighting and minimal shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057494968175888 + ], + "opposite_logprob_tokens": [ + -3.8955750465393066 + ], + "logprob_reward": -0.02057494968175888, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057494968175888, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1900154948234558, + 0.13178175687789917, + 0.06515531986951828, + 0.09622227400541306, + 0.16009005904197693, + 0.041451215744018555, + 0.13264533877372742, + 0.09462051838636398, + 0.04529237374663353, + 0.04272567853331566 + ], + "credit_map": [ + 0.02466205134987831, + 0.028258737176656723, + 0.039912063628435135, + 0.03394788131117821, + 0.01666591502726078, + 0.013431010767817497, + 0.016084905713796616, + 0.0170529093593359, + 0.019590988755226135, + 0.01979951746761799, + 0.024943752214312553, + 0.022697504609823227, + 0.009414220228791237, + 0.010644244030117989, + 0.012579970061779022, + 0.012111558578908443, + 0.009358162991702557, + 0.009452988393604755, + 0.007135665509849787, + 0.009625564329326153, + 0.005527711007744074, + 0.013102586381137371, + 0.006326120812445879, + 0.0046265097334980965, + 0.012214653193950653, + 0.01100972667336464, + 0.012499367818236351, + 0.010610895231366158, + 0.01022950280457735, + 0.01545774471014738, + 0.011614491231739521, + 0.01258588396012783, + 0.02172969840466976, + 0.008014154620468616, + 0.018512124195694923, + 0.04290333017706871, + 0.010425092652440071, + 0.028031492605805397, + 0.016380513086915016, + 0.014093644917011261, + 0.004494498483836651, + 0.004627550486475229, + 0.0053381528705358505, + 0.0074695716612041, + 0.004084068816155195, + 0.006592642515897751, + 0.005783969070762396, + 0.0030607592780143023, + 0.01740133762359619, + 0.009057252667844296, + 0.014678655192255974, + 0.030408095568418503, + 0.010141278617084026, + 0.020513111725449562, + 0.016140764579176903, + 0.014304821379482746, + 0.011777415871620178, + 0.008586423471570015, + 0.010967595502734184, + 0.025033125653862953, + 0.006617113947868347, + 0.014613835141062737, + 0.010969828814268112, + 0.0060551706701517105, + 0.005363460164517164, + 0.004869961645454168, + 0.005280328448861837, + 0.011805053800344467, + 0.0032048257999122143, + 0.007077193818986416, + 0.004449461121112108, + 0.003242081729695201, + 0.004723711404949427, + 0.00361750484444201, + 0.005195865407586098, + 0.009676036424934864, + 0.003257727948948741, + 0.007191454991698265, + 0.00453150412067771, + 0.0045318687334656715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.02057494968175888 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures a young Southeast Asian man in his early twenties, filled with infectious energy, enthusiastically explaining a complex idea to an unseen person off-camera. He has short, tousled black hair, and his expressive dark eyes are wide and excited. Wearing a simple, fitted olive green cotton T-shirt, he leans slightly forward at the waist, his body language focused. His hands are in constant, dynamic motion; one moment they dart through the air to emphasize a critical point, the next, his fingers carefully trace an invisible diagram, shaping abstract concepts with precise, clear gestures. The scene is set in a bright, modern creative studio, with a large whiteboard covered with a complex flowchart, and colorful sticky notes visible against a softly blurred background. The space is illuminated by clean, diffuse light from large, unseen windows, creating a vibrant, high-key look with minimal shadows. The atmosphere is filled with intellectual curiosity and creative enthusiasm, completely focused on the man's vibrant and engaging expression of his ideas.", + "seed": 9818, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_14/index_908_seed_9818_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.047034124843776226, + "vlm_logprob": -0.047034124843776226, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.047034124843776226, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with short, tousled black hair and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a fitted olive green cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning slightly forward at the waist with hands in dynamic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright modern studio with a large whiteboard covered in a flowchart and colorful sticky notes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with clean, diffuse lighting and minimal shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01227550208568573 + ], + "opposite_logprob_tokens": [ + -4.512275695800781 + ], + "logprob_reward": -0.01227550208568573, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01227550208568573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22417448461055756, + 0.1504395753145218, + 0.07650084048509598, + 0.1447359025478363, + 0.04550817981362343, + 0.05738908424973488, + 0.08285625278949738, + 0.10221325606107712, + 0.06284074485301971, + 0.05334171652793884 + ], + "credit_map": [ + 0.04347016289830208, + 0.024666037410497665, + 0.02433912642300129, + 0.026243217289447784, + 0.029135873541235924, + 0.024017687886953354, + 0.029175857082009315, + 0.023126529529690742, + 0.019339028745889664, + 0.016731558367609978, + 0.02202877588570118, + 0.020361945033073425, + 0.020776188001036644, + 0.01771574467420578, + 0.012885477393865585, + 0.020600849762558937, + 0.016297588124871254, + 0.007064398843795061, + 0.015185916796326637, + 0.006651643197983503, + 0.005675163120031357, + 0.013217149302363396, + 0.006295333616435528, + 0.006113645154982805, + 0.02005435898900032, + 0.02790258638560772, + 0.01993994414806366, + 0.013396772556006908, + 0.0110268983989954, + 0.02134908363223076, + 0.010680068284273148, + 0.02038619853556156, + 0.006663430016487837, + 0.008199217729270458, + 0.0051008895970880985, + 0.0056485142558813095, + 0.003087792545557022, + 0.003050923580303788, + 0.004962507635354996, + 0.008794903755187988, + 0.00850446242839098, + 0.0072688269428908825, + 0.013754268176853657, + 0.004651102237403393, + 0.003312129992991686, + 0.012804708443582058, + 0.0036180606111884117, + 0.0034755223896354437, + 0.011154083535075188, + 0.012199509888887405, + 0.01656152680516243, + 0.012120869010686874, + 0.005684634670615196, + 0.012684045359492302, + 0.006659307982772589, + 0.0057922834530472755, + 0.016141297295689583, + 0.018590088933706284, + 0.013821467757225037, + 0.011925692670047283, + 0.007025015540421009, + 0.01281734649091959, + 0.009655782021582127, + 0.012236560694873333, + 0.012556036002933979, + 0.008891144767403603, + 0.007132485508918762, + 0.00702696293592453, + 0.0044763837940990925, + 0.007056821137666702, + 0.006442004814743996, + 0.009258907288312912, + 0.012842950411140919, + 0.011068487539887428, + 0.007089001126587391, + 0.005138738546520472, + 0.005149082746356726, + 0.004541496746242046, + 0.0034113116562366486, + 0.0041006482206285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.01227550208568573 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with short, tousled black hair and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a fitted olive green cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning slightly forward at the waist with hands in dynamic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright modern studio with a large whiteboard covered in a flowchart and colorful sticky notes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with clean, diffuse lighting and minimal shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005966234486550093 + ], + "opposite_logprob_tokens": [ + -5.1309661865234375 + ], + "logprob_reward": -0.005966234486550093, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005966234486550093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.151300311088562, + 0.12314198911190033, + 0.07980797439813614, + 0.08133367449045181, + 0.05022707208991051, + 0.12055519223213196, + 0.1563834249973297, + 0.13873139023780823, + 0.055383067578077316, + 0.04313591867685318 + ], + "credit_map": [ + 0.024863075464963913, + 0.0174506064504385, + 0.016498439013957977, + 0.019171198830008507, + 0.019199512898921967, + 0.011020687408745289, + 0.02519392967224121, + 0.01790286973118782, + 0.015549789182841778, + 0.018673403188586235, + 0.021932452917099, + 0.01420669537037611, + 0.012142475694417953, + 0.018168918788433075, + 0.012670928612351418, + 0.009797340258955956, + 0.014576557092368603, + 0.009212011471390724, + 0.011209152638912201, + 0.008154426701366901, + 0.008981379680335522, + 0.012304021045565605, + 0.008379523642361164, + 0.006990907713770866, + 0.011731171049177647, + 0.01973743364214897, + 0.011719548143446445, + 0.007079262286424637, + 0.006609482690691948, + 0.01412907149642706, + 0.004982410930097103, + 0.005345300771296024, + 0.00613507442176342, + 0.007360388524830341, + 0.005591810215264559, + 0.004245194140821695, + 0.0046438840217888355, + 0.0036658053286373615, + 0.0068652029149234295, + 0.01171971671283245, + 0.01501525565981865, + 0.015015169978141785, + 0.03847398981451988, + 0.006203849799931049, + 0.004357714205980301, + 0.03088993951678276, + 0.004532103426754475, + 0.006067181937396526, + 0.02823975309729576, + 0.01667352020740509, + 0.02036435343325138, + 0.018093453720211983, + 0.014294659718871117, + 0.019915616139769554, + 0.012921946123242378, + 0.02588014118373394, + 0.021700767800211906, + 0.022645235061645508, + 0.020233260467648506, + 0.014535879716277122, + 0.01091423723846674, + 0.015654340386390686, + 0.015020257793366909, + 0.018027417361736298, + 0.010202725417912006, + 0.012174276635050774, + 0.007242346182465553, + 0.00515382643789053, + 0.005944292526692152, + 0.005280419252812862, + 0.0036304655950516462, + 0.0057547190226614475, + 0.012305554933845997, + 0.007630021311342716, + 0.006089192349463701, + 0.0034088999964296818, + 0.0032770875841379166, + 0.002733068773522973, + 0.0035565695725381374, + 0.0041355290450155735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.005966234486550093 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with short, tousled black hair and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a fitted olive green cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning slightly forward at the waist with hands in dynamic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright modern studio with a large whiteboard covered in a flowchart and colorful sticky notes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with clean, diffuse lighting and minimal shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20157617330551147 + ], + "opposite_logprob_tokens": [ + -1.7015762329101562 + ], + "logprob_reward": -0.20157617330551147, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20157617330551147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1210201159119606, + 0.1067822054028511, + 0.0924597680568695, + 0.07541064918041229, + 0.0675046518445015, + 0.039476968348026276, + 0.07301928848028183, + 0.3306063711643219, + 0.06027446314692497, + 0.03344548121094704 + ], + "credit_map": [ + 0.011630699969828129, + 0.012581972405314445, + 0.01319627184420824, + 0.015911169350147247, + 0.011596049182116985, + 0.007924058474600315, + 0.020105721428990364, + 0.02807418256998062, + 0.01602349802851677, + 0.0176223311573267, + 0.01767348125576973, + 0.012096477672457695, + 0.00885613914579153, + 0.015490012243390083, + 0.010350530035793781, + 0.008669737726449966, + 0.012168431654572487, + 0.006982713006436825, + 0.011459819972515106, + 0.0074166348204016685, + 0.010365071706473827, + 0.013650014065206051, + 0.013387718237936497, + 0.01702936552464962, + 0.009779103100299835, + 0.017983654513955116, + 0.01088206097483635, + 0.007709236349910498, + 0.00460071349516511, + 0.013201302848756313, + 0.004610625561326742, + 0.006643959786742926, + 0.011868911795318127, + 0.011057513765990734, + 0.009000818245112896, + 0.007311985362321138, + 0.007168122101575136, + 0.005504640284925699, + 0.006668021436780691, + 0.008924644440412521, + 0.00717983627691865, + 0.005261152517050505, + 0.006234404165297747, + 0.006027088034898043, + 0.0032467327546328306, + 0.005154964979737997, + 0.0032663436140865088, + 0.0031064492650330067, + 0.009035952389240265, + 0.012569822371006012, + 0.01076499093323946, + 0.010883712209761143, + 0.005971110891550779, + 0.011953803710639477, + 0.006509824655950069, + 0.005330076906830072, + 0.06678209453821182, + 0.05429710075259209, + 0.05042886361479759, + 0.029939904808998108, + 0.022771744057536125, + 0.041415441781282425, + 0.029372405260801315, + 0.03559885546565056, + 0.010102968662977219, + 0.012367471121251583, + 0.007832780480384827, + 0.0064475624822080135, + 0.006557959597557783, + 0.006647470407187939, + 0.0048984140157699585, + 0.005419843830168247, + 0.005579438991844654, + 0.003996619023382664, + 0.003474785014986992, + 0.0032927782740443945, + 0.003731134347617626, + 0.0025977445766329765, + 0.004436320159584284, + 0.006336664315313101 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.20157617330551147 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with short, tousled black hair and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a fitted olive green cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning slightly forward at the waist with hands in dynamic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright modern studio with a large whiteboard covered in a flowchart and colorful sticky notes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with clean, diffuse lighting and minimal shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002818067092448473 + ], + "opposite_logprob_tokens": [ + -5.8778181076049805 + ], + "logprob_reward": -0.002818067092448473, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002818067092448473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2192084938287735, + 0.18336902558803558, + 0.10161565244197845, + 0.08555199205875397, + 0.05121009051799774, + 0.1065596491098404, + 0.04668848589062691, + 0.07134686410427094, + 0.07626291364431381, + 0.05818677321076393 + ], + "credit_map": [ + 0.035431597381830215, + 0.04848773777484894, + 0.023165235295891762, + 0.022360943257808685, + 0.018873009830713272, + 0.02377009205520153, + 0.024898327887058258, + 0.02222156710922718, + 0.026807956397533417, + 0.03059419058263302, + 0.028103137388825417, + 0.018693752586841583, + 0.01686626300215721, + 0.02393813617527485, + 0.011619924567639828, + 0.02674567885696888, + 0.01996367610991001, + 0.01311727799475193, + 0.021671747788786888, + 0.007271282374858856, + 0.007409663405269384, + 0.019024880602955818, + 0.0061112940311431885, + 0.007045835256576538, + 0.012973873876035213, + 0.019797177985310555, + 0.013373931869864464, + 0.008601878769695759, + 0.005755119025707245, + 0.01533441711217165, + 0.004573687445372343, + 0.005141917150467634, + 0.010339387692511082, + 0.010271367616951466, + 0.006844738498330116, + 0.004357176832854748, + 0.00416187010705471, + 0.004530757199972868, + 0.004167334642261267, + 0.006537462119013071, + 0.018483946099877357, + 0.01579948328435421, + 0.019777635112404823, + 0.012101611122488976, + 0.008454278111457825, + 0.013852955773472786, + 0.00905702169984579, + 0.009032726287841797, + 0.006536591332405806, + 0.012078865431249142, + 0.006194015499204397, + 0.0048078796826303005, + 0.002434889320284128, + 0.007934587076306343, + 0.0031323295552283525, + 0.003569334279745817, + 0.017569521442055702, + 0.008470566011965275, + 0.011887185275554657, + 0.006258996203541756, + 0.004076095763593912, + 0.009869149886071682, + 0.006868948694318533, + 0.006346405483782291, + 0.016750626266002655, + 0.01692487671971321, + 0.009956425055861473, + 0.007464245427399874, + 0.004929613322019577, + 0.007192748598754406, + 0.00700650317594409, + 0.006037880200892687, + 0.01718289591372013, + 0.012261354364454746, + 0.008098353631794453, + 0.005306095816195011, + 0.003857184201478958, + 0.004241178743541241, + 0.003550107590854168, + 0.0036896143574267626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.002818067092448473 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with short, tousled black hair and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a fitted olive green cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning slightly forward at the waist with hands in dynamic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright modern studio with a large whiteboard covered in a flowchart and colorful sticky notes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with clean, diffuse lighting and minimal shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253464724868536 + ], + "opposite_logprob_tokens": [ + -4.3875346183776855 + ], + "logprob_reward": -0.01253464724868536, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253464724868536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24753466248512268, + 0.21351079642772675, + 0.06532862782478333, + 0.14602775871753693, + 0.10557565838098526, + 0.0525917187333107, + 0.0343606099486351, + 0.04814257100224495, + 0.042290735989809036, + 0.04463693127036095 + ], + "credit_map": [ + 0.024900877848267555, + 0.027611348778009415, + 0.026353755965828896, + 0.029386647045612335, + 0.02986643649637699, + 0.0165299903601408, + 0.042525094002485275, + 0.05036048963665962, + 0.027576914057135582, + 0.03513043001294136, + 0.040298059582710266, + 0.019900627434253693, + 0.017469340935349464, + 0.03584698215126991, + 0.015788298100233078, + 0.021500110626220703, + 0.005921203177422285, + 0.007419366855174303, + 0.012912092730402946, + 0.006456492934376001, + 0.00773895438760519, + 0.009227963164448738, + 0.006860555149614811, + 0.00879199244081974, + 0.018198898062109947, + 0.026722442358732224, + 0.02576526068150997, + 0.01323921699076891, + 0.016023142263293266, + 0.02367292158305645, + 0.00990185234695673, + 0.01250402070581913, + 0.023794464766979218, + 0.020511610433459282, + 0.012843281030654907, + 0.011774197220802307, + 0.008912834338843822, + 0.009446818381547928, + 0.006517942063510418, + 0.011774497106671333, + 0.011131762526929379, + 0.009064719080924988, + 0.008581224828958511, + 0.004088917281478643, + 0.0033664903603494167, + 0.007967384532094002, + 0.0038160057738423347, + 0.004575209692120552, + 0.004394978750497103, + 0.004668291658163071, + 0.008133627474308014, + 0.004366226028650999, + 0.0025459681637585163, + 0.00571992713958025, + 0.0022851310204714537, + 0.0022464559879153967, + 0.005346093326807022, + 0.0075432416051626205, + 0.007847843691706657, + 0.00601368909701705, + 0.0033566553611308336, + 0.006048097275197506, + 0.004913607612252235, + 0.007073340937495232, + 0.009602705016732216, + 0.00700995372608304, + 0.005423152819275856, + 0.0048214527778327465, + 0.003617669688537717, + 0.004599804989993572, + 0.0033914921805262566, + 0.0038244971074163914, + 0.011289897374808788, + 0.0070313094183802605, + 0.0062626139260828495, + 0.003941074479371309, + 0.0030405500438064337, + 0.004130715504288673, + 0.00364179746247828, + 0.005298970267176628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.01253464724868536 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_15/index_tensor([124, 124])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_15/index_tensor([124, 124])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0822128019e94f8dd18c03cee6e350b62cc3e3aa --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_15/index_tensor([124, 124])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "majestic low-angle wide shot, the camera panning steadily to the right, juxtaposes a massive dragon boat, crafted entirely from gleaming brass, gracefully sailing above a boundless, fluffy sea of clouds. The golden sunlight of late afternoon bathes the frame, gleaming brilliantly off the boat's serpentine hull, meticulously riveted and ornately carved. At the prow lies a majestic dragon head, its eyes blazing red. From its flared nostrils, thick, white steam erupts rhythmically and powerfully, creating a shimmering halo in the sunlight before dissipating into the clear, high-altitude air. Along its sides, massive, intermeshing brass gears rotate with a steady, unstoppable force, driving complex pistons and connecting rods. This exposed mechanism, with its precise rhythmic movement, provides the fantastical craft with an endless supply of power. The entire scene is filled with an epic and awe-inspiring steampunk atmosphere, perfectly blending the shape of ancient mythology with the majestic power of the Industrial Revolution, presenting an extraordinary and magnificent voyage.", + "seed": 15795, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_15/index_124_seed_15795_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.23845508047379554, + "vlm_logprob": -0.23845508047379554, + "vlm_yes_no_margin": 2.475000023841858, + "vlm_token_credit": -0.23845508047379554, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the dragon boat crafted entirely from gleaming brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dragon boat sailing above a boundless, fluffy sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon head with blazing red eyes visible at the prow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive intermeshing brass gears rotating along the sides of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15030448138713837 + ], + "opposite_logprob_tokens": [ + -2.0253045558929443 + ], + "logprob_reward": -0.15030448138713837, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.15030448138713837, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1457928866147995, + 0.11868924647569656, + 0.04782761260867119, + 0.0583270788192749, + 0.044863078743219376, + 0.061547134071588516, + 0.04348796233534813, + 0.03896322473883629, + 0.18052567541599274, + 0.2599761188030243 + ], + "credit_map": [ + 0.022884678095579147, + 0.019595371559262276, + 0.024149922654032707, + 0.016734784469008446, + 0.012404113076627254, + 0.01313003059476614, + 0.01871693879365921, + 0.018177034333348274, + 0.019866913557052612, + 0.009971309453248978, + 0.0238377396017313, + 0.011958741582930088, + 0.01529010571539402, + 0.00907943956553936, + 0.01224210299551487, + 0.016442880034446716, + 0.007386051118373871, + 0.005065527278929949, + 0.006572232116013765, + 0.004016040824353695, + 0.007427737582474947, + 0.004683764185756445, + 0.006929033435881138, + 0.005747221875935793, + 0.008173887617886066, + 0.009104237891733646, + 0.007624114863574505, + 0.0078052496537566185, + 0.005164661444723606, + 0.006358765531331301, + 0.0075510842725634575, + 0.006545075215399265, + 0.005770375486463308, + 0.006220846436917782, + 0.00629739835858345, + 0.0050707352347671986, + 0.005079213995486498, + 0.005189923569560051, + 0.0067595248110592365, + 0.004475056193768978, + 0.007640623487532139, + 0.006925560068339109, + 0.009090475738048553, + 0.007691216189414263, + 0.007057123351842165, + 0.006842527072876692, + 0.008595339953899384, + 0.007704269606620073, + 0.011447052471339703, + 0.005191084463149309, + 0.007965625263750553, + 0.0036071185022592545, + 0.004561664070934057, + 0.00363743444904685, + 0.0036206005606800318, + 0.003457379527390003, + 0.005132388323545456, + 0.004663580097258091, + 0.0056668855249881744, + 0.0033770303707569838, + 0.003509475849568844, + 0.0031456532888114452, + 0.004973305389285088, + 0.00849490612745285, + 0.03140820935368538, + 0.022555019706487656, + 0.04121020436286926, + 0.013399476185441017, + 0.015647558495402336, + 0.01853851228952408, + 0.01526711881160736, + 0.022499553859233856, + 0.02946464717388153, + 0.029304346069693565, + 0.03416799381375313, + 0.02162216790020466, + 0.03152577579021454, + 0.060421980917453766, + 0.02680715173482895, + 0.026662025600671768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.15030448138713837 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the dragon boat crafted entirely from gleaming brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dragon boat sailing above a boundless, fluffy sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon head with blazing red eyes visible at the prow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive intermeshing brass gears rotating along the sides of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036340404767543077 + ], + "opposite_logprob_tokens": [ + -5.628633975982666 + ], + "logprob_reward": -0.0036340404767543077, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036340404767543077, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18418218195438385, + 0.19296284019947052, + 0.09165555238723755, + 0.07717832922935486, + 0.0573442243039608, + 0.04459825158119202, + 0.09494616836309433, + 0.044883329421281815, + 0.038111649453639984, + 0.17413751780986786 + ], + "credit_map": [ + 0.03320344537496567, + 0.016844701021909714, + 0.04367922991514206, + 0.012762761674821377, + 0.024115923792123795, + 0.014367811381816864, + 0.017678556963801384, + 0.021529758349061012, + 0.03338894993066788, + 0.013894138857722282, + 0.042033497244119644, + 0.01848720572888851, + 0.025675004348158836, + 0.015986183658242226, + 0.020550694316625595, + 0.022947169840335846, + 0.011859603226184845, + 0.006716365925967693, + 0.012415400706231594, + 0.006587916519492865, + 0.01133606769144535, + 0.007490247022360563, + 0.011444809846580029, + 0.02380514331161976, + 0.009051176719367504, + 0.004591221921145916, + 0.01370365358889103, + 0.006087273359298706, + 0.005377932917326689, + 0.004851675126701593, + 0.008938404731452465, + 0.02457698993384838, + 0.006576559040695429, + 0.004386113490909338, + 0.010063223540782928, + 0.005231086630374193, + 0.004880806896835566, + 0.003961261361837387, + 0.009929838590323925, + 0.012315336614847183, + 0.008526635356247425, + 0.004075434524565935, + 0.006235156673938036, + 0.0035938608925789595, + 0.004327363800257444, + 0.0031053186394274235, + 0.005304678343236446, + 0.009429802186787128, + 0.021928705275058746, + 0.011922664940357208, + 0.01555815152823925, + 0.009112323634326458, + 0.007168647367507219, + 0.0072161429561674595, + 0.009834283031523228, + 0.012205248698592186, + 0.006906584370881319, + 0.0039679608307778835, + 0.009149632416665554, + 0.002605957444757223, + 0.003444192698225379, + 0.003911738749593496, + 0.005010862834751606, + 0.009886400774121284, + 0.005971421021968126, + 0.003207869827747345, + 0.007452929392457008, + 0.0032336399890482426, + 0.0050863297656178474, + 0.003136163577437401, + 0.00339270057156682, + 0.006630597170442343, + 0.022373469546437263, + 0.017247959971427917, + 0.025781061500310898, + 0.011963510885834694, + 0.017200253903865814, + 0.026182996109128, + 0.013890238478779793, + 0.03949802741408348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.0036340404767543077 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the dragon boat crafted entirely from gleaming brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dragon boat sailing above a boundless, fluffy sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon head with blazing red eyes visible at the prow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive intermeshing brass gears rotating along the sides of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000492513179779 + ], + "opposite_logprob_tokens": [ + -2.695004940032959 + ], + "logprob_reward": -0.07000492513179779, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000492513179779, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20934423804283142, + 0.10830409079790115, + 0.1660841703414917, + 0.09924691170454025, + 0.13546514511108398, + 0.029642252251505852, + 0.023434801027178764, + 0.051057327538728714, + 0.028024345636367798, + 0.14939671754837036 + ], + "credit_map": [ + 0.03558963164687157, + 0.021755080670118332, + 0.043519217520952225, + 0.0191995520144701, + 0.024102743715047836, + 0.021692687645554543, + 0.01633603125810623, + 0.02714928425848484, + 0.01645972765982151, + 0.011757336556911469, + 0.017153475433588028, + 0.010420650243759155, + 0.014439799822866917, + 0.010359881445765495, + 0.01138182170689106, + 0.0163313839584589, + 0.029500791803002357, + 0.01653173193335533, + 0.028837867081165314, + 0.009765361435711384, + 0.032758403569459915, + 0.016449863091111183, + 0.018501287326216698, + 0.013738859444856644, + 0.011693914420902729, + 0.010472842492163181, + 0.015269311144948006, + 0.01439328957349062, + 0.01015534158796072, + 0.009529362432658672, + 0.01294959057122469, + 0.014783264137804508, + 0.014208626933395863, + 0.014360703527927399, + 0.018200576305389404, + 0.015465540811419487, + 0.021301794797182083, + 0.015310111455619335, + 0.021448735147714615, + 0.015169056132435799, + 0.003190858056768775, + 0.0033301038201898336, + 0.004397182259708643, + 0.0042401328682899475, + 0.0037969935219734907, + 0.0024278329219669104, + 0.005274641327559948, + 0.0029845077078789473, + 0.004156855400651693, + 0.0026191622018814087, + 0.003703648690134287, + 0.0032738186419010162, + 0.0018300935626029968, + 0.0020618808921426535, + 0.0026844185777008533, + 0.0031049230601638556, + 0.0109492726624012, + 0.006790217012166977, + 0.007562438491731882, + 0.005370463244616985, + 0.005165998358279467, + 0.003827536478638649, + 0.005013014655560255, + 0.006378384307026863, + 0.0030425169970840216, + 0.0027414937503635883, + 0.00532723031938076, + 0.0038179235998541117, + 0.003122848691418767, + 0.0025474634021520615, + 0.003716440172865987, + 0.003708428470417857, + 0.022198889404535294, + 0.016376422718167305, + 0.02043173275887966, + 0.010934378951787949, + 0.01431260909885168, + 0.033514752984046936, + 0.012719407677650452, + 0.01890852488577366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.07000492513179779 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the dragon boat crafted entirely from gleaming brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dragon boat sailing above a boundless, fluffy sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon head with blazing red eyes visible at the prow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive intermeshing brass gears rotating along the sides of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003393024206161 + ], + "opposite_logprob_tokens": [ + -2.6950340270996094 + ], + "logprob_reward": -0.07003393024206161, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003393024206161, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17314539849758148, + 0.11414072662591934, + 0.16056948900222778, + 0.09795340150594711, + 0.10974302887916565, + 0.053913433104753494, + 0.03799935802817345, + 0.12550047039985657, + 0.04742712527513504, + 0.07960755378007889 + ], + "credit_map": [ + 0.028581466525793076, + 0.0186713095754385, + 0.036064986139535904, + 0.015394377522170544, + 0.017615418881177902, + 0.013986079953610897, + 0.022770296782255173, + 0.020061470568180084, + 0.016210373491048813, + 0.01467695739120245, + 0.018432073295116425, + 0.012283461168408394, + 0.01683518849313259, + 0.011009592562913895, + 0.013876300305128098, + 0.010816788300871849, + 0.027059826999902725, + 0.015594592317938805, + 0.027546701952815056, + 0.010470355860888958, + 0.02792324125766754, + 0.014775717630982399, + 0.020902952179312706, + 0.01629611663520336, + 0.011519541963934898, + 0.007166997995227575, + 0.01906939595937729, + 0.010392667725682259, + 0.009276334196329117, + 0.008824779652059078, + 0.0144812548533082, + 0.017222441732883453, + 0.012834643945097923, + 0.01260436326265335, + 0.014799180440604687, + 0.009993596002459526, + 0.01455887034535408, + 0.01086383406072855, + 0.021897783502936363, + 0.012190763838589191, + 0.005711121950298548, + 0.006084536202251911, + 0.009071077220141888, + 0.005730883684009314, + 0.005284484010189772, + 0.005635235458612442, + 0.008829786442220211, + 0.0075663127936422825, + 0.0058882334269583225, + 0.005229663569480181, + 0.005829515866935253, + 0.005080126225948334, + 0.0037914346903562546, + 0.0031176472548395395, + 0.004941995721310377, + 0.004120744299143553, + 0.020891709253191948, + 0.010603387840092182, + 0.034767698496580124, + 0.010530764237046242, + 0.010167031548917294, + 0.011595037765800953, + 0.013819606974720955, + 0.01312524825334549, + 0.010501359589397907, + 0.003820342943072319, + 0.014485727064311504, + 0.0038090150337666273, + 0.0039789085276424885, + 0.0033270088024437428, + 0.003690464189276099, + 0.0038143021520227194, + 0.009491377510130405, + 0.010006614960730076, + 0.01214191596955061, + 0.0066409907303750515, + 0.01095775980502367, + 0.012246900238096714, + 0.0067424229346215725, + 0.011379578150808811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.07003393024206161 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the dragon boat crafted entirely from gleaming brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dragon boat sailing above a boundless, fluffy sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon head with blazing red eyes visible at the prow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive intermeshing brass gears rotating along the sides of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8982980251312256 + ], + "opposite_logprob_tokens": [ + -0.5232980251312256 + ], + "logprob_reward": -0.8982980251312256, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8982980251312256, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1321086287498474, + 0.31685471534729004, + 0.05238224193453789, + 0.07531463354825974, + 0.06585083156824112, + 0.07125187665224075, + 0.09990274161100388, + 0.05098029598593712, + 0.07523917406797409, + 0.060114890336990356 + ], + "credit_map": [ + 0.024060864001512527, + 0.012991715222597122, + 0.03557823970913887, + 0.013803360983729362, + 0.010862342081964016, + 0.00872090458869934, + 0.014158354140818119, + 0.011932837776839733, + 0.07810501754283905, + 0.018834825605154037, + 0.09396614134311676, + 0.0199725441634655, + 0.02591184340417385, + 0.015117612667381763, + 0.024079114198684692, + 0.04086760804057121, + 0.006854658480733633, + 0.006659585516899824, + 0.00921337865293026, + 0.004820296075195074, + 0.007993916980922222, + 0.004354874137789011, + 0.005782528780400753, + 0.006703006103634834, + 0.00943775288760662, + 0.005618521012365818, + 0.014393295161426067, + 0.007388560101389885, + 0.005804791115224361, + 0.0051937950775027275, + 0.009967208839952946, + 0.017510714009404182, + 0.006700355559587479, + 0.006075299810618162, + 0.010414818301796913, + 0.006619520951062441, + 0.008248252794146538, + 0.005385197699069977, + 0.01105944998562336, + 0.011347933672368526, + 0.015091955661773682, + 0.00653471564874053, + 0.011364344507455826, + 0.0055276136845350266, + 0.006603351328521967, + 0.0047197178937494755, + 0.009749598801136017, + 0.011660580523312092, + 0.014669926837086678, + 0.011876157484948635, + 0.015650831162929535, + 0.009201238863170147, + 0.008710309863090515, + 0.006989107001572847, + 0.014225794933736324, + 0.01857936382293701, + 0.010629240423440933, + 0.004882965702563524, + 0.011824053712189198, + 0.002970772562548518, + 0.003133088583126664, + 0.003477800404652953, + 0.005276830401271582, + 0.008785545825958252, + 0.008893579244613647, + 0.011934896931052208, + 0.008868008852005005, + 0.005337987560778856, + 0.010628744028508663, + 0.007875856012105942, + 0.007157564628869295, + 0.014542537741363049, + 0.00946900062263012, + 0.006443460006266832, + 0.007918260060250759, + 0.006058680359274149, + 0.004340302664786577, + 0.01522759348154068, + 0.004863606300204992, + 0.005793987773358822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.8982980251312256 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "majestic low-angle wide shot, the camera panning steadily to the right, juxtaposes a massive dragon boat, crafted entirely from gleaming brass, gracefully sailing above a boundless, fluffy sea of clouds. The golden sunlight of late afternoon bathes the frame, gleaming brilliantly off the boat's serpentine hull, meticulously riveted and ornately carved. At the prow lies a majestic dragon head, its eyes blazing red. From its flared nostrils, thick, white steam erupts rhythmically and powerfully, creating a shimmering halo in the sunlight before dissipating into the clear, high-altitude air. Along its sides, massive, intermeshing brass gears rotate with a steady, unstoppable force, driving complex pistons and connecting rods. This exposed mechanism, with its precise rhythmic movement, provides the fantastical craft with an endless supply of power. The entire scene is filled with an epic and awe-inspiring steampunk atmosphere, perfectly blending the shape of ancient mythology with the majestic power of the Industrial Revolution, presenting an extraordinary and magnificent voyage.", + "seed": 15796, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_15/index_124_seed_15796_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.13307685621548443, + "vlm_logprob": -0.13307685621548443, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.13307685621548443, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the dragon boat crafted entirely from gleaming brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dragon boat sailing above a boundless, fluffy sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon head with blazing red eyes visible at the prow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive intermeshing brass gears rotating along the sides of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08356396108865738 + ], + "opposite_logprob_tokens": [ + -2.583564043045044 + ], + "logprob_reward": -0.08356396108865738, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08356396108865738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09198281168937683, + 0.051063355058431625, + 0.04361148178577423, + 0.12582702934741974, + 0.03802090138196945, + 0.13409423828125, + 0.11222492158412933, + 0.27395114302635193, + 0.0867094174027443, + 0.04251466691493988 + ], + "credit_map": [ + 0.010616873390972614, + 0.011144633404910564, + 0.008361085318028927, + 0.009666386991739273, + 0.021015550941228867, + 0.00918255839496851, + 0.012338845059275627, + 0.009656878188252449, + 0.007535968907177448, + 0.007252271752804518, + 0.006448829080909491, + 0.0054722377099096775, + 0.008268089033663273, + 0.005019182339310646, + 0.005633221007883549, + 0.005433558952063322, + 0.006257022265344858, + 0.005295279435813427, + 0.004948214162141085, + 0.0036986428312957287, + 0.004925966262817383, + 0.006374568212777376, + 0.0065983799286186695, + 0.005513411480933428, + 0.019245192408561707, + 0.01630076952278614, + 0.015877559781074524, + 0.011503098532557487, + 0.013737915083765984, + 0.016328440979123116, + 0.01704384572803974, + 0.015790224075317383, + 0.006515789311379194, + 0.004822536837309599, + 0.003913319204002619, + 0.00360920955426991, + 0.004591765813529491, + 0.004732989240437746, + 0.00526024354621768, + 0.004575049038976431, + 0.015165976248681545, + 0.014912963844835758, + 0.009848253801465034, + 0.026122163981199265, + 0.015445036813616753, + 0.017133379355072975, + 0.021888161078095436, + 0.013578313402831554, + 0.025575200095772743, + 0.019389720633625984, + 0.010896326042711735, + 0.010425666347146034, + 0.008779422380030155, + 0.017376141622662544, + 0.01194570492953062, + 0.007836748845875263, + 0.0629367083311081, + 0.031489770859479904, + 0.03375783935189247, + 0.03147988021373749, + 0.02760854922235012, + 0.023186907172203064, + 0.03629644587635994, + 0.02719505876302719, + 0.00552527979016304, + 0.005584223661571741, + 0.012517804279923439, + 0.023367099463939667, + 0.010053596459329128, + 0.009224925190210342, + 0.014860274270176888, + 0.005576219875365496, + 0.007627873681485653, + 0.0037855394184589386, + 0.004531116224825382, + 0.004671451169997454, + 0.004913856275379658, + 0.009381965734064579, + 0.00439150957390666, + 0.00321135763078928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 3 + ], + "selected_reward": -0.08356396108865738 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the dragon boat crafted entirely from gleaming brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dragon boat sailing above a boundless, fluffy sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon head with blazing red eyes visible at the prow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive intermeshing brass gears rotating along the sides of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019488405669108033 + ], + "opposite_logprob_tokens": [ + -6.251948833465576 + ], + "logprob_reward": -0.0019488405669108033, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019488405669108033, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.150673970580101, + 0.06170246750116348, + 0.025811566039919853, + 0.04413624480366707, + 0.06000091880559921, + 0.0819084495306015, + 0.2068256288766861, + 0.2804626524448395, + 0.027468664571642876, + 0.0610094889998436 + ], + "credit_map": [ + 0.018440397456288338, + 0.018889259546995163, + 0.02813219092786312, + 0.026985669508576393, + 0.015534319914877415, + 0.01155915018171072, + 0.01639677956700325, + 0.014736199751496315, + 0.008598006330430508, + 0.008663445711135864, + 0.017123496159911156, + 0.006971206050366163, + 0.0061441450379788876, + 0.0049429950304329395, + 0.004952612332999706, + 0.004306562710553408, + 0.0036345506086945534, + 0.0025357480626553297, + 0.006680553313344717, + 0.004140275530517101, + 0.0025000444147735834, + 0.0023592612706124783, + 0.00196929182857275, + 0.001991839613765478, + 0.006186827551573515, + 0.007976379245519638, + 0.006175185553729534, + 0.004319831263273954, + 0.004682144150137901, + 0.005179763305932283, + 0.003958522342145443, + 0.005657589063048363, + 0.009335464797914028, + 0.005476427264511585, + 0.014928165823221207, + 0.004447459243237972, + 0.008050858974456787, + 0.007602091412991285, + 0.005441480316221714, + 0.00471897516399622, + 0.013381966389715672, + 0.015124820172786713, + 0.009356508031487465, + 0.010456433519721031, + 0.007803803775459528, + 0.01104583777487278, + 0.008721786551177502, + 0.006017296575009823, + 0.03376572206616402, + 0.02190694399178028, + 0.022377341985702515, + 0.029055703431367874, + 0.03089194931089878, + 0.02436434105038643, + 0.03084375336766243, + 0.013619869947433472, + 0.04235284775495529, + 0.026515914127230644, + 0.03660696744918823, + 0.03917597234249115, + 0.03659440204501152, + 0.026728356257081032, + 0.039154693484306335, + 0.03333350270986557, + 0.0024977412540465593, + 0.0020884056575596333, + 0.003772290423512459, + 0.006673351861536503, + 0.003363080555573106, + 0.003053101245313883, + 0.0041589378379285336, + 0.0018617534078657627, + 0.007652780041098595, + 0.006563868373632431, + 0.007523755077272654, + 0.007812838070094585, + 0.010265449061989784, + 0.009522749111056328, + 0.006977243814617395, + 0.004690810572355986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.0019488405669108033 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the dragon boat crafted entirely from gleaming brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dragon boat sailing above a boundless, fluffy sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon head with blazing red eyes visible at the prow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive intermeshing brass gears rotating along the sides of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11290860176086426 + ], + "opposite_logprob_tokens": [ + -2.2379086017608643 + ], + "logprob_reward": -0.11290860176086426, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11290860176086426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22092436254024506, + 0.04443731531500816, + 0.04891905188560486, + 0.13395410776138306, + 0.06960421055555344, + 0.14687852561473846, + 0.049943458288908005, + 0.16746993362903595, + 0.06898565590381622, + 0.048883356153964996 + ], + "credit_map": [ + 0.026906223967671394, + 0.027716118842363358, + 0.02301471307873726, + 0.022271638736128807, + 0.03745388612151146, + 0.022466083988547325, + 0.035453300923109055, + 0.025642404332756996, + 0.007501474115997553, + 0.007219132035970688, + 0.010416329838335514, + 0.00505458377301693, + 0.004711316432803869, + 0.0036055520176887512, + 0.0032552643679082394, + 0.002673664828762412, + 0.007327768485993147, + 0.005766313057392836, + 0.006545763462781906, + 0.004410641733556986, + 0.007720073219388723, + 0.007143189199268818, + 0.004439433105289936, + 0.00556587427854538, + 0.021348416805267334, + 0.030439769849181175, + 0.02241402678191662, + 0.014397372491657734, + 0.012717673555016518, + 0.01110774651169777, + 0.009782612323760986, + 0.01174649316817522, + 0.009447094053030014, + 0.006439251825213432, + 0.010074452497065067, + 0.008950314484536648, + 0.007109376601874828, + 0.006474948953837156, + 0.01234439481049776, + 0.008764379657804966, + 0.014244458638131618, + 0.02284742332994938, + 0.021845674142241478, + 0.028355766087770462, + 0.011914161033928394, + 0.015940066426992416, + 0.019704055041074753, + 0.012026922777295113, + 0.005417102947831154, + 0.007254347670823336, + 0.0067466869950294495, + 0.006754336878657341, + 0.006039972882717848, + 0.007656303700059652, + 0.005939103662967682, + 0.004135605879127979, + 0.04613398015499115, + 0.021261870861053467, + 0.024867434054613113, + 0.012653643265366554, + 0.014228043146431446, + 0.01448028814047575, + 0.01969248615205288, + 0.014152185060083866, + 0.004089499823749065, + 0.00484258821234107, + 0.009728391654789448, + 0.017496420070528984, + 0.008479919284582138, + 0.007554675918072462, + 0.012014892883598804, + 0.004779272247105837, + 0.007880807854235172, + 0.004241079557687044, + 0.006411529146134853, + 0.006861767265945673, + 0.005309345666319132, + 0.009242089465260506, + 0.004917020909488201, + 0.0040197186172008514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.11290860176086426 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the dragon boat crafted entirely from gleaming brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dragon boat sailing above a boundless, fluffy sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon head with blazing red eyes visible at the prow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive intermeshing brass gears rotating along the sides of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038095925003290176 + ], + "opposite_logprob_tokens": [ + -3.2880959510803223 + ], + "logprob_reward": -0.038095925003290176, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038095925003290176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056703004986047745, + 0.03733210265636444, + 0.030666278675198555, + 0.042699385434389114, + 0.1214069202542305, + 0.1207379549741745, + 0.12967799603939056, + 0.1492069810628891, + 0.08164414763450623, + 0.22992518544197083 + ], + "credit_map": [ + 0.006310128141194582, + 0.0067567587830126286, + 0.006871463730931282, + 0.006576984189450741, + 0.008708993904292583, + 0.0053845057263970375, + 0.00823386013507843, + 0.007860311307013035, + 0.005432309117168188, + 0.004650014452636242, + 0.005462751258164644, + 0.0036995355039834976, + 0.005939783062785864, + 0.0027147727087140083, + 0.00424687284976244, + 0.005186063703149557, + 0.004198306705802679, + 0.003765057073906064, + 0.0039585973136126995, + 0.0031745729502290487, + 0.004922512918710709, + 0.004071153700351715, + 0.0033705560490489006, + 0.0032055231276899576, + 0.0063495333306491375, + 0.007653117645531893, + 0.006321093533188105, + 0.005028794053941965, + 0.003623929340392351, + 0.0041916146874427795, + 0.00457011116668582, + 0.00496118888258934, + 0.016087861731648445, + 0.011883669532835484, + 0.01575176604092121, + 0.011093411594629288, + 0.019903510808944702, + 0.016818316653370857, + 0.019168103113770485, + 0.010700284503400326, + 0.018198242411017418, + 0.023346664384007454, + 0.01151072233915329, + 0.018332982435822487, + 0.010692954063415527, + 0.018958307802677155, + 0.012163358740508556, + 0.007534720469266176, + 0.02608661912381649, + 0.020445534959435463, + 0.013955999165773392, + 0.012399431318044662, + 0.014505194500088692, + 0.017243951559066772, + 0.015168694779276848, + 0.009872576221823692, + 0.031651731580495834, + 0.014670221135020256, + 0.018781710416078568, + 0.0165767353028059, + 0.018207747489213943, + 0.013630390167236328, + 0.020866598933935165, + 0.014821852557361126, + 0.007712770719081163, + 0.006627196911722422, + 0.01442379504442215, + 0.014013847336173058, + 0.009618936106562614, + 0.009277336299419403, + 0.013006158173084259, + 0.006964107509702444, + 0.026223015040159225, + 0.02695247158408165, + 0.024911530315876007, + 0.028429580852389336, + 0.04104062169790268, + 0.03655463829636574, + 0.02698175422847271, + 0.018831579014658928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.038095925003290176 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the dragon boat crafted entirely from gleaming brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dragon boat sailing above a boundless, fluffy sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon head with blazing red eyes visible at the prow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive intermeshing brass gears rotating along the sides of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4288669526576996 + ], + "opposite_logprob_tokens": [ + -1.053866982460022 + ], + "logprob_reward": -0.4288669526576996, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4288669526576996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.094357430934906, + 0.03109622560441494, + 0.045380670577287674, + 0.045419059693813324, + 0.05469213426113129, + 0.08311777561903, + 0.2816362977027893, + 0.29841235280036926, + 0.045612070709466934, + 0.02027599699795246 + ], + "credit_map": [ + 0.010914171114563942, + 0.011445865035057068, + 0.013006390072405338, + 0.011155185289680958, + 0.016368230804800987, + 0.006706571206450462, + 0.01406850665807724, + 0.010692516341805458, + 0.003650995437055826, + 0.004011237993836403, + 0.00605800049379468, + 0.0036082572769373655, + 0.003706264542415738, + 0.0027558107394725084, + 0.0033569103106856346, + 0.003948749974370003, + 0.006693598814308643, + 0.004424178972840309, + 0.010839204303920269, + 0.007099245209246874, + 0.0047368998639285564, + 0.003923006355762482, + 0.0037315695080906153, + 0.003932966850697994, + 0.0062173763290047646, + 0.006809353828430176, + 0.007861216552555561, + 0.00451055821031332, + 0.004258306697010994, + 0.005873110145330429, + 0.00530781876295805, + 0.004581316839903593, + 0.007663722615689039, + 0.005177423823624849, + 0.010895502753555775, + 0.005254807882010937, + 0.008378433994948864, + 0.0075639598071575165, + 0.0046045370399951935, + 0.005153742153197527, + 0.012652921490371227, + 0.013566633686423302, + 0.0088104959577322, + 0.014124184846878052, + 0.007569433655589819, + 0.010669331066310406, + 0.009408125653862953, + 0.006316644139587879, + 0.05207560211420059, + 0.03202744200825691, + 0.029602471739053726, + 0.0375315323472023, + 0.03856699541211128, + 0.03322475031018257, + 0.03994767740368843, + 0.01865982636809349, + 0.04924416169524193, + 0.026891382411122322, + 0.038016244769096375, + 0.04057999327778816, + 0.03748331218957901, + 0.027870502322912216, + 0.041641637682914734, + 0.03668511286377907, + 0.004530610982328653, + 0.004078548867255449, + 0.007466999813914299, + 0.010826760903000832, + 0.003429287811741233, + 0.004708359949290752, + 0.00691487081348896, + 0.003656631102785468, + 0.0034329837653785944, + 0.0017055635107681155, + 0.003824667539447546, + 0.0024018946569412947, + 0.0022751837968826294, + 0.003599753137677908, + 0.0017576261889189482, + 0.0012783246347680688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.4288669526576996 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "majestic low-angle wide shot, the camera panning steadily to the right, juxtaposes a massive dragon boat, crafted entirely from gleaming brass, gracefully sailing above a boundless, fluffy sea of clouds. The golden sunlight of late afternoon bathes the frame, gleaming brilliantly off the boat's serpentine hull, meticulously riveted and ornately carved. At the prow lies a majestic dragon head, its eyes blazing red. From its flared nostrils, thick, white steam erupts rhythmically and powerfully, creating a shimmering halo in the sunlight before dissipating into the clear, high-altitude air. Along its sides, massive, intermeshing brass gears rotate with a steady, unstoppable force, driving complex pistons and connecting rods. This exposed mechanism, with its precise rhythmic movement, provides the fantastical craft with an endless supply of power. The entire scene is filled with an epic and awe-inspiring steampunk atmosphere, perfectly blending the shape of ancient mythology with the majestic power of the Industrial Revolution, presenting an extraordinary and magnificent voyage.", + "seed": 15797, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_15/index_124_seed_15797_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04691862240433693, + "vlm_logprob": -0.04691862240433693, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.04691862240433693, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the dragon boat crafted entirely from gleaming brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dragon boat sailing above a boundless, fluffy sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon head with blazing red eyes visible at the prow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive intermeshing brass gears rotating along the sides of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16979825496673584 + ], + "opposite_logprob_tokens": [ + -1.9197982549667358 + ], + "logprob_reward": -0.16979825496673584, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16979825496673584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19611729681491852, + 0.235295370221138, + 0.08543016016483307, + 0.03901204094290733, + 0.02905942127108574, + 0.03588591143488884, + 0.07874587178230286, + 0.07828566431999207, + 0.06110990419983864, + 0.16105830669403076 + ], + "credit_map": [ + 0.030696643516421318, + 0.015679342672228813, + 0.04032197222113609, + 0.03127570077776909, + 0.020969370380043983, + 0.021589353680610657, + 0.012826720252633095, + 0.022758206352591515, + 0.027691781520843506, + 0.029583541676402092, + 0.02838755212724209, + 0.033551570028066635, + 0.02860383503139019, + 0.01962081342935562, + 0.021686527878046036, + 0.04616975039243698, + 0.00792740285396576, + 0.009872142225503922, + 0.006671605631709099, + 0.007368758320808411, + 0.013286243192851543, + 0.00982944667339325, + 0.014365298673510551, + 0.016109265387058258, + 0.004643001593649387, + 0.004580877721309662, + 0.0031916650477796793, + 0.0032822086941450834, + 0.007759594824165106, + 0.006070360541343689, + 0.004543017130345106, + 0.004941313993185759, + 0.005451231263577938, + 0.004536300431936979, + 0.0023944976273924112, + 0.002484122524037957, + 0.0038486067205667496, + 0.003726329654455185, + 0.0035383354406803846, + 0.003079996444284916, + 0.003903446951881051, + 0.005051437299698591, + 0.004292525351047516, + 0.0036308413837105036, + 0.004588298033922911, + 0.00501500628888607, + 0.004701771307736635, + 0.004702585283666849, + 0.007391572929918766, + 0.01249451469630003, + 0.007911588996648788, + 0.008359909057617188, + 0.011003593914210796, + 0.00982038676738739, + 0.011687777005136013, + 0.01007652934640646, + 0.009239441715180874, + 0.00992991030216217, + 0.007654370740056038, + 0.00880197249352932, + 0.01016361452639103, + 0.011132003739476204, + 0.011340083554387093, + 0.010024267248809338, + 0.006452657748013735, + 0.007441021967679262, + 0.004501148127019405, + 0.006265639327466488, + 0.011078841984272003, + 0.00807297695428133, + 0.009385268203914165, + 0.007912348955869675, + 0.01733422465622425, + 0.021292494609951973, + 0.02935517206788063, + 0.0178107637912035, + 0.0211421400308609, + 0.021292272955179214, + 0.014995366334915161, + 0.017835870385169983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.16979825496673584 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the dragon boat crafted entirely from gleaming brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dragon boat sailing above a boundless, fluffy sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon head with blazing red eyes visible at the prow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive intermeshing brass gears rotating along the sides of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011091128922998905 + ], + "opposite_logprob_tokens": [ + -4.511091232299805 + ], + "logprob_reward": -0.011091128922998905, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011091128922998905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27471327781677246, + 0.08282048255205154, + 0.049078911542892456, + 0.03846421465277672, + 0.04293113574385643, + 0.058448124676942825, + 0.044936228543519974, + 0.10381536185741425, + 0.03440198674798012, + 0.27039027214050293 + ], + "credit_map": [ + 0.036517687141895294, + 0.02236085571348667, + 0.051097337156534195, + 0.03223361074924469, + 0.027766581624746323, + 0.02950620837509632, + 0.02359810471534729, + 0.05163288488984108, + 0.009911127388477325, + 0.008007404394447803, + 0.006673101801425219, + 0.010621195659041405, + 0.010440978221595287, + 0.007952234707772732, + 0.011943044140934944, + 0.017271392047405243, + 0.004900022875517607, + 0.005468738730996847, + 0.0049177855253219604, + 0.0035356907173991203, + 0.007302246522158384, + 0.006386671215295792, + 0.00865269172936678, + 0.007915066555142403, + 0.002997074043378234, + 0.0027429640758782625, + 0.0018831549677997828, + 0.002090337686240673, + 0.00860411487519741, + 0.008322575129568577, + 0.004656612407416105, + 0.007167384028434753, + 0.005054526496678591, + 0.005502012092620134, + 0.00346275232732296, + 0.0034599516075104475, + 0.0093426164239645, + 0.007403797935694456, + 0.004777306225150824, + 0.0039281705394387245, + 0.005647050682455301, + 0.006879960186779499, + 0.006573954597115517, + 0.005267546512186527, + 0.008781101554632187, + 0.008019577711820602, + 0.008519914932549, + 0.008759021759033203, + 0.006181758362799883, + 0.006998632103204727, + 0.004568879958242178, + 0.003999097738415003, + 0.006523031275719404, + 0.005681520793586969, + 0.005413038656115532, + 0.005570268724113703, + 0.011217632330954075, + 0.014897593297064304, + 0.01321397628635168, + 0.015353787690401077, + 0.013252714648842812, + 0.007752067409455776, + 0.015089398249983788, + 0.013038190081715584, + 0.0033732473384588957, + 0.0036807048600167036, + 0.003166492097079754, + 0.0031674534548074007, + 0.005838213488459587, + 0.00773129565641284, + 0.0032915216870605946, + 0.004153059795498848, + 0.026803530752658844, + 0.055139727890491486, + 0.027479764074087143, + 0.024440033361315727, + 0.04863327741622925, + 0.023915503174066544, + 0.025804445147514343, + 0.038174014538526535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.011091128922998905 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the dragon boat crafted entirely from gleaming brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dragon boat sailing above a boundless, fluffy sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon head with blazing red eyes visible at the prow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive intermeshing brass gears rotating along the sides of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012538178823888302 + ], + "opposite_logprob_tokens": [ + -4.387537956237793 + ], + "logprob_reward": -0.012538178823888302, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012538178823888302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20579802989959717, + 0.4570813477039337, + 0.04441710561513901, + 0.038541827350854874, + 0.03392060101032257, + 0.05778210610151291, + 0.03517945483326912, + 0.018816815689206123, + 0.05564215034246445, + 0.052820559591054916 + ], + "credit_map": [ + 0.031187038868665695, + 0.021563110873103142, + 0.029186511412262917, + 0.024988306686282158, + 0.030086807906627655, + 0.026271166279911995, + 0.02011273056268692, + 0.02240234799683094, + 0.05556905269622803, + 0.11359775066375732, + 0.03613021969795227, + 0.06218559294939041, + 0.05818954110145569, + 0.037227872759103775, + 0.043722230941057205, + 0.05045907944440842, + 0.005482299719005823, + 0.00800894945859909, + 0.002965284511446953, + 0.004184001591056585, + 0.007953495718538761, + 0.004993220325559378, + 0.004882934037595987, + 0.005946924444288015, + 0.005377352237701416, + 0.0038825024385005236, + 0.003594840643927455, + 0.002857285551726818, + 0.005779950879514217, + 0.00561130978167057, + 0.004948981571942568, + 0.006489605642855167, + 0.004099822603166103, + 0.006056069862097502, + 0.004728016443550587, + 0.004127668682485819, + 0.003940278198570013, + 0.004768692422658205, + 0.003180756466463208, + 0.003019295632839203, + 0.007160937413573265, + 0.008002731949090958, + 0.006125930696725845, + 0.005060710944235325, + 0.010185468941926956, + 0.006540223024785519, + 0.007005124352872372, + 0.0077009801752865314, + 0.003932961728423834, + 0.005235062912106514, + 0.0050577688962221146, + 0.004449802450835705, + 0.004985483828932047, + 0.003270870540291071, + 0.003807902103289962, + 0.0044396002776920795, + 0.0022294840309768915, + 0.0018075167899951339, + 0.0013902001082897186, + 0.0017487966688349843, + 0.0023685183841735125, + 0.003170439274981618, + 0.0029240241274237633, + 0.003177837235853076, + 0.004374849144369364, + 0.006795359309762716, + 0.006228088401257992, + 0.007117909379303455, + 0.00962721835821867, + 0.007841135375201702, + 0.0068245623260736465, + 0.006833027116954327, + 0.005990867502987385, + 0.006224615033715963, + 0.009695375338196754, + 0.0053674280643463135, + 0.009895778261125088, + 0.005273310001939535, + 0.005269187502563, + 0.005103999748826027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.012538178823888302 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the dragon boat crafted entirely from gleaming brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dragon boat sailing above a boundless, fluffy sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon head with blazing red eyes visible at the prow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive intermeshing brass gears rotating along the sides of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058347500860691 + ], + "opposite_logprob_tokens": [ + -3.895583391189575 + ], + "logprob_reward": -0.02058347500860691, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058347500860691, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11962312459945679, + 0.124026820063591, + 0.0554620623588562, + 0.06571850925683975, + 0.08690416812896729, + 0.049722734838724136, + 0.05547846853733063, + 0.19766157865524292, + 0.08627784997224808, + 0.15912464261054993 + ], + "credit_map": [ + 0.017317622900009155, + 0.010692419484257698, + 0.009346304461359978, + 0.011481582187116146, + 0.017075655981898308, + 0.01427630428224802, + 0.014548402279615402, + 0.024884840473532677, + 0.014322532340884209, + 0.01615907996892929, + 0.008897324092686176, + 0.015047363005578518, + 0.019544977694749832, + 0.010780959390103817, + 0.017166005447506905, + 0.022108573466539383, + 0.006254163570702076, + 0.008075632154941559, + 0.004187031649053097, + 0.004365075379610062, + 0.009359940886497498, + 0.007140294183045626, + 0.006522097159177065, + 0.009557832963764668, + 0.007164095062762499, + 0.003860211232677102, + 0.005133204627782106, + 0.0034555799793452024, + 0.015301957726478577, + 0.013339693658053875, + 0.008827351033687592, + 0.008636422455310822, + 0.010918657295405865, + 0.00872753281146288, + 0.007436311338096857, + 0.00697407266125083, + 0.011976541951298714, + 0.010880344547331333, + 0.016563216224312782, + 0.013427495025098324, + 0.005324116442352533, + 0.004811777267605066, + 0.004717017989605665, + 0.0053090001456439495, + 0.007794365286827087, + 0.008459938690066338, + 0.006051019765436649, + 0.007255500182509422, + 0.004669552203267813, + 0.008424841798841953, + 0.00394798070192337, + 0.006448423024266958, + 0.006831585895270109, + 0.008744592778384686, + 0.008429855108261108, + 0.007981639355421066, + 0.019129255786538124, + 0.017530368641018867, + 0.013918697834014893, + 0.022118378430604935, + 0.02244221605360508, + 0.0275572519749403, + 0.036504440009593964, + 0.03846099600195885, + 0.007698703091591597, + 0.007915369234979153, + 0.006289815530180931, + 0.008597353473305702, + 0.012189527042210102, + 0.01457828190177679, + 0.013815543614327908, + 0.015193254686892033, + 0.016172070056200027, + 0.018709572032094002, + 0.02276247926056385, + 0.018160870298743248, + 0.020281225442886353, + 0.016496265307068825, + 0.025945328176021576, + 0.020596826449036598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 1 + ], + "selected_reward": -0.02058347500860691 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the dragon boat crafted entirely from gleaming brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dragon boat sailing above a boundless, fluffy sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon head with blazing red eyes visible at the prow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive intermeshing brass gears rotating along the sides of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058207429945469 + ], + "opposite_logprob_tokens": [ + -3.8955819606781006 + ], + "logprob_reward": -0.02058207429945469, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058207429945469, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07189720124006271, + 0.07147655636072159, + 0.0682213231921196, + 0.09672893583774567, + 0.10419075936079025, + 0.05423934757709503, + 0.07623296231031418, + 0.08175840973854065, + 0.33567824959754944, + 0.039576202630996704 + ], + "credit_map": [ + 0.010021432302892208, + 0.006616555619984865, + 0.012809979729354382, + 0.007813326083123684, + 0.007235965691506863, + 0.006941677536815405, + 0.006514419801533222, + 0.013943853788077831, + 0.008549065329134464, + 0.008705935440957546, + 0.006988585460931063, + 0.010772915557026863, + 0.008909467607736588, + 0.006242789328098297, + 0.010414369404315948, + 0.010893434286117554, + 0.006714601535350084, + 0.008653771132230759, + 0.005069712642580271, + 0.006231891456991434, + 0.009767399169504642, + 0.0072975377552211285, + 0.012986304238438606, + 0.011500108055770397, + 0.014855977147817612, + 0.008964559063315392, + 0.010334430262446404, + 0.006450389511883259, + 0.01634748838841915, + 0.015356482937932014, + 0.010622058063745499, + 0.013797558844089508, + 0.009173164144158363, + 0.014512371271848679, + 0.010729622095823288, + 0.00949230045080185, + 0.018969912081956863, + 0.017818808555603027, + 0.01350833848118782, + 0.009986245073378086, + 0.00457912078127265, + 0.0063002160750329494, + 0.00583789125084877, + 0.0047532739117741585, + 0.006317879538983107, + 0.007927331142127514, + 0.009303473867475986, + 0.009220167994499207, + 0.007354419678449631, + 0.009883862920105457, + 0.007060835137963295, + 0.009088985621929169, + 0.009579477831721306, + 0.012110510841012001, + 0.010934893973171711, + 0.010219983756542206, + 0.009752589277923107, + 0.015585795044898987, + 0.00675587123259902, + 0.012512559071183205, + 0.012601143680512905, + 0.008847140707075596, + 0.007256327196955681, + 0.008446990512311459, + 0.02728365920484066, + 0.03460083156824112, + 0.024255061522126198, + 0.040844809263944626, + 0.05202849954366684, + 0.03853848576545715, + 0.057240549474954605, + 0.06088639050722122, + 0.004077022895216942, + 0.004786858335137367, + 0.006210130639374256, + 0.003342088544741273, + 0.00897647812962532, + 0.003835591720417142, + 0.00411426555365324, + 0.0042337714694440365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 3 + ], + "selected_reward": -0.02058207429945469 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "majestic low-angle wide shot, the camera panning steadily to the right, juxtaposes a massive dragon boat, crafted entirely from gleaming brass, gracefully sailing above a boundless, fluffy sea of clouds. The golden sunlight of late afternoon bathes the frame, gleaming brilliantly off the boat's serpentine hull, meticulously riveted and ornately carved. At the prow lies a majestic dragon head, its eyes blazing red. From its flared nostrils, thick, white steam erupts rhythmically and powerfully, creating a shimmering halo in the sunlight before dissipating into the clear, high-altitude air. Along its sides, massive, intermeshing brass gears rotate with a steady, unstoppable force, driving complex pistons and connecting rods. This exposed mechanism, with its precise rhythmic movement, provides the fantastical craft with an endless supply of power. The entire scene is filled with an epic and awe-inspiring steampunk atmosphere, perfectly blending the shape of ancient mythology with the majestic power of the Industrial Revolution, presenting an extraordinary and magnificent voyage.", + "seed": 15798, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_15/index_124_seed_15798_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.20983574464917182, + "vlm_logprob": -0.20983574464917182, + "vlm_yes_no_margin": 2.200000011920929, + "vlm_token_credit": -0.20983574464917182, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the dragon boat crafted entirely from gleaming brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dragon boat sailing above a boundless, fluffy sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon head with blazing red eyes visible at the prow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive intermeshing brass gears rotating along the sides of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5841845870018005 + ], + "opposite_logprob_tokens": [ + -0.8341845870018005 + ], + "logprob_reward": -0.5841845870018005, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5841845870018005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18117991089820862, + 0.06848833709955215, + 0.04716486856341362, + 0.042901113629341125, + 0.12158916145563126, + 0.05757703632116318, + 0.3567056655883789, + 0.04783805459737778, + 0.052683647722005844, + 0.02387225814163685 + ], + "credit_map": [ + 0.019530124962329865, + 0.026519574224948883, + 0.016226189211010933, + 0.012865758500993252, + 0.022970963269472122, + 0.028018584474921227, + 0.02827698551118374, + 0.02677171677350998, + 0.011490094475448132, + 0.009111061692237854, + 0.00849163718521595, + 0.0063396962359547615, + 0.007071677595376968, + 0.011470012366771698, + 0.00657226936891675, + 0.00794188678264618, + 0.005304798949509859, + 0.006529959850013256, + 0.004032684024423361, + 0.0029207223560661077, + 0.010848847217857838, + 0.008742135018110275, + 0.005574662238359451, + 0.0032110579777508974, + 0.005377635359764099, + 0.0049257115460932255, + 0.003953950013965368, + 0.006736030336469412, + 0.004820432048290968, + 0.0042456104420125484, + 0.007124013267457485, + 0.005717728286981583, + 0.023708054795861244, + 0.020488200709223747, + 0.012556334026157856, + 0.012780622579157352, + 0.019995076581835747, + 0.009710106998682022, + 0.012537617236375809, + 0.009813141077756882, + 0.008889579214155674, + 0.00856354832649231, + 0.007154608145356178, + 0.01081226859241724, + 0.008417978882789612, + 0.003941718488931656, + 0.004542630165815353, + 0.0052547091618180275, + 0.017792347818613052, + 0.0840841606259346, + 0.024484576657414436, + 0.029386959969997406, + 0.030785052105784416, + 0.10579875856637955, + 0.0296672023832798, + 0.03470660746097565, + 0.00512854615226388, + 0.007502848282456398, + 0.0032463662791997194, + 0.004235745873302221, + 0.005831333342939615, + 0.009696580469608307, + 0.004833937622606754, + 0.007362697273492813, + 0.004086700733751059, + 0.005213270895183086, + 0.004903168883174658, + 0.00614791177213192, + 0.0051138498820364475, + 0.006643929984420538, + 0.01157610397785902, + 0.008998711593449116, + 0.002799779875203967, + 0.0032349014654755592, + 0.0017703609773889184, + 0.0026787095703184605, + 0.003235533367842436, + 0.002833528444170952, + 0.004741276614367962, + 0.002578168176114559 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.5841845870018005 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the dragon boat crafted entirely from gleaming brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dragon boat sailing above a boundless, fluffy sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon head with blazing red eyes visible at the prow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive intermeshing brass gears rotating along the sides of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331206575036049 + ], + "opposite_logprob_tokens": [ + -3.7733120918273926 + ], + "logprob_reward": -0.02331206575036049, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331206575036049, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2628489136695862, + 0.14917488396167755, + 0.08647559583187103, + 0.03753986209630966, + 0.0828297808766365, + 0.029982220381498337, + 0.07320769131183624, + 0.07347226142883301, + 0.16214939951896667, + 0.0423194020986557 + ], + "credit_map": [ + 0.020823754370212555, + 0.024070216342806816, + 0.029813820496201515, + 0.02022830955684185, + 0.03544769808650017, + 0.03432874009013176, + 0.05398799851536751, + 0.04414839297533035, + 0.01187180820852518, + 0.012248441576957703, + 0.007756923791021109, + 0.013681670650839806, + 0.018191341310739517, + 0.023099787533283234, + 0.029047146439552307, + 0.03327776864171028, + 0.008789636194705963, + 0.01067723985761404, + 0.0064773051999509335, + 0.008277668617665768, + 0.011999841779470444, + 0.012410870753228664, + 0.0158084724098444, + 0.012034560553729534, + 0.003408854128792882, + 0.0034010952804237604, + 0.0031348196789622307, + 0.0038321439642459154, + 0.00542700570076704, + 0.00599122466519475, + 0.006806258577853441, + 0.005538462661206722, + 0.010745040141046047, + 0.01746537536382675, + 0.007385284174233675, + 0.010554011911153793, + 0.01490215864032507, + 0.005628136917948723, + 0.007971592247486115, + 0.00817817822098732, + 0.003287708153948188, + 0.0045048994943499565, + 0.0028223113622516394, + 0.006291999947279692, + 0.004270907957106829, + 0.0025389804504811764, + 0.003590293927118182, + 0.002675119088962674, + 0.006922232918441296, + 0.011293061077594757, + 0.0056232875213027, + 0.009397058747708797, + 0.009824860841035843, + 0.012804686091840267, + 0.008765396662056446, + 0.008577106520533562, + 0.0075825112871825695, + 0.012280864641070366, + 0.005453596357256174, + 0.010636338964104652, + 0.009133096784353256, + 0.009956694208085537, + 0.010153593495488167, + 0.008275561966001987, + 0.01493766438215971, + 0.01623050682246685, + 0.018436722457408905, + 0.01895304024219513, + 0.015466522425413132, + 0.016877399757504463, + 0.029340719804167747, + 0.03190680593252182, + 0.003556717187166214, + 0.005785685498267412, + 0.003110139165073633, + 0.0051803854294121265, + 0.006135828327387571, + 0.004727631341665983, + 0.0067503200843930244, + 0.007072694133967161 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.02331206575036049 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the dragon boat crafted entirely from gleaming brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dragon boat sailing above a boundless, fluffy sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon head with blazing red eyes visible at the prow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive intermeshing brass gears rotating along the sides of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809374198317528 + ], + "opposite_logprob_tokens": [ + -3.2880938053131104 + ], + "logprob_reward": -0.03809374198317528, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809374198317528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2524508833885193, + 0.2028159201145172, + 0.05957973375916481, + 0.08535096794366837, + 0.17089293897151947, + 0.09183965623378754, + 0.049915824085474014, + 0.021355852484703064, + 0.03929552808403969, + 0.02650269865989685 + ], + "credit_map": [ + 0.033652763813734055, + 0.04336061701178551, + 0.024580558761954308, + 0.02970770187675953, + 0.02927246503531933, + 0.03466722369194031, + 0.0262374859303236, + 0.030972059816122055, + 0.031009985134005547, + 0.031307902187108994, + 0.020795749500393867, + 0.01689893938601017, + 0.02518717758357525, + 0.03778257966041565, + 0.017845572903752327, + 0.021988017484545708, + 0.012278830632567406, + 0.008104993961751461, + 0.00667893560603261, + 0.004305207636207342, + 0.011291546747088432, + 0.008775768801569939, + 0.004055235534906387, + 0.0040892139077186584, + 0.016654320061206818, + 0.009138762019574642, + 0.012534703128039837, + 0.010234164074063301, + 0.010990038514137268, + 0.009986027143895626, + 0.008280538953840733, + 0.007532414048910141, + 0.033483829349279404, + 0.027649207040667534, + 0.024892514571547508, + 0.018788468092679977, + 0.023103486746549606, + 0.01289381179958582, + 0.01578965224325657, + 0.014291970059275627, + 0.02093530260026455, + 0.011631779372692108, + 0.015679875388741493, + 0.0103784603998065, + 0.013659286312758923, + 0.007925871759653091, + 0.005633351393043995, + 0.005995730869472027, + 0.004073933698236942, + 0.01105903834104538, + 0.0056529175490140915, + 0.003895633388310671, + 0.004652060568332672, + 0.010565645061433315, + 0.005137243773788214, + 0.004879351239651442, + 0.001854761503636837, + 0.003247736720368266, + 0.0025620863307267427, + 0.0024787113070487976, + 0.0025845167692750692, + 0.002791722072288394, + 0.002977066906169057, + 0.0028592520393431187, + 0.004020439460873604, + 0.003811540547758341, + 0.005045028403401375, + 0.00457732705399394, + 0.004508295562118292, + 0.0038966857828199863, + 0.005844778846949339, + 0.007591436617076397, + 0.0027835022192448378, + 0.003215064760297537, + 0.0027185450308024883, + 0.0027541297022253275, + 0.00341100268997252, + 0.0038704494945704937, + 0.0039015812799334526, + 0.0038484225515276194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.03809374198317528 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the dragon boat crafted entirely from gleaming brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dragon boat sailing above a boundless, fluffy sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon head with blazing red eyes visible at the prow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive intermeshing brass gears rotating along the sides of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054967496544122696 + ], + "opposite_logprob_tokens": [ + -2.9299674034118652 + ], + "logprob_reward": -0.054967496544122696, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054967496544122696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18322420120239258, + 0.08793669193983078, + 0.06299342215061188, + 0.12406712770462036, + 0.10751210153102875, + 0.08881807327270508, + 0.16502046585083008, + 0.03528265655040741, + 0.0931430384516716, + 0.05200226232409477 + ], + "credit_map": [ + 0.02028108574450016, + 0.018898077309131622, + 0.012455521151423454, + 0.01710592769086361, + 0.023740166798233986, + 0.026663215830922127, + 0.03769940510392189, + 0.026380807161331177, + 0.013727053999900818, + 0.013138948939740658, + 0.01001441478729248, + 0.00773769011721015, + 0.009599009528756142, + 0.014412506483495235, + 0.008294791914522648, + 0.011012273840606213, + 0.011851374059915543, + 0.007614183239638805, + 0.008217236027121544, + 0.005366991739720106, + 0.012207932770252228, + 0.009607364423573017, + 0.004151796922087669, + 0.0039765434339642525, + 0.019476858898997307, + 0.012544273398816586, + 0.01900480128824711, + 0.018198847770690918, + 0.013011861592531204, + 0.009493257850408554, + 0.01960800588130951, + 0.012729216367006302, + 0.015546874143183231, + 0.014873165637254715, + 0.010721832513809204, + 0.011955322697758675, + 0.014589340426027775, + 0.013983994722366333, + 0.013031425885856152, + 0.012810141779482365, + 0.020121866837143898, + 0.012151514180004597, + 0.008987624198198318, + 0.013838310725986958, + 0.012542780488729477, + 0.006598552688956261, + 0.007129683159291744, + 0.007447737269103527, + 0.012270326726138592, + 0.03297144174575806, + 0.0102113401517272, + 0.015459895133972168, + 0.018049510195851326, + 0.03229058161377907, + 0.015331982634961605, + 0.028435397893190384, + 0.00441371975466609, + 0.003853157628327608, + 0.004529823083430529, + 0.004718449432402849, + 0.0033123602624982595, + 0.003796611912548542, + 0.005678930319845676, + 0.0049796015955507755, + 0.009602786973118782, + 0.009438025765120983, + 0.013221150264143944, + 0.009183323942124844, + 0.008538920432329178, + 0.00816915463656187, + 0.014702382497489452, + 0.0202872846275568, + 0.003988784737884998, + 0.005951638333499432, + 0.004485949408262968, + 0.007244476117193699, + 0.0076392763294279575, + 0.007306376937776804, + 0.006745436694473028, + 0.008640321902930737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.054967496544122696 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the dragon boat crafted entirely from gleaming brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dragon boat sailing above a boundless, fluffy sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon head with blazing red eyes visible at the prow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive intermeshing brass gears rotating along the sides of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34862083196640015 + ], + "opposite_logprob_tokens": [ + -1.223620891571045 + ], + "logprob_reward": -0.34862083196640015, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.34862083196640015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23033463954925537, + 0.04885396733880043, + 0.102823406457901, + 0.12541678547859192, + 0.0592954121530056, + 0.055050116032361984, + 0.14734645187854767, + 0.04825877398252487, + 0.06289620697498322, + 0.1197243258357048 + ], + "credit_map": [ + 0.035791873931884766, + 0.03341572359204292, + 0.025784987956285477, + 0.02124905027449131, + 0.026225320994853973, + 0.033794060349464417, + 0.027936764061450958, + 0.026136839762330055, + 0.005400129593908787, + 0.0072635761462152, + 0.004481114447116852, + 0.004460140597075224, + 0.007066127844154835, + 0.009137959219515324, + 0.005687209777534008, + 0.005357707384973764, + 0.013279424048960209, + 0.015340551733970642, + 0.010232052765786648, + 0.008561474271118641, + 0.02234273962676525, + 0.017120467498898506, + 0.008859124965965748, + 0.00708756921812892, + 0.01734602078795433, + 0.015111412853002548, + 0.009587422013282776, + 0.016479946672916412, + 0.015387111343443394, + 0.011379996314644814, + 0.023353030905127525, + 0.016771847382187843, + 0.00559258321300149, + 0.0059693544171750546, + 0.004487946163862944, + 0.009551447816193104, + 0.005770840682089329, + 0.005565781611949205, + 0.011894711293280125, + 0.010462746024131775, + 0.013342675752937794, + 0.008567887358367443, + 0.0054313489235937595, + 0.00664321705698967, + 0.0078001306392252445, + 0.004443378187716007, + 0.003830735804513097, + 0.004990741610527039, + 0.007786422502249479, + 0.03996074199676514, + 0.007693609222769737, + 0.009235586039721966, + 0.011007501743733883, + 0.047462500631809235, + 0.008992926217615604, + 0.01520716492086649, + 0.004718324635177851, + 0.006330382078886032, + 0.0035039978101849556, + 0.004794500768184662, + 0.005554842296987772, + 0.00869614165276289, + 0.0050802710466086864, + 0.00958031415939331, + 0.00608762027695775, + 0.007195839658379555, + 0.008836700581014156, + 0.007923206314444542, + 0.006184936035424471, + 0.008006811141967773, + 0.01027940958738327, + 0.008381674997508526, + 0.011426039040088654, + 0.01759732887148857, + 0.012014995329082012, + 0.01639341190457344, + 0.01734255440533161, + 0.01582408882677555, + 0.014110716059803963, + 0.015015190467238426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.34862083196640015 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_2/index_tensor([1909, 1909])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_2/index_tensor([1909, 1909])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..06fa015a66a06505a3df1352d32ba35e4030360c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_2/index_tensor([1909, 1909])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium close-up, captured in exquisite high-frame-rate slow motion, establishes the hushed atmosphere of mutual success in a luxurious, wood-paneled bar. The static camera focuses on two well-dressed men gently bumping fists against the polished, dark-wood bartop. The first man, seen from the forearm down, wears the crisp white cuff of a shirt peeking from the sleeve of a well-tailored navy wool blazer. The second man, his wrist adorned with a classic gold watch, extends his hand from the sleeve of a charcoal-gray blazer. Their knuckles connect with soft, deliberate precision, the slow motion emphasizing the controlled and respectful nature of the celebratory act. In the immediate foreground, slightly out of focus, rests a crystal cognac decanter, its wide bowl filled with rich amber liquid that swirls gracefully with a syrupy consistency, capturing the warm, focused glow of a low-hanging lamp. The background blends into a soft bokeh, composed of deep, lustrous mahogany panels and backlit bottles of premium spirits in the distance, creating a refined and intimate setting. The entire scene is bathed in warm, understated lighting that casts long shadows and highlights the rich textures of wood, fabric and swirling cognac, perfectly embodying a refined, understated moment of celebration.", + "seed": 12454, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_2/index_1909_seed_12454_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.016816282738000156, + "vlm_logprob": -0.016816282738000156, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.016816282738000156, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two well-dressed men visible bumping fists against a polished dark-wood bartop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first man wearing a navy wool blazer with a crisp white shirt cuff visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second man wearing a charcoal-gray blazer with a classic gold watch on his wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a crystal cognac decanter filled with rich amber liquid visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a luxurious wood-paneled bar with a background of mahogany panels and backlit bottles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016532335430383682 + ], + "opposite_logprob_tokens": [ + -4.6415324211120605 + ], + "logprob_reward": -0.016532335430383682, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.016532335430383682, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18902789056301117, + 0.17780554294586182, + 0.09192980080842972, + 0.04211907833814621, + 0.10122550278902054, + 0.10436038672924042, + 0.11278792470693588, + 0.06598386913537979, + 0.06429717689752579, + 0.05046284943819046 + ], + "credit_map": [ + 0.01415396947413683, + 0.028488056734204292, + 0.03813678398728371, + 0.01296832226216793, + 0.05945706367492676, + 0.012792189605534077, + 0.010713142342865467, + 0.012318355962634087, + 0.011954295448958874, + 0.0239119790494442, + 0.027845751494169235, + 0.032179757952690125, + 0.032978929579257965, + 0.022931167855858803, + 0.016492720693349838, + 0.009510929696261883, + 0.012058382853865623, + 0.015317427925765514, + 0.01558513380587101, + 0.008745389059185982, + 0.014145923778414726, + 0.006250473670661449, + 0.013059026561677456, + 0.006768044084310532, + 0.004815139342099428, + 0.007167596835643053, + 0.006424860563129187, + 0.0042336913757026196, + 0.006157961208373308, + 0.004180164076387882, + 0.00553987268358469, + 0.0035997896920889616, + 0.01576722040772438, + 0.00948409829288721, + 0.028515830636024475, + 0.007826941087841988, + 0.008347623981535435, + 0.007043791469186544, + 0.01228064950555563, + 0.011959349736571312, + 0.011598650366067886, + 0.014281675219535828, + 0.015411105006933212, + 0.012900451198220253, + 0.0171683207154274, + 0.01001752633601427, + 0.013097193092107773, + 0.009885462932288647, + 0.016436142846941948, + 0.007864302955567837, + 0.026243502274155617, + 0.01070196833461523, + 0.013023275882005692, + 0.018156202509999275, + 0.00959325022995472, + 0.010769275948405266, + 0.007264818996191025, + 0.0098763732239604, + 0.010699538514018059, + 0.00794889498502016, + 0.00832583848387003, + 0.010868750512599945, + 0.00659675057977438, + 0.00440290616825223, + 0.00658388203009963, + 0.012382245622575283, + 0.009281588718295097, + 0.006170410197228193, + 0.013647101819515228, + 0.005143674556165934, + 0.006211581639945507, + 0.004876691848039627, + 0.00395742105320096, + 0.009354561567306519, + 0.00898171216249466, + 0.008317284286022186, + 0.006679494399577379, + 0.004678405821323395, + 0.004347389098256826, + 0.004146582912653685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.016532335430383682 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two well-dressed men visible bumping fists against a polished dark-wood bartop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first man wearing a navy wool blazer with a crisp white shirt cuff visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second man wearing a charcoal-gray blazer with a classic gold watch on his wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a crystal cognac decanter filled with rich amber liquid visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a luxurious wood-paneled bar with a background of mahogany panels and backlit bottles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607493683695793 + ], + "opposite_logprob_tokens": [ + -4.141075134277344 + ], + "logprob_reward": -0.01607493683695793, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607493683695793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15476016700267792, + 0.13604383170604706, + 0.2758605182170868, + 0.12369538098573685, + 0.03738096356391907, + 0.05832606554031372, + 0.05117813125252724, + 0.04967397078871727, + 0.04344014823436737, + 0.06964079290628433 + ], + "credit_map": [ + 0.01318007055670023, + 0.016179881989955902, + 0.039443109184503555, + 0.010890962556004524, + 0.04500924423336983, + 0.011838220059871674, + 0.010713125579059124, + 0.0075055453926324844, + 0.020650260150432587, + 0.023484287783503532, + 0.02179533988237381, + 0.013096966780722141, + 0.030917102470993996, + 0.008393317461013794, + 0.009990016929805279, + 0.007716536056250334, + 0.032484591007232666, + 0.043662771582603455, + 0.04611344262957573, + 0.026191841810941696, + 0.039621543139219284, + 0.012835296802222729, + 0.054824937134981155, + 0.020126109942793846, + 0.01576734520494938, + 0.01692652702331543, + 0.02225468121469021, + 0.011283917352557182, + 0.02754475176334381, + 0.008190765976905823, + 0.013326311483979225, + 0.008401085622608662, + 0.003963451366871595, + 0.003495714394375682, + 0.0063659511506557465, + 0.0029940905515104532, + 0.006326071452349424, + 0.0025781032163649797, + 0.007377841509878635, + 0.004279738757759333, + 0.008637766353785992, + 0.007806804962456226, + 0.0069832345470786095, + 0.008268256671726704, + 0.011301706545054913, + 0.006903049070388079, + 0.005183082073926926, + 0.0032421648502349854, + 0.010298176668584347, + 0.008984156884253025, + 0.008400932885706425, + 0.006888536270707846, + 0.00570373609662056, + 0.003925848752260208, + 0.0038196241948753595, + 0.003157116239890456, + 0.006183636374771595, + 0.007801390718668699, + 0.007751405239105225, + 0.005562124773859978, + 0.006901746615767479, + 0.006283424329012632, + 0.006359342951327562, + 0.0028308965265750885, + 0.0032008306588977575, + 0.007759530562907457, + 0.007826066575944424, + 0.0040512243285775185, + 0.009721634909510612, + 0.004470193292945623, + 0.0037935515865683556, + 0.0026171118952333927, + 0.007851366885006428, + 0.010707555338740349, + 0.013928364962339401, + 0.007932030595839024, + 0.01017783209681511, + 0.0055512916296720505, + 0.0066361925564706326, + 0.0068561541847884655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.01607493683695793 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two well-dressed men visible bumping fists against a polished dark-wood bartop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first man wearing a navy wool blazer with a crisp white shirt cuff visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second man wearing a charcoal-gray blazer with a classic gold watch on his wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a crystal cognac decanter filled with rich amber liquid visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a luxurious wood-paneled bar with a background of mahogany panels and backlit bottles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011071322485804558 + ], + "opposite_logprob_tokens": [ + -4.51107120513916 + ], + "logprob_reward": -0.011071322485804558, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011071322485804558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15187343955039978, + 0.10116709768772125, + 0.3165082335472107, + 0.13050788640975952, + 0.05408397316932678, + 0.1018296629190445, + 0.021930959075689316, + 0.06659476459026337, + 0.02590710110962391, + 0.029596885666251183 + ], + "credit_map": [ + 0.015037490986287594, + 0.015561949461698532, + 0.03385443985462189, + 0.014836625196039677, + 0.04393285512924194, + 0.01084061712026596, + 0.010666612535715103, + 0.007142866961658001, + 0.011012940667569637, + 0.01488060224801302, + 0.016129864379763603, + 0.013184251263737679, + 0.023723885416984558, + 0.008121730759739876, + 0.00725934561342001, + 0.0068544927053153515, + 0.03634195774793625, + 0.050264615565538406, + 0.051879607141017914, + 0.03182755783200264, + 0.04460933804512024, + 0.01853022910654545, + 0.059832531958818436, + 0.023222418501973152, + 0.016033610329031944, + 0.017212407663464546, + 0.023399367928504944, + 0.012019556015729904, + 0.03069966845214367, + 0.008974848315119743, + 0.013428840786218643, + 0.008739592507481575, + 0.006496707908809185, + 0.005352349486202002, + 0.012371045537292957, + 0.003228481626138091, + 0.010620110668241978, + 0.003862586338073015, + 0.005954123102128506, + 0.006198570132255554, + 0.009793887846171856, + 0.012489820830523968, + 0.015351386740803719, + 0.009162750095129013, + 0.016769837588071823, + 0.021628519520163536, + 0.010806028731167316, + 0.005827438086271286, + 0.002367888344451785, + 0.0027466011233627796, + 0.003057405585423112, + 0.0024608371313661337, + 0.0027898838743567467, + 0.0035816002637147903, + 0.003470662049949169, + 0.0014560840791091323, + 0.00577570078894496, + 0.009585629217326641, + 0.009972688741981983, + 0.006108791567385197, + 0.014102226123213768, + 0.011133749969303608, + 0.005929747596383095, + 0.0039862398989498615, + 0.0017415437614545226, + 0.0054914262145757675, + 0.0034178958740085363, + 0.0018182076746597886, + 0.005465109832584858, + 0.003419851651415229, + 0.0022322002332657576, + 0.0023208686616271734, + 0.0021907303016632795, + 0.004678085446357727, + 0.004724663682281971, + 0.003875692142173648, + 0.00406898558139801, + 0.002412118948996067, + 0.004907242022454739, + 0.0027393701020628214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.011071322485804558 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two well-dressed men visible bumping fists against a polished dark-wood bartop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first man wearing a navy wool blazer with a crisp white shirt cuff visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second man wearing a charcoal-gray blazer with a classic gold watch on his wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a crystal cognac decanter filled with rich amber liquid visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a luxurious wood-paneled bar with a background of mahogany panels and backlit bottles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03367551043629646 + ], + "opposite_logprob_tokens": [ + -3.4086754322052 + ], + "logprob_reward": -0.03367551043629646, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03367551043629646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14847460389137268, + 0.054820429533720016, + 0.1699870526790619, + 0.25004473328590393, + 0.19985619187355042, + 0.037866152822971344, + 0.033015407621860504, + 0.03645449876785278, + 0.03417770937085152, + 0.0353032611310482 + ], + "credit_map": [ + 0.013439612463116646, + 0.016918402165174484, + 0.028393108397722244, + 0.02190246805548668, + 0.029775628820061684, + 0.011739772744476795, + 0.013821469619870186, + 0.012484136037528515, + 0.006560484878718853, + 0.006834805011749268, + 0.007774355821311474, + 0.007607911713421345, + 0.009355337359011173, + 0.005080764181911945, + 0.007647530175745487, + 0.003959242254495621, + 0.019448474049568176, + 0.023976175114512444, + 0.030580727383494377, + 0.014573846012353897, + 0.025656770914793015, + 0.0113727618008852, + 0.029837680980563164, + 0.01454060710966587, + 0.035423748195171356, + 0.03810485824942589, + 0.043924104422330856, + 0.02306288108229637, + 0.051862847059965134, + 0.015078732743859291, + 0.027928980067372322, + 0.014658581465482712, + 0.015119917690753937, + 0.01998996175825596, + 0.03504367917776108, + 0.014817162416875362, + 0.04619866609573364, + 0.016323523595929146, + 0.030394727364182472, + 0.0219685398042202, + 0.0037771330680698156, + 0.004790202248841524, + 0.006527195684611797, + 0.005702003836631775, + 0.006603944581001997, + 0.005008339881896973, + 0.0030060827266424894, + 0.0024512496311217546, + 0.0036180540919303894, + 0.0027423056308180094, + 0.005347879137843847, + 0.005662493873387575, + 0.0034448341466486454, + 0.005665242671966553, + 0.003313471795991063, + 0.003221124177798629, + 0.002966308733448386, + 0.004284852650016546, + 0.0056266686879098415, + 0.0038956559728831053, + 0.0067768339067697525, + 0.006403052713721991, + 0.0035917425993829966, + 0.002909383736550808, + 0.004227158147841692, + 0.005319993942975998, + 0.005296860821545124, + 0.003649515099823475, + 0.004742061253637075, + 0.004136941861361265, + 0.004017227794975042, + 0.0027879495173692703, + 0.004143626894801855, + 0.005702365655452013, + 0.006801385432481766, + 0.0030060589779168367, + 0.004941295832395554, + 0.00299323839135468, + 0.003921633120626211, + 0.003793657524511218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 2 + ], + "selected_reward": -0.03367551043629646 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two well-dressed men visible bumping fists against a polished dark-wood bartop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first man wearing a navy wool blazer with a crisp white shirt cuff visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second man wearing a charcoal-gray blazer with a classic gold watch on his wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a crystal cognac decanter filled with rich amber liquid visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a luxurious wood-paneled bar with a background of mahogany panels and backlit bottles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006727308500558138 + ], + "opposite_logprob_tokens": [ + -5.00672721862793 + ], + "logprob_reward": -0.006727308500558138, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006727308500558138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4907852113246918, + 0.060273248702287674, + 0.13495312631130219, + 0.0576433427631855, + 0.03874947503209114, + 0.02592293731868267, + 0.029536474496126175, + 0.018749555572867393, + 0.05446850508451462, + 0.08891811221837997 + ], + "credit_map": [ + 0.043212130665779114, + 0.1190640926361084, + 0.08427967876195908, + 0.05203893408179283, + 0.07045168429613113, + 0.040170978754758835, + 0.04050266370177269, + 0.04106508195400238, + 0.00519594457000494, + 0.008458839729428291, + 0.010350730270147324, + 0.006353066302835941, + 0.01563824899494648, + 0.005549884866923094, + 0.004976412281394005, + 0.003750123782083392, + 0.01597857102751732, + 0.019932672381401062, + 0.021644961088895798, + 0.012956415303051472, + 0.018314743414521217, + 0.006617101840674877, + 0.028664804995059967, + 0.010843861848115921, + 0.007371869869530201, + 0.007638954557478428, + 0.01095443032681942, + 0.006115154828876257, + 0.012107857502996922, + 0.004615600220859051, + 0.005901305936276913, + 0.0029381695203483105, + 0.005529965739697218, + 0.00665678083896637, + 0.0059698112308979034, + 0.0041993651539087296, + 0.004446062259376049, + 0.0046193478628993034, + 0.004479185678064823, + 0.002848957199603319, + 0.0026504709385335445, + 0.0029423749074339867, + 0.0034233124461025, + 0.002190896775573492, + 0.0037347658071666956, + 0.005497761536389589, + 0.002983399434015155, + 0.002499955240637064, + 0.0025620218366384506, + 0.0029492753092199564, + 0.004462733399122953, + 0.0053999051451683044, + 0.003812693525105715, + 0.004517379682511091, + 0.003945521079003811, + 0.0018869424238801003, + 0.0017080075340345502, + 0.001957653323188424, + 0.0030041194986552, + 0.0024137047585099936, + 0.003256472758948803, + 0.0017543199937790632, + 0.0031656373757869005, + 0.0014896391658112407, + 0.0054400283843278885, + 0.0067787449806928635, + 0.007720183115452528, + 0.004532501567155123, + 0.004945877008140087, + 0.009174120612442493, + 0.0071749016642570496, + 0.008702149614691734, + 0.006366674322634935, + 0.016848783940076828, + 0.016107851639389992, + 0.01049226988106966, + 0.01319923996925354, + 0.008768662810325623, + 0.00875093787908554, + 0.00838368758559227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.006727308500558138 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up, captured in exquisite high-frame-rate slow motion, establishes the hushed atmosphere of mutual success in a luxurious, wood-paneled bar. The static camera focuses on two well-dressed men gently bumping fists against the polished, dark-wood bartop. The first man, seen from the forearm down, wears the crisp white cuff of a shirt peeking from the sleeve of a well-tailored navy wool blazer. The second man, his wrist adorned with a classic gold watch, extends his hand from the sleeve of a charcoal-gray blazer. Their knuckles connect with soft, deliberate precision, the slow motion emphasizing the controlled and respectful nature of the celebratory act. In the immediate foreground, slightly out of focus, rests a crystal cognac decanter, its wide bowl filled with rich amber liquid that swirls gracefully with a syrupy consistency, capturing the warm, focused glow of a low-hanging lamp. The background blends into a soft bokeh, composed of deep, lustrous mahogany panels and backlit bottles of premium spirits in the distance, creating a refined and intimate setting. The entire scene is bathed in warm, understated lighting that casts long shadows and highlights the rich textures of wood, fabric and swirling cognac, perfectly embodying a refined, understated moment of celebration.", + "seed": 12455, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_2/index_1909_seed_12455_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.015862732753157617, + "vlm_logprob": -0.015862732753157617, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.015862732753157617, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two well-dressed men visible bumping fists against a polished dark-wood bartop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first man wearing a navy wool blazer with a crisp white shirt cuff visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second man wearing a charcoal-gray blazer with a classic gold watch on his wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a crystal cognac decanter filled with rich amber liquid visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a luxurious wood-paneled bar with a background of mahogany panels and backlit bottles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018298154696822166 + ], + "opposite_logprob_tokens": [ + -4.268298149108887 + ], + "logprob_reward": -0.018298154696822166, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.018298154696822166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11890248209238052, + 0.238970085978508, + 0.05990641564130783, + 0.13071997463703156, + 0.1401972472667694, + 0.028067294508218765, + 0.06263898313045502, + 0.11202874034643173, + 0.08663364499807358, + 0.021935105323791504 + ], + "credit_map": [ + 0.01889052614569664, + 0.014834136702120304, + 0.023318517953157425, + 0.014645645394921303, + 0.012609491124749184, + 0.013536278158426285, + 0.011067165993154049, + 0.010000712238252163, + 0.015584695152938366, + 0.0148105937987566, + 0.048229895532131195, + 0.08253956586122513, + 0.014399318024516106, + 0.022265657782554626, + 0.02316385507583618, + 0.017976507544517517, + 0.005817420780658722, + 0.007601393386721611, + 0.008834206499159336, + 0.011435221880674362, + 0.006682445760816336, + 0.005623656325042248, + 0.008146623149514198, + 0.00576544227078557, + 0.01231897808611393, + 0.006803025491535664, + 0.03482523187994957, + 0.04372749105095863, + 0.006273540668189526, + 0.007753931451588869, + 0.009057080373167992, + 0.009960692375898361, + 0.01815793290734291, + 0.017347443848848343, + 0.01730409450829029, + 0.015509502030909061, + 0.016761524602770805, + 0.024656090885400772, + 0.014594543725252151, + 0.01586611196398735, + 0.00308259972371161, + 0.0037097360473126173, + 0.00305759790353477, + 0.003472308162599802, + 0.0036569153890013695, + 0.004615606740117073, + 0.003863278077915311, + 0.002609256189316511, + 0.01072581484913826, + 0.005884900689125061, + 0.005167083349078894, + 0.017189281061291695, + 0.005075287539511919, + 0.005726751871407032, + 0.006358157377690077, + 0.006511707324534655, + 0.014639466069638729, + 0.013822440057992935, + 0.01660425215959549, + 0.019146889448165894, + 0.008254761807620525, + 0.018424496054649353, + 0.011384548619389534, + 0.009751882404088974, + 0.012193604372441769, + 0.014922438189387321, + 0.006147539708763361, + 0.011570918373763561, + 0.01055789366364479, + 0.010970303788781166, + 0.008363950997591019, + 0.011906994506716728, + 0.004041409585624933, + 0.001875557005405426, + 0.002394532086327672, + 0.004150412045419216, + 0.0015210021520033479, + 0.001992234028875828, + 0.00309357768855989, + 0.0028663792181760073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 3 + ], + "selected_reward": -0.018298154696822166 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two well-dressed men visible bumping fists against a polished dark-wood bartop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first man wearing a navy wool blazer with a crisp white shirt cuff visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second man wearing a charcoal-gray blazer with a classic gold watch on his wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a crystal cognac decanter filled with rich amber liquid visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a luxurious wood-paneled bar with a background of mahogany panels and backlit bottles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607270911335945 + ], + "opposite_logprob_tokens": [ + -4.141072750091553 + ], + "logprob_reward": -0.01607270911335945, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607270911335945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16727577149868011, + 0.1556587666273117, + 0.04719929397106171, + 0.11601340770721436, + 0.12216683477163315, + 0.020028583705425262, + 0.04366403818130493, + 0.07495581358671188, + 0.14211171865463257, + 0.11092576384544373 + ], + "credit_map": [ + 0.025377614423632622, + 0.02188439480960369, + 0.029920024797320366, + 0.023272447288036346, + 0.02032381296157837, + 0.015761373564600945, + 0.014778895303606987, + 0.015957217663526535, + 0.011297227814793587, + 0.010204958729445934, + 0.02277952991425991, + 0.04168248176574707, + 0.012630194425582886, + 0.018357407301664352, + 0.022482827305793762, + 0.016224155202507973, + 0.003954097628593445, + 0.004420319572091103, + 0.007287340238690376, + 0.006404784973710775, + 0.004472444765269756, + 0.010444825515151024, + 0.005128094460815191, + 0.005087394267320633, + 0.01589731127023697, + 0.005837687291204929, + 0.02765788324177265, + 0.030369149520993233, + 0.007285892032086849, + 0.008359496481716633, + 0.010893750935792923, + 0.009712238796055317, + 0.018671076744794846, + 0.01425281073898077, + 0.013014407828450203, + 0.012090111151337624, + 0.010681938380002975, + 0.01751222088932991, + 0.020346704870462418, + 0.01559758372604847, + 0.0028068532701581717, + 0.002887072740122676, + 0.0023789943661540747, + 0.0020859581418335438, + 0.002424228470772505, + 0.002700407989323139, + 0.002223896561190486, + 0.0025211728643625975, + 0.006866679526865482, + 0.004053868353366852, + 0.010153328999876976, + 0.007688512094318867, + 0.004677602555602789, + 0.003528917208313942, + 0.0030606265645474195, + 0.0036345103289932013, + 0.009523462504148483, + 0.00816289708018303, + 0.010631079785525799, + 0.01495431549847126, + 0.005275920033454895, + 0.008875907398760319, + 0.009954738430678844, + 0.00757749704644084, + 0.02752126008272171, + 0.028706058859825134, + 0.01514402125030756, + 0.026970023289322853, + 0.016471071168780327, + 0.009248528629541397, + 0.008794892579317093, + 0.009255887009203434, + 0.02456778846681118, + 0.004994949325919151, + 0.008706717751920223, + 0.018341196700930595, + 0.004322736989706755, + 0.006766350008547306, + 0.005277243908494711, + 0.037948787212371826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.01607270911335945 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two well-dressed men visible bumping fists against a polished dark-wood bartop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first man wearing a navy wool blazer with a crisp white shirt cuff visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second man wearing a charcoal-gray blazer with a classic gold watch on his wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a crystal cognac decanter filled with rich amber liquid visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a luxurious wood-paneled bar with a background of mahogany panels and backlit bottles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107415184378624 + ], + "opposite_logprob_tokens": [ + -4.511074066162109 + ], + "logprob_reward": -0.01107415184378624, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107415184378624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19294939935207367, + 0.18374036252498627, + 0.07678186148405075, + 0.08149562776088715, + 0.06134781241416931, + 0.03422316163778305, + 0.10739538818597794, + 0.017133159562945366, + 0.14659333229064941, + 0.09833987057209015 + ], + "credit_map": [ + 0.024450717493891716, + 0.025721147656440735, + 0.033156730234622955, + 0.037764422595500946, + 0.014668512158095837, + 0.017896072939038277, + 0.02058444358408451, + 0.01870734989643097, + 0.01061161793768406, + 0.010704577900469303, + 0.03823060169816017, + 0.061430007219314575, + 0.010214351117610931, + 0.016301369294524193, + 0.02029101364314556, + 0.0159568153321743, + 0.008158856071531773, + 0.009336287155747414, + 0.010498609393835068, + 0.010872050188481808, + 0.01023575197905302, + 0.012165054678916931, + 0.009560572914779186, + 0.0059546842239797115, + 0.006782227661460638, + 0.006495254579931498, + 0.01822187751531601, + 0.024572519585490227, + 0.006488675717264414, + 0.004845303948968649, + 0.008293194696307182, + 0.005796571262180805, + 0.007402900140732527, + 0.005121628288179636, + 0.006472697947174311, + 0.005154907703399658, + 0.007219761610031128, + 0.01091743353754282, + 0.009327661246061325, + 0.009730824269354343, + 0.0039045019075274467, + 0.005418084561824799, + 0.0038931239396333694, + 0.004085847642272711, + 0.0060414099134504795, + 0.0041984268464148045, + 0.003384845796972513, + 0.003296919632703066, + 0.02947804145514965, + 0.005600215867161751, + 0.006221435032784939, + 0.044290926307439804, + 0.004376040771603584, + 0.00532293226569891, + 0.0056165121495723724, + 0.006489292252808809, + 0.002375482115894556, + 0.0020515103824436665, + 0.0023469848092645407, + 0.0019983903039246798, + 0.0018785690190270543, + 0.002191565465182066, + 0.0025500624906271696, + 0.001740594976581633, + 0.026619909331202507, + 0.029673807322978973, + 0.014455791562795639, + 0.02880905196070671, + 0.018109947443008423, + 0.010332290083169937, + 0.009467021562159061, + 0.009125512093305588, + 0.015258610248565674, + 0.006874305661767721, + 0.005157939624041319, + 0.02167341485619545, + 0.005246276967227459, + 0.006694499868899584, + 0.0055040703155100346, + 0.03193075954914093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.01107415184378624 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two well-dressed men visible bumping fists against a polished dark-wood bartop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first man wearing a navy wool blazer with a crisp white shirt cuff visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second man wearing a charcoal-gray blazer with a classic gold watch on his wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a crystal cognac decanter filled with rich amber liquid visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a luxurious wood-paneled bar with a background of mahogany panels and backlit bottles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029780270531773567 + ], + "opposite_logprob_tokens": [ + -3.529780387878418 + ], + "logprob_reward": -0.029780270531773567, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029780270531773567, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0844459980726242, + 0.346306174993515, + 0.15074218809604645, + 0.03287459909915924, + 0.01747804880142212, + 0.05968235433101654, + 0.041053105145692825, + 0.03834456577897072, + 0.04054506495594978, + 0.1885278820991516 + ], + "credit_map": [ + 0.010275867767632008, + 0.011097751557826996, + 0.014744384214282036, + 0.014309069141745567, + 0.006683072075247765, + 0.008997021242976189, + 0.009591641835868359, + 0.008747181855142117, + 0.019196832552552223, + 0.017518121749162674, + 0.096257284283638, + 0.13740307092666626, + 0.01070484984666109, + 0.02251640148460865, + 0.024063045158982277, + 0.018646562471985817, + 0.014949113130569458, + 0.017067743465304375, + 0.018771328032016754, + 0.019845493137836456, + 0.015157234854996204, + 0.031090544536709785, + 0.015191920101642609, + 0.01866881735622883, + 0.002534135477617383, + 0.0015986544312909245, + 0.007738817948848009, + 0.011298875324428082, + 0.001654970576055348, + 0.002924760105088353, + 0.002196133602410555, + 0.0029282504692673683, + 0.0017195155378431082, + 0.002104362240061164, + 0.0015326488064602017, + 0.00183486461173743, + 0.0027126416098326445, + 0.003158831736072898, + 0.0023978303652256727, + 0.002017353428527713, + 0.004821926821023226, + 0.004758606664836407, + 0.013140941970050335, + 0.014997348189353943, + 0.006962142419070005, + 0.005641599651426077, + 0.0055848038755357265, + 0.003774980315938592, + 0.009983829222619534, + 0.0033401090186089277, + 0.0023130953777581453, + 0.014781362377107143, + 0.002602687571197748, + 0.002249759156256914, + 0.002957217628136277, + 0.002825042698532343, + 0.006450658198446035, + 0.0066537256352603436, + 0.005287804640829563, + 0.0034478248562663794, + 0.0029977704398334026, + 0.003445955691859126, + 0.005650799255818129, + 0.00441002706065774, + 0.006576801184564829, + 0.006325520575046539, + 0.003835441544651985, + 0.004703951068222523, + 0.0062199742533266544, + 0.004724442958831787, + 0.004309710580855608, + 0.003849219996482134, + 0.0319482684135437, + 0.00913273636251688, + 0.016216564923524857, + 0.039530836045742035, + 0.008356336504220963, + 0.01375668402761221, + 0.008593329228460789, + 0.060993120074272156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 2 + ], + "selected_reward": -0.029780270531773567 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two well-dressed men visible bumping fists against a polished dark-wood bartop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first man wearing a navy wool blazer with a crisp white shirt cuff visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second man wearing a charcoal-gray blazer with a classic gold watch on his wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a crystal cognac decanter filled with rich amber liquid visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a luxurious wood-paneled bar with a background of mahogany panels and backlit bottles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004088377580046654 + ], + "opposite_logprob_tokens": [ + -5.504088401794434 + ], + "logprob_reward": -0.004088377580046654, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004088377580046654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15649086236953735, + 0.23625417053699493, + 0.024109207093715668, + 0.05504748597741127, + 0.03229928016662598, + 0.023761657997965813, + 0.05991969257593155, + 0.06804443150758743, + 0.06298449635505676, + 0.2810886800289154 + ], + "credit_map": [ + 0.020300444215536118, + 0.021281998604536057, + 0.019941020756959915, + 0.02041611447930336, + 0.01764342561364174, + 0.018530521541833878, + 0.022434217855334282, + 0.0159431342035532, + 0.013520649634301662, + 0.013113531284034252, + 0.05854468420147896, + 0.09066896885633469, + 0.010496034286916256, + 0.018471674993634224, + 0.017029235139489174, + 0.014409412629902363, + 0.00242417911067605, + 0.0021804189309477806, + 0.003143679816275835, + 0.0030918272677809, + 0.0029272192623466253, + 0.004939781967550516, + 0.0026992792263627052, + 0.00270282244309783, + 0.004695705603808165, + 0.0029410207644104958, + 0.011422150768339634, + 0.016342520713806152, + 0.004355825949460268, + 0.005017388146370649, + 0.005465907044708729, + 0.004806965589523315, + 0.004305247217416763, + 0.004899201449006796, + 0.0032488175202161074, + 0.0032740775495767593, + 0.003251731861382723, + 0.004930732771754265, + 0.004564293660223484, + 0.0038251769728958607, + 0.0034380732104182243, + 0.0028620536904782057, + 0.003449876792728901, + 0.0026870067231357098, + 0.0027554419357329607, + 0.0036285051610320807, + 0.0028390311636030674, + 0.002101669553667307, + 0.010376128368079662, + 0.006314938887953758, + 0.00522850826382637, + 0.012728335335850716, + 0.008531718514859676, + 0.006803946103900671, + 0.00527611467987299, + 0.00466000335291028, + 0.009435570798814297, + 0.00802501942962408, + 0.007178190164268017, + 0.007253446616232395, + 0.007047399412840605, + 0.011269956827163696, + 0.010580267757177353, + 0.007254575379192829, + 0.0077852220274508, + 0.009478793479502201, + 0.00559733621776104, + 0.009642831981182098, + 0.008923603221774101, + 0.0076601034961640835, + 0.007493221666663885, + 0.00640338659286499, + 0.028670769184827805, + 0.024221107363700867, + 0.01568424701690674, + 0.07064701616764069, + 0.020803283900022507, + 0.022685794159770012, + 0.018669767305254936, + 0.07970670610666275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.004088377580046654 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up, captured in exquisite high-frame-rate slow motion, establishes the hushed atmosphere of mutual success in a luxurious, wood-paneled bar. The static camera focuses on two well-dressed men gently bumping fists against the polished, dark-wood bartop. The first man, seen from the forearm down, wears the crisp white cuff of a shirt peeking from the sleeve of a well-tailored navy wool blazer. The second man, his wrist adorned with a classic gold watch, extends his hand from the sleeve of a charcoal-gray blazer. Their knuckles connect with soft, deliberate precision, the slow motion emphasizing the controlled and respectful nature of the celebratory act. In the immediate foreground, slightly out of focus, rests a crystal cognac decanter, its wide bowl filled with rich amber liquid that swirls gracefully with a syrupy consistency, capturing the warm, focused glow of a low-hanging lamp. The background blends into a soft bokeh, composed of deep, lustrous mahogany panels and backlit bottles of premium spirits in the distance, creating a refined and intimate setting. The entire scene is bathed in warm, understated lighting that casts long shadows and highlights the rich textures of wood, fabric and swirling cognac, perfectly embodying a refined, understated moment of celebration.", + "seed": 12456, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_2/index_1909_seed_12456_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01988385245203972, + "vlm_logprob": -0.01988385245203972, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.01988385245203972, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two well-dressed men visible bumping fists against a polished dark-wood bartop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first man wearing a navy wool blazer with a crisp white shirt cuff visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second man wearing a charcoal-gray blazer with a classic gold watch on his wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a crystal cognac decanter filled with rich amber liquid visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a luxurious wood-paneled bar with a background of mahogany panels and backlit bottles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03312014043331146 + ], + "opposite_logprob_tokens": [ + -3.6581201553344727 + ], + "logprob_reward": -0.03312014043331146, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03312014043331146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3768354654312134, + 0.0625472217798233, + 0.051729749888181686, + 0.0740962028503418, + 0.08805472403764725, + 0.030542055144906044, + 0.07219405472278595, + 0.03524266183376312, + 0.12777884304523468, + 0.08097905665636063 + ], + "credit_map": [ + 0.04969853162765503, + 0.041152194142341614, + 0.04948178678750992, + 0.03565302491188049, + 0.029896562919020653, + 0.03879835456609726, + 0.05449725687503815, + 0.07765773683786392, + 0.007477506063878536, + 0.006467630621045828, + 0.007434616331011057, + 0.008951842784881592, + 0.008238846436142921, + 0.005619629751890898, + 0.011988031677901745, + 0.006369113922119141, + 0.005558446981012821, + 0.003765132511034608, + 0.006906292866915464, + 0.010007771663367748, + 0.00503011979162693, + 0.00570023013278842, + 0.010091320611536503, + 0.004670439753681421, + 0.012208373285830021, + 0.008056385442614555, + 0.010236099362373352, + 0.010030921548604965, + 0.004969548434019089, + 0.007853096351027489, + 0.01265102718025446, + 0.008090748451650143, + 0.010717570781707764, + 0.006864815019071102, + 0.010262379422783852, + 0.008577319793403149, + 0.009072650223970413, + 0.010122020728886127, + 0.017480194568634033, + 0.014957777224481106, + 0.004025711677968502, + 0.00283264834433794, + 0.0037000211887061596, + 0.005023610778152943, + 0.0033133123070001602, + 0.0023340536281466484, + 0.006012367084622383, + 0.003300328738987446, + 0.008345693349838257, + 0.008007369004189968, + 0.008723560720682144, + 0.010776272974908352, + 0.007723094429820776, + 0.006261853501200676, + 0.011700076051056385, + 0.010656137019395828, + 0.006091360002756119, + 0.003297709859907627, + 0.005262197460979223, + 0.00548661220818758, + 0.002884147921577096, + 0.0026112631894648075, + 0.0062872921116650105, + 0.0033220823388546705, + 0.027792252600193024, + 0.014271873980760574, + 0.020225999876856804, + 0.011777200736105442, + 0.01385006494820118, + 0.01067501213401556, + 0.014362707734107971, + 0.014823731034994125, + 0.01576963998377323, + 0.010464497841894627, + 0.011038361117243767, + 0.00704707857221365, + 0.010934034362435341, + 0.004469458479434252, + 0.015096457675099373, + 0.006159525830298662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.03312014043331146 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two well-dressed men visible bumping fists against a polished dark-wood bartop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first man wearing a navy wool blazer with a crisp white shirt cuff visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second man wearing a charcoal-gray blazer with a classic gold watch on his wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a crystal cognac decanter filled with rich amber liquid visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a luxurious wood-paneled bar with a background of mahogany panels and backlit bottles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014211048372089863 + ], + "opposite_logprob_tokens": [ + -4.264211177825928 + ], + "logprob_reward": -0.014211048372089863, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014211048372089863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1924780011177063, + 0.07519770413637161, + 0.1174730435013771, + 0.16461870074272156, + 0.09562975913286209, + 0.04795194789767265, + 0.06138173118233681, + 0.07181335985660553, + 0.12843206524848938, + 0.04502374678850174 + ], + "credit_map": [ + 0.026261568069458008, + 0.0214986689388752, + 0.02126840315759182, + 0.021139437332749367, + 0.02509389817714691, + 0.018576206639409065, + 0.035389505326747894, + 0.023250317201018333, + 0.008992636576294899, + 0.009642718359827995, + 0.01038562972098589, + 0.008901769295334816, + 0.007466097828000784, + 0.009975702501833439, + 0.011754759587347507, + 0.008078396320343018, + 0.017571911215782166, + 0.011032622307538986, + 0.016350330784916878, + 0.014383112080395222, + 0.011212198063731194, + 0.015305332839488983, + 0.01863100938498974, + 0.012986531481146812, + 0.01801072247326374, + 0.0199038814753294, + 0.020122405141592026, + 0.025494728237390518, + 0.01921945810317993, + 0.01340953167527914, + 0.028641540557146072, + 0.019816428422927856, + 0.019692184403538704, + 0.012749426998198032, + 0.01248154602944851, + 0.012347142212092876, + 0.010723262093961239, + 0.007074224296957254, + 0.012500842101871967, + 0.008061129599809647, + 0.007015971001237631, + 0.003513502422720194, + 0.006490708794444799, + 0.005036765709519386, + 0.0032147581223398447, + 0.00853270385414362, + 0.006531524006277323, + 0.00761601235717535, + 0.008869917131960392, + 0.005366000346839428, + 0.009055799804627895, + 0.013053981587290764, + 0.004321597050875425, + 0.006443734280765057, + 0.009133213199675083, + 0.005137488711625338, + 0.012345847673714161, + 0.006549927871674299, + 0.013023932464420795, + 0.007965669967234135, + 0.005893537774682045, + 0.008304606191813946, + 0.008312088437378407, + 0.009417753666639328, + 0.03302141651511192, + 0.011607291176915169, + 0.022382235154509544, + 0.01805523782968521, + 0.008430981077253819, + 0.006914026103913784, + 0.016607509925961494, + 0.01141336653381586, + 0.007500385865569115, + 0.006691596936434507, + 0.006165584549307823, + 0.004917419049888849, + 0.005071857012808323, + 0.00351240043528378, + 0.006811537779867649, + 0.004352964460849762 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.014211048372089863 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two well-dressed men visible bumping fists against a polished dark-wood bartop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first man wearing a navy wool blazer with a crisp white shirt cuff visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second man wearing a charcoal-gray blazer with a classic gold watch on his wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a crystal cognac decanter filled with rich amber liquid visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a luxurious wood-paneled bar with a background of mahogany panels and backlit bottles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977916456758976 + ], + "opposite_logprob_tokens": [ + -4.63477897644043 + ], + "logprob_reward": -0.00977916456758976, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977916456758976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15114612877368927, + 0.19650918245315552, + 0.08224166929721832, + 0.0761834904551506, + 0.0693025216460228, + 0.05662675201892853, + 0.056899916380643845, + 0.0904771089553833, + 0.13817375898361206, + 0.08243948221206665 + ], + "credit_map": [ + 0.02271103486418724, + 0.014441307634115219, + 0.01764431595802307, + 0.014509118162095547, + 0.010669928044080734, + 0.009283328428864479, + 0.027825556695461273, + 0.03406154364347458, + 0.02504383772611618, + 0.030454179272055626, + 0.028482653200626373, + 0.022522330284118652, + 0.022206202149391174, + 0.02223343960940838, + 0.02596576325595379, + 0.019600773230195045, + 0.01094829011708498, + 0.007601011544466019, + 0.009609036147594452, + 0.009248078800737858, + 0.008628999814391136, + 0.013653885573148727, + 0.011667528189718723, + 0.010884840041399002, + 0.009904361329972744, + 0.006583597511053085, + 0.009453958831727505, + 0.013525499030947685, + 0.0062860157340765, + 0.008549301885068417, + 0.014910692349076271, + 0.0069700609892606735, + 0.01122905034571886, + 0.007351618725806475, + 0.008600495755672455, + 0.009038602001965046, + 0.006749757099896669, + 0.005497409962117672, + 0.012780885212123394, + 0.008054697886109352, + 0.008751627057790756, + 0.009718706831336021, + 0.007435151841491461, + 0.00639004074037075, + 0.0069928159937262535, + 0.0040714796632528305, + 0.008224318735301495, + 0.005042610689997673, + 0.009992032311856747, + 0.0054151397198438644, + 0.009177964180707932, + 0.010956589132547379, + 0.004756838083267212, + 0.005157687701284885, + 0.006802732590585947, + 0.004640928003937006, + 0.014152116142213345, + 0.008196807466447353, + 0.01375028770416975, + 0.01337985135614872, + 0.006782627198845148, + 0.008809168823063374, + 0.016830425709486008, + 0.00857582874596119, + 0.04275626316666603, + 0.013605515472590923, + 0.027548471465706825, + 0.01184176281094551, + 0.008488987572491169, + 0.007287904154509306, + 0.014652020297944546, + 0.011992838233709335, + 0.013897035270929337, + 0.013031644746661186, + 0.011963041499257088, + 0.010445789434015751, + 0.009612907655537128, + 0.005831321235746145, + 0.010999907739460468, + 0.0066578262485563755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.00977916456758976 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two well-dressed men visible bumping fists against a polished dark-wood bartop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first man wearing a navy wool blazer with a crisp white shirt cuff visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second man wearing a charcoal-gray blazer with a classic gold watch on his wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a crystal cognac decanter filled with rich amber liquid visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a luxurious wood-paneled bar with a background of mahogany panels and backlit bottles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029780618846416473 + ], + "opposite_logprob_tokens": [ + -3.529780626296997 + ], + "logprob_reward": -0.029780618846416473, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029780618846416473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09078233689069748, + 0.4476722478866577, + 0.07462408393621445, + 0.07667046040296555, + 0.08893181383609772, + 0.04581018164753914, + 0.04516472667455673, + 0.052681390196084976, + 0.04702814668416977, + 0.03063458949327469 + ], + "credit_map": [ + 0.010570299811661243, + 0.01553263794630766, + 0.010201995261013508, + 0.010042911395430565, + 0.010835318826138973, + 0.010728221386671066, + 0.014191465452313423, + 0.008679487742483616, + 0.05054980143904686, + 0.062329456210136414, + 0.061855826526880264, + 0.04324183240532875, + 0.06757152080535889, + 0.044504303485155106, + 0.06739642471075058, + 0.050223104655742645, + 0.006994871888309717, + 0.005909716244786978, + 0.009445659816265106, + 0.013543280772864819, + 0.00593691598623991, + 0.009609558619558811, + 0.01581253483891487, + 0.007371547631919384, + 0.010177875868976116, + 0.006573984399437904, + 0.010726841166615486, + 0.017998235300183296, + 0.004845979157835245, + 0.0063840183429419994, + 0.01212576124817133, + 0.007837767712771893, + 0.009796791709959507, + 0.009079270996153355, + 0.01133731473237276, + 0.015478456392884254, + 0.006879483349621296, + 0.005913198925554752, + 0.01582745648920536, + 0.014619839377701283, + 0.006324967835098505, + 0.008165066130459309, + 0.006302503403276205, + 0.005015262868255377, + 0.004840684123337269, + 0.004624254535883665, + 0.006530966144055128, + 0.004006475675851107, + 0.006391754373908043, + 0.004680695477873087, + 0.006783292628824711, + 0.006213594693690538, + 0.0037293327040970325, + 0.007017351686954498, + 0.006309823133051395, + 0.00403887964785099, + 0.009868059307336807, + 0.005066802259534597, + 0.008243255317211151, + 0.006143177859485149, + 0.005434687715023756, + 0.004384967498481274, + 0.00881468877196312, + 0.004725751467049122, + 0.009359871037304401, + 0.004114117007702589, + 0.006894745863974094, + 0.009695412591099739, + 0.0034956014715135098, + 0.002542775357142091, + 0.0075444793328642845, + 0.003381140064448118, + 0.006881079636514187, + 0.002230219542980194, + 0.005041339434683323, + 0.002629803027957678, + 0.0024328595027327538, + 0.002315608784556389, + 0.006621318403631449, + 0.0024823618587106466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.029780618846416473 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two well-dressed men visible bumping fists against a polished dark-wood bartop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first man wearing a navy wool blazer with a crisp white shirt cuff visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second man wearing a charcoal-gray blazer with a classic gold watch on his wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a crystal cognac decanter filled with rich amber liquid visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a luxurious wood-paneled bar with a background of mahogany panels and backlit bottles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012528290040791035 + ], + "opposite_logprob_tokens": [ + -4.387528419494629 + ], + "logprob_reward": -0.012528290040791035, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012528290040791035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10024659335613251, + 0.43067413568496704, + 0.08864440023899078, + 0.06320139765739441, + 0.053178977221250534, + 0.04351954907178879, + 0.047181449830532074, + 0.03201989829540253, + 0.06262725591659546, + 0.07870631664991379 + ], + "credit_map": [ + 0.012224236503243446, + 0.011240234598517418, + 0.01328815333545208, + 0.0085503114387393, + 0.009793850593268871, + 0.011059354059398174, + 0.014658922329545021, + 0.01943153887987137, + 0.05326656997203827, + 0.04824153706431389, + 0.05265543982386589, + 0.04545126110315323, + 0.049767110496759415, + 0.04184071719646454, + 0.0829768255352974, + 0.05647468566894531, + 0.009681090712547302, + 0.006974378600716591, + 0.011423805728554726, + 0.013878763653337955, + 0.0075091891922056675, + 0.007826142013072968, + 0.01927255280315876, + 0.012078477069735527, + 0.007355862762778997, + 0.005068509839475155, + 0.008518588729202747, + 0.01498559769243002, + 0.003904427168890834, + 0.005034412257373333, + 0.00986810028553009, + 0.0084659019485116, + 0.006214930675923824, + 0.00460104551166296, + 0.00569820636883378, + 0.004504653159528971, + 0.00640222430229187, + 0.0065045892260968685, + 0.010760588571429253, + 0.008492744527757168, + 0.006206434220075607, + 0.004589569289237261, + 0.0055808499455451965, + 0.006877658888697624, + 0.0029657369013875723, + 0.005927750375121832, + 0.006689038593322039, + 0.004682510159909725, + 0.005383225157856941, + 0.005288843531161547, + 0.006255284883081913, + 0.0054493797942996025, + 0.005389704369008541, + 0.0043769399635493755, + 0.008189229294657707, + 0.006848843302577734, + 0.00529694277793169, + 0.0039418586529791355, + 0.004414352122694254, + 0.004262447357177734, + 0.002894673263654113, + 0.00213381415233016, + 0.0060353693552315235, + 0.0030404406134039164, + 0.010441708378493786, + 0.008413144387304783, + 0.008977794088423252, + 0.005158542655408382, + 0.006803310941904783, + 0.008364127017557621, + 0.005892091430723667, + 0.008576540276408195, + 0.014370962046086788, + 0.012150873430073261, + 0.011062905192375183, + 0.005334113258868456, + 0.011249094270169735, + 0.005235807038843632, + 0.012588604353368282, + 0.006713956128805876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.012528290040791035 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up, captured in exquisite high-frame-rate slow motion, establishes the hushed atmosphere of mutual success in a luxurious, wood-paneled bar. The static camera focuses on two well-dressed men gently bumping fists against the polished, dark-wood bartop. The first man, seen from the forearm down, wears the crisp white cuff of a shirt peeking from the sleeve of a well-tailored navy wool blazer. The second man, his wrist adorned with a classic gold watch, extends his hand from the sleeve of a charcoal-gray blazer. Their knuckles connect with soft, deliberate precision, the slow motion emphasizing the controlled and respectful nature of the celebratory act. In the immediate foreground, slightly out of focus, rests a crystal cognac decanter, its wide bowl filled with rich amber liquid that swirls gracefully with a syrupy consistency, capturing the warm, focused glow of a low-hanging lamp. The background blends into a soft bokeh, composed of deep, lustrous mahogany panels and backlit bottles of premium spirits in the distance, creating a refined and intimate setting. The entire scene is bathed in warm, understated lighting that casts long shadows and highlights the rich textures of wood, fabric and swirling cognac, perfectly embodying a refined, understated moment of celebration.", + "seed": 12457, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_2/index_1909_seed_12457_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02550455890595913, + "vlm_logprob": -0.02550455890595913, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.02550455890595913, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two well-dressed men visible bumping fists against a polished dark-wood bartop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first man wearing a navy wool blazer with a crisp white shirt cuff visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second man wearing a charcoal-gray blazer with a classic gold watch on his wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a crystal cognac decanter filled with rich amber liquid visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a luxurious wood-paneled bar with a background of mahogany panels and backlit bottles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04718108847737312 + ], + "opposite_logprob_tokens": [ + -3.1721811294555664 + ], + "logprob_reward": -0.04718108847737312, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04718108847737312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06256497651338577, + 0.13053762912750244, + 0.14517225325107574, + 0.2626992166042328, + 0.059977974742650986, + 0.023843219503760338, + 0.048461977392435074, + 0.1130632758140564, + 0.12419842183589935, + 0.02948102355003357 + ], + "credit_map": [ + 0.006823383271694183, + 0.006536014843732119, + 0.009138264693319798, + 0.009404514916241169, + 0.007992973551154137, + 0.006892747711390257, + 0.009306712076067924, + 0.006470366846770048, + 0.018315739929676056, + 0.008164689876139164, + 0.019907528534531593, + 0.01802728697657585, + 0.016739320009946823, + 0.013734388165175915, + 0.024019887670874596, + 0.01162879727780819, + 0.02420872077345848, + 0.009195705875754356, + 0.007921594195067883, + 0.04008937254548073, + 0.021090427413582802, + 0.013463381677865982, + 0.022003576159477234, + 0.0071994648315012455, + 0.04299568757414818, + 0.010671385563910007, + 0.007883635349571705, + 0.04308628663420677, + 0.03249478340148926, + 0.09597915410995483, + 0.01819176971912384, + 0.011396517977118492, + 0.010357062332332134, + 0.004291652236133814, + 0.005033605266362429, + 0.014110444113612175, + 0.00634973356500268, + 0.005701560992747545, + 0.010346774011850357, + 0.0037871396634727716, + 0.0033733209129422903, + 0.001765602733939886, + 0.0031750944908708334, + 0.004453741014003754, + 0.002575550228357315, + 0.0029492517933249474, + 0.0036093099042773247, + 0.001941348658874631, + 0.00838999543339014, + 0.0041156732477247715, + 0.003926106728613377, + 0.007742059882730246, + 0.004027693532407284, + 0.00603494793176651, + 0.009107249788939953, + 0.005118249915540218, + 0.016697382554411888, + 0.02158331871032715, + 0.012367787770926952, + 0.021437644958496094, + 0.011281730607151985, + 0.010038566775619984, + 0.012035987339913845, + 0.007620856165885925, + 0.018162189051508904, + 0.010782352648675442, + 0.0070816269144415855, + 0.016343114897608757, + 0.01401582546532154, + 0.027439741417765617, + 0.017473217099905014, + 0.012900357134640217, + 0.003990236669778824, + 0.0019727337639778852, + 0.003149393480271101, + 0.004755720961838961, + 0.0037820693105459213, + 0.004406650085002184, + 0.004196213558316231, + 0.003228006884455681 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 1 + ], + "selected_reward": -0.04718108847737312 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two well-dressed men visible bumping fists against a polished dark-wood bartop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first man wearing a navy wool blazer with a crisp white shirt cuff visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second man wearing a charcoal-gray blazer with a classic gold watch on his wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a crystal cognac decanter filled with rich amber liquid visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a luxurious wood-paneled bar with a background of mahogany panels and backlit bottles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018197840079665184 + ], + "opposite_logprob_tokens": [ + -4.018198013305664 + ], + "logprob_reward": -0.018197840079665184, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018197840079665184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13122963905334473, + 0.11220985651016235, + 0.06602680683135986, + 0.2742939591407776, + 0.028585586696863174, + 0.022892720997333527, + 0.050622113049030304, + 0.14795126020908356, + 0.10832623392343521, + 0.05786179006099701 + ], + "credit_map": [ + 0.020955976098775864, + 0.014524987898766994, + 0.021648533642292023, + 0.01934996247291565, + 0.015737146139144897, + 0.015283922664821148, + 0.014128393493592739, + 0.009600727818906307, + 0.0169976819306612, + 0.011306770145893097, + 0.01747617870569229, + 0.02400469034910202, + 0.011652044020593166, + 0.009996920824050903, + 0.011729457415640354, + 0.009046116843819618, + 0.01224547903984785, + 0.003963603638112545, + 0.0037958126049488783, + 0.013934298418462276, + 0.011096497066318989, + 0.007508448790758848, + 0.010308419354259968, + 0.003174249082803726, + 0.04345635324716568, + 0.011429880745708942, + 0.009115795604884624, + 0.046463966369628906, + 0.033452630043029785, + 0.09479578584432602, + 0.02262985147535801, + 0.012949745170772076, + 0.005165998823940754, + 0.002178165828809142, + 0.001772028161212802, + 0.0069372099824249744, + 0.00312838819809258, + 0.0024902287404984236, + 0.005289801396429539, + 0.0016237668460235, + 0.0034779265988618135, + 0.0016523497179150581, + 0.0015217927284538746, + 0.004953362047672272, + 0.0034633802715688944, + 0.0029496634379029274, + 0.003237456548959017, + 0.0016367920907214284, + 0.009629734791815281, + 0.004411724396049976, + 0.008222299627959728, + 0.010234725661575794, + 0.004193033557385206, + 0.004226693417876959, + 0.006014029961079359, + 0.0036898725666105747, + 0.024128824472427368, + 0.017548667266964912, + 0.010421832092106342, + 0.02442825958132744, + 0.015952356159687042, + 0.01333134900778532, + 0.03191588446497917, + 0.01022409088909626, + 0.01916918158531189, + 0.013989635743200779, + 0.009802731685340405, + 0.01935514621436596, + 0.01312593650072813, + 0.013682398945093155, + 0.011011078022420406, + 0.008190140128135681, + 0.009761935099959373, + 0.0033856683876365423, + 0.007224623579531908, + 0.0075361342169344425, + 0.008479922078549862, + 0.007495481986552477, + 0.006050858646631241, + 0.007927171885967255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.018197840079665184 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two well-dressed men visible bumping fists against a polished dark-wood bartop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first man wearing a navy wool blazer with a crisp white shirt cuff visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second man wearing a charcoal-gray blazer with a classic gold watch on his wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a crystal cognac decanter filled with rich amber liquid visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a luxurious wood-paneled bar with a background of mahogany panels and backlit bottles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023286793380975723 + ], + "opposite_logprob_tokens": [ + -3.773286819458008 + ], + "logprob_reward": -0.023286793380975723, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023286793380975723, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09030726552009583, + 0.16612054407596588, + 0.05749162659049034, + 0.24699187278747559, + 0.039879459887742996, + 0.020577311515808105, + 0.06330207735300064, + 0.16386453807353973, + 0.08959509432315826, + 0.0618702657520771 + ], + "credit_map": [ + 0.014979026280343533, + 0.00781850703060627, + 0.009500673972070217, + 0.010608324781060219, + 0.0186002217233181, + 0.010065846145153046, + 0.008904548361897469, + 0.00983010046184063, + 0.025594303384423256, + 0.015016925521194935, + 0.03184930235147476, + 0.03307617828249931, + 0.014159571379423141, + 0.015445929020643234, + 0.01761113852262497, + 0.013367164880037308, + 0.01016203686594963, + 0.0037505615036934614, + 0.003742144675925374, + 0.016712073236703873, + 0.005765073467046022, + 0.009054897353053093, + 0.005359491799026728, + 0.002945343963801861, + 0.039186205714941025, + 0.009971916675567627, + 0.008132952265441418, + 0.04107657074928284, + 0.030484097078442574, + 0.08419017493724823, + 0.021704383194446564, + 0.012245546095073223, + 0.007239336613565683, + 0.002645466709509492, + 0.002378332195803523, + 0.008250889368355274, + 0.0047393194399774075, + 0.0032130901236087084, + 0.008731278590857983, + 0.0026817426551133394, + 0.0034084238577634096, + 0.0013636827934533358, + 0.003111697267740965, + 0.004310606513172388, + 0.002165403449907899, + 0.0024989608209580183, + 0.0023187336046248674, + 0.0013998004142194986, + 0.011870881542563438, + 0.005234087817370892, + 0.008778596296906471, + 0.009881005622446537, + 0.005263492930680513, + 0.007221550680696964, + 0.009243748150765896, + 0.005808709189295769, + 0.023245403543114662, + 0.022438619285821915, + 0.02002613991498947, + 0.03126352280378342, + 0.016210485249757767, + 0.01658950373530388, + 0.020542435348033905, + 0.013548411428928375, + 0.0123756043612957, + 0.012038344517350197, + 0.005171896889805794, + 0.012178721837699413, + 0.00943823717534542, + 0.01823924295604229, + 0.011283238418400288, + 0.008869800716638565, + 0.006723285187035799, + 0.004722305573523045, + 0.00916270911693573, + 0.010950200259685516, + 0.006065757013857365, + 0.00799811165779829, + 0.005971962586045265, + 0.010275926440954208 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 7 + ], + "selected_reward": -0.023286793380975723 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two well-dressed men visible bumping fists against a polished dark-wood bartop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first man wearing a navy wool blazer with a crisp white shirt cuff visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second man wearing a charcoal-gray blazer with a classic gold watch on his wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a crystal cognac decanter filled with rich amber liquid visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a luxurious wood-paneled bar with a background of mahogany panels and backlit bottles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026328429579734802 + ], + "opposite_logprob_tokens": [ + -3.6513283252716064 + ], + "logprob_reward": -0.026328429579734802, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026328429579734802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1019309014081955, + 0.3019518554210663, + 0.08411340415477753, + 0.05791207775473595, + 0.03609941527247429, + 0.025287019088864326, + 0.10281924158334732, + 0.057576365768909454, + 0.1425146907567978, + 0.08979499340057373 + ], + "credit_map": [ + 0.021190766245126724, + 0.009940141811966896, + 0.011833997443318367, + 0.011842970736324787, + 0.020021354779601097, + 0.010428131558001041, + 0.01026476826518774, + 0.0064087738282978535, + 0.050266820937395096, + 0.020113812759518623, + 0.013980595394968987, + 0.03786606714129448, + 0.06425536423921585, + 0.026987433433532715, + 0.07352984696626663, + 0.014951917342841625, + 0.012769137509167194, + 0.005808641202747822, + 0.006069007329642773, + 0.018813256174325943, + 0.010679857805371284, + 0.012201759032905102, + 0.01366336364299059, + 0.004108377732336521, + 0.009713717736303806, + 0.002507902914658189, + 0.0028361419681459665, + 0.00878210924565792, + 0.006400300655514002, + 0.020198184996843338, + 0.004642510786652565, + 0.0028312106151133776, + 0.005561447702348232, + 0.0030686305835843086, + 0.0032637319527566433, + 0.008410269394516945, + 0.003195195458829403, + 0.005447640549391508, + 0.004083615727722645, + 0.003068882506340742, + 0.0038855718448758125, + 0.0016189711168408394, + 0.0039010022301226854, + 0.004422447644174099, + 0.0027407759334892035, + 0.0033445667941123247, + 0.0027568270452320576, + 0.0026168571785092354, + 0.022309135645627975, + 0.008518227376043797, + 0.017744364216923714, + 0.016846181824803352, + 0.008093832060694695, + 0.013704229146242142, + 0.008579893968999386, + 0.007023376412689686, + 0.009459780529141426, + 0.009308520704507828, + 0.004136200062930584, + 0.010444398038089275, + 0.006081896834075451, + 0.005716064944863319, + 0.008018890395760536, + 0.004410608671605587, + 0.02180272713303566, + 0.015342099592089653, + 0.010486266575753689, + 0.019691944122314453, + 0.018306700512766838, + 0.02522362768650055, + 0.01682346686720848, + 0.014837861061096191, + 0.011619621887803078, + 0.0062126838602125645, + 0.008227638900279999, + 0.01060442440211773, + 0.011122899129986763, + 0.010330782271921635, + 0.01623726263642311, + 0.015439677983522415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 6 + ], + "selected_reward": -0.026328429579734802 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two well-dressed men visible bumping fists against a polished dark-wood bartop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first man wearing a navy wool blazer with a crisp white shirt cuff visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second man wearing a charcoal-gray blazer with a classic gold watch on his wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a crystal cognac decanter filled with rich amber liquid visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a luxurious wood-paneled bar with a background of mahogany panels and backlit bottles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012528643012046814 + ], + "opposite_logprob_tokens": [ + -4.387528419494629 + ], + "logprob_reward": -0.012528643012046814, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012528643012046814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1487520933151245, + 0.2140830010175705, + 0.1686696857213974, + 0.09465519338846207, + 0.04688991233706474, + 0.07395276427268982, + 0.13996003568172455, + 0.025657247751951218, + 0.05576961487531662, + 0.03161044791340828 + ], + "credit_map": [ + 0.013971593230962753, + 0.011016675271093845, + 0.01039152592420578, + 0.020716995000839233, + 0.021044041961431503, + 0.02146931178867817, + 0.024673283100128174, + 0.025468671694397926, + 0.030642330646514893, + 0.01490054838359356, + 0.02766156941652298, + 0.03282500058412552, + 0.031208094209432602, + 0.022554360330104828, + 0.039394937455654144, + 0.014896153472363949, + 0.027486231178045273, + 0.011685692705214024, + 0.01239705365151167, + 0.02644023858010769, + 0.019434822723269463, + 0.028225736692547798, + 0.03242858126759529, + 0.010571340098977089, + 0.01335633173584938, + 0.0060754031874239445, + 0.007099443580955267, + 0.014164216816425323, + 0.01158694364130497, + 0.0269851665943861, + 0.009349552914500237, + 0.006038136314600706, + 0.007348822429776192, + 0.0033304591197520494, + 0.005409155040979385, + 0.007576265837997198, + 0.005622330121695995, + 0.008626492694020271, + 0.0043988460674881935, + 0.004577541723847389, + 0.012193799950182438, + 0.005516353994607925, + 0.00452263280749321, + 0.011060158722102642, + 0.007888562977313995, + 0.013970435597002506, + 0.010219044052064419, + 0.008581776171922684, + 0.020890934392809868, + 0.014584660530090332, + 0.012079396285116673, + 0.020340364426374435, + 0.01442318968474865, + 0.026051774621009827, + 0.017627423629164696, + 0.013962304219603539, + 0.004089464899152517, + 0.00271605234593153, + 0.002569949021562934, + 0.0047040218487381935, + 0.0026954335626214743, + 0.0028881870675832033, + 0.0036278225015848875, + 0.0023663160391151905, + 0.00838091503828764, + 0.005072388797998428, + 0.0052746920846402645, + 0.006968718953430653, + 0.006369410548359156, + 0.011459915898740292, + 0.007779689040035009, + 0.004463882185518742, + 0.0041767871007323265, + 0.003389450255781412, + 0.0031401512678712606, + 0.0050880806520581245, + 0.0037008984945714474, + 0.00424081040546298, + 0.002816948341205716, + 0.005057319067418575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.012528643012046814 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_3/index_tensor([9142, 9142])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_3/index_tensor([9142, 9142])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3cac5cabea6111a46bbcbc9cba5c98fe62315f2d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_3/index_tensor([9142, 9142])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "tracking shot, beginning at a very low, upward angle, follows a robotic arm, defined entirely by sharp, clean black lines. It exists within a clean, black and white world, set against a backdrop of infinitely stretching white space. Each segment of the arm is connected by complex spherical joints and a sophisticated biomimetic structure. Its surface is devoid of any sheen or shading, dominated by pure white fill and decisive black outlines. Every sense of dimension and structure is expressed through the density, thickness, and precise perspective of the lines. As the camera steadily spirals upward, we witness the arm ascending with the graceful grace of a vine, smoothly and silently unfolding, twisting, and relocking section by section in a precise and rhythmic movement, as if a living creature is growing, reaching resolutely toward the unknown beyond the frame. The entire process, framed by a minimalist visual style, creates a unique atmosphere that blends futuristic technology, abstract dynamic beauty, and a sense of boundless exploration.", + "seed": 17988, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_3/index_9142_seed_17988_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01624805168248713, + "vlm_logprob": -0.01624805168248713, + "vlm_yes_no_margin": 4.55, + "vlm_token_credit": -0.01624805168248713, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a robotic arm with a biomimetic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by sharp black lines and pure white fill without shading or sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinitely stretching white space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the robotic arm ascend while twisting and unfolding section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low upward angle and steadily spiral upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029198534786701202 + ], + "opposite_logprob_tokens": [ + -3.90419864654541 + ], + "logprob_reward": -0.029198534786701202, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.029198534786701202, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30684417486190796, + 0.10159637033939362, + 0.0729825347661972, + 0.07975750416517258, + 0.10237262398004532, + 0.05632273852825165, + 0.06702914088964462, + 0.05976666137576103, + 0.07158999890089035, + 0.08173830062150955 + ], + "credit_map": [ + 0.030830280855298042, + 0.038964953273534775, + 0.04191431775689125, + 0.03879716247320175, + 0.03423967584967613, + 0.03717806935310364, + 0.05361759662628174, + 0.0313020795583725, + 0.01468750275671482, + 0.008864360861480236, + 0.01474086195230484, + 0.01502982247620821, + 0.007644389756023884, + 0.009081508032977581, + 0.013528070412576199, + 0.01801985874772072, + 0.008887325413525105, + 0.008929800242185593, + 0.01144935842603445, + 0.009346295148134232, + 0.006971513852477074, + 0.00646928371861577, + 0.009229477494955063, + 0.011699484661221504, + 0.010794434696435928, + 0.009689834900200367, + 0.009420257061719894, + 0.008811326697468758, + 0.01342522632330656, + 0.00879786629229784, + 0.009554028511047363, + 0.009264525026082993, + 0.01311050821095705, + 0.015656281262636185, + 0.01696752943098545, + 0.007789153140038252, + 0.00808646809309721, + 0.009452160447835922, + 0.007897582836449146, + 0.02341294288635254, + 0.014696375466883183, + 0.0035912857856601477, + 0.004748333711177111, + 0.002920519793406129, + 0.004864073824137449, + 0.015176298096776009, + 0.005078588146716356, + 0.0052472613751888275, + 0.010997358709573746, + 0.005130951292812824, + 0.011641211807727814, + 0.006616917438805103, + 0.005823522340506315, + 0.0057477629743516445, + 0.012152673676609993, + 0.008918738923966885, + 0.008361690677702427, + 0.010039878077805042, + 0.005762502085417509, + 0.006182132288813591, + 0.005114573519676924, + 0.004281486384570599, + 0.01384482067078352, + 0.00617957254871726, + 0.01138429343700409, + 0.005325039848685265, + 0.008958058431744576, + 0.00713385408744216, + 0.008292177692055702, + 0.015275768004357815, + 0.006207296159118414, + 0.009013517759740353, + 0.015679677948355675, + 0.005439367610961199, + 0.01029776968061924, + 0.00580610753968358, + 0.008216879330575466, + 0.01553640142083168, + 0.013411480002105236, + 0.007350610103458166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.029198534786701202 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a robotic arm with a biomimetic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by sharp black lines and pure white fill without shading or sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinitely stretching white space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the robotic arm ascend while twisting and unfolding section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low upward angle and steadily spiral upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014212694019079208 + ], + "opposite_logprob_tokens": [ + -4.264212608337402 + ], + "logprob_reward": -0.014212694019079208, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014212694019079208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12705577909946442, + 0.2710121273994446, + 0.10985949635505676, + 0.052186112850904465, + 0.05494249239563942, + 0.0879799872636795, + 0.04536832123994827, + 0.04043733328580856, + 0.06987212598323822, + 0.1412862092256546 + ], + "credit_map": [ + 0.012559503316879272, + 0.01368994265794754, + 0.02033039927482605, + 0.01880987174808979, + 0.010966183617711067, + 0.013637125492095947, + 0.02026800625026226, + 0.016794752329587936, + 0.0354110412299633, + 0.021517891436815262, + 0.043919455260038376, + 0.03829478845000267, + 0.018170548602938652, + 0.02267470210790634, + 0.04455694556236267, + 0.04646674543619156, + 0.011781333014369011, + 0.011811373755335808, + 0.016573935747146606, + 0.013058459386229515, + 0.015799200162291527, + 0.008591447956860065, + 0.017701061442494392, + 0.014542686752974987, + 0.005394784267991781, + 0.004920826759189367, + 0.007503247819840908, + 0.006401827558875084, + 0.009202810935676098, + 0.005806522909551859, + 0.008188358508050442, + 0.004767728503793478, + 0.0039078183472156525, + 0.00559233920648694, + 0.012883445248007774, + 0.004116204101592302, + 0.004165505524724722, + 0.0042728702537715435, + 0.004429937340319157, + 0.015574377030134201, + 0.025347303599119186, + 0.005146730691194534, + 0.008390662260353565, + 0.0052430168725550175, + 0.006261971779167652, + 0.020502036437392235, + 0.009119110181927681, + 0.007969152182340622, + 0.0066536953672766685, + 0.004592900164425373, + 0.004946803208440542, + 0.004975659307092428, + 0.004558661486953497, + 0.0039019195828586817, + 0.009693664498627186, + 0.0060450173914432526, + 0.006009629461914301, + 0.0059876092709600925, + 0.00340531999245286, + 0.004342753905802965, + 0.0029414596501737833, + 0.003331640036776662, + 0.01000250969082117, + 0.004416412673890591, + 0.01030312292277813, + 0.005515589378774166, + 0.00895408820360899, + 0.007279573939740658, + 0.005764619912952185, + 0.013670682907104492, + 0.006887500640004873, + 0.011496949940919876, + 0.025480302050709724, + 0.010989933274686337, + 0.022625375539064407, + 0.012076989747583866, + 0.014306715689599514, + 0.023274686187505722, + 0.014848494902253151, + 0.017683714628219604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.014212694019079208 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a robotic arm with a biomimetic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by sharp black lines and pure white fill without shading or sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinitely stretching white space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the robotic arm ascend while twisting and unfolding section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low upward angle and steadily spiral upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780935011804104 + ], + "opposite_logprob_tokens": [ + -4.6347808837890625 + ], + "logprob_reward": -0.009780935011804104, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780935011804104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11321960389614105, + 0.2209509164094925, + 0.039868906140327454, + 0.03389904648065567, + 0.046933140605688095, + 0.14053913950920105, + 0.03646151348948479, + 0.054916705936193466, + 0.058554671704769135, + 0.2546563744544983 + ], + "credit_map": [ + 0.020916368812322617, + 0.01226005144417286, + 0.016746800392866135, + 0.016902999952435493, + 0.012053938582539558, + 0.010462403297424316, + 0.010348389856517315, + 0.013528653420507908, + 0.03075275756418705, + 0.01613890565931797, + 0.033370669931173325, + 0.03376145288348198, + 0.019022878259420395, + 0.021759407594799995, + 0.034348372370004654, + 0.03179647773504257, + 0.004754374735057354, + 0.0038748052902519703, + 0.0062928637489676476, + 0.004718820098787546, + 0.0044167786836624146, + 0.0033265254460275173, + 0.005973496474325657, + 0.006511240266263485, + 0.005123058799654245, + 0.004389809910207987, + 0.004773661494255066, + 0.0036027811001986265, + 0.004529206547886133, + 0.003120227251201868, + 0.004788815509527922, + 0.0035714854020625353, + 0.0034285762812942266, + 0.0040836026892066, + 0.011104180477559566, + 0.002544337883591652, + 0.0039481851272284985, + 0.0033447605092078447, + 0.003701210953295231, + 0.014778287149965763, + 0.04777190834283829, + 0.004828353878110647, + 0.05345065891742706, + 0.007201167289167643, + 0.005657013040035963, + 0.004292353056371212, + 0.009729452431201935, + 0.007608239538967609, + 0.006453092209994793, + 0.0038658652920275927, + 0.004624675959348679, + 0.0030493238009512424, + 0.0044313897378742695, + 0.003177331294864416, + 0.007623704150319099, + 0.0032361275516450405, + 0.007603670936077833, + 0.006786549463868141, + 0.003957743290811777, + 0.00690887775272131, + 0.0036670586559921503, + 0.005973477382212877, + 0.013361730612814426, + 0.006657598540186882, + 0.009677546098828316, + 0.004432663321495056, + 0.0060484278947114944, + 0.007765612564980984, + 0.005648257676512003, + 0.013100190088152885, + 0.00504589919000864, + 0.006836074870079756, + 0.047134727239608765, + 0.02055284194648266, + 0.0398249588906765, + 0.023099901154637337, + 0.0249880813062191, + 0.044650349766016006, + 0.02614721842110157, + 0.028258299455046654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 5 + ], + "selected_reward": -0.009780935011804104 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a robotic arm with a biomimetic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by sharp black lines and pure white fill without shading or sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinitely stretching white space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the robotic arm ascend while twisting and unfolding section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low upward angle and steadily spiral upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017149040941148996 + ], + "opposite_logprob_tokens": [ + -6.376714706420898 + ], + "logprob_reward": -0.0017149040941148996, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017149040941148996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29631131887435913, + 0.06789473444223404, + 0.04986650496721268, + 0.04507077857851982, + 0.11640651524066925, + 0.14496061205863953, + 0.04078134894371033, + 0.027385635301470757, + 0.0630059614777565, + 0.14831656217575073 + ], + "credit_map": [ + 0.05429486557841301, + 0.03541624918580055, + 0.04042353481054306, + 0.047505252063274384, + 0.02091922052204609, + 0.026409221813082695, + 0.02998959831893444, + 0.041353367269039154, + 0.007933446206152439, + 0.00987937580794096, + 0.009288357570767403, + 0.00936164055019617, + 0.005818347446620464, + 0.0062627834267914295, + 0.009103028103709221, + 0.010247753001749516, + 0.005115729756653309, + 0.004640843253582716, + 0.008590097539126873, + 0.00604936433956027, + 0.005717982538044453, + 0.004750360734760761, + 0.006399643607437611, + 0.00860248040407896, + 0.00502765690907836, + 0.003703819354996085, + 0.00637618126347661, + 0.006405558902770281, + 0.00657677510753274, + 0.004098547622561455, + 0.008032012730836868, + 0.004850225057452917, + 0.009759173728525639, + 0.011201830580830574, + 0.031012455001473427, + 0.006355229765176773, + 0.005395088344812393, + 0.009044583886861801, + 0.006327926181256771, + 0.03731023520231247, + 0.017952922731637955, + 0.005986832547932863, + 0.0469769723713398, + 0.007165129762142897, + 0.00998789444565773, + 0.03574518859386444, + 0.00815347395837307, + 0.012992208823561668, + 0.007042296230792999, + 0.004234864376485348, + 0.005354446824640036, + 0.0035360080655664206, + 0.003591354237869382, + 0.0034126429818570614, + 0.009716815315186977, + 0.003892922308295965, + 0.003804638748988509, + 0.0031161699444055557, + 0.003382049733772874, + 0.003232513088732958, + 0.0018251561559736729, + 0.0028452437836676836, + 0.005281873047351837, + 0.00389799103140831, + 0.009367356076836586, + 0.01015464961528778, + 0.007471080403774977, + 0.006862659007310867, + 0.006890533026307821, + 0.009555963799357414, + 0.00588054396212101, + 0.006823173724114895, + 0.0291582141071558, + 0.017065050080418587, + 0.025435836985707283, + 0.013227459974586964, + 0.014326149597764015, + 0.019143935292959213, + 0.015499980188906193, + 0.014459945261478424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.0017149040941148996 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a robotic arm with a biomimetic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by sharp black lines and pure white fill without shading or sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinitely stretching white space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the robotic arm ascend while twisting and unfolding section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low upward angle and steadily spiral upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026333190500736237 + ], + "opposite_logprob_tokens": [ + -3.6513330936431885 + ], + "logprob_reward": -0.026333190500736237, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026333190500736237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12520916759967804, + 0.042611658573150635, + 0.02662985399365425, + 0.04097211733460426, + 0.06809158623218536, + 0.3661006987094879, + 0.07023927569389343, + 0.05479002371430397, + 0.15085755288600922, + 0.05449806898832321 + ], + "credit_map": [ + 0.01627342216670513, + 0.012661215849220753, + 0.017575684934854507, + 0.015481572598218918, + 0.012525164522230625, + 0.017424190416932106, + 0.014703071676194668, + 0.018564846366643906, + 0.005337497219443321, + 0.005849745590239763, + 0.006491272244602442, + 0.006514992564916611, + 0.0031260403338819742, + 0.004174548666924238, + 0.004541228991001844, + 0.006576329469680786, + 0.003781566396355629, + 0.002981537487357855, + 0.004167979583144188, + 0.003018174320459366, + 0.0031406194902956486, + 0.002437487244606018, + 0.003548469627276063, + 0.003554019145667553, + 0.005232212133705616, + 0.004262655973434448, + 0.005048027262091637, + 0.005636362358927727, + 0.0060437931679189205, + 0.004105915781110525, + 0.004506790544837713, + 0.006136360112577677, + 0.00845049973577261, + 0.006075829267501831, + 0.015340807847678661, + 0.00352831045165658, + 0.0035835641901940107, + 0.005478677339851856, + 0.0036472666542977095, + 0.021986624225974083, + 0.1388307809829712, + 0.012272624298930168, + 0.14085976779460907, + 0.015945814549922943, + 0.009053532965481281, + 0.016748866066336632, + 0.016810446977615356, + 0.015578847378492355, + 0.012357049621641636, + 0.00622852286323905, + 0.009556606411933899, + 0.006385673303157091, + 0.005789358634501696, + 0.006001560017466545, + 0.014889356680214405, + 0.00903114303946495, + 0.007433717604726553, + 0.004339275415986776, + 0.0063148383051157, + 0.007905282080173492, + 0.003425546921789646, + 0.0049104648642241955, + 0.014017091132700443, + 0.0064438022673130035, + 0.024359894916415215, + 0.01361838448792696, + 0.013038991950452328, + 0.012796180322766304, + 0.01589684747159481, + 0.042460035532712936, + 0.010520664975047112, + 0.01816653087735176, + 0.009887331165373325, + 0.006341719068586826, + 0.007454360835254192, + 0.004645670764148235, + 0.004697734490036964, + 0.01071581244468689, + 0.005123611073940992, + 0.0056318249553442 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.026333190500736237 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tracking shot, beginning at a very low, upward angle, follows a robotic arm, defined entirely by sharp, clean black lines. It exists within a clean, black and white world, set against a backdrop of infinitely stretching white space. Each segment of the arm is connected by complex spherical joints and a sophisticated biomimetic structure. Its surface is devoid of any sheen or shading, dominated by pure white fill and decisive black outlines. Every sense of dimension and structure is expressed through the density, thickness, and precise perspective of the lines. As the camera steadily spirals upward, we witness the arm ascending with the graceful grace of a vine, smoothly and silently unfolding, twisting, and relocking section by section in a precise and rhythmic movement, as if a living creature is growing, reaching resolutely toward the unknown beyond the frame. The entire process, framed by a minimalist visual style, creates a unique atmosphere that blends futuristic technology, abstract dynamic beauty, and a sense of boundless exploration.", + "seed": 17989, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_3/index_9142_seed_17989_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06637393203563988, + "vlm_logprob": -0.06637393203563988, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.06637393203563988, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a robotic arm with a biomimetic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by sharp black lines and pure white fill without shading or sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinitely stretching white space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the robotic arm ascend while twisting and unfolding section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low upward angle and steadily spiral upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.037364792078733444 + ], + "opposite_logprob_tokens": [ + -3.6623647212982178 + ], + "logprob_reward": -0.037364792078733444, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.037364792078733444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25359728932380676, + 0.058160822838544846, + 0.17548076808452606, + 0.05628637224435806, + 0.13055546581745148, + 0.03874148055911064, + 0.053915683180093765, + 0.08252675831317902, + 0.12575511634349823, + 0.024980265647172928 + ], + "credit_map": [ + 0.02164560556411743, + 0.024439549073576927, + 0.03180846944451332, + 0.028052108362317085, + 0.02727462537586689, + 0.019445110112428665, + 0.056989818811416626, + 0.04394199699163437, + 0.0034871867392212152, + 0.005601781420409679, + 0.013538564555346966, + 0.005131441634148359, + 0.005726024508476257, + 0.004642852116376162, + 0.011024015955626965, + 0.009008953347802162, + 0.013133144937455654, + 0.017888527363538742, + 0.032923221588134766, + 0.013543505221605301, + 0.014873584732413292, + 0.015257399529218674, + 0.03426974639296532, + 0.03359163925051689, + 0.0039212461560964584, + 0.003913738299161196, + 0.012188292108476162, + 0.0038176618982106447, + 0.004502672236412764, + 0.003780182683840394, + 0.01171191968023777, + 0.012450658716261387, + 0.008402782492339611, + 0.008313019759953022, + 0.02752959541976452, + 0.007578561082482338, + 0.008711859583854675, + 0.021952293813228607, + 0.012589864432811737, + 0.03547748923301697, + 0.005733105354011059, + 0.004042921122163534, + 0.004514231346547604, + 0.004485939629375935, + 0.003649673890322447, + 0.005084322765469551, + 0.005573740229010582, + 0.005657549481838942, + 0.006105865817517042, + 0.005690844729542732, + 0.01177461352199316, + 0.00766382459551096, + 0.004950555041432381, + 0.007102399133145809, + 0.005987375974655151, + 0.004640205297619104, + 0.007589306216686964, + 0.007878883741796017, + 0.006699883844703436, + 0.008584513328969479, + 0.01603519916534424, + 0.01714080013334751, + 0.0067135365679860115, + 0.011884629726409912, + 0.01744873635470867, + 0.010944495908915997, + 0.01425654161721468, + 0.010393199510872364, + 0.014014470390975475, + 0.0260679442435503, + 0.0113235367462039, + 0.021306199952960014, + 0.002763821743428707, + 0.0030493736267089844, + 0.002882773522287607, + 0.0031338571570813656, + 0.0026002514641731977, + 0.003780147060751915, + 0.0034949115943163633, + 0.003275129944086075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.037364792078733444 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a robotic arm with a biomimetic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by sharp black lines and pure white fill without shading or sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinitely stretching white space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the robotic arm ascend while twisting and unfolding section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low upward angle and steadily spiral upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00980312842875719 + ], + "opposite_logprob_tokens": [ + -4.634803295135498 + ], + "logprob_reward": -0.00980312842875719, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00980312842875719, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09055682271718979, + 0.21460823714733124, + 0.09172859787940979, + 0.06391666829586029, + 0.2099630981683731, + 0.09651144593954086, + 0.04019036516547203, + 0.046609144657850266, + 0.04743671417236328, + 0.09847896546125412 + ], + "credit_map": [ + 0.009576062671840191, + 0.008913416415452957, + 0.010956871323287487, + 0.008374153636395931, + 0.009795506484806538, + 0.010105010122060776, + 0.019058378413319588, + 0.013777420856058598, + 0.01499891933053732, + 0.02103094570338726, + 0.030348781496286392, + 0.01986609771847725, + 0.024526255205273628, + 0.024895936250686646, + 0.04529959708452225, + 0.033641695976257324, + 0.006424856372177601, + 0.01180688664317131, + 0.015926538035273552, + 0.008009743876755238, + 0.007512554991990328, + 0.012699590995907784, + 0.013964868150651455, + 0.01538356114178896, + 0.00534029072150588, + 0.006297520827502012, + 0.009421649388968945, + 0.006341369356960058, + 0.006824991200119257, + 0.014072753489017487, + 0.007412266451865435, + 0.008205828256905079, + 0.011955684050917625, + 0.009208320640027523, + 0.04635670781135559, + 0.009750026278197765, + 0.011236063204705715, + 0.04050583764910698, + 0.01842607371509075, + 0.06252438575029373, + 0.01173867005854845, + 0.0106077054515481, + 0.014240504242479801, + 0.012328622862696648, + 0.008901680819690228, + 0.011702945455908775, + 0.014749882742762566, + 0.012241438962519169, + 0.004571558441966772, + 0.005170741584151983, + 0.006287280470132828, + 0.0050276946276426315, + 0.004405531566590071, + 0.005239587742835283, + 0.004636118188500404, + 0.004851848818361759, + 0.005019571632146835, + 0.004995913710445166, + 0.004225100390613079, + 0.004953925032168627, + 0.005747437011450529, + 0.009734747000038624, + 0.003293254179880023, + 0.008639195933938026, + 0.005159738473594189, + 0.005126536823809147, + 0.004644925706088543, + 0.004759009461849928, + 0.004659139085561037, + 0.00942312739789486, + 0.005907290615141392, + 0.007756948936730623, + 0.018263692036271095, + 0.007585193030536175, + 0.009546956047415733, + 0.010339902713894844, + 0.009399519301950932, + 0.018706360831856728, + 0.008457580581307411, + 0.016179751604795456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 9 + ], + "selected_reward": -0.00980312842875719 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a robotic arm with a biomimetic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by sharp black lines and pure white fill without shading or sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinitely stretching white space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the robotic arm ascend while twisting and unfolding section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low upward angle and steadily spiral upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25209563970565796 + ], + "opposite_logprob_tokens": [ + -1.5020956993103027 + ], + "logprob_reward": -0.25209563970565796, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25209563970565796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08780886977910995, + 0.39227208495140076, + 0.07712847739458084, + 0.08604919910430908, + 0.1974320113658905, + 0.027040909975767136, + 0.06698912382125854, + 0.025367140769958496, + 0.023288791999220848, + 0.016623372212052345 + ], + "credit_map": [ + 0.007303192280232906, + 0.009011643007397652, + 0.010967577807605267, + 0.00893904734402895, + 0.01017822790890932, + 0.006433692295104265, + 0.019855080172419548, + 0.015120411291718483, + 0.027134645730257034, + 0.04046441614627838, + 0.05816684290766716, + 0.035670261830091476, + 0.04474213719367981, + 0.04527497664093971, + 0.07992634177207947, + 0.060892488807439804, + 0.005158254876732826, + 0.007609257008880377, + 0.012008643709123135, + 0.006584168411791325, + 0.00638806214556098, + 0.013223906978964806, + 0.013059941120445728, + 0.013096252456307411, + 0.008348488248884678, + 0.009985094889998436, + 0.011080075055360794, + 0.007962830364704132, + 0.009114648215472698, + 0.018601857125759125, + 0.010433477349579334, + 0.010522738099098206, + 0.017158618196845055, + 0.014275097288191319, + 0.03910201042890549, + 0.020120153203606606, + 0.016926484182476997, + 0.026492828503251076, + 0.0303674153983593, + 0.03298942372202873, + 0.00250658905133605, + 0.0026138541288673878, + 0.004125524777919054, + 0.003038342110812664, + 0.00217807712033391, + 0.0036321256775408983, + 0.0029935468919575214, + 0.005952851381152868, + 0.005440703127533197, + 0.005048849619925022, + 0.021294105798006058, + 0.005103239789605141, + 0.004481688141822815, + 0.01253960095345974, + 0.007398991379886866, + 0.005681947804987431, + 0.0033131970558315516, + 0.0020298794843256474, + 0.0027062532026320696, + 0.0026647199410945177, + 0.004225810989737511, + 0.004045681096613407, + 0.0021494899410754442, + 0.004232111386954784, + 0.0025947873946279287, + 0.002233352279290557, + 0.003948522731661797, + 0.0020793916191905737, + 0.0026151728816330433, + 0.004459857475012541, + 0.0016581028467044234, + 0.0036996062844991684, + 0.0036453627981245518, + 0.0017334874719381332, + 0.0018430729396641254, + 0.0016474229050800204, + 0.0015894609969109297, + 0.002188384532928467, + 0.0013361838646233082, + 0.002639999147504568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.25209563970565796 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a robotic arm with a biomimetic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by sharp black lines and pure white fill without shading or sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinitely stretching white space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the robotic arm ascend while twisting and unfolding section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low upward angle and steadily spiral upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002818423556163907 + ], + "opposite_logprob_tokens": [ + -5.877818584442139 + ], + "logprob_reward": -0.002818423556163907, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002818423556163907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18796974420547485, + 0.06428337097167969, + 0.08496955782175064, + 0.04415325075387955, + 0.08448127657175064, + 0.03773076459765434, + 0.29831400513648987, + 0.06290437281131744, + 0.09731021523475647, + 0.03788347542285919 + ], + "credit_map": [ + 0.017474975436925888, + 0.016291402280330658, + 0.030820632353425026, + 0.0193890780210495, + 0.01653515361249447, + 0.020165061578154564, + 0.038545362651348114, + 0.02874806895852089, + 0.004293147940188646, + 0.008982748724520206, + 0.011998563073575497, + 0.0050393459387123585, + 0.00795698817819357, + 0.0038035502657294273, + 0.012216790579259396, + 0.009992232546210289, + 0.006455667782574892, + 0.008375387638807297, + 0.016290374100208282, + 0.008945755660533905, + 0.00798935629427433, + 0.013175338506698608, + 0.010095110163092613, + 0.013642572797834873, + 0.003684402210637927, + 0.00386993121355772, + 0.009697276167571545, + 0.002578123239800334, + 0.0037653138861060143, + 0.00400681933388114, + 0.007877017371356487, + 0.008674368262290955, + 0.010546445846557617, + 0.010951319709420204, + 0.012194623239338398, + 0.008767182938754559, + 0.008127804845571518, + 0.010751071386039257, + 0.011157487519085407, + 0.011985345743596554, + 0.0043239109218120575, + 0.004513840191066265, + 0.004764441400766373, + 0.006314513739198446, + 0.003039122326299548, + 0.005372556392103434, + 0.004772686865180731, + 0.004629690200090408, + 0.021568218246102333, + 0.020044371485710144, + 0.08662668615579605, + 0.02538847178220749, + 0.02328294888138771, + 0.06064226105809212, + 0.03317977115511894, + 0.02758130244910717, + 0.008442966267466545, + 0.006739613134413958, + 0.00500163808465004, + 0.007825781591236591, + 0.008349712938070297, + 0.008461405523121357, + 0.005491810385137796, + 0.012591446749866009, + 0.012372546829283237, + 0.00805775634944439, + 0.013912491500377655, + 0.007936707697808743, + 0.011091658845543861, + 0.019709402695298195, + 0.007761097513139248, + 0.016468549147248268, + 0.00453999824821949, + 0.00363521627150476, + 0.0032934958580881357, + 0.005251383408904076, + 0.005856670904904604, + 0.006502438802272081, + 0.0035783550702035427, + 0.005225915927439928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.002818423556163907 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a robotic arm with a biomimetic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by sharp black lines and pure white fill without shading or sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinitely stretching white space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the robotic arm ascend while twisting and unfolding section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low upward angle and steadily spiral upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978767640888691 + ], + "opposite_logprob_tokens": [ + -3.52978777885437 + ], + "logprob_reward": -0.02978767640888691, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978767640888691, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19034792482852936, + 0.1240767315030098, + 0.10654149949550629, + 0.0510062612593174, + 0.09946125745773315, + 0.0913718193769455, + 0.16943001747131348, + 0.085819773375988, + 0.04769138991832733, + 0.03425333648920059 + ], + "credit_map": [ + 0.015004066750407219, + 0.013051108457148075, + 0.028266457840800285, + 0.0168822780251503, + 0.0158193651586771, + 0.017016040161252022, + 0.045788101851940155, + 0.03852050378918648, + 0.011189536191523075, + 0.015348180197179317, + 0.023944800719618797, + 0.009201657958328724, + 0.011625089682638645, + 0.010888474993407726, + 0.022070854902267456, + 0.01980813406407833, + 0.007556964177638292, + 0.012217800132930279, + 0.022349951788783073, + 0.00920721422880888, + 0.008759543299674988, + 0.013223731890320778, + 0.015475191175937653, + 0.017751097679138184, + 0.004103624727576971, + 0.0051356032490730286, + 0.00871584378182888, + 0.003905970137566328, + 0.005423384718596935, + 0.007967171259224415, + 0.007488918956369162, + 0.008265743032097816, + 0.00900921132415533, + 0.00953421276062727, + 0.016745707020163536, + 0.009182491339743137, + 0.007451229263097048, + 0.013167155906558037, + 0.014212334528565407, + 0.020158924162387848, + 0.013300036080181599, + 0.011191170662641525, + 0.011732622981071472, + 0.012353655882179737, + 0.006827501580119133, + 0.01277687307447195, + 0.013334894552826881, + 0.0098550571128726, + 0.015097052790224552, + 0.012568779289722443, + 0.0491473525762558, + 0.014340464025735855, + 0.0134507417678833, + 0.03016665019094944, + 0.01989150047302246, + 0.014767464250326157, + 0.011020153760910034, + 0.008353076875209808, + 0.008193226531147957, + 0.013230151496827602, + 0.012316345237195492, + 0.011424394324421883, + 0.008489555679261684, + 0.01279287226498127, + 0.0057831271551549435, + 0.00474699167534709, + 0.009489655494689941, + 0.00418200995773077, + 0.005097220651805401, + 0.0058292304165661335, + 0.0041941991075873375, + 0.008368952199816704, + 0.004278568085283041, + 0.0028177788481116295, + 0.0034821960143744946, + 0.0045095765963196754, + 0.004176192916929722, + 0.006484901532530785, + 0.0032074889168143272, + 0.005296635907143354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.02978767640888691 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tracking shot, beginning at a very low, upward angle, follows a robotic arm, defined entirely by sharp, clean black lines. It exists within a clean, black and white world, set against a backdrop of infinitely stretching white space. Each segment of the arm is connected by complex spherical joints and a sophisticated biomimetic structure. Its surface is devoid of any sheen or shading, dominated by pure white fill and decisive black outlines. Every sense of dimension and structure is expressed through the density, thickness, and precise perspective of the lines. As the camera steadily spirals upward, we witness the arm ascending with the graceful grace of a vine, smoothly and silently unfolding, twisting, and relocking section by section in a precise and rhythmic movement, as if a living creature is growing, reaching resolutely toward the unknown beyond the frame. The entire process, framed by a minimalist visual style, creates a unique atmosphere that blends futuristic technology, abstract dynamic beauty, and a sense of boundless exploration.", + "seed": 17990, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_3/index_9142_seed_17990_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.13891587159596383, + "vlm_logprob": -0.13891587159596383, + "vlm_yes_no_margin": 2.749999988079071, + "vlm_token_credit": -0.13891587159596383, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a robotic arm with a biomimetic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by sharp black lines and pure white fill without shading or sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinitely stretching white space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the robotic arm ascend while twisting and unfolding section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low upward angle and steadily spiral upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1098143607378006 + ], + "opposite_logprob_tokens": [ + -2.359814405441284 + ], + "logprob_reward": -0.1098143607378006, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1098143607378006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3649175763130188, + 0.0819431021809578, + 0.04247505962848663, + 0.14493967592716217, + 0.06487017869949341, + 0.13688358664512634, + 0.061756253242492676, + 0.036606237292289734, + 0.03823208808898926, + 0.02737628109753132 + ], + "credit_map": [ + 0.030867066234350204, + 0.04592188075184822, + 0.05426449328660965, + 0.03067191131412983, + 0.042865876108407974, + 0.09113135933876038, + 0.02503320947289467, + 0.044161759316921234, + 0.008755878545343876, + 0.007701915223151445, + 0.019553499296307564, + 0.005673260428011417, + 0.006131002679467201, + 0.01652655191719532, + 0.012988573871552944, + 0.004612425807863474, + 0.00606720382347703, + 0.002999101299792528, + 0.0045669665560126305, + 0.006727183237671852, + 0.0039003053680062294, + 0.008042012341320515, + 0.003940379247069359, + 0.00623190775513649, + 0.015000753104686737, + 0.012916358187794685, + 0.025624316185712814, + 0.03656797111034393, + 0.008152447640895844, + 0.02585800550878048, + 0.012207289226353168, + 0.008612552657723427, + 0.005266559310257435, + 0.0019858991727232933, + 0.023989379405975342, + 0.0020987745374441147, + 0.002995237475261092, + 0.01306870486587286, + 0.0025014092680066824, + 0.0129642179235816, + 0.005970806814730167, + 0.005291495937854052, + 0.017686009407043457, + 0.003276543226093054, + 0.00800901185721159, + 0.05289482697844505, + 0.005110742058604956, + 0.03864414989948273, + 0.007586066611111164, + 0.006738653872162104, + 0.009453858248889446, + 0.0043648709543049335, + 0.005363463424146175, + 0.015210730023682117, + 0.004919655155390501, + 0.008118954487144947, + 0.004269670695066452, + 0.0031105263624340296, + 0.006407191976904869, + 0.0025755742099136114, + 0.003571556182578206, + 0.009085082449018955, + 0.0037131421267986298, + 0.0038734942208975554, + 0.004057397600263357, + 0.0036605203058570623, + 0.005254223942756653, + 0.007901521399617195, + 0.0027920042630285025, + 0.00785901304334402, + 0.0030693616718053818, + 0.0036380449309945107, + 0.0032417969778180122, + 0.004503376316279173, + 0.0029050554148852825, + 0.004090307280421257, + 0.002832202473655343, + 0.004476292058825493, + 0.002409616718068719, + 0.002917637350037694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.1098143607378006 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a robotic arm with a biomimetic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by sharp black lines and pure white fill without shading or sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinitely stretching white space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the robotic arm ascend while twisting and unfolding section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low upward angle and steadily spiral upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371596708893776 + ], + "opposite_logprob_tokens": [ + -3.4087159633636475 + ], + "logprob_reward": -0.03371596708893776, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371596708893776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02241644077003002, + 0.022779302671551704, + 0.0179941076785326, + 0.016477707773447037, + 0.803554117679596, + 0.011843160726130009, + 0.012022624723613262, + 0.021508025005459785, + 0.008907422423362732, + 0.06249711662530899 + ], + "credit_map": [ + 0.0029485076665878296, + 0.003369779558852315, + 0.0026691581588238478, + 0.0018715834012255073, + 0.002402224577963352, + 0.0047329142689704895, + 0.0016947052208706737, + 0.002727569080889225, + 0.0038382282946258783, + 0.0013848963426426053, + 0.0055154720321297646, + 0.0012223869562149048, + 0.0011524821165949106, + 0.0055869026109576225, + 0.0026450599543750286, + 0.0014338723849505186, + 0.004634360782802105, + 0.001134807476773858, + 0.003392831888049841, + 0.001735607162117958, + 0.0012422894360497594, + 0.0030380189418792725, + 0.001176884863525629, + 0.0016393063124269247, + 0.0017903101397678256, + 0.001693354221060872, + 0.0023493622429668903, + 0.0038666639011353254, + 0.0011716883163899183, + 0.0025426531210541725, + 0.0018085851334035397, + 0.0012550895335152745, + 0.031003275886178017, + 0.008694778196513653, + 0.35033875703811646, + 0.015485582873225212, + 0.012891719117760658, + 0.3500320613384247, + 0.01260033342987299, + 0.02250758558511734, + 0.00209682690910995, + 0.0008949959883466363, + 0.001608438091352582, + 0.0006332853808999062, + 0.001168923918157816, + 0.0027107340283691883, + 0.0007523181266151369, + 0.001977638341486454, + 0.0017073039198294282, + 0.0019131054868921638, + 0.002159158932045102, + 0.0011114176595583558, + 0.001124504255130887, + 0.0017617989797145128, + 0.0013845524517819285, + 0.0008607830968685448, + 0.004950657486915588, + 0.0014613675884902477, + 0.004289484117180109, + 0.0011245672358199954, + 0.0015456555411219597, + 0.00510360486805439, + 0.0010296737309545279, + 0.0020030152518302202, + 0.0007049545529298484, + 0.0006862508016638458, + 0.0007952490705065429, + 0.0013620208483189344, + 0.0005526828463189304, + 0.002558227628469467, + 0.0005294228903949261, + 0.0017186134355142713, + 0.015063163824379444, + 0.005094415042549372, + 0.01257243286818266, + 0.005675767548382282, + 0.004831498488783836, + 0.010910870507359505, + 0.0034852579701691866, + 0.004863705951720476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 1 + ], + "selected_reward": -0.03371596708893776 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a robotic arm with a biomimetic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by sharp black lines and pure white fill without shading or sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinitely stretching white space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the robotic arm ascend while twisting and unfolding section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low upward angle and steadily spiral upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38709622621536255 + ], + "opposite_logprob_tokens": [ + -1.1370961666107178 + ], + "logprob_reward": -0.38709622621536255, + "yes_no_margin_reward": 0.7499999403953552, + "token_credit_reward": -0.38709622621536255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.013489454053342342, + 0.01695205271244049, + 0.01104889065027237, + 0.011808880604803562, + 0.8263790607452393, + 0.013786724768579006, + 0.005627820733934641, + 0.032632481306791306, + 0.003658175002783537, + 0.06461650878190994 + ], + "credit_map": [ + 0.0017058767843991518, + 0.001446008333005011, + 0.002644031774252653, + 0.0012328862212598324, + 0.00131462630815804, + 0.0029548280872404575, + 0.0009118879679590464, + 0.0012793071800842881, + 0.003638324560597539, + 0.0011303677456453443, + 0.004225826356559992, + 0.0009204484522342682, + 0.0009174742735922337, + 0.0034560589119791985, + 0.0015576689038425684, + 0.001105880830436945, + 0.002575414488092065, + 0.0006677479832433164, + 0.0017543528228998184, + 0.0010039160260930657, + 0.000671472167596221, + 0.002041121246293187, + 0.0006751847686246037, + 0.0016596791101619601, + 0.0011505603324621916, + 0.0011772796278819442, + 0.001992369070649147, + 0.00300640519708395, + 0.0007902294164523482, + 0.0018344689160585403, + 0.00103916903026402, + 0.0008183982572518289, + 0.03453855589032173, + 0.009125650860369205, + 0.35290688276290894, + 0.01573271118104458, + 0.014014811255037785, + 0.3610793948173523, + 0.01302112452685833, + 0.025959819555282593, + 0.0030334172770380974, + 0.0008660528692416847, + 0.002497717970982194, + 0.0008354337769560516, + 0.0009860985446721315, + 0.002538990695029497, + 0.0006690857699140906, + 0.0023599262349307537, + 0.0006439589196816087, + 0.0005743704969063401, + 0.0007731076912023127, + 0.0006275569903664291, + 0.0006445826729759574, + 0.0009856994729489088, + 0.0008922202978283167, + 0.000486323464429006, + 0.008163618855178356, + 0.0018582508200779557, + 0.006343381013721228, + 0.0014505116268992424, + 0.002240278059616685, + 0.008373054675757885, + 0.0012860679998993874, + 0.0029173139482736588, + 0.0007257033721543849, + 0.0002611699455883354, + 0.0004888487164862454, + 0.0005852982867509127, + 0.0001789920061128214, + 0.0006591689307242632, + 0.00020772409334313124, + 0.000551269156858325, + 0.01714266650378704, + 0.004240357782691717, + 0.014000753872096539, + 0.004734070971608162, + 0.004384415224194527, + 0.011689548380672932, + 0.00319480337202549, + 0.005229886621236801 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 7 + ], + "selected_reward": -0.38709622621536255 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a robotic arm with a biomimetic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by sharp black lines and pure white fill without shading or sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinitely stretching white space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the robotic arm ascend while twisting and unfolding section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low upward angle and steadily spiral upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003617292968556285 + ], + "opposite_logprob_tokens": [ + -5.628617286682129 + ], + "logprob_reward": -0.003617292968556285, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003617292968556285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06404169648885727, + 0.015565235167741776, + 0.03883301094174385, + 0.030975090339779854, + 0.6884486675262451, + 0.033009275794029236, + 0.015434029512107372, + 0.023580987006425858, + 0.023222506046295166, + 0.06688952445983887 + ], + "credit_map": [ + 0.007343030069023371, + 0.00822015292942524, + 0.009917140938341618, + 0.004795079585164785, + 0.005951410625129938, + 0.016854427754878998, + 0.0036592953838407993, + 0.007301156874746084, + 0.0019983085803687572, + 0.0021091154776513577, + 0.0022896381560713053, + 0.0014064846327528358, + 0.001331249251961708, + 0.0028455443680286407, + 0.0020836545154452324, + 0.0015012389048933983, + 0.006330551579594612, + 0.0024657784961163998, + 0.007398612797260284, + 0.003114106133580208, + 0.0023423624224960804, + 0.010020381771028042, + 0.0025765541940927505, + 0.004584663547575474, + 0.003962086047977209, + 0.0032872590236365795, + 0.005968377459794283, + 0.0033387006260454655, + 0.0027306212577968836, + 0.006153361406177282, + 0.0031727447640150785, + 0.0023619418498128653, + 0.03137469291687012, + 0.007961119525134563, + 0.28755801916122437, + 0.013053231872618198, + 0.01250563282519579, + 0.30024877190589905, + 0.011330942623317242, + 0.024416213855147362, + 0.0028146763797849417, + 0.0016651279293000698, + 0.008085380308330059, + 0.0012256629997864366, + 0.0025245104916393757, + 0.011116455309092999, + 0.001432946533896029, + 0.0041445158421993256, + 0.002121149329468608, + 0.002832190366461873, + 0.0019285593880340457, + 0.0018025229219347239, + 0.0013305589091032743, + 0.002175713889300823, + 0.0016155934426933527, + 0.0016277414979413152, + 0.005431859754025936, + 0.0013647308805957437, + 0.004314308520406485, + 0.0010025926167145371, + 0.0015193219296634197, + 0.006203977856785059, + 0.0010372482938691974, + 0.0027069489005953074, + 0.001340533490292728, + 0.0012628278927877545, + 0.005881267599761486, + 0.0023917006328701973, + 0.0009644498932175338, + 0.008082590065896511, + 0.000996060436591506, + 0.0023030771408230066, + 0.01636943779885769, + 0.005820001009851694, + 0.012947328388690948, + 0.005427839700132608, + 0.004240006674081087, + 0.012466670013964176, + 0.0034968513064086437, + 0.006121387705206871 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.003617292968556285 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a robotic arm with a biomimetic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by sharp black lines and pure white fill without shading or sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinitely stretching white space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the robotic arm ascend while twisting and unfolding section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low upward angle and steadily spiral upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.160335510969162 + ], + "opposite_logprob_tokens": [ + -1.9103355407714844 + ], + "logprob_reward": -0.160335510969162, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.160335510969162, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14639149606227875, + 0.03955567255616188, + 0.04869431257247925, + 0.08771500736474991, + 0.436809241771698, + 0.03908696398139, + 0.06771686673164368, + 0.0320502370595932, + 0.03596149757504463, + 0.06601870059967041 + ], + "credit_map": [ + 0.012621722184121609, + 0.018323112279176712, + 0.017593884840607643, + 0.010466794483363628, + 0.013869410380721092, + 0.03733174502849579, + 0.00986467394977808, + 0.026320168748497963, + 0.0047622425481677055, + 0.004106924403458834, + 0.00663016689941287, + 0.0037464872002601624, + 0.0031039617024362087, + 0.006418030243366957, + 0.0064394851215183735, + 0.00434837443754077, + 0.004870160017162561, + 0.0021913396194577217, + 0.005076058208942413, + 0.005233914591372013, + 0.0021523628383874893, + 0.01388382539153099, + 0.0026629501953721046, + 0.012623704969882965, + 0.00882735475897789, + 0.007792833726853132, + 0.015472084283828735, + 0.020731648430228233, + 0.0049980683252215385, + 0.015579349361360073, + 0.008662641979753971, + 0.0056510260328650475, + 0.02535099908709526, + 0.006156843155622482, + 0.1692798137664795, + 0.00805326271802187, + 0.012105628848075867, + 0.1869572550058365, + 0.007433475460857153, + 0.021471988409757614, + 0.0038639288395643234, + 0.0023878684733062983, + 0.004375975113362074, + 0.0018471034709364176, + 0.002523801987990737, + 0.010135618038475513, + 0.0019247630843892694, + 0.012027903459966183, + 0.00875160563737154, + 0.008378012105822563, + 0.010517693124711514, + 0.005958931986242533, + 0.008044279180467129, + 0.009704536758363247, + 0.009720765054225922, + 0.006641041487455368, + 0.0044598509557545185, + 0.00175067491363734, + 0.003452787408605218, + 0.0018184129148721695, + 0.0017039260128512979, + 0.009935164824128151, + 0.001620509079657495, + 0.007308911066502333, + 0.0034336731769144535, + 0.0017165145836770535, + 0.009128517471253872, + 0.0045812372118234634, + 0.0013720296556130052, + 0.011274689808487892, + 0.0016154269687831402, + 0.0028394043911248446, + 0.01686219871044159, + 0.004474612884223461, + 0.014302890747785568, + 0.004462017677724361, + 0.0049041383899748325, + 0.012187380343675613, + 0.0033309331629425287, + 0.0054945312440395355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.160335510969162 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tracking shot, beginning at a very low, upward angle, follows a robotic arm, defined entirely by sharp, clean black lines. It exists within a clean, black and white world, set against a backdrop of infinitely stretching white space. Each segment of the arm is connected by complex spherical joints and a sophisticated biomimetic structure. Its surface is devoid of any sheen or shading, dominated by pure white fill and decisive black outlines. Every sense of dimension and structure is expressed through the density, thickness, and precise perspective of the lines. As the camera steadily spirals upward, we witness the arm ascending with the graceful grace of a vine, smoothly and silently unfolding, twisting, and relocking section by section in a precise and rhythmic movement, as if a living creature is growing, reaching resolutely toward the unknown beyond the frame. The entire process, framed by a minimalist visual style, creates a unique atmosphere that blends futuristic technology, abstract dynamic beauty, and a sense of boundless exploration.", + "seed": 17991, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_3/index_9142_seed_17991_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03489210554398596, + "vlm_logprob": -0.03489210554398596, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.03489210554398596, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a robotic arm with a biomimetic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by sharp black lines and pure white fill without shading or sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinitely stretching white space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the robotic arm ascend while twisting and unfolding section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low upward angle and steadily spiral upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08479810506105423 + ], + "opposite_logprob_tokens": [ + -2.8347980976104736 + ], + "logprob_reward": -0.08479810506105423, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.08479810506105423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11373428255319595, + 0.19137443602085114, + 0.1747160702943802, + 0.03649286553263664, + 0.02206825092434883, + 0.15863236784934998, + 0.07303638011217117, + 0.03534740209579468, + 0.10733726620674133, + 0.08726069331169128 + ], + "credit_map": [ + 0.009458230808377266, + 0.007876974530518055, + 0.01150477398186922, + 0.010496078990399837, + 0.029330724850296974, + 0.013771658763289452, + 0.009324866347014904, + 0.021970974281430244, + 0.015663180500268936, + 0.010675621218979359, + 0.01172475703060627, + 0.01417876873165369, + 0.04805087298154831, + 0.05865759029984474, + 0.01637338474392891, + 0.016050254926085472, + 0.005783231928944588, + 0.003679845482110977, + 0.0034216258209198713, + 0.004486686550080776, + 0.06969358772039413, + 0.07367927581071854, + 0.005364578682929277, + 0.0086072301492095, + 0.0049958610907197, + 0.0029070063028484583, + 0.0021938669960945845, + 0.0025201020762324333, + 0.00680945161730051, + 0.009307091124355793, + 0.0027006911113858223, + 0.00505879707634449, + 0.002618583617731929, + 0.001782468636520207, + 0.0026625976897776127, + 0.0030123395845294, + 0.0027729382272809744, + 0.003245923900976777, + 0.0022356552071869373, + 0.0037377439439296722, + 0.017164118587970734, + 0.013031406328082085, + 0.023234620690345764, + 0.026422495022416115, + 0.01692919433116913, + 0.025363335385918617, + 0.02180340141057968, + 0.014683798886835575, + 0.008627341128885746, + 0.006389929447323084, + 0.010045956820249557, + 0.011899751611053944, + 0.005272767972201109, + 0.013288530521094799, + 0.009639021009206772, + 0.007873076014220715, + 0.004697644151747227, + 0.004647879861295223, + 0.003928076941519976, + 0.004012142773717642, + 0.003615347435697913, + 0.00484183756634593, + 0.004776080138981342, + 0.004828391596674919, + 0.019943274557590485, + 0.011065623722970486, + 0.01398327387869358, + 0.009074825793504715, + 0.009727496653795242, + 0.016167089343070984, + 0.011825150810182095, + 0.015550526790320873, + 0.021127974614501, + 0.008896925486624241, + 0.011332406662404537, + 0.007188384421169758, + 0.006592802703380585, + 0.009646015241742134, + 0.006731825415045023, + 0.01574436016380787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 5 + ], + "selected_reward": -0.08479810506105423 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a robotic arm with a biomimetic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by sharp black lines and pure white fill without shading or sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinitely stretching white space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the robotic arm ascend while twisting and unfolding section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low upward angle and steadily spiral upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014224446378648281 + ], + "opposite_logprob_tokens": [ + -4.264224529266357 + ], + "logprob_reward": -0.014224446378648281, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014224446378648281, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15702474117279053, + 0.09463059157133102, + 0.3386882245540619, + 0.13182461261749268, + 0.03950710594654083, + 0.02297315187752247, + 0.01139235682785511, + 0.023024236783385277, + 0.03126055747270584, + 0.1496744155883789 + ], + "credit_map": [ + 0.003693855833262205, + 0.0019332222873345017, + 0.005261908285319805, + 0.0025457858573645353, + 0.04650728031992912, + 0.06891278177499771, + 0.0037532178685069084, + 0.024416686967015266, + 0.012540408410131931, + 0.008796626701951027, + 0.010025326162576675, + 0.011873573996126652, + 0.015783993527293205, + 0.013702266849577427, + 0.008158357813954353, + 0.013750038109719753, + 0.009700886905193329, + 0.0069087170995771885, + 0.0054420772939920425, + 0.006577807944267988, + 0.13515260815620422, + 0.1474466770887375, + 0.008116143755614758, + 0.019343310967087746, + 0.0017940594116225839, + 0.0016635186038911343, + 0.001128318952396512, + 0.0012075667036697268, + 0.060653168708086014, + 0.059918079525232315, + 0.002196315908804536, + 0.0032635866664350033, + 0.0022916283924132586, + 0.0023611171636730433, + 0.0028706486336886883, + 0.004863077774643898, + 0.0044476636685431, + 0.007941460236907005, + 0.004952704533934593, + 0.009778806008398533, + 0.0021189088001847267, + 0.0012120867613703012, + 0.0012096784776076674, + 0.0027088376227766275, + 0.0018342436524108052, + 0.003831845475360751, + 0.004123900551348925, + 0.005933650303632021, + 0.000988880987279117, + 0.0006444596801884472, + 0.001047385623678565, + 0.0013253152137622237, + 0.0007704853196628392, + 0.002355380682274699, + 0.0016943704104050994, + 0.0025660782121121883, + 0.00250625959597528, + 0.004084005486220121, + 0.002289596013724804, + 0.0027733470778912306, + 0.0027655933517962694, + 0.002942584687843919, + 0.0029295498970896006, + 0.002733300905674696, + 0.006458597723394632, + 0.003333419095724821, + 0.0031362103763967752, + 0.0022481216583400965, + 0.002816217951476574, + 0.004299562890082598, + 0.004251403734087944, + 0.004717023577541113, + 0.034359950572252274, + 0.014762355014681816, + 0.016428133472800255, + 0.013833150267601013, + 0.014434068463742733, + 0.015307058580219746, + 0.013237401843070984, + 0.027312295511364937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.014224446378648281 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a robotic arm with a biomimetic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by sharp black lines and pure white fill without shading or sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinitely stretching white space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the robotic arm ascend while twisting and unfolding section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low upward angle and steadily spiral upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009783532470464706 + ], + "opposite_logprob_tokens": [ + -4.634783744812012 + ], + "logprob_reward": -0.009783532470464706, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783532470464706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6275976300239563, + 0.16940365731716156, + 0.050649262964725494, + 0.07357423007488251, + 0.029022283852100372, + 0.015352736227214336, + 0.003570862812921405, + 0.006031437776982784, + 0.010451661422848701, + 0.01434614509344101 + ], + "credit_map": [ + 0.0050114914774894714, + 0.004929948132485151, + 0.008545170538127422, + 0.004724820144474506, + 0.27826353907585144, + 0.2933136820793152, + 0.006515404209494591, + 0.026293665170669556, + 0.0017178535927087069, + 0.0016156042693182826, + 0.0016350080259144306, + 0.0018505060579627752, + 0.06939427554607391, + 0.07899988442659378, + 0.0022180425003170967, + 0.011972489766776562, + 0.0011051575420424342, + 0.0007885440136305988, + 0.0009704947588033974, + 0.001031784457154572, + 0.02062160335481167, + 0.021019820123910904, + 0.0010472211288288236, + 0.00406463909894228, + 0.0011107290629297495, + 0.0009949096711352468, + 0.0005593067035079002, + 0.0007070153951644897, + 0.03411445394158363, + 0.03274853155016899, + 0.001187753863632679, + 0.0021515334956347942, + 0.001663444098085165, + 0.0017934772185981274, + 0.0022876285947859287, + 0.0036548760253936052, + 0.003231324255466461, + 0.005772988777607679, + 0.00359417125582695, + 0.007024374790489674, + 0.0009598167962394655, + 0.0007557269418612123, + 0.0009614668088033795, + 0.0009953906992450356, + 0.0010890952544286847, + 0.0032980963587760925, + 0.0035428714472800493, + 0.0037502734921872616, + 0.00037133091245777905, + 0.0002428377774776891, + 0.0005025526043027639, + 0.0005632721586152911, + 0.00039618226583115757, + 0.00048605367192067206, + 0.0003866888291668147, + 0.000621945015154779, + 0.000821922265458852, + 0.0008123178849928081, + 0.0005486680311150849, + 0.0006756776128895581, + 0.0006643545348197222, + 0.0008319158805534244, + 0.000759065558668226, + 0.0009175162995234132, + 0.002251921221613884, + 0.0008848868892528117, + 0.00110138941090554, + 0.000761949282605201, + 0.0007489213021472096, + 0.001875321613624692, + 0.0013554919278249145, + 0.0014717797748744488, + 0.004014734644442797, + 0.0010215671500191092, + 0.0014922205591574311, + 0.000883460626937449, + 0.0008965244051069021, + 0.0022241335827857256, + 0.0009924123296514153, + 0.0028210917953401804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.009783532470464706 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a robotic arm with a biomimetic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by sharp black lines and pure white fill without shading or sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinitely stretching white space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the robotic arm ascend while twisting and unfolding section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low upward angle and steadily spiral upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036204999778419733 + ], + "opposite_logprob_tokens": [ + -5.628620624542236 + ], + "logprob_reward": -0.0036204999778419733, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036204999778419733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27510780096054077, + 0.23099924623966217, + 0.11263281106948853, + 0.15183088183403015, + 0.026004483923316002, + 0.03842132166028023, + 0.0246272049844265, + 0.04036739841103554, + 0.05827790126204491, + 0.04173099994659424 + ], + "credit_map": [ + 0.009068994782865047, + 0.008317354135215282, + 0.00904229748994112, + 0.008479933254420757, + 0.11459960043430328, + 0.09928080439567566, + 0.00894547626376152, + 0.01737334206700325, + 0.00537771126255393, + 0.0030008498579263687, + 0.004557514563202858, + 0.00444137305021286, + 0.08537887781858444, + 0.10341758280992508, + 0.004160079639405012, + 0.020665248855948448, + 0.007720475550740957, + 0.005407510790973902, + 0.0057287332601845264, + 0.006071561947464943, + 0.0397041030228138, + 0.030582387000322342, + 0.006441845558583736, + 0.010976193472743034, + 0.002952250652015209, + 0.002962455153465271, + 0.0026990738697350025, + 0.0027899364940822124, + 0.06579450517892838, + 0.06535321474075317, + 0.004129447042942047, + 0.005149987060576677, + 0.0019546656403690577, + 0.001519895507954061, + 0.0019261453999206424, + 0.0027181883342564106, + 0.0036505055613815784, + 0.0046999454498291016, + 0.0033536364790052176, + 0.006181500852108002, + 0.0035687799099832773, + 0.002520733280107379, + 0.003525984939187765, + 0.003992292098701, + 0.0034174348693341017, + 0.00825674831867218, + 0.0059198010712862015, + 0.00721954507753253, + 0.0029502231627702713, + 0.0019545855466276407, + 0.0028996116016060114, + 0.0032537681981921196, + 0.0017725899815559387, + 0.004955764394253492, + 0.0034969577100127935, + 0.0033437067177146673, + 0.005931014660745859, + 0.004473621025681496, + 0.003468747716397047, + 0.003479377133771777, + 0.0040686242282390594, + 0.007611291948705912, + 0.003881785087287426, + 0.007452933117747307, + 0.006741389166563749, + 0.005656201858073473, + 0.007471143733710051, + 0.003509008791297674, + 0.0055486466735601425, + 0.012766879983246326, + 0.008184337057173252, + 0.008400293067097664, + 0.009272989816963673, + 0.003564665326848626, + 0.004930311813950539, + 0.004528384190052748, + 0.0034189887810498476, + 0.004981796257197857, + 0.0036064342129975557, + 0.007427426520735025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.0036204999778419733 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a robotic arm with a biomimetic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by sharp black lines and pure white fill without shading or sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinitely stretching white space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the robotic arm ascend while twisting and unfolding section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low upward angle and steadily spiral upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062033943831920624 + ], + "opposite_logprob_tokens": [ + -2.8120338916778564 + ], + "logprob_reward": -0.062033943831920624, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062033943831920624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47332194447517395, + 0.06132916361093521, + 0.02796052023768425, + 0.2852274775505066, + 0.00989406369626522, + 0.01133827306330204, + 0.013623627834022045, + 0.017596231773495674, + 0.043075013905763626, + 0.05663357302546501 + ], + "credit_map": [ + 0.008437537588179111, + 0.0068195476196706295, + 0.011661499738693237, + 0.006682475097477436, + 0.21622668206691742, + 0.19443850219249725, + 0.00864698551595211, + 0.020408743992447853, + 0.005326878745108843, + 0.003037619637325406, + 0.0038192966021597385, + 0.0037254930939525366, + 0.017533862963318825, + 0.018447235226631165, + 0.0041937511414289474, + 0.0052450280636549, + 0.0021169446408748627, + 0.0014964987058192492, + 0.001401263289153576, + 0.001229019369930029, + 0.008948646485805511, + 0.00887787714600563, + 0.0012721485691145062, + 0.002618122845888138, + 0.0029231104999780655, + 0.002787003992125392, + 0.0018900742288678885, + 0.0025248725432902575, + 0.13035492599010468, + 0.13246862590312958, + 0.004914334509521723, + 0.007364536635577679, + 0.0010760263539850712, + 0.0007525196997448802, + 0.0007866078522056341, + 0.0007271552458405495, + 0.0012347446754574776, + 0.0017980942502617836, + 0.00125150999519974, + 0.0022674044594168663, + 0.0008863208931870759, + 0.0006811546045355499, + 0.0007993447943590581, + 0.00121248175855726, + 0.0012386930175125599, + 0.0025184722617268562, + 0.0017518902895972133, + 0.002249915851280093, + 0.0013039528857916594, + 0.0009651517611928284, + 0.001373262144625187, + 0.0015366568695753813, + 0.0009167765383608639, + 0.0032177383545786142, + 0.0017101949779316783, + 0.0025998943019658327, + 0.001652589300647378, + 0.0016456841258332133, + 0.0018921996233984828, + 0.0017957219388335943, + 0.00138096883893013, + 0.002901224186643958, + 0.0023823159281164408, + 0.003945526201277971, + 0.005648365244269371, + 0.005437721963971853, + 0.0035312511026859283, + 0.003431762335821986, + 0.0035513255279511213, + 0.007894403301179409, + 0.0071436637081205845, + 0.006436520721763372, + 0.015118317678570747, + 0.004495200235396624, + 0.006386586464941502, + 0.004766600206494331, + 0.00436865771189332, + 0.005438384134322405, + 0.004355330485850573, + 0.011704494245350361 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.062033943831920624 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_4/index_tensor([7036, 7036])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_4/index_tensor([7036, 7036])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c10b894aa6f8e3fd9e2141638b9e0654297661e4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_4/index_tensor([7036, 7036])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In the surreal, sensual style of Hajime Sorayama, the camera orbits a perfect lotus flower in a slow, deliberate, and perfectly smooth 360-degree orbiting panorama. The shot is a medium close-up, maintaining a consistent, eye-level angle with the subject. This magnificent flower, an unlikely fusion of biology and machinery, floats serenely on the surface of a still pool of viscous black liquid, absorbing all light to create an infinite, unreflective void. The flower's petals are perfectly rendered with airbrushed paint; their inner surface is a soft, organic pink with a subtle skin-like texture, seamlessly transitioning at incredibly sharp edges to a liquid chrome appearance, so polished that it resembles a distorted, shifting mirror. Perfectly nestled within the flower's core, amidst a cluster of polished gold, sits a tiny, perfectly spherical ball of polished metal, reflecting its surroundings in miniature. The entire scene is bathed in a soft, ethereal, diffuse light from an unseen source. As the camera rotates, these reflections glide across the chrome surface, creating long, clean, mirror-like highlights that emphasize every perfect curve and contour. This deliberate camera movement and the play of light on impossible textures produces a hypnotic, clinical, yet deeply sensory atmosphere that celebrates form and finish with otherworldly precision.", + "seed": 14938, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_4/index_7036_seed_14938_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.11936967205256224, + "vlm_logprob": -0.11936967205256224, + "vlm_yes_no_margin": 3.200000011920929, + "vlm_token_credit": -0.11936967205256224, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lotus flower that appears as a fusion of biology and machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flower float on a still pool of viscous black liquid that absorbs all light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flower's petals transition from soft pink with skin-like texture to liquid chrome with sharp edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny, perfectly spherical polished metal ball nestled within the flower's core amidst polished gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate, smooth 360-degree orbiting panorama around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010460784658789635 + ], + "opposite_logprob_tokens": [ + -4.88546085357666 + ], + "logprob_reward": -0.010460784658789635, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.010460784658789635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06179501861333847, + 0.12731601297855377, + 0.03474294766783714, + 0.04501841589808464, + 0.05913909524679184, + 0.05852796509861946, + 0.05710937827825546, + 0.11463552713394165, + 0.05653603374958038, + 0.38517963886260986 + ], + "credit_map": [ + 0.015155531466007233, + 0.005937117617577314, + 0.007894862443208694, + 0.00982910767197609, + 0.004890103358775377, + 0.007111776620149612, + 0.0061660464853048325, + 0.0048104701563715935, + 0.024549612775444984, + 0.01518789492547512, + 0.013305178843438625, + 0.014866877347230911, + 0.01554790697991848, + 0.018734781071543694, + 0.012946061789989471, + 0.012177684344351292, + 0.005689581390470266, + 0.0072515313513576984, + 0.0026881417725235224, + 0.005209680180996656, + 0.005522978492081165, + 0.0029923729598522186, + 0.0028372525703161955, + 0.0025514056906104088, + 0.009433440864086151, + 0.003512351308017969, + 0.002741847652941942, + 0.0029070714954286814, + 0.005458442028611898, + 0.010480652563273907, + 0.0023829422425478697, + 0.008101662620902061, + 0.022155387327075005, + 0.0037991469725966454, + 0.0027722809463739395, + 0.002844007220119238, + 0.001998727908357978, + 0.003307684324681759, + 0.002417653566226363, + 0.01984419859945774, + 0.013873725198209286, + 0.006693548988550901, + 0.005609550978988409, + 0.005122547969222069, + 0.007752180099487305, + 0.006526718381792307, + 0.005982034374028444, + 0.006967655383050442, + 0.012717888690531254, + 0.004909876734018326, + 0.0041211736388504505, + 0.003697957145050168, + 0.007496001664549112, + 0.0054269758984446526, + 0.003803142113611102, + 0.014936354011297226, + 0.03809456527233124, + 0.028121834620833397, + 0.006737177725881338, + 0.009452519938349724, + 0.012050896883010864, + 0.007947426289319992, + 0.005563606973737478, + 0.0066674984991550446, + 0.007706707343459129, + 0.007367524318397045, + 0.004140209406614304, + 0.0036847111769020557, + 0.015097524970769882, + 0.008296224288642406, + 0.005394876003265381, + 0.004848252981901169, + 0.09716936945915222, + 0.12731356918811798, + 0.025228336453437805, + 0.012589868158102036, + 0.05025921389460564, + 0.024219267070293427, + 0.009931000880897045, + 0.03846897929906845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 7 + ], + "selected_reward": -0.010460784658789635 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lotus flower that appears as a fusion of biology and machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flower float on a still pool of viscous black liquid that absorbs all light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flower's petals transition from soft pink with skin-like texture to liquid chrome with sharp edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny, perfectly spherical polished metal ball nestled within the flower's core amidst polished gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate, smooth 360-degree orbiting panorama around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12709026038646698 + ], + "opposite_logprob_tokens": [ + -2.1270902156829834 + ], + "logprob_reward": -0.12709026038646698, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12709026038646698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08123817294836044, + 0.06698879599571228, + 0.04012376815080643, + 0.010897346772253513, + 0.08493751287460327, + 0.198920339345932, + 0.06178377568721771, + 0.04681103676557541, + 0.024018608033657074, + 0.38428056240081787 + ], + "credit_map": [ + 0.01693355105817318, + 0.008848646655678749, + 0.009321468882262707, + 0.014075405895709991, + 0.007463826797902584, + 0.007633229251950979, + 0.009556397795677185, + 0.007405651733279228, + 0.012940507382154465, + 0.007479544263333082, + 0.008178366348147392, + 0.005709952209144831, + 0.014299237169325352, + 0.0063927751034498215, + 0.006203961092978716, + 0.005784457083791494, + 0.007304391358047724, + 0.008433621376752853, + 0.0028013817500323057, + 0.00628653634339571, + 0.005623624660074711, + 0.003162813140079379, + 0.003759288229048252, + 0.002752115484327078, + 0.0014253034023568034, + 0.0015498257707804441, + 0.001271297107450664, + 0.0012147546513006091, + 0.002006803872063756, + 0.0010447051608934999, + 0.0012176643358543515, + 0.0011669929372146726, + 0.02285255491733551, + 0.009234781377017498, + 0.007466825190931559, + 0.008480332791805267, + 0.006299077533185482, + 0.006546773947775364, + 0.005390250124037266, + 0.01866692490875721, + 0.05874386429786682, + 0.01065036840736866, + 0.017679421231150627, + 0.014522133395075798, + 0.012007753364741802, + 0.016661597415804863, + 0.011657171882689, + 0.056998033076524734, + 0.013498176820576191, + 0.0054415552876889706, + 0.006264926865696907, + 0.005322298500686884, + 0.007828036323189735, + 0.006588980555534363, + 0.00403595669195056, + 0.012803844176232815, + 0.017023734748363495, + 0.009979510679841042, + 0.0028589994180947542, + 0.0029761490877717733, + 0.003528667613863945, + 0.0037047225050628185, + 0.002357612829655409, + 0.004381641745567322, + 0.0028263814747333527, + 0.0031708229798823595, + 0.0017071717884391546, + 0.0023482944816350937, + 0.003727444913238287, + 0.002885999623686075, + 0.0035313682164996862, + 0.003821125254034996, + 0.0974607914686203, + 0.12218170613050461, + 0.025403399020433426, + 0.013636710122227669, + 0.04761916399002075, + 0.026848886162042618, + 0.01085514947772026, + 0.040274783968925476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 4 + ], + "selected_reward": -0.12709026038646698 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lotus flower that appears as a fusion of biology and machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flower float on a still pool of viscous black liquid that absorbs all light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flower's petals transition from soft pink with skin-like texture to liquid chrome with sharp edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny, perfectly spherical polished metal ball nestled within the flower's core amidst polished gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate, smooth 360-degree orbiting panorama around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009808204136788845 + ], + "opposite_logprob_tokens": [ + -4.63480806350708 + ], + "logprob_reward": -0.009808204136788845, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009808204136788845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05489100515842438, + 0.11573729664087296, + 0.0904870554804802, + 0.042595501989126205, + 0.05872712656855583, + 0.05753044784069061, + 0.016369836404919624, + 0.06569784134626389, + 0.03188736364245415, + 0.4660766124725342 + ], + "credit_map": [ + 0.013142397627234459, + 0.004437399562448263, + 0.006371806841343641, + 0.007913855835795403, + 0.004392513073980808, + 0.007476571947336197, + 0.005747655406594276, + 0.005408802069723606, + 0.023590320721268654, + 0.013500302098691463, + 0.011688769795000553, + 0.01100866124033928, + 0.019637543708086014, + 0.015521344728767872, + 0.011141874827444553, + 0.009648467414081097, + 0.017368189990520477, + 0.01918795518577099, + 0.006196904927492142, + 0.0133584039285779, + 0.013304082676768303, + 0.00748396385461092, + 0.007917710579931736, + 0.005669836886227131, + 0.010112919844686985, + 0.003118105698376894, + 0.002256763866171241, + 0.0033854893408715725, + 0.0051976353861391544, + 0.00963873416185379, + 0.0020536731462925673, + 0.0068321749567985535, + 0.02335338108241558, + 0.003829743480309844, + 0.005402368027716875, + 0.004480451811105013, + 0.0033176362048834562, + 0.004314575809985399, + 0.0027255844324827194, + 0.01130338292568922, + 0.008028612472116947, + 0.010734684765338898, + 0.0027353097684681416, + 0.0063269659876823425, + 0.012418885715305805, + 0.005764744710177183, + 0.006162053905427456, + 0.005359184462577105, + 0.004185696132481098, + 0.0015308463480323553, + 0.0014958373503759503, + 0.001220853184349835, + 0.00157044583465904, + 0.0017900815000757575, + 0.0012949766824021935, + 0.003281097626313567, + 0.023768479004502296, + 0.015794554725289345, + 0.004292024299502373, + 0.003805264364928007, + 0.0069718570448458195, + 0.004407066851854324, + 0.002185023622587323, + 0.00447356142103672, + 0.005348014645278454, + 0.004850757773965597, + 0.0018919503781944513, + 0.002099502831697464, + 0.008651877753436565, + 0.003924175165593624, + 0.0029753679409623146, + 0.0021457138936966658, + 0.11492675542831421, + 0.14421285688877106, + 0.03242690861225128, + 0.016132673248648643, + 0.05626562237739563, + 0.035571373999118805, + 0.013560788705945015, + 0.05297962203621864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 2 + ], + "selected_reward": -0.009808204136788845 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lotus flower that appears as a fusion of biology and machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flower float on a still pool of viscous black liquid that absorbs all light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flower's petals transition from soft pink with skin-like texture to liquid chrome with sharp edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny, perfectly spherical polished metal ball nestled within the flower's core amidst polished gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate, smooth 360-degree orbiting panorama around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020584875717759132 + ], + "opposite_logprob_tokens": [ + -3.89558482170105 + ], + "logprob_reward": -0.020584875717759132, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020584875717759132, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22352896630764008, + 0.06400104612112045, + 0.09002076834440231, + 0.024677759036421776, + 0.15314525365829468, + 0.043059512972831726, + 0.02998502552509308, + 0.11666828393936157, + 0.05723889544606209, + 0.19767455756664276 + ], + "credit_map": [ + 0.0563279390335083, + 0.02211754024028778, + 0.02198629640042782, + 0.036402344703674316, + 0.01804470643401146, + 0.02184184081852436, + 0.023853547871112823, + 0.02295473776757717, + 0.012952026911079884, + 0.010494126006960869, + 0.00537869380787015, + 0.008625353686511517, + 0.007493681274354458, + 0.0075815534219145775, + 0.006413915194571018, + 0.005061696749180555, + 0.01749742589890957, + 0.018413474783301353, + 0.005806487984955311, + 0.013725724071264267, + 0.013497400097548962, + 0.006898797117173672, + 0.008445155806839466, + 0.005736295133829117, + 0.004017426632344723, + 0.001990551594644785, + 0.001693588332273066, + 0.0016205799765884876, + 0.0029449781868606806, + 0.006216313224285841, + 0.0013148169964551926, + 0.004879503510892391, + 0.052764300256967545, + 0.017721733078360558, + 0.011972709558904171, + 0.011453546583652496, + 0.011238175444304943, + 0.011375166475772858, + 0.009732806123793125, + 0.026886818930506706, + 0.012164091691374779, + 0.0032018607016652822, + 0.0041963024996221066, + 0.004569300916045904, + 0.004675367381423712, + 0.003781757550314069, + 0.0037458983715623617, + 0.00672493688762188, + 0.007353322580456734, + 0.0021021843422204256, + 0.0024132428225129843, + 0.0018056336557492614, + 0.0023302717600017786, + 0.0027627390809357166, + 0.001422991044819355, + 0.009794640354812145, + 0.042411886155605316, + 0.02662482298910618, + 0.006841896567493677, + 0.009245763532817364, + 0.00982002541422844, + 0.006907064002007246, + 0.005017239600419998, + 0.009799591265618801, + 0.009693876840174198, + 0.006678505800664425, + 0.003766506677493453, + 0.003698374144732952, + 0.013561329804360867, + 0.010506486520171165, + 0.0043710158206522465, + 0.004962801001966, + 0.05263223499059677, + 0.06268896162509918, + 0.013082161545753479, + 0.007027709390968084, + 0.023762807250022888, + 0.013705478981137276, + 0.005357385613024235, + 0.019417818635702133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.020584875717759132 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lotus flower that appears as a fusion of biology and machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flower float on a still pool of viscous black liquid that absorbs all light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flower's petals transition from soft pink with skin-like texture to liquid chrome with sharp edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny, perfectly spherical polished metal ball nestled within the flower's core amidst polished gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate, smooth 360-degree orbiting panorama around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289042353630066 + ], + "opposite_logprob_tokens": [ + -1.0539042949676514 + ], + "logprob_reward": -0.4289042353630066, + "yes_no_margin_reward": 0.6250000596046448, + "token_credit_reward": -0.4289042353630066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05389738827943802, + 0.05031294375658035, + 0.017270389944314957, + 0.024711724370718002, + 0.050931598991155624, + 0.043957002460956573, + 0.008923610672354698, + 0.07345014810562134, + 0.09883011132478714, + 0.5777150988578796 + ], + "credit_map": [ + 0.016557974740862846, + 0.004510811064392328, + 0.006304929964244366, + 0.0063729193061590195, + 0.004434933885931969, + 0.005442261230200529, + 0.004515041597187519, + 0.005758519284427166, + 0.010104546323418617, + 0.005748912692070007, + 0.004954883363097906, + 0.004872685763984919, + 0.007559144403785467, + 0.0073464298620820045, + 0.005059836432337761, + 0.0046665119007229805, + 0.0032991592306643724, + 0.003507066983729601, + 0.001158517086878419, + 0.0026672594249248505, + 0.0021876406390219927, + 0.0016307801706716418, + 0.0016256433445960283, + 0.0011943253921344876, + 0.006027676165103912, + 0.0015293820761144161, + 0.002003506990149617, + 0.0014948257012292743, + 0.0027348895091563463, + 0.005857537966221571, + 0.0015587074449285865, + 0.0035052027087658644, + 0.020893212407827377, + 0.004475768655538559, + 0.0039865258149802685, + 0.0025680926628410816, + 0.0029863242525607347, + 0.003132188692688942, + 0.0020626678597182035, + 0.01082682516425848, + 0.01216132566332817, + 0.006267429795116186, + 0.003239288227632642, + 0.0033131232485175133, + 0.004557613283395767, + 0.0032990791369229555, + 0.002371145412325859, + 0.008748002350330353, + 0.001804318162612617, + 0.0008465687278658152, + 0.0009065817575901747, + 0.0008641028543934226, + 0.0013310834765434265, + 0.0010573274921625853, + 0.0008951480849646032, + 0.001218480640091002, + 0.025575684383511543, + 0.01832319237291813, + 0.004569938872009516, + 0.004892696160823107, + 0.007736500818282366, + 0.004709470085799694, + 0.0033925336319953203, + 0.004250137601047754, + 0.014542113058269024, + 0.013264077715575695, + 0.005272329319268465, + 0.00631178542971611, + 0.02781163714826107, + 0.01577441394329071, + 0.009321981109678745, + 0.006531774066388607, + 0.14066557586193085, + 0.1790749877691269, + 0.04052025452256203, + 0.019718537107110023, + 0.07022907584905624, + 0.04400105029344559, + 0.017175737768411636, + 0.06632989645004272 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.4289042353630066 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the surreal, sensual style of Hajime Sorayama, the camera orbits a perfect lotus flower in a slow, deliberate, and perfectly smooth 360-degree orbiting panorama. The shot is a medium close-up, maintaining a consistent, eye-level angle with the subject. This magnificent flower, an unlikely fusion of biology and machinery, floats serenely on the surface of a still pool of viscous black liquid, absorbing all light to create an infinite, unreflective void. The flower's petals are perfectly rendered with airbrushed paint; their inner surface is a soft, organic pink with a subtle skin-like texture, seamlessly transitioning at incredibly sharp edges to a liquid chrome appearance, so polished that it resembles a distorted, shifting mirror. Perfectly nestled within the flower's core, amidst a cluster of polished gold, sits a tiny, perfectly spherical ball of polished metal, reflecting its surroundings in miniature. The entire scene is bathed in a soft, ethereal, diffuse light from an unseen source. As the camera rotates, these reflections glide across the chrome surface, creating long, clean, mirror-like highlights that emphasize every perfect curve and contour. This deliberate camera movement and the play of light on impossible textures produces a hypnotic, clinical, yet deeply sensory atmosphere that celebrates form and finish with otherworldly precision.", + "seed": 14939, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_4/index_7036_seed_14939_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07801693473011255, + "vlm_logprob": -0.07801693473011255, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.07801693473011255, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lotus flower that appears as a fusion of biology and machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flower float on a still pool of viscous black liquid that absorbs all light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flower's petals transition from soft pink with skin-like texture to liquid chrome with sharp edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny, perfectly spherical polished metal ball nestled within the flower's core amidst polished gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate, smooth 360-degree orbiting panorama around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005817031487822533 + ], + "opposite_logprob_tokens": [ + -6.13081693649292 + ], + "logprob_reward": -0.005817031487822533, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.005817031487822533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03725700452923775, + 0.039206892251968384, + 0.03399360924959183, + 0.06936883181333542, + 0.08566906303167343, + 0.191374734044075, + 0.014163958840072155, + 0.033074405044317245, + 0.16983164846897125, + 0.32605990767478943 + ], + "credit_map": [ + 0.004450165666639805, + 0.005292977672070265, + 0.003030698047950864, + 0.005131477024406195, + 0.006097202189266682, + 0.004940359387546778, + 0.004039386752992868, + 0.004274739418178797, + 0.00485944002866745, + 0.004853463266044855, + 0.0031709889881312847, + 0.003346359124407172, + 0.006773449946194887, + 0.005400647409260273, + 0.007631433662027121, + 0.0031711100600659847, + 0.007232566364109516, + 0.0038691346999257803, + 0.0029829449485987425, + 0.003443900030106306, + 0.003791918745264411, + 0.006242883391678333, + 0.003638641443103552, + 0.0027916200924664736, + 0.011742349714040756, + 0.004880337975919247, + 0.005028223618865013, + 0.0065932669676840305, + 0.007628674153238535, + 0.016928238794207573, + 0.011135355569422245, + 0.005432388745248318, + 0.01321857888251543, + 0.009824392385780811, + 0.014400608837604523, + 0.008716246113181114, + 0.011444436386227608, + 0.01159459538757801, + 0.007647232618182898, + 0.008822969160974026, + 0.03210558742284775, + 0.019930783659219742, + 0.03869592398405075, + 0.01924150064587593, + 0.01675509288907051, + 0.021755097433924675, + 0.018577510491013527, + 0.024313222616910934, + 0.0030405051074922085, + 0.0014995768433436751, + 0.0018489977810531855, + 0.0010587053839117289, + 0.0016960317734628916, + 0.001979346852749586, + 0.00098716770298779, + 0.0020536277443170547, + 0.0044159661047160625, + 0.00442140270024538, + 0.00455175107344985, + 0.002922507468611002, + 0.004754762630909681, + 0.004495841450989246, + 0.0023443319369107485, + 0.005167840048670769, + 0.05853626877069473, + 0.012730492278933525, + 0.004050583578646183, + 0.004151975270360708, + 0.007323230616748333, + 0.07124914973974228, + 0.004733225796371698, + 0.007056732662022114, + 0.0940675362944603, + 0.031153732910752296, + 0.012715459801256657, + 0.02096981182694435, + 0.01781669445335865, + 0.11177849024534225, + 0.022766567766666412, + 0.014791588298976421 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.005817031487822533 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lotus flower that appears as a fusion of biology and machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flower float on a still pool of viscous black liquid that absorbs all light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flower's petals transition from soft pink with skin-like texture to liquid chrome with sharp edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny, perfectly spherical polished metal ball nestled within the flower's core amidst polished gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate, smooth 360-degree orbiting panorama around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905690908432007 + ], + "opposite_logprob_tokens": [ + -2.464056968688965 + ], + "logprob_reward": -0.08905690908432007, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905690908432007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04770375415682793, + 0.05094030499458313, + 0.05575947463512421, + 0.08685914427042007, + 0.07584558427333832, + 0.22206643223762512, + 0.07691299170255661, + 0.04068808630108833, + 0.10022565722465515, + 0.24299857020378113 + ], + "credit_map": [ + 0.00780814653262496, + 0.006669885013252497, + 0.004408705048263073, + 0.005652791820466518, + 0.008196917362511158, + 0.005606312770396471, + 0.005119509994983673, + 0.004241484217345715, + 0.008598427288234234, + 0.007048876956105232, + 0.005104329902678728, + 0.004446475766599178, + 0.006666277535259724, + 0.007207855116575956, + 0.007252140901982784, + 0.004615921061486006, + 0.007628679741173983, + 0.007190370466560125, + 0.005460379645228386, + 0.00667701056227088, + 0.00598999485373497, + 0.007400002796202898, + 0.007147844880819321, + 0.008265189826488495, + 0.012307843193411827, + 0.007768768351525068, + 0.022058507427573204, + 0.00851135328412056, + 0.007686743047088385, + 0.011756292544305325, + 0.007523156702518463, + 0.009246480651199818, + 0.021691294386982918, + 0.008667586371302605, + 0.005514509975910187, + 0.006607651710510254, + 0.014993660151958466, + 0.008260133676230907, + 0.005453895777463913, + 0.004656849894672632, + 0.03525546193122864, + 0.0210840106010437, + 0.050119251012802124, + 0.022745877504348755, + 0.018045971170067787, + 0.02650604024529457, + 0.01988867297768593, + 0.028421130031347275, + 0.014978949911892414, + 0.009999621659517288, + 0.007894718088209629, + 0.00812667515128851, + 0.007928727194666862, + 0.009869320318102837, + 0.0075930459424853325, + 0.010521933436393738, + 0.007092946209013462, + 0.0044700754806399345, + 0.005711494944989681, + 0.003791028866544366, + 0.00466097192838788, + 0.006662821397185326, + 0.0038020871579647064, + 0.0044966572895646095, + 0.0345994271337986, + 0.007517651654779911, + 0.0033686640672385693, + 0.002887543523684144, + 0.004796975292265415, + 0.0394216924905777, + 0.0030370247550308704, + 0.004596680402755737, + 0.07562737166881561, + 0.021720385178923607, + 0.008046775124967098, + 0.011404270306229591, + 0.01303689181804657, + 0.09183558821678162, + 0.012256154790520668, + 0.009071128442883492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.08905690908432007 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lotus flower that appears as a fusion of biology and machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flower float on a still pool of viscous black liquid that absorbs all light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flower's petals transition from soft pink with skin-like texture to liquid chrome with sharp edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny, perfectly spherical polished metal ball nestled within the flower's core amidst polished gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate, smooth 360-degree orbiting panorama around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005268499255180359 + ], + "opposite_logprob_tokens": [ + -5.255268573760986 + ], + "logprob_reward": -0.005268499255180359, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005268499255180359, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04898402467370033, + 0.04391748458147049, + 0.0326915979385376, + 0.15344737470149994, + 0.052872076630592346, + 0.08081815391778946, + 0.03265324980020523, + 0.03027629293501377, + 0.3574462831020355, + 0.16689345240592957 + ], + "credit_map": [ + 0.007770202588289976, + 0.006896509788930416, + 0.005940438713878393, + 0.005954512394964695, + 0.00505467364564538, + 0.006706110667437315, + 0.004900218918919563, + 0.005761361215263605, + 0.00563918799161911, + 0.007716938387602568, + 0.0024871707428246737, + 0.0033108771312981844, + 0.008149434812366962, + 0.0071327262558043, + 0.006373791955411434, + 0.0031073554418981075, + 0.004424354061484337, + 0.005263722967356443, + 0.003063105745241046, + 0.004255147185176611, + 0.003320597345009446, + 0.004810686223208904, + 0.0040699755772948265, + 0.0034840088337659836, + 0.03801732510328293, + 0.013072025962173939, + 0.012163263745605946, + 0.006230782717466354, + 0.01641809567809105, + 0.04315507411956787, + 0.013626694679260254, + 0.010764105245471, + 0.014179944060742855, + 0.007262231782078743, + 0.002765517681837082, + 0.003753582015633583, + 0.015247990377247334, + 0.004515760578215122, + 0.0031853842083364725, + 0.001961667789146304, + 0.013798465952277184, + 0.008700637146830559, + 0.01503762137144804, + 0.008393104188144207, + 0.007103465497493744, + 0.009062864817678928, + 0.008864757604897022, + 0.009857230819761753, + 0.004940614569932222, + 0.003772121388465166, + 0.004290474578738213, + 0.0033982915338128805, + 0.00369049864821136, + 0.0050074318423867226, + 0.0029946593567728996, + 0.0045591555535793304, + 0.004812743980437517, + 0.0035869211424142122, + 0.005328682251274586, + 0.0037021690513938665, + 0.0037393809761852026, + 0.0033140224404633045, + 0.002289495663717389, + 0.0035028778947889805, + 0.13137568533420563, + 0.023485956713557243, + 0.011113328859210014, + 0.00658927857875824, + 0.01762773096561432, + 0.1449591964483261, + 0.008078687824308872, + 0.014216414652764797, + 0.04690461978316307, + 0.014528506435453892, + 0.006581509951502085, + 0.013208732008934021, + 0.00971484649926424, + 0.05602133646607399, + 0.013026424683630466, + 0.006907470989972353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.005268499255180359 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lotus flower that appears as a fusion of biology and machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flower float on a still pool of viscous black liquid that absorbs all light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flower's petals transition from soft pink with skin-like texture to liquid chrome with sharp edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny, perfectly spherical polished metal ball nestled within the flower's core amidst polished gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate, smooth 360-degree orbiting panorama around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640049956738949 + ], + "opposite_logprob_tokens": [ + -4.758639812469482 + ], + "logprob_reward": -0.008640049956738949, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640049956738949, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061910420656204224, + 0.05547453090548515, + 0.04209059104323387, + 0.20664523541927338, + 0.05885716527700424, + 0.023768315091729164, + 0.019389182329177856, + 0.03320680931210518, + 0.1666775345802307, + 0.3319801986217499 + ], + "credit_map": [ + 0.009218438528478146, + 0.008238982409238815, + 0.00559602677822113, + 0.004535086918622255, + 0.015591307543218136, + 0.007700177375227213, + 0.00672899279743433, + 0.004301406443119049, + 0.00729351956397295, + 0.007497909478843212, + 0.003873524721711874, + 0.00449729710817337, + 0.011187626980245113, + 0.008605051785707474, + 0.008516822010278702, + 0.004002781119197607, + 0.008745469152927399, + 0.004065875895321369, + 0.003885920625180006, + 0.003582024248316884, + 0.003564497921615839, + 0.00833786278963089, + 0.004040858708322048, + 0.005868081469088793, + 0.04796380549669266, + 0.017562469467520714, + 0.01509897131472826, + 0.010991454124450684, + 0.022357232868671417, + 0.05817700922489166, + 0.016186809167265892, + 0.01830749586224556, + 0.012052956037223339, + 0.00878856796771288, + 0.00490137143060565, + 0.005049350205808878, + 0.008631875738501549, + 0.010553932748734951, + 0.004007996525615454, + 0.0048711118288338184, + 0.003678057109937072, + 0.0025929680559784174, + 0.005199871025979519, + 0.0022308933548629284, + 0.0021174054127186537, + 0.002581694396212697, + 0.0023453969042748213, + 0.0030220288317650557, + 0.00267747207544744, + 0.002192009473219514, + 0.0029469355940818787, + 0.0017576190875843167, + 0.0019711446948349476, + 0.003767878282815218, + 0.0016584176337346435, + 0.002417703391984105, + 0.006361253559589386, + 0.0024983081966638565, + 0.009440181776881218, + 0.0027059530839323997, + 0.0022938072215765715, + 0.0031323810108006, + 0.003046942176297307, + 0.0037279827520251274, + 0.059569306671619415, + 0.011567666195333004, + 0.005578944925218821, + 0.003376740263774991, + 0.00838939007371664, + 0.06698795408010483, + 0.0038780609611421824, + 0.007329463958740234, + 0.09342119842767715, + 0.0283312126994133, + 0.013985009863972664, + 0.025170037522912025, + 0.020512878894805908, + 0.10693472623825073, + 0.028623072430491447, + 0.015002061612904072 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 8 + ], + "selected_reward": -0.008640049956738949 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lotus flower that appears as a fusion of biology and machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flower float on a still pool of viscous black liquid that absorbs all light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flower's petals transition from soft pink with skin-like texture to liquid chrome with sharp edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny, perfectly spherical polished metal ball nestled within the flower's core amidst polished gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate, smooth 360-degree orbiting panorama around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28130218386650085 + ], + "opposite_logprob_tokens": [ + -1.4063022136688232 + ], + "logprob_reward": -0.28130218386650085, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28130218386650085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.012411773204803467, + 0.011244571767747402, + 0.013571185059845448, + 0.03928441181778908, + 0.020770492032170296, + 0.02960943803191185, + 0.01612524688243866, + 0.03697814792394638, + 0.4108847379684448, + 0.40911996364593506 + ], + "credit_map": [ + 0.002106679370626807, + 0.001718510058708489, + 0.001412465120665729, + 0.0015256801852956414, + 0.0016308830818161368, + 0.0015305120032280684, + 0.0016529500717297196, + 0.0008340937783941627, + 0.0014341616770252585, + 0.0018070504302158952, + 0.0007148378645069897, + 0.0009012204245664179, + 0.001729044015519321, + 0.0018450033385306597, + 0.0017618632409721613, + 0.0010513912420719862, + 0.002025901572778821, + 0.0019764539320021868, + 0.0015029330970719457, + 0.0018175923032686114, + 0.0015064750332385302, + 0.0019816027488559484, + 0.0016166481655091047, + 0.0011435779742896557, + 0.008564780466258526, + 0.004274031613022089, + 0.004213923122733831, + 0.0021174047142267227, + 0.0045622666366398335, + 0.008276627399027348, + 0.003799755359068513, + 0.003475625067949295, + 0.003710090881213546, + 0.0025855456478893757, + 0.002382969716563821, + 0.002011954551562667, + 0.002554516540840268, + 0.0033110727090388536, + 0.0014394557802006602, + 0.002774887951090932, + 0.0047522736713290215, + 0.003142992965877056, + 0.005868006497621536, + 0.0033780483063310385, + 0.0025410579983145, + 0.0031894685234874487, + 0.0030423756688833237, + 0.0036952169612050056, + 0.0021446857135742903, + 0.0014293909771367908, + 0.0023907239083200693, + 0.001935968641191721, + 0.0014866425190120935, + 0.0023705754429101944, + 0.001132969162426889, + 0.003234292147681117, + 0.005545260384678841, + 0.005252913571894169, + 0.0053054457530379295, + 0.004890631418675184, + 0.004218193236738443, + 0.004088112618774176, + 0.0038215802051126957, + 0.0038560116663575172, + 0.14742517471313477, + 0.028586370870471, + 0.01280547771602869, + 0.0072014220058918, + 0.02073039673268795, + 0.16789910197257996, + 0.009304883889853954, + 0.016931932419538498, + 0.13128936290740967, + 0.036460768431425095, + 0.012856844812631607, + 0.0158427432179451, + 0.02148778736591339, + 0.1602529138326645, + 0.016968274489045143, + 0.013961299322545528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.28130218386650085 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the surreal, sensual style of Hajime Sorayama, the camera orbits a perfect lotus flower in a slow, deliberate, and perfectly smooth 360-degree orbiting panorama. The shot is a medium close-up, maintaining a consistent, eye-level angle with the subject. This magnificent flower, an unlikely fusion of biology and machinery, floats serenely on the surface of a still pool of viscous black liquid, absorbing all light to create an infinite, unreflective void. The flower's petals are perfectly rendered with airbrushed paint; their inner surface is a soft, organic pink with a subtle skin-like texture, seamlessly transitioning at incredibly sharp edges to a liquid chrome appearance, so polished that it resembles a distorted, shifting mirror. Perfectly nestled within the flower's core, amidst a cluster of polished gold, sits a tiny, perfectly spherical ball of polished metal, reflecting its surroundings in miniature. The entire scene is bathed in a soft, ethereal, diffuse light from an unseen source. As the camera rotates, these reflections glide across the chrome surface, creating long, clean, mirror-like highlights that emphasize every perfect curve and contour. This deliberate camera movement and the play of light on impossible textures produces a hypnotic, clinical, yet deeply sensory atmosphere that celebrates form and finish with otherworldly precision.", + "seed": 14940, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_4/index_7036_seed_14940_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1694441234692931, + "vlm_logprob": -0.1694441234692931, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.1694441234692931, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lotus flower that appears as a fusion of biology and machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flower float on a still pool of viscous black liquid that absorbs all light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flower's petals transition from soft pink with skin-like texture to liquid chrome with sharp edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny, perfectly spherical polished metal ball nestled within the flower's core amidst polished gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate, smooth 360-degree orbiting panorama around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00874877255409956 + ], + "opposite_logprob_tokens": [ + -5.383749008178711 + ], + "logprob_reward": -0.00874877255409956, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00874877255409956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1328141987323761, + 0.11851738393306732, + 0.10847876220941544, + 0.17593123018741608, + 0.07546128332614899, + 0.04741963371634483, + 0.06258057802915573, + 0.09827116876840591, + 0.04335849732160568, + 0.13716734945774078 + ], + "credit_map": [ + 0.018375879153609276, + 0.017823033034801483, + 0.01714671589434147, + 0.012495300732553005, + 0.01685018464922905, + 0.014082028530538082, + 0.01502712070941925, + 0.021013913676142693, + 0.009646727703511715, + 0.008385913446545601, + 0.006240702699869871, + 0.01212538406252861, + 0.008149385452270508, + 0.036733757704496384, + 0.009253989905118942, + 0.027981506660580635, + 0.01978766918182373, + 0.01780211180448532, + 0.010596832260489464, + 0.012754927389323711, + 0.012419207952916622, + 0.011354593560099602, + 0.007712500169873238, + 0.016050904989242554, + 0.033120229840278625, + 0.02061210572719574, + 0.01657075621187687, + 0.01705850660800934, + 0.02040991187095642, + 0.022778671234846115, + 0.02101445011794567, + 0.024366596713662148, + 0.012294886633753777, + 0.007406671531498432, + 0.007565475068986416, + 0.009089757688343525, + 0.007239131722599268, + 0.008940606378018856, + 0.010447164066135883, + 0.012477578595280647, + 0.004926260095089674, + 0.004033699631690979, + 0.01344844326376915, + 0.004271776415407658, + 0.004755611065775156, + 0.005031659733504057, + 0.0047579677775502205, + 0.0061942096799612045, + 0.007718343287706375, + 0.0074583436362445354, + 0.0067652431316673756, + 0.005177101586014032, + 0.007419499568641186, + 0.007125406991690397, + 0.006850528996437788, + 0.014066102914512157, + 0.009732842445373535, + 0.014305698685348034, + 0.008555831387639046, + 0.012387135997414589, + 0.008908869698643684, + 0.0163778904825449, + 0.015038296580314636, + 0.012964592315256596, + 0.004987790714949369, + 0.006373197305947542, + 0.007406703662127256, + 0.006773435510694981, + 0.0036228476092219353, + 0.0035461916122585535, + 0.005522394087165594, + 0.0051259323954582214, + 0.010488498955965042, + 0.029637359082698822, + 0.020386576652526855, + 0.02260560914874077, + 0.012503060512244701, + 0.01299135759472847, + 0.012976515106856823, + 0.015578355640172958 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.00874877255409956 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lotus flower that appears as a fusion of biology and machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flower float on a still pool of viscous black liquid that absorbs all light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flower's petals transition from soft pink with skin-like texture to liquid chrome with sharp edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny, perfectly spherical polished metal ball nestled within the flower's core amidst polished gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate, smooth 360-degree orbiting panorama around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22563011944293976 + ], + "opposite_logprob_tokens": [ + -1.6006301641464233 + ], + "logprob_reward": -0.22563011944293976, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22563011944293976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07596093416213989, + 0.13431914150714874, + 0.07423670589923859, + 0.05262397602200508, + 0.32953888177871704, + 0.054188262671232224, + 0.049909818917512894, + 0.10347254574298859, + 0.049324315041303635, + 0.07642538845539093 + ], + "credit_map": [ + 0.01181037724018097, + 0.009560005739331245, + 0.009996357373893261, + 0.008265489712357521, + 0.012621278874576092, + 0.005390620790421963, + 0.010146121494472027, + 0.008170690387487411, + 0.007788855116814375, + 0.006337797734886408, + 0.0056813787668943405, + 0.008491148240864277, + 0.006625818088650703, + 0.05233018845319748, + 0.008531024679541588, + 0.03853295370936394, + 0.015947701409459114, + 0.010827850550413132, + 0.006550882942974567, + 0.006556421052664518, + 0.007279536221176386, + 0.008393843658268452, + 0.005402704235166311, + 0.013277772814035416, + 0.007582542020827532, + 0.00682308804243803, + 0.004787059500813484, + 0.003837693715468049, + 0.005477019585669041, + 0.00989059079438448, + 0.004915927536785603, + 0.009310061112046242, + 0.040277790278196335, + 0.009314878843724728, + 0.015262614004313946, + 0.01356149185448885, + 0.010170039720833302, + 0.10864261537790298, + 0.016122328117489815, + 0.11618717759847641, + 0.008245609700679779, + 0.005939080845564604, + 0.010561077855527401, + 0.00636068033054471, + 0.005038178991526365, + 0.0060132332146167755, + 0.006431396584957838, + 0.005599013529717922, + 0.013341174460947514, + 0.005367707461118698, + 0.005345042794942856, + 0.0044412617571651936, + 0.004073706921190023, + 0.008128834888339043, + 0.003814829047769308, + 0.005397266708314419, + 0.012849072925746441, + 0.010041596367955208, + 0.009078461676836014, + 0.007073883432894945, + 0.00820801593363285, + 0.030793026089668274, + 0.00870825257152319, + 0.016720248386263847, + 0.0069955806247889996, + 0.0031385987531393766, + 0.013431680388748646, + 0.005672633182257414, + 0.003628454403951764, + 0.005756910890340805, + 0.005072609521448612, + 0.005627852398902178, + 0.005665023345500231, + 0.018576251342892647, + 0.007042194250971079, + 0.013829069212079048, + 0.0074312882497906685, + 0.006673378869891167, + 0.009410882368683815, + 0.0077973101288080215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 7 + ], + "selected_reward": -0.22563011944293976 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lotus flower that appears as a fusion of biology and machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flower float on a still pool of viscous black liquid that absorbs all light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flower's petals transition from soft pink with skin-like texture to liquid chrome with sharp edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny, perfectly spherical polished metal ball nestled within the flower's core amidst polished gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate, smooth 360-degree orbiting panorama around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016099220141768456 + ], + "opposite_logprob_tokens": [ + -4.141099452972412 + ], + "logprob_reward": -0.016099220141768456, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016099220141768456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14329369366168976, + 0.09240125864744186, + 0.04938303679227829, + 0.06160106509923935, + 0.29420721530914307, + 0.02694636769592762, + 0.0573757030069828, + 0.11180999875068665, + 0.044748369604349136, + 0.11823329329490662 + ], + "credit_map": [ + 0.026708681136369705, + 0.015657803043723106, + 0.016873512417078018, + 0.01725887320935726, + 0.012851199135184288, + 0.01595064252614975, + 0.021324004977941513, + 0.01666896790266037, + 0.005374335218220949, + 0.004639777820557356, + 0.003036825219169259, + 0.007682353723794222, + 0.0040967161767184734, + 0.03353989124298096, + 0.0049950540997087955, + 0.029036298394203186, + 0.005372332409024239, + 0.00848478265106678, + 0.0026209652423858643, + 0.004743681289255619, + 0.004488728940486908, + 0.012842475436627865, + 0.0036746214609593153, + 0.0071554407477378845, + 0.011058741249144077, + 0.005628456361591816, + 0.005145268049091101, + 0.00481728371232748, + 0.0057575032114982605, + 0.012648037634789944, + 0.00601371843367815, + 0.010532046668231487, + 0.03366806358098984, + 0.008646118454635143, + 0.012957891449332237, + 0.014108546078205109, + 0.009130830876529217, + 0.09615322202444077, + 0.01668831706047058, + 0.10285419970750809, + 0.0036938227713108063, + 0.0033357772044837475, + 0.004570924676954746, + 0.0030977551359683275, + 0.002831327961757779, + 0.0025231477338820696, + 0.00408274307847023, + 0.0028108663391321898, + 0.0127192921936512, + 0.006772004067897797, + 0.005530616734176874, + 0.004794325679540634, + 0.007995371706783772, + 0.008449521847069263, + 0.005026662722229958, + 0.006087904330343008, + 0.01373064611107111, + 0.021864688023924828, + 0.01134557370096445, + 0.016044780611991882, + 0.011456193402409554, + 0.012489509768784046, + 0.011590019799768925, + 0.013288574293255806, + 0.005078251473605633, + 0.006761167664080858, + 0.005915066227316856, + 0.00760173425078392, + 0.00415007583796978, + 0.0037905480712652206, + 0.006072657182812691, + 0.00537886843085289, + 0.010610749013721943, + 0.020726535469293594, + 0.015263058245182037, + 0.017545342445373535, + 0.012334146536886692, + 0.014384443871676922, + 0.012080010026693344, + 0.0152890058234334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.016099220141768456 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lotus flower that appears as a fusion of biology and machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flower float on a still pool of viscous black liquid that absorbs all light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flower's petals transition from soft pink with skin-like texture to liquid chrome with sharp edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny, perfectly spherical polished metal ball nestled within the flower's core amidst polished gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate, smooth 360-degree orbiting panorama around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020584875717759132 + ], + "opposite_logprob_tokens": [ + -3.89558482170105 + ], + "logprob_reward": -0.020584875717759132, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020584875717759132, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1003030315041542, + 0.033413708209991455, + 0.07748255878686905, + 0.11141755431890488, + 0.27498888969421387, + 0.03490325063467026, + 0.05071374773979187, + 0.08859629184007645, + 0.030417846515774727, + 0.19776315987110138 + ], + "credit_map": [ + 0.011487122625112534, + 0.019463248550891876, + 0.008617584593594074, + 0.012936127372086048, + 0.012932614423334599, + 0.006839643232524395, + 0.015615911222994328, + 0.012410769239068031, + 0.0039769732393324375, + 0.004957018885761499, + 0.0027095326222479343, + 0.004724241327494383, + 0.003329005092382431, + 0.004572015721350908, + 0.002941232407465577, + 0.00620368542149663, + 0.008367527276277542, + 0.019606389105319977, + 0.004136613570153713, + 0.008562931790947914, + 0.00945281982421875, + 0.010545349679887295, + 0.005174096208065748, + 0.011636827141046524, + 0.0176810622215271, + 0.016531994566321373, + 0.011419197544455528, + 0.011653422378003597, + 0.013239342719316483, + 0.012285053730010986, + 0.01427772268652916, + 0.014329750090837479, + 0.02780924178659916, + 0.0073057785630226135, + 0.012155459262430668, + 0.012548881582915783, + 0.008005335927009583, + 0.09511064738035202, + 0.014149483293294907, + 0.09790404886007309, + 0.005668726284056902, + 0.003952100872993469, + 0.007945752702653408, + 0.003521125065162778, + 0.003321262774989009, + 0.0031928648240864277, + 0.0037790227215737104, + 0.0035223921295255423, + 0.012128722853958607, + 0.006064382381737232, + 0.00496272649616003, + 0.004138641059398651, + 0.004956081975251436, + 0.005581595469266176, + 0.005071406252682209, + 0.00781018752604723, + 0.0075671072117984295, + 0.016459524631500244, + 0.008928467519581318, + 0.009516824036836624, + 0.009643914178013802, + 0.009909959509968758, + 0.009995679371058941, + 0.016574809327721596, + 0.005058827809989452, + 0.002504863776266575, + 0.0061879465356469154, + 0.004049200098961592, + 0.002579767256975174, + 0.002818455221131444, + 0.0034296121448278427, + 0.0037891685497015715, + 0.0149266691878438, + 0.0387917086482048, + 0.018541384488344193, + 0.04358291998505592, + 0.02305746078491211, + 0.01334519125521183, + 0.02481367439031601, + 0.020704127848148346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 3 + ], + "selected_reward": -0.020584875717759132 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lotus flower that appears as a fusion of biology and machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flower float on a still pool of viscous black liquid that absorbs all light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flower's petals transition from soft pink with skin-like texture to liquid chrome with sharp edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny, perfectly spherical polished metal ball nestled within the flower's core amidst polished gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate, smooth 360-degree orbiting panorama around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761576294898987 + ], + "opposite_logprob_tokens": [ + -0.8261576294898987 + ], + "logprob_reward": -0.5761576294898987, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761576294898987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11223620176315308, + 0.14324143528938293, + 0.01860768347978592, + 0.10790342837572098, + 0.3755943477153778, + 0.025107810273766518, + 0.07575348019599915, + 0.04380720481276512, + 0.04368782043457031, + 0.054060645401477814 + ], + "credit_map": [ + 0.03022274747490883, + 0.01036216039210558, + 0.015875501558184624, + 0.012390621937811375, + 0.006826984696090221, + 0.015687325969338417, + 0.008333730511367321, + 0.012537118047475815, + 0.011648951098322868, + 0.0058741336688399315, + 0.005163167603313923, + 0.005726438481360674, + 0.005255176220089197, + 0.055288996547460556, + 0.00714012561365962, + 0.04714443162083626, + 0.0025357012636959553, + 0.0030843971762806177, + 0.0013320805737748742, + 0.0017562552820891142, + 0.0021579046733677387, + 0.0035062141250818968, + 0.0015108175575733185, + 0.0027243122458457947, + 0.018920572474598885, + 0.018110183998942375, + 0.010027924552559853, + 0.01078174076974392, + 0.012911062687635422, + 0.013175330124795437, + 0.010341420769691467, + 0.013635183684527874, + 0.0393156073987484, + 0.010840438306331635, + 0.016597434878349304, + 0.016546843573451042, + 0.010359402745962143, + 0.13018709421157837, + 0.01677028276026249, + 0.13497717678546906, + 0.004073659423738718, + 0.002875132719054818, + 0.004006101284176111, + 0.0033523610327392817, + 0.002276856452226639, + 0.00220541306771338, + 0.0026980757247656584, + 0.0036202077753841877, + 0.021797431632876396, + 0.007032714318484068, + 0.009091407991945744, + 0.00526159442961216, + 0.007650155574083328, + 0.012478862889111042, + 0.004674173891544342, + 0.007767133414745331, + 0.006470753811299801, + 0.004438824951648712, + 0.00579862343147397, + 0.005147457122802734, + 0.0038758073933422565, + 0.006594453006982803, + 0.00674542086198926, + 0.004735860973596573, + 0.008018304593861103, + 0.003331159008666873, + 0.011382272467017174, + 0.0045426771976053715, + 0.002783447038382292, + 0.004463500808924437, + 0.0038844284135848284, + 0.005282026715576649, + 0.006080914754420519, + 0.007454027887433767, + 0.00957975722849369, + 0.006414397154003382, + 0.004777596797794104, + 0.007698023226112127, + 0.0058419532142579556, + 0.006213970948010683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.5761576294898987 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the surreal, sensual style of Hajime Sorayama, the camera orbits a perfect lotus flower in a slow, deliberate, and perfectly smooth 360-degree orbiting panorama. The shot is a medium close-up, maintaining a consistent, eye-level angle with the subject. This magnificent flower, an unlikely fusion of biology and machinery, floats serenely on the surface of a still pool of viscous black liquid, absorbing all light to create an infinite, unreflective void. The flower's petals are perfectly rendered with airbrushed paint; their inner surface is a soft, organic pink with a subtle skin-like texture, seamlessly transitioning at incredibly sharp edges to a liquid chrome appearance, so polished that it resembles a distorted, shifting mirror. Perfectly nestled within the flower's core, amidst a cluster of polished gold, sits a tiny, perfectly spherical ball of polished metal, reflecting its surroundings in miniature. The entire scene is bathed in a soft, ethereal, diffuse light from an unseen source. As the camera rotates, these reflections glide across the chrome surface, creating long, clean, mirror-like highlights that emphasize every perfect curve and contour. This deliberate camera movement and the play of light on impossible textures produces a hypnotic, clinical, yet deeply sensory atmosphere that celebrates form and finish with otherworldly precision.", + "seed": 14941, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_4/index_7036_seed_14941_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05135696223005652, + "vlm_logprob": -0.05135696223005652, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.05135696223005652, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lotus flower that appears as a fusion of biology and machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flower float on a still pool of viscous black liquid that absorbs all light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flower's petals transition from soft pink with skin-like texture to liquid chrome with sharp edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny, perfectly spherical polished metal ball nestled within the flower's core amidst polished gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate, smooth 360-degree orbiting panorama around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006587338168174028 + ], + "opposite_logprob_tokens": [ + -6.631587505340576 + ], + "logprob_reward": -0.006587338168174028, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.006587338168174028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029244299978017807, + 0.018224617466330528, + 0.008646400645375252, + 0.007516537792980671, + 0.0073797800578176975, + 0.06226903945207596, + 0.016340600326657295, + 0.7320517897605896, + 0.09321736544370651, + 0.025109635666012764 + ], + "credit_map": [ + 0.0022578760981559753, + 0.0025590925943106413, + 0.001938738627359271, + 0.0025095560122281313, + 0.003595545422285795, + 0.005380692891776562, + 0.0033419299870729446, + 0.007660865318030119, + 0.0025852567050606012, + 0.0013375224079936743, + 0.0012642431538552046, + 0.0013174108462408185, + 0.001631954568438232, + 0.0037806641776114702, + 0.001201049773953855, + 0.005106511991471052, + 0.0005896209040656686, + 0.0012068698415532708, + 0.0007497224723920226, + 0.0011484575225040317, + 0.001734676188789308, + 0.0014313411666080356, + 0.0008829007856547832, + 0.0009028108906932175, + 0.000910768227186054, + 0.0011025179410353303, + 0.0006559789180755615, + 0.0007739872089587152, + 0.0009819583501666784, + 0.0011989233316853642, + 0.0006589977419935167, + 0.0012334047351032495, + 0.001184986555017531, + 0.0007839243044145405, + 0.001514045288786292, + 0.0005319431656971574, + 0.0006423087324947119, + 0.0008112973882816732, + 0.0009747991571202874, + 0.000936474185436964, + 0.006698554381728172, + 0.010381578467786312, + 0.004477447830140591, + 0.007775662932544947, + 0.012847562320530415, + 0.007090427912771702, + 0.00765446899458766, + 0.005343326833099127, + 0.0023546500597149134, + 0.002124385442584753, + 0.0011593230301514268, + 0.0027212845161557198, + 0.00239345571026206, + 0.0021006944589316845, + 0.0024091152008622885, + 0.0010776910930871964, + 0.30169230699539185, + 0.030685536563396454, + 0.3172454535961151, + 0.024186931550502777, + 0.011637477204203606, + 0.020507030189037323, + 0.012889595702290535, + 0.01320741232484579, + 0.03671583533287048, + 0.002920448547229171, + 0.03896461799740791, + 0.003674059174954891, + 0.002339489059522748, + 0.0033691104035824537, + 0.0024660159833729267, + 0.0027677856851369143, + 0.002629366470500827, + 0.003150986274704337, + 0.0029050535522401333, + 0.006258431822061539, + 0.002402474405243993, + 0.0026460005901753902, + 0.0024459047708660364, + 0.0026714149862527847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.006587338168174028 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lotus flower that appears as a fusion of biology and machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flower float on a still pool of viscous black liquid that absorbs all light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flower's petals transition from soft pink with skin-like texture to liquid chrome with sharp edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny, perfectly spherical polished metal ball nestled within the flower's core amidst polished gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate, smooth 360-degree orbiting panorama around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981220744550228 + ], + "opposite_logprob_tokens": [ + -3.5298120975494385 + ], + "logprob_reward": -0.02981220744550228, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981220744550228, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12413112819194794, + 0.02740362659096718, + 0.07070703059434891, + 0.03798753395676613, + 0.11756965517997742, + 0.14569392800331116, + 0.029157696291804314, + 0.326150119304657, + 0.03729740157723427, + 0.08390186727046967 + ], + "credit_map": [ + 0.01332398783415556, + 0.0077457912266254425, + 0.011619948782026768, + 0.006815074477344751, + 0.0087832510471344, + 0.024528007954359055, + 0.007991115562617779, + 0.043323952704668045, + 0.002453738823533058, + 0.002887914888560772, + 0.0021336290519684553, + 0.0027757862117141485, + 0.003370224265381694, + 0.005282284691929817, + 0.0021508883219212294, + 0.006349160335958004, + 0.009542413987219334, + 0.007506294641643763, + 0.006557380314916372, + 0.0061311544850468636, + 0.01665509305894375, + 0.011860385537147522, + 0.0050449613481760025, + 0.007409349083900452, + 0.005307068582624197, + 0.004754810594022274, + 0.0033977720886468887, + 0.004343759268522263, + 0.006775302812457085, + 0.0056792679242789745, + 0.003554106457158923, + 0.004175448790192604, + 0.017396794632077217, + 0.013194430619478226, + 0.02280047908425331, + 0.00836203619837761, + 0.01676705665886402, + 0.011902117170393467, + 0.012656033970415592, + 0.014490713365375996, + 0.01742486096918583, + 0.03033445216715336, + 0.013139229267835617, + 0.013496617786586285, + 0.03332724794745445, + 0.015873001888394356, + 0.01170351356267929, + 0.010395008139312267, + 0.0035763420164585114, + 0.0036301780492067337, + 0.002803168026730418, + 0.004882581066340208, + 0.004351226147264242, + 0.0027388466987758875, + 0.004804349038749933, + 0.002371006179600954, + 0.13123220205307007, + 0.01523558422923088, + 0.13876847922801971, + 0.011563392356038094, + 0.007304702885448933, + 0.008898349478840828, + 0.007158861495554447, + 0.0059885443188250065, + 0.007314628455787897, + 0.0036447399761527777, + 0.007601963821798563, + 0.004809658043086529, + 0.003089145990088582, + 0.0025315405800938606, + 0.004929074551910162, + 0.0033766485285013914, + 0.010919126681983471, + 0.010053202509880066, + 0.010977977886795998, + 0.01741945929825306, + 0.005325294565409422, + 0.009864614345133305, + 0.008969754911959171, + 0.010372431948781013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.02981220744550228 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lotus flower that appears as a fusion of biology and machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flower float on a still pool of viscous black liquid that absorbs all light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flower's petals transition from soft pink with skin-like texture to liquid chrome with sharp edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny, perfectly spherical polished metal ball nestled within the flower's core amidst polished gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate, smooth 360-degree orbiting panorama around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004654288291931152 + ], + "opposite_logprob_tokens": [ + -5.379654407501221 + ], + "logprob_reward": -0.004654288291931152, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004654288291931152, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.015085508115589619, + 0.08902250975370407, + 0.02677786909043789, + 0.005277743563055992, + 0.007662744726985693, + 0.055369384586811066, + 0.024655697867274284, + 0.7568903565406799, + 0.006411307491362095, + 0.012846875004470348 + ], + "credit_map": [ + 0.001640254631638527, + 0.0013837580336257815, + 0.0031860247254371643, + 0.0012637609615921974, + 0.0017501588445156813, + 0.0022184220142662525, + 0.001165124704129994, + 0.0024780044332146645, + 0.012253856286406517, + 0.007054741960018873, + 0.008606787770986557, + 0.005225903354585171, + 0.00824347510933876, + 0.01795222982764244, + 0.006891856901347637, + 0.022793663665652275, + 0.0036459669936448336, + 0.00229359557852149, + 0.002700303215533495, + 0.002322235843166709, + 0.006071854382753372, + 0.004406807478517294, + 0.002257011830806732, + 0.0030800956301391125, + 0.0006860035355202854, + 0.0006376227829605341, + 0.0005238964222371578, + 0.0005837641074322164, + 0.0008653675322420895, + 0.0007354866247624159, + 0.0006109181558713317, + 0.0006346847512759268, + 0.0007634664652869105, + 0.00122574414126575, + 0.0008131648064590991, + 0.0007003214559517801, + 0.0013584548141807318, + 0.0009375596418976784, + 0.0009491518721915781, + 0.0009148821118287742, + 0.006393129006028175, + 0.010603738948702812, + 0.004202225245535374, + 0.0052796402014791965, + 0.012878083623945713, + 0.006029506679624319, + 0.005277419928461313, + 0.004705644212663174, + 0.004740708973258734, + 0.002919434569776058, + 0.002395538380369544, + 0.002925821579992771, + 0.0028414842672646046, + 0.00437059486284852, + 0.0027055733371526003, + 0.0017565410817041993, + 0.31155675649642944, + 0.031484462320804596, + 0.3293025493621826, + 0.025229748338460922, + 0.01197581086307764, + 0.02061975933611393, + 0.013159344904124737, + 0.013561978004872799, + 0.001164724351838231, + 0.0007063493249006569, + 0.001080070505850017, + 0.0007695886888541281, + 0.0007681357092224061, + 0.0006578026222996414, + 0.0007755039841867983, + 0.0004891327116638422, + 0.0012412559008225799, + 0.002910566283389926, + 0.0017424023244529963, + 0.0016394315753132105, + 0.001325580757111311, + 0.0014697276055812836, + 0.001727055641822517, + 0.0007908556144684553 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 5 + ], + "selected_reward": -0.004654288291931152 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lotus flower that appears as a fusion of biology and machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flower float on a still pool of viscous black liquid that absorbs all light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flower's petals transition from soft pink with skin-like texture to liquid chrome with sharp edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny, perfectly spherical polished metal ball nestled within the flower's core amidst polished gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate, smooth 360-degree orbiting panorama around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419494766741991 + ], + "opposite_logprob_tokens": [ + -4.264194965362549 + ], + "logprob_reward": -0.01419494766741991, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419494766741991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.00803451333194971, + 0.00792138371616602, + 0.007022878620773554, + 0.0051907203160226345, + 0.008554832078516483, + 0.017173519358038902, + 0.013652851805090904, + 0.8715613484382629, + 0.013186633586883545, + 0.047701314091682434 + ], + "credit_map": [ + 0.0009302367689087987, + 0.0005923721473664045, + 0.0007676960667595267, + 0.000817584921605885, + 0.0006866038893349469, + 0.001387468189932406, + 0.00120727124158293, + 0.001645280048251152, + 0.0013422496849671006, + 0.0005258689052425325, + 0.0007675025844946504, + 0.0004479046619962901, + 0.0006323060370050371, + 0.0015812232159078121, + 0.0005439048982225358, + 0.0020804235246032476, + 0.0008725483785383403, + 0.0008945759036578238, + 0.0005845559062436223, + 0.0008848911384120584, + 0.0014178160345181823, + 0.0010517401387915015, + 0.000576995313167572, + 0.0007397555746138096, + 0.0007389530073851347, + 0.0007252712966874242, + 0.0006653415039181709, + 0.00042316626058891416, + 0.0007627664599567652, + 0.0008838094654493034, + 0.0004025126399938017, + 0.0005888998275622725, + 0.0008763883961364627, + 0.0012481915764510632, + 0.0010828226804733276, + 0.0005848178407177329, + 0.0013654676731675863, + 0.0010979643557220697, + 0.0010362423490732908, + 0.0012629374396055937, + 0.0026526707224547863, + 0.003572349203750491, + 0.0019963295198976994, + 0.0011749567929655313, + 0.0028998563066124916, + 0.002066450659185648, + 0.001384715549647808, + 0.001426190952770412, + 0.003125772811472416, + 0.0017636548727750778, + 0.00164233916439116, + 0.0010961550287902355, + 0.0012378098908811808, + 0.0028712917119264603, + 0.0008612617384642363, + 0.0010545675177127123, + 0.3588905930519104, + 0.03630928695201874, + 0.37721961736679077, + 0.029161985963582993, + 0.014053551480174065, + 0.02526378631591797, + 0.01466451957821846, + 0.01599797233939171, + 0.0030712182633578777, + 0.0013717428082600236, + 0.003007160732522607, + 0.0011644294718280435, + 0.0012646818067878485, + 0.0011878011282533407, + 0.0012023448944091797, + 0.0009172549471259117, + 0.005120028275996447, + 0.009881803765892982, + 0.007195849437266588, + 0.006519999820739031, + 0.005116330459713936, + 0.005713903345167637, + 0.004455165937542915, + 0.003698233515024185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 5 + ], + "selected_reward": -0.01419494766741991 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lotus flower that appears as a fusion of biology and machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flower float on a still pool of viscous black liquid that absorbs all light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flower's petals transition from soft pink with skin-like texture to liquid chrome with sharp edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny, perfectly spherical polished metal ball nestled within the flower's core amidst polished gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate, smooth 360-degree orbiting panorama around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20153602957725525 + ], + "opposite_logprob_tokens": [ + -1.7015360593795776 + ], + "logprob_reward": -0.20153602957725525, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20153602957725525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0357394739985466, + 0.031160593032836914, + 0.0369788222014904, + 0.014820989221334457, + 0.008738141506910324, + 0.14793652296066284, + 0.036594826728105545, + 0.6222956776618958, + 0.047334492206573486, + 0.018400514498353004 + ], + "credit_map": [ + 0.002718542004004121, + 0.0033673110883682966, + 0.0032667522318661213, + 0.0027193198911845684, + 0.003573104040697217, + 0.006138213444501162, + 0.003297580638900399, + 0.01065864972770214, + 0.003936485853046179, + 0.0023826705291867256, + 0.002797223860397935, + 0.0017723329365253448, + 0.0030243974179029465, + 0.00673838471993804, + 0.002108236774802208, + 0.008400862105190754, + 0.004791285376995802, + 0.0036183507181704044, + 0.0038482663221657276, + 0.002855474827811122, + 0.008627897128462791, + 0.006217677611857653, + 0.0026872328016906977, + 0.0043326388113200665, + 0.0013121505035087466, + 0.0019408869557082653, + 0.0013220759574323893, + 0.002402768237516284, + 0.002157198265194893, + 0.0021931196097284555, + 0.0017845256952568889, + 0.0017082634149119258, + 0.0011450093006715178, + 0.0009419103153049946, + 0.0014539744006469846, + 0.0006780580151826143, + 0.0008980392012745142, + 0.0009088999358937144, + 0.0014597696717828512, + 0.0012524800840765238, + 0.019419819116592407, + 0.03266437351703644, + 0.0144461989402771, + 0.010863278061151505, + 0.032592520117759705, + 0.01598382368683815, + 0.011275820434093475, + 0.010690677911043167, + 0.005581302102655172, + 0.004028509836643934, + 0.004140375182032585, + 0.005549644120037556, + 0.004284348804503679, + 0.004363984800875187, + 0.005048398859798908, + 0.0035982616245746613, + 0.24881167709827423, + 0.03026636689901352, + 0.2607615888118744, + 0.022308632731437683, + 0.015007859095931053, + 0.02022465132176876, + 0.013041932135820389, + 0.011872966773808002, + 0.012793113477528095, + 0.0052546183578670025, + 0.012942598201334476, + 0.002701388206332922, + 0.0036425229627639055, + 0.0034465778153389692, + 0.003571701468899846, + 0.0029819693882018328, + 0.003042713738977909, + 0.002159405965358019, + 0.0026586069725453854, + 0.002578956540673971, + 0.002001470187678933, + 0.002383551327511668, + 0.0017831525765359402, + 0.0017926573054865003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 8 + ], + "selected_reward": -0.20153602957725525 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_5/index_tensor([7086, 7086])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_5/index_tensor([7086, 7086])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e313ea7d1f8599cfa166f2ccb62e60c8fd28824f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_5/index_tensor([7086, 7086])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, shot from a slightly upward angle, initially stabilizes, focusing on an emperor penguin standing alone at the edge of a vast, ancient glacier. The low, evening sun casts long, soft golden rays, shading the penguin's signature black and white feathers into a sharp silhouette. This silhouette contrasts sharply against the deep blue water and pale sky in the background, while the glacier's surface reflects the cold light. Then, a surreal transformation occurs: the penguin's outline begins to vibrate slightly and blur. As if touched by an invisible force, its solid form slowly begins to liquefy. The black portion dissolves first, transforming into streams of a thick, inky fluid that steadily and deliberately flows down its body, like a drop of thick ink falling into clear water. As these inky substances converge at the edge of the glacier, they drip silently, and the camera slowly tilts downward, following these drops into the water below. The moment they hit the icy surface, the droplets quickly spread into ever-expanding black ripples before ultimately vanishing into thin air. The striking contrast of warm and cool tones creates a dreamlike atmosphere of melancholy and tranquility, imbued with a poetic sense of extinction.", + "seed": 16037, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_5/index_7086_seed_16037_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.038819140288978815, + "vlm_logprob": -0.038819140288978815, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.038819140288978815, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the edge of a vast, ancient glacier with deep blue water and a pale sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does an emperor penguin initially appear as a sharp black and white silhouette against the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the penguin's solid form slowly liquefy, with the black portion dissolving into thick, inky fluid that flows down its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the inky droplets hit the icy surface, spread into expanding black ripples, and then vanish into thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward to follow the inky drops into the water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.044651247560977936 + ], + "opposite_logprob_tokens": [ + -3.4196512699127197 + ], + "logprob_reward": -0.044651247560977936, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.044651247560977936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1081613302230835, + 0.02001473680138588, + 0.02927962690591812, + 0.02171880379319191, + 0.03006603941321373, + 0.023104317486286163, + 0.060346074402332306, + 0.48069050908088684, + 0.12481948733329773, + 0.1017991304397583 + ], + "credit_map": [ + 0.008797301910817623, + 0.011427325196564198, + 0.03501323238015175, + 0.015287338756024837, + 0.0062337834388017654, + 0.016046738252043724, + 0.004176812246441841, + 0.011178781278431416, + 0.0019564919639378786, + 0.0022301599383354187, + 0.005041292402893305, + 0.002553801517933607, + 0.0018286186968907714, + 0.0025358321145176888, + 0.0013341184239834547, + 0.0025344204623252153, + 0.0022517077159136534, + 0.0026878570206463337, + 0.008728792890906334, + 0.003884436096996069, + 0.0019284238805994391, + 0.003989605698734522, + 0.000988685293123126, + 0.004820115864276886, + 0.0021472962107509375, + 0.002452630316838622, + 0.006512935739010572, + 0.0033617368899285793, + 0.0005649938248097897, + 0.0028118088375777006, + 0.0010627032024785876, + 0.002804695861414075, + 0.0024063135497272015, + 0.002414934802800417, + 0.008461374789476395, + 0.003428142983466387, + 0.0011920398101210594, + 0.004020741209387779, + 0.0010836266446858644, + 0.007058863528072834, + 0.0024277237243950367, + 0.0032514361664652824, + 0.0032879242207854986, + 0.002053791657090187, + 0.0011181880254298449, + 0.0034595548640936613, + 0.0012961049797013402, + 0.006209591869264841, + 0.0052312221378088, + 0.009285900741815567, + 0.012953582219779491, + 0.007563411258161068, + 0.004560454282909632, + 0.008744139224290848, + 0.004774417262524366, + 0.00723293935880065, + 0.03260278329253197, + 0.05344270542263985, + 0.1538650244474411, + 0.057375598698854446, + 0.0055280569940805435, + 0.09339938312768936, + 0.011430053971707821, + 0.07304686307907104, + 0.022293847054243088, + 0.009986945427954197, + 0.026422064751386642, + 0.016083605587482452, + 0.006163818296045065, + 0.014655196107923985, + 0.008528984151780605, + 0.020685020834207535, + 0.005485984962433577, + 0.005132954567670822, + 0.03450078144669533, + 0.013217577710747719, + 0.004250945523381233, + 0.004391020629554987, + 0.005911801941692829, + 0.028908051550388336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.044651247560977936 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the edge of a vast, ancient glacier with deep blue water and a pale sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does an emperor penguin initially appear as a sharp black and white silhouette against the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the penguin's solid form slowly liquefy, with the black portion dissolving into thick, inky fluid that flows down its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the inky droplets hit the icy surface, spread into expanding black ripples, and then vanish into thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward to follow the inky drops into the water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002215313958004117 + ], + "opposite_logprob_tokens": [ + -6.127215385437012 + ], + "logprob_reward": -0.002215313958004117, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002215313958004117, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08599602431058884, + 0.07143722474575043, + 0.05676315352320671, + 0.07100382447242737, + 0.06795923411846161, + 0.02555977925658226, + 0.030586427077651024, + 0.3975837230682373, + 0.05913936719298363, + 0.13397128880023956 + ], + "credit_map": [ + 0.011915978975594044, + 0.01465526595711708, + 0.016418280079960823, + 0.008564239367842674, + 0.006576875690370798, + 0.01313699409365654, + 0.004841720685362816, + 0.009886665269732475, + 0.006920502986758947, + 0.006366767920553684, + 0.010264603421092033, + 0.008379601873457432, + 0.005477928556501865, + 0.0092381052672863, + 0.005967709701508284, + 0.018822001293301582, + 0.003401438007131219, + 0.00616515101864934, + 0.012298114597797394, + 0.006875579245388508, + 0.007503190543502569, + 0.008427001535892487, + 0.004821852780878544, + 0.007270826492458582, + 0.006578541360795498, + 0.007759678643196821, + 0.008514405228197575, + 0.00932648777961731, + 0.008214536122977734, + 0.009828252717852592, + 0.004256931599229574, + 0.016524987295269966, + 0.005985337309539318, + 0.005787246860563755, + 0.012580333277583122, + 0.008678341284394264, + 0.005056958645582199, + 0.009188325144350529, + 0.0038636340759694576, + 0.01681906171143055, + 0.002568267984315753, + 0.0027648916002362967, + 0.003755198558792472, + 0.00352924270555377, + 0.0036274180747568607, + 0.002831066260114312, + 0.003781962441280484, + 0.0027017320971935987, + 0.0030314174946397543, + 0.0042330860160291195, + 0.006086976733058691, + 0.004070424474775791, + 0.0022718252148479223, + 0.0045440783724188805, + 0.0027107198256999254, + 0.003637897316366434, + 0.03714734688401222, + 0.04404406622052193, + 0.0959852784872055, + 0.033138468861579895, + 0.021151958033442497, + 0.0395621657371521, + 0.024924634024500847, + 0.1016298159956932, + 0.008016240783035755, + 0.004864298272877932, + 0.012153806164860725, + 0.008822761476039886, + 0.004292631987482309, + 0.005950205028057098, + 0.004309200681746006, + 0.01073022186756134, + 0.012597364373505116, + 0.013251623138785362, + 0.019990501925349236, + 0.022041698917746544, + 0.01521119475364685, + 0.011615082621574402, + 0.02054395154118538, + 0.018719876185059547 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.002215313958004117 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the edge of a vast, ancient glacier with deep blue water and a pale sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does an emperor penguin initially appear as a sharp black and white silhouette against the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the penguin's solid form slowly liquefy, with the black portion dissolving into thick, inky fluid that flows down its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the inky droplets hit the icy surface, spread into expanding black ripples, and then vanish into thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward to follow the inky drops into the water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031986285466700792 + ], + "opposite_logprob_tokens": [ + -5.753198623657227 + ], + "logprob_reward": -0.0031986285466700792, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031986285466700792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04318387433886528, + 0.031322743743658066, + 0.018725629895925522, + 0.020967628806829453, + 0.028390677645802498, + 0.036974478513002396, + 0.024632882326841354, + 0.6256141662597656, + 0.07715819031000137, + 0.09302975982427597 + ], + "credit_map": [ + 0.004266746342182159, + 0.005137916188687086, + 0.010233228094875813, + 0.007510860450565815, + 0.0036947207991033792, + 0.006007558200508356, + 0.0026337066665291786, + 0.0036991345696151257, + 0.0035455310717225075, + 0.004554830025881529, + 0.004126688465476036, + 0.003382513765245676, + 0.0011674846755340695, + 0.004731961991637945, + 0.0015258212806656957, + 0.008287910372018814, + 0.001552260247990489, + 0.0020184130407869816, + 0.0030292219016700983, + 0.001953437691554427, + 0.0018137519946321845, + 0.0027126253116875887, + 0.001508842222392559, + 0.004137078300118446, + 0.002402222016826272, + 0.002284692833200097, + 0.004194845911115408, + 0.0029260357841849327, + 0.001176459132693708, + 0.0027360732201486826, + 0.0009712615865282714, + 0.004276039078831673, + 0.0026605352759361267, + 0.002610478550195694, + 0.004252422600984573, + 0.003766498062759638, + 0.002215202432125807, + 0.0035002711229026318, + 0.002318721730262041, + 0.007066547404974699, + 0.004145771265029907, + 0.004786709323525429, + 0.005559342913329601, + 0.0028233136981725693, + 0.002496750559657812, + 0.00500472541898489, + 0.002989261643961072, + 0.009168602526187897, + 0.0022310756612569094, + 0.00238777045160532, + 0.0050779483281075954, + 0.0033358330838382244, + 0.002250014105811715, + 0.0028666823636740446, + 0.0037996512837707996, + 0.0026839047204703093, + 0.03228480741381645, + 0.026417210698127747, + 0.25104472041130066, + 0.07695544511079788, + 0.016046511009335518, + 0.04131912812590599, + 0.02481778897345066, + 0.156728595495224, + 0.013210899196565151, + 0.007084914483129978, + 0.015146248042583466, + 0.00984729453921318, + 0.007725575938820839, + 0.007270846515893936, + 0.006800435017794371, + 0.01007197517901659, + 0.006118860561400652, + 0.008794414810836315, + 0.014549743384122849, + 0.01748662441968918, + 0.011729142628610134, + 0.007553357630968094, + 0.015070636756718159, + 0.011726981028914452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.0031986285466700792 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the edge of a vast, ancient glacier with deep blue water and a pale sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does an emperor penguin initially appear as a sharp black and white silhouette against the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the penguin's solid form slowly liquefy, with the black portion dissolving into thick, inky fluid that flows down its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the inky droplets hit the icy surface, spread into expanding black ripples, and then vanish into thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward to follow the inky drops into the water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890728309750557 + ], + "opposite_logprob_tokens": [ + -2.4640729427337646 + ], + "logprob_reward": -0.0890728309750557, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890728309750557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07577407360076904, + 0.04162658751010895, + 0.021990563720464706, + 0.035416290163993835, + 0.01893746666610241, + 0.037599094212055206, + 0.0632694810628891, + 0.3601830005645752, + 0.13963480293750763, + 0.20556862652301788 + ], + "credit_map": [ + 0.007120831869542599, + 0.008236692287027836, + 0.014290649443864822, + 0.010373112745583057, + 0.009827187284827232, + 0.011309563182294369, + 0.008071407675743103, + 0.006544632837176323, + 0.005174790509045124, + 0.006183367222547531, + 0.007746717892587185, + 0.0045174905098974705, + 0.0029046679846942425, + 0.00559282349422574, + 0.003946991171687841, + 0.0055597391910851, + 0.0027886463794857264, + 0.0034348629415035248, + 0.004641873296350241, + 0.0022897603921592236, + 0.0016597385983914137, + 0.0029887608252465725, + 0.0018328757723793387, + 0.002354046795517206, + 0.0036555789411067963, + 0.004158422350883484, + 0.005774234421551228, + 0.004456238355487585, + 0.003195921890437603, + 0.005074211861938238, + 0.0027491727378219366, + 0.006352510303258896, + 0.0022167826537042856, + 0.0028061128687113523, + 0.0036475779488682747, + 0.0020674041006714106, + 0.0023075034841895103, + 0.0026378831826150417, + 0.0010088459821417928, + 0.002245354698970914, + 0.003187554655596614, + 0.0038998599629849195, + 0.007730358280241489, + 0.003394403960555792, + 0.0014481963589787483, + 0.005276314914226532, + 0.0017454540356993675, + 0.010916952975094318, + 0.008025200106203556, + 0.006326677277684212, + 0.011622607707977295, + 0.010235845111310482, + 0.004087654873728752, + 0.006762148346751928, + 0.009564916603267193, + 0.0066444301046431065, + 0.03520640358328819, + 0.013984269462525845, + 0.15038323402404785, + 0.05402940511703491, + 0.004575460217893124, + 0.01644679717719555, + 0.012456804513931274, + 0.07310061901807785, + 0.012169034220278263, + 0.006697563920170069, + 0.04130895063281059, + 0.017462212592363358, + 0.007208286318928003, + 0.012970982119441032, + 0.007605977822095156, + 0.03421179950237274, + 0.010256391018629074, + 0.008739367127418518, + 0.07597461342811584, + 0.016905589029192924, + 0.008651068434119225, + 0.009209349751472473, + 0.010857056826353073, + 0.06497519463300705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.0890728309750557 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the edge of a vast, ancient glacier with deep blue water and a pale sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does an emperor penguin initially appear as a sharp black and white silhouette against the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the penguin's solid form slowly liquefy, with the black portion dissolving into thick, inky fluid that flows down its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the inky droplets hit the icy surface, spread into expanding black ripples, and then vanish into thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward to follow the inky drops into the water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495768040418625 + ], + "opposite_logprob_tokens": [ + -2.929957628250122 + ], + "logprob_reward": -0.05495768040418625, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495768040418625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03001958131790161, + 0.01359375286847353, + 0.012898389250040054, + 0.01598450168967247, + 0.03588554263114929, + 0.016635170206427574, + 0.04338414967060089, + 0.42791125178337097, + 0.055455271154642105, + 0.3482324779033661 + ], + "credit_map": [ + 0.0035441573709249496, + 0.005225272383540869, + 0.005287462379783392, + 0.003849360393360257, + 0.0034680324606597424, + 0.004189071245491505, + 0.0026093849446624517, + 0.001846836181357503, + 0.0010095554171130061, + 0.001583235920406878, + 0.0023611593060195446, + 0.0013987207785248756, + 0.0014936274383217096, + 0.0016719987615942955, + 0.0011730125406756997, + 0.002902440959587693, + 0.0011941323755308986, + 0.0013875090517103672, + 0.002230443060398102, + 0.0014057036023586988, + 0.0011200113222002983, + 0.0018695022445172071, + 0.001248607411980629, + 0.002442479133605957, + 0.0012336141662672162, + 0.0017487145960330963, + 0.0033944319002330303, + 0.001804147264920175, + 0.001020402298308909, + 0.0023189689964056015, + 0.0008320292108692229, + 0.0036321929655969143, + 0.0031919432803988457, + 0.003771277144551277, + 0.00590129941701889, + 0.0035301551688462496, + 0.0013563742395490408, + 0.005179818253964186, + 0.0014859576476737857, + 0.011468714103102684, + 0.0019222695846110582, + 0.002012690994888544, + 0.003280005417764187, + 0.0018761837854981422, + 0.0013763055903837085, + 0.0024169860407710075, + 0.001586439087986946, + 0.002164288889616728, + 0.004709439352154732, + 0.004555166233330965, + 0.007141024339944124, + 0.004566601477563381, + 0.005120155401527882, + 0.004600356798619032, + 0.006654756609350443, + 0.006036645267158747, + 0.025854121893644333, + 0.06366758793592453, + 0.12683743238449097, + 0.06041814014315605, + 0.008723141625523567, + 0.0879589393734932, + 0.014558159746229649, + 0.03989367559552193, + 0.004794207401573658, + 0.004304687958210707, + 0.013275153934955597, + 0.007273025810718536, + 0.004220300354063511, + 0.004131126683205366, + 0.007094815839082003, + 0.010361949913203716, + 0.01510163489729166, + 0.010691264644265175, + 0.1363035887479782, + 0.022711675614118576, + 0.01373211294412613, + 0.013857242651283741, + 0.019244180992245674, + 0.11659076064825058 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.05495768040418625 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a slightly upward angle, initially stabilizes, focusing on an emperor penguin standing alone at the edge of a vast, ancient glacier. The low, evening sun casts long, soft golden rays, shading the penguin's signature black and white feathers into a sharp silhouette. This silhouette contrasts sharply against the deep blue water and pale sky in the background, while the glacier's surface reflects the cold light. Then, a surreal transformation occurs: the penguin's outline begins to vibrate slightly and blur. As if touched by an invisible force, its solid form slowly begins to liquefy. The black portion dissolves first, transforming into streams of a thick, inky fluid that steadily and deliberately flows down its body, like a drop of thick ink falling into clear water. As these inky substances converge at the edge of the glacier, they drip silently, and the camera slowly tilts downward, following these drops into the water below. The moment they hit the icy surface, the droplets quickly spread into ever-expanding black ripples before ultimately vanishing into thin air. The striking contrast of warm and cool tones creates a dreamlike atmosphere of melancholy and tranquility, imbued with a poetic sense of extinction.", + "seed": 16038, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_5/index_7086_seed_16038_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07014823565259576, + "vlm_logprob": -0.07014823565259576, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.07014823565259576, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the edge of a vast, ancient glacier with deep blue water and a pale sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does an emperor penguin initially appear as a sharp black and white silhouette against the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the penguin's solid form slowly liquefy, with the black portion dissolving into thick, inky fluid that flows down its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the inky droplets hit the icy surface, spread into expanding black ripples, and then vanish into thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward to follow the inky drops into the water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0349021777510643 + ], + "opposite_logprob_tokens": [ + -3.6599020957946777 + ], + "logprob_reward": -0.0349021777510643, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0349021777510643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16415992379188538, + 0.08759082853794098, + 0.017685728147625923, + 0.03891756013035774, + 0.10661796480417252, + 0.03468700498342514, + 0.018977681174874306, + 0.023877589032053947, + 0.018718527629971504, + 0.48876720666885376 + ], + "credit_map": [ + 0.02217649668455124, + 0.015315264463424683, + 0.02816065214574337, + 0.014776772819459438, + 0.012188661843538284, + 0.019825365394353867, + 0.011384946294128895, + 0.04033178463578224, + 0.013618702068924904, + 0.007712024264037609, + 0.015258091501891613, + 0.009644346311688423, + 0.006554876919835806, + 0.010601361282169819, + 0.005416315980255604, + 0.01878512091934681, + 0.002692562062293291, + 0.001013979665003717, + 0.0031395137775689363, + 0.0012198881013318896, + 0.0012844956945627928, + 0.0023935313802212477, + 0.0009831479983404279, + 0.004958611447364092, + 0.007074784953147173, + 0.003177808364853263, + 0.007050178945064545, + 0.0028756572864949703, + 0.0032822280190885067, + 0.005086413584649563, + 0.0025714191142469645, + 0.007799073122441769, + 0.013634645380079746, + 0.005258107092231512, + 0.025525787845253944, + 0.007405862677842379, + 0.00445780623704195, + 0.012204470112919807, + 0.009196378290653229, + 0.028934916481375694, + 0.005702387075871229, + 0.0020114376675337553, + 0.006284436676651239, + 0.0022664142306894064, + 0.0021536084823310375, + 0.004195068031549454, + 0.0031161480583250523, + 0.008957508951425552, + 0.0018835174851119518, + 0.0012683438835665584, + 0.0027045519091188908, + 0.002033416647464037, + 0.0014916042564436793, + 0.0033642472699284554, + 0.002219847636297345, + 0.004012154880911112, + 0.003573231864720583, + 0.0018182231578975916, + 0.003111061407253146, + 0.0025709508918225765, + 0.0025223041884601116, + 0.002823204966261983, + 0.0032025128602981567, + 0.004256099462509155, + 0.0025159253273159266, + 0.001357781351543963, + 0.0023868188727647066, + 0.0022119740024209023, + 0.002040733117610216, + 0.0029929126612842083, + 0.002280744258314371, + 0.0029316390864551067, + 0.09266798943281174, + 0.032802514731884, + 0.08172522485256195, + 0.031176086515188217, + 0.041595131158828735, + 0.08437004685401917, + 0.03825856000185013, + 0.08617163449525833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.0349021777510643 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the edge of a vast, ancient glacier with deep blue water and a pale sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does an emperor penguin initially appear as a sharp black and white silhouette against the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the penguin's solid form slowly liquefy, with the black portion dissolving into thick, inky fluid that flows down its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the inky droplets hit the icy surface, spread into expanding black ripples, and then vanish into thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward to follow the inky drops into the water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009806905873119831 + ], + "opposite_logprob_tokens": [ + -4.634807109832764 + ], + "logprob_reward": -0.009806905873119831, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009806905873119831, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1460084319114685, + 0.11033686995506287, + 0.10706868022680283, + 0.05995514988899231, + 0.1000073030591011, + 0.101683109998703, + 0.04693012312054634, + 0.05168335512280464, + 0.034866128116846085, + 0.2414608597755432 + ], + "credit_map": [ + 0.019179724156856537, + 0.028513096272945404, + 0.013463684357702732, + 0.018604297190904617, + 0.016389530152082443, + 0.017564093694090843, + 0.015533887781202793, + 0.016760116443037987, + 0.01624065823853016, + 0.01615694910287857, + 0.014016651548445225, + 0.01668795384466648, + 0.012103337794542313, + 0.01128994021564722, + 0.009939299896359444, + 0.01390207838267088, + 0.01682847924530506, + 0.018721407279372215, + 0.011439958587288857, + 0.01520940288901329, + 0.013811728917062283, + 0.013179561123251915, + 0.008619830943644047, + 0.009258311241865158, + 0.01001024805009365, + 0.00930142030119896, + 0.006478455848991871, + 0.007898426614701748, + 0.0064614275470376015, + 0.0065963733941316605, + 0.006234032567590475, + 0.0069747623056173325, + 0.010321694426238537, + 0.005437348037958145, + 0.017972758039832115, + 0.006940168794244528, + 0.00613082479685545, + 0.01150747574865818, + 0.013132714666426182, + 0.028564320877194405, + 0.009839177131652832, + 0.009330007247626781, + 0.01577124185860157, + 0.011261929757893085, + 0.007117453962564468, + 0.009752601385116577, + 0.012444774620234966, + 0.026165921241044998, + 0.005104031879454851, + 0.0042953877709805965, + 0.005806187633424997, + 0.0069930823519825935, + 0.004395873751491308, + 0.004885273985564709, + 0.00696247024461627, + 0.008487817831337452, + 0.006689971312880516, + 0.006084995809942484, + 0.008010296151041985, + 0.00686135608702898, + 0.006380314473062754, + 0.007320386357605457, + 0.005157096311450005, + 0.005178937688469887, + 0.00376959890127182, + 0.005544133018702269, + 0.0036774163600057364, + 0.005343751981854439, + 0.004466430749744177, + 0.004355731885880232, + 0.0040305559523403645, + 0.0036785094998776913, + 0.04685027524828911, + 0.019560744985938072, + 0.03187723830342293, + 0.01545919943600893, + 0.02298343926668167, + 0.040026865899562836, + 0.019595222547650337, + 0.0451078824698925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.009806905873119831 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the edge of a vast, ancient glacier with deep blue water and a pale sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does an emperor penguin initially appear as a sharp black and white silhouette against the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the penguin's solid form slowly liquefy, with the black portion dissolving into thick, inky fluid that flows down its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the inky droplets hit the icy surface, spread into expanding black ripples, and then vanish into thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward to follow the inky drops into the water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004101674538105726 + ], + "opposite_logprob_tokens": [ + -5.504101753234863 + ], + "logprob_reward": -0.004101674538105726, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004101674538105726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1250264197587967, + 0.03878701105713844, + 0.03749601170420647, + 0.047889869660139084, + 0.10468603670597076, + 0.06153020262718201, + 0.1347065418958664, + 0.13393272459506989, + 0.06088852137327194, + 0.255056768655777 + ], + "credit_map": [ + 0.023953065276145935, + 0.01814177818596363, + 0.012770109809935093, + 0.014564920216798782, + 0.010168366134166718, + 0.014912926591932774, + 0.009477085433900356, + 0.021038144826889038, + 0.007135918363928795, + 0.004587635863572359, + 0.004309301730245352, + 0.004389021545648575, + 0.003519185120239854, + 0.005149646662175655, + 0.003168160794302821, + 0.006528133992105722, + 0.008811145089566708, + 0.0028877134900540113, + 0.004787353333085775, + 0.004609454423189163, + 0.003430439392104745, + 0.0046136039309203625, + 0.0028208745643496513, + 0.005535422824323177, + 0.007865558378398418, + 0.006323427893221378, + 0.005058677867054939, + 0.0066688996739685535, + 0.004246541298925877, + 0.005584760569036007, + 0.005530261900275946, + 0.006611733231693506, + 0.012779559008777142, + 0.007157699670642614, + 0.01633341796696186, + 0.007186375558376312, + 0.007324167527258396, + 0.010440478101372719, + 0.01667425036430359, + 0.0267900712788105, + 0.00690031424164772, + 0.00611296109855175, + 0.008663143962621689, + 0.006408981513231993, + 0.006002847105264664, + 0.01051405631005764, + 0.006469484884291887, + 0.01045839674770832, + 0.012695962563157082, + 0.011186293326318264, + 0.023589495569467545, + 0.014177735894918442, + 0.008599110879004002, + 0.02347075566649437, + 0.023150905966758728, + 0.017836254090070724, + 0.01853061094880104, + 0.016831140965223312, + 0.01703055202960968, + 0.018732264637947083, + 0.017909502610564232, + 0.0177545715123415, + 0.014325132593512535, + 0.01281892042607069, + 0.006207496393471956, + 0.008850459940731525, + 0.0079636350274086, + 0.010764467529952526, + 0.007143617141991854, + 0.008122359402477741, + 0.006327169947326183, + 0.005509299226105213, + 0.05397608503699303, + 0.012819458730518818, + 0.05337466299533844, + 0.022741774097085, + 0.01725625805556774, + 0.04268377646803856, + 0.022107917815446854, + 0.03009679727256298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.004101674538105726 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the edge of a vast, ancient glacier with deep blue water and a pale sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does an emperor penguin initially appear as a sharp black and white silhouette against the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the penguin's solid form slowly liquefy, with the black portion dissolving into thick, inky fluid that flows down its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the inky droplets hit the icy surface, spread into expanding black ripples, and then vanish into thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward to follow the inky drops into the water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10034536570310593 + ], + "opposite_logprob_tokens": [ + -2.3503453731536865 + ], + "logprob_reward": -0.10034536570310593, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10034536570310593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.042688410729169846, + 0.017676839604973793, + 0.015659313648939133, + 0.017148658633232117, + 0.05176244303584099, + 0.05399831384420395, + 0.10135331004858017, + 0.04836266115307808, + 0.04301220923662186, + 0.6083378195762634 + ], + "credit_map": [ + 0.009716792963445187, + 0.0031376078259199858, + 0.007098995614796877, + 0.0032921249512583017, + 0.0033497365657240152, + 0.004996748641133308, + 0.002537649357691407, + 0.00855875201523304, + 0.003472638549283147, + 0.0012774959905073047, + 0.003419633721932769, + 0.0018104477785527706, + 0.001283898949623108, + 0.0022421195171773434, + 0.0011901003308594227, + 0.0029805051162838936, + 0.002521784510463476, + 0.002715881448239088, + 0.001800383673980832, + 0.0018330310704186559, + 0.00125309347640723, + 0.0021223046351224184, + 0.0012445192551240325, + 0.0021683145314455032, + 0.0027462029829621315, + 0.0029737891163676977, + 0.0016346265329048038, + 0.002419584197923541, + 0.0018606698140501976, + 0.0018245544051751494, + 0.0018201852217316628, + 0.0018690453143790364, + 0.006698284763842821, + 0.0037791281938552856, + 0.007565282750874758, + 0.0033721071667969227, + 0.004102851264178753, + 0.005001787096261978, + 0.007898365147411823, + 0.013344638980925083, + 0.005779959261417389, + 0.005308841820806265, + 0.008375064469873905, + 0.005383452400565147, + 0.004916635807603598, + 0.006428165826946497, + 0.005940337665379047, + 0.011865860782563686, + 0.009187692776322365, + 0.008072504773736, + 0.013668105937540531, + 0.015264958143234253, + 0.007905450649559498, + 0.015560907311737537, + 0.016389071941375732, + 0.015304614789783955, + 0.007894604466855526, + 0.0055636256001889706, + 0.00621233182027936, + 0.0068155680783092976, + 0.005555703304708004, + 0.006052044220268726, + 0.005544018000364304, + 0.004724766127765179, + 0.006168708670884371, + 0.003818789729848504, + 0.008574082516133785, + 0.004413565155118704, + 0.00339702726341784, + 0.006660194136202335, + 0.005217946600168943, + 0.004761893767863512, + 0.11894097924232483, + 0.03262769803404808, + 0.11833541840314865, + 0.039633847773075104, + 0.04589761421084404, + 0.10699290782213211, + 0.04836968705058098, + 0.09753969311714172 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 5 + ], + "selected_reward": -0.10034536570310593 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the edge of a vast, ancient glacier with deep blue water and a pale sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does an emperor penguin initially appear as a sharp black and white silhouette against the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the penguin's solid form slowly liquefy, with the black portion dissolving into thick, inky fluid that flows down its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the inky droplets hit the icy surface, spread into expanding black ripples, and then vanish into thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward to follow the inky drops into the water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.201585054397583 + ], + "opposite_logprob_tokens": [ + -1.701585054397583 + ], + "logprob_reward": -0.201585054397583, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.201585054397583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028552202507853508, + 0.023216433823108673, + 0.022360021248459816, + 0.04064426198601723, + 0.0377735011279583, + 0.10725130140781403, + 0.13191349804401398, + 0.12906615436077118, + 0.07117179781198502, + 0.4080507755279541 + ], + "credit_map": [ + 0.005822521168738604, + 0.0036916094832122326, + 0.003042912110686302, + 0.002809084951877594, + 0.0032687580678611994, + 0.0037698326632380486, + 0.0027133554685860872, + 0.0034341292921453714, + 0.004182539880275726, + 0.002112304326146841, + 0.0023498868104070425, + 0.002579617314040661, + 0.003130465280264616, + 0.0029987148009240627, + 0.002504284493625164, + 0.003358622780069709, + 0.004730544053018093, + 0.002024475485086441, + 0.0023993966169655323, + 0.0029781428165733814, + 0.0017942881677299738, + 0.00283647864125669, + 0.0023121165577322245, + 0.0032845817040652037, + 0.0066996607929468155, + 0.004671012982726097, + 0.004192121792584658, + 0.004144275560975075, + 0.004178978037089109, + 0.0050737312994897366, + 0.005425842944532633, + 0.006258642300963402, + 0.0050450097769498825, + 0.0032756312284618616, + 0.006439352408051491, + 0.0029476862400770187, + 0.003655665321275592, + 0.003841013414785266, + 0.006151864305138588, + 0.0064172809943556786, + 0.010429611429572105, + 0.009157227352261543, + 0.017456768080592155, + 0.011739581823348999, + 0.00874707754701376, + 0.012173009105026722, + 0.012419815175235271, + 0.025128215551376343, + 0.012455041520297527, + 0.010984404943883419, + 0.015763701871037483, + 0.022588325664401054, + 0.009999946691095829, + 0.017605429515242577, + 0.020552493631839752, + 0.02196415513753891, + 0.018405267968773842, + 0.016261106356978416, + 0.017836831510066986, + 0.016263457015156746, + 0.013260331004858017, + 0.013752514496445656, + 0.014899482019245625, + 0.018387174233794212, + 0.008437782526016235, + 0.006751516368240118, + 0.010460914112627506, + 0.010895958170294762, + 0.009153083898127079, + 0.010427060537040234, + 0.008426583372056484, + 0.006618902552872896, + 0.08174385130405426, + 0.026468787342309952, + 0.07067300379276276, + 0.02664257027208805, + 0.0372878760099411, + 0.07192779332399368, + 0.030765987932682037, + 0.0625409483909607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.201585054397583 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a slightly upward angle, initially stabilizes, focusing on an emperor penguin standing alone at the edge of a vast, ancient glacier. The low, evening sun casts long, soft golden rays, shading the penguin's signature black and white feathers into a sharp silhouette. This silhouette contrasts sharply against the deep blue water and pale sky in the background, while the glacier's surface reflects the cold light. Then, a surreal transformation occurs: the penguin's outline begins to vibrate slightly and blur. As if touched by an invisible force, its solid form slowly begins to liquefy. The black portion dissolves first, transforming into streams of a thick, inky fluid that steadily and deliberately flows down its body, like a drop of thick ink falling into clear water. As these inky substances converge at the edge of the glacier, they drip silently, and the camera slowly tilts downward, following these drops into the water below. The moment they hit the icy surface, the droplets quickly spread into ever-expanding black ripples before ultimately vanishing into thin air. The striking contrast of warm and cool tones creates a dreamlike atmosphere of melancholy and tranquility, imbued with a poetic sense of extinction.", + "seed": 16039, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_5/index_7086_seed_16039_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07156212795525789, + "vlm_logprob": -0.07156212795525789, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.07156212795525789, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the edge of a vast, ancient glacier with deep blue water and a pale sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does an emperor penguin initially appear as a sharp black and white silhouette against the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the penguin's solid form slowly liquefy, with the black portion dissolving into thick, inky fluid that flows down its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the inky droplets hit the icy surface, spread into expanding black ripples, and then vanish into thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward to follow the inky drops into the water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015362387523055077 + ], + "opposite_logprob_tokens": [ + -4.76536226272583 + ], + "logprob_reward": -0.015362387523055077, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.015362387523055077, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09302547574043274, + 0.06163213402032852, + 0.06917835772037506, + 0.07547442615032196, + 0.03378250449895859, + 0.3061339855194092, + 0.07179509103298187, + 0.16094660758972168, + 0.06972528249025345, + 0.05830610543489456 + ], + "credit_map": [ + 0.012832192704081535, + 0.0053259264677762985, + 0.019368257373571396, + 0.005546221975237131, + 0.007294958457350731, + 0.013125129975378513, + 0.00759037584066391, + 0.021942414343357086, + 0.009763895533978939, + 0.003779509337618947, + 0.011356779374182224, + 0.0031524314545094967, + 0.004133682698011398, + 0.010030006058514118, + 0.004116029478609562, + 0.015299799852073193, + 0.01000992301851511, + 0.005188602022826672, + 0.01483309268951416, + 0.004027273505926132, + 0.005050709471106529, + 0.009471974335610867, + 0.005109521094709635, + 0.015487258322536945, + 0.01096657570451498, + 0.0044468967244029045, + 0.01728331297636032, + 0.00589790428057313, + 0.004727084655314684, + 0.00842332188040018, + 0.00629533501341939, + 0.017433995380997658, + 0.004915107507258654, + 0.0022068964317440987, + 0.006976428907364607, + 0.003369793761521578, + 0.002340837614610791, + 0.004651533905416727, + 0.0035831276327371597, + 0.005738778039813042, + 0.034533653408288956, + 0.015817083418369293, + 0.0701441690325737, + 0.019248804077506065, + 0.024347536265850067, + 0.06654825806617737, + 0.02944631688296795, + 0.046048153191804886, + 0.008103776723146439, + 0.004323017783463001, + 0.013775816187262535, + 0.008696254342794418, + 0.006002687383443117, + 0.010859484784305096, + 0.009588309563696384, + 0.01044574473053217, + 0.008687054738402367, + 0.005653313361108303, + 0.0591752752661705, + 0.008462448604404926, + 0.005941471084952354, + 0.009238749742507935, + 0.008147450163960457, + 0.05564084276556969, + 0.007384523283690214, + 0.004156006500124931, + 0.01044660247862339, + 0.006047219503670931, + 0.009119750931859016, + 0.011607074178755283, + 0.008924758993089199, + 0.012039347551763058, + 0.008836112916469574, + 0.0054109483025968075, + 0.00840869452804327, + 0.0043044728226959705, + 0.00774574326351285, + 0.00954031478613615, + 0.00694878539070487, + 0.00711103668436408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.015362387523055077 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the edge of a vast, ancient glacier with deep blue water and a pale sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does an emperor penguin initially appear as a sharp black and white silhouette against the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the penguin's solid form slowly liquefy, with the black portion dissolving into thick, inky fluid that flows down its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the inky droplets hit the icy surface, spread into expanding black ripples, and then vanish into thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward to follow the inky drops into the water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011095726862549782 + ], + "opposite_logprob_tokens": [ + -4.5110955238342285 + ], + "logprob_reward": -0.011095726862549782, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011095726862549782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07468587160110474, + 0.0349138081073761, + 0.04732512682676315, + 0.04133743420243263, + 0.03759746998548508, + 0.25717583298683167, + 0.033420395106077194, + 0.2981433570384979, + 0.0949070081114769, + 0.08049365133047104 + ], + "credit_map": [ + 0.012437919154763222, + 0.009450504556298256, + 0.01284792646765709, + 0.006757815834134817, + 0.006225306075066328, + 0.012296978384256363, + 0.005723014008253813, + 0.00894641038030386, + 0.004806618206202984, + 0.005975970532745123, + 0.0049471003003418446, + 0.0032466670963913202, + 0.003895080415531993, + 0.006036016158759594, + 0.0027143016923218966, + 0.0032920590601861477, + 0.007269593887031078, + 0.00735848443582654, + 0.007837659679353237, + 0.0043816645629704, + 0.003991672769188881, + 0.007434995844960213, + 0.004078726749867201, + 0.004972334019839764, + 0.00599267240613699, + 0.003976944833993912, + 0.007793833967298269, + 0.0037910251412540674, + 0.0029595617670565844, + 0.0059041595086455345, + 0.003376169828698039, + 0.0075430721044540405, + 0.005886645056307316, + 0.004136091098189354, + 0.00481750164180994, + 0.004407778382301331, + 0.0029105418361723423, + 0.006149036809802055, + 0.0038655754178762436, + 0.005424300208687782, + 0.028025196865200996, + 0.014118662104010582, + 0.05504760518670082, + 0.02057727612555027, + 0.020745214074850082, + 0.0516120120882988, + 0.027820412069559097, + 0.039229489862918854, + 0.002751594176515937, + 0.0030897820834070444, + 0.006732671055942774, + 0.0038745030760765076, + 0.0031237825751304626, + 0.003969462588429451, + 0.004988400265574455, + 0.004890204407274723, + 0.020044654607772827, + 0.013128438964486122, + 0.09995771199464798, + 0.015146893449127674, + 0.014586877077817917, + 0.024955444037914276, + 0.015171613544225693, + 0.09515176713466644, + 0.008716224692761898, + 0.009384437464177608, + 0.013020911253988743, + 0.012147271074354649, + 0.013090397231280804, + 0.0122545026242733, + 0.011952575296163559, + 0.014340702444314957, + 0.010588238015770912, + 0.0070704021491110325, + 0.01590709201991558, + 0.004048185423016548, + 0.011050593107938766, + 0.011952854692935944, + 0.005846455693244934, + 0.01402983721345663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 8 + ], + "selected_reward": -0.011095726862549782 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the edge of a vast, ancient glacier with deep blue water and a pale sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does an emperor penguin initially appear as a sharp black and white silhouette against the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the penguin's solid form slowly liquefy, with the black portion dissolving into thick, inky fluid that flows down its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the inky droplets hit the icy surface, spread into expanding black ripples, and then vanish into thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward to follow the inky drops into the water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864406768232584 + ], + "opposite_logprob_tokens": [ + -4.758644104003906 + ], + "logprob_reward": -0.00864406768232584, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864406768232584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.024264952167868614, + 0.011296403594315052, + 0.014653240330517292, + 0.01996820978820324, + 0.012541968375444412, + 0.20882517099380493, + 0.04685872793197632, + 0.3877914249897003, + 0.19425024092197418, + 0.07954965531826019 + ], + "credit_map": [ + 0.0029676666017621756, + 0.002500498667359352, + 0.005511515773832798, + 0.0017756590386852622, + 0.0023665905464440584, + 0.0032924243714660406, + 0.0023541306145489216, + 0.0034964655060321093, + 0.0017294666031375527, + 0.0012198355980217457, + 0.002248582663014531, + 0.0007724685128778219, + 0.0010712817311286926, + 0.0017774406587705016, + 0.0009182380745187402, + 0.0015590899856761098, + 0.002584090456366539, + 0.0015678098425269127, + 0.0025121006183326244, + 0.0011980676790699363, + 0.0014202403835952282, + 0.0022213938646018505, + 0.0011972414795309305, + 0.0019522958900779486, + 0.0019476236775517464, + 0.0023847478441894054, + 0.0035216608084738255, + 0.0028100062627345324, + 0.0015893745003268123, + 0.0023207450285553932, + 0.001805702573619783, + 0.003588349325582385, + 0.0020583653822541237, + 0.0010756280971691012, + 0.0015920901205390692, + 0.001156339538283646, + 0.0010622006375342607, + 0.00238144607283175, + 0.0012487332569435239, + 0.001967165619134903, + 0.023095425218343735, + 0.012794418260455132, + 0.039527155458927155, + 0.019417788833379745, + 0.01796702668070793, + 0.038632795214653015, + 0.027644556015729904, + 0.029746009036898613, + 0.004760876763612032, + 0.0061644637025892735, + 0.007186098489910364, + 0.005378909409046173, + 0.005332842003554106, + 0.005519821308553219, + 0.0052145072259008884, + 0.007301211357116699, + 0.02485443651676178, + 0.014091335237026215, + 0.13761572539806366, + 0.01837364211678505, + 0.015565680339932442, + 0.029116153717041016, + 0.01655460335314274, + 0.13161984086036682, + 0.013416706584393978, + 0.021313730627298355, + 0.03302547708153725, + 0.025084370747208595, + 0.035498254001140594, + 0.018396994099020958, + 0.025411156937479973, + 0.02210354618728161, + 0.008785719983279705, + 0.006393216084688902, + 0.017791910097002983, + 0.006256592459976673, + 0.006545166485011578, + 0.007866811007261276, + 0.010798974893987179, + 0.01511126197874546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 8 + ], + "selected_reward": -0.00864406768232584 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the edge of a vast, ancient glacier with deep blue water and a pale sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does an emperor penguin initially appear as a sharp black and white silhouette against the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the penguin's solid form slowly liquefy, with the black portion dissolving into thick, inky fluid that flows down its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the inky droplets hit the icy surface, spread into expanding black ripples, and then vanish into thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward to follow the inky drops into the water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14281730353832245 + ], + "opposite_logprob_tokens": [ + -2.017817258834839 + ], + "logprob_reward": -0.14281730353832245, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14281730353832245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.025862352922558784, + 0.01672995463013649, + 0.018822334706783295, + 0.018253546208143234, + 0.0335608534514904, + 0.23554503917694092, + 0.03323898836970329, + 0.4670824408531189, + 0.04817414656281471, + 0.1027304157614708 + ], + "credit_map": [ + 0.0034342396538704634, + 0.002413318958133459, + 0.005329814273864031, + 0.001998053165152669, + 0.0018956501735374331, + 0.0037278044037520885, + 0.0015594915021210909, + 0.0055039809085428715, + 0.00214012386277318, + 0.0014033224433660507, + 0.0038153380155563354, + 0.0011888279113918543, + 0.0011933285277336836, + 0.0025638295337557793, + 0.0012611417332664132, + 0.003164042020216584, + 0.002353852614760399, + 0.0015248199924826622, + 0.003555554896593094, + 0.0008883296395651996, + 0.00124099082313478, + 0.002586039714515209, + 0.0015498613938689232, + 0.005122884642332792, + 0.002212133491411805, + 0.002216892782598734, + 0.002735643880441785, + 0.0031391961965709925, + 0.0014013390755280852, + 0.0019547741394490004, + 0.001732442993670702, + 0.0028611235320568085, + 0.0056842309422791, + 0.0018684259848669171, + 0.005240166559815407, + 0.002897941041737795, + 0.002468687715008855, + 0.004648229572921991, + 0.0032898508943617344, + 0.007463320158421993, + 0.025861503556370735, + 0.01330344844609499, + 0.054700274020433426, + 0.016543937847018242, + 0.018815144896507263, + 0.0498628206551075, + 0.02370811253786087, + 0.03274979814887047, + 0.0034413349349051714, + 0.002914767013862729, + 0.006050595548003912, + 0.002712400397285819, + 0.003008026396855712, + 0.006559219677001238, + 0.0036893030628561974, + 0.0048633418045938015, + 0.029904857277870178, + 0.013668620958924294, + 0.17146839201450348, + 0.02169628068804741, + 0.015567301772534847, + 0.03147733584046364, + 0.020495640113949776, + 0.16280396282672882, + 0.005037922877818346, + 0.004267873242497444, + 0.008220819756388664, + 0.0053718662820756435, + 0.004909711889922619, + 0.00806114636361599, + 0.005235683172941208, + 0.007069121580570936, + 0.010555775836110115, + 0.01132257655262947, + 0.010398334823548794, + 0.009692348539829254, + 0.014396220445632935, + 0.01728944294154644, + 0.016470763832330704, + 0.012604949064552784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 9 + ], + "selected_reward": -0.14281730353832245 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the edge of a vast, ancient glacier with deep blue water and a pale sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does an emperor penguin initially appear as a sharp black and white silhouette against the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the penguin's solid form slowly liquefy, with the black portion dissolving into thick, inky fluid that flows down its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the inky droplets hit the icy surface, spread into expanding black ripples, and then vanish into thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward to follow the inky drops into the water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17989115417003632 + ], + "opposite_logprob_tokens": [ + -1.8048911094665527 + ], + "logprob_reward": -0.17989115417003632, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17989115417003632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02261187881231308, + 0.017387988045811653, + 0.039299990981817245, + 0.05069905146956444, + 0.033173978328704834, + 0.2764909267425537, + 0.04940962418913841, + 0.3231038451194763, + 0.14645712077617645, + 0.04136558249592781 + ], + "credit_map": [ + 0.003303809091448784, + 0.0017192303203046322, + 0.004540205933153629, + 0.001102012232877314, + 0.0015310710296034813, + 0.0032507190480828285, + 0.001762118423357606, + 0.005402715876698494, + 0.0018718817736953497, + 0.0013782661408185959, + 0.0033708091359585524, + 0.0008846058044582605, + 0.0010773608228191733, + 0.002396939555183053, + 0.0013823582557961345, + 0.005025768652558327, + 0.005241990555077791, + 0.0019580426160246134, + 0.009076935239136219, + 0.0026588488835841417, + 0.002064760308712721, + 0.005632891319692135, + 0.002599808620288968, + 0.010066716931760311, + 0.006003283429890871, + 0.0036726428661495447, + 0.011211407370865345, + 0.004753700457513332, + 0.003182073589414358, + 0.005564313847571611, + 0.0042670294642448425, + 0.012044605799019337, + 0.003214814467355609, + 0.0038407007232308388, + 0.005174572113901377, + 0.004592721350491047, + 0.0033494997769594193, + 0.004977941978722811, + 0.0033224010840058327, + 0.004701331723481417, + 0.029352791607379913, + 0.016481982544064522, + 0.058127161115407944, + 0.022420067340135574, + 0.021399879828095436, + 0.05576053634285927, + 0.02941957861185074, + 0.04352894797921181, + 0.0051306672394275665, + 0.003223360050469637, + 0.010926301591098309, + 0.005438174586743116, + 0.004310018382966518, + 0.006150423549115658, + 0.007611531764268875, + 0.006619148887693882, + 0.02142512798309326, + 0.011234378442168236, + 0.11516907811164856, + 0.015689793974161148, + 0.012090894393622875, + 0.02445400133728981, + 0.013346510007977486, + 0.10969411581754684, + 0.016606172546744347, + 0.009657447226345539, + 0.021591944620013237, + 0.012374959886074066, + 0.016732683405280113, + 0.028464116156101227, + 0.016657693311572075, + 0.02437211573123932, + 0.003833823138847947, + 0.00373640563338995, + 0.0048508611507713795, + 0.0034858463332057, + 0.004937428515404463, + 0.005923200864344835, + 0.008679313585162163, + 0.0059187039732933044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 8 + ], + "selected_reward": -0.17989115417003632 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a slightly upward angle, initially stabilizes, focusing on an emperor penguin standing alone at the edge of a vast, ancient glacier. The low, evening sun casts long, soft golden rays, shading the penguin's signature black and white feathers into a sharp silhouette. This silhouette contrasts sharply against the deep blue water and pale sky in the background, while the glacier's surface reflects the cold light. Then, a surreal transformation occurs: the penguin's outline begins to vibrate slightly and blur. As if touched by an invisible force, its solid form slowly begins to liquefy. The black portion dissolves first, transforming into streams of a thick, inky fluid that steadily and deliberately flows down its body, like a drop of thick ink falling into clear water. As these inky substances converge at the edge of the glacier, they drip silently, and the camera slowly tilts downward, following these drops into the water below. The moment they hit the icy surface, the droplets quickly spread into ever-expanding black ripples before ultimately vanishing into thin air. The striking contrast of warm and cool tones creates a dreamlike atmosphere of melancholy and tranquility, imbued with a poetic sense of extinction.", + "seed": 16040, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_5/index_7086_seed_16040_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.045716024655848744, + "vlm_logprob": -0.045716024655848744, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.045716024655848744, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the edge of a vast, ancient glacier with deep blue water and a pale sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does an emperor penguin initially appear as a sharp black and white silhouette against the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the penguin's solid form slowly liquefy, with the black portion dissolving into thick, inky fluid that flows down its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the inky droplets hit the icy surface, spread into expanding black ripples, and then vanish into thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward to follow the inky drops into the water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07420476526021957 + ], + "opposite_logprob_tokens": [ + -2.824204683303833 + ], + "logprob_reward": -0.07420476526021957, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07420476526021957, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04420202970504761, + 0.08202952146530151, + 0.06947162747383118, + 0.606736421585083, + 0.013173806481063366, + 0.02435656078159809, + 0.012424714863300323, + 0.0346563421189785, + 0.015742428600788116, + 0.09720660746097565 + ], + "credit_map": [ + 0.00710971700027585, + 0.0043945228680968285, + 0.006604149006307125, + 0.0035232577938586473, + 0.0033883622381836176, + 0.006819551344960928, + 0.005044788122177124, + 0.007317685522139072, + 0.020834995433688164, + 0.006102112587541342, + 0.006158746313303709, + 0.006058730650693178, + 0.007697982247918844, + 0.020041573792696, + 0.007265022024512291, + 0.007870358414947987, + 0.024618981406092644, + 0.0035929062869399786, + 0.002862068824470043, + 0.0019765873439610004, + 0.00504003930836916, + 0.02480735071003437, + 0.0029726396314799786, + 0.0036010490730404854, + 0.25544440746307373, + 0.024214763194322586, + 0.012365895323455334, + 0.008585307747125626, + 0.01329432800412178, + 0.2638918161392212, + 0.013318035751581192, + 0.015621859580278397, + 0.002012170385569334, + 0.001013853820040822, + 0.002291254000738263, + 0.0010559253860265017, + 0.0011114001972600818, + 0.0018844606820493937, + 0.0013458856847137213, + 0.002458855276927352, + 0.004136414732784033, + 0.0017613532254472375, + 0.005054753739386797, + 0.0016560893272981048, + 0.0015055577969178557, + 0.0038568649906665087, + 0.0013246731832623482, + 0.005060853436589241, + 0.001531129819341004, + 0.0007341491873376071, + 0.0013354562688618898, + 0.0020895744673907757, + 0.0011548656038939953, + 0.002271774224936962, + 0.001842984464019537, + 0.0014647803036496043, + 0.004320619627833366, + 0.0019569622818380594, + 0.006755943410098553, + 0.003971587400883436, + 0.00371459755115211, + 0.005157223902642727, + 0.002753068692982197, + 0.006026336457580328, + 0.0020180849824100733, + 0.001260465127415955, + 0.003164710011333227, + 0.0017499111127108335, + 0.0012604794465005398, + 0.0016102000372484326, + 0.0022149626165628433, + 0.0024636154994368553, + 0.00639092642813921, + 0.002558283507823944, + 0.03561846539378166, + 0.0063695660792291164, + 0.0025583994574844837, + 0.00755084864795208, + 0.008063784800469875, + 0.028096331283450127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 1 + ], + "selected_reward": -0.07420476526021957 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the edge of a vast, ancient glacier with deep blue water and a pale sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does an emperor penguin initially appear as a sharp black and white silhouette against the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the penguin's solid form slowly liquefy, with the black portion dissolving into thick, inky fluid that flows down its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the inky droplets hit the icy surface, spread into expanding black ripples, and then vanish into thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward to follow the inky drops into the water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008668295107781887 + ], + "opposite_logprob_tokens": [ + -4.758668422698975 + ], + "logprob_reward": -0.008668295107781887, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008668295107781887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08027570694684982, + 0.059853192418813705, + 0.05758845806121826, + 0.6163442730903625, + 0.023700730875134468, + 0.017671246081590652, + 0.010172918438911438, + 0.02711968496441841, + 0.06985097378492355, + 0.03742283210158348 + ], + "credit_map": [ + 0.021779701113700867, + 0.004777727648615837, + 0.0077428934164345264, + 0.004465220496058464, + 0.0066871605813503265, + 0.023447193205356598, + 0.004458876326680183, + 0.006916929967701435, + 0.014542641118168831, + 0.003390373196452856, + 0.005384001415222883, + 0.0043410807847976685, + 0.005091304890811443, + 0.01553089264780283, + 0.004177380353212357, + 0.007395518012344837, + 0.008126565255224705, + 0.006031295750290155, + 0.004413546994328499, + 0.00498106237500906, + 0.01149707194417715, + 0.009934873320162296, + 0.007164811249822378, + 0.005439231172204018, + 0.25626152753829956, + 0.024570006877183914, + 0.015389924868941307, + 0.00892108678817749, + 0.013495968654751778, + 0.2663823366165161, + 0.015023349784314632, + 0.016300082206726074, + 0.003363483352586627, + 0.0021102966275066137, + 0.0018754932098090649, + 0.0025429679080843925, + 0.0038149163592606783, + 0.0040027606301009655, + 0.0035465664695948362, + 0.0024442486464977264, + 0.0019983872771263123, + 0.001498165656812489, + 0.0033781270030885935, + 0.0014611174119636416, + 0.0018678962951526046, + 0.0026287974324077368, + 0.0013862005434930325, + 0.0034525550436228514, + 0.0012202290818095207, + 0.0010462640784680843, + 0.001439662417396903, + 0.0014140121638774872, + 0.0007970365113578737, + 0.0012228518025949597, + 0.0012384423753246665, + 0.001794420531950891, + 0.0029969727620482445, + 0.0010295540560036898, + 0.0028034525457769632, + 0.0015580628532916307, + 0.0019446000223979354, + 0.006898979190737009, + 0.0023988308385014534, + 0.007489233277738094, + 0.005464827176183462, + 0.0020853911992162466, + 0.0067377714440226555, + 0.005019267555326223, + 0.0022691728081554174, + 0.015852266922593117, + 0.007461425848305225, + 0.024960851296782494, + 0.006050312891602516, + 0.002528748707845807, + 0.008659697137773037, + 0.004065377172082663, + 0.001739611616358161, + 0.003839921671897173, + 0.0035765115171670914, + 0.006962653715163469 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.008668295107781887 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the edge of a vast, ancient glacier with deep blue water and a pale sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does an emperor penguin initially appear as a sharp black and white silhouette against the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the penguin's solid form slowly liquefy, with the black portion dissolving into thick, inky fluid that flows down its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the inky droplets hit the icy surface, spread into expanding black ripples, and then vanish into thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward to follow the inky drops into the water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004647168796509504 + ], + "opposite_logprob_tokens": [ + -5.379647254943848 + ], + "logprob_reward": -0.004647168796509504, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004647168796509504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06183340400457382, + 0.059956472367048264, + 0.024334516376256943, + 0.06752745807170868, + 0.02021956443786621, + 0.017686426639556885, + 0.022828558459877968, + 0.02896052598953247, + 0.40124279260635376, + 0.2954103350639343 + ], + "credit_map": [ + 0.014404690824449062, + 0.0048610069788992405, + 0.006333363242447376, + 0.0035940518137067556, + 0.005200580228120089, + 0.016843222081661224, + 0.004166613332927227, + 0.006429870147258043, + 0.014126038178801537, + 0.0047766175121068954, + 0.0046134814620018005, + 0.00433502858504653, + 0.0063466960564255714, + 0.014584709890186787, + 0.004604438319802284, + 0.006569457706063986, + 0.0049522374756634235, + 0.0017992345383390784, + 0.0029533347114920616, + 0.001565986662171781, + 0.002408938715234399, + 0.004692814778536558, + 0.0022759176790714264, + 0.0036860485561192036, + 0.024033578112721443, + 0.0031971619464457035, + 0.003778808517381549, + 0.0017930972389876842, + 0.002426315564662218, + 0.023900723084807396, + 0.0027844849973917007, + 0.005613294430077076, + 0.004733066540211439, + 0.0013349225046113133, + 0.001757061225362122, + 0.0015222064685076475, + 0.0020815155003219843, + 0.004771476145833731, + 0.002088341396301985, + 0.0019309750059619546, + 0.0019635639619082212, + 0.0018296187045052648, + 0.002426183084025979, + 0.0016257582465186715, + 0.002569988602772355, + 0.002208923688158393, + 0.0025769323110580444, + 0.0024854582734405994, + 0.0026585652958601713, + 0.0020640285219997168, + 0.0031704504508525133, + 0.0032355068251490593, + 0.0027100786101073027, + 0.0027790777385234833, + 0.0025628472212702036, + 0.0036480012349784374, + 0.004185589030385017, + 0.0014439536025747657, + 0.004865834955126047, + 0.0018667579861357808, + 0.0023639665450900793, + 0.00638053473085165, + 0.002602271270006895, + 0.0052516162395477295, + 0.03197990357875824, + 0.0073412624187767506, + 0.026634592562913895, + 0.02531069703400135, + 0.010897913947701454, + 0.09937120229005814, + 0.043442901223897934, + 0.1562643051147461, + 0.019825365394353867, + 0.007473035249859095, + 0.10576985776424408, + 0.01746230013668537, + 0.008497519418597221, + 0.02366819977760315, + 0.025321612134575844, + 0.0873924046754837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.004647168796509504 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the edge of a vast, ancient glacier with deep blue water and a pale sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does an emperor penguin initially appear as a sharp black and white silhouette against the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the penguin's solid form slowly liquefy, with the black portion dissolving into thick, inky fluid that flows down its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the inky droplets hit the icy surface, spread into expanding black ripples, and then vanish into thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward to follow the inky drops into the water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0790102630853653 + ], + "opposite_logprob_tokens": [ + -2.579010248184204 + ], + "logprob_reward": -0.0790102630853653, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0790102630853653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06708436459302902, + 0.08576735854148865, + 0.047624923288822174, + 0.03261186182498932, + 0.02206839621067047, + 0.011260660365223885, + 0.027440981939435005, + 0.02891930565237999, + 0.3555617928504944, + 0.3216603398323059 + ], + "credit_map": [ + 0.020480526611208916, + 0.0044560725800693035, + 0.004755410831421614, + 0.0028983207885175943, + 0.004577061627060175, + 0.02138744480907917, + 0.003223877400159836, + 0.005305644124746323, + 0.016253935173153877, + 0.006928944028913975, + 0.008216270245611668, + 0.0072279442101716995, + 0.009396720677614212, + 0.01782982237637043, + 0.008771915920078754, + 0.011141804046928883, + 0.010179631412029266, + 0.003492987249046564, + 0.0040383716113865376, + 0.003477846970781684, + 0.005223906598985195, + 0.012214344926178455, + 0.00432822247967124, + 0.004669609945267439, + 0.007299942430108786, + 0.0022387884091585875, + 0.0030271257273852825, + 0.002636427292600274, + 0.0032024220563471317, + 0.007627091836184263, + 0.0033269436098635197, + 0.003253120696172118, + 0.003257843665778637, + 0.002055311808362603, + 0.0036231132689863443, + 0.00167481554672122, + 0.002617446705698967, + 0.003605844918638468, + 0.002445221645757556, + 0.0027887977194041014, + 0.0020058676600456238, + 0.001005090307444334, + 0.001653298269957304, + 0.001015746034681797, + 0.0008878172957338393, + 0.0017827940173447132, + 0.0011424756376072764, + 0.0017675708513706923, + 0.003360641421750188, + 0.0015499049331992865, + 0.004248547833412886, + 0.004435097333043814, + 0.0019164211116731167, + 0.004734302405267954, + 0.0034715153742581606, + 0.003724549198523164, + 0.0041931974701583385, + 0.0013304848689585924, + 0.005431510508060455, + 0.002311557065695524, + 0.0023987225722521544, + 0.0057143266312778, + 0.0022787810303270817, + 0.00526072271168232, + 0.02972925826907158, + 0.006545133423060179, + 0.021487755700945854, + 0.02240452915430069, + 0.009833266027271748, + 0.08869475871324539, + 0.03889908641576767, + 0.1379680037498474, + 0.018735555931925774, + 0.006511964835226536, + 0.12062150239944458, + 0.020683135837316513, + 0.007249038200825453, + 0.024596989154815674, + 0.027140796184539795, + 0.09612135589122772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 1 + ], + "selected_reward": -0.0790102630853653 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the edge of a vast, ancient glacier with deep blue water and a pale sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does an emperor penguin initially appear as a sharp black and white silhouette against the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the penguin's solid form slowly liquefy, with the black portion dissolving into thick, inky fluid that flows down its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the inky droplets hit the icy surface, spread into expanding black ripples, and then vanish into thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward to follow the inky drops into the water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204963102936745 + ], + "opposite_logprob_tokens": [ + -2.812049627304077 + ], + "logprob_reward": -0.06204963102936745, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204963102936745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03926333039999008, + 0.020651983097195625, + 0.024230025708675385, + 0.16159945726394653, + 0.01649477705359459, + 0.012178219854831696, + 0.04137295484542847, + 0.06858987361192703, + 0.0960763469338417, + 0.5195429921150208 + ], + "credit_map": [ + 0.012688633985817432, + 0.0021979911252856255, + 0.00296569149941206, + 0.0018278384814038873, + 0.002575001446530223, + 0.012998029589653015, + 0.0016239081742241979, + 0.002386233536526561, + 0.005257792770862579, + 0.0012852904619649053, + 0.0023616584949195385, + 0.0014041116228327155, + 0.0014422606909647584, + 0.005458156578242779, + 0.0015165579970926046, + 0.001926154363900423, + 0.008188138715922832, + 0.0011151207145303488, + 0.0014307367382571101, + 0.0008612281526438892, + 0.0011367429979145527, + 0.008716537617146969, + 0.001128177740611136, + 0.0016533449525013566, + 0.06383493542671204, + 0.006463567726314068, + 0.0057625374756753445, + 0.0032323135528713465, + 0.004924198612570763, + 0.06638913601636887, + 0.00480539770796895, + 0.00618738541379571, + 0.003180104773491621, + 0.0013824228662997484, + 0.0015741171082481742, + 0.0016250170301645994, + 0.0015885768225416541, + 0.002852507634088397, + 0.0020244214683771133, + 0.0022676093503832817, + 0.001837650896050036, + 0.0009419957059435546, + 0.002048209309577942, + 0.0011243978515267372, + 0.0012692790478467941, + 0.0018661388894543052, + 0.0012692437740042806, + 0.0018213038565590978, + 0.006026185117661953, + 0.0026883594691753387, + 0.0068161021918058395, + 0.0027052152436226606, + 0.00470874784514308, + 0.007695864886045456, + 0.0033569675870239735, + 0.0073755113407969475, + 0.010595166124403477, + 0.003387502860277891, + 0.012469463050365448, + 0.00543841440230608, + 0.005333711393177509, + 0.012107610702514648, + 0.006601095665246248, + 0.01265690941363573, + 0.010488295927643776, + 0.006868922151625156, + 0.01611311174929142, + 0.011174955405294895, + 0.007477202918380499, + 0.011265171691775322, + 0.011165847070515156, + 0.021522842347621918, + 0.033731717616319656, + 0.011331599205732346, + 0.19061648845672607, + 0.032058991491794586, + 0.012358679436147213, + 0.04094230383634567, + 0.04354333505034447, + 0.15495985746383667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 8 + ], + "selected_reward": -0.06204963102936745 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_6/index_tensor([773, 773])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_6/index_tensor([773, 773])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..68d29b49f3896c54808dc196d45a53c451ec2c80 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_6/index_tensor([773, 773])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, close-up shot, viewed from a slightly upward angle, captures a scene imbued with the artistic conception of a traditional Chinese ink painting. In a deep, mist-shrouded valley, the outlines of distant peaks loom in a gray haze of varying shades, like smudges of ink on rice paper. An elder, resembling a sage, stands alone on a flat, moss-covered rock jutting from the cliff face. He wears a pure white silk robe with wide sleeves, his long, silvery hair and flowing beard gently swaying in the mountain breeze. With his eyes slightly closed, his expression serene and focused, he slowly and effortlessly practices Tai Chi with the rhythmic grace of flowing water. His arms draw steady and deliberate curves, his body gently rotating in response, each movement imbued with power and harmony. The wide sleeves and hem of his robe gently rise and fall with his movements and the mountain breeze, a gesture of exquisite grace. Soft diffuse light penetrates the mist, creating a tranquil and transcendent atmosphere for the entire scene. The main colors of the picture are elegant black, white and gray, with only moss and the dark green of ancient pines in the distance dotted in between, creating a sense of depth and tranquility of harmony between man and nature.", + "seed": 6237, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_6/index_773_seed_6237_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.20118502117693424, + "vlm_logprob": -0.20118502117693424, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.20118502117693424, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a traditional Chinese ink painting with black, white, and gray tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elder with long silvery hair and a flowing beard wearing a pure white silk robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elder standing on a flat, moss-covered rock jutting from a cliff face in a mist-shrouded valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elder performing slow, deliberate Tai Chi movements with his arms drawing steady curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow close-up viewed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7590634226799011 + ], + "opposite_logprob_tokens": [ + -0.6340634226799011 + ], + "logprob_reward": -0.7590634226799011, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7590634226799011, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.14890959858894348, + 0.03403427079319954, + 0.036500487476587296, + 0.04024777188897133, + 0.037694621831178665, + 0.044781506061553955, + 0.32126155495643616, + 0.06664346158504486, + 0.13242852687835693, + 0.1374981850385666 + ], + "credit_map": [ + 0.013199793174862862, + 0.017773687839508057, + 0.018234161660075188, + 0.015422563999891281, + 0.023643115535378456, + 0.019605452194809914, + 0.02452346682548523, + 0.016507357358932495, + 0.003686741227284074, + 0.0038625109009444714, + 0.00624507712200284, + 0.003464913461357355, + 0.004791122395545244, + 0.0034618538338690996, + 0.004400237929075956, + 0.004121812991797924, + 0.0034775950480252504, + 0.003953937441110611, + 0.0045174527913331985, + 0.004078795202076435, + 0.006524418015033007, + 0.004180647898465395, + 0.005849342793226242, + 0.003918296191841364, + 0.004247609060257673, + 0.004875353071838617, + 0.006633060518652201, + 0.004676086362451315, + 0.005144424270838499, + 0.004016653634607792, + 0.007225779816508293, + 0.003428804688155651, + 0.004353672731667757, + 0.004421974997967482, + 0.005061055067926645, + 0.004903427325189114, + 0.004566310439258814, + 0.005197723396122456, + 0.005224057473242283, + 0.00396640133112669, + 0.004114128183573484, + 0.00610737781971693, + 0.0062128049321472645, + 0.004257528576999903, + 0.010258282534778118, + 0.006298137363046408, + 0.00487002357840538, + 0.0026632242370396852, + 0.029458822682499886, + 0.04804276302456856, + 0.01908218115568161, + 0.05703376978635788, + 0.05011642351746559, + 0.052103545516729355, + 0.034866813570261, + 0.03055724687874317, + 0.008235378190875053, + 0.009907634928822517, + 0.009601186029613018, + 0.015259629115462303, + 0.007038081530481577, + 0.005129205994307995, + 0.00705665722489357, + 0.0044156890362501144, + 0.012267138808965683, + 0.024468064308166504, + 0.00872261542826891, + 0.03022405505180359, + 0.019424542784690857, + 0.007415675092488527, + 0.021815257146954536, + 0.008091186173260212, + 0.031024301424622536, + 0.01933298446238041, + 0.020412396639585495, + 0.019695479422807693, + 0.01344526931643486, + 0.01217818446457386, + 0.015146863646805286, + 0.006262699142098427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.7590634226799011 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a traditional Chinese ink painting with black, white, and gray tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elder with long silvery hair and a flowing beard wearing a pure white silk robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elder standing on a flat, moss-covered rock jutting from a cliff face in a mist-shrouded valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elder performing slow, deliberate Tai Chi movements with his arms drawing steady curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow close-up viewed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008651750162243843 + ], + "opposite_logprob_tokens": [ + -4.7586517333984375 + ], + "logprob_reward": -0.008651750162243843, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008651750162243843, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24497239291667938, + 0.046938370913267136, + 0.034660257399082184, + 0.03808864206075668, + 0.04333262890577316, + 0.03306788206100464, + 0.29193568229675293, + 0.03825034201145172, + 0.1458594650030136, + 0.08289438486099243 + ], + "credit_map": [ + 0.022511225193738937, + 0.03143151476979256, + 0.019943786785006523, + 0.028164120391011238, + 0.04966891556978226, + 0.030284924432635307, + 0.03326078876852989, + 0.029707102105021477, + 0.004236631095409393, + 0.005093052983283997, + 0.0060634920373559, + 0.0050122966058552265, + 0.008208578452467918, + 0.005483947694301605, + 0.007469602860510349, + 0.005370768718421459, + 0.003263219492509961, + 0.003842771053314209, + 0.004380689933896065, + 0.0035069503355771303, + 0.006835960317403078, + 0.004039858002215624, + 0.005111133214086294, + 0.003679675282910466, + 0.0041166855953633785, + 0.00461368216201663, + 0.005248584784567356, + 0.004723145626485348, + 0.005867558065801859, + 0.0039289891719818115, + 0.006508732680231333, + 0.003081263741478324, + 0.005149259697645903, + 0.005613694898784161, + 0.004931049887090921, + 0.005184274632483721, + 0.006386023014783859, + 0.006250826641917229, + 0.005990455858409405, + 0.003827042644843459, + 0.0030658752657473087, + 0.004487178288400173, + 0.005049589555710554, + 0.0029665015172213316, + 0.007370947394520044, + 0.004106324166059494, + 0.00399400619789958, + 0.0020274603739380836, + 0.023623453453183174, + 0.043035924434661865, + 0.016751626506447792, + 0.0512787401676178, + 0.051408033818006516, + 0.04342355579137802, + 0.03438848257064819, + 0.02802588976919651, + 0.0036691720597445965, + 0.004404151346534491, + 0.004636953119188547, + 0.005456702318042517, + 0.004942120518535376, + 0.005320745985955, + 0.005955039523541927, + 0.003865455510094762, + 0.009307187981903553, + 0.03749195113778114, + 0.008473215624690056, + 0.03841711953282356, + 0.01720130443572998, + 0.0059978836216032505, + 0.021086232736706734, + 0.007884579710662365, + 0.007372949738055468, + 0.008078496903181076, + 0.007366693578660488, + 0.011579000391066074, + 0.012758748605847359, + 0.007136405911296606, + 0.017685730010271072, + 0.01091635413467884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.008651750162243843 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a traditional Chinese ink painting with black, white, and gray tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elder with long silvery hair and a flowing beard wearing a pure white silk robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elder standing on a flat, moss-covered rock jutting from a cliff face in a mist-shrouded valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elder performing slow, deliberate Tai Chi movements with his arms drawing steady curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow close-up viewed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607036218047142 + ], + "opposite_logprob_tokens": [ + -4.141070365905762 + ], + "logprob_reward": -0.01607036218047142, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607036218047142, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1562262326478958, + 0.05394114553928375, + 0.03505922108888626, + 0.02666151337325573, + 0.06806919723749161, + 0.054657820612192154, + 0.3863956034183502, + 0.04974884167313576, + 0.10092389583587646, + 0.06831659376621246 + ], + "credit_map": [ + 0.015712769702076912, + 0.019665906205773354, + 0.02238154597580433, + 0.01654921844601631, + 0.02291032299399376, + 0.01849181391298771, + 0.024228626862168312, + 0.016286034137010574, + 0.005672215484082699, + 0.0051120282150805, + 0.00876587349921465, + 0.0066729094833135605, + 0.007410014048218727, + 0.005602051038295031, + 0.008328345604240894, + 0.00637770863249898, + 0.002964991144835949, + 0.0036866185255348682, + 0.004403612110763788, + 0.004553390666842461, + 0.00578899635002017, + 0.003561248304322362, + 0.005836128257215023, + 0.004264235030859709, + 0.0032708814833313227, + 0.003172072349116206, + 0.003890222404152155, + 0.003762266132980585, + 0.0025124731473624706, + 0.0029180312994867563, + 0.004871580749750137, + 0.002263987436890602, + 0.01016357634216547, + 0.0072282785549759865, + 0.009061481803655624, + 0.006935243494808674, + 0.007031977642327547, + 0.007320382632315159, + 0.010761037468910217, + 0.009567220695316792, + 0.0069597880356013775, + 0.006841741502285004, + 0.012656962499022484, + 0.0059096054174005985, + 0.005328711587935686, + 0.007281558588147163, + 0.005410039331763983, + 0.004269416909664869, + 0.03596588224172592, + 0.05748177319765091, + 0.020717794075608253, + 0.0650634691119194, + 0.0659555196762085, + 0.06333516538143158, + 0.04020439460873604, + 0.03767158091068268, + 0.005089731886982918, + 0.007526955101639032, + 0.0048795524053275585, + 0.009212449193000793, + 0.007120048627257347, + 0.003489354159682989, + 0.009040473029017448, + 0.0033902793657034636, + 0.008378691039979458, + 0.026473958045244217, + 0.00942288339138031, + 0.023003360256552696, + 0.010891074314713478, + 0.005540893878787756, + 0.012156166136264801, + 0.005056873895227909, + 0.006890427321195602, + 0.010143349878489971, + 0.009187544696033001, + 0.009999766945838928, + 0.00902540422976017, + 0.007645176723599434, + 0.010962425731122494, + 0.004462498240172863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.01607036218047142 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a traditional Chinese ink painting with black, white, and gray tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elder with long silvery hair and a flowing beard wearing a pure white silk robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elder standing on a flat, moss-covered rock jutting from a cliff face in a mist-shrouded valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elder performing slow, deliberate Tai Chi movements with his arms drawing steady curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow close-up viewed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581606775522232 + ], + "opposite_logprob_tokens": [ + -3.8955817222595215 + ], + "logprob_reward": -0.020581606775522232, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581606775522232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10960571467876434, + 0.04002341255545616, + 0.03976745530962944, + 0.0972818061709404, + 0.04246173053979874, + 0.036629289388656616, + 0.39633727073669434, + 0.07789215445518494, + 0.10471104085445404, + 0.05529008060693741 + ], + "credit_map": [ + 0.01195853017270565, + 0.012054369784891605, + 0.00969785638153553, + 0.011390017345547676, + 0.018857523798942566, + 0.01363887544721365, + 0.018109146505594254, + 0.013899384997785091, + 0.004388740751892328, + 0.004367209039628506, + 0.005273836199194193, + 0.004921896383166313, + 0.005274932831525803, + 0.004234489984810352, + 0.007101195864379406, + 0.004461109638214111, + 0.003059228416532278, + 0.0048708561807870865, + 0.0036637396551668644, + 0.005464551039040089, + 0.007780136540532112, + 0.004897314589470625, + 0.006304542999714613, + 0.0037270875182002783, + 0.004920109175145626, + 0.01347642857581377, + 0.0061265998519957066, + 0.01236766204237938, + 0.026736751198768616, + 0.01335626095533371, + 0.014715089462697506, + 0.005582907702773809, + 0.003407869255170226, + 0.005609237123280764, + 0.0039369226433336735, + 0.005770415533334017, + 0.007646831218153238, + 0.004164782352745533, + 0.008110232651233673, + 0.003815440693870187, + 0.003699592547491193, + 0.004942154977470636, + 0.0035683319438248873, + 0.0047638025134801865, + 0.006184816360473633, + 0.005231919698417187, + 0.006054023280739784, + 0.002184648532420397, + 0.04025351628661156, + 0.05810084193944931, + 0.020357081666588783, + 0.0612107515335083, + 0.0654463842511177, + 0.06698966026306152, + 0.04491191357374191, + 0.03906712681055069, + 0.006205772515386343, + 0.01266649179160595, + 0.00710156187415123, + 0.020385604351758957, + 0.006928134243935347, + 0.005560361314564943, + 0.012511588633060455, + 0.006532639265060425, + 0.007832635194063187, + 0.029955074191093445, + 0.008047020994126797, + 0.025050578638911247, + 0.009728129021823406, + 0.005684094503521919, + 0.011921564117074013, + 0.006491943262517452, + 0.007051480934023857, + 0.0067750439047813416, + 0.005193368997424841, + 0.00902948435395956, + 0.009191307239234447, + 0.004189582541584969, + 0.009511761367321014, + 0.00434804568067193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.020581606775522232 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a traditional Chinese ink painting with black, white, and gray tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elder with long silvery hair and a flowing beard wearing a pure white silk robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elder standing on a flat, moss-covered rock jutting from a cliff face in a mist-shrouded valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elder performing slow, deliberate Tai Chi movements with his arms drawing steady curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow close-up viewed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015579640865326 + ], + "opposite_logprob_tokens": [ + -1.701557993888855 + ], + "logprob_reward": -0.2015579640865326, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015579640865326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11863048374652863, + 0.03328139707446098, + 0.026626290753483772, + 0.02769855596125126, + 0.08503514528274536, + 0.03932187333703041, + 0.23974566161632538, + 0.12571363151073456, + 0.21247142553329468, + 0.09147550165653229 + ], + "credit_map": [ + 0.012240784242749214, + 0.015054921619594097, + 0.013767633587121964, + 0.011636567302048206, + 0.022589582949876785, + 0.017469212412834167, + 0.014889891259372234, + 0.01098189502954483, + 0.004255261272192001, + 0.0033785656560212374, + 0.005044125020503998, + 0.00398099422454834, + 0.0046594757586717606, + 0.003518210956826806, + 0.0050238436087965965, + 0.0034209240693598986, + 0.003946356475353241, + 0.0026926719583570957, + 0.004264554940164089, + 0.003455184632912278, + 0.0034686303697526455, + 0.0027865981683135033, + 0.0032190720085054636, + 0.002793222665786743, + 0.0035408074036240578, + 0.003225896041840315, + 0.004111080896109343, + 0.003409601515159011, + 0.004094479605555534, + 0.002514817286282778, + 0.003876203205436468, + 0.0029256734997034073, + 0.008206550031900406, + 0.011037842370569706, + 0.00744016095995903, + 0.010563334450125694, + 0.01471725944429636, + 0.009993348270654678, + 0.014436539262533188, + 0.008640113286674023, + 0.004155291244387627, + 0.005398799665272236, + 0.006349612958729267, + 0.004194744862616062, + 0.005717666819691658, + 0.006050981115549803, + 0.004355093464255333, + 0.0030996869318187237, + 0.020298218354582787, + 0.03427765518426895, + 0.011842777021229267, + 0.03946109116077423, + 0.03959575667977333, + 0.036491792649030685, + 0.03425465151667595, + 0.023523731157183647, + 0.009696843102574348, + 0.01716739684343338, + 0.011984034441411495, + 0.0327458456158638, + 0.01478076633065939, + 0.009849951602518559, + 0.02048330008983612, + 0.00900549627840519, + 0.02066146954894066, + 0.05563386529684067, + 0.01598251610994339, + 0.04297064617276192, + 0.01938808709383011, + 0.015760354697704315, + 0.02558361180126667, + 0.016490887850522995, + 0.008986568078398705, + 0.015953635796904564, + 0.008892840705811977, + 0.012577737681567669, + 0.013625255785882473, + 0.006733736954629421, + 0.015330882743000984, + 0.00937485508620739 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.2015579640865326 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, close-up shot, viewed from a slightly upward angle, captures a scene imbued with the artistic conception of a traditional Chinese ink painting. In a deep, mist-shrouded valley, the outlines of distant peaks loom in a gray haze of varying shades, like smudges of ink on rice paper. An elder, resembling a sage, stands alone on a flat, moss-covered rock jutting from the cliff face. He wears a pure white silk robe with wide sleeves, his long, silvery hair and flowing beard gently swaying in the mountain breeze. With his eyes slightly closed, his expression serene and focused, he slowly and effortlessly practices Tai Chi with the rhythmic grace of flowing water. His arms draw steady and deliberate curves, his body gently rotating in response, each movement imbued with power and harmony. The wide sleeves and hem of his robe gently rise and fall with his movements and the mountain breeze, a gesture of exquisite grace. Soft diffuse light penetrates the mist, creating a tranquil and transcendent atmosphere for the entire scene. The main colors of the picture are elegant black, white and gray, with only moss and the dark green of ancient pines in the distance dotted in between, creating a sense of depth and tranquility of harmony between man and nature.", + "seed": 6238, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_6/index_773_seed_6238_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.5534447904676199, + "vlm_logprob": -0.5534447904676199, + "vlm_yes_no_margin": 1.975, + "vlm_token_credit": -0.5534447904676199, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a traditional Chinese ink painting with black, white, and gray tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elder with long silvery hair and a flowing beard wearing a pure white silk robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elder standing on a flat, moss-covered rock jutting from a cliff face in a mist-shrouded valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elder performing slow, deliberate Tai Chi movements with his arms drawing steady curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow close-up viewed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.580465793609619 + ], + "opposite_logprob_tokens": [ + -0.08046575635671616 + ], + "logprob_reward": -2.580465793609619, + "yes_no_margin_reward": -2.5, + "token_credit_reward": -2.580465793609619, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.479912132024765, + 0.10335543751716614, + 0.07080622762441635, + 0.04533956199884415, + 0.03985337167978287, + 0.05209483578801155, + 0.05984874069690704, + 0.04508819803595543, + 0.03928162530064583, + 0.06441982835531235 + ], + "credit_map": [ + 0.03884442523121834, + 0.05651278421282768, + 0.0707351341843605, + 0.08349877595901489, + 0.049132123589515686, + 0.0567731037735939, + 0.06705926358699799, + 0.05735652893781662, + 0.012032284401357174, + 0.014674792997539043, + 0.013722115196287632, + 0.009579100646078587, + 0.016247417777776718, + 0.015332822687923908, + 0.01219695620238781, + 0.009569954127073288, + 0.006498390808701515, + 0.009267859160900116, + 0.012769021093845367, + 0.010678332298994064, + 0.006749016232788563, + 0.01054496131837368, + 0.007754444610327482, + 0.00654420955106616, + 0.004427555948495865, + 0.004752575885504484, + 0.008554444648325443, + 0.004291625693440437, + 0.004782582633197308, + 0.006218923721462488, + 0.006813652813434601, + 0.005498205777257681, + 0.004529220052063465, + 0.004050636198371649, + 0.006618549581617117, + 0.003448622301220894, + 0.004451591521501541, + 0.006299259141087532, + 0.005261892452836037, + 0.005193604622036219, + 0.004450553562492132, + 0.006725601386278868, + 0.006880726665258408, + 0.0065706707537174225, + 0.009683765470981598, + 0.0038084168918430805, + 0.00891619548201561, + 0.005058910697698593, + 0.00853906199336052, + 0.005901806056499481, + 0.005205092951655388, + 0.010368088260293007, + 0.010568181052803993, + 0.004150249995291233, + 0.009607899002730846, + 0.005508366506546736, + 0.006494131404906511, + 0.004972519818693399, + 0.006771173793822527, + 0.005105376709252596, + 0.005960139445960522, + 0.0030613879207521677, + 0.007303569931536913, + 0.005419899243861437, + 0.003779631108045578, + 0.004852612968534231, + 0.004495172295719385, + 0.003612347412854433, + 0.0054754880256950855, + 0.0030562742613255978, + 0.009724907577037811, + 0.004285192117094994, + 0.008011052384972572, + 0.008709417656064034, + 0.012942846864461899, + 0.004532915074378252, + 0.007564948406070471, + 0.00637233117595315, + 0.00789131224155426, + 0.008395012468099594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -2.580465793609619 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a traditional Chinese ink painting with black, white, and gray tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elder with long silvery hair and a flowing beard wearing a pure white silk robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elder standing on a flat, moss-covered rock jutting from a cliff face in a mist-shrouded valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elder performing slow, deliberate Tai Chi movements with his arms drawing steady curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow close-up viewed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03820045664906502 + ], + "opposite_logprob_tokens": [ + -3.2882003784179688 + ], + "logprob_reward": -0.03820045664906502, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03820045664906502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4768730103969574, + 0.10011613368988037, + 0.06964890658855438, + 0.06764201074838638, + 0.0425059050321579, + 0.05673939734697342, + 0.05304044112563133, + 0.04007070139050484, + 0.041548632085323334, + 0.051814913749694824 + ], + "credit_map": [ + 0.035276561975479126, + 0.04893400892615318, + 0.07040169835090637, + 0.09202404320240021, + 0.0549074225127697, + 0.056090839207172394, + 0.06323755532503128, + 0.05600084736943245, + 0.012792701832950115, + 0.015376887284219265, + 0.014903152361512184, + 0.010412914678454399, + 0.011493555270135403, + 0.014031689614057541, + 0.010621801018714905, + 0.010483422316610813, + 0.006307799369096756, + 0.006150317844003439, + 0.012809410691261292, + 0.01305397693067789, + 0.006977387238293886, + 0.00881864596158266, + 0.008666081354022026, + 0.006865285336971283, + 0.006254436913877726, + 0.006298734340816736, + 0.011889277026057243, + 0.007799243088811636, + 0.00797310657799244, + 0.008294155821204185, + 0.011953589506447315, + 0.007179461885243654, + 0.0036537826526910067, + 0.003566145896911621, + 0.006768336985260248, + 0.004357544705271721, + 0.006536510773003101, + 0.006126160267740488, + 0.006270175334066153, + 0.005227245856076479, + 0.005010298453271389, + 0.006384291220456362, + 0.007908377796411514, + 0.007883683778345585, + 0.008203846402466297, + 0.004477907903492451, + 0.010008790530264378, + 0.006862195208668709, + 0.007630188018083572, + 0.005717426538467407, + 0.005761925131082535, + 0.00882660411298275, + 0.006732694339007139, + 0.0036335692275315523, + 0.010114978067576885, + 0.004623051732778549, + 0.006244420539587736, + 0.005049433093518019, + 0.006841080728918314, + 0.0028251560870558023, + 0.004203885328024626, + 0.004787051118910313, + 0.0041784425266087055, + 0.005941225215792656, + 0.0031751529313623905, + 0.005837262142449617, + 0.0037172159645706415, + 0.0036562259774655104, + 0.007232740521430969, + 0.004614377859979868, + 0.0077967569231987, + 0.005518896505236626, + 0.006990618538111448, + 0.00755241559818387, + 0.010555798187851906, + 0.0032295563723891973, + 0.0038468611892312765, + 0.005553592462092638, + 0.006183444522321224, + 0.00790262222290039 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.03820045664906502 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a traditional Chinese ink painting with black, white, and gray tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elder with long silvery hair and a flowing beard wearing a pure white silk robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elder standing on a flat, moss-covered rock jutting from a cliff face in a mist-shrouded valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elder performing slow, deliberate Tai Chi movements with his arms drawing steady curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow close-up viewed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029824355617165565 + ], + "opposite_logprob_tokens": [ + -3.5298242568969727 + ], + "logprob_reward": -0.029824355617165565, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029824355617165565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4267543852329254, + 0.12890318036079407, + 0.046195365488529205, + 0.055434029549360275, + 0.049504008144140244, + 0.06658776849508286, + 0.056159574538469315, + 0.08577977120876312, + 0.03341217339038849, + 0.05126972496509552 + ], + "credit_map": [ + 0.04036280885338783, + 0.0609297938644886, + 0.07180671393871307, + 0.0669763907790184, + 0.031892720609903336, + 0.05717236176133156, + 0.04964029788970947, + 0.04797327145934105, + 0.012022562325000763, + 0.01707838475704193, + 0.019266096875071526, + 0.01216852106153965, + 0.018103431910276413, + 0.023209823295474052, + 0.013264201581478119, + 0.013790163211524487, + 0.0045067425817251205, + 0.005164317321032286, + 0.0054725161753594875, + 0.006224957760423422, + 0.008822496049106121, + 0.004572358913719654, + 0.006730818189680576, + 0.004701157100498676, + 0.004480074625462294, + 0.0050380597822368145, + 0.010077406652271748, + 0.0049804821610450745, + 0.006862056907266378, + 0.008301159366965294, + 0.00917744915932417, + 0.00651734322309494, + 0.004681635648012161, + 0.004259802401065826, + 0.008531942963600159, + 0.006797378417104483, + 0.007656006142497063, + 0.00527526019141078, + 0.007519851438701153, + 0.004782133270055056, + 0.00670347036793828, + 0.008018200285732746, + 0.010379289276897907, + 0.00697743846103549, + 0.013110450468957424, + 0.005445071030408144, + 0.009453061036765575, + 0.006500786170363426, + 0.007355835288763046, + 0.0058586266823112965, + 0.006426800973713398, + 0.008000473491847515, + 0.006452939007431269, + 0.004477991722524166, + 0.012208295054733753, + 0.005378608591854572, + 0.020915087312459946, + 0.010223730467259884, + 0.009918228723108768, + 0.00750354677438736, + 0.00805695727467537, + 0.007111149840056896, + 0.012215458787977695, + 0.00983561109751463, + 0.002822399837896228, + 0.004044277593493462, + 0.004435803275555372, + 0.002619720296934247, + 0.004049486014991999, + 0.0025045047514140606, + 0.009155039675533772, + 0.0037809398490935564, + 0.006346052512526512, + 0.008059337735176086, + 0.008243074640631676, + 0.00403146306052804, + 0.0062224725261330605, + 0.0039322152733802795, + 0.008184182457625866, + 0.00625093188136816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.029824355617165565 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a traditional Chinese ink painting with black, white, and gray tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elder with long silvery hair and a flowing beard wearing a pure white silk robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elder standing on a flat, moss-covered rock jutting from a cliff face in a mist-shrouded valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elder performing slow, deliberate Tai Chi movements with his arms drawing steady curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow close-up viewed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868418723344803 + ], + "opposite_logprob_tokens": [ + -3.0486841201782227 + ], + "logprob_reward": -0.04868418723344803, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868418723344803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4427875280380249, + 0.1268281787633896, + 0.0810977965593338, + 0.07840963453054428, + 0.059673111885786057, + 0.05944614112377167, + 0.038975320756435394, + 0.03660661727190018, + 0.03446245938539505, + 0.041713226586580276 + ], + "credit_map": [ + 0.03717479109764099, + 0.0421515516936779, + 0.05614488199353218, + 0.06858476996421814, + 0.05930722504854202, + 0.0512826181948185, + 0.07853931933641434, + 0.049602359533309937, + 0.01759284734725952, + 0.020067226141691208, + 0.017759893089532852, + 0.013160749338567257, + 0.016175799071788788, + 0.015506355091929436, + 0.014794966205954552, + 0.011770336888730526, + 0.009831300005316734, + 0.008419937454164028, + 0.01002793200314045, + 0.01044527254998684, + 0.01327764242887497, + 0.00875360332429409, + 0.012064315378665924, + 0.008277795277535915, + 0.010024260729551315, + 0.009669772349298, + 0.008741253986954689, + 0.006639061961323023, + 0.01606357842683792, + 0.007455207873135805, + 0.011074588634073734, + 0.008741914294660091, + 0.005333998240530491, + 0.006916145794093609, + 0.007309406530112028, + 0.008562509901821613, + 0.01220926083624363, + 0.005116071552038193, + 0.010224385187029839, + 0.004001334309577942, + 0.005915804300457239, + 0.0068779438734054565, + 0.00630171038210392, + 0.004426419734954834, + 0.015926698222756386, + 0.0056754513643682, + 0.007751492317765951, + 0.006570623256266117, + 0.004707508720457554, + 0.0037921296898275614, + 0.004351092968136072, + 0.005024114158004522, + 0.004206777084618807, + 0.002844474511221051, + 0.010561707429587841, + 0.0034875166602432728, + 0.0049330019392073154, + 0.004453965928405523, + 0.006116775330156088, + 0.0026356689631938934, + 0.003887052880600095, + 0.0036901829298585653, + 0.006039510015398264, + 0.004850461613386869, + 0.0025529025588184595, + 0.005017424002289772, + 0.004285001661628485, + 0.002993931993842125, + 0.005457640625536442, + 0.0036186163779348135, + 0.007045144680887461, + 0.0034917956218123436, + 0.0034302049316465855, + 0.005228875670582056, + 0.006756120827049017, + 0.004134087823331356, + 0.005634258035570383, + 0.005599064752459526, + 0.00576020497828722, + 0.00517041003331542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.04868418723344803 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a traditional Chinese ink painting with black, white, and gray tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elder with long silvery hair and a flowing beard wearing a pure white silk robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elder standing on a flat, moss-covered rock jutting from a cliff face in a mist-shrouded valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elder performing slow, deliberate Tai Chi movements with his arms drawing steady curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow close-up viewed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004915922880173 + ], + "opposite_logprob_tokens": [ + -2.6950490474700928 + ], + "logprob_reward": -0.07004915922880173, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004915922880173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3973917067050934, + 0.11080149561166763, + 0.04551577568054199, + 0.07763523608446121, + 0.0681709572672844, + 0.07199458032846451, + 0.0412871278822422, + 0.06285794824361801, + 0.02889307774603367, + 0.09545205533504486 + ], + "credit_map": [ + 0.03590453788638115, + 0.047684021294116974, + 0.04882434755563736, + 0.06235840916633606, + 0.051841966807842255, + 0.0481235533952713, + 0.06266957521438599, + 0.039985306560993195, + 0.013194690458476543, + 0.018011346459388733, + 0.01659363880753517, + 0.01009456068277359, + 0.011275731027126312, + 0.020833924412727356, + 0.010514765977859497, + 0.010282839648425579, + 0.0040011219680309296, + 0.004710944835096598, + 0.006472891662269831, + 0.008933145552873611, + 0.004853792954236269, + 0.006052063778042793, + 0.004655339289456606, + 0.005836482159793377, + 0.006870985962450504, + 0.007649294566363096, + 0.011428612284362316, + 0.015464403666555882, + 0.010811062529683113, + 0.0074830674566328526, + 0.008988168090581894, + 0.00893964059650898, + 0.005822170991450548, + 0.007467909716069698, + 0.008490747772157192, + 0.011442511342465878, + 0.012654726393520832, + 0.00675491476431489, + 0.010836299508810043, + 0.004701676312834024, + 0.006474799010902643, + 0.009401896968483925, + 0.007633866276592016, + 0.011301473714411259, + 0.013127844780683517, + 0.00726123945787549, + 0.009275191463530064, + 0.007518266327679157, + 0.004131815861910582, + 0.0047989957965910435, + 0.005180714186280966, + 0.006189905107021332, + 0.007510792929679155, + 0.003356959205120802, + 0.006027864757925272, + 0.004090083763003349, + 0.008315050043165684, + 0.006751293782144785, + 0.005850536748766899, + 0.006198884453624487, + 0.010836309753358364, + 0.0044011930003762245, + 0.014800808392465115, + 0.005703871604055166, + 0.0029985690489411354, + 0.004513641819357872, + 0.0031962969806045294, + 0.0023289043456315994, + 0.004644135013222694, + 0.0023003872483968735, + 0.005005571059882641, + 0.003905573394149542, + 0.008959039114415646, + 0.00931697990745306, + 0.008341943845152855, + 0.009783660992980003, + 0.017304832115769386, + 0.011727717705070972, + 0.014974883757531643, + 0.01504299882799387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.07004915922880173 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, close-up shot, viewed from a slightly upward angle, captures a scene imbued with the artistic conception of a traditional Chinese ink painting. In a deep, mist-shrouded valley, the outlines of distant peaks loom in a gray haze of varying shades, like smudges of ink on rice paper. An elder, resembling a sage, stands alone on a flat, moss-covered rock jutting from the cliff face. He wears a pure white silk robe with wide sleeves, his long, silvery hair and flowing beard gently swaying in the mountain breeze. With his eyes slightly closed, his expression serene and focused, he slowly and effortlessly practices Tai Chi with the rhythmic grace of flowing water. His arms draw steady and deliberate curves, his body gently rotating in response, each movement imbued with power and harmony. The wide sleeves and hem of his robe gently rise and fall with his movements and the mountain breeze, a gesture of exquisite grace. Soft diffuse light penetrates the mist, creating a tranquil and transcendent atmosphere for the entire scene. The main colors of the picture are elegant black, white and gray, with only moss and the dark green of ancient pines in the distance dotted in between, creating a sense of depth and tranquility of harmony between man and nature.", + "seed": 6239, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_6/index_773_seed_6239_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.206609296053648, + "vlm_logprob": -0.206609296053648, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.206609296053648, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a traditional Chinese ink painting with black, white, and gray tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elder with long silvery hair and a flowing beard wearing a pure white silk robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elder standing on a flat, moss-covered rock jutting from a cliff face in a mist-shrouded valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elder performing slow, deliberate Tai Chi movements with his arms drawing steady curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow close-up viewed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8994782567024231 + ], + "opposite_logprob_tokens": [ + -0.5244782567024231 + ], + "logprob_reward": -0.8994782567024231, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8994782567024231, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2502668499946594, + 0.11669077724218369, + 0.04274560138583183, + 0.03740564361214638, + 0.05981416627764702, + 0.049512818455696106, + 0.11465638130903244, + 0.20981526374816895, + 0.08138042688369751, + 0.03771212697029114 + ], + "credit_map": [ + 0.025033660233020782, + 0.04389100521802902, + 0.032281629741191864, + 0.05480048432946205, + 0.024943653494119644, + 0.02383239008486271, + 0.022288737818598747, + 0.023195302113890648, + 0.023838480934500694, + 0.015017819590866566, + 0.03729661926627159, + 0.013545561581850052, + 0.00583904841914773, + 0.008966878056526184, + 0.0048463065177202225, + 0.007340065203607082, + 0.00491354800760746, + 0.0053955442272126675, + 0.005116673652082682, + 0.007794019766151905, + 0.003715016646310687, + 0.0058451429940760136, + 0.004057885613292456, + 0.0059077683836221695, + 0.003380345180630684, + 0.004317660350352526, + 0.00435050530359149, + 0.00620450172573328, + 0.0044853175058960915, + 0.004797878675162792, + 0.005106136202812195, + 0.004763295873999596, + 0.006857349071651697, + 0.006924217101186514, + 0.011592617258429527, + 0.006945926230400801, + 0.006592044606804848, + 0.00522590009495616, + 0.007904122583568096, + 0.007771986536681652, + 0.007530391216278076, + 0.004876427818089724, + 0.007378859911113977, + 0.005620107054710388, + 0.00660960515961051, + 0.00461354898288846, + 0.009030930697917938, + 0.003852947149425745, + 0.012214950285851955, + 0.016912143677473068, + 0.01683000661432743, + 0.01604614593088627, + 0.015089639462530613, + 0.013885867781937122, + 0.012983758933842182, + 0.010693879798054695, + 0.039251234382390976, + 0.030015939846634865, + 0.05705321207642555, + 0.023001179099082947, + 0.01371318195015192, + 0.012669072486460209, + 0.017143847420811653, + 0.016967590898275375, + 0.01235733088105917, + 0.010122009553015232, + 0.020377719774842262, + 0.010460834950208664, + 0.006298469379544258, + 0.005536608397960663, + 0.008934322744607925, + 0.007293132599443197, + 0.008782221004366875, + 0.004248099867254496, + 0.009071220643818378, + 0.00385857280343771, + 0.0024596250150352716, + 0.002790923463180661, + 0.003932514693588018, + 0.002568949246779084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.8994782567024231 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a traditional Chinese ink painting with black, white, and gray tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elder with long silvery hair and a flowing beard wearing a pure white silk robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elder standing on a flat, moss-covered rock jutting from a cliff face in a mist-shrouded valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elder performing slow, deliberate Tai Chi movements with his arms drawing steady curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow close-up viewed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007646694779396057 + ], + "opposite_logprob_tokens": [ + -4.882646560668945 + ], + "logprob_reward": -0.007646694779396057, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007646694779396057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25954824686050415, + 0.17895248532295227, + 0.05639667436480522, + 0.05039278790354729, + 0.03827846422791481, + 0.05948100984096527, + 0.1158667579293251, + 0.11375141143798828, + 0.07247792184352875, + 0.05485427752137184 + ], + "credit_map": [ + 0.028461892157793045, + 0.04300999268889427, + 0.04114297032356262, + 0.056634511798620224, + 0.025613414123654366, + 0.02105523645877838, + 0.01947491057217121, + 0.02415531687438488, + 0.03318155184388161, + 0.02285807952284813, + 0.06186123564839363, + 0.01804833672940731, + 0.00964102242141962, + 0.014027277939021587, + 0.007997117005288601, + 0.011337869800627232, + 0.007679815869778395, + 0.0071235233917832375, + 0.009284362196922302, + 0.008462528698146343, + 0.0046280669048428535, + 0.007152909878641367, + 0.004694323521107435, + 0.007371144834905863, + 0.004935670178383589, + 0.005829560570418835, + 0.0062379976734519005, + 0.006383245345205069, + 0.008895612321794033, + 0.006205123383551836, + 0.007437548600137234, + 0.004468030761927366, + 0.004137703683227301, + 0.004946546163409948, + 0.005465714726597071, + 0.005167555063962936, + 0.004921914543956518, + 0.0032150978222489357, + 0.005225954577326775, + 0.005197977181524038, + 0.007012395188212395, + 0.005765206180512905, + 0.00887156929820776, + 0.00746108777821064, + 0.007629789412021637, + 0.005547014065086842, + 0.011087159626185894, + 0.006106790620833635, + 0.014867805875837803, + 0.015140390954911709, + 0.019137071445584297, + 0.017087571322917938, + 0.013830963522195816, + 0.012452603317797184, + 0.01357455924153328, + 0.00977579690515995, + 0.01539774052798748, + 0.015971723943948746, + 0.014657435938715935, + 0.022199412807822227, + 0.01525556854903698, + 0.009590892121195793, + 0.009644663892686367, + 0.011033971793949604, + 0.00954209640622139, + 0.008836353197693825, + 0.00877007283270359, + 0.011841177009046078, + 0.008249884471297264, + 0.008935614489018917, + 0.009538162499666214, + 0.006764557212591171, + 0.007489428389817476, + 0.00509093701839447, + 0.007728072348982096, + 0.010639416053891182, + 0.006811155006289482, + 0.006126112304627895, + 0.006679864600300789, + 0.004289289005100727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.007646694779396057 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a traditional Chinese ink painting with black, white, and gray tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elder with long silvery hair and a flowing beard wearing a pure white silk robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elder standing on a flat, moss-covered rock jutting from a cliff face in a mist-shrouded valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elder performing slow, deliberate Tai Chi movements with his arms drawing steady curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow close-up viewed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020583825185894966 + ], + "opposite_logprob_tokens": [ + -3.8955838680267334 + ], + "logprob_reward": -0.020583825185894966, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020583825185894966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2830806076526642, + 0.29567646980285645, + 0.045353829860687256, + 0.03851712867617607, + 0.04146483168005943, + 0.030995190143585205, + 0.043267250061035156, + 0.09538572281599045, + 0.07656063139438629, + 0.04969833418726921 + ], + "credit_map": [ + 0.03531204164028168, + 0.041664332151412964, + 0.05181938782334328, + 0.0635724738240242, + 0.017225731164216995, + 0.034536246210336685, + 0.01621081680059433, + 0.022739585489034653, + 0.060912806540727615, + 0.038062769919633865, + 0.1066984012722969, + 0.027819687500596046, + 0.010657386854290962, + 0.02637525461614132, + 0.009100544266402721, + 0.01604962907731533, + 0.004758669063448906, + 0.00556408753618598, + 0.005000136326998472, + 0.010169022716581821, + 0.0032522922847419977, + 0.00676692696288228, + 0.005004862789064646, + 0.004837831482291222, + 0.004328654147684574, + 0.004820242058485746, + 0.0065919868648052216, + 0.008662065491080284, + 0.0029112729243934155, + 0.0033765300177037716, + 0.003735477337613702, + 0.004090900532901287, + 0.004468243103474379, + 0.006181797478348017, + 0.006798983085900545, + 0.008574122563004494, + 0.0035709941294044256, + 0.005098677705973387, + 0.0036106337793171406, + 0.0031613793689757586, + 0.004208900034427643, + 0.0035481476224958897, + 0.00380619871430099, + 0.006195009686052799, + 0.003419304033741355, + 0.003408583579584956, + 0.004279232118278742, + 0.0021298150531947613, + 0.005497591570019722, + 0.006790020503103733, + 0.006526183802634478, + 0.007026494946330786, + 0.004310673102736473, + 0.004911625292152166, + 0.00468260096386075, + 0.0035220591817051172, + 0.0167233943939209, + 0.014266209676861763, + 0.022498179227113724, + 0.01416951883584261, + 0.007471322547644377, + 0.006143085192888975, + 0.006874144542962313, + 0.007239866070449352, + 0.013795718550682068, + 0.009526905603706837, + 0.021151741966605186, + 0.009052397683262825, + 0.005364593118429184, + 0.004983358085155487, + 0.007018120493739843, + 0.0056677935644984245, + 0.006654030177742243, + 0.009377934038639069, + 0.005281058605760336, + 0.005491824820637703, + 0.005537398159503937, + 0.004003694746643305, + 0.006835937034338713, + 0.006516453810036182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.020583825185894966 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a traditional Chinese ink painting with black, white, and gray tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elder with long silvery hair and a flowing beard wearing a pure white silk robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elder standing on a flat, moss-covered rock jutting from a cliff face in a mist-shrouded valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elder performing slow, deliberate Tai Chi movements with his arms drawing steady curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow close-up viewed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026352116838097572 + ], + "opposite_logprob_tokens": [ + -3.6513521671295166 + ], + "logprob_reward": -0.026352116838097572, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026352116838097572, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11676210910081863, + 0.049154624342918396, + 0.037774551659822464, + 0.054909735918045044, + 0.04590173438191414, + 0.04168906435370445, + 0.10018715262413025, + 0.1506200134754181, + 0.26468977332115173, + 0.13831131160259247 + ], + "credit_map": [ + 0.01686040125787258, + 0.01644301787018776, + 0.018725069239735603, + 0.02175399847328663, + 0.01337291207164526, + 0.009372856467962265, + 0.009509921073913574, + 0.01072392612695694, + 0.006816925946623087, + 0.006183028221130371, + 0.01178441196680069, + 0.009201889857649803, + 0.004199039191007614, + 0.0040742200799286366, + 0.0034762986470013857, + 0.0034188071731477976, + 0.004075413569808006, + 0.004622314125299454, + 0.004613665863871574, + 0.007377848494797945, + 0.0034724147990345955, + 0.005628423765301704, + 0.004293127916753292, + 0.003691344289109111, + 0.007659544702619314, + 0.0063099609687924385, + 0.01206329558044672, + 0.005852352827787399, + 0.006340457126498222, + 0.004735610447824001, + 0.008270453661680222, + 0.003678064327687025, + 0.005238933023065329, + 0.005695618223398924, + 0.005393527448177338, + 0.007224945817142725, + 0.0078003439120948315, + 0.0034479605965316296, + 0.007581979501992464, + 0.003518424928188324, + 0.004789516795426607, + 0.005938409361988306, + 0.006650274153798819, + 0.0071301269344985485, + 0.004302073270082474, + 0.004002424888312817, + 0.00604757247492671, + 0.002828662982210517, + 0.01610204204916954, + 0.015180153772234917, + 0.01840967684984207, + 0.01624225080013275, + 0.007403487805277109, + 0.0069610546343028545, + 0.013306344859302044, + 0.006582141388207674, + 0.01726396754384041, + 0.025160836055874825, + 0.017089542001485825, + 0.01990196667611599, + 0.018958037719130516, + 0.016657162457704544, + 0.01894429326057434, + 0.01664421334862709, + 0.03527430444955826, + 0.02942131832242012, + 0.02951819635927677, + 0.04462684318423271, + 0.026409123092889786, + 0.04124715179204941, + 0.03340351954102516, + 0.02478930912911892, + 0.021381564438343048, + 0.01208754163235426, + 0.017728248611092567, + 0.024913229048252106, + 0.017435763031244278, + 0.014145473949611187, + 0.019823675975203514, + 0.010795816779136658 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.026352116838097572 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a traditional Chinese ink painting with black, white, and gray tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elder with long silvery hair and a flowing beard wearing a pure white silk robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elder standing on a flat, moss-covered rock jutting from a cliff face in a mist-shrouded valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elder performing slow, deliberate Tai Chi movements with his arms drawing steady curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow close-up viewed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898558676242828 + ], + "opposite_logprob_tokens": [ + -2.5789856910705566 + ], + "logprob_reward": -0.07898558676242828, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898558676242828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15206190943717957, + 0.08716605603694916, + 0.10143692046403885, + 0.05068100616335869, + 0.049261972308158875, + 0.03421333059668541, + 0.0810982882976532, + 0.1828199326992035, + 0.1631396859884262, + 0.09812091290950775 + ], + "credit_map": [ + 0.015811830759048462, + 0.026723267510533333, + 0.02181410603225231, + 0.03548477962613106, + 0.01413075253367424, + 0.015580807812511921, + 0.010855836793780327, + 0.011660520918667316, + 0.013807626441121101, + 0.011725297197699547, + 0.022003816440701485, + 0.01011496502906084, + 0.006509257480502129, + 0.00677880784496665, + 0.008425027132034302, + 0.0078012580052018166, + 0.014627999626100063, + 0.012315480969846249, + 0.020199675112962723, + 0.014092321507632732, + 0.008919530548155308, + 0.013985357247292995, + 0.009067532606422901, + 0.008229016326367855, + 0.00488056056201458, + 0.006308783311396837, + 0.006226377096027136, + 0.008089092560112476, + 0.006477152928709984, + 0.0067727044224739075, + 0.006295164115726948, + 0.005631172563880682, + 0.006731728091835976, + 0.00655521871522069, + 0.009894067421555519, + 0.005969102960079908, + 0.0028170612640678883, + 0.004002502653747797, + 0.004579216707497835, + 0.008713075891137123, + 0.005695497151464224, + 0.004559650085866451, + 0.008247221820056438, + 0.005680473521351814, + 0.002749019768089056, + 0.0027007723692804575, + 0.0025910832919180393, + 0.0019896128214895725, + 0.014417297206819057, + 0.012006266973912716, + 0.01936119794845581, + 0.008000113070011139, + 0.005948162637650967, + 0.006359403021633625, + 0.010219010524451733, + 0.004786841105669737, + 0.028756864368915558, + 0.02190021425485611, + 0.03986911103129387, + 0.02278224006295204, + 0.021873803809285164, + 0.012891310267150402, + 0.023516204208135605, + 0.011230189353227615, + 0.01905904710292816, + 0.02155132219195366, + 0.016332415863871574, + 0.019577931612730026, + 0.023744169622659683, + 0.019816622138023376, + 0.024223238229751587, + 0.01883494295179844, + 0.015145804733037949, + 0.010630610398948193, + 0.013170153833925724, + 0.01344524510204792, + 0.016532938927412033, + 0.005652578081935644, + 0.01452442817389965, + 0.009019151329994202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.07898558676242828 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, close-up shot, viewed from a slightly upward angle, captures a scene imbued with the artistic conception of a traditional Chinese ink painting. In a deep, mist-shrouded valley, the outlines of distant peaks loom in a gray haze of varying shades, like smudges of ink on rice paper. An elder, resembling a sage, stands alone on a flat, moss-covered rock jutting from the cliff face. He wears a pure white silk robe with wide sleeves, his long, silvery hair and flowing beard gently swaying in the mountain breeze. With his eyes slightly closed, his expression serene and focused, he slowly and effortlessly practices Tai Chi with the rhythmic grace of flowing water. His arms draw steady and deliberate curves, his body gently rotating in response, each movement imbued with power and harmony. The wide sleeves and hem of his robe gently rise and fall with his movements and the mountain breeze, a gesture of exquisite grace. Soft diffuse light penetrates the mist, creating a tranquil and transcendent atmosphere for the entire scene. The main colors of the picture are elegant black, white and gray, with only moss and the dark green of ancient pines in the distance dotted in between, creating a sense of depth and tranquility of harmony between man and nature.", + "seed": 6240, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_6/index_773_seed_6240_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.11868666969239712, + "vlm_logprob": -0.11868666969239712, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.11868666969239712, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a traditional Chinese ink painting with black, white, and gray tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elder with long silvery hair and a flowing beard wearing a pure white silk robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elder standing on a flat, moss-covered rock jutting from a cliff face in a mist-shrouded valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elder performing slow, deliberate Tai Chi movements with his arms drawing steady curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow close-up viewed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3152291178703308 + ], + "opposite_logprob_tokens": [ + -1.3152291774749756 + ], + "logprob_reward": -0.3152291178703308, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3152291178703308, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26012441515922546, + 0.07027409225702286, + 0.1470252126455307, + 0.05241496488451958, + 0.1378876119852066, + 0.0830288827419281, + 0.04786377027630806, + 0.0766800120472908, + 0.05903278663754463, + 0.06566822528839111 + ], + "credit_map": [ + 0.022699136286973953, + 0.023093877360224724, + 0.02865215763449669, + 0.05043574050068855, + 0.02540196105837822, + 0.0363125316798687, + 0.029165560379624367, + 0.044363465160131454, + 0.008244900964200497, + 0.005728178657591343, + 0.008682944811880589, + 0.005207506474107504, + 0.008822512812912464, + 0.016360461711883545, + 0.008392631076276302, + 0.008834960870444775, + 0.017149748280644417, + 0.012730181217193604, + 0.016834411770105362, + 0.01862887106835842, + 0.013637821190059185, + 0.019427737221121788, + 0.018412811681628227, + 0.03020363114774227, + 0.0068567306734621525, + 0.005614036228507757, + 0.005255583673715591, + 0.00715095829218626, + 0.004990062210708857, + 0.005573096685111523, + 0.0061463359743356705, + 0.010828160680830479, + 0.014969159848988056, + 0.010971521027386189, + 0.018005093559622765, + 0.017735904082655907, + 0.024604083970189095, + 0.010192724876105785, + 0.021850498393177986, + 0.019558623433113098, + 0.008052405901253223, + 0.01038467325270176, + 0.010674642398953438, + 0.015795623883605003, + 0.010513152927160263, + 0.007709341589361429, + 0.011190660297870636, + 0.00870838388800621, + 0.004903298337012529, + 0.005377802066504955, + 0.004621859174221754, + 0.009835833683609962, + 0.007677000947296619, + 0.004036727361381054, + 0.007110555190593004, + 0.004300692118704319, + 0.01118084043264389, + 0.010578126646578312, + 0.007305578328669071, + 0.008926085196435452, + 0.0153406523168087, + 0.0074481968767941, + 0.011192080564796925, + 0.004708452150225639, + 0.006467188708484173, + 0.006887157913297415, + 0.006840585730969906, + 0.010388948023319244, + 0.00863695703446865, + 0.007626908831298351, + 0.007245018146932125, + 0.00494002178311348, + 0.00549985421821475, + 0.009631537832319736, + 0.005865379702299833, + 0.006900121457874775, + 0.009629316627979279, + 0.007042589131742716, + 0.014382898807525635, + 0.006716533098369837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.3152291178703308 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a traditional Chinese ink painting with black, white, and gray tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elder with long silvery hair and a flowing beard wearing a pure white silk robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elder standing on a flat, moss-covered rock jutting from a cliff face in a mist-shrouded valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elder performing slow, deliberate Tai Chi movements with his arms drawing steady curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow close-up viewed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006747201085090637 + ], + "opposite_logprob_tokens": [ + -5.006747245788574 + ], + "logprob_reward": -0.006747201085090637, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006747201085090637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30717602372169495, + 0.10468190163373947, + 0.060556039214134216, + 0.0788884162902832, + 0.0613015741109848, + 0.06451846659183502, + 0.07987160980701447, + 0.12772135436534882, + 0.08469235897064209, + 0.030592266470193863 + ], + "credit_map": [ + 0.02738005667924881, + 0.05907776579260826, + 0.03731385990977287, + 0.04930061101913452, + 0.034124962985515594, + 0.03097645379602909, + 0.040410131216049194, + 0.028592193499207497, + 0.010472343303263187, + 0.018023043870925903, + 0.014838818460702896, + 0.011706393212080002, + 0.012229224666953087, + 0.012896910309791565, + 0.014220712706446648, + 0.010294447652995586, + 0.005030325148254633, + 0.007978687062859535, + 0.009745226241648197, + 0.011030657216906548, + 0.0062197791412472725, + 0.005048319231718779, + 0.009179666638374329, + 0.006323374807834625, + 0.006984031293541193, + 0.009177293628454208, + 0.008408496156334877, + 0.01697140373289585, + 0.008739576674997807, + 0.00615955563262105, + 0.013494147919118404, + 0.008953912183642387, + 0.0052785370498895645, + 0.006550242193043232, + 0.007447024807333946, + 0.011361069045960903, + 0.009658370167016983, + 0.007008566986769438, + 0.006556098815053701, + 0.007441666442900896, + 0.007485010661184788, + 0.009162957780063152, + 0.010030006058514118, + 0.008579886518418789, + 0.008600442670285702, + 0.008068528026342392, + 0.007111978251487017, + 0.005479659419506788, + 0.0055572339333593845, + 0.008653338998556137, + 0.005829580593854189, + 0.013149100355803967, + 0.019982412457466125, + 0.009195901453495026, + 0.011445592157542706, + 0.0060584587045013905, + 0.02108289673924446, + 0.015006288886070251, + 0.014537941664457321, + 0.02015889808535576, + 0.020505445078015327, + 0.010993732139468193, + 0.017287202179431915, + 0.00814895424991846, + 0.007830034010112286, + 0.008449013344943523, + 0.008885035291314125, + 0.020025845617055893, + 0.012372292578220367, + 0.008653944358229637, + 0.011551380157470703, + 0.006924808956682682, + 0.0019201015820726752, + 0.003099213121458888, + 0.0028465574141591787, + 0.008443228900432587, + 0.004109105560928583, + 0.0031358960550278425, + 0.004399426747113466, + 0.0026387369725853205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.006747201085090637 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a traditional Chinese ink painting with black, white, and gray tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elder with long silvery hair and a flowing beard wearing a pure white silk robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elder standing on a flat, moss-covered rock jutting from a cliff face in a mist-shrouded valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elder performing slow, deliberate Tai Chi movements with his arms drawing steady curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow close-up viewed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607493683695793 + ], + "opposite_logprob_tokens": [ + -4.141075134277344 + ], + "logprob_reward": -0.01607493683695793, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607493683695793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26926276087760925, + 0.10806993395090103, + 0.0890750139951706, + 0.08094637840986252, + 0.08748374134302139, + 0.070899598300457, + 0.06182087957859039, + 0.09174961596727371, + 0.10538029670715332, + 0.03531177341938019 + ], + "credit_map": [ + 0.027906347066164017, + 0.04344702139496803, + 0.026195477694272995, + 0.03553907573223114, + 0.02996346540749073, + 0.040879204869270325, + 0.02409777045249939, + 0.04123440012335777, + 0.010489558801054955, + 0.013078052550554276, + 0.013049164786934853, + 0.00958422850817442, + 0.011960036121308804, + 0.021771591156721115, + 0.014486843720078468, + 0.013650448992848396, + 0.010357523337006569, + 0.008930441923439503, + 0.012576074339449406, + 0.009387415833771229, + 0.01083964854478836, + 0.009567854925990105, + 0.015111328102648258, + 0.012304726056754589, + 0.009506622329354286, + 0.010116461664438248, + 0.008561838418245316, + 0.01620272919535637, + 0.008775683119893074, + 0.006307459436357021, + 0.011380977928638458, + 0.010094600729644299, + 0.007971533574163914, + 0.00924256443977356, + 0.009084681048989296, + 0.015660852193832397, + 0.015219669789075851, + 0.009067399427294731, + 0.010504142381250858, + 0.010732901282608509, + 0.006805585231631994, + 0.008636697195470333, + 0.009409734047949314, + 0.007838286459445953, + 0.013072176836431026, + 0.009682096540927887, + 0.010040496475994587, + 0.0054145194590091705, + 0.004960326943546534, + 0.006389846559613943, + 0.005815800745040178, + 0.00818115659058094, + 0.015384628437459469, + 0.006111512426286936, + 0.009403862059116364, + 0.00557375093922019, + 0.01308965589851141, + 0.01203971728682518, + 0.01079632993787527, + 0.02137112244963646, + 0.010845969431102276, + 0.005737702362239361, + 0.011976111680269241, + 0.005893002729862928, + 0.010612288489937782, + 0.009275882504880428, + 0.011289235204458237, + 0.018937500193715096, + 0.01948309689760208, + 0.010676994919776917, + 0.018075209110975266, + 0.007030095439404249, + 0.002638800535351038, + 0.003235558047890663, + 0.003546205349266529, + 0.009095404297113419, + 0.0055952295660972595, + 0.0032535232603549957, + 0.004775229375809431, + 0.003171823685988784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.01607493683695793 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a traditional Chinese ink painting with black, white, and gray tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elder with long silvery hair and a flowing beard wearing a pure white silk robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elder standing on a flat, moss-covered rock jutting from a cliff face in a mist-shrouded valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elder performing slow, deliberate Tai Chi movements with his arms drawing steady curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow close-up viewed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029801445081830025 + ], + "opposite_logprob_tokens": [ + -3.529801368713379 + ], + "logprob_reward": -0.029801445081830025, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029801445081830025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2588632106781006, + 0.08604993671178818, + 0.06640013307332993, + 0.07029744982719421, + 0.06271258741617203, + 0.17713508009910583, + 0.03798018768429756, + 0.08649279922246933, + 0.12441295385360718, + 0.02965562976896763 + ], + "credit_map": [ + 0.02238226868212223, + 0.027634240686893463, + 0.02874291129410267, + 0.0437781848013401, + 0.027076480910182, + 0.03846457973122597, + 0.02685210295021534, + 0.043932441622018814, + 0.007567061576992273, + 0.015530455857515335, + 0.00928290281444788, + 0.01155396644026041, + 0.008499493822455406, + 0.012712436728179455, + 0.01031236071139574, + 0.010591262020170689, + 0.006230633705854416, + 0.006823791656643152, + 0.008811174891889095, + 0.014632083475589752, + 0.007551147602498531, + 0.007835384458303452, + 0.009603393264114857, + 0.004912519361823797, + 0.007277097553014755, + 0.006652471609413624, + 0.005535999312996864, + 0.014847097918391228, + 0.009545106440782547, + 0.006432835012674332, + 0.012050211429595947, + 0.00795662496238947, + 0.006371796131134033, + 0.005914683453738689, + 0.007371081970632076, + 0.012944504618644714, + 0.009941993281245232, + 0.005939982365816832, + 0.006625790148973465, + 0.0076027559116482735, + 0.01559378020465374, + 0.017485691234469414, + 0.021523136645555496, + 0.05954601615667343, + 0.018246738240122795, + 0.012262793257832527, + 0.022150686010718346, + 0.010326249524950981, + 0.004293872509151697, + 0.004363523330539465, + 0.005208710208535194, + 0.007121379487216473, + 0.004692952148616314, + 0.003840881399810314, + 0.004763272125273943, + 0.003695594146847725, + 0.013366559520363808, + 0.008916467428207397, + 0.010221651755273342, + 0.013212553225457668, + 0.014835872687399387, + 0.0075518898665905, + 0.01277497224509716, + 0.005612833891063929, + 0.011305759660899639, + 0.01174309104681015, + 0.013646366074681282, + 0.025046000257134438, + 0.021515842527151108, + 0.01383076049387455, + 0.016673533245921135, + 0.0106516033411026, + 0.0022477111779153347, + 0.003048835089430213, + 0.002560841618105769, + 0.007070203777402639, + 0.004221245180815458, + 0.0028471641708165407, + 0.005443778354674578, + 0.0022158499341458082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.029801445081830025 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a traditional Chinese ink painting with black, white, and gray tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elder with long silvery hair and a flowing beard wearing a pure white silk robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elder standing on a flat, moss-covered rock jutting from a cliff face in a mist-shrouded valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elder performing slow, deliberate Tai Chi movements with his arms drawing steady curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow close-up viewed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22558064758777618 + ], + "opposite_logprob_tokens": [ + -1.6005806922912598 + ], + "logprob_reward": -0.22558064758777618, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22558064758777618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13788114488124847, + 0.06172072887420654, + 0.1030285656452179, + 0.21583472192287445, + 0.03902534395456314, + 0.09022615104913712, + 0.14307914674282074, + 0.08565520495176315, + 0.10807344317436218, + 0.015475600026547909 + ], + "credit_map": [ + 0.016687192022800446, + 0.018890084698796272, + 0.012988029047846794, + 0.01688959449529648, + 0.016603048890829086, + 0.018549466505646706, + 0.009169140830636024, + 0.02810458466410637, + 0.006376404780894518, + 0.011536240577697754, + 0.0059301541186869144, + 0.00881604477763176, + 0.007477174047380686, + 0.008638855069875717, + 0.006627349648624659, + 0.006318497937172651, + 0.010074090212583542, + 0.010549474507570267, + 0.010525817051529884, + 0.01838066801428795, + 0.010663735680282116, + 0.015560870058834553, + 0.010937172919511795, + 0.016336726024746895, + 0.020698752254247665, + 0.0241683442145586, + 0.015944549813866615, + 0.04773765802383423, + 0.027664612978696823, + 0.013304694555699825, + 0.0433570072054863, + 0.02295907959342003, + 0.004254660103470087, + 0.00399802764877677, + 0.0046677072532474995, + 0.008126907981932163, + 0.0038872226141393185, + 0.003929594997316599, + 0.004138215910643339, + 0.006023004651069641, + 0.007757548708468676, + 0.00810716487467289, + 0.011909103021025658, + 0.027441661804914474, + 0.009446647949516773, + 0.007220184430480003, + 0.010587753728032112, + 0.007756083272397518, + 0.013190047815442085, + 0.015290122479200363, + 0.012843392789363861, + 0.02475329302251339, + 0.03245559334754944, + 0.014121966436505318, + 0.02099670097231865, + 0.009428021498024464, + 0.010916408151388168, + 0.01229711715131998, + 0.013422535732388496, + 0.014759398996829987, + 0.010328754782676697, + 0.006781045347452164, + 0.01045229285955429, + 0.00669764494523406, + 0.012493210844695568, + 0.010962344706058502, + 0.01134786382317543, + 0.016661211848258972, + 0.022989848628640175, + 0.010575056076049805, + 0.015487085096538067, + 0.007556811440736055, + 0.0011641117744147778, + 0.0014882257673889399, + 0.0021112633403390646, + 0.0036307168193161488, + 0.002171212574467063, + 0.0014881823444738984, + 0.0021327913273125887, + 0.0012890944490209222 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.22558064758777618 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_7/index_tensor([3572, 3572])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_7/index_tensor([3572, 3572])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f7fa5a1bcbe6d4352f5e5aa4ae145d182804897d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_7/index_tensor([3572, 3572])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, medium shot, shot from a slightly upward angle, draws the viewer's attention to a mysterious hacker in a dimly lit room. He wears a black, high-tech jacket with a massive hood, its deep shadow almost completely obscuring his features, only occasionally revealing his resolute jawline in the cold glow of the holographic projection. His hands, clad in sleek black fingerless gloves, manipulate a complex, three-dimensional holographic DJ booth suspended before him with precise and rhythmic control. This virtual device, composed of glowing geometric shapes, pulsating sound waves, and data streams, shifts in real time with his swipes, pinches, and taps, leaving fleeting afterimages behind his fingertips. The background is a tangle of exposed pipes and cables, vine-like, creeping across the walls. In the distance, LED indicators on server racks flicker silently in the darkness. The main cold-toned light comes from the holographic projection itself, casting a sharp light on the hacker, while the magenta and cyan neon lights from the giant billboards outside the window in the eternal rainy night slowly flow on the reflective material of his high-tech jacket, forming dynamic light and shadow. The whole scene is filled with a highly focused, futuristic and slightly decadent cyberpunk atmosphere.", + "seed": 3089, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_7/index_3572_seed_3089_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.17766975611448288, + "vlm_logprob": -0.17766975611448288, + "vlm_yes_no_margin": 2.050000011920929, + "vlm_token_credit": -0.17766975611448288, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hacker wearing a black high-tech jacket with a massive hood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hacker manipulating a complex three-dimensional holographic DJ booth with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background filled with exposed pipes and cables that look like vines creeping across the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dimly lit room with flickering LED indicators on server racks in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle in a medium shot framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25463470816612244 + ], + "opposite_logprob_tokens": [ + -1.5046347379684448 + ], + "logprob_reward": -0.25463470816612244, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25463470816612244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20273950695991516, + 0.12377765029668808, + 0.08133266121149063, + 0.0671834871172905, + 0.12502142786979675, + 0.22269272804260254, + 0.05775384232401848, + 0.01938379555940628, + 0.04363749548792839, + 0.0564773753285408 + ], + "credit_map": [ + 0.026541652157902718, + 0.022353840991854668, + 0.03461357206106186, + 0.023589350283145905, + 0.022276563569903374, + 0.018474062904715538, + 0.020569657906889915, + 0.03432081639766693, + 0.012429066002368927, + 0.01422541681677103, + 0.016089478507637978, + 0.024822505190968513, + 0.012902055867016315, + 0.015642058104276657, + 0.01567644625902176, + 0.011990615166723728, + 0.005678202025592327, + 0.013638819567859173, + 0.018938004970550537, + 0.0064658839255571365, + 0.009050529450178146, + 0.00933698657900095, + 0.010288232006132603, + 0.007936006411910057, + 0.0129063306376338, + 0.00421508401632309, + 0.006725216284394264, + 0.003768199821934104, + 0.005961726885288954, + 0.007449829950928688, + 0.019136585295200348, + 0.0070205130614340305, + 0.04485350847244263, + 0.03977162018418312, + 0.01430552452802658, + 0.005678410641849041, + 0.004382025916129351, + 0.003114778781309724, + 0.008205597288906574, + 0.004709962289780378, + 0.06133444234728813, + 0.03448204696178436, + 0.04691515490412712, + 0.01604459062218666, + 0.01273325178772211, + 0.016775354743003845, + 0.02314073219895363, + 0.011267143301665783, + 0.006654710974544287, + 0.008733577094972134, + 0.008158492855727673, + 0.006628539878875017, + 0.003953170031309128, + 0.005766598042100668, + 0.010888241231441498, + 0.006970510818064213, + 0.003016810864210129, + 0.0029894153121858835, + 0.002873287070542574, + 0.001751640229485929, + 0.0014310500118881464, + 0.0029130587354302406, + 0.002265513874590397, + 0.0021430193446576595, + 0.005509198643267155, + 0.007219902239739895, + 0.0062260739505290985, + 0.005493672098964453, + 0.005225959233939648, + 0.004048788454383612, + 0.005850867833942175, + 0.004063029307872057, + 0.007400442846119404, + 0.0054252902045845985, + 0.008258449845016003, + 0.007921489886939526, + 0.005623293109238148, + 0.0045088473707437515, + 0.011242094449698925, + 0.006097470410168171 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.25463470816612244 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hacker wearing a black high-tech jacket with a massive hood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hacker manipulating a complex three-dimensional holographic DJ booth with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background filled with exposed pipes and cables that look like vines creeping across the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dimly lit room with flickering LED indicators on server racks in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle in a medium shot framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42894071340560913 + ], + "opposite_logprob_tokens": [ + -1.053940773010254 + ], + "logprob_reward": -0.42894071340560913, + "yes_no_margin_reward": 0.6250000596046448, + "token_credit_reward": -0.42894071340560913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0889548510313034, + 0.051045000553131104, + 0.04196283221244812, + 0.13012109696865082, + 0.27982959151268005, + 0.21855652332305908, + 0.03814518824219704, + 0.014884479343891144, + 0.05523715913295746, + 0.0812632292509079 + ], + "credit_map": [ + 0.016690600663423538, + 0.018370142206549644, + 0.012910546734929085, + 0.008930220268666744, + 0.006868619937449694, + 0.006662094034254551, + 0.009117158129811287, + 0.009405468590557575, + 0.005556360352784395, + 0.00767541816458106, + 0.006499275099486113, + 0.012161361053586006, + 0.0041678063571453094, + 0.004388702567666769, + 0.005353433545678854, + 0.00524264108389616, + 0.005062495823949575, + 0.006555137690156698, + 0.00614594342187047, + 0.00430797366425395, + 0.0037295734509825706, + 0.006006231065839529, + 0.004821496549993753, + 0.005333981476724148, + 0.019913626834750175, + 0.01426197960972786, + 0.013932277448475361, + 0.013638263568282127, + 0.02301056869328022, + 0.015080593526363373, + 0.018915411084890366, + 0.011368374340236187, + 0.09345007687807083, + 0.08772054314613342, + 0.02959620952606201, + 0.01258161198347807, + 0.01485377550125122, + 0.008219122886657715, + 0.022092361003160477, + 0.011315906420350075, + 0.07920532673597336, + 0.0649222657084465, + 0.02494240738451481, + 0.010813470929861069, + 0.008891420438885689, + 0.008479115553200245, + 0.013442924246191978, + 0.007859595119953156, + 0.005495763849467039, + 0.008764320984482765, + 0.005091843195259571, + 0.0029782303608953953, + 0.0023246905766427517, + 0.005332144442945719, + 0.004378707613795996, + 0.003779485123232007, + 0.002236118307337165, + 0.0018465900793671608, + 0.002313415752723813, + 0.001394915976561606, + 0.0013475068844854832, + 0.0015383076388388872, + 0.0024880398996174335, + 0.0017195857362821698, + 0.00491366209462285, + 0.006143295206129551, + 0.00868828035891056, + 0.009233090095221996, + 0.0049629807472229, + 0.006649920716881752, + 0.00901404395699501, + 0.005631887819617987, + 0.01292665209621191, + 0.008847955614328384, + 0.012395722791552544, + 0.012142844498157501, + 0.008535369299352169, + 0.006625341717153788, + 0.013445786200463772, + 0.006343556102365255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.42894071340560913 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hacker wearing a black high-tech jacket with a massive hood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hacker manipulating a complex three-dimensional holographic DJ booth with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background filled with exposed pipes and cables that look like vines creeping across the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dimly lit room with flickering LED indicators on server racks in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle in a medium shot framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293117702007294 + ], + "opposite_logprob_tokens": [ + -2.237931251525879 + ], + "logprob_reward": -0.11293117702007294, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293117702007294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1569487303495407, + 0.09844426065683365, + 0.06124668940901756, + 0.08618850260972977, + 0.2169051468372345, + 0.1085788756608963, + 0.03759736940264702, + 0.025493836030364037, + 0.048327527940273285, + 0.160269096493721 + ], + "credit_map": [ + 0.02809896692633629, + 0.02185267210006714, + 0.022565383464097977, + 0.018616940826177597, + 0.019362689927220345, + 0.01203656755387783, + 0.016064563766121864, + 0.018350958824157715, + 0.017677703872323036, + 0.01804867386817932, + 0.01648562401533127, + 0.013789394870400429, + 0.005363603122532368, + 0.006378747522830963, + 0.009139038622379303, + 0.01156147662550211, + 0.011497276835143566, + 0.007689660415053368, + 0.009280298836529255, + 0.011272981762886047, + 0.004662603605538607, + 0.005371992010623217, + 0.005817752797156572, + 0.00565411988645792, + 0.01040654443204403, + 0.012740747071802616, + 0.01476739440113306, + 0.013591145165264606, + 0.007273765746504068, + 0.007687101140618324, + 0.01271022018045187, + 0.007011584937572479, + 0.0761336162686348, + 0.07239881157875061, + 0.02140168845653534, + 0.00924594048410654, + 0.006976178381592035, + 0.006340398453176022, + 0.015030940063297749, + 0.009377590380609035, + 0.02622336708009243, + 0.032653700560331345, + 0.01537367608398199, + 0.00616720225661993, + 0.005809444468468428, + 0.006178915500640869, + 0.01036499347537756, + 0.005807583220303059, + 0.005630042403936386, + 0.0074213361367583275, + 0.004228152800351381, + 0.004595461767166853, + 0.002465779660269618, + 0.003449441632255912, + 0.005719972774386406, + 0.004087182693183422, + 0.004368811845779419, + 0.006377566140145063, + 0.0038768043741583824, + 0.002373236231505871, + 0.001465102657675743, + 0.0016968236304819584, + 0.003476737067103386, + 0.0018587536178529263, + 0.005344904027879238, + 0.006740344688296318, + 0.007318408694118261, + 0.008932492695748806, + 0.005037498194724321, + 0.0044763716869056225, + 0.005178343504667282, + 0.00529916537925601, + 0.025431742891669273, + 0.01785469427704811, + 0.020640820264816284, + 0.02994818240404129, + 0.013559795916080475, + 0.014052586629986763, + 0.027370305731892586, + 0.011410964652895927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.11293117702007294 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hacker wearing a black high-tech jacket with a massive hood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hacker manipulating a complex three-dimensional holographic DJ booth with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background filled with exposed pipes and cables that look like vines creeping across the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dimly lit room with flickering LED indicators on server racks in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle in a medium shot framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204559653997421 + ], + "opposite_logprob_tokens": [ + -2.8120455741882324 + ], + "logprob_reward": -0.06204559653997421, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204559653997421, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08565443009138107, + 0.10096807777881622, + 0.18521295487880707, + 0.05195002257823944, + 0.25023558735847473, + 0.06179007887840271, + 0.08577818423509598, + 0.07832901924848557, + 0.030819734558463097, + 0.06926185637712479 + ], + "credit_map": [ + 0.01581890508532524, + 0.016474958509206772, + 0.012117774225771427, + 0.008883249945938587, + 0.011841956526041031, + 0.007286376319825649, + 0.006711803376674652, + 0.006519405636936426, + 0.016485348343849182, + 0.016344208270311356, + 0.018378987908363342, + 0.01382587943226099, + 0.0062608071602880955, + 0.008646831847727299, + 0.010034308768808842, + 0.010991713032126427, + 0.02329728752374649, + 0.03503044694662094, + 0.05310704559087753, + 0.0241924449801445, + 0.010238624177873135, + 0.012631080113351345, + 0.013703456148505211, + 0.013012563809752464, + 0.006500049959868193, + 0.005429759155958891, + 0.005404654890298843, + 0.0036047901958227158, + 0.007963497191667557, + 0.006065021269023418, + 0.01165314856916666, + 0.005329101346433163, + 0.08475115150213242, + 0.0803692564368248, + 0.026254285126924515, + 0.011356815695762634, + 0.01300263311713934, + 0.009247121401131153, + 0.016174541786313057, + 0.009079794399440289, + 0.011002575047314167, + 0.015076922252774239, + 0.010425692424178123, + 0.00532272644340992, + 0.00489344447851181, + 0.006152908317744732, + 0.004921038635075092, + 0.003994773142039776, + 0.013068933971226215, + 0.02345576323568821, + 0.012065471149981022, + 0.005805957596749067, + 0.005732393357902765, + 0.01065045315772295, + 0.00705779530107975, + 0.007941409945487976, + 0.010211174376308918, + 0.014782321639358997, + 0.013243640772998333, + 0.008556799031794071, + 0.008369862101972103, + 0.007418405264616013, + 0.009438950568437576, + 0.006307860836386681, + 0.0034899828024208546, + 0.005442631896585226, + 0.0039945365861058235, + 0.004458236042410135, + 0.003404516726732254, + 0.003249170957133174, + 0.0034091274719685316, + 0.003371533704921603, + 0.008885731920599937, + 0.007761272601783276, + 0.00930777657777071, + 0.011118362657725811, + 0.006253569386899471, + 0.00761208264157176, + 0.011741678230464458, + 0.006581385154277086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.06204559653997421 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hacker wearing a black high-tech jacket with a massive hood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hacker manipulating a complex three-dimensional holographic DJ booth with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background filled with exposed pipes and cables that look like vines creeping across the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dimly lit room with flickering LED indicators on server racks in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle in a medium shot framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979658544063568 + ], + "opposite_logprob_tokens": [ + -3.529796600341797 + ], + "logprob_reward": -0.02979658544063568, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979658544063568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16599592566490173, + 0.1222943365573883, + 0.04765535891056061, + 0.07231953740119934, + 0.05729673430323601, + 0.2987546920776367, + 0.10297111421823502, + 0.06394080072641373, + 0.03300463780760765, + 0.03576694056391716 + ], + "credit_map": [ + 0.0205756276845932, + 0.023434827104210854, + 0.0248232651501894, + 0.020969271659851074, + 0.019485879689455032, + 0.016242090612649918, + 0.01646740362048149, + 0.023997558280825615, + 0.015009392984211445, + 0.016335364431142807, + 0.027419084683060646, + 0.014173186384141445, + 0.010729349218308926, + 0.011597736738622189, + 0.01386339496821165, + 0.013166822493076324, + 0.004838109947741032, + 0.006689722184091806, + 0.006596478633582592, + 0.00547427823767066, + 0.005868665874004364, + 0.005563328508287668, + 0.006743447855114937, + 0.0058813272044062614, + 0.008117618970572948, + 0.009734723716974258, + 0.010007145814597607, + 0.0076042041182518005, + 0.014158016070723534, + 0.007310414686799049, + 0.009240524843335152, + 0.006146889179944992, + 0.008674108423292637, + 0.008447553962469101, + 0.0100298672914505, + 0.004866724833846092, + 0.00525572570040822, + 0.004497564863413572, + 0.010005109012126923, + 0.0055200811475515366, + 0.10661903768777847, + 0.09460020810365677, + 0.03365037590265274, + 0.013935395516455173, + 0.009597184136509895, + 0.013777488842606544, + 0.013367288745939732, + 0.013207713142037392, + 0.010004816576838493, + 0.017505185678601265, + 0.014262977987527847, + 0.009610417298972607, + 0.005943892523646355, + 0.012612342834472656, + 0.01784413680434227, + 0.015187343582510948, + 0.008285670541226864, + 0.017326481640338898, + 0.007817288860678673, + 0.004909926559776068, + 0.004457340110093355, + 0.007244758307933807, + 0.008474025875329971, + 0.005425306037068367, + 0.004636261612176895, + 0.004717640113085508, + 0.005385103169828653, + 0.004231438972055912, + 0.0032322718761861324, + 0.0040234667249023914, + 0.0034319297410547733, + 0.0033465276937931776, + 0.0033633990678936243, + 0.003800096921622753, + 0.005495443474501371, + 0.007575681433081627, + 0.002553361002355814, + 0.003329262603074312, + 0.0048909238539636135, + 0.004758771508932114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.02979658544063568 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, medium shot, shot from a slightly upward angle, draws the viewer's attention to a mysterious hacker in a dimly lit room. He wears a black, high-tech jacket with a massive hood, its deep shadow almost completely obscuring his features, only occasionally revealing his resolute jawline in the cold glow of the holographic projection. His hands, clad in sleek black fingerless gloves, manipulate a complex, three-dimensional holographic DJ booth suspended before him with precise and rhythmic control. This virtual device, composed of glowing geometric shapes, pulsating sound waves, and data streams, shifts in real time with his swipes, pinches, and taps, leaving fleeting afterimages behind his fingertips. The background is a tangle of exposed pipes and cables, vine-like, creeping across the walls. In the distance, LED indicators on server racks flicker silently in the darkness. The main cold-toned light comes from the holographic projection itself, casting a sharp light on the hacker, while the magenta and cyan neon lights from the giant billboards outside the window in the eternal rainy night slowly flow on the reflective material of his high-tech jacket, forming dynamic light and shadow. The whole scene is filled with a highly focused, futuristic and slightly decadent cyberpunk atmosphere.", + "seed": 3090, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_7/index_3572_seed_3090_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1781411524862051, + "vlm_logprob": -0.1781411524862051, + "vlm_yes_no_margin": 2.325, + "vlm_token_credit": -0.1781411524862051, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hacker wearing a black high-tech jacket with a massive hood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hacker manipulating a complex three-dimensional holographic DJ booth with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background filled with exposed pipes and cables that look like vines creeping across the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dimly lit room with flickering LED indicators on server racks in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle in a medium shot framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08144234120845795 + ], + "opposite_logprob_tokens": [ + -2.581442356109619 + ], + "logprob_reward": -0.08144234120845795, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08144234120845795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3686618208885193, + 0.07917194813489914, + 0.16069132089614868, + 0.035452231764793396, + 0.049912065267562866, + 0.02996484935283661, + 0.027364082634449005, + 0.10665849596261978, + 0.10482229292392731, + 0.03730088844895363 + ], + "credit_map": [ + 0.03633659705519676, + 0.06588615477085114, + 0.0652577206492424, + 0.04571659117937088, + 0.027373889461159706, + 0.022980744019150734, + 0.052648596465587616, + 0.05246153846383095, + 0.006425085943192244, + 0.017040682956576347, + 0.007983486168086529, + 0.00821931567043066, + 0.008799081668257713, + 0.009485555812716484, + 0.009156197309494019, + 0.012062550522387028, + 0.020674219354987144, + 0.038365960121154785, + 0.01743505522608757, + 0.026534350588917732, + 0.010770602151751518, + 0.019134696573019028, + 0.014490197412669659, + 0.013286235742270947, + 0.005097664892673492, + 0.004015614278614521, + 0.006145560182631016, + 0.0037871424574404955, + 0.002497716573998332, + 0.0032335896976292133, + 0.005791564006358385, + 0.004883380141109228, + 0.007068102713674307, + 0.0029654642567038536, + 0.0038556442596018314, + 0.012191197834908962, + 0.003657638793811202, + 0.011612361297011375, + 0.0036517754197120667, + 0.004909884184598923, + 0.004485272336751223, + 0.003610563464462757, + 0.0033844306599348783, + 0.004124171566218138, + 0.0027656282763928175, + 0.003408789401873946, + 0.0041107977740466595, + 0.004075198899954557, + 0.0038689379580318928, + 0.0033158024307340384, + 0.0035241490695625544, + 0.003934122156351805, + 0.0035976695362478495, + 0.0038388047832995653, + 0.0029960847459733486, + 0.0022885154467076063, + 0.010310609824955463, + 0.006473074201494455, + 0.004900170955806971, + 0.02819022350013256, + 0.007759730331599712, + 0.03491374850273132, + 0.005309530533850193, + 0.008801409974694252, + 0.009383762255311012, + 0.009405797347426414, + 0.005552316550165415, + 0.01770600862801075, + 0.02050919085741043, + 0.024146486073732376, + 0.01078653335571289, + 0.007332195527851582, + 0.008280903100967407, + 0.0038285607006400824, + 0.005043690092861652, + 0.004133457783609629, + 0.0026306170038878918, + 0.005270875059068203, + 0.003392024664208293, + 0.004720764700323343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.08144234120845795 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hacker wearing a black high-tech jacket with a massive hood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hacker manipulating a complex three-dimensional holographic DJ booth with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background filled with exposed pipes and cables that look like vines creeping across the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dimly lit room with flickering LED indicators on server racks in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle in a medium shot framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328502893447876 + ], + "opposite_logprob_tokens": [ + -0.7578502893447876 + ], + "logprob_reward": -0.6328502893447876, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328502893447876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20705407857894897, + 0.09067942947149277, + 0.10474079102277756, + 0.09492263197898865, + 0.10170835256576538, + 0.053544480353593826, + 0.037466101348400116, + 0.17397098243236542, + 0.08205743134021759, + 0.05385565757751465 + ], + "credit_map": [ + 0.021974246948957443, + 0.03120993636548519, + 0.027872921898961067, + 0.024089232087135315, + 0.02084089070558548, + 0.019470056518912315, + 0.03638097643852234, + 0.025215819478034973, + 0.00805594865232706, + 0.020879385992884636, + 0.008914877660572529, + 0.01188138872385025, + 0.011485214345157146, + 0.010989849455654621, + 0.01020055916160345, + 0.008272210136055946, + 0.01825343817472458, + 0.01697421446442604, + 0.011713325046002865, + 0.013047563843429089, + 0.007015821989625692, + 0.013030248694121838, + 0.013191931881010532, + 0.011514244601130486, + 0.010220760479569435, + 0.008671257644891739, + 0.010976018384099007, + 0.01003341656178236, + 0.007590268738567829, + 0.011879161931574345, + 0.019004158675670624, + 0.01654759794473648, + 0.008747106418013573, + 0.008587883785367012, + 0.005542422644793987, + 0.03525259718298912, + 0.004987685009837151, + 0.027728848159313202, + 0.0062661855481565, + 0.004595631267875433, + 0.008618989959359169, + 0.004807686898857355, + 0.007149253971874714, + 0.007558750919997692, + 0.004279036074876785, + 0.006080221850425005, + 0.008599224500358105, + 0.006451322231441736, + 0.005366695579141378, + 0.003437320701777935, + 0.005523700267076492, + 0.004133386537432671, + 0.0029016397893428802, + 0.004846001043915749, + 0.0067728860303759575, + 0.004484475590288639, + 0.013614893890917301, + 0.010959801264107227, + 0.007859369739890099, + 0.04196440428495407, + 0.01853785291314125, + 0.05582118034362793, + 0.010546990670263767, + 0.01466650515794754, + 0.006739482749253511, + 0.006527187302708626, + 0.005877222865819931, + 0.015651263296604156, + 0.013973740860819817, + 0.019373269751667976, + 0.008041790686547756, + 0.00587347662076354, + 0.009067540057003498, + 0.006791241466999054, + 0.004093538038432598, + 0.007415944244712591, + 0.006741341669112444, + 0.011236057616770267, + 0.005055256653577089, + 0.0034547445829957724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.6328502893447876 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hacker wearing a black high-tech jacket with a massive hood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hacker manipulating a complex three-dimensional holographic DJ booth with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background filled with exposed pipes and cables that look like vines creeping across the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dimly lit room with flickering LED indicators on server racks in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle in a medium shot framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292190849781036 + ], + "opposite_logprob_tokens": [ + -2.237921953201294 + ], + "logprob_reward": -0.11292190849781036, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292190849781036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19519555568695068, + 0.0910736545920372, + 0.08848360925912857, + 0.12189565598964691, + 0.0881321057677269, + 0.19664663076400757, + 0.06289879977703094, + 0.03702254965901375, + 0.08377578109502792, + 0.034875690937042236 + ], + "credit_map": [ + 0.021633872762322426, + 0.037520989775657654, + 0.029138365760445595, + 0.022482044994831085, + 0.014488373883068562, + 0.022701021283864975, + 0.024805771186947823, + 0.022425122559070587, + 0.007960365153849125, + 0.01280940417200327, + 0.009457460604608059, + 0.009939015842974186, + 0.008108587935566902, + 0.011850150302052498, + 0.008834648877382278, + 0.022114017978310585, + 0.010433022864162922, + 0.014696268364787102, + 0.007599309086799622, + 0.01492004469037056, + 0.0062700496055185795, + 0.012636808678507805, + 0.01251312904059887, + 0.009414980188012123, + 0.015433748252689838, + 0.013062749989330769, + 0.018380766734480858, + 0.014273064211010933, + 0.009571176022291183, + 0.011785056442022324, + 0.016944704577326775, + 0.02244439721107483, + 0.00649971142411232, + 0.008462849073112011, + 0.00477965222671628, + 0.024610847234725952, + 0.006233645603060722, + 0.02102467231452465, + 0.007967062294483185, + 0.008553663268685341, + 0.026951566338539124, + 0.01698625646531582, + 0.024306591600179672, + 0.030670901760458946, + 0.011693122796714306, + 0.03050757199525833, + 0.02553839609026909, + 0.029992224648594856, + 0.01120703388005495, + 0.004467719234526157, + 0.0082330834120512, + 0.006825038697570562, + 0.004726695362478495, + 0.009721019305288792, + 0.009902064688503742, + 0.007816142402589321, + 0.004657726734876633, + 0.0034431188832968473, + 0.0029931303579360247, + 0.0072598764672875404, + 0.005066966637969017, + 0.0067397719249129295, + 0.003283261088654399, + 0.003578696632757783, + 0.0073663885705173016, + 0.008640419691801071, + 0.005040270742028952, + 0.013834362849593163, + 0.015574706718325615, + 0.018085571005940437, + 0.009110833518207073, + 0.006123235449194908, + 0.006540087517350912, + 0.002828823169693351, + 0.0057213446125388145, + 0.004033945966511965, + 0.0038171119522303343, + 0.004995577037334442, + 0.002773849293589592, + 0.004164950922131538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.11292190849781036 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hacker wearing a black high-tech jacket with a massive hood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hacker manipulating a complex three-dimensional holographic DJ booth with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background filled with exposed pipes and cables that look like vines creeping across the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dimly lit room with flickering LED indicators on server racks in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle in a medium shot framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029794966802001 + ], + "opposite_logprob_tokens": [ + -3.529794931411743 + ], + "logprob_reward": -0.029794966802001, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029794966802001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1714283674955368, + 0.05953822284936905, + 0.06533509492874146, + 0.280612587928772, + 0.06452785432338715, + 0.09136001020669937, + 0.0331275500357151, + 0.035937294363975525, + 0.15754090249538422, + 0.04059205204248428 + ], + "credit_map": [ + 0.015518112108111382, + 0.018704714253544807, + 0.015878453850746155, + 0.022004902362823486, + 0.012155403383076191, + 0.02341686189174652, + 0.02347441390156746, + 0.040275514125823975, + 0.006698423996567726, + 0.00970284640789032, + 0.0074328831396996975, + 0.00804547406733036, + 0.005465826950967312, + 0.006575861480087042, + 0.0054566687904298306, + 0.010160239413380623, + 0.007974671199917793, + 0.009926311671733856, + 0.005046191159635782, + 0.010447495616972446, + 0.00808047316968441, + 0.0085929399356246, + 0.006468208506703377, + 0.008798801340162754, + 0.03442343324422836, + 0.027974089607596397, + 0.045068804174661636, + 0.032913319766521454, + 0.022299595177173615, + 0.03898351639509201, + 0.036381762474775314, + 0.042568083852529526, + 0.007215762976557016, + 0.0052353092469275, + 0.003924615681171417, + 0.015420317649841309, + 0.005587006453424692, + 0.014220372773706913, + 0.0058867186307907104, + 0.007037756498903036, + 0.012921277433633804, + 0.008373296819627285, + 0.009520079009234905, + 0.016155049204826355, + 0.00594032509252429, + 0.014443667605519295, + 0.009343050420284271, + 0.014663263224065304, + 0.003471457865089178, + 0.0028410782106220722, + 0.005067501682788134, + 0.004222958814352751, + 0.0026341970078647137, + 0.004384740721434355, + 0.006101294420659542, + 0.004404319915920496, + 0.005139478016644716, + 0.0030493009835481644, + 0.0035184689331799746, + 0.007067237049341202, + 0.0036850799806416035, + 0.007972577586770058, + 0.0021660171914845705, + 0.0033391346223652363, + 0.017223134636878967, + 0.017400702461600304, + 0.011315608397126198, + 0.01996658183634281, + 0.035837747156620026, + 0.022784210741519928, + 0.019958507269620895, + 0.01305441651493311, + 0.006109892390668392, + 0.0026610277127474546, + 0.0034490216057747602, + 0.004064731299877167, + 0.006478023715317249, + 0.0056402175687253475, + 0.004000718239694834, + 0.008188419044017792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.029794966802001 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hacker wearing a black high-tech jacket with a massive hood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hacker manipulating a complex three-dimensional holographic DJ booth with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background filled with exposed pipes and cables that look like vines creeping across the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dimly lit room with flickering LED indicators on server racks in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle in a medium shot framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0336962565779686 + ], + "opposite_logprob_tokens": [ + -3.408696174621582 + ], + "logprob_reward": -0.0336962565779686, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0336962565779686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.213446706533432, + 0.12150807678699493, + 0.07609842717647552, + 0.11603749543428421, + 0.2100398987531662, + 0.14281201362609863, + 0.031182169914245605, + 0.02487832121551037, + 0.025487350299954414, + 0.03850960731506348 + ], + "credit_map": [ + 0.02042211964726448, + 0.02589966543018818, + 0.03646956384181976, + 0.025890236720442772, + 0.02502058818936348, + 0.01896345242857933, + 0.022450922057032585, + 0.03833014890551567, + 0.00954584963619709, + 0.03185541182756424, + 0.015764158219099045, + 0.013839206658303738, + 0.013812113553285599, + 0.01278273481875658, + 0.01194141898304224, + 0.011967168189585209, + 0.008810434490442276, + 0.01946152001619339, + 0.011827528476715088, + 0.01016637496650219, + 0.0056425584480166435, + 0.00625912519171834, + 0.007694921921938658, + 0.006235957611352205, + 0.009328057058155537, + 0.017781144008040428, + 0.020288536325097084, + 0.014738463796675205, + 0.013260486535727978, + 0.010294291190803051, + 0.02133815363049507, + 0.009008354507386684, + 0.015353082679212093, + 0.013185657560825348, + 0.015869002789258957, + 0.0747765600681305, + 0.008249644190073013, + 0.0618249885737896, + 0.010268333368003368, + 0.010512628592550755, + 0.014501916244626045, + 0.020436372607946396, + 0.027914343401789665, + 0.018466297537088394, + 0.013626030646264553, + 0.012994215823709965, + 0.023446816951036453, + 0.011426017619669437, + 0.004920788109302521, + 0.0029544641729444265, + 0.0060111405327916145, + 0.004289970733225346, + 0.002584638074040413, + 0.0035837837494909763, + 0.0041799405589699745, + 0.0026574430521577597, + 0.006411869078874588, + 0.002249679062515497, + 0.002863543340936303, + 0.002524896990507841, + 0.002061912789940834, + 0.0037121528293937445, + 0.001780520542524755, + 0.003273742739111185, + 0.004893355071544647, + 0.0033538665156811476, + 0.0019366191700100899, + 0.002624272368848324, + 0.004312878008931875, + 0.0031378192361444235, + 0.002818714827299118, + 0.0024098202120512724, + 0.00585835101082921, + 0.003849772969260812, + 0.008745793253183365, + 0.004982380662113428, + 0.003916377667337656, + 0.005027203820645809, + 0.002928669797256589, + 0.0032010539434850216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.0336962565779686 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, medium shot, shot from a slightly upward angle, draws the viewer's attention to a mysterious hacker in a dimly lit room. He wears a black, high-tech jacket with a massive hood, its deep shadow almost completely obscuring his features, only occasionally revealing his resolute jawline in the cold glow of the holographic projection. His hands, clad in sleek black fingerless gloves, manipulate a complex, three-dimensional holographic DJ booth suspended before him with precise and rhythmic control. This virtual device, composed of glowing geometric shapes, pulsating sound waves, and data streams, shifts in real time with his swipes, pinches, and taps, leaving fleeting afterimages behind his fingertips. The background is a tangle of exposed pipes and cables, vine-like, creeping across the walls. In the distance, LED indicators on server racks flicker silently in the darkness. The main cold-toned light comes from the holographic projection itself, casting a sharp light on the hacker, while the magenta and cyan neon lights from the giant billboards outside the window in the eternal rainy night slowly flow on the reflective material of his high-tech jacket, forming dynamic light and shadow. The whole scene is filled with a highly focused, futuristic and slightly decadent cyberpunk atmosphere.", + "seed": 3091, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_7/index_3572_seed_3091_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07843343168497086, + "vlm_logprob": -0.07843343168497086, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.07843343168497086, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hacker wearing a black high-tech jacket with a massive hood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hacker manipulating a complex three-dimensional holographic DJ booth with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background filled with exposed pipes and cables that look like vines creeping across the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dimly lit room with flickering LED indicators on server racks in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle in a medium shot framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10429709404706955 + ], + "opposite_logprob_tokens": [ + -2.354297161102295 + ], + "logprob_reward": -0.10429709404706955, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10429709404706955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11848568171262741, + 0.5136117339134216, + 0.17683228850364685, + 0.016780424863100052, + 0.026529377326369286, + 0.022658757865428925, + 0.027612263336777687, + 0.04728417098522186, + 0.026561425998806953, + 0.023643858730793 + ], + "credit_map": [ + 0.010673624463379383, + 0.016811709851026535, + 0.00991073064506054, + 0.018243856728076935, + 0.022925980389118195, + 0.013891338370740414, + 0.015779586508870125, + 0.010248847305774689, + 0.043922893702983856, + 0.08200836926698685, + 0.06763946264982224, + 0.1022217720746994, + 0.10850001871585846, + 0.03689129278063774, + 0.04758823662996292, + 0.024839716032147408, + 0.0206897035241127, + 0.03504807874560356, + 0.02376358024775982, + 0.04623707756400108, + 0.020434530451893806, + 0.008094092831015587, + 0.015767183154821396, + 0.006798042915761471, + 0.0020612101070582867, + 0.0026791165582835674, + 0.0019682468846440315, + 0.0025120992213487625, + 0.0019851878751069307, + 0.002925039501860738, + 0.0014766593230888247, + 0.0011728658573701978, + 0.0038349241949617863, + 0.0024816375225782394, + 0.0034266775473952293, + 0.003285994753241539, + 0.004169930703938007, + 0.005127272102981806, + 0.0016416035359725356, + 0.0025613359175622463, + 0.0034652124159038067, + 0.00288950907997787, + 0.003606614889577031, + 0.004438626579940319, + 0.0016495195450261235, + 0.001945684663951397, + 0.0021588043309748173, + 0.002504785545170307, + 0.0028993512969464064, + 0.003273666137829423, + 0.0046953014098107815, + 0.0027276419568806887, + 0.0031328757759183645, + 0.0031508549582213163, + 0.002607530914247036, + 0.005125041119754314, + 0.006881886161863804, + 0.010056805796921253, + 0.007585440296679735, + 0.004324585665017366, + 0.006770532112568617, + 0.003675402607768774, + 0.0039084493182599545, + 0.004081066697835922, + 0.004000286106020212, + 0.002826861571520567, + 0.004132765345275402, + 0.002663704566657543, + 0.005104266572743654, + 0.0029547670856118202, + 0.0027647018432617188, + 0.0021140750031918287, + 0.0019253614591434598, + 0.003131355158984661, + 0.003310413798317313, + 0.0037362107541412115, + 0.002468518679961562, + 0.003593807341530919, + 0.0031080972403287888, + 0.002370096044614911 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.10429709404706955 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hacker wearing a black high-tech jacket with a massive hood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hacker manipulating a complex three-dimensional holographic DJ booth with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background filled with exposed pipes and cables that look like vines creeping across the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dimly lit room with flickering LED indicators on server racks in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle in a medium shot framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905211091041565 + ], + "opposite_logprob_tokens": [ + -2.464052200317383 + ], + "logprob_reward": -0.08905211091041565, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905211091041565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17834262549877167, + 0.22542080283164978, + 0.18156065046787262, + 0.03549853712320328, + 0.0864696130156517, + 0.062043603509664536, + 0.06229569390416145, + 0.05764669179916382, + 0.06714944541454315, + 0.043572358787059784 + ], + "credit_map": [ + 0.019785672426223755, + 0.02095748856663704, + 0.017731400206685066, + 0.022952986881136894, + 0.025715406984090805, + 0.016307422891259193, + 0.022887784987688065, + 0.0320044681429863, + 0.020038731396198273, + 0.03492435812950134, + 0.02975986897945404, + 0.04509862884879112, + 0.022603396326303482, + 0.021137826144695282, + 0.030023697763681412, + 0.02183428220450878, + 0.020802156999707222, + 0.029215026646852493, + 0.02346281334757805, + 0.042105067521333694, + 0.02807038277387619, + 0.009915117174386978, + 0.018941326066851616, + 0.009048756211996078, + 0.0036504999734461308, + 0.004540783818811178, + 0.0035053445026278496, + 0.005311221349984407, + 0.00475360406562686, + 0.0065065291710197926, + 0.003732450306415558, + 0.0034981060307472944, + 0.011258450336754322, + 0.009012146852910519, + 0.011621948331594467, + 0.0121371541172266, + 0.01174326241016388, + 0.013834374025464058, + 0.006923086475580931, + 0.009939195588231087, + 0.006449486128985882, + 0.006428983993828297, + 0.008557838387787342, + 0.007340684998780489, + 0.005202521104365587, + 0.006250531878322363, + 0.013644888065755367, + 0.00816867221146822, + 0.004515905864536762, + 0.006644285749644041, + 0.006705739069730043, + 0.00545572629198432, + 0.006746890489012003, + 0.00791015475988388, + 0.0073316507041454315, + 0.016985341906547546, + 0.006177236791700125, + 0.009000815451145172, + 0.0059891873970627785, + 0.005563872866332531, + 0.01044277474284172, + 0.005071750842034817, + 0.006491024512797594, + 0.008910027332603931, + 0.008893238380551338, + 0.005582334939390421, + 0.008034128695726395, + 0.008301359601318836, + 0.011840108782052994, + 0.007189155090600252, + 0.01076348964124918, + 0.006545624695718288, + 0.0036134887486696243, + 0.006941447965800762, + 0.005255845841020346, + 0.0077622439712285995, + 0.0061136893928050995, + 0.004039355553686619, + 0.006408489774912596, + 0.0034377961419522762 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.08905211091041565 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hacker wearing a black high-tech jacket with a massive hood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hacker manipulating a complex three-dimensional holographic DJ booth with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background filled with exposed pipes and cables that look like vines creeping across the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dimly lit room with flickering LED indicators on server racks in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle in a medium shot framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703566253185272 + ], + "opposite_logprob_tokens": [ + -2.127035617828369 + ], + "logprob_reward": -0.12703566253185272, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12703566253185272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1263623833656311, + 0.19076767563819885, + 0.0707717090845108, + 0.045500773936510086, + 0.058038391172885895, + 0.0561675950884819, + 0.05848832428455353, + 0.11889199912548065, + 0.17774684727191925, + 0.09726427495479584 + ], + "credit_map": [ + 0.013612701557576656, + 0.014907806180417538, + 0.013711145147681236, + 0.021779529750347137, + 0.014857198111712933, + 0.01578187197446823, + 0.013341175392270088, + 0.01837095431983471, + 0.016606220975518227, + 0.026245322078466415, + 0.023540286347270012, + 0.030873700976371765, + 0.04260038584470749, + 0.014530656859278679, + 0.02118605375289917, + 0.015185053460299969, + 0.007358985021710396, + 0.0070208050310611725, + 0.00917789340019226, + 0.008390840142965317, + 0.016611771658062935, + 0.007449626456946135, + 0.009584110230207443, + 0.005177672021090984, + 0.0052741277031600475, + 0.006336925085633993, + 0.004634879529476166, + 0.006705576088279486, + 0.007444304879754782, + 0.005757886450737715, + 0.0045735202729702, + 0.0047735534608364105, + 0.009068095125257969, + 0.00414535216987133, + 0.009240582585334778, + 0.009037374518811703, + 0.006257606670260429, + 0.0095821563154459, + 0.005022719036787748, + 0.005684500094503164, + 0.01100621186196804, + 0.005499298218637705, + 0.007671673316508532, + 0.006625575013458729, + 0.0034127545077353716, + 0.005600729025900364, + 0.008375216275453568, + 0.007976136170327663, + 0.0063957409001886845, + 0.007161459419876337, + 0.00827568955719471, + 0.004616506397724152, + 0.007767850067466497, + 0.006392362527549267, + 0.007218085695058107, + 0.010660628788173199, + 0.011872113682329655, + 0.012966199778020382, + 0.014090344309806824, + 0.013929626904428005, + 0.020851075649261475, + 0.01282724179327488, + 0.016196392476558685, + 0.01615900918841362, + 0.014800051227211952, + 0.010983259417116642, + 0.022769739851355553, + 0.02357209287583828, + 0.03125505521893501, + 0.01640324853360653, + 0.03405418246984482, + 0.023909205570816994, + 0.006452005822211504, + 0.008995207026600838, + 0.009593266993761063, + 0.013100064359605312, + 0.013472304679453373, + 0.01895984075963497, + 0.013595782220363617, + 0.013095801696181297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.12703566253185272 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hacker wearing a black high-tech jacket with a massive hood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hacker manipulating a complex three-dimensional holographic DJ booth with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background filled with exposed pipes and cables that look like vines creeping across the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dimly lit room with flickering LED indicators on server racks in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle in a medium shot framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809018433094025 + ], + "opposite_logprob_tokens": [ + -3.288090229034424 + ], + "logprob_reward": -0.03809018433094025, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809018433094025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08385829627513885, + 0.11130347847938538, + 0.09580916166305542, + 0.04355401545763016, + 0.052525680512189865, + 0.058489952236413956, + 0.02923715114593506, + 0.2772195041179657, + 0.10392722487449646, + 0.14407551288604736 + ], + "credit_map": [ + 0.00920671783387661, + 0.010545718483626842, + 0.00961499847471714, + 0.011972584761679173, + 0.014242362231016159, + 0.007405669428408146, + 0.010259871371090412, + 0.010610373690724373, + 0.010790902189910412, + 0.012900173664093018, + 0.0135575570166111, + 0.02062140218913555, + 0.018303049728274345, + 0.012716186232864857, + 0.01290794275701046, + 0.009506268426775932, + 0.00794895924627781, + 0.008977701887488365, + 0.012040681205689907, + 0.008612522855401039, + 0.0225216094404459, + 0.011201432906091213, + 0.017360685393214226, + 0.007145570125430822, + 0.005816992837935686, + 0.005465256050229073, + 0.00492168590426445, + 0.007558449171483517, + 0.005883447825908661, + 0.005327229388058186, + 0.004987603519111872, + 0.0035933502949774265, + 0.00775189371779561, + 0.00495558138936758, + 0.007910406216979027, + 0.006302747875452042, + 0.005973983556032181, + 0.00808852817863226, + 0.00588533841073513, + 0.005657202564179897, + 0.009992708452045918, + 0.006816954351961613, + 0.007630686275660992, + 0.006085013039410114, + 0.004021394066512585, + 0.005779325496405363, + 0.008992964401841164, + 0.009170908480882645, + 0.0028044767677783966, + 0.004398792516440153, + 0.0034094927832484245, + 0.002590161981061101, + 0.0033272334840148687, + 0.0028805267065763474, + 0.0032691454980522394, + 0.00655731838196516, + 0.03568106144666672, + 0.02850443124771118, + 0.029127640649676323, + 0.03127765655517578, + 0.05331721529364586, + 0.030764928087592125, + 0.030285952612757683, + 0.03826062008738518, + 0.009960493072867393, + 0.009993890300393105, + 0.013249835930764675, + 0.009611018002033234, + 0.022145401686429977, + 0.011483190581202507, + 0.01838875189423561, + 0.009094636887311935, + 0.012083273380994797, + 0.01650051400065422, + 0.016454003751277924, + 0.01685856096446514, + 0.018647221848368645, + 0.0215571578592062, + 0.02350860834121704, + 0.018466182053089142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 1 + ], + "selected_reward": -0.03809018433094025 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hacker wearing a black high-tech jacket with a massive hood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hacker manipulating a complex three-dimensional holographic DJ booth with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background filled with exposed pipes and cables that look like vines creeping across the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dimly lit room with flickering LED indicators on server racks in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle in a medium shot framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369210660457611 + ], + "opposite_logprob_tokens": [ + -3.4086921215057373 + ], + "logprob_reward": -0.03369210660457611, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369210660457611, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10051379352807999, + 0.2417488694190979, + 0.11936430633068085, + 0.05452389642596245, + 0.054680753499269485, + 0.07568798214197159, + 0.05235766991972923, + 0.16517303884029388, + 0.05204924941062927, + 0.08390045911073685 + ], + "credit_map": [ + 0.012149938382208347, + 0.013100087642669678, + 0.010398571379482746, + 0.020144158974289894, + 0.014844069257378578, + 0.008459514938294888, + 0.011550751514732838, + 0.009866700507700443, + 0.020142365247011185, + 0.03339022770524025, + 0.030450575053691864, + 0.04694293439388275, + 0.052628930658102036, + 0.014714245684444904, + 0.02759084478020668, + 0.01588873192667961, + 0.012523933313786983, + 0.015600909478962421, + 0.016847100108861923, + 0.03611225262284279, + 0.012092980556190014, + 0.00795051921159029, + 0.012768074870109558, + 0.005468535237014294, + 0.005904458928853273, + 0.009104261174798012, + 0.006803428288549185, + 0.010671823285520077, + 0.006131932605057955, + 0.006552790757268667, + 0.005327466409653425, + 0.004027732647955418, + 0.007454256992787123, + 0.004253961145877838, + 0.008555276319384575, + 0.0071861958131194115, + 0.005898942239582539, + 0.007844499312341213, + 0.007497384212911129, + 0.00599023187533021, + 0.011408182792365551, + 0.011575507931411266, + 0.012361085042357445, + 0.014617446810007095, + 0.0058423238806426525, + 0.007934554480016232, + 0.006576094310730696, + 0.005372784100472927, + 0.00566073739901185, + 0.007052158936858177, + 0.007644449360668659, + 0.006293365731835365, + 0.005861084442585707, + 0.0058767045848071575, + 0.004822744056582451, + 0.009146420285105705, + 0.020439129322767258, + 0.030151234939694405, + 0.01984369568526745, + 0.025978002697229385, + 0.02228786237537861, + 0.01810583472251892, + 0.015760967507958412, + 0.012606297619640827, + 0.005942972842603922, + 0.005922649521380663, + 0.007877876050770283, + 0.00499087106436491, + 0.010256076231598854, + 0.0055281370878219604, + 0.006039774511009455, + 0.005490891635417938, + 0.00663650082424283, + 0.018878143280744553, + 0.011604621075093746, + 0.009592493996024132, + 0.010366750881075859, + 0.01153592485934496, + 0.008655179291963577, + 0.006630849093198776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 2 + ], + "selected_reward": -0.03369210660457611 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, medium shot, shot from a slightly upward angle, draws the viewer's attention to a mysterious hacker in a dimly lit room. He wears a black, high-tech jacket with a massive hood, its deep shadow almost completely obscuring his features, only occasionally revealing his resolute jawline in the cold glow of the holographic projection. His hands, clad in sleek black fingerless gloves, manipulate a complex, three-dimensional holographic DJ booth suspended before him with precise and rhythmic control. This virtual device, composed of glowing geometric shapes, pulsating sound waves, and data streams, shifts in real time with his swipes, pinches, and taps, leaving fleeting afterimages behind his fingertips. The background is a tangle of exposed pipes and cables, vine-like, creeping across the walls. In the distance, LED indicators on server racks flicker silently in the darkness. The main cold-toned light comes from the holographic projection itself, casting a sharp light on the hacker, while the magenta and cyan neon lights from the giant billboards outside the window in the eternal rainy night slowly flow on the reflective material of his high-tech jacket, forming dynamic light and shadow. The whole scene is filled with a highly focused, futuristic and slightly decadent cyberpunk atmosphere.", + "seed": 3092, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_7/index_3572_seed_3092_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10936436392366886, + "vlm_logprob": -0.10936436392366886, + "vlm_yes_no_margin": 2.7250000119209288, + "vlm_token_credit": -0.10936436392366886, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hacker wearing a black high-tech jacket with a massive hood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hacker manipulating a complex three-dimensional holographic DJ booth with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background filled with exposed pipes and cables that look like vines creeping across the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dimly lit room with flickering LED indicators on server racks in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle in a medium shot framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14548909664154053 + ], + "opposite_logprob_tokens": [ + -2.02048921585083 + ], + "logprob_reward": -0.14548909664154053, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14548909664154053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1547684669494629, + 0.1010744720697403, + 0.07830235362052917, + 0.1434382051229477, + 0.0600174181163311, + 0.10279575735330582, + 0.039574310183525085, + 0.045383043587207794, + 0.1002088114619255, + 0.17443718016147614 + ], + "credit_map": [ + 0.0181779433041811, + 0.015300629660487175, + 0.018059395253658295, + 0.018357248976826668, + 0.022802313789725304, + 0.013874886557459831, + 0.029346121475100517, + 0.018849922344088554, + 0.019718047231435776, + 0.011072533205151558, + 0.011432662606239319, + 0.012945734895765781, + 0.010580536909401417, + 0.0073619382455945015, + 0.01661190204322338, + 0.011351116932928562, + 0.012866719625890255, + 0.006867530290037394, + 0.014424232766032219, + 0.0103586595505476, + 0.008872448466718197, + 0.007976457476615906, + 0.00801708921790123, + 0.008919213898479939, + 0.013567768968641758, + 0.012996290810406208, + 0.016999073326587677, + 0.017923453822731972, + 0.026207027956843376, + 0.017864279448986053, + 0.02087913453578949, + 0.017001181840896606, + 0.005593420006334782, + 0.008977076970040798, + 0.006560158915817738, + 0.011455927044153214, + 0.00821581482887268, + 0.004108050838112831, + 0.011296940967440605, + 0.0038100306410342455, + 0.02515524812042713, + 0.010889152064919472, + 0.01561255194246769, + 0.014171650633215904, + 0.009084665216505527, + 0.006469112820923328, + 0.015232408419251442, + 0.006180969532579184, + 0.008960580453276634, + 0.0035688006319105625, + 0.007923427037894726, + 0.005338634364306927, + 0.003908810205757618, + 0.003034562338143587, + 0.003734576515853405, + 0.0031049163080751896, + 0.006248745135962963, + 0.0033606509678065777, + 0.005811575334519148, + 0.008926470763981342, + 0.0039114137180149555, + 0.004804365336894989, + 0.00790391769260168, + 0.004415905103087425, + 0.008454074151813984, + 0.01190934143960476, + 0.011461730115115643, + 0.009287780150771141, + 0.026598306372761726, + 0.008304922841489315, + 0.015814589336514473, + 0.00837806798517704, + 0.02333681471645832, + 0.01276316773146391, + 0.06161489710211754, + 0.030647004023194313, + 0.018445251509547234, + 0.00637592701241374, + 0.015643510967493057, + 0.005610609892755747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.14548909664154053 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hacker wearing a black high-tech jacket with a massive hood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hacker manipulating a complex three-dimensional holographic DJ booth with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background filled with exposed pipes and cables that look like vines creeping across the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dimly lit room with flickering LED indicators on server racks in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle in a medium shot framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.313479483127594 + ], + "opposite_logprob_tokens": [ + -1.3134794235229492 + ], + "logprob_reward": -0.313479483127594, + "yes_no_margin_reward": 0.9999999403953552, + "token_credit_reward": -0.313479483127594, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09882984310388565, + 0.1606358289718628, + 0.15098917484283447, + 0.17914822697639465, + 0.07823805510997772, + 0.08448556065559387, + 0.053908571600914, + 0.07187719643115997, + 0.05057896301150322, + 0.07130859792232513 + ], + "credit_map": [ + 0.015116497874259949, + 0.01023071352392435, + 0.011468527838587761, + 0.011400450021028519, + 0.019214196130633354, + 0.009216172620654106, + 0.012819353491067886, + 0.009363928809762001, + 0.021754782646894455, + 0.017589673399925232, + 0.02577778324484825, + 0.024423114955425262, + 0.017863020300865173, + 0.012383723631501198, + 0.023309284821152687, + 0.017534425482153893, + 0.035239722579717636, + 0.013706338591873646, + 0.03473939001560211, + 0.021060941740870476, + 0.014305484481155872, + 0.008832131512463093, + 0.012619850225746632, + 0.010485310107469559, + 0.018911214545369148, + 0.017791304737329483, + 0.020574714988470078, + 0.022724710404872894, + 0.02977869287133217, + 0.028510142117738724, + 0.019354747608304024, + 0.021502677351236343, + 0.013240596279501915, + 0.01050169300287962, + 0.008191687054932117, + 0.012766839936375618, + 0.012168102897703648, + 0.007027860730886459, + 0.009412127546966076, + 0.004929136484861374, + 0.01713656820356846, + 0.008315014652907848, + 0.016170747578144073, + 0.012907403521239758, + 0.009043624624609947, + 0.005429188255220652, + 0.010331529192626476, + 0.005151475314050913, + 0.008645797148346901, + 0.00570050161331892, + 0.008340932428836823, + 0.007391071878373623, + 0.006961671635508537, + 0.004323170054703951, + 0.008547117933630943, + 0.003998305648565292, + 0.01109335757791996, + 0.0059700654819607735, + 0.008544648997485638, + 0.011786186136305332, + 0.006506058387458324, + 0.007407706696540117, + 0.012286962941288948, + 0.008282210677862167, + 0.005525621585547924, + 0.005343560129404068, + 0.00554594025015831, + 0.005941873881965876, + 0.011457529850304127, + 0.004076682031154633, + 0.007503422908484936, + 0.005184330977499485, + 0.006167995743453503, + 0.0055483486503362656, + 0.017356107011437416, + 0.01856720820069313, + 0.011987948790192604, + 0.0031405494082719088, + 0.005286623723804951, + 0.0032538133673369884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.313479483127594 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hacker wearing a black high-tech jacket with a massive hood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hacker manipulating a complex three-dimensional holographic DJ booth with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background filled with exposed pipes and cables that look like vines creeping across the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dimly lit room with flickering LED indicators on server racks in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle in a medium shot framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809374198317528 + ], + "opposite_logprob_tokens": [ + -3.2880938053131104 + ], + "logprob_reward": -0.03809374198317528, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809374198317528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09730946272611618, + 0.07937320321798325, + 0.278707891702652, + 0.06854161620140076, + 0.040444061160087585, + 0.06104658916592598, + 0.24493739008903503, + 0.028583485633134842, + 0.03946748003363609, + 0.061588820070028305 + ], + "credit_map": [ + 0.016754647716879845, + 0.013856823556125164, + 0.013541119173169136, + 0.012660561129450798, + 0.013719132170081139, + 0.007060889154672623, + 0.012557437643408775, + 0.007158849388360977, + 0.009989513084292412, + 0.0075734881684184074, + 0.011380642652511597, + 0.014796302653849125, + 0.010491281747817993, + 0.005319388583302498, + 0.011833998374640942, + 0.007988587021827698, + 0.03593020141124725, + 0.026805618777871132, + 0.04121416434645653, + 0.04059898480772972, + 0.037492379546165466, + 0.02718871831893921, + 0.03482997789978981, + 0.03464784473180771, + 0.006888231262564659, + 0.0062849633395671844, + 0.007945012301206589, + 0.0086695970967412, + 0.010068817995488644, + 0.008597771637141705, + 0.01235747430473566, + 0.007729743607342243, + 0.005422687623649836, + 0.005223628133535385, + 0.004254282917827368, + 0.007515236269682646, + 0.004913637414574623, + 0.0032033459283411503, + 0.006096857599914074, + 0.0038143869023770094, + 0.00963697861880064, + 0.006929297931492329, + 0.009162782691419125, + 0.008569210767745972, + 0.006014036945998669, + 0.00723283551633358, + 0.007186100352555513, + 0.006315344478935003, + 0.04528037831187248, + 0.027321847155690193, + 0.044934097677469254, + 0.03489600867033005, + 0.028300585225224495, + 0.01589883118867874, + 0.03325732424855232, + 0.015048305504024029, + 0.004666244611144066, + 0.0023431337904185057, + 0.003409696277230978, + 0.006687736604362726, + 0.0026786033995449543, + 0.0016065703239291906, + 0.0045744916424155235, + 0.0026170082855969667, + 0.006818596739321947, + 0.004009584430605173, + 0.004871591925621033, + 0.004542035050690174, + 0.008549812249839306, + 0.0021434705704450607, + 0.0057110232301056385, + 0.0028213676996529102, + 0.007337604649364948, + 0.004355464596301317, + 0.017176039516925812, + 0.009775148704648018, + 0.00959151703864336, + 0.003295555943623185, + 0.0060154469683766365, + 0.004042043350636959 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.03809374198317528 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hacker wearing a black high-tech jacket with a massive hood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hacker manipulating a complex three-dimensional holographic DJ booth with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background filled with exposed pipes and cables that look like vines creeping across the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dimly lit room with flickering LED indicators on server racks in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle in a medium shot framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0160652007907629 + ], + "opposite_logprob_tokens": [ + -4.1410651206970215 + ], + "logprob_reward": -0.0160652007907629, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0160652007907629, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07605016976594925, + 0.0508427768945694, + 0.05124441906809807, + 0.07766223698854446, + 0.051338061690330505, + 0.050932444632053375, + 0.09341252595186234, + 0.2663566470146179, + 0.08408897370100021, + 0.19807176291942596 + ], + "credit_map": [ + 0.011631887406110764, + 0.007389561738818884, + 0.009905186481773853, + 0.009915993548929691, + 0.013369870372116566, + 0.007498031482100487, + 0.00881310272961855, + 0.007526535075157881, + 0.008310964331030846, + 0.004971468821167946, + 0.008810862898826599, + 0.007691199891269207, + 0.006550300400704145, + 0.0030871503986418247, + 0.006937576457858086, + 0.004483249504119158, + 0.006762776058167219, + 0.004944656044244766, + 0.008156871423125267, + 0.006746952421963215, + 0.007730806712061167, + 0.004628405440598726, + 0.007215726189315319, + 0.005058224778622389, + 0.010186945088207722, + 0.006416833493858576, + 0.010069058276712894, + 0.009329764172434807, + 0.013849603943526745, + 0.0077435895800590515, + 0.01199996005743742, + 0.008066476322710514, + 0.009200523607432842, + 0.006403191480785608, + 0.004845886956900358, + 0.007852860726416111, + 0.00913269817829132, + 0.004176288843154907, + 0.00664941780269146, + 0.0030771936289966106, + 0.01095773559063673, + 0.004961081314831972, + 0.007870059460401535, + 0.008272855542600155, + 0.004272901453077793, + 0.003598021576181054, + 0.0061923242174088955, + 0.004807465244084597, + 0.012566761113703251, + 0.010184546001255512, + 0.016058512032032013, + 0.013500173576176167, + 0.01169215515255928, + 0.007145566400140524, + 0.012410417199134827, + 0.009854388423264027, + 0.04264526069164276, + 0.02939910814166069, + 0.02974921464920044, + 0.048252761363983154, + 0.026956472545862198, + 0.02037040702998638, + 0.043096572160720825, + 0.02588685415685177, + 0.014466676861047745, + 0.012033714912831783, + 0.010111331008374691, + 0.014673089608550072, + 0.011837710626423359, + 0.004936121869832277, + 0.00955919362604618, + 0.006471120752394199, + 0.02617955021560192, + 0.015262902714312077, + 0.06529536098241806, + 0.04380271956324577, + 0.02236883156001568, + 0.006182540208101273, + 0.013739368878304958, + 0.005240483209490776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.0160652007907629 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hacker wearing a black high-tech jacket with a massive hood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hacker manipulating a complex three-dimensional holographic DJ booth with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background filled with exposed pipes and cables that look like vines creeping across the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dimly lit room with flickering LED indicators on server racks in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle in a medium shot framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033694297075271606 + ], + "opposite_logprob_tokens": [ + -3.408694267272949 + ], + "logprob_reward": -0.033694297075271606, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033694297075271606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24528037011623383, + 0.07538574934005737, + 0.1264970451593399, + 0.06818174570798874, + 0.06000975891947746, + 0.07013529539108276, + 0.0612386018037796, + 0.07274796068668365, + 0.1008443608880043, + 0.11967911571264267 + ], + "credit_map": [ + 0.034494418650865555, + 0.01943785510957241, + 0.033485785126686096, + 0.032979566603899, + 0.050296850502491, + 0.015720335766673088, + 0.03730392083525658, + 0.0215616337954998, + 0.009894688613712788, + 0.010259335860610008, + 0.009597483091056347, + 0.011405331082642078, + 0.010992280207574368, + 0.00515366718173027, + 0.011734923347830772, + 0.00634803855791688, + 0.01959419623017311, + 0.009737656451761723, + 0.0200963132083416, + 0.017866015434265137, + 0.01737123541533947, + 0.011083461344242096, + 0.017054405063390732, + 0.013693764805793762, + 0.007823362946510315, + 0.006196440663188696, + 0.008852378465235233, + 0.0072263567708432674, + 0.010124089196324348, + 0.009256097488105297, + 0.011345680803060532, + 0.007357338443398476, + 0.007062825374305248, + 0.00678036455065012, + 0.006733429618179798, + 0.009638356044888496, + 0.010698814876377583, + 0.004808932542800903, + 0.009497661143541336, + 0.004789375234395266, + 0.019192317500710487, + 0.00803365372121334, + 0.007992926985025406, + 0.0090170968323946, + 0.007546175271272659, + 0.004114380106329918, + 0.008174763061106205, + 0.0060639772564172745, + 0.007907286286354065, + 0.008824977092444897, + 0.009445765055716038, + 0.008256882429122925, + 0.009744334034621716, + 0.005119298119097948, + 0.007923820987343788, + 0.004016237799078226, + 0.010947105474770069, + 0.005951376631855965, + 0.008771686814725399, + 0.01204385980963707, + 0.010502064600586891, + 0.005981041118502617, + 0.012097165919840336, + 0.006453660782426596, + 0.01120343804359436, + 0.01111633237451315, + 0.011676314286887646, + 0.01063037384301424, + 0.020154351368546486, + 0.009922659955918789, + 0.01650596596300602, + 0.009634929709136486, + 0.011131016537547112, + 0.011390997096896172, + 0.02358846366405487, + 0.02905125729739666, + 0.020990906283259392, + 0.005706298165023327, + 0.011860033497214317, + 0.005960135255008936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.033694297075271606 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_8/index_tensor([8934, 8934])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_8/index_tensor([8934, 8934])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f524c1bbb55de515b9f69b2beade1c70d455575f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_8/index_tensor([8934, 8934])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An epic cosmic panorama, initially presented in soft focus, slowly zooms in on two doomed stars, pulled toward each other by the relentless pull of gravity. Amidst the deep, serene void of space, on the left stands a massive red supergiant star, its massive, swollen surface roiling with a deep crimson of fiery plasma. On the right, a dense, piercing blue-white neutron star, like a deadly cosmic projectile, its immense gravity subtly warping the surrounding space. As the two stars collide in an ever-accelerating spiral, a silent yet visually stunning supernova explosion is triggered at the moment of contact. A spherical shockwave expands outward at near-light speed, instantly engulfing and ripping apart the massive mass of the red giant, unleashing an all-consuming, dazzling radiance whose intensity momentarily surpasses the entire galaxy, illuminating the entire frame in a pure, searing white, creating an atmosphere of cosmic wonder, both majestic and destructive.", + "seed": 2549, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_8/index_8934_seed_2549_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.15249258673284202, + "vlm_logprob": -0.15249258673284202, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.15249258673284202, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial view of the scene presented in soft focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive red supergiant star visible on the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense blue-white neutron star visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two stars collide in an accelerating spiral motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene end with a spherical shockwave expanding outward and illuminating the frame in pure white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.48979252576828003 + ], + "opposite_logprob_tokens": [ + -0.98979252576828 + ], + "logprob_reward": -0.48979252576828003, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.48979252576828003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10052869468927383, + 0.053424593061208725, + 0.05402754247188568, + 0.045598261058330536, + 0.2471911609172821, + 0.01638314686715603, + 0.02991396375000477, + 0.05373525619506836, + 0.09088846296072006, + 0.308308869600296 + ], + "credit_map": [ + 0.014919420704245567, + 0.009531971998512745, + 0.01997934840619564, + 0.011196914128959179, + 0.011507400311529636, + 0.010725819505751133, + 0.008267248049378395, + 0.014400571584701538, + 0.004374078940600157, + 0.002701112302020192, + 0.0076430425979197025, + 0.0029973320197314024, + 0.004880629014223814, + 0.013949728570878506, + 0.003157947212457657, + 0.01372072659432888, + 0.008079799823462963, + 0.007210844196379185, + 0.014366181567311287, + 0.008092955686151981, + 0.004267213866114616, + 0.003809859976172447, + 0.0030256141908466816, + 0.005175070837140083, + 0.008041422814130783, + 0.0044958386570215225, + 0.01196747925132513, + 0.005220348481088877, + 0.002586763584986329, + 0.0045219724997878075, + 0.002560993889346719, + 0.006203442346304655, + 0.009991977363824844, + 0.005832191556692123, + 0.013409756124019623, + 0.008615589700639248, + 0.01113609317690134, + 0.10215921700000763, + 0.005184921436011791, + 0.09086141735315323, + 0.0018508953507989645, + 0.0008158029522746801, + 0.001789212692528963, + 0.0015713867032900453, + 0.0026828336995095015, + 0.0030837859958410263, + 0.0014889207668602467, + 0.003100308822467923, + 0.003981383051723242, + 0.003419956425204873, + 0.004265450406819582, + 0.006746373139321804, + 0.003329368308186531, + 0.002743449294939637, + 0.0022352200467139482, + 0.0031927640084177256, + 0.004039686638861895, + 0.0036473465152084827, + 0.005320587661117315, + 0.00937001034617424, + 0.005531962029635906, + 0.009996911510825157, + 0.004260518122464418, + 0.01156823243945837, + 0.012529013678431511, + 0.0111104566603899, + 0.01568624936044216, + 0.013806432485580444, + 0.007753286976367235, + 0.008384312503039837, + 0.0110397320240736, + 0.010578986257314682, + 0.011467316187918186, + 0.005961907561868429, + 0.04868517443537712, + 0.010439112782478333, + 0.011503594927489758, + 0.12770971655845642, + 0.010194561444222927, + 0.08234748989343643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.48979252576828003 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial view of the scene presented in soft focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive red supergiant star visible on the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense blue-white neutron star visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two stars collide in an accelerating spiral motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene end with a spherical shockwave expanding outward and illuminating the frame in pure white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009657248156145215 + ], + "opposite_logprob_tokens": [ + -7.000965595245361 + ], + "logprob_reward": -0.0009657248156145215, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009657248156145215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17275509238243103, + 0.23037083446979523, + 0.03548285737633705, + 0.017280256375670433, + 0.19034096598625183, + 0.01888820342719555, + 0.020604023709893227, + 0.04264627397060394, + 0.07004136592149734, + 0.20159012079238892 + ], + "credit_map": [ + 0.01635916531085968, + 0.008014125749468803, + 0.01771271787583828, + 0.008861007168889046, + 0.022188877686858177, + 0.03808579966425896, + 0.024793123826384544, + 0.036740273237228394, + 0.03491787612438202, + 0.010931102558970451, + 0.022044828161597252, + 0.01259115431457758, + 0.012254349887371063, + 0.0603928379714489, + 0.00936479214578867, + 0.06787388026714325, + 0.005112317856401205, + 0.0031005002092570066, + 0.00750563433393836, + 0.002977584721520543, + 0.0038755659479647875, + 0.0048858425579965115, + 0.002313098404556513, + 0.005712312646210194, + 0.00237823068164289, + 0.0015394299989566207, + 0.003419942455366254, + 0.001970688346773386, + 0.0017268407391384244, + 0.0022202134132385254, + 0.001580313895829022, + 0.0024445971939712763, + 0.007494690828025341, + 0.003998531959950924, + 0.004127394407987595, + 0.003898253198713064, + 0.009355485439300537, + 0.08212912082672119, + 0.004787379410117865, + 0.07455011457204819, + 0.002057792618870735, + 0.0012368151219561696, + 0.002973865484818816, + 0.0018517757998779416, + 0.003462879452854395, + 0.002993619069457054, + 0.0014782245270907879, + 0.002833231817930937, + 0.0031942452769726515, + 0.0019102629739791155, + 0.0028525181114673615, + 0.0029796857852488756, + 0.002514522522687912, + 0.0026108729653060436, + 0.001234154449775815, + 0.0033077625557780266, + 0.005619938485324383, + 0.003639388596639037, + 0.006740754935890436, + 0.007339507807046175, + 0.003367065917700529, + 0.0058483462780714035, + 0.0030074212700128555, + 0.007083851844072342, + 0.006934416480362415, + 0.0035606049932539463, + 0.006371698342263699, + 0.006777101196348667, + 0.007113146595656872, + 0.016357924789190292, + 0.005785071756690741, + 0.017141399905085564, + 0.011139479465782642, + 0.004996090661734343, + 0.020581107586622238, + 0.006632637232542038, + 0.01698365807533264, + 0.07399558275938034, + 0.008324635215103626, + 0.05893692746758461 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 4 + ], + "selected_reward": -0.0009657248156145215 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial view of the scene presented in soft focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive red supergiant star visible on the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense blue-white neutron star visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two stars collide in an accelerating spiral motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene end with a spherical shockwave expanding outward and illuminating the frame in pure white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009795219637453556 + ], + "opposite_logprob_tokens": [ + -4.634795188903809 + ], + "logprob_reward": -0.009795219637453556, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009795219637453556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23587891459465027, + 0.16541820764541626, + 0.06940687447786331, + 0.028667081147432327, + 0.06256774812936783, + 0.05509091913700104, + 0.03372333198785782, + 0.03634011745452881, + 0.06800694018602371, + 0.2448999434709549 + ], + "credit_map": [ + 0.02096603438258171, + 0.009927015751600266, + 0.020206263288855553, + 0.01160495262593031, + 0.03628842532634735, + 0.05301239714026451, + 0.03536849096417427, + 0.04850533604621887, + 0.021795032545924187, + 0.009920746088027954, + 0.017122212797403336, + 0.012721310369670391, + 0.03176265209913254, + 0.02462136745452881, + 0.01726871356368065, + 0.030206166207790375, + 0.008041027933359146, + 0.005079942289739847, + 0.011065163649618626, + 0.004525535274296999, + 0.007087241858243942, + 0.013819627463817596, + 0.005159348715096712, + 0.014628992415964603, + 0.0046824561432003975, + 0.0017090003238990903, + 0.006926495116204023, + 0.002786835888400674, + 0.0030881729908287525, + 0.003248987253755331, + 0.002799915848299861, + 0.0034252170007675886, + 0.006625884212553501, + 0.0031130583956837654, + 0.011799321509897709, + 0.005746473558247089, + 0.005186752416193485, + 0.012332085520029068, + 0.004191983491182327, + 0.01357219647616148, + 0.00331668509170413, + 0.001789009547792375, + 0.005759607534855604, + 0.0027066790498793125, + 0.007962089963257313, + 0.01495806872844696, + 0.004037606064230204, + 0.014561176300048828, + 0.004087415058165789, + 0.003032551845535636, + 0.004671910777688026, + 0.003989796619862318, + 0.006694943644106388, + 0.0037462390027940273, + 0.0033174497075378895, + 0.00418302696198225, + 0.0030699269846081734, + 0.003841774770990014, + 0.004071338567882776, + 0.005013493355363607, + 0.00454375147819519, + 0.005626964848488569, + 0.003245952073484659, + 0.0069269160740077496, + 0.007595638744533062, + 0.0035384921357035637, + 0.00696334196254611, + 0.006018885876983404, + 0.005915519781410694, + 0.016479285433888435, + 0.004576840903609991, + 0.016918936744332314, + 0.01047486625611782, + 0.005982696078717709, + 0.04446648061275482, + 0.008095679804682732, + 0.009779857471585274, + 0.09979759156703949, + 0.008937444537878036, + 0.0573653019964695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.009795219637453556 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial view of the scene presented in soft focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive red supergiant star visible on the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense blue-white neutron star visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two stars collide in an accelerating spiral motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene end with a spherical shockwave expanding outward and illuminating the frame in pure white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.252124160528183 + ], + "opposite_logprob_tokens": [ + -1.5021241903305054 + ], + "logprob_reward": -0.252124160528183, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.252124160528183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.022294702008366585, + 0.04818229377269745, + 0.025983521714806557, + 0.02232571877539158, + 0.13603559136390686, + 0.019541990011930466, + 0.021811483427882195, + 0.034664053469896317, + 0.016680367290973663, + 0.6524803638458252 + ], + "credit_map": [ + 0.0027939206920564175, + 0.0016190001042559743, + 0.003201741725206375, + 0.002831874182447791, + 0.0033116256818175316, + 0.0022462636698037386, + 0.0034181650262326, + 0.0028721054550260305, + 0.008110938593745232, + 0.0026303057093173265, + 0.006463202182203531, + 0.002925264183431864, + 0.003975945059210062, + 0.01014538574963808, + 0.002251553349196911, + 0.011679688468575478, + 0.004854380618780851, + 0.002114818897098303, + 0.0049707526341080666, + 0.0025223293341696262, + 0.0021378465462476015, + 0.00420876732096076, + 0.0014657212886959314, + 0.003708897391334176, + 0.003526936983689666, + 0.0022983993403613567, + 0.005451112985610962, + 0.00275151408277452, + 0.002140582073479891, + 0.002363587263971567, + 0.0014387770788744092, + 0.002354804892092943, + 0.014617254957556725, + 0.006315671838819981, + 0.020551282912492752, + 0.01154678501188755, + 0.00610917154699564, + 0.03881622478365898, + 0.004000909626483917, + 0.034078262746334076, + 0.002549444092437625, + 0.0021316634956747293, + 0.0050422921776771545, + 0.0021804366260766983, + 0.0015886761248111725, + 0.0022066431120038033, + 0.0011106299934908748, + 0.0027322007808834314, + 0.0032544503919780254, + 0.002318399026989937, + 0.0031584997195750475, + 0.003936480265110731, + 0.002005983842536807, + 0.0024692534934729338, + 0.0014839625218883157, + 0.0031844486948102713, + 0.0040741777047514915, + 0.0029900837689638138, + 0.00614817114546895, + 0.008937749080359936, + 0.0020957202650606632, + 0.003503474174067378, + 0.0021734070032835007, + 0.004741263575851917, + 0.0011336627649143338, + 0.0008003920083865523, + 0.0013355185510590672, + 0.0011045691790059209, + 0.0018193136202171445, + 0.004975634627044201, + 0.0011466442374512553, + 0.0043646288104355335, + 0.024694077670574188, + 0.011445802636444569, + 0.09307524561882019, + 0.016113199293613434, + 0.023303266614675522, + 0.27543744444847107, + 0.0206606425344944, + 0.18775056302547455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 1 + ], + "selected_reward": -0.252124160528183 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial view of the scene presented in soft focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive red supergiant star visible on the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense blue-white neutron star visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two stars collide in an accelerating spiral motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene end with a spherical shockwave expanding outward and illuminating the frame in pure white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978530291467905 + ], + "opposite_logprob_tokens": [ + -4.634785175323486 + ], + "logprob_reward": -0.00978530291467905, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978530291467905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.014586804434657097, + 0.05704488232731819, + 0.00693900603801012, + 0.037172745913267136, + 0.013953992165625095, + 0.007175416685640812, + 0.01401669904589653, + 0.056154970079660416, + 0.0552494078874588, + 0.7377061247825623 + ], + "credit_map": [ + 0.0017863212851807475, + 0.0015419019619002938, + 0.0021369413007050753, + 0.002296135062351823, + 0.0014472028706222773, + 0.00189616484567523, + 0.0011724219657480717, + 0.002309714909642935, + 0.008617745712399483, + 0.002470324980095029, + 0.005342969670891762, + 0.002626323141157627, + 0.0028169904835522175, + 0.015531192533671856, + 0.0025859379675239325, + 0.01705339550971985, + 0.0012223481899127364, + 0.000570432806853205, + 0.0009580980404280126, + 0.0008275422151200473, + 0.0008258754387497902, + 0.0009382680873386562, + 0.0004783649055752903, + 0.0011180751025676727, + 0.0024445760063827038, + 0.0017416843911632895, + 0.003334444249048829, + 0.002009226009249687, + 0.002039362443611026, + 0.012712477706372738, + 0.0012312621111050248, + 0.011659707874059677, + 0.0008225002675317228, + 0.0006975830183364451, + 0.0010770553490146995, + 0.000769425998441875, + 0.0007407699595205486, + 0.004970337729901075, + 0.0004498258640524, + 0.004426492378115654, + 0.0011103370925411582, + 0.0007714140811003745, + 0.0014126959722489119, + 0.0008431774913333356, + 0.0007900662603788078, + 0.0008832530002109706, + 0.00045073244837112725, + 0.0009137391461990774, + 0.0017835638718679547, + 0.0013171336613595486, + 0.00181973644066602, + 0.0028627479914575815, + 0.0018288964638486505, + 0.001641549519263208, + 0.001154083525761962, + 0.0016089859418570995, + 0.005750320386141539, + 0.006453522946685553, + 0.0077770426869392395, + 0.01225473452359438, + 0.005069969221949577, + 0.0058535682037472725, + 0.004593079909682274, + 0.008402721025049686, + 0.003020852105692029, + 0.0023035199847072363, + 0.005777142010629177, + 0.003846050938591361, + 0.00230595376342535, + 0.018439626321196556, + 0.0026582079008221626, + 0.016898049041628838, + 0.029872754588723183, + 0.020170075818896294, + 0.10231076180934906, + 0.02157621644437313, + 0.03052828460931778, + 0.29824626445770264, + 0.0321408174932003, + 0.20286092162132263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 7 + ], + "selected_reward": -0.00978530291467905 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic cosmic panorama, initially presented in soft focus, slowly zooms in on two doomed stars, pulled toward each other by the relentless pull of gravity. Amidst the deep, serene void of space, on the left stands a massive red supergiant star, its massive, swollen surface roiling with a deep crimson of fiery plasma. On the right, a dense, piercing blue-white neutron star, like a deadly cosmic projectile, its immense gravity subtly warping the surrounding space. As the two stars collide in an ever-accelerating spiral, a silent yet visually stunning supernova explosion is triggered at the moment of contact. A spherical shockwave expands outward at near-light speed, instantly engulfing and ripping apart the massive mass of the red giant, unleashing an all-consuming, dazzling radiance whose intensity momentarily surpasses the entire galaxy, illuminating the entire frame in a pure, searing white, creating an atmosphere of cosmic wonder, both majestic and destructive.", + "seed": 2550, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_8/index_8934_seed_2550_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.13081414001062514, + "vlm_logprob": -0.13081414001062514, + "vlm_yes_no_margin": 3.449999952316284, + "vlm_token_credit": -0.13081414001062514, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial view of the scene presented in soft focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive red supergiant star visible on the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense blue-white neutron star visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two stars collide in an accelerating spiral motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene end with a spherical shockwave expanding outward and illuminating the frame in pure white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5429266691207886 + ], + "opposite_logprob_tokens": [ + -0.9179266691207886 + ], + "logprob_reward": -0.5429266691207886, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5429266691207886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13389694690704346, + 0.08091840147972107, + 0.06037294864654541, + 0.09487961232662201, + 0.07763523608446121, + 0.08382349461317062, + 0.07560128718614578, + 0.12992073595523834, + 0.10419034957885742, + 0.15876099467277527 + ], + "credit_map": [ + 0.027881450951099396, + 0.016565615311264992, + 0.018316345289349556, + 0.015230925753712654, + 0.017924118787050247, + 0.015057441778481007, + 0.008793117478489876, + 0.014127942733466625, + 0.009416485205292702, + 0.011559685692191124, + 0.01152181439101696, + 0.015217102132737637, + 0.007093507330864668, + 0.009303653612732887, + 0.005733543541282415, + 0.011072609573602676, + 0.008057009428739548, + 0.00744260149076581, + 0.008205304853618145, + 0.01202043890953064, + 0.005818680860102177, + 0.007599043659865856, + 0.0050430805422365665, + 0.006186790764331818, + 0.011475042439997196, + 0.012201620265841484, + 0.011189121752977371, + 0.016338873654603958, + 0.008236578665673733, + 0.014609011821448803, + 0.007091050036251545, + 0.013738310895860195, + 0.011019708588719368, + 0.00580142717808485, + 0.011127334088087082, + 0.008498473092913628, + 0.005254361778497696, + 0.014062432572245598, + 0.0063172997906804085, + 0.015554200857877731, + 0.011807623319327831, + 0.007028952240943909, + 0.010145936161279678, + 0.013903333805501461, + 0.006785503122955561, + 0.013784837909042835, + 0.007185927126556635, + 0.013181379996240139, + 0.00877876952290535, + 0.00687553221359849, + 0.00994300190359354, + 0.0078028845600783825, + 0.00905732624232769, + 0.012217146344482899, + 0.00995034258812666, + 0.010976282879710197, + 0.022195272147655487, + 0.009592272341251373, + 0.013859246857464314, + 0.013337035663425922, + 0.007664686534553766, + 0.0430411659181118, + 0.00772992055863142, + 0.012501147575676441, + 0.013532769866287708, + 0.007724021095782518, + 0.022607196122407913, + 0.010374057106673717, + 0.008948950096964836, + 0.018282651901245117, + 0.005711657926440239, + 0.01700904220342636, + 0.017283188179135323, + 0.011805073358118534, + 0.02034062333405018, + 0.022807534784078598, + 0.022918701171875, + 0.019080160185694695, + 0.02003209851682186, + 0.024493621662259102 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.5429266691207886 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial view of the scene presented in soft focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive red supergiant star visible on the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense blue-white neutron star visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two stars collide in an accelerating spiral motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene end with a spherical shockwave expanding outward and illuminating the frame in pure white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018234478309750557 + ], + "opposite_logprob_tokens": [ + -4.0182342529296875 + ], + "logprob_reward": -0.018234478309750557, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018234478309750557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3292281925678253, + 0.16242137551307678, + 0.07326982915401459, + 0.05533508583903313, + 0.0771755799651146, + 0.05940811708569527, + 0.06992678344249725, + 0.055273525416851044, + 0.07712366431951523, + 0.04083786904811859 + ], + "credit_map": [ + 0.06729301065206528, + 0.04382498934864998, + 0.0392773412168026, + 0.040587544441223145, + 0.051750678569078445, + 0.04032493382692337, + 0.018663549795746803, + 0.027506142854690552, + 0.0189400315284729, + 0.025634143501520157, + 0.0224494319409132, + 0.03457465022802353, + 0.015467533841729164, + 0.0158479455858469, + 0.01203743927180767, + 0.017470192164182663, + 0.008198730647563934, + 0.009786388836801052, + 0.010988631285727024, + 0.012633607722818851, + 0.006728800479322672, + 0.009231016971170902, + 0.005503165535628796, + 0.010199488140642643, + 0.0054541719146072865, + 0.008671583607792854, + 0.007475594989955425, + 0.011997747235000134, + 0.00389750930480659, + 0.007845309562981129, + 0.003229758469387889, + 0.006763407960534096, + 0.008909658528864384, + 0.0070745935663580894, + 0.010208630934357643, + 0.01625349372625351, + 0.006847726181149483, + 0.01225083228200674, + 0.007495072204619646, + 0.008135567419230938, + 0.006955655291676521, + 0.004974401090294123, + 0.007155403960496187, + 0.00748061528429389, + 0.004062595311552286, + 0.00979425385594368, + 0.005483116023242474, + 0.013502072542905807, + 0.008274722844362259, + 0.006972015369683504, + 0.009135342203080654, + 0.006062205880880356, + 0.0064687649719417095, + 0.011061889119446278, + 0.006387591827660799, + 0.01556425541639328, + 0.008696396835148335, + 0.00286883139051497, + 0.008791830390691757, + 0.005992680322378874, + 0.0028128738049417734, + 0.012190374545753002, + 0.0031397517304867506, + 0.010780785232782364, + 0.009788677096366882, + 0.005393168888986111, + 0.013441631570458412, + 0.007039369083940983, + 0.007946227677166462, + 0.01924276165664196, + 0.004620324820280075, + 0.009651505388319492, + 0.005761357489973307, + 0.0033583305776119232, + 0.00553548289462924, + 0.004972261842340231, + 0.0044544716365635395, + 0.007893563248217106, + 0.0033909850753843784, + 0.0054714130237698555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.018234478309750557 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial view of the scene presented in soft focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive red supergiant star visible on the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense blue-white neutron star visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two stars collide in an accelerating spiral motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene end with a spherical shockwave expanding outward and illuminating the frame in pure white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011083112098276615 + ], + "opposite_logprob_tokens": [ + -4.511083126068115 + ], + "logprob_reward": -0.011083112098276615, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011083112098276615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33491021394729614, + 0.0724027156829834, + 0.08057858794927597, + 0.06329146772623062, + 0.08330834656953812, + 0.035039566457271576, + 0.0849209651350975, + 0.09294260293245316, + 0.052755195647478104, + 0.09985032677650452 + ], + "credit_map": [ + 0.06832854449748993, + 0.03912807255983353, + 0.03498261049389839, + 0.033911533653736115, + 0.061479661613702774, + 0.044494010508060455, + 0.021302493289113045, + 0.031283289194107056, + 0.005950131919234991, + 0.005587801802903414, + 0.006426106207072735, + 0.006410358939319849, + 0.010985835455358028, + 0.010821620002388954, + 0.009808980859816074, + 0.016411880031228065, + 0.010535478591918945, + 0.010827946476638317, + 0.010167009197175503, + 0.01535132434219122, + 0.005643736105412245, + 0.009662709198892117, + 0.006020709872245789, + 0.012369673699140549, + 0.0064733293838799, + 0.008285445161163807, + 0.013003518804907799, + 0.011934391222894192, + 0.004425307735800743, + 0.00870671309530735, + 0.004263334441930056, + 0.006199429742991924, + 0.008990024216473103, + 0.008379440754652023, + 0.010978429578244686, + 0.013870754279196262, + 0.005729974247515202, + 0.012903103604912758, + 0.005864692851901054, + 0.016591930761933327, + 0.0039199525490403175, + 0.0028565656393766403, + 0.0068830810487270355, + 0.0035827201791107655, + 0.003004964906722307, + 0.004921608604490757, + 0.003478808794170618, + 0.0063918656669557095, + 0.01021137461066246, + 0.004660719074308872, + 0.02213696762919426, + 0.006692815106362104, + 0.004451255314052105, + 0.018325375393033028, + 0.005198665428906679, + 0.013243797235190868, + 0.014827465638518333, + 0.0058797248639166355, + 0.015424792654812336, + 0.008647095412015915, + 0.004804480820894241, + 0.02205146662890911, + 0.00519240228459239, + 0.016115177422761917, + 0.006421930156648159, + 0.004133622162044048, + 0.009816494770348072, + 0.0074554020538926125, + 0.004487889818847179, + 0.007803601212799549, + 0.004486903548240662, + 0.008149348199367523, + 0.012048543430864811, + 0.007862946018576622, + 0.017326822504401207, + 0.010206217877566814, + 0.008723906241357327, + 0.012476573698222637, + 0.00858028419315815, + 0.022625036537647247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.011083112098276615 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial view of the scene presented in soft focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive red supergiant star visible on the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense blue-white neutron star visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two stars collide in an accelerating spiral motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene end with a spherical shockwave expanding outward and illuminating the frame in pure white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900409400463104 + ], + "opposite_logprob_tokens": [ + -2.5790040493011475 + ], + "logprob_reward": -0.07900409400463104, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900409400463104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13583487272262573, + 0.0695987418293953, + 0.14743727445602417, + 0.08520934730768204, + 0.0873568207025528, + 0.1451350748538971, + 0.09610705822706223, + 0.09781070053577423, + 0.07809067517518997, + 0.057419437915086746 + ], + "credit_map": [ + 0.027564389631152153, + 0.013614716939628124, + 0.018115639686584473, + 0.01309845969080925, + 0.02364693209528923, + 0.01588013768196106, + 0.01040150411427021, + 0.013513069599866867, + 0.01066705770790577, + 0.009069984778761864, + 0.008726820349693298, + 0.011627952568233013, + 0.006847410928457975, + 0.007796084973961115, + 0.006706972140818834, + 0.008156458847224712, + 0.018712153658270836, + 0.028372948989272118, + 0.015258101746439934, + 0.029745861887931824, + 0.010904594324529171, + 0.016095729544758797, + 0.008636577054858208, + 0.01971128210425377, + 0.010007370263338089, + 0.013262730091810226, + 0.010802468284964561, + 0.018585961312055588, + 0.005628017242997885, + 0.012890879064798355, + 0.0058076088316738605, + 0.008224311284720898, + 0.011822512373328209, + 0.009611970745027065, + 0.013581676408648491, + 0.013930773362517357, + 0.006330745294690132, + 0.016131453216075897, + 0.005307164508849382, + 0.010640518739819527, + 0.02147616632282734, + 0.013405568897724152, + 0.01578734628856182, + 0.022451499477028847, + 0.009221306070685387, + 0.027767596766352654, + 0.012280827388167381, + 0.02274474687874317, + 0.011641934514045715, + 0.008465026505291462, + 0.01334056444466114, + 0.010712871327996254, + 0.010355504229664803, + 0.010862165130674839, + 0.011622542515397072, + 0.019106432795524597, + 0.01455905381590128, + 0.008636271581053734, + 0.011803833767771721, + 0.014916502870619297, + 0.005585997831076384, + 0.025112643837928772, + 0.006247862242162228, + 0.010948524810373783, + 0.011047257110476494, + 0.005249336827546358, + 0.00950400810688734, + 0.009482432156801224, + 0.007593447342514992, + 0.01734805479645729, + 0.0065201641991734505, + 0.011345973238348961, + 0.006456837989389896, + 0.00340771465562284, + 0.009354335255920887, + 0.005769737530499697, + 0.0065995012409985065, + 0.008579809218645096, + 0.005666053853929043, + 0.011585444211959839 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.07900409400463104 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial view of the scene presented in soft focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive red supergiant star visible on the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense blue-white neutron star visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two stars collide in an accelerating spiral motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene end with a spherical shockwave expanding outward and illuminating the frame in pure white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002822346519678831 + ], + "opposite_logprob_tokens": [ + -5.877822399139404 + ], + "logprob_reward": -0.002822346519678831, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002822346519678831, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1760668009519577, + 0.027137553319334984, + 0.04889023303985596, + 0.0283052995800972, + 0.05107804387807846, + 0.037615492939949036, + 0.03905590623617172, + 0.0775151252746582, + 0.15144768357276917, + 0.3628878593444824 + ], + "credit_map": [ + 0.04019268974661827, + 0.02157854661345482, + 0.02010217122733593, + 0.014560319483280182, + 0.03661492094397545, + 0.02365553379058838, + 0.009169085882604122, + 0.010193523950874805, + 0.003761837724596262, + 0.003640878014266491, + 0.0033709148410707712, + 0.004757311195135117, + 0.0027562668547034264, + 0.003037471557036042, + 0.002616605721414089, + 0.0031962681096047163, + 0.0067039052955806255, + 0.007384099531918764, + 0.005859040189534426, + 0.009250108152627945, + 0.0040752519853413105, + 0.005647426471114159, + 0.00346427783370018, + 0.006506126839667559, + 0.003969691693782806, + 0.0032938062213361263, + 0.003165693487972021, + 0.004896730650216341, + 0.0022125961259007454, + 0.004201905336230993, + 0.0019972200971096754, + 0.004567655269056559, + 0.0064675030298531055, + 0.006435296963900328, + 0.007734895683825016, + 0.007152646780014038, + 0.004180673509836197, + 0.008477381430566311, + 0.003453535493463278, + 0.00717611238360405, + 0.004548309836536646, + 0.003482749918475747, + 0.006370334420353174, + 0.005563612561672926, + 0.003110828809440136, + 0.005260929930955172, + 0.0037912146653980017, + 0.005487514194101095, + 0.0047029973939061165, + 0.003932506311684847, + 0.003977505024522543, + 0.007486697286367416, + 0.002487089252099395, + 0.005048074759542942, + 0.004254551604390144, + 0.007166486699134111, + 0.011284113861620426, + 0.004305362235754728, + 0.017224343493580818, + 0.006010147277265787, + 0.004622314125299454, + 0.012667680159211159, + 0.004163168370723724, + 0.01723800040781498, + 0.018008053302764893, + 0.015107395127415657, + 0.027248617261648178, + 0.0216832235455513, + 0.015718499198555946, + 0.019815528765320778, + 0.009800013154745102, + 0.02406635694205761, + 0.049602679908275604, + 0.018652521073818207, + 0.0756632536649704, + 0.022666320204734802, + 0.02510235831141472, + 0.05702845752239227, + 0.02357725240290165, + 0.09059502929449081 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.002822346519678831 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic cosmic panorama, initially presented in soft focus, slowly zooms in on two doomed stars, pulled toward each other by the relentless pull of gravity. Amidst the deep, serene void of space, on the left stands a massive red supergiant star, its massive, swollen surface roiling with a deep crimson of fiery plasma. On the right, a dense, piercing blue-white neutron star, like a deadly cosmic projectile, its immense gravity subtly warping the surrounding space. As the two stars collide in an ever-accelerating spiral, a silent yet visually stunning supernova explosion is triggered at the moment of contact. A spherical shockwave expands outward at near-light speed, instantly engulfing and ripping apart the massive mass of the red giant, unleashing an all-consuming, dazzling radiance whose intensity momentarily surpasses the entire galaxy, illuminating the entire frame in a pure, searing white, creating an atmosphere of cosmic wonder, both majestic and destructive.", + "seed": 2551, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_8/index_8934_seed_2551_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09866554923355579, + "vlm_logprob": -0.09866554923355579, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.09866554923355579, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial view of the scene presented in soft focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive red supergiant star visible on the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense blue-white neutron star visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two stars collide in an accelerating spiral motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene end with a spherical shockwave expanding outward and illuminating the frame in pure white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3778289556503296 + ], + "opposite_logprob_tokens": [ + -1.2528289556503296 + ], + "logprob_reward": -0.3778289556503296, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3778289556503296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08487026393413544, + 0.04778720811009407, + 0.06366585195064545, + 0.08329540491104126, + 0.04254686459898949, + 0.14190328121185303, + 0.0952553004026413, + 0.06929749250411987, + 0.08413578569889069, + 0.287242591381073 + ], + "credit_map": [ + 0.009005011059343815, + 0.0074981385841965675, + 0.01207792665809393, + 0.010328018106520176, + 0.008072975091636181, + 0.006557545159012079, + 0.015307756140828133, + 0.01602289080619812, + 0.009783698245882988, + 0.0034404080361127853, + 0.006990748457610607, + 0.00560973584651947, + 0.0032771280966699123, + 0.0032517756335437298, + 0.0062331934459507465, + 0.00920051708817482, + 0.007791975978761911, + 0.004331070929765701, + 0.01196453906595707, + 0.012657181359827518, + 0.0036728251725435257, + 0.010966595262289047, + 0.0037254109047353268, + 0.008556249551475048, + 0.00880367774516344, + 0.00775372888892889, + 0.010945035144686699, + 0.009265168569982052, + 0.007605261635035276, + 0.01938893459737301, + 0.006326464470475912, + 0.013207117095589638, + 0.0026850912254303694, + 0.0035875693429261446, + 0.008162621408700943, + 0.008816382847726345, + 0.005047852639108896, + 0.004991735797375441, + 0.005810619331896305, + 0.003444985020905733, + 0.012945816852152348, + 0.010727541521191597, + 0.027713578194379807, + 0.017183097079396248, + 0.016492383554577827, + 0.02044404111802578, + 0.021439770236611366, + 0.014957049861550331, + 0.01356157474219799, + 0.007483536843210459, + 0.009213410317897797, + 0.016614854335784912, + 0.009089706465601921, + 0.017099566757678986, + 0.011595803312957287, + 0.01059685181826353, + 0.007958591915667057, + 0.006738412193953991, + 0.007825622335076332, + 0.012459788471460342, + 0.004433509893715382, + 0.010364855639636517, + 0.00960540771484375, + 0.009911289438605309, + 0.01325888279825449, + 0.0049296170473098755, + 0.012549398466944695, + 0.014578050002455711, + 0.006105567794293165, + 0.007232577074319124, + 0.012096807360649109, + 0.013384874910116196, + 0.03699972853064537, + 0.030785949900746346, + 0.1293395459651947, + 0.013524954207241535, + 0.008055991493165493, + 0.009804286062717438, + 0.012313492596149445, + 0.04641861841082573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.3778289556503296 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial view of the scene presented in soft focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive red supergiant star visible on the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense blue-white neutron star visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two stars collide in an accelerating spiral motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene end with a spherical shockwave expanding outward and illuminating the frame in pure white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001750485971570015 + ], + "opposite_logprob_tokens": [ + -6.376750469207764 + ], + "logprob_reward": -0.001750485971570015, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001750485971570015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19257882237434387, + 0.0931326374411583, + 0.17685072124004364, + 0.08431719243526459, + 0.05933329463005066, + 0.08534582704305649, + 0.07254993170499802, + 0.09843514114618301, + 0.055718835443258286, + 0.08173761516809464 + ], + "credit_map": [ + 0.03266303613781929, + 0.012629055418074131, + 0.0330427810549736, + 0.023093417286872864, + 0.015596532262861729, + 0.01647520810365677, + 0.028516145423054695, + 0.030562639236450195, + 0.017059221863746643, + 0.006259736604988575, + 0.013534950092434883, + 0.01578373648226261, + 0.005572964437305927, + 0.006770613603293896, + 0.010831180959939957, + 0.01732023060321808, + 0.019403908401727676, + 0.011262478306889534, + 0.03789550065994263, + 0.037057630717754364, + 0.01125891599804163, + 0.03060230426490307, + 0.01189038623124361, + 0.017479589208960533, + 0.010370520874857903, + 0.006264619994908571, + 0.012068545445799828, + 0.00747996149584651, + 0.008124938234686852, + 0.019550342112779617, + 0.006270120851695538, + 0.014188149012625217, + 0.004054762423038483, + 0.004979869816452265, + 0.011217034421861172, + 0.006784322205930948, + 0.008417388424277306, + 0.008837008848786354, + 0.008491113781929016, + 0.006551791448146105, + 0.0071954322047531605, + 0.009762141853570938, + 0.014191832393407822, + 0.01161977183073759, + 0.011257634498178959, + 0.009123418480157852, + 0.014538784511387348, + 0.007656815927475691, + 0.008943031542003155, + 0.0057624997571110725, + 0.006281679030507803, + 0.012555881403386593, + 0.006509027443826199, + 0.013925242237746716, + 0.00834518764168024, + 0.010227383114397526, + 0.013151384890079498, + 0.005794193129986525, + 0.009355561807751656, + 0.019546054303646088, + 0.004792623687535524, + 0.019066642969846725, + 0.00890628807246685, + 0.01782238483428955, + 0.013354969210922718, + 0.0034334042575210333, + 0.005062389187514782, + 0.007200832013040781, + 0.0032397452741861343, + 0.0050299218855798244, + 0.006045314017683268, + 0.01235225796699524, + 0.006082138512283564, + 0.005587871186435223, + 0.030478736385703087, + 0.005962879862636328, + 0.004054900724440813, + 0.003874976886436343, + 0.005120167508721352, + 0.020575949922204018 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.001750485971570015 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial view of the scene presented in soft focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive red supergiant star visible on the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense blue-white neutron star visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two stars collide in an accelerating spiral motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene end with a spherical shockwave expanding outward and illuminating the frame in pure white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003628220409154892 + ], + "opposite_logprob_tokens": [ + -5.628628253936768 + ], + "logprob_reward": -0.003628220409154892, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003628220409154892, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19900402426719666, + 0.07544475048780441, + 0.1082230657339096, + 0.08038643002510071, + 0.057883162051439285, + 0.10104981064796448, + 0.04207993298768997, + 0.04298882931470871, + 0.04942636936903, + 0.24351361393928528 + ], + "credit_map": [ + 0.03114054538309574, + 0.01192425936460495, + 0.03310319036245346, + 0.01945248246192932, + 0.01683771051466465, + 0.019883988425135612, + 0.031556323170661926, + 0.035105545073747635, + 0.011890572495758533, + 0.004784435965120792, + 0.010344075970351696, + 0.00958947092294693, + 0.007458761800080538, + 0.008121751248836517, + 0.008676525205373764, + 0.01457916758954525, + 0.009230577386915684, + 0.007531259674578905, + 0.011367760598659515, + 0.012577290646731853, + 0.005502792075276375, + 0.024274414405226707, + 0.009005640633404255, + 0.028733335435390472, + 0.008759468793869019, + 0.00597041193395853, + 0.009290275163948536, + 0.00840496551245451, + 0.0068247150629758835, + 0.018705962225794792, + 0.00798380933701992, + 0.01444682851433754, + 0.0042173597030341625, + 0.004927035421133041, + 0.008484241552650928, + 0.008735223673284054, + 0.009243487380445004, + 0.009278221987187862, + 0.007365609519183636, + 0.0056319874711334705, + 0.009020394645631313, + 0.008376284502446651, + 0.02444893680512905, + 0.010708987712860107, + 0.012895827181637287, + 0.011478076688945293, + 0.01234092190861702, + 0.011780383065342903, + 0.004413223825395107, + 0.0033710633870214224, + 0.0054954602383077145, + 0.005867171101272106, + 0.004089548252522945, + 0.007706889417022467, + 0.005094787571579218, + 0.006041791755706072, + 0.006231959443539381, + 0.003186522051692009, + 0.004100825637578964, + 0.006896951701492071, + 0.002362194936722517, + 0.0061104791238904, + 0.004100480582565069, + 0.009999420493841171, + 0.006441473960876465, + 0.004863230977207422, + 0.005609577521681786, + 0.008459771052002907, + 0.003672280814498663, + 0.005238256882876158, + 0.008265108801424503, + 0.006876674480736256, + 0.027764562517404556, + 0.028274158015847206, + 0.10535212606191635, + 0.014018477872014046, + 0.00980475451797247, + 0.009377175010740757, + 0.010535151697695255, + 0.038387231528759 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.003628220409154892 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial view of the scene presented in soft focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive red supergiant star visible on the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense blue-white neutron star visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two stars collide in an accelerating spiral motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene end with a spherical shockwave expanding outward and illuminating the frame in pure white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003350168466568 + ], + "opposite_logprob_tokens": [ + -2.350335121154785 + ], + "logprob_reward": -0.1003350168466568, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003350168466568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06522919237613678, + 0.0380186028778553, + 0.1977900266647339, + 0.0566185861825943, + 0.1676008403301239, + 0.09942242503166199, + 0.09363351762294769, + 0.05684775859117508, + 0.06397269666194916, + 0.160866379737854 + ], + "credit_map": [ + 0.007821277715265751, + 0.0059425607323646545, + 0.0075896563939750195, + 0.008976299315690994, + 0.005844309460371733, + 0.006572422105818987, + 0.010206090286374092, + 0.012276574037969112, + 0.004832631442695856, + 0.0028230291791260242, + 0.004852510057389736, + 0.003912843763828278, + 0.004825405776500702, + 0.005067402496933937, + 0.0039041510317474604, + 0.007800631690770388, + 0.0228124912828207, + 0.009836244396865368, + 0.046303313225507736, + 0.030713673681020737, + 0.011350052431225777, + 0.027823707088828087, + 0.01855045184493065, + 0.030400091782212257, + 0.004131514113396406, + 0.003896262962371111, + 0.004411275964230299, + 0.004216791596263647, + 0.005941824521869421, + 0.017461299896240234, + 0.003910965286195278, + 0.012648648582398891, + 0.014235064387321472, + 0.012479491531848907, + 0.041180506348609924, + 0.02583245001733303, + 0.016387401148676872, + 0.020195845514535904, + 0.02083412930369377, + 0.01645595021545887, + 0.008341493085026741, + 0.010785774327814579, + 0.023975027725100517, + 0.010519676841795444, + 0.010523051023483276, + 0.011833366937935352, + 0.012412804178893566, + 0.011031227186322212, + 0.011534431017935276, + 0.007708499673753977, + 0.009901818819344044, + 0.015105608850717545, + 0.008090411312878132, + 0.016435818746685982, + 0.011074753478169441, + 0.013782169669866562, + 0.006596582941710949, + 0.003293869085609913, + 0.005601339507848024, + 0.013078431598842144, + 0.003312387503683567, + 0.008262302726507187, + 0.005218186881393194, + 0.011484656482934952, + 0.008350460790097713, + 0.005986191798001528, + 0.010214796289801598, + 0.010591496713459492, + 0.005623181816190481, + 0.006229673977941275, + 0.007714746054261923, + 0.00926214735955, + 0.011200420558452606, + 0.010608329437673092, + 0.05851052328944206, + 0.011487717740237713, + 0.009208641946315765, + 0.009334429167211056, + 0.009434270672500134, + 0.041082046926021576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 9 + ], + "selected_reward": -0.1003350168466568 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial view of the scene presented in soft focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive red supergiant star visible on the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense blue-white neutron star visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two stars collide in an accelerating spiral motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene end with a spherical shockwave expanding outward and illuminating the frame in pure white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009785067290067673 + ], + "opposite_logprob_tokens": [ + -4.634785175323486 + ], + "logprob_reward": -0.009785067290067673, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009785067290067673, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03555067628622055, + 0.01732097938656807, + 0.026307612657546997, + 0.021767986938357353, + 0.02247009053826332, + 0.04318320006132126, + 0.06289654225111008, + 0.042287684977054596, + 0.157069593667984, + 0.5711455941200256 + ], + "credit_map": [ + 0.006290560122579336, + 0.002964574610814452, + 0.004646638873964548, + 0.005472512915730476, + 0.0030945134349167347, + 0.0024958555586636066, + 0.0048845382407307625, + 0.00570148229598999, + 0.002294693375006318, + 0.001017085975036025, + 0.0023347740061581135, + 0.003519420512020588, + 0.0016135243931785226, + 0.0012620665365830064, + 0.002185852499678731, + 0.0030935630202293396, + 0.0038057398051023483, + 0.0022016381844878197, + 0.004747174214571714, + 0.004980452358722687, + 0.0022172185126692057, + 0.0032667231280356646, + 0.0021285032853484154, + 0.0029601617716252804, + 0.0028379166033118963, + 0.0017435357440263033, + 0.0037862861063331366, + 0.0027547942008823156, + 0.002324509434401989, + 0.0034973653964698315, + 0.0022485200315713882, + 0.0025750589556992054, + 0.001583216479048133, + 0.0018431898206472397, + 0.004501833580434322, + 0.004122241400182247, + 0.0027911581564694643, + 0.00249510258436203, + 0.002935719210654497, + 0.0021976283751428127, + 0.0037179409991949797, + 0.003850538283586502, + 0.007959136739373207, + 0.0052377465181052685, + 0.0043827202171087265, + 0.005073572974652052, + 0.008100810460746288, + 0.004860736429691315, + 0.0075388820841908455, + 0.00538556557148695, + 0.007086015772074461, + 0.011540969833731651, + 0.006734034977853298, + 0.008696896024048328, + 0.007750356569886208, + 0.008163818158209324, + 0.006618275307118893, + 0.0038143254350870848, + 0.005018074531108141, + 0.007966556586325169, + 0.003394997911527753, + 0.00527899619191885, + 0.003532473463565111, + 0.0066639878787100315, + 0.029600679874420166, + 0.010700702667236328, + 0.022593539208173752, + 0.027360592037439346, + 0.00812771450728178, + 0.010573925450444221, + 0.016845446079969406, + 0.03126699477434158, + 0.05515522509813309, + 0.044060010462999344, + 0.2483212798833847, + 0.026946257799863815, + 0.019091365858912468, + 0.02108508162200451, + 0.02396557852625847, + 0.13252080976963043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.009785067290067673 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic cosmic panorama, initially presented in soft focus, slowly zooms in on two doomed stars, pulled toward each other by the relentless pull of gravity. Amidst the deep, serene void of space, on the left stands a massive red supergiant star, its massive, swollen surface roiling with a deep crimson of fiery plasma. On the right, a dense, piercing blue-white neutron star, like a deadly cosmic projectile, its immense gravity subtly warping the surrounding space. As the two stars collide in an ever-accelerating spiral, a silent yet visually stunning supernova explosion is triggered at the moment of contact. A spherical shockwave expands outward at near-light speed, instantly engulfing and ripping apart the massive mass of the red giant, unleashing an all-consuming, dazzling radiance whose intensity momentarily surpasses the entire galaxy, illuminating the entire frame in a pure, searing white, creating an atmosphere of cosmic wonder, both majestic and destructive.", + "seed": 2552, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_8/index_8934_seed_2552_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.3422635061666369, + "vlm_logprob": -0.3422635061666369, + "vlm_yes_no_margin": 2.199999952316284, + "vlm_token_credit": -0.3422635061666369, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial view of the scene presented in soft focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive red supergiant star visible on the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense blue-white neutron star visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two stars collide in an accelerating spiral motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene end with a spherical shockwave expanding outward and illuminating the frame in pure white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3668729364871979 + ], + "opposite_logprob_tokens": [ + -1.2418729066848755 + ], + "logprob_reward": -0.3668729364871979, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3668729364871979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5163037180900574, + 0.13363686203956604, + 0.04556359723210335, + 0.027535708621144295, + 0.024233415722846985, + 0.04523678123950958, + 0.04831269383430481, + 0.09176758676767349, + 0.0404568687081337, + 0.02695283479988575 + ], + "credit_map": [ + 0.034457962960004807, + 0.019141996279358864, + 0.02965928241610527, + 0.03689522668719292, + 0.03978920355439186, + 0.025009071454405785, + 0.16409820318222046, + 0.16725273430347443, + 0.008241571485996246, + 0.005537286400794983, + 0.006411432288587093, + 0.016613435000181198, + 0.020641474053263664, + 0.009843967854976654, + 0.043717168271541595, + 0.02263053134083748, + 0.005545484833419323, + 0.004668247886002064, + 0.00400598393753171, + 0.0067620412446558475, + 0.006989629473537207, + 0.0047418540343642235, + 0.007474214769899845, + 0.005376138724386692, + 0.002183042000979185, + 0.0035933780018240213, + 0.003003231482580304, + 0.004757000133395195, + 0.0044489516876637936, + 0.002731455024331808, + 0.003906281664967537, + 0.0029123681597411633, + 0.002588511910289526, + 0.0029759248718619347, + 0.002084024017676711, + 0.0034944142680615187, + 0.005099443253129721, + 0.0027538693975657225, + 0.003379005705937743, + 0.0018582235788926482, + 0.006290164310485125, + 0.0034691670443862677, + 0.004356961231678724, + 0.005811760202050209, + 0.005734540522098541, + 0.010781971737742424, + 0.005143825896084309, + 0.0036483893636614084, + 0.007826060988008976, + 0.003463359782472253, + 0.0050038425251841545, + 0.007192280143499374, + 0.0032637310214340687, + 0.009885312058031559, + 0.0068195778876543045, + 0.004858528729528189, + 0.028692010790109634, + 0.0035275090485811234, + 0.003959454596042633, + 0.0033653751015663147, + 0.0027278419584035873, + 0.00906936265528202, + 0.004708358086645603, + 0.0357176810503006, + 0.0029167253524065018, + 0.005849769804626703, + 0.004594665020704269, + 0.005887279752641916, + 0.0074408818036317825, + 0.002676489297300577, + 0.007500994950532913, + 0.0035900596994906664, + 0.002747819758951664, + 0.0032278106082230806, + 0.003466559574007988, + 0.005579335615038872, + 0.00343241635710001, + 0.002692232374101877, + 0.003274360438808799, + 0.0025322993751615286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.3668729364871979 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial view of the scene presented in soft focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive red supergiant star visible on the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense blue-white neutron star visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two stars collide in an accelerating spiral motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene end with a spherical shockwave expanding outward and illuminating the frame in pure white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9743388295173645 + ], + "opposite_logprob_tokens": [ + -0.4743388295173645 + ], + "logprob_reward": -0.9743388295173645, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9743388295173645, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.47662413120269775, + 0.20238567888736725, + 0.08786380290985107, + 0.038164928555488586, + 0.09216983616352081, + 0.026017187163233757, + 0.02000037394464016, + 0.03893358260393143, + 0.009524771012365818, + 0.008315669372677803 + ], + "credit_map": [ + 0.05236732214689255, + 0.027833469212055206, + 0.056054919958114624, + 0.046413443982601166, + 0.06220676377415657, + 0.03480759263038635, + 0.15374985337257385, + 0.04319077357649803, + 0.014286764897406101, + 0.008905255235731602, + 0.010714787989854813, + 0.024772290140390396, + 0.029289644211530685, + 0.016669591888785362, + 0.06292718648910522, + 0.03482016921043396, + 0.010024290531873703, + 0.00911561120301485, + 0.007366142701357603, + 0.013034079223871231, + 0.014088604599237442, + 0.010939907282590866, + 0.013071388937532902, + 0.01022377610206604, + 0.002987832995131612, + 0.004271713551133871, + 0.004265218041837215, + 0.006536278873682022, + 0.006377736106514931, + 0.004266772419214249, + 0.006154177710413933, + 0.0033051990903913975, + 0.015578998252749443, + 0.009434353560209274, + 0.0059231119230389595, + 0.010875537060201168, + 0.013066796585917473, + 0.02108008600771427, + 0.009954242967069149, + 0.006256715394556522, + 0.003889942541718483, + 0.002233371837064624, + 0.0023523771669715643, + 0.0035647519398480654, + 0.00310957757756114, + 0.005591479130089283, + 0.002749938052147627, + 0.002525751478970051, + 0.003157627070322633, + 0.0015889040660113096, + 0.0025545416865497828, + 0.002622418338432908, + 0.0019847627263516188, + 0.0044400994665920734, + 0.0023120963014662266, + 0.0013399263843894005, + 0.01157323457300663, + 0.0015262673841789365, + 0.0015862605068832636, + 0.0012829079059883952, + 0.0017539110267534852, + 0.0040513016283512115, + 0.00242704339325428, + 0.014732660725712776, + 0.0007627049344591796, + 0.0013054363662377, + 0.001652859733439982, + 0.0013765925541520119, + 0.0014287150697782636, + 0.0008989272173494101, + 0.0008554839878343046, + 0.0012440518476068974, + 0.0007489901036024094, + 0.001276167924515903, + 0.000752535997889936, + 0.0011849245056509972, + 0.001544890576042235, + 0.0009467679192312062, + 0.0010982771636918187, + 0.0007631154148839414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.9743388295173645 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial view of the scene presented in soft focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive red supergiant star visible on the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense blue-white neutron star visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two stars collide in an accelerating spiral motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene end with a spherical shockwave expanding outward and illuminating the frame in pure white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34869620203971863 + ], + "opposite_logprob_tokens": [ + -1.223696231842041 + ], + "logprob_reward": -0.34869620203971863, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34869620203971863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.420422226190567, + 0.10717658698558807, + 0.11301389336585999, + 0.037176888436079025, + 0.06553450226783752, + 0.03826788440346718, + 0.07777590304613113, + 0.10951539129018784, + 0.020402023568749428, + 0.010714633390307426 + ], + "credit_map": [ + 0.05316196009516716, + 0.02899613231420517, + 0.058464325964450836, + 0.03870683163404465, + 0.04741157591342926, + 0.030991286039352417, + 0.0883113220334053, + 0.0743788480758667, + 0.012535041198134422, + 0.009287300519645214, + 0.010919426567852497, + 0.013234068639576435, + 0.009426209144294262, + 0.017586883157491684, + 0.020526288077235222, + 0.013661383651196957, + 0.012366191484034061, + 0.008481952361762524, + 0.010562594048678875, + 0.011496440507471561, + 0.013353909365832806, + 0.026242244988679886, + 0.014458158053457737, + 0.01605241745710373, + 0.003994803410023451, + 0.004607334733009338, + 0.0053759473375976086, + 0.004845624789595604, + 0.006064063403755426, + 0.003698123386129737, + 0.0054414356127381325, + 0.003149563679471612, + 0.011059438809752464, + 0.006141732446849346, + 0.005665711127221584, + 0.0060705579817295074, + 0.005297783296555281, + 0.018975108861923218, + 0.005732250865548849, + 0.006591931916773319, + 0.006202804856002331, + 0.00399284902960062, + 0.0029421828221529722, + 0.006252497900277376, + 0.0043649980798363686, + 0.004512378480285406, + 0.005912438966333866, + 0.004087741952389479, + 0.013742979615926743, + 0.004979231394827366, + 0.008489913307130337, + 0.009863409213721752, + 0.0053347996436059475, + 0.02150067687034607, + 0.007712203543633223, + 0.00615270622074604, + 0.03393403813242912, + 0.004049492534250021, + 0.004618773236870766, + 0.003164758440107107, + 0.0024569514207541943, + 0.011709862388670444, + 0.005632386077195406, + 0.043949149549007416, + 0.0016953462036326528, + 0.0017097144154831767, + 0.004756391979753971, + 0.003193024778738618, + 0.0027119978331029415, + 0.0014715340221300721, + 0.003324573626741767, + 0.001539443852379918, + 0.0010205225553363562, + 0.0010597455548122525, + 0.0015830331249162555, + 0.002181439660489559, + 0.0014826369006186724, + 0.0009847497567534447, + 0.0013462937204167247, + 0.0010562147945165634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.34869620203971863 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial view of the scene presented in soft focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive red supergiant star visible on the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense blue-white neutron star visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two stars collide in an accelerating spiral motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene end with a spherical shockwave expanding outward and illuminating the frame in pure white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018212003633379936 + ], + "opposite_logprob_tokens": [ + -4.018211841583252 + ], + "logprob_reward": -0.018212003633379936, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018212003633379936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4933178126811981, + 0.11319126188755035, + 0.07743526250123978, + 0.03290541097521782, + 0.03470967710018158, + 0.03026489168405533, + 0.036578018218278885, + 0.03100338578224182, + 0.13531352579593658, + 0.015280747786164284 + ], + "credit_map": [ + 0.051321763545274734, + 0.032125212252140045, + 0.05699937418103218, + 0.035124771296978, + 0.03294114023447037, + 0.03506409376859665, + 0.08132098615169525, + 0.1684204787015915, + 0.00971745140850544, + 0.006306060124188662, + 0.007721558213233948, + 0.013346212916076183, + 0.014570650644600391, + 0.011389478109776974, + 0.0314834900200367, + 0.01865636743605137, + 0.008599129505455494, + 0.00769776152446866, + 0.005880917422473431, + 0.010440591722726822, + 0.01213082019239664, + 0.01271816249936819, + 0.010139217600226402, + 0.009828670881688595, + 0.003218467812985182, + 0.004721254575997591, + 0.0044113583862781525, + 0.005272249691188335, + 0.004903400782495737, + 0.0035525416024029255, + 0.004451317712664604, + 0.002374818082898855, + 0.003710700897499919, + 0.003849297296255827, + 0.004082747735083103, + 0.004209802020341158, + 0.004475054796785116, + 0.006056732963770628, + 0.004305108916014433, + 0.004020229913294315, + 0.004340385086834431, + 0.0026940887328237295, + 0.002777357120066881, + 0.004332253709435463, + 0.003699822584167123, + 0.006781182251870632, + 0.0035772514529526234, + 0.0020625514443963766, + 0.005103024188429117, + 0.0035031249281018972, + 0.0030314770992845297, + 0.004846619907766581, + 0.003954427782446146, + 0.005985018331557512, + 0.006947766523808241, + 0.003206560155376792, + 0.002933628624305129, + 0.0037886081263422966, + 0.004279796965420246, + 0.005128163378685713, + 0.004624101798981428, + 0.0026684694457799196, + 0.004463253077119589, + 0.0031173638999462128, + 0.0065840017050504684, + 0.01800413616001606, + 0.011163336224853992, + 0.03103419579565525, + 0.018436171114444733, + 0.006006248760968447, + 0.038166146725416183, + 0.005919283255934715, + 0.0018012482905760407, + 0.002052354160696268, + 0.0020002240780740976, + 0.0028188598807901144, + 0.0021190906409174204, + 0.0010242494754493237, + 0.0021083285100758076, + 0.0013563919346779585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.018212003633379936 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial view of the scene presented in soft focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive red supergiant star visible on the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense blue-white neutron star visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two stars collide in an accelerating spiral motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene end with a spherical shockwave expanding outward and illuminating the frame in pure white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003197559155523777 + ], + "opposite_logprob_tokens": [ + -5.75319766998291 + ], + "logprob_reward": -0.003197559155523777, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003197559155523777, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4756810665130615, + 0.16495011746883392, + 0.053381916135549545, + 0.03441598638892174, + 0.04435141757130623, + 0.018221747130155563, + 0.018600402399897575, + 0.039144646376371384, + 0.06766945123672485, + 0.0835832953453064 + ], + "credit_map": [ + 0.05236309766769409, + 0.030315585434436798, + 0.060768503695726395, + 0.03553761541843414, + 0.032175350934267044, + 0.0347541905939579, + 0.07018056511878967, + 0.1595861315727234, + 0.012746525928378105, + 0.007632410619407892, + 0.00839329231530428, + 0.019305862486362457, + 0.02462157979607582, + 0.011998884379863739, + 0.05216633155941963, + 0.02808520756661892, + 0.007099918555468321, + 0.0058770556934177876, + 0.004896068014204502, + 0.007273938041180372, + 0.008653903380036354, + 0.006984966807067394, + 0.006736631505191326, + 0.005859425291419029, + 0.0029561349656432867, + 0.003697529435157776, + 0.0035073263570666313, + 0.0051990593783557415, + 0.005822193808853626, + 0.004856848157942295, + 0.0052480450831353664, + 0.003128846175968647, + 0.007182786241173744, + 0.004149060230702162, + 0.0035981154069304466, + 0.0058944872580468655, + 0.006852783262729645, + 0.00877195131033659, + 0.004964357241988182, + 0.0029378756880760193, + 0.0024681726936250925, + 0.001368221128359437, + 0.0023770194966346025, + 0.002361746272072196, + 0.002305760979652405, + 0.003926469013094902, + 0.0020858743228018284, + 0.0013284803135320544, + 0.00329801463522017, + 0.001873771077953279, + 0.0015175972366705537, + 0.002718371106311679, + 0.0014641847228631377, + 0.002317592967301607, + 0.0026691602542996407, + 0.002741709351539612, + 0.00853489339351654, + 0.0018320099916309118, + 0.0032772934064269066, + 0.00520113343372941, + 0.0023777943570166826, + 0.0036493027582764626, + 0.0032259919680655003, + 0.011046221479773521, + 0.005850541405379772, + 0.008835311979055405, + 0.007638515438884497, + 0.007235715165734291, + 0.013572701252996922, + 0.006526173558086157, + 0.00762175302952528, + 0.01038872916251421, + 0.009520182386040688, + 0.008992322720587254, + 0.009822280146181583, + 0.01295226439833641, + 0.01334631908684969, + 0.010266425088047981, + 0.01077619194984436, + 0.007907294668257236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.003197559155523777 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_9/index_tensor([7491, 7491])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_9/index_tensor([7491, 7491])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7d7c514b758cacbecf48bd7759ab937bde319fcd --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_9/index_tensor([7491, 7491])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In an extreme macro close-up, the camera moves forward at an almost imperceptible pace, focusing on a butterfly perched on the petals of a pink peony, drenched in morning dew. Soft morning sunlight filters through a hazy green garden in the background, precisely outlining the butterfly's silhouette. Its wings, initially closed, resemble a delicate, folded jewel, covered in microscopic iridescent scales. A crisscross of sapphire, emerald green, and flame orange creates a symmetrical and intricate pattern like a cathedral stained glass window, threaded with delicate black veins. With a barely perceptible flutter, its paper-like, fragile wings begin to unfold, inch by inch, with a steady, deliberate rhythm, a slow, ritualistic process, until they are fully extended, revealing their breathtaking full range of color and pattern. The edges of the wings glow softly in the golden light, bathing the entire image in a warm glow, creating a tranquil, magical atmosphere filled with the wonder of life.", + "seed": 22318, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_9/index_7491_seed_22318_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04941128771752119, + "vlm_logprob": -0.04941128771752119, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.04941128771752119, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a butterfly perched on a pink peony?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a hazy green garden with soft morning sunlight filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the butterfly's wings initially closed, resembling a folded jewel with microscopic iridescent scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wings unfold slowly and deliberately from a closed state to a fully extended state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward at an almost imperceptible pace throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01339931134134531 + ], + "opposite_logprob_tokens": [ + -5.638399124145508 + ], + "logprob_reward": -0.01339931134134531, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.01339931134134531, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09225717931985855, + 0.04642793536186218, + 0.025055283680558205, + 0.036957673728466034, + 0.03830001503229141, + 0.05500207841396332, + 0.04568314924836159, + 0.31477805972099304, + 0.11878310889005661, + 0.22675549983978271 + ], + "credit_map": [ + 0.012979441322386265, + 0.012362932786345482, + 0.019495386630296707, + 0.007822935469448566, + 0.012620407156646252, + 0.011791751720011234, + 0.008718017488718033, + 0.006466309539973736, + 0.00430441414937377, + 0.0081191910430789, + 0.006490756291896105, + 0.0057053216733038425, + 0.007622161880135536, + 0.005179558414965868, + 0.004868146497756243, + 0.004138384945690632, + 0.0038888435810804367, + 0.0037340247072279453, + 0.0035920373629778624, + 0.002736806869506836, + 0.0033257068134844303, + 0.0028883921913802624, + 0.002715826965868473, + 0.002173646120354533, + 0.0028379864525049925, + 0.00557008758187294, + 0.005479676648974419, + 0.006671030540019274, + 0.003935929853469133, + 0.005185706540942192, + 0.0034800488501787186, + 0.0037972053978592157, + 0.003502852050587535, + 0.007066745311021805, + 0.005150610115379095, + 0.005496414843946695, + 0.0063622803427278996, + 0.004982131067663431, + 0.00310450978577137, + 0.0026344729121774435, + 0.004761733114719391, + 0.00936963316053152, + 0.011574474163353443, + 0.006225689314305782, + 0.007557875011116266, + 0.007434105034917593, + 0.004000956658273935, + 0.004077611491084099, + 0.004186928737908602, + 0.007630256004631519, + 0.006999204400926828, + 0.004635432735085487, + 0.008666322566568851, + 0.006193104200065136, + 0.0038925777189433575, + 0.0034793252125382423, + 0.04860245808959007, + 0.04212549701333046, + 0.05587829276919365, + 0.020393697544932365, + 0.051310937851667404, + 0.03957543149590492, + 0.027139615267515182, + 0.02975212223827839, + 0.009895853698253632, + 0.00992554146796465, + 0.02529461309313774, + 0.010005645453929901, + 0.02344338782131672, + 0.025759533047676086, + 0.007817844860255718, + 0.006640682462602854, + 0.012282035313546658, + 0.011063461191952229, + 0.07893747836351395, + 0.011722533963620663, + 0.017020242288708687, + 0.07647687196731567, + 0.009727900847792625, + 0.009524978697299957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.01339931134134531 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a butterfly perched on a pink peony?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a hazy green garden with soft morning sunlight filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the butterfly's wings initially closed, resembling a folded jewel with microscopic iridescent scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wings unfold slowly and deliberately from a closed state to a fully extended state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward at an almost imperceptible pace throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060437947511673 + ], + "opposite_logprob_tokens": [ + -3.895604372024536 + ], + "logprob_reward": -0.02060437947511673, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060437947511673, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03930114954710007, + 0.015123381279408932, + 0.025759806856513023, + 0.02504354529082775, + 0.03362754359841347, + 0.017151545733213425, + 0.014048530720174313, + 0.7659534215927124, + 0.01181701384484768, + 0.05217406526207924 + ], + "credit_map": [ + 0.0038877902552485466, + 0.0071114483289420605, + 0.006913234479725361, + 0.002900614170357585, + 0.006759587675333023, + 0.00786492507904768, + 0.0019698627293109894, + 0.0018936831038445234, + 0.0013993923785164952, + 0.0023185547906905413, + 0.0017436824273318052, + 0.0014742545317858458, + 0.003196484176442027, + 0.002141625387594104, + 0.0015558608574792743, + 0.00129352614749223, + 0.002734400797635317, + 0.004232462495565414, + 0.005899421405047178, + 0.0022480946499854326, + 0.0030541548039764166, + 0.003731840057298541, + 0.0018947155913338065, + 0.001964715775102377, + 0.0018939401488751173, + 0.00431849667802453, + 0.00449020741507411, + 0.004393347073346376, + 0.0037420853041112423, + 0.0032406167592853308, + 0.001769230468198657, + 0.0011956201633438468, + 0.0027820984832942486, + 0.005725198425352573, + 0.006827364210039377, + 0.0035113231278955936, + 0.005952584557235241, + 0.0047512901946902275, + 0.0022984202951192856, + 0.0017792588332667947, + 0.0018581913318485022, + 0.0024952797684818506, + 0.0034864037297666073, + 0.0013536070473492146, + 0.002682191552594304, + 0.00230166083201766, + 0.0012574531137943268, + 0.0017167555633932352, + 0.0016353242099285126, + 0.0021387289743870497, + 0.002431359374895692, + 0.0015166501980274916, + 0.001988881966099143, + 0.0016961131477728486, + 0.0015435523819178343, + 0.0010979194194078445, + 0.11907033622264862, + 0.10167604684829712, + 0.14155443012714386, + 0.04989950731396675, + 0.12619289755821228, + 0.09943686425685883, + 0.06462696939706802, + 0.06349635869264603, + 0.0015974147245287895, + 0.0013582449173554778, + 0.001917174900881946, + 0.0015424346784129739, + 0.0021088842768222094, + 0.0015651206485927105, + 0.0009510267991572618, + 0.0007767115603201091, + 0.006493150256574154, + 0.004604884423315525, + 0.01398716401308775, + 0.003470203373581171, + 0.005561464000493288, + 0.012020668014883995, + 0.0033644915092736483, + 0.002672035712748766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.02060437947511673 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a butterfly perched on a pink peony?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a hazy green garden with soft morning sunlight filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the butterfly's wings initially closed, resembling a folded jewel with microscopic iridescent scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wings unfold slowly and deliberately from a closed state to a fully extended state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward at an almost imperceptible pace throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635989524424076 + ], + "opposite_logprob_tokens": [ + -3.651359796524048 + ], + "logprob_reward": -0.02635989524424076, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635989524424076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11725769191980362, + 0.08152070641517639, + 0.0750415101647377, + 0.0454576276242733, + 0.02005651593208313, + 0.03288283944129944, + 0.1012520119547844, + 0.18397749960422516, + 0.15195202827453613, + 0.19060157239437103 + ], + "credit_map": [ + 0.01278553158044815, + 0.024090267717838287, + 0.019929679110646248, + 0.012108659371733665, + 0.01441859919577837, + 0.01636664941906929, + 0.009845569729804993, + 0.007712731137871742, + 0.0075787934474647045, + 0.017091432586312294, + 0.012124327942728996, + 0.013300764374434948, + 0.010724151507019997, + 0.009173674508929253, + 0.0067820558324456215, + 0.004745509475469589, + 0.0060900249518454075, + 0.013770782388746738, + 0.014136233367025852, + 0.012976499274373055, + 0.009836374782025814, + 0.009116916917264462, + 0.005609082989394665, + 0.003505595261231065, + 0.0038099910598248243, + 0.010005315765738487, + 0.006749776657670736, + 0.007439462933689356, + 0.006905390881001949, + 0.0047873565927147865, + 0.0034057567827403545, + 0.002354577649384737, + 0.0020052429754287004, + 0.0028960539493709803, + 0.0033755244221538305, + 0.0028819481376558542, + 0.003145553870126605, + 0.0026622037403285503, + 0.0018992472905665636, + 0.0011907407315447927, + 0.002938091056421399, + 0.006807658355683088, + 0.005995933897793293, + 0.002988806925714016, + 0.004898732528090477, + 0.004447480663657188, + 0.002049047499895096, + 0.002757088514044881, + 0.0062414673157036304, + 0.010487268678843975, + 0.02922816015779972, + 0.005443098489195108, + 0.012370915152132511, + 0.0279936995357275, + 0.004707830026745796, + 0.004779573064297438, + 0.02570086158812046, + 0.029446996748447418, + 0.033809829503297806, + 0.013473386876285076, + 0.027465997263789177, + 0.02342902310192585, + 0.014421295374631882, + 0.016230111941695213, + 0.010971518233418465, + 0.0091393506154418, + 0.03026241436600685, + 0.010153786279261112, + 0.04112972319126129, + 0.033551234751939774, + 0.009383718483150005, + 0.007360272109508514, + 0.010590660385787487, + 0.012434644624590874, + 0.05670705437660217, + 0.01368884090334177, + 0.01725098118185997, + 0.057189226150512695, + 0.01103894878178835, + 0.011701235547661781 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.02635989524424076 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a butterfly perched on a pink peony?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a hazy green garden with soft morning sunlight filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the butterfly's wings initially closed, resembling a folded jewel with microscopic iridescent scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wings unfold slowly and deliberately from a closed state to a fully extended state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward at an almost imperceptible pace throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026342827826738358 + ], + "opposite_logprob_tokens": [ + -3.6513428688049316 + ], + "logprob_reward": -0.026342827826738358, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026342827826738358, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1107902079820633, + 0.03543951362371445, + 0.038800884038209915, + 0.026414675638079643, + 0.04058927297592163, + 0.03999654948711395, + 0.05931403115391731, + 0.11016054451465607, + 0.259560227394104, + 0.2789340913295746 + ], + "credit_map": [ + 0.012798444367945194, + 0.02107093296945095, + 0.021487709134817123, + 0.010435310192406178, + 0.013270415365695953, + 0.015042715705931187, + 0.009101726114749908, + 0.007582955062389374, + 0.0024876040406525135, + 0.0060439156368374825, + 0.003994823433458805, + 0.005188203416764736, + 0.006602001842111349, + 0.005200239364057779, + 0.0034172600135207176, + 0.002505464944988489, + 0.0029973804485052824, + 0.006662651430815458, + 0.007246233522891998, + 0.005296782590448856, + 0.006152919493615627, + 0.004896897356957197, + 0.0030735600739717484, + 0.002474457025527954, + 0.002574379090219736, + 0.004412222653627396, + 0.005764340981841087, + 0.002129239495843649, + 0.004412595182657242, + 0.003239703131839633, + 0.002035797806456685, + 0.0018463967135176063, + 0.003890693187713623, + 0.006322301458567381, + 0.008378962986171246, + 0.00543479947373271, + 0.004876752384006977, + 0.006321846041828394, + 0.0030311662703752518, + 0.0023327546659857035, + 0.0037209137808531523, + 0.007628718391060829, + 0.007769907359033823, + 0.003671499202027917, + 0.007483422756195068, + 0.005443600006401539, + 0.0022533617448061705, + 0.002025127410888672, + 0.004471793305128813, + 0.0061118509620428085, + 0.01564960926771164, + 0.004045996814966202, + 0.007427010219544172, + 0.015307684428989887, + 0.002931916620582342, + 0.0033681674394756556, + 0.015380577184259892, + 0.01634579710662365, + 0.016165221109986305, + 0.009818056598305702, + 0.02186398208141327, + 0.013845431618392467, + 0.009113349951803684, + 0.007628129329532385, + 0.01809300109744072, + 0.016709767282009125, + 0.05181976780295372, + 0.017446177080273628, + 0.07146115601062775, + 0.05736243724822998, + 0.014276968315243721, + 0.012390938587486744, + 0.015167342498898506, + 0.01810397394001484, + 0.08504417538642883, + 0.019450588151812553, + 0.02311309613287449, + 0.08546057343482971, + 0.015154874883592129, + 0.01743946224451065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.026342827826738358 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a butterfly perched on a pink peony?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a hazy green garden with soft morning sunlight filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the butterfly's wings initially closed, resembling a folded jewel with microscopic iridescent scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wings unfold slowly and deliberately from a closed state to a fully extended state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward at an almost imperceptible pace throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603500247001648 + ], + "opposite_logprob_tokens": [ + -1.9103500843048096 + ], + "logprob_reward": -0.1603500247001648, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603500247001648, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04126367345452309, + 0.02317696437239647, + 0.0321229062974453, + 0.05018937960267067, + 0.02742108330130577, + 0.03701040521264076, + 0.06974753737449646, + 0.5729424357414246, + 0.051743876188993454, + 0.09438171237707138 + ], + "credit_map": [ + 0.004785784985870123, + 0.00961389858275652, + 0.004613079596310854, + 0.004126960877329111, + 0.007833980955183506, + 0.00462215906009078, + 0.0033453877549618483, + 0.0023224251344799995, + 0.0022431754041463137, + 0.0033342649694532156, + 0.0031357642728835344, + 0.002751431195065379, + 0.004787186626344919, + 0.0025377729907631874, + 0.0023230724036693573, + 0.0020642962772399187, + 0.0033724522218108177, + 0.00370943546295166, + 0.005900327581912279, + 0.005233370698988438, + 0.00382631435059011, + 0.003946348093450069, + 0.0031144351232796907, + 0.003020223695784807, + 0.0036242178175598383, + 0.01112110260874033, + 0.008245973847806454, + 0.008529463782906532, + 0.007213670760393143, + 0.005985990632325411, + 0.0034339979756623507, + 0.002034959848970175, + 0.0023381009232252836, + 0.005530295893549919, + 0.004401406738907099, + 0.0031181261874735355, + 0.004646942485123873, + 0.0036045811139047146, + 0.0019695127848535776, + 0.0018121175235137343, + 0.003714250400662422, + 0.00626123184338212, + 0.006925774272531271, + 0.003823326900601387, + 0.006171850487589836, + 0.00499561196193099, + 0.0024648220278322697, + 0.0026535361539572477, + 0.00590237881988287, + 0.011294670403003693, + 0.013696991838514805, + 0.0054911058396101, + 0.012725429609417915, + 0.010601303540170193, + 0.0051175630651414394, + 0.004918097984045744, + 0.09022178500890732, + 0.07722444087266922, + 0.11122433841228485, + 0.03795553371310234, + 0.09423109143972397, + 0.07856108248233795, + 0.04281734675168991, + 0.0407068133354187, + 0.006294238846749067, + 0.004718568176031113, + 0.006102173589169979, + 0.0062612988986074924, + 0.012994096614420414, + 0.007402737159281969, + 0.0044706943444907665, + 0.0035000741481781006, + 0.008066078647971153, + 0.006289791781455278, + 0.03099588118493557, + 0.0049430448561906815, + 0.007134712301194668, + 0.028814880177378654, + 0.0045957076363265514, + 0.0035416195169091225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.1603500247001648 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme macro close-up, the camera moves forward at an almost imperceptible pace, focusing on a butterfly perched on the petals of a pink peony, drenched in morning dew. Soft morning sunlight filters through a hazy green garden in the background, precisely outlining the butterfly's silhouette. Its wings, initially closed, resemble a delicate, folded jewel, covered in microscopic iridescent scales. A crisscross of sapphire, emerald green, and flame orange creates a symmetrical and intricate pattern like a cathedral stained glass window, threaded with delicate black veins. With a barely perceptible flutter, its paper-like, fragile wings begin to unfold, inch by inch, with a steady, deliberate rhythm, a slow, ritualistic process, until they are fully extended, revealing their breathtaking full range of color and pattern. The edges of the wings glow softly in the golden light, bathing the entire image in a warm glow, creating a tranquil, magical atmosphere filled with the wonder of life.", + "seed": 22319, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_9/index_7491_seed_22319_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05512512307614088, + "vlm_logprob": -0.05512512307614088, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.05512512307614088, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a butterfly perched on a pink peony?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a hazy green garden with soft morning sunlight filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the butterfly's wings initially closed, resembling a folded jewel with microscopic iridescent scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wings unfold slowly and deliberately from a closed state to a fully extended state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward at an almost imperceptible pace throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009136062115430832 + ], + "opposite_logprob_tokens": [ + -5.759136199951172 + ], + "logprob_reward": -0.009136062115430832, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.009136062115430832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08717811852693558, + 0.09807959198951721, + 0.10674866288900375, + 0.16881440579891205, + 0.06411315500736237, + 0.13976339995861053, + 0.06849329173564911, + 0.12684288620948792, + 0.0632852092385292, + 0.07668131589889526 + ], + "credit_map": [ + 0.010440894402563572, + 0.01158828753978014, + 0.023606104776263237, + 0.009314976632595062, + 0.005704415962100029, + 0.0069257952272892, + 0.011349707841873169, + 0.008247928693890572, + 0.010861963964998722, + 0.011031296104192734, + 0.01982322707772255, + 0.013278720900416374, + 0.01133723184466362, + 0.006460161413997412, + 0.007376049179583788, + 0.017910940572619438, + 0.013022893108427525, + 0.022816888988018036, + 0.010422848165035248, + 0.013921665959060192, + 0.013140336610376835, + 0.012447775341570377, + 0.007373633794486523, + 0.013602613471448421, + 0.020881090313196182, + 0.02236648090183735, + 0.046973127871751785, + 0.015900928527116776, + 0.012587277218699455, + 0.016647139564156532, + 0.012672380544245243, + 0.0207859855145216, + 0.0076018827967345715, + 0.012121709063649178, + 0.008975094184279442, + 0.006926565431058407, + 0.006977801211178303, + 0.005456242244690657, + 0.006059079430997372, + 0.00999478530138731, + 0.017923157662153244, + 0.016745667904615402, + 0.031351782381534576, + 0.013266650959849358, + 0.010569159872829914, + 0.017790209501981735, + 0.015202926471829414, + 0.016913847997784615, + 0.008116009645164013, + 0.010303404182195663, + 0.009464050643146038, + 0.008006620220839977, + 0.006814845371991396, + 0.0050034658052027225, + 0.01242958102375269, + 0.00835531484335661, + 0.01988445594906807, + 0.01491206232458353, + 0.029027875512838364, + 0.01280053611844778, + 0.010674451477825642, + 0.01503787375986576, + 0.009494158439338207, + 0.015011472627520561, + 0.009014335460960865, + 0.010473149828612804, + 0.0094654131680727, + 0.0068346597254276276, + 0.006325507536530495, + 0.005778108257800341, + 0.007336521055549383, + 0.008057517930865288, + 0.008757529780268669, + 0.01669960469007492, + 0.011665946803987026, + 0.0081711420789361, + 0.0074006724171340466, + 0.0053503443486988544, + 0.008931633085012436, + 0.009704438969492912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 7 + ], + "selected_reward": -0.009136062115430832 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a butterfly perched on a pink peony?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a hazy green garden with soft morning sunlight filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the butterfly's wings initially closed, resembling a folded jewel with microscopic iridescent scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wings unfold slowly and deliberately from a closed state to a fully extended state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward at an almost imperceptible pace throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011085470207035542 + ], + "opposite_logprob_tokens": [ + -4.511085510253906 + ], + "logprob_reward": -0.011085470207035542, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011085470207035542, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17908518016338348, + 0.1009456068277359, + 0.0793858990073204, + 0.07359109073877335, + 0.05339515954256058, + 0.19672055542469025, + 0.05657922476530075, + 0.1005706638097763, + 0.09408839046955109, + 0.0656382218003273 + ], + "credit_map": [ + 0.024547209963202477, + 0.024979189038276672, + 0.037365302443504333, + 0.01723567768931389, + 0.017685115337371826, + 0.02710939384996891, + 0.01493470836430788, + 0.015228573232889175, + 0.012857109308242798, + 0.009960787370800972, + 0.020817061886191368, + 0.01278518047183752, + 0.011097041890025139, + 0.010829114355146885, + 0.009811312891542912, + 0.012787993997335434, + 0.008161715231835842, + 0.019946886226534843, + 0.009742291644215584, + 0.010514537803828716, + 0.011018426157534122, + 0.006083900574594736, + 0.0062367417849600315, + 0.007681393995881081, + 0.008649979718029499, + 0.009778025560081005, + 0.01802908629179001, + 0.006090214010328054, + 0.0072383638471364975, + 0.00767366960644722, + 0.005175727419555187, + 0.010956031270325184, + 0.006010727491229773, + 0.011284793727099895, + 0.006650193594396114, + 0.005939922295510769, + 0.006834565661847591, + 0.004139947704970837, + 0.006200375501066446, + 0.006334634497761726, + 0.029475504532456398, + 0.020025556907057762, + 0.05225043743848801, + 0.015545743517577648, + 0.011326136067509651, + 0.02681795507669449, + 0.021224744617938995, + 0.02005448378622532, + 0.007526678033173084, + 0.008787602186203003, + 0.011553655378520489, + 0.006451055407524109, + 0.004411661997437477, + 0.004903868306428194, + 0.006278999149799347, + 0.006665701512247324, + 0.014049341902136803, + 0.012988871894776821, + 0.015972385182976723, + 0.013223259709775448, + 0.009683417156338692, + 0.007492943666875362, + 0.012663341127336025, + 0.014497106894850731, + 0.012705126777291298, + 0.0103512704372406, + 0.013674162328243256, + 0.010360068641602993, + 0.010110767558217049, + 0.009755178354680538, + 0.012804235331714153, + 0.014327584765851498, + 0.010365819558501244, + 0.011668034829199314, + 0.007831186056137085, + 0.00702183973044157, + 0.005464758723974228, + 0.00603396724909544, + 0.00692311767488718, + 0.010329493321478367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.011085470207035542 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a butterfly perched on a pink peony?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a hazy green garden with soft morning sunlight filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the butterfly's wings initially closed, resembling a folded jewel with microscopic iridescent scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wings unfold slowly and deliberately from a closed state to a fully extended state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward at an almost imperceptible pace throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496998131275177 + ], + "opposite_logprob_tokens": [ + -2.9299700260162354 + ], + "logprob_reward": -0.05496998131275177, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496998131275177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09296105802059174, + 0.04240834712982178, + 0.21975472569465637, + 0.26073122024536133, + 0.07259263098239899, + 0.11901921033859253, + 0.06260548532009125, + 0.04370545223355293, + 0.02981967106461525, + 0.05640223249793053 + ], + "credit_map": [ + 0.01209868025034666, + 0.008678543381392956, + 0.012467795051634312, + 0.013062066398561, + 0.014134582132101059, + 0.009513610042631626, + 0.0139474431052804, + 0.009058334864675999, + 0.006156126037240028, + 0.005480729043483734, + 0.006691308692097664, + 0.004625960253179073, + 0.0046024383045732975, + 0.002897509839385748, + 0.003793204203248024, + 0.008161070756614208, + 0.022586004808545113, + 0.05177111178636551, + 0.02436143532395363, + 0.03189133480191231, + 0.024412719532847404, + 0.018778713420033455, + 0.016593826934695244, + 0.029359573498368263, + 0.036579761654138565, + 0.029271353036165237, + 0.08235904574394226, + 0.0231480672955513, + 0.01768801361322403, + 0.022635124623775482, + 0.01744409091770649, + 0.031605761498212814, + 0.009173440746963024, + 0.016361817717552185, + 0.008427711203694344, + 0.006992029957473278, + 0.007382766343653202, + 0.007175867911428213, + 0.0071935043670237064, + 0.009885494597256184, + 0.012203686870634556, + 0.019895140081644058, + 0.021074295043945312, + 0.01436739694327116, + 0.011996155604720116, + 0.01122650783509016, + 0.013389782048761845, + 0.014866242185235023, + 0.008943242952227592, + 0.009176318533718586, + 0.012518719770014286, + 0.00816196110099554, + 0.005735652055591345, + 0.0073544648475945, + 0.004970551934093237, + 0.005744575522840023, + 0.006563908886164427, + 0.005602966994047165, + 0.010005911812186241, + 0.005009390413761139, + 0.0037114189472049475, + 0.004310045391321182, + 0.0031911926344037056, + 0.0053106131963431835, + 0.0036462487187236547, + 0.004778963513672352, + 0.005328747443854809, + 0.0037401923909783363, + 0.0037182311061769724, + 0.003474822035059333, + 0.0026127316523343325, + 0.002519734436646104, + 0.008309866301715374, + 0.010105694644153118, + 0.007673837244510651, + 0.006022064946591854, + 0.0054708924144506454, + 0.004513157997280359, + 0.006478057708591223, + 0.007828659377992153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 5 + ], + "selected_reward": -0.05496998131275177 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a butterfly perched on a pink peony?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a hazy green garden with soft morning sunlight filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the butterfly's wings initially closed, resembling a folded jewel with microscopic iridescent scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wings unfold slowly and deliberately from a closed state to a fully extended state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward at an almost imperceptible pace throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057611756026745 + ], + "opposite_logprob_tokens": [ + -3.895576000213623 + ], + "logprob_reward": -0.02057611756026745, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057611756026745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12407520413398743, + 0.1682891845703125, + 0.35654914379119873, + 0.0467572882771492, + 0.049124136567115784, + 0.09697473794221878, + 0.06702107936143875, + 0.04028754308819771, + 0.01274244487285614, + 0.0381791889667511 + ], + "credit_map": [ + 0.019332608208060265, + 0.011654137633740902, + 0.01375760231167078, + 0.017066044732928276, + 0.018608512356877327, + 0.013834194280207157, + 0.018197545781731606, + 0.011624566279351711, + 0.023524314165115356, + 0.013888660818338394, + 0.031874798238277435, + 0.02032081037759781, + 0.01700802519917488, + 0.015694547444581985, + 0.013684424571692944, + 0.03229359909892082, + 0.03669673204421997, + 0.08479828387498856, + 0.043665383011102676, + 0.05185957998037338, + 0.04042676091194153, + 0.026356419548392296, + 0.02748759277164936, + 0.04525841772556305, + 0.005403073038905859, + 0.006402330938726664, + 0.012089242227375507, + 0.00464461138471961, + 0.004638082813471556, + 0.004180500283837318, + 0.004277953878045082, + 0.005121494643390179, + 0.006196520756930113, + 0.008796285837888718, + 0.0077807060442864895, + 0.004434399306774139, + 0.004887325223535299, + 0.004488753154873848, + 0.006581014022231102, + 0.005959130357950926, + 0.01500737201422453, + 0.012057921849191189, + 0.026328133419156075, + 0.008438479155302048, + 0.0056938352063298225, + 0.009463048540055752, + 0.011939801275730133, + 0.00804614182561636, + 0.008042178116738796, + 0.01351280976086855, + 0.008590564131736755, + 0.008417997509241104, + 0.007767410483211279, + 0.005724373273551464, + 0.007440253626555204, + 0.007525491062551737, + 0.005477046128362417, + 0.005763739347457886, + 0.007951809093356133, + 0.005044286604970694, + 0.003928052261471748, + 0.003697125008329749, + 0.0036999008152633905, + 0.004725583363324404, + 0.001322038471698761, + 0.001685533206909895, + 0.0024364327546209097, + 0.0011478769592940807, + 0.001347058336250484, + 0.0020066110882908106, + 0.001209526788443327, + 0.001587366801686585, + 0.0042988937348127365, + 0.009453889913856983, + 0.004709136206656694, + 0.0036024884320795536, + 0.004060840234160423, + 0.0029528909362852573, + 0.0048041632398962975, + 0.0042968872003257275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.02057611756026745 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a butterfly perched on a pink peony?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a hazy green garden with soft morning sunlight filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the butterfly's wings initially closed, resembling a folded jewel with microscopic iridescent scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wings unfold slowly and deliberately from a closed state to a fully extended state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward at an almost imperceptible pace throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798579841852188 + ], + "opposite_logprob_tokens": [ + -1.8048579692840576 + ], + "logprob_reward": -0.1798579841852188, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798579841852188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26057472825050354, + 0.07192838191986084, + 0.13706181943416595, + 0.10769171267747879, + 0.1082274317741394, + 0.07688263058662415, + 0.0609419047832489, + 0.04851863905787468, + 0.0750250369310379, + 0.053147733211517334 + ], + "credit_map": [ + 0.06814572215080261, + 0.02358490601181984, + 0.03584324195981026, + 0.021326960995793343, + 0.02179197408258915, + 0.05463341996073723, + 0.01841067336499691, + 0.016837820410728455, + 0.007837650366127491, + 0.009220391511917114, + 0.008363015949726105, + 0.011045394465327263, + 0.008341537788510323, + 0.010051000863313675, + 0.007557882461696863, + 0.009511502459645271, + 0.014125533401966095, + 0.024827655404806137, + 0.0184064619243145, + 0.01983136683702469, + 0.01383974589407444, + 0.01098708063364029, + 0.013654304668307304, + 0.021389663219451904, + 0.01567227579653263, + 0.012329994700849056, + 0.03400508314371109, + 0.00898713804781437, + 0.008177673444151878, + 0.00893571600317955, + 0.007901737466454506, + 0.011682086624205112, + 0.011389435268938541, + 0.022802654653787613, + 0.012500097043812275, + 0.009338503703474998, + 0.013075046241283417, + 0.008549649268388748, + 0.016223689541220665, + 0.014348349533975124, + 0.00796919222921133, + 0.008668339811265469, + 0.011412350460886955, + 0.008609949611127377, + 0.0056017800234258175, + 0.008298702538013458, + 0.012545198202133179, + 0.013777120038866997, + 0.008388671092689037, + 0.009543541818857193, + 0.00930615235120058, + 0.006395801901817322, + 0.006899507716298103, + 0.0049948925152421, + 0.009074215777218342, + 0.006339121609926224, + 0.006818798370659351, + 0.004663106985390186, + 0.008667988702654839, + 0.0052365586161613464, + 0.003817281685769558, + 0.005184294190257788, + 0.005998678505420685, + 0.008131930604577065, + 0.010410270653665066, + 0.009132959879934788, + 0.010397129692137241, + 0.009206372313201427, + 0.008840321563184261, + 0.007535053417086601, + 0.008164525032043457, + 0.011338393203914165, + 0.008525990881025791, + 0.007816211320459843, + 0.00718892365694046, + 0.005605026613920927, + 0.003930197563022375, + 0.004839469213038683, + 0.007327580358833075, + 0.007914328947663307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.1798579841852188 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme macro close-up, the camera moves forward at an almost imperceptible pace, focusing on a butterfly perched on the petals of a pink peony, drenched in morning dew. Soft morning sunlight filters through a hazy green garden in the background, precisely outlining the butterfly's silhouette. Its wings, initially closed, resemble a delicate, folded jewel, covered in microscopic iridescent scales. A crisscross of sapphire, emerald green, and flame orange creates a symmetrical and intricate pattern like a cathedral stained glass window, threaded with delicate black veins. With a barely perceptible flutter, its paper-like, fragile wings begin to unfold, inch by inch, with a steady, deliberate rhythm, a slow, ritualistic process, until they are fully extended, revealing their breathtaking full range of color and pattern. The edges of the wings glow softly in the golden light, bathing the entire image in a warm glow, creating a tranquil, magical atmosphere filled with the wonder of life.", + "seed": 22320, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_9/index_7491_seed_22320_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0569383948110044, + "vlm_logprob": -0.0569383948110044, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.0569383948110044, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a butterfly perched on a pink peony?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a hazy green garden with soft morning sunlight filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the butterfly's wings initially closed, resembling a folded jewel with microscopic iridescent scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wings unfold slowly and deliberately from a closed state to a fully extended state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward at an almost imperceptible pace throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007464257534593344 + ], + "opposite_logprob_tokens": [ + -5.882464408874512 + ], + "logprob_reward": -0.007464257534593344, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.007464257534593344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1921854168176651, + 0.1036079153418541, + 0.06779836863279343, + 0.1306307315826416, + 0.07060317695140839, + 0.058806803077459335, + 0.06662243604660034, + 0.0540490485727787, + 0.14819137752056122, + 0.10750473290681839 + ], + "credit_map": [ + 0.019598880782723427, + 0.035711269825696945, + 0.011037957854568958, + 0.009940804913640022, + 0.049770183861255646, + 0.03386994078755379, + 0.020847206935286522, + 0.0114091532304883, + 0.01313353143632412, + 0.022737080231308937, + 0.006139227654784918, + 0.0076454924419522285, + 0.01795956864953041, + 0.01582256704568863, + 0.012610194273293018, + 0.00756025081500411, + 0.006059400737285614, + 0.013177634216845036, + 0.0032238797284662724, + 0.004226311109960079, + 0.017791664227843285, + 0.011767206713557243, + 0.006488867104053497, + 0.005063400138169527, + 0.021961083635687828, + 0.026666082441806793, + 0.0049940128810703754, + 0.006308763287961483, + 0.03150010108947754, + 0.023294687271118164, + 0.00931463111191988, + 0.00659135403111577, + 0.012689932249486446, + 0.015394773334264755, + 0.0040881093591451645, + 0.0038388592656701803, + 0.01132021751254797, + 0.011766891926527023, + 0.005196771584451199, + 0.006307614967226982, + 0.00872830394655466, + 0.007382422685623169, + 0.0037890756502747536, + 0.003742191242054105, + 0.01408754475414753, + 0.00803208164870739, + 0.00674194423481822, + 0.006303239148110151, + 0.007516009733080864, + 0.012933144345879555, + 0.00413003284484148, + 0.004514239262789488, + 0.014291209168732166, + 0.010518631897866726, + 0.007836207747459412, + 0.00488295778632164, + 0.009569096378982067, + 0.005994043312966824, + 0.004441026598215103, + 0.004043424502015114, + 0.010863334871828556, + 0.0074530052952468395, + 0.00596943125128746, + 0.005715679377317429, + 0.02071201428771019, + 0.016066789627075195, + 0.010533930733799934, + 0.010822897776961327, + 0.0352635532617569, + 0.0250391848385334, + 0.016796093434095383, + 0.012956902384757996, + 0.013918592594563961, + 0.02029529958963394, + 0.005906304344534874, + 0.008775634691119194, + 0.01982872746884823, + 0.017907608300447464, + 0.01088565494865179, + 0.009986906312406063 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.007464257534593344 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a butterfly perched on a pink peony?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a hazy green garden with soft morning sunlight filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the butterfly's wings initially closed, resembling a folded jewel with microscopic iridescent scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wings unfold slowly and deliberately from a closed state to a fully extended state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward at an almost imperceptible pace throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029821811243891716 + ], + "opposite_logprob_tokens": [ + -3.5298218727111816 + ], + "logprob_reward": -0.029821811243891716, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029821811243891716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1595115065574646, + 0.060562096536159515, + 0.0628981813788414, + 0.10854992270469666, + 0.3451288342475891, + 0.04062899947166443, + 0.06846709549427032, + 0.07312240451574326, + 0.02959245629608631, + 0.05153852701187134 + ], + "credit_map": [ + 0.03014376014471054, + 0.02662757597863674, + 0.011605726554989815, + 0.010449063964188099, + 0.024414684623479843, + 0.023863738402724266, + 0.01984248496592045, + 0.012564464472234249, + 0.008162815123796463, + 0.011693531647324562, + 0.004027260933071375, + 0.003782777115702629, + 0.011109001003205776, + 0.009449907578527927, + 0.006624674424529076, + 0.005712119396775961, + 0.006095778197050095, + 0.012119214050471783, + 0.003689478151500225, + 0.003233779687434435, + 0.017020536586642265, + 0.010703775100409985, + 0.005469535943120718, + 0.004566079471260309, + 0.014951889403164387, + 0.030565205961465836, + 0.004491258878260851, + 0.006825587246567011, + 0.020414212718605995, + 0.018416956067085266, + 0.007084809709340334, + 0.005800001323223114, + 0.051185496151447296, + 0.0602722242474556, + 0.019193459302186966, + 0.02148658223450184, + 0.07682377845048904, + 0.05724039673805237, + 0.02756219543516636, + 0.03136468306183815, + 0.0070785777643322945, + 0.009006849490106106, + 0.002049832371994853, + 0.0026810760609805584, + 0.00814293697476387, + 0.004948023706674576, + 0.0036616704892367125, + 0.0030600279569625854, + 0.008992007002234459, + 0.009957605041563511, + 0.0051074218936264515, + 0.004589266609400511, + 0.012666343711316586, + 0.010104906745254993, + 0.010485623963177204, + 0.0065639144740998745, + 0.009797075763344765, + 0.00873413123190403, + 0.008182394318282604, + 0.0064843278378248215, + 0.011087644845247269, + 0.010053951293230057, + 0.006154142785817385, + 0.012628733180463314, + 0.0027267509140074253, + 0.0028786358889192343, + 0.0016221607802435756, + 0.001903254771605134, + 0.00934638548642397, + 0.005648130550980568, + 0.003062762785702944, + 0.0024043729063123465, + 0.005327002145349979, + 0.006751305423676968, + 0.004278325475752354, + 0.004148391075432301, + 0.00883976649492979, + 0.007111712358891964, + 0.004788656253367662, + 0.01029336266219616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.029821811243891716 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a butterfly perched on a pink peony?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a hazy green garden with soft morning sunlight filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the butterfly's wings initially closed, resembling a folded jewel with microscopic iridescent scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wings unfold slowly and deliberately from a closed state to a fully extended state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward at an almost imperceptible pace throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498543754220009 + ], + "opposite_logprob_tokens": [ + -2.929985523223877 + ], + "logprob_reward": -0.05498543754220009, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498543754220009, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08686783164739609, + 0.054432015866041183, + 0.07267238199710846, + 0.06234696879982948, + 0.36638882756233215, + 0.08014660328626633, + 0.06536340713500977, + 0.06372733414173126, + 0.09044650942087173, + 0.05760808661580086 + ], + "credit_map": [ + 0.011047055013477802, + 0.01205427199602127, + 0.006278600078076124, + 0.007910973392426968, + 0.016895774751901627, + 0.014070802368223667, + 0.011521169915795326, + 0.007089190650731325, + 0.0060004848055541515, + 0.012924745678901672, + 0.004453913774341345, + 0.0035109056625515223, + 0.01056143268942833, + 0.009356653317809105, + 0.0038665453903377056, + 0.0037573352456092834, + 0.008637909777462482, + 0.012354828417301178, + 0.004339795559644699, + 0.00420707231387496, + 0.0203081127256155, + 0.012141606770455837, + 0.006344854366034269, + 0.0043382057920098305, + 0.008578180335462093, + 0.014428544789552689, + 0.002908350434154272, + 0.0037004873156547546, + 0.014678362756967545, + 0.010134948417544365, + 0.00412119273096323, + 0.0037969050463289022, + 0.04003189131617546, + 0.06345845758914948, + 0.015359620563685894, + 0.024280594661831856, + 0.10465529561042786, + 0.0683349147439003, + 0.03356793522834778, + 0.016700180247426033, + 0.010326813906431198, + 0.014058093540370464, + 0.002985144266858697, + 0.004229206591844559, + 0.024548398330807686, + 0.01382480189204216, + 0.005093698855489492, + 0.005080454982817173, + 0.008458140306174755, + 0.009764940477907658, + 0.0037434857804328203, + 0.004943706560879946, + 0.014995492063462734, + 0.010909998789429665, + 0.006000050343573093, + 0.006547603756189346, + 0.006483830511569977, + 0.005839645862579346, + 0.006545646116137505, + 0.004295619670301676, + 0.011146451346576214, + 0.009699800051748753, + 0.006638210266828537, + 0.01307813823223114, + 0.011185390874743462, + 0.008429240435361862, + 0.009297135286033154, + 0.00608617439866066, + 0.01606288179755211, + 0.012646527960896492, + 0.008076907135546207, + 0.018662260845303535, + 0.0058012534864246845, + 0.010897497646510601, + 0.003514215350151062, + 0.003074107924476266, + 0.012102384120225906, + 0.010555615648627281, + 0.003584017977118492, + 0.008078999817371368 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.05498543754220009 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a butterfly perched on a pink peony?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a hazy green garden with soft morning sunlight filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the butterfly's wings initially closed, resembling a folded jewel with microscopic iridescent scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wings unfold slowly and deliberately from a closed state to a fully extended state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward at an almost imperceptible pace throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012536883354187012 + ], + "opposite_logprob_tokens": [ + -4.387537002563477 + ], + "logprob_reward": -0.012536883354187012, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012536883354187012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13665476441383362, + 0.04842288792133331, + 0.06728704273700714, + 0.13649024069309235, + 0.13049161434173584, + 0.07455475628376007, + 0.08037839084863663, + 0.053094491362571716, + 0.11362044513225555, + 0.15900537371635437 + ], + "credit_map": [ + 0.015982039272785187, + 0.02258462831377983, + 0.009131582453846931, + 0.008109193295240402, + 0.031618136912584305, + 0.02312563918530941, + 0.015759753063321114, + 0.010343782603740692, + 0.005541475024074316, + 0.009159156121313572, + 0.002972629386931658, + 0.0024886878672987223, + 0.01006758026778698, + 0.008665231987833977, + 0.004245417192578316, + 0.005282705649733543, + 0.007756467442959547, + 0.014151121489703655, + 0.0031645114067941904, + 0.0034784809686243534, + 0.0154277253895998, + 0.01100145187228918, + 0.007300262339413166, + 0.005007017403841019, + 0.020188605412840843, + 0.032390888780355453, + 0.0052587296813726425, + 0.008416254073381424, + 0.03126052767038345, + 0.02254554256796837, + 0.00997972022742033, + 0.006449975073337555, + 0.014763377606868744, + 0.023478073999285698, + 0.005429706536233425, + 0.009181779809296131, + 0.035334717482328415, + 0.023778358474373817, + 0.01151589211076498, + 0.007009697146713734, + 0.010616245679557323, + 0.011386066675186157, + 0.0027195275761187077, + 0.0034530938137322664, + 0.02264033630490303, + 0.012806424871087074, + 0.004683088976889849, + 0.006249965168535709, + 0.008228576742112637, + 0.009749210439622402, + 0.00511931674554944, + 0.006473326589912176, + 0.017378276214003563, + 0.012347903102636337, + 0.013779859989881516, + 0.007301910314708948, + 0.0052913520485162735, + 0.00561120267957449, + 0.005230677779763937, + 0.0032414786983281374, + 0.010297018103301525, + 0.007457290776073933, + 0.005765581037849188, + 0.010199882090091705, + 0.01603461243212223, + 0.014854860492050648, + 0.008815297856926918, + 0.008339972235262394, + 0.02599479630589485, + 0.01775115355849266, + 0.009949596598744392, + 0.011880150996148586, + 0.02109197899699211, + 0.0341746024787426, + 0.00824817456305027, + 0.010035600513219833, + 0.030182387679815292, + 0.028672995045781136, + 0.011508839204907417, + 0.015090793371200562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.012536883354187012 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a butterfly perched on a pink peony?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a hazy green garden with soft morning sunlight filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the butterfly's wings initially closed, resembling a folded jewel with microscopic iridescent scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wings unfold slowly and deliberately from a closed state to a fully extended state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward at an almost imperceptible pace throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17988358438014984 + ], + "opposite_logprob_tokens": [ + -1.804883599281311 + ], + "logprob_reward": -0.17988358438014984, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17988358438014984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1304924637079239, + 0.12723763287067413, + 0.11708956211805344, + 0.10310912877321243, + 0.11339490115642548, + 0.25540611147880554, + 0.03549417853355408, + 0.027432110160589218, + 0.03811610862612724, + 0.05222773551940918 + ], + "credit_map": [ + 0.01469450630247593, + 0.0178646482527256, + 0.00879785418510437, + 0.010742856189608574, + 0.031013812869787216, + 0.022239213809370995, + 0.014433999545872211, + 0.010705572552978992, + 0.01355833187699318, + 0.03196096047759056, + 0.007479436695575714, + 0.006300907116383314, + 0.024274613708257675, + 0.021804260089993477, + 0.013385480269789696, + 0.008473653346300125, + 0.01268050353974104, + 0.02964378520846367, + 0.005994764622300863, + 0.0054161627776920795, + 0.026109322905540466, + 0.018420297652482986, + 0.009784656576812267, + 0.00904006976634264, + 0.01436756644397974, + 0.023173673078417778, + 0.004953678231686354, + 0.00677068717777729, + 0.02190607227385044, + 0.015539216808974743, + 0.009334766305983067, + 0.007063470780849457, + 0.010794262401759624, + 0.020058320835232735, + 0.006543617695569992, + 0.008116815239191055, + 0.02162892371416092, + 0.017315248027443886, + 0.013302430510520935, + 0.015635285526514053, + 0.03737511858344078, + 0.04361167177557945, + 0.011449857614934444, + 0.016295300796628, + 0.06875962764024734, + 0.03821258991956711, + 0.01866201125085354, + 0.0210399329662323, + 0.004899142310023308, + 0.0059485347010195255, + 0.00237313867546618, + 0.00273962807841599, + 0.006171816494315863, + 0.0050867581740021706, + 0.005494420882314444, + 0.002780740614980459, + 0.003919074311852455, + 0.0029574178624898195, + 0.002889938885346055, + 0.002458995208144188, + 0.004226379096508026, + 0.0037730224430561066, + 0.0025707047898322344, + 0.004636575002223253, + 0.0064530386589467525, + 0.0041885823011398315, + 0.003474265569821, + 0.00238455249927938, + 0.007145042065531015, + 0.005124633200466633, + 0.002713774563744664, + 0.006632218603044748, + 0.007874460890889168, + 0.00927477702498436, + 0.003470039227977395, + 0.0026148478500545025, + 0.008667394518852234, + 0.009287814609706402, + 0.0033528029453009367, + 0.007685602642595768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.17988358438014984 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme macro close-up, the camera moves forward at an almost imperceptible pace, focusing on a butterfly perched on the petals of a pink peony, drenched in morning dew. Soft morning sunlight filters through a hazy green garden in the background, precisely outlining the butterfly's silhouette. Its wings, initially closed, resemble a delicate, folded jewel, covered in microscopic iridescent scales. A crisscross of sapphire, emerald green, and flame orange creates a symmetrical and intricate pattern like a cathedral stained glass window, threaded with delicate black veins. With a barely perceptible flutter, its paper-like, fragile wings begin to unfold, inch by inch, with a steady, deliberate rhythm, a slow, ritualistic process, until they are fully extended, revealing their breathtaking full range of color and pattern. The edges of the wings glow softly in the golden light, bathing the entire image in a warm glow, creating a tranquil, magical atmosphere filled with the wonder of life.", + "seed": 22321, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_9/index_7491_seed_22321_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05091134421527386, + "vlm_logprob": -0.05091134421527386, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.05091134421527386, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a butterfly perched on a pink peony?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a hazy green garden with soft morning sunlight filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the butterfly's wings initially closed, resembling a folded jewel with microscopic iridescent scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wings unfold slowly and deliberately from a closed state to a fully extended state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward at an almost imperceptible pace throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010834204033017159 + ], + "opposite_logprob_tokens": [ + -6.135834217071533 + ], + "logprob_reward": -0.010834204033017159, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.010834204033017159, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14600208401679993, + 0.045235708355903625, + 0.0508413128554821, + 0.03749432414770126, + 0.07181588560342789, + 0.08118695765733719, + 0.21762242913246155, + 0.15318118035793304, + 0.08160462230443954, + 0.1150154247879982 + ], + "credit_map": [ + 0.014821656979620457, + 0.046901050955057144, + 0.02388402819633484, + 0.009393212385475636, + 0.01901034079492092, + 0.012306011281907558, + 0.009234437718987465, + 0.010451363399624825, + 0.004989588167518377, + 0.008910711854696274, + 0.0064134299755096436, + 0.0030056568793952465, + 0.004727724473923445, + 0.006284741684794426, + 0.00345345470122993, + 0.007450400851666927, + 0.005212635267525911, + 0.008439159952104092, + 0.006287275347858667, + 0.0038432604633271694, + 0.009497140534222126, + 0.007757018320262432, + 0.005265648011118174, + 0.004539179150015116, + 0.003563587786629796, + 0.0089547298848629, + 0.005195075646042824, + 0.0026897345669567585, + 0.005350992549210787, + 0.0049877162091434, + 0.003111054189503193, + 0.0036414344795048237, + 0.009869168512523174, + 0.014859596267342567, + 0.009054685942828655, + 0.005311686545610428, + 0.00979533139616251, + 0.010165665298700333, + 0.00597482081502676, + 0.00678493594750762, + 0.0113996472209692, + 0.012177354656159878, + 0.010204305872321129, + 0.008300913497805595, + 0.012671589851379395, + 0.011528311297297478, + 0.006483657751232386, + 0.008421183563768864, + 0.020394330844283104, + 0.04708566144108772, + 0.02408255636692047, + 0.01243086438626051, + 0.04919761046767235, + 0.03283339738845825, + 0.01994655281305313, + 0.011651454493403435, + 0.015966707840561867, + 0.02383328787982464, + 0.022725360468029976, + 0.009367433376610279, + 0.028763772919774055, + 0.026040688157081604, + 0.014731061644852161, + 0.01175288949161768, + 0.008502029813826084, + 0.011185266077518463, + 0.010836673900485039, + 0.004762523807585239, + 0.009821870364248753, + 0.015273078344762325, + 0.011103113181889057, + 0.010120077058672905, + 0.013958626426756382, + 0.011461314745247364, + 0.01584736630320549, + 0.02149103581905365, + 0.010553253814578056, + 0.016606135293841362, + 0.015046348795294762, + 0.010051353834569454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.010834204033017159 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a butterfly perched on a pink peony?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a hazy green garden with soft morning sunlight filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the butterfly's wings initially closed, resembling a folded jewel with microscopic iridescent scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wings unfold slowly and deliberately from a closed state to a fully extended state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward at an almost imperceptible pace throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330181747674942 + ], + "opposite_logprob_tokens": [ + -3.773301839828491 + ], + "logprob_reward": -0.02330181747674942, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330181747674942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10859797894954681, + 0.026034139096736908, + 0.0160833653062582, + 0.017563849687576294, + 0.06700816750526428, + 0.019286053255200386, + 0.18282100558280945, + 0.306255042552948, + 0.21759331226348877, + 0.03875710815191269 + ], + "credit_map": [ + 0.018985547125339508, + 0.019555171951651573, + 0.01505263615399599, + 0.01000157929956913, + 0.011513816192746162, + 0.014719336293637753, + 0.010855235159397125, + 0.007914647459983826, + 0.003561572404578328, + 0.0053794607520103455, + 0.0032884820830076933, + 0.001914691412821412, + 0.0030787514988332987, + 0.003396152751520276, + 0.002016168087720871, + 0.0033988601062446833, + 0.0025100873317569494, + 0.0024467140901833773, + 0.0023652040399610996, + 0.001228311681188643, + 0.0021192114800214767, + 0.002296361606568098, + 0.0015633252914994955, + 0.0015541505999863148, + 0.002917371690273285, + 0.0034143985249102116, + 0.002318852348253131, + 0.0013675539521500468, + 0.0018229394918307662, + 0.0024580180179327726, + 0.001591070438735187, + 0.0016736449906602502, + 0.010687422007322311, + 0.011911594308912754, + 0.01075468584895134, + 0.004594225902110338, + 0.009737001731991768, + 0.010286976583302021, + 0.0035039635840803385, + 0.005532291252166033, + 0.002134117530658841, + 0.0028277987148612738, + 0.0028543169610202312, + 0.001744721201248467, + 0.0032886238768696785, + 0.002903195796534419, + 0.001616042573004961, + 0.0019172360189259052, + 0.020150592550635338, + 0.029796184971928596, + 0.027970241382718086, + 0.01847781613469124, + 0.026289822533726692, + 0.02785199135541916, + 0.016010764986276627, + 0.01627359353005886, + 0.06603725254535675, + 0.03462393209338188, + 0.05560877174139023, + 0.015054977498948574, + 0.05175666883587837, + 0.042444899678230286, + 0.019571926444768906, + 0.02115660533308983, + 0.024545006453990936, + 0.02260216698050499, + 0.028565334156155586, + 0.017679573968052864, + 0.02634594775736332, + 0.04453935846686363, + 0.017764514312148094, + 0.035551395267248154, + 0.005469160154461861, + 0.0068467347882688046, + 0.005720675457268953, + 0.006280265748500824, + 0.0037899676244705915, + 0.005322279408574104, + 0.003226396394893527, + 0.002101626479998231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.02330181747674942 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a butterfly perched on a pink peony?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a hazy green garden with soft morning sunlight filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the butterfly's wings initially closed, resembling a folded jewel with microscopic iridescent scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wings unfold slowly and deliberately from a closed state to a fully extended state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward at an almost imperceptible pace throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026360824704170227 + ], + "opposite_logprob_tokens": [ + -3.6513607501983643 + ], + "logprob_reward": -0.026360824704170227, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026360824704170227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10748613625764847, + 0.03337768465280533, + 0.04084709659218788, + 0.08180252462625504, + 0.08623377233743668, + 0.06481165438890457, + 0.09667707234621048, + 0.14109483361244202, + 0.253597229719162, + 0.09407197684049606 + ], + "credit_map": [ + 0.01538202166557312, + 0.028749367222189903, + 0.017005441710352898, + 0.006843241397291422, + 0.015002608299255371, + 0.011253507807850838, + 0.0062438263557851315, + 0.007006123661994934, + 0.0037445419002324343, + 0.007827728055417538, + 0.004234857391566038, + 0.0026769053656607866, + 0.003472760086879134, + 0.005192665383219719, + 0.0027637318708002567, + 0.0034644966945052147, + 0.0041753314435482025, + 0.009010234847664833, + 0.004537389613687992, + 0.003433576785027981, + 0.007762135472148657, + 0.00527024082839489, + 0.003586373059079051, + 0.003071810817345977, + 0.009892820380628109, + 0.019719937816262245, + 0.012592167593538761, + 0.005625239107757807, + 0.00897869374603033, + 0.013894696719944477, + 0.004655695986002684, + 0.006443271413445473, + 0.008935497142374516, + 0.01934451051056385, + 0.007955411449074745, + 0.0074461679905653, + 0.013381659984588623, + 0.011326789855957031, + 0.009214870631694794, + 0.008628868497908115, + 0.007757310289889574, + 0.012424389831721783, + 0.007834568619728088, + 0.004828011617064476, + 0.01222993154078722, + 0.008749556727707386, + 0.006389656104147434, + 0.004598231054842472, + 0.010118440724909306, + 0.01622619666159153, + 0.015443206764757633, + 0.00777796097099781, + 0.01696101203560829, + 0.014818066731095314, + 0.008486484177410603, + 0.006845696363598108, + 0.01637078821659088, + 0.015091290697455406, + 0.023263294249773026, + 0.008021892979741096, + 0.031062059104442596, + 0.027921514585614204, + 0.01281631551682949, + 0.006547683849930763, + 0.02705240622162819, + 0.020417332649230957, + 0.034166570752859116, + 0.020294969901442528, + 0.029450111091136932, + 0.04846956208348274, + 0.02238459698855877, + 0.05136168748140335, + 0.010445881634950638, + 0.01596710830926895, + 0.01357716228812933, + 0.010344461537897587, + 0.010860277339816093, + 0.016098618507385254, + 0.009882686659693718, + 0.006895772647112608 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.026360824704170227 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a butterfly perched on a pink peony?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a hazy green garden with soft morning sunlight filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the butterfly's wings initially closed, resembling a folded jewel with microscopic iridescent scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wings unfold slowly and deliberately from a closed state to a fully extended state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward at an almost imperceptible pace throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014190834015607834 + ], + "opposite_logprob_tokens": [ + -4.264190673828125 + ], + "logprob_reward": -0.014190834015607834, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014190834015607834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1499282419681549, + 0.027958689257502556, + 0.028179528191685677, + 0.026879223063588142, + 0.0574987456202507, + 0.09769825637340546, + 0.09060708433389664, + 0.09646040201187134, + 0.3424449861049652, + 0.08234477788209915 + ], + "credit_map": [ + 0.018164653331041336, + 0.036033544689416885, + 0.022403832525014877, + 0.011450385674834251, + 0.024904008954763412, + 0.016780253499746323, + 0.009745809249579906, + 0.010445772670209408, + 0.0032887249253690243, + 0.005781128071248531, + 0.0038381440099328756, + 0.0022791405208408833, + 0.00353953056037426, + 0.004292962606996298, + 0.001992642180994153, + 0.0029464156832545996, + 0.0027474549133330584, + 0.005594635382294655, + 0.0043806955218315125, + 0.0024018995463848114, + 0.003892031265422702, + 0.0042028180323541164, + 0.0025056933518499136, + 0.0024542990140616894, + 0.0030562528409063816, + 0.0070357914082705975, + 0.003291626460850239, + 0.0019833489786833525, + 0.0031385209877043962, + 0.004282140638679266, + 0.001957123400643468, + 0.0021344181150197983, + 0.006678141187876463, + 0.008851160295307636, + 0.005252279341220856, + 0.0046698241494596004, + 0.009414358995854855, + 0.006893659010529518, + 0.007462814915925264, + 0.008276508189737797, + 0.00910941418260336, + 0.019604744389653206, + 0.010420616716146469, + 0.007898837327957153, + 0.020249737426638603, + 0.014156132936477661, + 0.009745451621711254, + 0.006513325963169336, + 0.00950116291642189, + 0.014708178117871284, + 0.01139491330832243, + 0.005812343209981918, + 0.02190573327243328, + 0.01361848320811987, + 0.008218579925596714, + 0.005447691306471825, + 0.016249138861894608, + 0.013272748328745365, + 0.013436522334814072, + 0.005996389780193567, + 0.01618018187582493, + 0.01154986023902893, + 0.009608126245439053, + 0.0101674385368824, + 0.03661563992500305, + 0.028507370501756668, + 0.04545402526855469, + 0.028468264266848564, + 0.040477994829416275, + 0.06468310207128525, + 0.02980654686689377, + 0.0684320479631424, + 0.008516773581504822, + 0.01263530645519495, + 0.011743339709937572, + 0.012104014866054058, + 0.009031794965267181, + 0.014326870441436768, + 0.008431386202573776, + 0.005555290263146162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.014190834015607834 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a butterfly perched on a pink peony?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a hazy green garden with soft morning sunlight filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the butterfly's wings initially closed, resembling a folded jewel with microscopic iridescent scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wings unfold slowly and deliberately from a closed state to a fully extended state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward at an almost imperceptible pace throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17986904084682465 + ], + "opposite_logprob_tokens": [ + -1.8048690557479858 + ], + "logprob_reward": -0.17986904084682465, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17986904084682465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04214822128415108, + 0.017781870439648628, + 0.022684533149003983, + 0.035825490951538086, + 0.08504002541303635, + 0.0700971931219101, + 0.0997835248708725, + 0.10216990858316422, + 0.4844314157962799, + 0.04003778472542763 + ], + "credit_map": [ + 0.006574829109013081, + 0.00682948250323534, + 0.007471289951354265, + 0.0028001947794109583, + 0.004932178184390068, + 0.007269585970789194, + 0.0025436896830797195, + 0.0037269718013703823, + 0.0020774458535015583, + 0.003867292543873191, + 0.002675737952813506, + 0.0014642280293628573, + 0.0022094494197517633, + 0.002641939092427492, + 0.0013021244667470455, + 0.0015436530811712146, + 0.0025692544877529144, + 0.004638844635337591, + 0.0030486758332699537, + 0.001906762132421136, + 0.0036413653288036585, + 0.0032384260557591915, + 0.0016330901999026537, + 0.0020081147085875273, + 0.0048707821406424046, + 0.007979907095432281, + 0.00545002706348896, + 0.003057314781472087, + 0.0032177409157156944, + 0.0056501044891774654, + 0.0028977792244404554, + 0.002701834309846163, + 0.01611081324517727, + 0.011608573608100414, + 0.011584541760385036, + 0.006026184652000666, + 0.0116801206022501, + 0.012015331536531448, + 0.0064696804620325565, + 0.009544774889945984, + 0.007179683074355125, + 0.011769947595894337, + 0.007251301314681768, + 0.007456893101334572, + 0.013083305209875107, + 0.010221840813755989, + 0.007847012020647526, + 0.005287208594381809, + 0.011620701290667057, + 0.01616753451526165, + 0.012570529244840145, + 0.006590598728507757, + 0.022171832621097565, + 0.015353986993432045, + 0.008755701594054699, + 0.006552632432430983, + 0.012434316799044609, + 0.012071719393134117, + 0.013092994689941406, + 0.005473725963383913, + 0.028620243072509766, + 0.012589381076395512, + 0.012320558540523052, + 0.0055669695138931274, + 0.05596436560153961, + 0.04196321964263916, + 0.06379904597997665, + 0.04126463085412979, + 0.05530497059226036, + 0.08809802681207657, + 0.04173309728503227, + 0.09630408883094788, + 0.005573641043156385, + 0.006687877234071493, + 0.006030473858118057, + 0.0060628317296504974, + 0.004866102710366249, + 0.005134588107466698, + 0.0027543227188289165, + 0.0029279470909386873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.17986904084682465 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_0/index_tensor([9953, 9953])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_0/index_tensor([9953, 9953])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9f0267c504726467313a5e94fd1d01edcbfd22d3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_0/index_tensor([9953, 9953])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, medium-length shot slowly zooms in, focusing on the gnarled trunk of a southern orchid tree, which dominates a desolate white space. The entire scene is rendered in JoJo's signature style: a world of extreme high-contrast black and white, with deep, absolute black carved into brilliant, pure white. The trunk itself is a menacing and muscular figure, its twisted and convex forms striking a dramatic pose, defined by rough, graphic shading that emphasizes every contour. Its ancient bark audibly cracks and splinters under immense internal pressure, cracks spreading as rapidly as black lightning. From these newly formed fractures, smooth, perfectly spherical black spheres erupt directly from the wood, growing rapidly and powerfully. They expand explosively and decisively, pushing the broken bark into its full size in an uninterrupted, unnatural sequence. Their polished, inky black surfaces catch the harsh light with a sharp, mirror-like radiance, creating an atmosphere of tension and otherworldliness.", + "seed": 18575, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_0/index_9953_seed_18575_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.11826987422537058, + "vlm_logprob": -0.11826987422537058, + "vlm_yes_no_margin": 3.750000011920929, + "vlm_token_credit": -0.11826987422537058, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium-length shot focusing on a gnarled southern orchid tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in extreme high-contrast black and white with deep black and pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tree trunk exhibit twisted, convex forms with rough, graphic shading emphasizing contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do smooth, perfectly spherical black spheres erupt from cracks in the bark and expand rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the tree trunk throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4318006634712219 + ], + "opposite_logprob_tokens": [ + -1.0568006038665771 + ], + "logprob_reward": -0.4318006634712219, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.4318006634712219, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21157394349575043, + 0.0828891322016716, + 0.1177932620048523, + 0.06799096614122391, + 0.017487645149230957, + 0.018914788961410522, + 0.0859604924917221, + 0.15996086597442627, + 0.11328242719173431, + 0.12414649873971939 + ], + "credit_map": [ + 0.024787383154034615, + 0.02294543944299221, + 0.0375162810087204, + 0.02152751199901104, + 0.03276806324720383, + 0.02070249617099762, + 0.015223509632050991, + 0.036103256046772, + 0.010635164566338062, + 0.008641662076115608, + 0.013871323317289352, + 0.006531630642712116, + 0.013439998961985111, + 0.00983469933271408, + 0.006913703866302967, + 0.013020956888794899, + 0.014086342416703701, + 0.013034422881901264, + 0.025666525587439537, + 0.010762914083898067, + 0.015091774053871632, + 0.009280879981815815, + 0.010763444937765598, + 0.01910696178674698, + 0.007636892609298229, + 0.0085530960932374, + 0.006055645179003477, + 0.010629735887050629, + 0.007060185074806213, + 0.01021923404186964, + 0.007614423055201769, + 0.010221749544143677, + 0.0016089295968413353, + 0.0016782052116468549, + 0.0028991953004151583, + 0.002338480670005083, + 0.002867642790079117, + 0.0015563307097181678, + 0.0017989148618653417, + 0.002739946125075221, + 0.0015568226808682084, + 0.0017497459193691611, + 0.0027032410725951195, + 0.002686842577531934, + 0.0029274446424096823, + 0.0020624124445021152, + 0.0025304839946329594, + 0.002697795396670699, + 0.013191578909754753, + 0.008405926637351513, + 0.009906936436891556, + 0.010034077800810337, + 0.011735929176211357, + 0.01136096566915512, + 0.011062772013247013, + 0.010262303054332733, + 0.015656163915991783, + 0.022956140339374542, + 0.022920042276382446, + 0.01835748739540577, + 0.019253289327025414, + 0.022676287218928337, + 0.02102987840771675, + 0.01711159013211727, + 0.017031501978635788, + 0.01524173654615879, + 0.010468545369803905, + 0.016262605786323547, + 0.009997222572565079, + 0.014116051606833935, + 0.0189664326608181, + 0.011198331601917744, + 0.012452693656086922, + 0.02541898936033249, + 0.01668688841164112, + 0.020231910049915314, + 0.012414301745593548, + 0.01787358522415161, + 0.009112061001360416, + 0.00995606929063797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.4318006634712219 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium-length shot focusing on a gnarled southern orchid tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in extreme high-contrast black and white with deep black and pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tree trunk exhibit twisted, convex forms with rough, graphic shading emphasizing contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do smooth, perfectly spherical black spheres erupt from cracks in the bark and expand rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the tree trunk throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005263874307274818 + ], + "opposite_logprob_tokens": [ + -5.255263805389404 + ], + "logprob_reward": -0.005263874307274818, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005263874307274818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15019303560256958, + 0.12631826102733612, + 0.11114108562469482, + 0.06046513840556145, + 0.034554269164800644, + 0.027060477063059807, + 0.14529159665107727, + 0.1444641649723053, + 0.0926387682557106, + 0.10787323117256165 + ], + "credit_map": [ + 0.020812388509511948, + 0.01523076742887497, + 0.02992931380867958, + 0.013719997368752956, + 0.0161319337785244, + 0.014177899807691574, + 0.009197808802127838, + 0.030992915853857994, + 0.01977570354938507, + 0.013005625456571579, + 0.022921521216630936, + 0.010898648761212826, + 0.015559669584035873, + 0.013717223890125751, + 0.009762924164533615, + 0.02067693881690502, + 0.013633149676024914, + 0.011266009882092476, + 0.02670850232243538, + 0.010092087090015411, + 0.012163849547505379, + 0.007930533960461617, + 0.010919935069978237, + 0.018427027389407158, + 0.00648557860404253, + 0.006574537139385939, + 0.007278104778379202, + 0.006971970200538635, + 0.006338790990412235, + 0.009841464459896088, + 0.00754574267193675, + 0.009428952820599079, + 0.0041496059857308865, + 0.00355488620698452, + 0.0070776427164673805, + 0.0039881207048892975, + 0.0037422673776745796, + 0.0034703873097896576, + 0.003276237053796649, + 0.0052951183170080185, + 0.003427363932132721, + 0.0020823522936552763, + 0.004295262508094311, + 0.004180799704045057, + 0.0034827995114028454, + 0.0029733094852417707, + 0.00285698683001101, + 0.003761602332815528, + 0.02292614057660103, + 0.012150916270911694, + 0.018215298652648926, + 0.01280501764267683, + 0.017239542677998543, + 0.022367795929312706, + 0.022044111043214798, + 0.0175427608191967, + 0.013212026096880436, + 0.021012509241700172, + 0.023805933073163033, + 0.015043655410408974, + 0.01441408134996891, + 0.02209308184683323, + 0.018731670454144478, + 0.016151200979948044, + 0.011613004840910435, + 0.012279987335205078, + 0.008082033134996891, + 0.014451500028371811, + 0.010207637213170528, + 0.008482403121888638, + 0.018667353317141533, + 0.00885484367609024, + 0.01249060407280922, + 0.02670411765575409, + 0.011517922393977642, + 0.017116490751504898, + 0.007932320237159729, + 0.010981068015098572, + 0.006460284814238548, + 0.014670418575406075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.005263874307274818 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium-length shot focusing on a gnarled southern orchid tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in extreme high-contrast black and white with deep black and pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tree trunk exhibit twisted, convex forms with rough, graphic shading emphasizing contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do smooth, perfectly spherical black spheres erupt from cracks in the bark and expand rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the tree trunk throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00171538011636585 + ], + "opposite_logprob_tokens": [ + -6.376715183258057 + ], + "logprob_reward": -0.00171538011636585, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.00171538011636585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23228266835212708, + 0.12763117253780365, + 0.11584984511137009, + 0.08054542541503906, + 0.02637600526213646, + 0.02986280247569084, + 0.040434159338474274, + 0.14064446091651917, + 0.09465064108371735, + 0.11172283440828323 + ], + "credit_map": [ + 0.026403993368148804, + 0.02569456771016121, + 0.03996151685714722, + 0.024963613599538803, + 0.035738974809646606, + 0.022945158183574677, + 0.01698416657745838, + 0.03959067165851593, + 0.01612180843949318, + 0.013575437478721142, + 0.020601384341716766, + 0.011015473864972591, + 0.02088840678334236, + 0.015042486600577831, + 0.010547973215579987, + 0.019838210195302963, + 0.013569850474596024, + 0.014343018643558025, + 0.02185307815670967, + 0.010112444870173931, + 0.01707104966044426, + 0.009963824413716793, + 0.010691884905099869, + 0.01824468933045864, + 0.008656476624310017, + 0.009751318022608757, + 0.007206839509308338, + 0.01184283196926117, + 0.00820714607834816, + 0.012273043394088745, + 0.009901413694024086, + 0.012706355191767216, + 0.0034369537606835365, + 0.0025220918469130993, + 0.004990794230252504, + 0.0032895503100007772, + 0.0031336054671555758, + 0.0024401284754276276, + 0.0025508578401058912, + 0.004012024961411953, + 0.0026418433990329504, + 0.0035996213555336, + 0.00321693136356771, + 0.004125763662159443, + 0.004588805139064789, + 0.0036876611411571503, + 0.003661268623545766, + 0.00434090755879879, + 0.005416483618319035, + 0.0040850164368748665, + 0.004989387467503548, + 0.004351678304374218, + 0.0055241272784769535, + 0.005696721840649843, + 0.004758785013109446, + 0.005611961241811514, + 0.012663749977946281, + 0.019356537610292435, + 0.017194505780935287, + 0.01591408997774124, + 0.018447939306497574, + 0.022371485829353333, + 0.019405998289585114, + 0.015290153212845325, + 0.013904990628361702, + 0.011647098697721958, + 0.008306729607284069, + 0.012706546112895012, + 0.008868752047419548, + 0.012569129467010498, + 0.017054736614227295, + 0.009592654183506966, + 0.011482678353786469, + 0.022305866703391075, + 0.013835146091878414, + 0.017834186553955078, + 0.010525474324822426, + 0.018209239467978477, + 0.00765148364007473, + 0.00987876858562231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.00171538011636585 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium-length shot focusing on a gnarled southern orchid tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in extreme high-contrast black and white with deep black and pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tree trunk exhibit twisted, convex forms with rough, graphic shading emphasizing contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do smooth, perfectly spherical black spheres erupt from cracks in the bark and expand rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the tree trunk throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009782115928828716 + ], + "opposite_logprob_tokens": [ + -4.634782314300537 + ], + "logprob_reward": -0.009782115928828716, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009782115928828716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08866586536169052, + 0.08510735630989075, + 0.04067109897732735, + 0.05914655700325966, + 0.02954448014497757, + 0.028676843270659447, + 0.04163595661520958, + 0.1990279108285904, + 0.19526644051074982, + 0.23225755989551544 + ], + "credit_map": [ + 0.01167670264840126, + 0.008978566154837608, + 0.01796179637312889, + 0.008403673768043518, + 0.010035544633865356, + 0.008700991049408913, + 0.0059507847763597965, + 0.016957810148596764, + 0.013148178346455097, + 0.009616411291062832, + 0.013479485176503658, + 0.006610375363379717, + 0.012630181387066841, + 0.01028240192681551, + 0.006933107506483793, + 0.012407216243445873, + 0.0047029671259224415, + 0.0032721813768148422, + 0.010429436340928078, + 0.004498012829571962, + 0.0027702143415808678, + 0.004150733817368746, + 0.0037464615888893604, + 0.007101090624928474, + 0.007189519703388214, + 0.008734697476029396, + 0.005887115839868784, + 0.007453637197613716, + 0.006600989960134029, + 0.007553704082965851, + 0.006760768126696348, + 0.008966122753918171, + 0.0051133171655237675, + 0.003123397007584572, + 0.004843509290367365, + 0.0034232481848448515, + 0.003209238639101386, + 0.0025604197289794683, + 0.003000382799655199, + 0.004270969424396753, + 0.004143241792917252, + 0.0038956794887781143, + 0.0034478907473385334, + 0.004951084032654762, + 0.0029236399568617344, + 0.002324772533029318, + 0.0029982656706124544, + 0.0039922697469592094, + 0.0036255917511880398, + 0.005592982750386, + 0.005300348624587059, + 0.0056610689498484135, + 0.004310809541493654, + 0.005284133832901716, + 0.005997392814606428, + 0.005863628350198269, + 0.02448241040110588, + 0.02406540885567665, + 0.018346456810832024, + 0.019781911745667458, + 0.02480235882103443, + 0.03355487808585167, + 0.03344038873910904, + 0.020554089918732643, + 0.015617863275110722, + 0.02807251363992691, + 0.020391566678881645, + 0.027200067415833473, + 0.02068139798939228, + 0.02315794676542282, + 0.03945715352892876, + 0.020687932148575783, + 0.02534232661128044, + 0.05234914645552635, + 0.02806008607149124, + 0.03436656296253204, + 0.02031835913658142, + 0.015868395566940308, + 0.016440950334072113, + 0.039511729031801224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.009782115928828716 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium-length shot focusing on a gnarled southern orchid tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in extreme high-contrast black and white with deep black and pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tree trunk exhibit twisted, convex forms with rough, graphic shading emphasizing contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do smooth, perfectly spherical black spheres erupt from cracks in the bark and expand rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the tree trunk throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14278733730316162 + ], + "opposite_logprob_tokens": [ + -2.017787456512451 + ], + "logprob_reward": -0.14278733730316162, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14278733730316162, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09799370169639587, + 0.0569029226899147, + 0.04363837093114853, + 0.09612981230020523, + 0.05021350830793381, + 0.10759112238883972, + 0.18805687129497528, + 0.1239510253071785, + 0.13421592116355896, + 0.10130675137042999 + ], + "credit_map": [ + 0.010962924920022488, + 0.010156293399631977, + 0.018774939700961113, + 0.008893122896552086, + 0.010727915912866592, + 0.012842816300690174, + 0.0069212960079312325, + 0.018714386969804764, + 0.005788605194538832, + 0.006164758466184139, + 0.010767290368676186, + 0.006565944757312536, + 0.006358569022268057, + 0.005904701072722673, + 0.005447921808809042, + 0.009905131533741951, + 0.0048249694518744946, + 0.0044400510378181934, + 0.01024786289781332, + 0.005153965670615435, + 0.0034414164256304502, + 0.003947701305150986, + 0.004515179432928562, + 0.007067226339131594, + 0.008741272613406181, + 0.009301646612584591, + 0.010128890164196491, + 0.009503993205726147, + 0.017434552311897278, + 0.014692421071231365, + 0.011002972722053528, + 0.015324062667787075, + 0.005848984699696302, + 0.005636922083795071, + 0.009931654669344425, + 0.005654046777635813, + 0.00530591094866395, + 0.00439231563359499, + 0.005768350325524807, + 0.00767532316967845, + 0.013630207628011703, + 0.014191609807312489, + 0.009829433634877205, + 0.012401708401739597, + 0.017188919708132744, + 0.012163104489445686, + 0.012912077829241753, + 0.01527406182140112, + 0.025195203721523285, + 0.02345871925354004, + 0.02501622773706913, + 0.02700343169271946, + 0.02245716005563736, + 0.0207035094499588, + 0.022521212697029114, + 0.021701425313949585, + 0.01585499569773674, + 0.013821438886225224, + 0.02201908826828003, + 0.013628230430185795, + 0.012179301120340824, + 0.017051562666893005, + 0.017864881083369255, + 0.011531530879437923, + 0.016789909452199936, + 0.01842845045030117, + 0.016516223549842834, + 0.015503105707466602, + 0.019510628655552864, + 0.015825070440769196, + 0.018684905022382736, + 0.012957632541656494, + 0.009908394888043404, + 0.023216908797621727, + 0.011732651852071285, + 0.01512600015848875, + 0.011519495397806168, + 0.010665406472980976, + 0.008461840450763702, + 0.010676061734557152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.14278733730316162 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium-length shot slowly zooms in, focusing on the gnarled trunk of a southern orchid tree, which dominates a desolate white space. The entire scene is rendered in JoJo's signature style: a world of extreme high-contrast black and white, with deep, absolute black carved into brilliant, pure white. The trunk itself is a menacing and muscular figure, its twisted and convex forms striking a dramatic pose, defined by rough, graphic shading that emphasizes every contour. Its ancient bark audibly cracks and splinters under immense internal pressure, cracks spreading as rapidly as black lightning. From these newly formed fractures, smooth, perfectly spherical black spheres erupt directly from the wood, growing rapidly and powerfully. They expand explosively and decisively, pushing the broken bark into its full size in an uninterrupted, unnatural sequence. Their polished, inky black surfaces catch the harsh light with a sharp, mirror-like radiance, creating an atmosphere of tension and otherworldliness.", + "seed": 18576, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_0/index_9953_seed_18576_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.16341900443658233, + "vlm_logprob": -0.16341900443658233, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.16341900443658233, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium-length shot focusing on a gnarled southern orchid tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in extreme high-contrast black and white with deep black and pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tree trunk exhibit twisted, convex forms with rough, graphic shading emphasizing contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do smooth, perfectly spherical black spheres erupt from cracks in the bark and expand rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the tree trunk throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6995223760604858 + ], + "opposite_logprob_tokens": [ + -0.6995223760604858 + ], + "logprob_reward": -0.6995223760604858, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6995223760604858, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.15480612218379974, + 0.2025197446346283, + 0.06793754547834396, + 0.05856587365269661, + 0.1086617186665535, + 0.02701273374259472, + 0.0719800740480423, + 0.17381949722766876, + 0.09823020547628403, + 0.03646644949913025 + ], + "credit_map": [ + 0.015126467682421207, + 0.020329369232058525, + 0.021228861063718796, + 0.019614271819591522, + 0.020633956417441368, + 0.01406078040599823, + 0.018855703994631767, + 0.0249567162245512, + 0.028286198154091835, + 0.027787314727902412, + 0.018911072984337807, + 0.029652606695890427, + 0.029323169961571693, + 0.018654577434062958, + 0.025030042976140976, + 0.024874761700630188, + 0.01017784234136343, + 0.007686874829232693, + 0.00702863372862339, + 0.007514966651797295, + 0.007828738540410995, + 0.008754686452448368, + 0.007488373667001724, + 0.011457426473498344, + 0.009283838793635368, + 0.013010339811444283, + 0.005888680461794138, + 0.009379624389111996, + 0.005317335482686758, + 0.004731034394353628, + 0.004956919699907303, + 0.005998099222779274, + 0.016000475734472275, + 0.01441087108105421, + 0.023982902988791466, + 0.010079316794872284, + 0.011484354734420776, + 0.011678940616548061, + 0.009276948869228363, + 0.011747907847166061, + 0.004077686462551355, + 0.00435214675962925, + 0.004201638512313366, + 0.003206120105460286, + 0.0025766820181161165, + 0.002194412751123309, + 0.0031395473051816225, + 0.003264500293880701, + 0.010518362745642662, + 0.011211797595024109, + 0.008775732479989529, + 0.0060240780003368855, + 0.008033768273890018, + 0.009432634338736534, + 0.010670056566596031, + 0.0073136440478265285, + 0.026743842288851738, + 0.02714703232049942, + 0.02644496038556099, + 0.023141376674175262, + 0.019272999837994576, + 0.017530804499983788, + 0.01567964069545269, + 0.017858844250440598, + 0.015896709635853767, + 0.016028178855776787, + 0.009192479774355888, + 0.010697583667933941, + 0.007250187918543816, + 0.01439393125474453, + 0.01785067282617092, + 0.0069204592145979404, + 0.006064657587558031, + 0.005585185252130032, + 0.004175208508968353, + 0.004864561837166548, + 0.0035676280967891216, + 0.0036391341127455235, + 0.003766499226912856, + 0.004803576041013002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.6995223760604858 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium-length shot focusing on a gnarled southern orchid tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in extreme high-contrast black and white with deep black and pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tree trunk exhibit twisted, convex forms with rough, graphic shading emphasizing contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do smooth, perfectly spherical black spheres erupt from cracks in the bark and expand rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the tree trunk throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005275969859212637 + ], + "opposite_logprob_tokens": [ + -5.255276203155518 + ], + "logprob_reward": -0.005275969859212637, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005275969859212637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06215552240610123, + 0.07367780059576035, + 0.07873965054750443, + 0.09707251936197281, + 0.08874382078647614, + 0.11392717808485031, + 0.025169838219881058, + 0.26317498087882996, + 0.06107465922832489, + 0.13626408576965332 + ], + "credit_map": [ + 0.010452711023390293, + 0.007386323064565659, + 0.007831980474293232, + 0.006704523228108883, + 0.007548344787210226, + 0.0063895611092448235, + 0.006613400764763355, + 0.00922867376357317, + 0.00924571231007576, + 0.009233703836798668, + 0.010456384159624577, + 0.011134558357298374, + 0.010405002161860466, + 0.005869725253432989, + 0.009442443959414959, + 0.007890264503657818, + 0.01019346434623003, + 0.009220626205205917, + 0.010501785203814507, + 0.01113989669829607, + 0.007718367502093315, + 0.00961916521191597, + 0.008582871407270432, + 0.011763469316065311, + 0.01619775779545307, + 0.01642291434109211, + 0.01401596236974001, + 0.011234348639845848, + 0.008339152671396732, + 0.0076853553764522076, + 0.009873160161077976, + 0.013303863815963268, + 0.012566987425088882, + 0.010548027232289314, + 0.013991511426866055, + 0.01620401255786419, + 0.008906549774110317, + 0.010914376936852932, + 0.006302877329289913, + 0.009309468790888786, + 0.01764325238764286, + 0.01889181137084961, + 0.015282672829926014, + 0.01079592201858759, + 0.011142721399664879, + 0.007854960858821869, + 0.01351193618029356, + 0.018803896382451057, + 0.0038790106773376465, + 0.003808018285781145, + 0.0034993637818843126, + 0.0026152057107537985, + 0.00357917626388371, + 0.0018482058076187968, + 0.0030288659036159515, + 0.0029119898099452257, + 0.02901488170027733, + 0.030750064179301262, + 0.022191593423485756, + 0.04344833269715309, + 0.02145957201719284, + 0.024839403107762337, + 0.019593195989727974, + 0.07187792658805847, + 0.01068951841443777, + 0.012565537355840206, + 0.005788265261799097, + 0.008547917939722538, + 0.003848036052659154, + 0.009179643355309963, + 0.004027221351861954, + 0.0064285192638635635, + 0.023380598053336143, + 0.025941120460629463, + 0.018944982439279556, + 0.016139427199959755, + 0.008489927276968956, + 0.014458537101745605, + 0.011841390281915665, + 0.01706809736788273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 5 + ], + "selected_reward": -0.005275969859212637 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium-length shot focusing on a gnarled southern orchid tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in extreme high-contrast black and white with deep black and pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tree trunk exhibit twisted, convex forms with rough, graphic shading emphasizing contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do smooth, perfectly spherical black spheres erupt from cracks in the bark and expand rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the tree trunk throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022019920870661736 + ], + "opposite_logprob_tokens": [ + -6.127202033996582 + ], + "logprob_reward": -0.0022019920870661736, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022019920870661736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18526363372802734, + 0.20714187622070312, + 0.08033948391675949, + 0.035898998379707336, + 0.08349166065454483, + 0.03352759778499603, + 0.0698598325252533, + 0.14257431030273438, + 0.08757521212100983, + 0.07432746142148972 + ], + "credit_map": [ + 0.023891068994998932, + 0.024284686893224716, + 0.025399383157491684, + 0.02221001870930195, + 0.023772509768605232, + 0.015757068991661072, + 0.021529173478484154, + 0.028419731184840202, + 0.02941843308508396, + 0.027756787836551666, + 0.019520454108715057, + 0.03090060129761696, + 0.02973001077771187, + 0.01849564164876938, + 0.02568797394633293, + 0.025631966069340706, + 0.01184934563934803, + 0.010488051921129227, + 0.00995030626654625, + 0.008868306875228882, + 0.008754953742027283, + 0.009473703801631927, + 0.00864446721971035, + 0.012310352176427841, + 0.005248312838375568, + 0.006763418670743704, + 0.004517814610153437, + 0.005072980187833309, + 0.003131407080218196, + 0.00300853094086051, + 0.003442079294472933, + 0.004714454989880323, + 0.012629074975848198, + 0.012010722421109676, + 0.018447505310177803, + 0.00806334987282753, + 0.008541030809283257, + 0.008036598563194275, + 0.006940447725355625, + 0.008822928182780743, + 0.004996553063392639, + 0.005359905306249857, + 0.004297490697354078, + 0.004297262988984585, + 0.0037162071093916893, + 0.002482384443283081, + 0.00367580633610487, + 0.004701991565525532, + 0.009990602731704712, + 0.01128102745860815, + 0.008266917429864407, + 0.007277287542819977, + 0.0076721059158444405, + 0.007684838958084583, + 0.011435221880674362, + 0.006251832470297813, + 0.023136094212532043, + 0.023825181648135185, + 0.02276572771370411, + 0.018518703058362007, + 0.015514855273067951, + 0.013291563838720322, + 0.012496436014771461, + 0.013025754131376743, + 0.014232007786631584, + 0.015282140113413334, + 0.007680192124098539, + 0.015319904312491417, + 0.005498572718352079, + 0.009317793883383274, + 0.01244093757122755, + 0.007803662680089474, + 0.011904128827154636, + 0.013091339729726315, + 0.010401017963886261, + 0.007975688204169273, + 0.0048835319466888905, + 0.010161448270082474, + 0.003991262521594763, + 0.011919046752154827 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.0022019920870661736 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium-length shot focusing on a gnarled southern orchid tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in extreme high-contrast black and white with deep black and pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tree trunk exhibit twisted, convex forms with rough, graphic shading emphasizing contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do smooth, perfectly spherical black spheres erupt from cracks in the bark and expand rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the tree trunk throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978435855358839 + ], + "opposite_logprob_tokens": [ + -4.63478422164917 + ], + "logprob_reward": -0.00978435855358839, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978435855358839, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05653794854879379, + 0.03374182805418968, + 0.07650981098413467, + 0.05523579940199852, + 0.13969621062278748, + 0.037945233285427094, + 0.025685397908091545, + 0.14139679074287415, + 0.27449122071266174, + 0.15875981748104095 + ], + "credit_map": [ + 0.009558097459375858, + 0.00839259009808302, + 0.010411243885755539, + 0.004733064211905003, + 0.005584080703556538, + 0.004469332285225391, + 0.005566238425672054, + 0.007823298685252666, + 0.00599061232060194, + 0.00483162235468626, + 0.004997631534934044, + 0.005239567719399929, + 0.004161369055509567, + 0.0028279100079089403, + 0.0028194787446409464, + 0.0028736363165080547, + 0.010512035340070724, + 0.010614595375955105, + 0.011152612045407295, + 0.012796817347407341, + 0.0061518000438809395, + 0.008540977723896503, + 0.006969138514250517, + 0.009771828539669514, + 0.00958755798637867, + 0.010898742824792862, + 0.007456619292497635, + 0.00665206927806139, + 0.005656805820763111, + 0.004478896036744118, + 0.004826577380299568, + 0.005678525660187006, + 0.019665762782096863, + 0.019262263551354408, + 0.03733888640999794, + 0.014128344133496284, + 0.013816946186125278, + 0.012834788300096989, + 0.011614435352385044, + 0.011034770868718624, + 0.005737672094255686, + 0.005270888563245535, + 0.005404010880738497, + 0.0048532611690461636, + 0.004400190431624651, + 0.0032823116052895784, + 0.0039848871529102325, + 0.005012011621147394, + 0.004129013977944851, + 0.004676357842981815, + 0.0036641580518335104, + 0.002833261853083968, + 0.0032759676687419415, + 0.0024568219669163227, + 0.002643364015966654, + 0.0020064502023160458, + 0.016402367502450943, + 0.018491247668862343, + 0.014989804476499557, + 0.017328958958387375, + 0.013270964846014977, + 0.017476215958595276, + 0.012135237455368042, + 0.031301967799663544, + 0.043098557740449905, + 0.04114324226975441, + 0.02125738002359867, + 0.06375040858983994, + 0.01845911517739296, + 0.03941277042031288, + 0.020146217197179794, + 0.0272234994918108, + 0.025143228471279144, + 0.027198662981390953, + 0.019330937415361404, + 0.015617497265338898, + 0.011418987065553665, + 0.028330976143479347, + 0.00911613367497921, + 0.022603394463658333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.00978435855358839 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium-length shot focusing on a gnarled southern orchid tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in extreme high-contrast black and white with deep black and pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tree trunk exhibit twisted, convex forms with rough, graphic shading emphasizing contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do smooth, perfectly spherical black spheres erupt from cracks in the bark and expand rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the tree trunk throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003103256225586 + ], + "opposite_logprob_tokens": [ + -2.3503103256225586 + ], + "logprob_reward": -0.1003103256225586, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003103256225586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0711047574877739, + 0.054549094289541245, + 0.058492958545684814, + 0.07430484145879745, + 0.07842491567134857, + 0.14700184762477875, + 0.0669570043683052, + 0.03696146234869957, + 0.3203025460243225, + 0.09190060943365097 + ], + "credit_map": [ + 0.012718948535621166, + 0.007721095811575651, + 0.008946934714913368, + 0.010202147997915745, + 0.009229305200278759, + 0.005952183622866869, + 0.00781275238841772, + 0.00852139014750719, + 0.0074097043834626675, + 0.009074226953089237, + 0.0064501953311264515, + 0.0064697228372097015, + 0.0070626502856612206, + 0.004915738943964243, + 0.005786337424069643, + 0.007380516268312931, + 0.008566041477024555, + 0.009325729683041573, + 0.009778722189366817, + 0.008260197006165981, + 0.004366712644696236, + 0.005941343959420919, + 0.006183846388012171, + 0.006070366594940424, + 0.013758099637925625, + 0.010961146093904972, + 0.013364126905798912, + 0.00831756740808487, + 0.005973245482891798, + 0.00527877826243639, + 0.006864138878881931, + 0.009787731803953648, + 0.011837455444037914, + 0.011391556821763515, + 0.020205562934279442, + 0.008845007978379726, + 0.006253757048398256, + 0.005991722922772169, + 0.006210786756128073, + 0.007689058315008879, + 0.02202521078288555, + 0.021423280239105225, + 0.02109590359032154, + 0.017697962000966072, + 0.014582237228751183, + 0.01336281280964613, + 0.01759224385023117, + 0.019222185015678406, + 0.008146732114255428, + 0.009244509972631931, + 0.00725542800500989, + 0.006448558997362852, + 0.007664044387638569, + 0.007745683658868074, + 0.010859416797757149, + 0.009592629037797451, + 0.006288787815719843, + 0.007739574182778597, + 0.00516468845307827, + 0.004988483153283596, + 0.0032408342231065035, + 0.003250998444855213, + 0.0030100224539637566, + 0.003278071526437998, + 0.05333273112773895, + 0.05357695370912552, + 0.027289647608995438, + 0.07209637016057968, + 0.020361704751849174, + 0.03196367621421814, + 0.028792038559913635, + 0.03288940340280533, + 0.014104715548455715, + 0.016896938905119896, + 0.014543772675096989, + 0.009694802574813366, + 0.006311628036201, + 0.010989694856107235, + 0.006206849589943886, + 0.013152207247912884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -0.1003103256225586 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium-length shot slowly zooms in, focusing on the gnarled trunk of a southern orchid tree, which dominates a desolate white space. The entire scene is rendered in JoJo's signature style: a world of extreme high-contrast black and white, with deep, absolute black carved into brilliant, pure white. The trunk itself is a menacing and muscular figure, its twisted and convex forms striking a dramatic pose, defined by rough, graphic shading that emphasizes every contour. Its ancient bark audibly cracks and splinters under immense internal pressure, cracks spreading as rapidly as black lightning. From these newly formed fractures, smooth, perfectly spherical black spheres erupt directly from the wood, growing rapidly and powerfully. They expand explosively and decisively, pushing the broken bark into its full size in an uninterrupted, unnatural sequence. Their polished, inky black surfaces catch the harsh light with a sharp, mirror-like radiance, creating an atmosphere of tension and otherworldliness.", + "seed": 18577, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_0/index_9953_seed_18577_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.21810492295771838, + "vlm_logprob": -0.21810492295771838, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.21810492295771838, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium-length shot focusing on a gnarled southern orchid tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in extreme high-contrast black and white with deep black and pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tree trunk exhibit twisted, convex forms with rough, graphic shading emphasizing contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do smooth, perfectly spherical black spheres erupt from cracks in the bark and expand rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the tree trunk throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8350120186805725 + ], + "opposite_logprob_tokens": [ + -0.5850120186805725 + ], + "logprob_reward": -0.8350120186805725, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8350120186805725, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10483463853597641, + 0.05376926809549332, + 0.14453081786632538, + 0.0883457362651825, + 0.05505989491939545, + 0.10232771933078766, + 0.22859451174736023, + 0.04371533915400505, + 0.07469009608030319, + 0.10413195937871933 + ], + "credit_map": [ + 0.012906803749501705, + 0.012818219140172005, + 0.01569950580596924, + 0.017945583909749985, + 0.00940272118896246, + 0.01407969556748867, + 0.013001742772758007, + 0.008980371989309788, + 0.01099781971424818, + 0.0034768003970384598, + 0.005798645317554474, + 0.006640433333814144, + 0.00312766688875854, + 0.012605126947164536, + 0.00731905410066247, + 0.0038037218619138002, + 0.02610691636800766, + 0.01209181547164917, + 0.01058545708656311, + 0.013524255715310574, + 0.01088277343660593, + 0.0426383838057518, + 0.01799112930893898, + 0.010710087604820728, + 0.010564325377345085, + 0.011334977112710476, + 0.007139136083424091, + 0.019574226811528206, + 0.010790141299366951, + 0.012109946459531784, + 0.009967049583792686, + 0.00686593446880579, + 0.00852337945252657, + 0.005020204000174999, + 0.005578923039138317, + 0.01736210286617279, + 0.004257218446582556, + 0.005274043418467045, + 0.0042767878621816635, + 0.004767239093780518, + 0.00901380181312561, + 0.010491074994206429, + 0.006850388366729021, + 0.007309074513614178, + 0.005674935411661863, + 0.030878055840730667, + 0.005549682304263115, + 0.026560703292489052, + 0.02482464537024498, + 0.012151098810136318, + 0.01833629608154297, + 0.009150531142950058, + 0.012745705433189869, + 0.07746239006519318, + 0.0076425508596003056, + 0.0662812888622284, + 0.009677746333181858, + 0.003499117447063327, + 0.008720900863409042, + 0.009874328970909119, + 0.0022788962814956903, + 0.002636068034917116, + 0.004243049304932356, + 0.0027852333150804043, + 0.006411571055650711, + 0.015435513108968735, + 0.00945128221064806, + 0.010140849277377129, + 0.0069453236646950245, + 0.006933975033462048, + 0.008212410844862461, + 0.011159173212945461, + 0.015339575707912445, + 0.028176037594676018, + 0.010621609166264534, + 0.009126344695687294, + 0.009164202958345413, + 0.009959659539163113, + 0.009128873236477375, + 0.012615658342838287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.8350120186805725 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium-length shot focusing on a gnarled southern orchid tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in extreme high-contrast black and white with deep black and pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tree trunk exhibit twisted, convex forms with rough, graphic shading emphasizing contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do smooth, perfectly spherical black spheres erupt from cracks in the bark and expand rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the tree trunk throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006771592888981104 + ], + "opposite_logprob_tokens": [ + -5.006771564483643 + ], + "logprob_reward": -0.006771592888981104, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006771592888981104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03275291994214058, + 0.1823458969593048, + 0.021223261952400208, + 0.026243731379508972, + 0.1849510669708252, + 0.025207780301570892, + 0.22442038357257843, + 0.012818350456655025, + 0.06351761519908905, + 0.22651897370815277 + ], + "credit_map": [ + 0.005409836303442717, + 0.004279958549886942, + 0.00592469098046422, + 0.003896442474797368, + 0.0033327590208500624, + 0.002699293429031968, + 0.0038203697185963392, + 0.003389572026208043, + 0.03187942877411842, + 0.009982460178434849, + 0.07599759846925735, + 0.025260087102651596, + 0.007924131117761135, + 0.010797393508255482, + 0.012215636670589447, + 0.008289149031043053, + 0.0032742309849709272, + 0.0017719727475196123, + 0.0021214361768215895, + 0.0024099787697196007, + 0.0016207725275307894, + 0.0045478977262973785, + 0.0038818484172225, + 0.0015951271634548903, + 0.0030668724793940783, + 0.0026938028167933226, + 0.0030023944564163685, + 0.0029907419811934233, + 0.002921292558312416, + 0.002645653672516346, + 0.005777166690677404, + 0.0031458057928830385, + 0.02673167921602726, + 0.0061187585815787315, + 0.008531304076313972, + 0.0205867700278759, + 0.010617528110742569, + 0.07173631340265274, + 0.009416230954229832, + 0.031212469562888145, + 0.0027146614156663418, + 0.0017386905383318663, + 0.0015296111814677715, + 0.001748921349644661, + 0.0009163654176518321, + 0.008425364270806313, + 0.0012603214709088206, + 0.006873843725770712, + 0.02456207200884819, + 0.013322181068360806, + 0.01412589754909277, + 0.036280665546655655, + 0.008384871296584606, + 0.046186648309230804, + 0.01063035149127245, + 0.0709277018904686, + 0.001639217254705727, + 0.0014309595571830869, + 0.002729720203205943, + 0.002175743691623211, + 0.0008954913355410099, + 0.0014792741276323795, + 0.0011323078069835901, + 0.0013356368290260434, + 0.006258094683289528, + 0.01490892842411995, + 0.00887623243033886, + 0.0068442365154623985, + 0.006345352157950401, + 0.0069703697226941586, + 0.004964051768183708, + 0.008350350894033909, + 0.028820907697081566, + 0.030456528067588806, + 0.023851916193962097, + 0.02380104549229145, + 0.024259498342871666, + 0.02112317830324173, + 0.0360841304063797, + 0.038121771067380905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 4 + ], + "selected_reward": -0.006771592888981104 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium-length shot focusing on a gnarled southern orchid tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in extreme high-contrast black and white with deep black and pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tree trunk exhibit twisted, convex forms with rough, graphic shading emphasizing contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do smooth, perfectly spherical black spheres erupt from cracks in the bark and expand rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the tree trunk throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004101912025362253 + ], + "opposite_logprob_tokens": [ + -5.504101753234863 + ], + "logprob_reward": -0.004101912025362253, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004101912025362253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10730083286762238, + 0.055059678852558136, + 0.12594103813171387, + 0.06445140391588211, + 0.2104291468858719, + 0.07881748676300049, + 0.1320710927248001, + 0.03747803717851639, + 0.08809197694063187, + 0.10035935789346695 + ], + "credit_map": [ + 0.01789119280874729, + 0.014721809886395931, + 0.017987631261348724, + 0.015275768004357815, + 0.009366343729197979, + 0.010939035564661026, + 0.012853255495429039, + 0.008265785872936249, + 0.01127894502133131, + 0.0040219551883637905, + 0.005010406952351332, + 0.0072976225055754185, + 0.0033760324586182833, + 0.012776130810379982, + 0.007259245030581951, + 0.004039337392896414, + 0.021806925535202026, + 0.01053223479539156, + 0.009448576718568802, + 0.011846794746816158, + 0.00956258550286293, + 0.037442922592163086, + 0.01600780338048935, + 0.009293178096413612, + 0.007873192429542542, + 0.00916755199432373, + 0.00598836038261652, + 0.010206014849245548, + 0.007078567519783974, + 0.007659436669200659, + 0.011164123192429543, + 0.005314160138368607, + 0.03590448200702667, + 0.007168249227106571, + 0.011020475998520851, + 0.016006935387849808, + 0.011588435620069504, + 0.09059097617864609, + 0.009471064433455467, + 0.028678501024842262, + 0.013435575179755688, + 0.007334438152611256, + 0.005162984598428011, + 0.005303066223859787, + 0.003102858318015933, + 0.027143070474267006, + 0.004140680655837059, + 0.013194811530411243, + 0.014192083850502968, + 0.008462995290756226, + 0.00854732096195221, + 0.02898668684065342, + 0.004301420412957668, + 0.018332406878471375, + 0.007577067706733942, + 0.04167109727859497, + 0.008644470945000648, + 0.0035825143568217754, + 0.006248160731047392, + 0.009881083853542805, + 0.0024043433368206024, + 0.0017572133801877499, + 0.0025117986369878054, + 0.0024484475143253803, + 0.011153646744787693, + 0.01192477811127901, + 0.016502294689416885, + 0.009170396253466606, + 0.0075096245855093, + 0.009670767933130264, + 0.009965457953512669, + 0.012195004150271416, + 0.01421895157545805, + 0.018707020208239555, + 0.012728840112686157, + 0.010840541683137417, + 0.008364522829651833, + 0.008043073117733002, + 0.015086296014487743, + 0.012370112352073193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 2 + ], + "selected_reward": -0.004101912025362253 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium-length shot focusing on a gnarled southern orchid tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in extreme high-contrast black and white with deep black and pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tree trunk exhibit twisted, convex forms with rough, graphic shading emphasizing contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do smooth, perfectly spherical black spheres erupt from cracks in the bark and expand rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the tree trunk throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307333379983902 + ], + "opposite_logprob_tokens": [ + -3.1680734157562256 + ], + "logprob_reward": -0.04307333379983902, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307333379983902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.025350047275424004, + 0.140449658036232, + 0.024094123393297195, + 0.028791669756174088, + 0.38695085048675537, + 0.03928523138165474, + 0.11444742977619171, + 0.028266869485378265, + 0.17953623831272125, + 0.03282788395881653 + ], + "credit_map": [ + 0.0032800801564007998, + 0.0029485805425792933, + 0.0032775152940303087, + 0.005448338575661182, + 0.002095752162858844, + 0.0030279792845249176, + 0.0031133503653109074, + 0.0021584518253803253, + 0.022232631221413612, + 0.005222050938755274, + 0.0665435940027237, + 0.014634085819125175, + 0.004718730691820383, + 0.012837802991271019, + 0.009566621854901314, + 0.00469414284452796, + 0.0033929585479199886, + 0.002518194494768977, + 0.002598952269181609, + 0.002867372939363122, + 0.0019875105936080217, + 0.004775941371917725, + 0.004099354147911072, + 0.0018538384465500712, + 0.0022731763310730457, + 0.004293865989893675, + 0.002848232863470912, + 0.005143295973539352, + 0.0038078667130321264, + 0.0029053501784801483, + 0.005096428096294403, + 0.002423454076051712, + 0.04743657261133194, + 0.010201008059084415, + 0.01461801864206791, + 0.04580029100179672, + 0.02093832567334175, + 0.18417319655418396, + 0.01295210886746645, + 0.05083131790161133, + 0.009495358914136887, + 0.0049056559801101685, + 0.0041616749949753284, + 0.0027027528267353773, + 0.00213633687235415, + 0.010046680457890034, + 0.002876081271097064, + 0.0029606877360492945, + 0.01252780482172966, + 0.006226395256817341, + 0.008123794570565224, + 0.010316229425370693, + 0.0055587412789464, + 0.032285697758197784, + 0.004621373023837805, + 0.03478740155696869, + 0.0037289797328412533, + 0.005164800677448511, + 0.0032194622326642275, + 0.002400863217189908, + 0.003988209646195173, + 0.0038299523293972015, + 0.0026630847714841366, + 0.0032715180423110723, + 0.02735082618892193, + 0.03464054316282272, + 0.03673654794692993, + 0.014361880719661713, + 0.012870809994637966, + 0.012995264492928982, + 0.01960613764822483, + 0.020974231883883476, + 0.0031474768184125423, + 0.0057207513600587845, + 0.004243555013090372, + 0.002839587163180113, + 0.0030547804199159145, + 0.0043058148585259914, + 0.002590740565210581, + 0.006925180088728666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 1 + ], + "selected_reward": -0.04307333379983902 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium-length shot focusing on a gnarled southern orchid tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in extreme high-contrast black and white with deep black and pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tree trunk exhibit twisted, convex forms with rough, graphic shading emphasizing contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do smooth, perfectly spherical black spheres erupt from cracks in the bark and expand rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the tree trunk throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20156575739383698 + ], + "opposite_logprob_tokens": [ + -1.7015657424926758 + ], + "logprob_reward": -0.20156575739383698, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20156575739383698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01687663421034813, + 0.08320990949869156, + 0.025762854143977165, + 0.05377865210175514, + 0.21329089999198914, + 0.14978690445423126, + 0.2679833471775055, + 0.027232469990849495, + 0.0932391807436943, + 0.06883920729160309 + ], + "credit_map": [ + 0.0026594705414026976, + 0.0016822069883346558, + 0.002670466899871826, + 0.0017863484099507332, + 0.0016441593179479241, + 0.0019965318497270346, + 0.0024541416205465794, + 0.0019833077676594257, + 0.012956898659467697, + 0.0030727037228643894, + 0.04007171094417572, + 0.008803348056972027, + 0.002479230286553502, + 0.007898315787315369, + 0.0050355298444628716, + 0.002892183605581522, + 0.0035500777885317802, + 0.0026704976335167885, + 0.002266436582431197, + 0.003006919752806425, + 0.002352975308895111, + 0.006344241090118885, + 0.0034050224348902702, + 0.0021666823886334896, + 0.008024624548852444, + 0.006131511181592941, + 0.004538129549473524, + 0.007470116019248962, + 0.00783118698745966, + 0.006988371256738901, + 0.006196042988449335, + 0.006598668172955513, + 0.035794612020254135, + 0.008221187628805637, + 0.011923880316317081, + 0.02424946241080761, + 0.011484697461128235, + 0.09533428400754929, + 0.00859497394412756, + 0.017687806859612465, + 0.02475033886730671, + 0.012198930606245995, + 0.008410717360675335, + 0.007280666846781969, + 0.007089680526405573, + 0.05555960536003113, + 0.0047265514731407166, + 0.02977040410041809, + 0.02151891030371189, + 0.012749388813972473, + 0.021422695368528366, + 0.0443064384162426, + 0.01641078293323517, + 0.10392799973487854, + 0.007720337249338627, + 0.039926789700984955, + 0.0030110282823443413, + 0.004277313128113747, + 0.002733769593760371, + 0.005361910443753004, + 0.002709300722926855, + 0.004122880753129721, + 0.0023102157283574343, + 0.002706051804125309, + 0.012815930880606174, + 0.025118131190538406, + 0.007343821693211794, + 0.007179267704486847, + 0.00975608080625534, + 0.009078227914869785, + 0.00870414637029171, + 0.013243579305708408, + 0.007777436636388302, + 0.013417069800198078, + 0.005008778069168329, + 0.0054730139672756195, + 0.006596360355615616, + 0.010204746387898922, + 0.01029936783015728, + 0.01006243471056223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 5 + ], + "selected_reward": -0.20156575739383698 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium-length shot slowly zooms in, focusing on the gnarled trunk of a southern orchid tree, which dominates a desolate white space. The entire scene is rendered in JoJo's signature style: a world of extreme high-contrast black and white, with deep, absolute black carved into brilliant, pure white. The trunk itself is a menacing and muscular figure, its twisted and convex forms striking a dramatic pose, defined by rough, graphic shading that emphasizes every contour. Its ancient bark audibly cracks and splinters under immense internal pressure, cracks spreading as rapidly as black lightning. From these newly formed fractures, smooth, perfectly spherical black spheres erupt directly from the wood, growing rapidly and powerfully. They expand explosively and decisively, pushing the broken bark into its full size in an uninterrupted, unnatural sequence. Their polished, inky black surfaces catch the harsh light with a sharp, mirror-like radiance, creating an atmosphere of tension and otherworldliness.", + "seed": 18578, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_0/index_9953_seed_18578_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.039740964444354175, + "vlm_logprob": -0.039740964444354175, + "vlm_yes_no_margin": 4.074999976158142, + "vlm_token_credit": -0.039740964444354175, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium-length shot focusing on a gnarled southern orchid tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in extreme high-contrast black and white with deep black and pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tree trunk exhibit twisted, convex forms with rough, graphic shading emphasizing contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do smooth, perfectly spherical black spheres erupt from cracks in the bark and expand rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the tree trunk throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1344204843044281 + ], + "opposite_logprob_tokens": [ + -2.134420394897461 + ], + "logprob_reward": -0.1344204843044281, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1344204843044281, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2375091314315796, + 0.24595408141613007, + 0.10408983379602432, + 0.060747936367988586, + 0.0848594531416893, + 0.04414943605661392, + 0.04921312630176544, + 0.04921415075659752, + 0.05039461329579353, + 0.07386822253465652 + ], + "credit_map": [ + 0.03674967214465141, + 0.04368346184492111, + 0.024403635412454605, + 0.03873799368739128, + 0.024446791037917137, + 0.018131449818611145, + 0.029548903927206993, + 0.021807242184877396, + 0.05919281393289566, + 0.027825331315398216, + 0.03765304759144783, + 0.027330154553055763, + 0.02130601927638054, + 0.01945670321583748, + 0.02210775949060917, + 0.031082268804311752, + 0.022724641487002373, + 0.01496750209480524, + 0.010169442743062973, + 0.012043231166899204, + 0.008488821797072887, + 0.006828250829130411, + 0.012484724633395672, + 0.016383230686187744, + 0.010276713408529758, + 0.006231075618416071, + 0.004939535167068243, + 0.007636338006705046, + 0.0066762687638401985, + 0.006811104249209166, + 0.010892468504607677, + 0.0072844382375478745, + 0.011237544007599354, + 0.009266426786780357, + 0.006906130351126194, + 0.01352198887616396, + 0.009967987425625324, + 0.009909720160067081, + 0.014347145333886147, + 0.009702516719698906, + 0.005136593710631132, + 0.004075014032423496, + 0.004085318185389042, + 0.004164340905845165, + 0.004970399197191, + 0.00895389448851347, + 0.003879058873280883, + 0.008884821087121964, + 0.006554348859935999, + 0.004896549507975578, + 0.0050523472018539906, + 0.00834747590124607, + 0.004041817504912615, + 0.006916866172105074, + 0.007681559305638075, + 0.005722168367356062, + 0.0095203323289752, + 0.005569116212427616, + 0.005404248833656311, + 0.0057814097963273525, + 0.0037455824203789234, + 0.003857498988509178, + 0.007611953653395176, + 0.007724010851234198, + 0.007582561112940311, + 0.004347214475274086, + 0.004383586347103119, + 0.0040080975741147995, + 0.0030576707795262337, + 0.01094752736389637, + 0.0036517386324703693, + 0.01241622120141983, + 0.01123862061649561, + 0.005749763920903206, + 0.008062172681093216, + 0.008164267055690289, + 0.005773589480668306, + 0.014535377733409405, + 0.009772008284926414, + 0.010572425089776516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.1344204843044281 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium-length shot focusing on a gnarled southern orchid tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in extreme high-contrast black and white with deep black and pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tree trunk exhibit twisted, convex forms with rough, graphic shading emphasizing contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do smooth, perfectly spherical black spheres erupt from cracks in the bark and expand rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the tree trunk throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012562546879053116 + ], + "opposite_logprob_tokens": [ + -4.3875627517700195 + ], + "logprob_reward": -0.012562546879053116, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012562546879053116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04608358070254326, + 0.03378111496567726, + 0.02889838069677353, + 0.012325236573815346, + 0.03783620893955231, + 0.03238576278090477, + 0.06474752724170685, + 0.08175142109394073, + 0.21600836515426636, + 0.4461824893951416 + ], + "credit_map": [ + 0.006791712716221809, + 0.007205929607152939, + 0.008005484938621521, + 0.0059045408852398396, + 0.003930920735001564, + 0.004466869402676821, + 0.004648284520953894, + 0.005129834171384573, + 0.007018792908638716, + 0.0039229197427630424, + 0.005096136592328548, + 0.0035431738942861557, + 0.004908368922770023, + 0.0029055860359221697, + 0.0024362364783883095, + 0.003949895966798067, + 0.005863983649760485, + 0.004158613737672567, + 0.003047253703698516, + 0.0028313982766121626, + 0.0033589210361242294, + 0.0030388450250029564, + 0.0023064538836479187, + 0.004292909521609545, + 0.0018110719975084066, + 0.0015427980106323957, + 0.0012157626915723085, + 0.001448701717890799, + 0.0014308224199339747, + 0.0016199973179027438, + 0.0018396669765934348, + 0.0014164140447974205, + 0.004048366565257311, + 0.004260634072124958, + 0.003948940429836512, + 0.005336871836334467, + 0.004319679923355579, + 0.006507207173854113, + 0.005242083687335253, + 0.004172423854470253, + 0.003370126709342003, + 0.0025478277821093798, + 0.0035131704062223434, + 0.0028621454257518053, + 0.002202589763328433, + 0.0074467845261096954, + 0.0024039281997829676, + 0.008039186708629131, + 0.00702632823958993, + 0.006569042801856995, + 0.007539889309555292, + 0.012022861279547215, + 0.0055519999004900455, + 0.008700491860508919, + 0.009558040648698807, + 0.007778863422572613, + 0.014798990450799465, + 0.008591737598180771, + 0.009087512269616127, + 0.012603961862623692, + 0.007354182656854391, + 0.008946659043431282, + 0.0093687130138278, + 0.010999652557075024, + 0.06890610605478287, + 0.015637889504432678, + 0.015803704038262367, + 0.012643329799175262, + 0.009666803292930126, + 0.05387388914823532, + 0.007480606436729431, + 0.031996045261621475, + 0.035574205219745636, + 0.025271590799093246, + 0.025366077199578285, + 0.03455096110701561, + 0.022634826600551605, + 0.1361558884382248, + 0.02956083044409752, + 0.1370680332183838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.012562546879053116 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium-length shot focusing on a gnarled southern orchid tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in extreme high-contrast black and white with deep black and pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tree trunk exhibit twisted, convex forms with rough, graphic shading emphasizing contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do smooth, perfectly spherical black spheres erupt from cracks in the bark and expand rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the tree trunk throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013394916895776987 + ], + "opposite_logprob_tokens": [ + -6.626339435577393 + ], + "logprob_reward": -0.0013394916895776987, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013394916895776987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17443647980690002, + 0.1649123877286911, + 0.06820249557495117, + 0.03342286869883537, + 0.060412026941776276, + 0.05798126012086868, + 0.05669194832444191, + 0.06061292439699173, + 0.14636491239070892, + 0.1769627332687378 + ], + "credit_map": [ + 0.025593122467398643, + 0.03179172798991203, + 0.019083639606833458, + 0.030812816694378853, + 0.017565228044986725, + 0.013108339160680771, + 0.020002063363790512, + 0.016479527577757835, + 0.04282339662313461, + 0.017054060474038124, + 0.024921957403421402, + 0.015361188910901546, + 0.015488755889236927, + 0.014367152005434036, + 0.011893614195287228, + 0.023002250120043755, + 0.012903689406812191, + 0.008911137469112873, + 0.0057375202886760235, + 0.008726535364985466, + 0.007680234033614397, + 0.0066819810308516026, + 0.008237980306148529, + 0.009323414415121078, + 0.00454606581479311, + 0.004188482649624348, + 0.0023274323903024197, + 0.005151957273483276, + 0.0035240917932242155, + 0.00348590943031013, + 0.006338291335850954, + 0.00386063358746469, + 0.006845857948064804, + 0.0053860098123550415, + 0.005751042161136866, + 0.009620812721550465, + 0.007323315367102623, + 0.008731486275792122, + 0.009869405068457127, + 0.006884096190333366, + 0.006079890299588442, + 0.0058038379065692425, + 0.00646651117131114, + 0.007717742584645748, + 0.006019331514835358, + 0.010418037883937359, + 0.0048563238233327866, + 0.010619574226439, + 0.008278378285467625, + 0.00638945447281003, + 0.006107298657298088, + 0.011113353073596954, + 0.0042050667107105255, + 0.00788588635623455, + 0.008064934983849525, + 0.004647568799555302, + 0.009303472004830837, + 0.005379611160606146, + 0.007525495253503323, + 0.007514063268899918, + 0.0070470175705850124, + 0.008702188730239868, + 0.006437214557081461, + 0.008703857660293579, + 0.04263361915946007, + 0.009976053610444069, + 0.010373544879257679, + 0.008312401361763477, + 0.006779234390705824, + 0.05211273208260536, + 0.004596005659550428, + 0.011581305414438248, + 0.022473225370049477, + 0.016839127987623215, + 0.02190011367201805, + 0.030502276495099068, + 0.015360635705292225, + 0.020173510536551476, + 0.025603504851460457, + 0.024110330268740654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.0013394916895776987 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium-length shot focusing on a gnarled southern orchid tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in extreme high-contrast black and white with deep black and pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tree trunk exhibit twisted, convex forms with rough, graphic shading emphasizing contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do smooth, perfectly spherical black spheres erupt from cracks in the bark and expand rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the tree trunk throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058791182935238 + ], + "opposite_logprob_tokens": [ + -3.895587921142578 + ], + "logprob_reward": -0.02058791182935238, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058791182935238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06465381383895874, + 0.042787060141563416, + 0.050928544253110886, + 0.022462211549282074, + 0.0433826819062233, + 0.08984040468931198, + 0.13127192854881287, + 0.05145851522684097, + 0.04538102447986603, + 0.457833856344223 + ], + "credit_map": [ + 0.010117684490978718, + 0.011193370446562767, + 0.006926459725946188, + 0.010509561747312546, + 0.007506112568080425, + 0.005534360185265541, + 0.007270808797329664, + 0.005595456808805466, + 0.008296353742480278, + 0.0045654973946511745, + 0.0038460856303572655, + 0.004639806225895882, + 0.007840126752853394, + 0.005175278056412935, + 0.003049429738894105, + 0.005374483298510313, + 0.007002768572419882, + 0.007455103099346161, + 0.00495353527367115, + 0.006376094184815884, + 0.0064590973779559135, + 0.006928835995495319, + 0.005274366121739149, + 0.006478747818619013, + 0.0030690249986946583, + 0.0037099088076502085, + 0.002560604829341173, + 0.0024352618493139744, + 0.002065955428406596, + 0.0025473779533058405, + 0.0035745699424296618, + 0.002499508438631892, + 0.005096732638776302, + 0.004661960527300835, + 0.003594361012801528, + 0.010495103895664215, + 0.003919221460819244, + 0.005820123944431543, + 0.005085750482976437, + 0.004709429573267698, + 0.009557158686220646, + 0.01035650447010994, + 0.006897947750985622, + 0.01547389104962349, + 0.007824369706213474, + 0.015204215422272682, + 0.010556508786976337, + 0.013969814404845238, + 0.01868448778986931, + 0.011981470510363579, + 0.01298567745834589, + 0.02713927812874317, + 0.009392782114446163, + 0.018415790051221848, + 0.019788527861237526, + 0.012883918359875679, + 0.006552089471369982, + 0.005442148074507713, + 0.004947226494550705, + 0.007398880086839199, + 0.004292545840144157, + 0.005231339950114489, + 0.00826259981840849, + 0.009331687353551388, + 0.01231773104518652, + 0.0039447275921702385, + 0.003651292994618416, + 0.003670522477477789, + 0.002435806207358837, + 0.00640422897413373, + 0.002395998453721404, + 0.01056071650236845, + 0.03415156155824661, + 0.0343000665307045, + 0.03732377663254738, + 0.05348917841911316, + 0.026524901390075684, + 0.11816385388374329, + 0.04128757491707802, + 0.11259295046329498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 5 + ], + "selected_reward": -0.02058791182935238 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium-length shot focusing on a gnarled southern orchid tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in extreme high-contrast black and white with deep black and pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tree trunk exhibit twisted, convex forms with rough, graphic shading emphasizing contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do smooth, perfectly spherical black spheres erupt from cracks in the bark and expand rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the tree trunk throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979438751935959 + ], + "opposite_logprob_tokens": [ + -3.529794454574585 + ], + "logprob_reward": -0.02979438751935959, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979438751935959, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04603315144777298, + 0.04773494973778725, + 0.03759730979800224, + 0.02181086875498295, + 0.04673131927847862, + 0.1066281720995903, + 0.20676013827323914, + 0.21757741272449493, + 0.23220373690128326, + 0.036922939121723175 + ], + "credit_map": [ + 0.007561556063592434, + 0.009993298910558224, + 0.003971182741224766, + 0.0069947573356330395, + 0.005143338814377785, + 0.00416224543005228, + 0.004492811392992735, + 0.0037139663472771645, + 0.008898363448679447, + 0.005756180267781019, + 0.004769598133862019, + 0.005558802280575037, + 0.006952343508601189, + 0.004709009081125259, + 0.005389142315834761, + 0.005701512563973665, + 0.006721910554915667, + 0.004879895597696304, + 0.0042029209434986115, + 0.0039824326522648335, + 0.004245459567755461, + 0.004726190119981766, + 0.00344637269154191, + 0.005392131395637989, + 0.003197187790647149, + 0.0021838215179741383, + 0.002208384685218334, + 0.003077614586800337, + 0.0028484256472438574, + 0.0032174375373870134, + 0.002343498868867755, + 0.0027344992849975824, + 0.005334042012691498, + 0.006771481595933437, + 0.005267904605716467, + 0.00864424929022789, + 0.005171989090740681, + 0.004804595373570919, + 0.006733270362019539, + 0.004003792069852352, + 0.010762481018900871, + 0.006027475465089083, + 0.007238215766847134, + 0.01029619388282299, + 0.00979822687804699, + 0.025818290188908577, + 0.009665103629231453, + 0.027022188529372215, + 0.029705338180065155, + 0.020716240629553795, + 0.022849280387163162, + 0.029381660744547844, + 0.019491614773869514, + 0.029112545773386955, + 0.02576386369764805, + 0.029739616438746452, + 0.04250962287187576, + 0.017035391181707382, + 0.031302496790885925, + 0.029137983918190002, + 0.01777411252260208, + 0.02887905202805996, + 0.02114529348909855, + 0.029793476685881615, + 0.04107833281159401, + 0.0184060949832201, + 0.02146659977734089, + 0.01881604827940464, + 0.015788570046424866, + 0.0625486746430397, + 0.015153532847762108, + 0.03894590958952904, + 0.003071429906412959, + 0.0023371458519250154, + 0.0028306704480201006, + 0.003521069884300232, + 0.00163507170509547, + 0.00871002022176981, + 0.0029579896945506334, + 0.011859545484185219 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.02979438751935959 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_1/index_tensor([7729, 7729])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_1/index_tensor([7729, 7729])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3bfc7ce3ee8e2b72f4df1fe7fcfbf71f1d04044a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_1/index_tensor([7729, 7729])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, dynamic tracking shot follows a fully armored knight and his steed, the camera synchronizing them as they speed through a vast birch forest. The knight, clad in a full suit of meticulously polished, shimmering steel plate armor and a closed-faced helmet, his face obscured. A heavy crimson cape billows behind him at full speed, like a streak of flame. He leans forward, gripping the reins tightly, steadily guiding a powerful, muscular warhorse with a coat as sleek as obsidian. Golden afternoon sunlight penetrates the towering silvery-white birch canopy, casting a shifting, dappled light on the forest floor carpeted with golden and russet fallen leaves. With every powerful kick of the warhorse, its iron-shod hooves would suddenly lift the thick layer of dry fallen leaves on the ground backwards, forming a golden trail that spun, danced, and lingered behind them. The whole picture was full of a sense of speed and power. The rapid changes in light and shadow were intertwined with the flying fallen leaves, creating a tense and magnificent atmosphere of epic adventure.", + "seed": 3222, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_1/index_7729_seed_3222_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.2801338840276003, + "vlm_logprob": -0.2801338840276003, + "vlm_yes_no_margin": 2.125, + "vlm_token_credit": -0.2801338840276003, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a full suit of polished steel plate armor with a closed-faced helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warhorse's coat black and sleek like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy crimson cape billow behind the knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the horse's iron-shod hooves kick up dry fallen leaves forming a spinning golden trail behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4431177079677582 + ], + "opposite_logprob_tokens": [ + -1.3181177377700806 + ], + "logprob_reward": -0.4431177079677582, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.4431177079677582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1869623214006424, + 0.041726723313331604, + 0.07252795994281769, + 0.05703793838620186, + 0.09283450990915298, + 0.049139078706502914, + 0.188070610165596, + 0.04934334754943848, + 0.1552213430404663, + 0.10713616758584976 + ], + "credit_map": [ + 0.02414127066731453, + 0.011258675716817379, + 0.022617245092988014, + 0.0331970751285553, + 0.034692082554101944, + 0.018597161397337914, + 0.0247640460729599, + 0.017694778740406036, + 0.004454866982996464, + 0.0029512064065784216, + 0.006063577253371477, + 0.01147768646478653, + 0.003186124609783292, + 0.0049285609275102615, + 0.0037814450915902853, + 0.004883254878222942, + 0.009387749247252941, + 0.004179785028100014, + 0.016302073374390602, + 0.009030989371240139, + 0.005282869096845388, + 0.004334891680628061, + 0.014075912535190582, + 0.009933692403137684, + 0.006717774085700512, + 0.005382210947573185, + 0.00842117890715599, + 0.010539689101278782, + 0.004605933558195829, + 0.004249919671565294, + 0.00907041970640421, + 0.008050814270973206, + 0.008158503100275993, + 0.006561344489455223, + 0.026738053187727928, + 0.007907696068286896, + 0.005013200920075178, + 0.006114520598202944, + 0.007757943123579025, + 0.024583250284194946, + 0.005230551585555077, + 0.003463509026914835, + 0.008402243256568909, + 0.005620251875370741, + 0.0069106705486774445, + 0.008293594233691692, + 0.005081832874566317, + 0.006136422976851463, + 0.052115801721811295, + 0.008052002638578415, + 0.022544249892234802, + 0.00821267906576395, + 0.013327364809811115, + 0.058427706360816956, + 0.01072611752897501, + 0.01466468907892704, + 0.005676956847310066, + 0.004437482915818691, + 0.00640587555244565, + 0.004875833168625832, + 0.007400503847748041, + 0.005738293286412954, + 0.0050625731237232685, + 0.009745829738676548, + 0.03439569100737572, + 0.018282826989889145, + 0.019002439454197884, + 0.017320172861218452, + 0.015077563002705574, + 0.011144284158945084, + 0.020164798945188522, + 0.019833559170365334, + 0.014575407840311527, + 0.013457601889967918, + 0.014145788736641407, + 0.011949006468057632, + 0.01181088574230671, + 0.009833397343754768, + 0.01623508334159851, + 0.015128996223211288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.4431177079677582 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a full suit of polished steel plate armor with a closed-faced helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warhorse's coat black and sleek like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy crimson cape billow behind the knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the horse's iron-shod hooves kick up dry fallen leaves forming a spinning golden trail behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261942863464355 + ], + "opposite_logprob_tokens": [ + -0.5761942863464355 + ], + "logprob_reward": -0.8261942863464355, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261942863464355, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11718736588954926, + 0.11797240376472473, + 0.13499242067337036, + 0.10497895628213882, + 0.21483469009399414, + 0.12244664132595062, + 0.036487698554992676, + 0.07218284904956818, + 0.025721438229084015, + 0.05319558456540108 + ], + "credit_map": [ + 0.013985252939164639, + 0.0073328884318470955, + 0.013750181533396244, + 0.02563529647886753, + 0.01337064616382122, + 0.011630343273282051, + 0.0158760454505682, + 0.015606706030666828, + 0.019348112866282463, + 0.007632224354892969, + 0.015948167070746422, + 0.023842062801122665, + 0.010288581252098083, + 0.01328545156866312, + 0.01390447560697794, + 0.013723339885473251, + 0.012130610644817352, + 0.008862984366714954, + 0.022088374942541122, + 0.017177600413560867, + 0.01483769528567791, + 0.009199456311762333, + 0.029725953936576843, + 0.020969735458493233, + 0.00920162070542574, + 0.010053498670458794, + 0.010835402645170689, + 0.02423851378262043, + 0.00938117504119873, + 0.006673457100987434, + 0.018824875354766846, + 0.015770411118865013, + 0.018231293186545372, + 0.009533271193504333, + 0.06824307888746262, + 0.017648592591285706, + 0.010073252953588963, + 0.012077421881258488, + 0.009858337230980396, + 0.06916943937540054, + 0.02194175310432911, + 0.01327078603208065, + 0.015735706314444542, + 0.011293225921690464, + 0.01411435380578041, + 0.023492757230997086, + 0.01126823015511036, + 0.011329836212098598, + 0.00665463600307703, + 0.0026941222604364157, + 0.007172858342528343, + 0.0031466083601117134, + 0.0050735981203615665, + 0.003645262448117137, + 0.004041233565658331, + 0.004059378989040852, + 0.008446627296507359, + 0.008974645286798477, + 0.009656976908445358, + 0.00814948882907629, + 0.011954368092119694, + 0.009497324004769325, + 0.008924578316509724, + 0.006578846834599972, + 0.004394304007291794, + 0.0029566576704382896, + 0.003909258171916008, + 0.003536309814080596, + 0.0025893242564052343, + 0.0015583839267492294, + 0.003901675110682845, + 0.0028755252715200186, + 0.008930322714149952, + 0.006226849742233753, + 0.00594341242685914, + 0.0049184286035597324, + 0.009897108189761639, + 0.0037249643355607986, + 0.006642399355769157, + 0.0069120945408940315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 5 + ], + "selected_reward": -0.8261942863464355 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a full suit of polished steel plate armor with a closed-faced helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warhorse's coat black and sleek like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy crimson cape billow behind the knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the horse's iron-shod hooves kick up dry fallen leaves forming a spinning golden trail behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07904914766550064 + ], + "opposite_logprob_tokens": [ + -2.5790491104125977 + ], + "logprob_reward": -0.07904914766550064, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07904914766550064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2685069739818573, + 0.09307170659303665, + 0.144374281167984, + 0.05487920343875885, + 0.09628438204526901, + 0.0865471139550209, + 0.07193002104759216, + 0.10491837561130524, + 0.04316531494259834, + 0.03632268309593201 + ], + "credit_map": [ + 0.03135865926742554, + 0.017692113295197487, + 0.028385983780026436, + 0.03781130537390709, + 0.045575156807899475, + 0.029597900807857513, + 0.04546476900577545, + 0.03262108564376831, + 0.011339214630424976, + 0.005846613552421331, + 0.013118335977196693, + 0.023146353662014008, + 0.009833273477852345, + 0.007879551500082016, + 0.00966999214142561, + 0.012238376773893833, + 0.013982672244310379, + 0.008260872215032578, + 0.022999271750450134, + 0.01934882625937462, + 0.016615355387330055, + 0.011921528726816177, + 0.025389615446329117, + 0.025856127962470055, + 0.00806443952023983, + 0.0065859523601830006, + 0.006788450758904219, + 0.00880384910851717, + 0.005716762039810419, + 0.004374275915324688, + 0.00844509620219469, + 0.006100375670939684, + 0.009912027977406979, + 0.00467647286131978, + 0.02175355888903141, + 0.0073975506238639355, + 0.00597353046759963, + 0.006760827731341124, + 0.00791066512465477, + 0.03189975023269653, + 0.01777101494371891, + 0.008971489034593105, + 0.01052990835160017, + 0.007247368339449167, + 0.009846330620348454, + 0.01653708517551422, + 0.007993926294147968, + 0.007649997249245644, + 0.007886811159551144, + 0.006475756876170635, + 0.010884001851081848, + 0.008899413980543613, + 0.00922528374940157, + 0.009883569553494453, + 0.009251624345779419, + 0.009423554874956608, + 0.012410659343004227, + 0.0137307308614254, + 0.013381581753492355, + 0.0103954141959548, + 0.015948282554745674, + 0.01432814821600914, + 0.014459123834967613, + 0.010264434851706028, + 0.004520530812442303, + 0.004799974150955677, + 0.005781199783086777, + 0.007632634602487087, + 0.006176829803735018, + 0.0030865748412907124, + 0.007198865059763193, + 0.0039687021635472775, + 0.00574645958840847, + 0.00555347790941596, + 0.005000877194106579, + 0.004746290855109692, + 0.003654616652056575, + 0.0031612429302185774, + 0.004380637314170599, + 0.004079077858477831 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.07904914766550064 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a full suit of polished steel plate armor with a closed-faced helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warhorse's coat black and sleek like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy crimson cape billow behind the knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the horse's iron-shod hooves kick up dry fallen leaves forming a spinning golden trail behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810510411858559 + ], + "opposite_logprob_tokens": [ + -3.288105010986328 + ], + "logprob_reward": -0.03810510411858559, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810510411858559, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12866093218326569, + 0.08425495028495789, + 0.07422546297311783, + 0.061077017337083817, + 0.07321130484342575, + 0.1206660270690918, + 0.10030625015497208, + 0.16940724849700928, + 0.07443622499704361, + 0.11375461518764496 + ], + "credit_map": [ + 0.018999159336090088, + 0.010214154608547688, + 0.018411584198474884, + 0.02603645622730255, + 0.013708926737308502, + 0.014125898480415344, + 0.01219707727432251, + 0.014967665076255798, + 0.012758866883814335, + 0.005328591447323561, + 0.010755951516330242, + 0.015298478305339813, + 0.008824612945318222, + 0.008118907921016216, + 0.010138954035937786, + 0.013030573725700378, + 0.005127847660332918, + 0.008494500070810318, + 0.010755213908851147, + 0.012298078276216984, + 0.0064964210614562035, + 0.005212928634136915, + 0.014312188141047955, + 0.011528272181749344, + 0.005827232263982296, + 0.007213886361569166, + 0.007199365179985762, + 0.0123891057446599, + 0.004576817620545626, + 0.0030943062156438828, + 0.011279100552201271, + 0.009497194550931454, + 0.010766328312456608, + 0.003942356910556555, + 0.012506276369094849, + 0.006005339790135622, + 0.004049824085086584, + 0.008059117943048477, + 0.0070968689396977425, + 0.02078518457710743, + 0.020706376060843468, + 0.011221892200410366, + 0.013111472129821777, + 0.013192038983106613, + 0.024187948554754257, + 0.016568293794989586, + 0.010785113088786602, + 0.010892882943153381, + 0.015793152153491974, + 0.005731870420277119, + 0.013835964724421501, + 0.006823903415352106, + 0.013554135337471962, + 0.02858859673142433, + 0.007569898385554552, + 0.008408717811107635, + 0.01499959360808134, + 0.027093615382909775, + 0.01441179495304823, + 0.015863122418522835, + 0.04645270109176636, + 0.020878050476312637, + 0.013776092790067196, + 0.015932250767946243, + 0.011350772343575954, + 0.009653721004724503, + 0.0081636942923069, + 0.00749235600233078, + 0.015451041050255299, + 0.005989660508930683, + 0.009006754495203495, + 0.0073282187804579735, + 0.021411918103694916, + 0.0126589210703969, + 0.010837090201675892, + 0.011563980020582676, + 0.023649491369724274, + 0.008074012584984303, + 0.013709884136915207, + 0.01184931956231594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.03810510411858559 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a full suit of polished steel plate armor with a closed-faced helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warhorse's coat black and sleek like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy crimson cape billow behind the knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the horse's iron-shod hooves kick up dry fallen leaves forming a spinning golden trail behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014203174039721489 + ], + "opposite_logprob_tokens": [ + -4.264203071594238 + ], + "logprob_reward": -0.014203174039721489, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014203174039721489, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19562941789627075, + 0.05787261947989464, + 0.11776093393564224, + 0.0525185689330101, + 0.04490721970796585, + 0.10100267082452774, + 0.09282924979925156, + 0.07862726598978043, + 0.19175966084003448, + 0.06709246337413788 + ], + "credit_map": [ + 0.03225775435566902, + 0.013254391960799694, + 0.025514274835586548, + 0.04167265444993973, + 0.016197727993130684, + 0.012313717976212502, + 0.02450808323919773, + 0.02991080842912197, + 0.00781262293457985, + 0.003514836775138974, + 0.00854059774428606, + 0.013593386858701706, + 0.005606824532151222, + 0.00486123189330101, + 0.006015722639858723, + 0.007927393540740013, + 0.01618329994380474, + 0.015732409432530403, + 0.020406296476721764, + 0.01858079992234707, + 0.010365585796535015, + 0.009526928886771202, + 0.014602765440940857, + 0.01236285176128149, + 0.00787684042006731, + 0.005713151767849922, + 0.006246058735996485, + 0.00997745431959629, + 0.0036960646975785494, + 0.0043254429474473, + 0.006186665967106819, + 0.008496887981891632, + 0.004314215388149023, + 0.0029813076835125685, + 0.00920437928289175, + 0.00519590312615037, + 0.005259677302092314, + 0.0035634974483400583, + 0.00510176969692111, + 0.009286468848586082, + 0.012834295630455017, + 0.004987041465938091, + 0.023222798481583595, + 0.00872961524873972, + 0.00952176284044981, + 0.024258039891719818, + 0.008674542419612408, + 0.008774573914706707, + 0.010980721563100815, + 0.007089204154908657, + 0.014318671077489853, + 0.007095751818269491, + 0.01525403093546629, + 0.017892425879836082, + 0.010891621932387352, + 0.009306825697422028, + 0.018342455849051476, + 0.006967491935938597, + 0.009886300191283226, + 0.00739047396928072, + 0.00993999931961298, + 0.006753767374902964, + 0.007136811967939138, + 0.012209967710077763, + 0.035278793424367905, + 0.01930169016122818, + 0.024398425593972206, + 0.027661582455039024, + 0.026170143857598305, + 0.013001810759305954, + 0.026939675211906433, + 0.01900753378868103, + 0.010645061731338501, + 0.00781832356005907, + 0.006882253102958202, + 0.006879058200865984, + 0.013798709027469158, + 0.004750668536871672, + 0.00878496840596199, + 0.007533418480306864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.014203174039721489 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, dynamic tracking shot follows a fully armored knight and his steed, the camera synchronizing them as they speed through a vast birch forest. The knight, clad in a full suit of meticulously polished, shimmering steel plate armor and a closed-faced helmet, his face obscured. A heavy crimson cape billows behind him at full speed, like a streak of flame. He leans forward, gripping the reins tightly, steadily guiding a powerful, muscular warhorse with a coat as sleek as obsidian. Golden afternoon sunlight penetrates the towering silvery-white birch canopy, casting a shifting, dappled light on the forest floor carpeted with golden and russet fallen leaves. With every powerful kick of the warhorse, its iron-shod hooves would suddenly lift the thick layer of dry fallen leaves on the ground backwards, forming a golden trail that spun, danced, and lingered behind them. The whole picture was full of a sense of speed and power. The rapid changes in light and shadow were intertwined with the flying fallen leaves, creating a tense and magnificent atmosphere of epic adventure.", + "seed": 3223, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_1/index_7729_seed_3223_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2087904341518879, + "vlm_logprob": -0.2087904341518879, + "vlm_yes_no_margin": 2.2749999523162843, + "vlm_token_credit": -0.2087904341518879, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a full suit of polished steel plate armor with a closed-faced helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warhorse's coat black and sleek like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy crimson cape billow behind the knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the horse's iron-shod hooves kick up dry fallen leaves forming a spinning golden trail behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3493977189064026 + ], + "opposite_logprob_tokens": [ + -1.4743976593017578 + ], + "logprob_reward": -0.3493977189064026, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.3493977189064026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.022283989936113358, + 0.009610174223780632, + 0.02973155304789543, + 0.05622214823961258, + 0.013950154185295105, + 0.005805745255202055, + 0.004966842941939831, + 0.00664766039699316, + 0.8398988842964172, + 0.010882865637540817 + ], + "credit_map": [ + 0.0024190139956772327, + 0.001963412156328559, + 0.0024945628829300404, + 0.0034050929825752974, + 0.0021352367475628853, + 0.0039731902070343494, + 0.0030966743361204863, + 0.002796805463731289, + 0.0007312821689993143, + 0.0007613636553287506, + 0.001019527087919414, + 0.0019336596596986055, + 0.0010184366255998611, + 0.0014001799281686544, + 0.0017157623078674078, + 0.001029962906613946, + 0.0021903591696172953, + 0.0026140750851482153, + 0.005568271037191153, + 0.01075789611786604, + 0.0011758811306208372, + 0.0016612311592325568, + 0.00203719618730247, + 0.0037266442086547613, + 0.013045930303633213, + 0.0024744055699557066, + 0.020413264632225037, + 0.007306678686290979, + 0.0020231616217643023, + 0.002702336525544524, + 0.003972128964960575, + 0.004284243565052748, + 0.0018571445252746344, + 0.000852509867399931, + 0.002101931953802705, + 0.0015483754687011242, + 0.0009347847662866116, + 0.0009774601785466075, + 0.0015212767757475376, + 0.004156670067459345, + 0.0009063109173439443, + 0.001066804165020585, + 0.0007689099875278771, + 0.0006312282057479024, + 0.0005764569505117834, + 0.0007980672526173294, + 0.0005366367404349148, + 0.0005213316762819886, + 0.0006936541176401079, + 0.00060818757629022, + 0.0007323879981413484, + 0.0007050578715279698, + 0.0006194861489348114, + 0.0003954295825678855, + 0.00045110596693120897, + 0.0007615338545292616, + 0.0005020000971853733, + 0.000597227131947875, + 0.001134626567363739, + 0.0009024576283991337, + 0.0007227056194096804, + 0.0005696896696463227, + 0.0013862897176295519, + 0.0008326642564497888, + 0.07729550451040268, + 0.1743042916059494, + 0.27002325654029846, + 0.04278339073061943, + 0.033379580825567245, + 0.04095054045319557, + 0.16246874630451202, + 0.038693539798259735, + 0.0009318138472735882, + 0.0006312301848083735, + 0.0015568326925858855, + 0.0013194542843848467, + 0.0011705659562721848, + 0.0009290337329730392, + 0.0015179701149463654, + 0.0028259647078812122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 2 + ], + "selected_reward": -0.3493977189064026 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a full suit of polished steel plate armor with a closed-faced helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warhorse's coat black and sleek like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy crimson cape billow behind the knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the horse's iron-shod hooves kick up dry fallen leaves forming a spinning golden trail behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761847496032715 + ], + "opposite_logprob_tokens": [ + -0.8261847496032715 + ], + "logprob_reward": -0.5761847496032715, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761847496032715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02118297666311264, + 0.01387268491089344, + 0.026767153292894363, + 0.05073864012956619, + 0.02205575630068779, + 0.007335493806749582, + 0.004158842843025923, + 0.00996648520231247, + 0.8057793378829956, + 0.03814256191253662 + ], + "credit_map": [ + 0.0017062288243323565, + 0.0016277381218969822, + 0.002227113349363208, + 0.004019498825073242, + 0.0022512958385050297, + 0.004131678491830826, + 0.002521870657801628, + 0.002697556046769023, + 0.0014797261683270335, + 0.0012847963953390718, + 0.0019056398887187243, + 0.001964581198990345, + 0.0015673070447519422, + 0.0018990003736689687, + 0.002004592213779688, + 0.0017670427914708853, + 0.0020510812755674124, + 0.002398282755166292, + 0.004713669419288635, + 0.007946802303195, + 0.002061164937913418, + 0.0019750408828258514, + 0.002320758532732725, + 0.0033003559801727533, + 0.010885636322200298, + 0.003433268517255783, + 0.017291059717535973, + 0.006206225138157606, + 0.00267944298684597, + 0.002630770904943347, + 0.00407613068819046, + 0.003536105854436755, + 0.003299076110124588, + 0.001287944964133203, + 0.0032707122154533863, + 0.0017909799935296178, + 0.0014874690677970648, + 0.0016693680081516504, + 0.0022926197852939367, + 0.0069575863890349865, + 0.000748181133531034, + 0.0008722388884052634, + 0.0009257845231331885, + 0.0006863723392598331, + 0.0011880702804774046, + 0.0008305135415866971, + 0.0011913558701053262, + 0.0008929777541197836, + 0.000452014384791255, + 0.00039589652442373335, + 0.0006056299898773432, + 0.0005255606374703348, + 0.0005064845900051296, + 0.00040716593503020704, + 0.0007695001550018787, + 0.000496590742841363, + 0.0007737422711215913, + 0.0011413806350901723, + 0.001545171719044447, + 0.0013106412952765822, + 0.002031933516263962, + 0.00086074328282848, + 0.0013980544172227383, + 0.0009048185311257839, + 0.07408171147108078, + 0.16498315334320068, + 0.2599281966686249, + 0.041019655764102936, + 0.03190455585718155, + 0.03897278010845184, + 0.15738148987293243, + 0.03750789165496826, + 0.0029342479538172483, + 0.002003173343837261, + 0.00568374851718545, + 0.004548634402453899, + 0.003960661590099335, + 0.0032387312967330217, + 0.005546294152736664, + 0.010227073915302753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 9 + ], + "selected_reward": -0.5761847496032715 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a full suit of polished steel plate armor with a closed-faced helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warhorse's coat black and sleek like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy crimson cape billow behind the knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the horse's iron-shod hooves kick up dry fallen leaves forming a spinning golden trail behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062081895768642426 + ], + "opposite_logprob_tokens": [ + -2.812081813812256 + ], + "logprob_reward": -0.062081895768642426, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062081895768642426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0306538175791502, + 0.012974460609257221, + 0.02482346072793007, + 0.03686072677373886, + 0.011679778806865215, + 0.006083338987082243, + 0.025228196755051613, + 0.02287261001765728, + 0.7312480211257935, + 0.09757557511329651 + ], + "credit_map": [ + 0.0027560852468013763, + 0.002244174713268876, + 0.0031464220955967903, + 0.005629114340990782, + 0.005131194833666086, + 0.004716676659882069, + 0.0035772414412349463, + 0.003452905686572194, + 0.0014866754645481706, + 0.0011427364079281688, + 0.0019455071305856109, + 0.0020996276289224625, + 0.0013721334980800748, + 0.0018220603233203292, + 0.0015479158610105515, + 0.0015578035963699222, + 0.00325088482350111, + 0.0030060389544814825, + 0.003602278884500265, + 0.0037291839253157377, + 0.0026767610106617212, + 0.00332133867777884, + 0.0024589975364506245, + 0.002777976216748357, + 0.006413168273866177, + 0.004132813774049282, + 0.008845632895827293, + 0.004669873509556055, + 0.002740690717473626, + 0.002768438309431076, + 0.0031706735026091337, + 0.00411943718791008, + 0.0011271369876340032, + 0.0012550043174996972, + 0.0015136254951357841, + 0.001773105701431632, + 0.001981977839022875, + 0.0014274456771090627, + 0.0014746097149327397, + 0.0011268726084381342, + 0.0010054674930870533, + 0.0009121556067839265, + 0.0007522863452322781, + 0.0007626333390362561, + 0.000631328031886369, + 0.0008395119803026319, + 0.0006478716968558729, + 0.000532084668520838, + 0.0017940397374331951, + 0.0020133161451667547, + 0.003447491442784667, + 0.0035227984189987183, + 0.0037337657995522022, + 0.0021387257147580385, + 0.004767789505422115, + 0.0038102699909359217, + 0.0022150517906993628, + 0.0020928061567246914, + 0.0033350379671901464, + 0.0025112975854426622, + 0.0028096302412450314, + 0.0018011560896411538, + 0.004228606354445219, + 0.0038790239486843348, + 0.06642595678567886, + 0.1382514089345932, + 0.23968809843063354, + 0.0370798297226429, + 0.028059925884008408, + 0.033731281757354736, + 0.15150152146816254, + 0.03650997206568718, + 0.00751426862552762, + 0.006928702350705862, + 0.01332774292677641, + 0.010349712334573269, + 0.013542565517127514, + 0.00812444742769003, + 0.014341888949275017, + 0.023446252569556236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.062081895768642426 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a full suit of polished steel plate armor with a closed-faced helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warhorse's coat black and sleek like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy crimson cape billow behind the knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the horse's iron-shod hooves kick up dry fallen leaves forming a spinning golden trail behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038101088255643845 + ], + "opposite_logprob_tokens": [ + -3.2881011962890625 + ], + "logprob_reward": -0.038101088255643845, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038101088255643845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04131738469004631, + 0.0049067297950387, + 0.028391879051923752, + 0.020269840955734253, + 0.01117817685008049, + 0.004220794420689344, + 0.0067000119015574455, + 0.002584368223324418, + 0.8601197600364685, + 0.020311059430241585 + ], + "credit_map": [ + 0.003365828189998865, + 0.005674938205629587, + 0.003666076809167862, + 0.0043189809657633305, + 0.0032171239145100117, + 0.01391518209129572, + 0.003426960203796625, + 0.0037322943098843098, + 0.0007251522038131952, + 0.00042599154403433204, + 0.0005924516590312123, + 0.0006580593180842698, + 0.0005159923457540572, + 0.0007790614617988467, + 0.0005604263860732317, + 0.0006495946436189115, + 0.0019771154038608074, + 0.002722719917073846, + 0.0054969703778624535, + 0.010572746396064758, + 0.001005988335236907, + 0.0016820576274767518, + 0.001638541230931878, + 0.003295738948509097, + 0.0049276771023869514, + 0.001134140184149146, + 0.0070683276280760765, + 0.0024623076897114515, + 0.0009588965331204236, + 0.0009850726928561926, + 0.001491308561526239, + 0.0012421098072081804, + 0.0014073991915211082, + 0.0015181476483121514, + 0.0013655002694576979, + 0.0011448545847088099, + 0.0009247015696018934, + 0.0025400014128535986, + 0.000817581545561552, + 0.001459990395233035, + 0.00041468243580311537, + 0.0005063786520622671, + 0.000655492243822664, + 0.00041077169589698315, + 0.0006511661340482533, + 0.000454021617770195, + 0.0005731162382289767, + 0.0005551656940951943, + 0.0006339296232908964, + 0.0004966869601048529, + 0.0008770344429649413, + 0.0009876488475129008, + 0.0005576817784458399, + 0.0006994783761911094, + 0.0013546899426728487, + 0.0010928622214123607, + 0.00042788346763700247, + 0.00019698972755577415, + 0.00040687574073672295, + 0.0003061148745473474, + 0.0003788527683354914, + 0.00022085908858571202, + 0.00035433904849924147, + 0.00029245339101180434, + 0.07874781638383865, + 0.17215760052204132, + 0.2789303958415985, + 0.043782494962215424, + 0.033665429800748825, + 0.0410061813890934, + 0.17099598050117493, + 0.04083382338285446, + 0.0015985078644007444, + 0.0011365534737706184, + 0.002851399825885892, + 0.0025997066404670477, + 0.0019112563459202647, + 0.0016718258848413825, + 0.0032407601829618216, + 0.0053010485135018826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.038101088255643845 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a full suit of polished steel plate armor with a closed-faced helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warhorse's coat black and sleek like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy crimson cape billow behind the knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the horse's iron-shod hooves kick up dry fallen leaves forming a spinning golden trail behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018186718225479126 + ], + "opposite_logprob_tokens": [ + -4.018186569213867 + ], + "logprob_reward": -0.018186718225479126, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018186718225479126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09118036180734634, + 0.02342977374792099, + 0.0755816325545311, + 0.09086223691701889, + 0.08352137356996536, + 0.015061222948133945, + 0.05687348544597626, + 0.02364019677042961, + 0.49414271116256714, + 0.04570700600743294 + ], + "credit_map": [ + 0.009213586337864399, + 0.01449182815849781, + 0.011180410161614418, + 0.01135635282844305, + 0.009603532031178474, + 0.011893329210579395, + 0.011628581210970879, + 0.011812744662165642, + 0.003392681945115328, + 0.001727192080579698, + 0.003132755635306239, + 0.004482594784349203, + 0.0021274269092828035, + 0.0027423901483416557, + 0.0034088101238012314, + 0.002415922237560153, + 0.00485650822520256, + 0.006031920202076435, + 0.01504366286098957, + 0.028832383453845978, + 0.0024468975607305765, + 0.003194267163053155, + 0.0036431089974939823, + 0.01153288222849369, + 0.01723373308777809, + 0.006698726676404476, + 0.02894868515431881, + 0.012237345799803734, + 0.005663592368364334, + 0.005363755859434605, + 0.006683812011033297, + 0.00803259015083313, + 0.009903866797685623, + 0.0113918362185359, + 0.013387328945100307, + 0.011194942519068718, + 0.005249685607850552, + 0.014295754954218864, + 0.0057762386277318, + 0.012321721762418747, + 0.00247014919295907, + 0.002981312572956085, + 0.001782162580639124, + 0.0015804305439814925, + 0.0021541924215853214, + 0.0012951868120580912, + 0.0016807717038318515, + 0.001117016770876944, + 0.011737813241779804, + 0.011070750653743744, + 0.006296668201684952, + 0.007490171119570732, + 0.005912626627832651, + 0.003998208325356245, + 0.004825579468160868, + 0.00554166454821825, + 0.002107435604557395, + 0.002710629254579544, + 0.003326406702399254, + 0.0034090811386704445, + 0.0032444384414702654, + 0.0034849520307034254, + 0.0031847753562033176, + 0.0021724787075072527, + 0.04778233915567398, + 0.1241321936249733, + 0.14787468314170837, + 0.025344721972942352, + 0.02219243533909321, + 0.027865171432495117, + 0.08032682538032532, + 0.018624354153871536, + 0.00905096996575594, + 0.005873038433492184, + 0.005708605982363224, + 0.003727520117536187, + 0.004893730394542217, + 0.006257257424294949, + 0.005096192471683025, + 0.005099690984934568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.018186718225479126 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, dynamic tracking shot follows a fully armored knight and his steed, the camera synchronizing them as they speed through a vast birch forest. The knight, clad in a full suit of meticulously polished, shimmering steel plate armor and a closed-faced helmet, his face obscured. A heavy crimson cape billows behind him at full speed, like a streak of flame. He leans forward, gripping the reins tightly, steadily guiding a powerful, muscular warhorse with a coat as sleek as obsidian. Golden afternoon sunlight penetrates the towering silvery-white birch canopy, casting a shifting, dappled light on the forest floor carpeted with golden and russet fallen leaves. With every powerful kick of the warhorse, its iron-shod hooves would suddenly lift the thick layer of dry fallen leaves on the ground backwards, forming a golden trail that spun, danced, and lingered behind them. The whole picture was full of a sense of speed and power. The rapid changes in light and shadow were intertwined with the flying fallen leaves, creating a tense and magnificent atmosphere of epic adventure.", + "seed": 3224, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_1/index_7729_seed_3224_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.23288221172988416, + "vlm_logprob": -0.23288221172988416, + "vlm_yes_no_margin": 2.55, + "vlm_token_credit": -0.23288221172988416, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a full suit of polished steel plate armor with a closed-faced helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warhorse's coat black and sleek like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy crimson cape billow behind the knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the horse's iron-shod hooves kick up dry fallen leaves forming a spinning golden trail behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.325928270816803 + ], + "opposite_logprob_tokens": [ + -1.7009282112121582 + ], + "logprob_reward": -0.325928270816803, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.325928270816803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2323807328939438, + 0.08072064816951752, + 0.04283379018306732, + 0.05066798999905586, + 0.06949884444475174, + 0.055730387568473816, + 0.08525347709655762, + 0.23327814042568207, + 0.06618702411651611, + 0.08344894647598267 + ], + "credit_map": [ + 0.04054497927427292, + 0.016036394983530045, + 0.026065245270729065, + 0.028298329561948776, + 0.011813140474259853, + 0.04002286121249199, + 0.026978369802236557, + 0.04262138903141022, + 0.012366785667836666, + 0.009616052731871605, + 0.013458473607897758, + 0.007903257384896278, + 0.011895454488694668, + 0.009159068576991558, + 0.009310822933912277, + 0.007010734640061855, + 0.005177357699722052, + 0.004807899706065655, + 0.005285941064357758, + 0.0052225482650101185, + 0.005131624173372984, + 0.00628944905474782, + 0.003947002813220024, + 0.006971967406570911, + 0.007801681291311979, + 0.004745129495859146, + 0.007907911203801632, + 0.005195631645619869, + 0.005747104063630104, + 0.007517350371927023, + 0.006560821086168289, + 0.005192360375076532, + 0.009772933088243008, + 0.004384815692901611, + 0.01292247325181961, + 0.007345721125602722, + 0.00855542253702879, + 0.009000846184790134, + 0.007215890102088451, + 0.010300738736987114, + 0.014302978292107582, + 0.00434627803042531, + 0.009512588381767273, + 0.005075430031865835, + 0.004047649912536144, + 0.007302434183657169, + 0.006197920069098473, + 0.004945112857967615, + 0.010431934148073196, + 0.009619852527976036, + 0.0084269093349576, + 0.005450282245874405, + 0.025425149127840996, + 0.014683334156870842, + 0.005648352671414614, + 0.005567655898630619, + 0.029419563710689545, + 0.03437640517950058, + 0.03785524517297745, + 0.029863515868782997, + 0.029282759875059128, + 0.022274689748883247, + 0.029093550518155098, + 0.021112404763698578, + 0.006638104096055031, + 0.006197802256792784, + 0.009619578719139099, + 0.011336921714246273, + 0.00903516635298729, + 0.006862055975943804, + 0.008838971145451069, + 0.007658420596271753, + 0.006509888917207718, + 0.018634585663676262, + 0.01685721054673195, + 0.009140356443822384, + 0.006477294489741325, + 0.010333976708352566, + 0.007816491648554802, + 0.00767914392054081 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.325928270816803 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a full suit of polished steel plate armor with a closed-faced helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warhorse's coat black and sleek like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy crimson cape billow behind the knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the horse's iron-shod hooves kick up dry fallen leaves forming a spinning golden trail behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578458786010742 + ], + "opposite_logprob_tokens": [ + -0.6328458786010742 + ], + "logprob_reward": -0.7578458786010742, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578458786010742, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3033594489097595, + 0.10998804867267609, + 0.07274612784385681, + 0.06714209169149399, + 0.09043663740158081, + 0.0474577322602272, + 0.028385456651449203, + 0.06563693284988403, + 0.08208417147397995, + 0.13276340067386627 + ], + "credit_map": [ + 0.04803191125392914, + 0.02250722609460354, + 0.0675913393497467, + 0.03755160793662071, + 0.016863204538822174, + 0.056819479912519455, + 0.028040222823619843, + 0.025954434648156166, + 0.02556166984140873, + 0.01032477617263794, + 0.015018964186310768, + 0.012869362719357014, + 0.010386116802692413, + 0.014641732908785343, + 0.010834219865500927, + 0.01035118754953146, + 0.008669789880514145, + 0.006625724490731955, + 0.008425724692642689, + 0.009972150437533855, + 0.00794128980487585, + 0.011575731448829174, + 0.007947160862386227, + 0.011588551104068756, + 0.008573860861361027, + 0.008978604339063168, + 0.006163223646581173, + 0.006834674160927534, + 0.012898417189717293, + 0.00910420436412096, + 0.008745755068957806, + 0.005843345075845718, + 0.010068568401038647, + 0.006767056882381439, + 0.01599946804344654, + 0.010946767404675484, + 0.012732602655887604, + 0.01542807836085558, + 0.009718513116240501, + 0.008775577880442142, + 0.008328850381076336, + 0.004983979277312756, + 0.007703318260610104, + 0.004105671774595976, + 0.005781258922070265, + 0.005456367041915655, + 0.005253321025520563, + 0.0058449613861739635, + 0.004657897166907787, + 0.002670092275366187, + 0.004355498123914003, + 0.0021151271648705006, + 0.004158345051109791, + 0.005093391984701157, + 0.0025515968445688486, + 0.0027835052460432053, + 0.009201363660395145, + 0.006924195680767298, + 0.01176097895950079, + 0.008590638637542725, + 0.00940344762057066, + 0.007434851489961147, + 0.005733324680477381, + 0.006588130723685026, + 0.008948036469519138, + 0.013537425547838211, + 0.010499113239347935, + 0.00896355789154768, + 0.013397211208939552, + 0.01345091126859188, + 0.006145576946437359, + 0.007142337039113045, + 0.007448848802596331, + 0.028949234634637833, + 0.04411464184522629, + 0.012987812049686909, + 0.011396260000765324, + 0.011009241454303265, + 0.0071099563501775265, + 0.00974738597869873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.7578458786010742 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a full suit of polished steel plate armor with a closed-faced helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warhorse's coat black and sleek like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy crimson cape billow behind the knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the horse's iron-shod hooves kick up dry fallen leaves forming a spinning golden trail behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04870246723294258 + ], + "opposite_logprob_tokens": [ + -3.0487024784088135 + ], + "logprob_reward": -0.04870246723294258, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04870246723294258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20195098221302032, + 0.16661188006401062, + 0.04644602909684181, + 0.0746774673461914, + 0.13740123808383942, + 0.07786475867033005, + 0.07977620512247086, + 0.035896461457014084, + 0.0817812830209732, + 0.09759365022182465 + ], + "credit_map": [ + 0.03111785463988781, + 0.016558654606342316, + 0.04030180349946022, + 0.02749720588326454, + 0.013917309232056141, + 0.03490191698074341, + 0.020546671003103256, + 0.017109565436840057, + 0.021592993289232254, + 0.024904126301407814, + 0.013859141618013382, + 0.016002804040908813, + 0.04431140795350075, + 0.01702483370900154, + 0.016475027427077293, + 0.012441552244126797, + 0.005397422704845667, + 0.00556755717843771, + 0.005838444922119379, + 0.006940709892660379, + 0.005529697053134441, + 0.007253680378198624, + 0.004126997664570808, + 0.005791514180600643, + 0.010037641040980816, + 0.009764197282493114, + 0.007525273133069277, + 0.006785635836422443, + 0.014450383372604847, + 0.009448203258216381, + 0.010268235579133034, + 0.006397892255336046, + 0.01796753890812397, + 0.016085466369986534, + 0.01458839513361454, + 0.01001473143696785, + 0.03409774973988533, + 0.020227815955877304, + 0.013527155853807926, + 0.010892383754253387, + 0.013615400530397892, + 0.008581376634538174, + 0.012091631069779396, + 0.00796778779476881, + 0.011473417282104492, + 0.008916039019823074, + 0.00709730526432395, + 0.008121799677610397, + 0.013124192133545876, + 0.006368260830640793, + 0.008066603913903236, + 0.006637436803430319, + 0.0160562451928854, + 0.017974097281694412, + 0.005640193819999695, + 0.005909172352403402, + 0.004651292227208614, + 0.005351573694497347, + 0.004474975634366274, + 0.004683356732130051, + 0.004090084228664637, + 0.004793500993400812, + 0.00455595413222909, + 0.003295724280178547, + 0.009291245602071285, + 0.015318102203309536, + 0.009994424879550934, + 0.007758864667266607, + 0.011688610538840294, + 0.014868676662445068, + 0.006713140290230513, + 0.0061482153832912445, + 0.0070123616605997086, + 0.02061057835817337, + 0.02467931993305683, + 0.012892403639853, + 0.009348368272185326, + 0.009237456135451794, + 0.0060338969342410564, + 0.007779271341860294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.04870246723294258 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a full suit of polished steel plate armor with a closed-faced helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warhorse's coat black and sleek like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy crimson cape billow behind the knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the horse's iron-shod hooves kick up dry fallen leaves forming a spinning golden trail behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008642058819532394 + ], + "opposite_logprob_tokens": [ + -4.758642196655273 + ], + "logprob_reward": -0.008642058819532394, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008642058819532394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06466695666313171, + 0.07556646317243576, + 0.1270553320646286, + 0.3451751470565796, + 0.24413348734378815, + 0.013601901941001415, + 0.019446274265646935, + 0.04281776770949364, + 0.019833382219076157, + 0.047703247517347336 + ], + "credit_map": [ + 0.007878044620156288, + 0.0057653808034956455, + 0.011654012836515903, + 0.010336481966078281, + 0.005552052054554224, + 0.010145612061023712, + 0.007390623912215233, + 0.005944745149463415, + 0.009855065494775772, + 0.01032475009560585, + 0.007052806671708822, + 0.011967351660132408, + 0.012744604609906673, + 0.006539660505950451, + 0.00952576007694006, + 0.007556462194770575, + 0.013109763152897358, + 0.014213012531399727, + 0.013580022379755974, + 0.013691622763872147, + 0.025620127096772194, + 0.016391491517424583, + 0.011258028447628021, + 0.019191252067685127, + 0.0425778292119503, + 0.045255884528160095, + 0.0326848141849041, + 0.033825602382421494, + 0.07342442870140076, + 0.041919756680727005, + 0.04840866103768349, + 0.027078183367848396, + 0.023858388885855675, + 0.018313545733690262, + 0.03373988717794418, + 0.018865792080760002, + 0.06452440470457077, + 0.0523342490196228, + 0.016653699800372124, + 0.015843534842133522, + 0.0017651363741606474, + 0.0013447016244754195, + 0.0020529909525066614, + 0.001655298168770969, + 0.0013051044661551714, + 0.001472323201596737, + 0.002291229786351323, + 0.0017151178326457739, + 0.002877185819670558, + 0.0030880384147167206, + 0.0019518900662660599, + 0.0018388782627880573, + 0.0029040256049484015, + 0.0026118531823158264, + 0.0017492008628323674, + 0.002425203798338771, + 0.005619177594780922, + 0.0054893940687179565, + 0.007292690221220255, + 0.0052729384042322636, + 0.005617334507405758, + 0.004254672676324844, + 0.005089023616164923, + 0.004182534292340279, + 0.0019739759154617786, + 0.0027238288894295692, + 0.002494226675480604, + 0.002191021805629134, + 0.0038220961578190327, + 0.0027677095495164394, + 0.0021752314642071724, + 0.0016852907137945294, + 0.0037161726504564285, + 0.005268617067486048, + 0.014664022251963615, + 0.005461504682898521, + 0.005180399399250746, + 0.003597759176045656, + 0.0030215333681553602, + 0.0067932396195828915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 2 + ], + "selected_reward": -0.008642058819532394 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a full suit of polished steel plate armor with a closed-faced helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warhorse's coat black and sleek like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy crimson cape billow behind the knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the horse's iron-shod hooves kick up dry fallen leaves forming a spinning golden trail behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023292383179068565 + ], + "opposite_logprob_tokens": [ + -3.773292303085327 + ], + "logprob_reward": -0.023292383179068565, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023292383179068565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21401606500148773, + 0.08424783498048782, + 0.029918242245912552, + 0.07556936889886856, + 0.026972180232405663, + 0.016926830634474754, + 0.03226775303483009, + 0.13657955825328827, + 0.03874260559678078, + 0.34475958347320557 + ], + "credit_map": [ + 0.03847803175449371, + 0.013916755095124245, + 0.026465315371751785, + 0.021327409893274307, + 0.014624718576669693, + 0.02923843450844288, + 0.020568490028381348, + 0.049396924674510956, + 0.012415703386068344, + 0.009862864390015602, + 0.011305518448352814, + 0.01018235832452774, + 0.01467639859765768, + 0.009094067849218845, + 0.008933833800256252, + 0.0077770850621163845, + 0.0030088485218584538, + 0.0035534584894776344, + 0.003345862263813615, + 0.005121112335473299, + 0.0027891036588698626, + 0.0037224770057946444, + 0.0029055874329060316, + 0.005471791606396437, + 0.010597209446132183, + 0.009877214208245277, + 0.007600278593599796, + 0.006018802057951689, + 0.018133658915758133, + 0.009817337617278099, + 0.007470674812793732, + 0.006054196041077375, + 0.004934611730277538, + 0.002527039498090744, + 0.0038719968870282173, + 0.0024903269950300455, + 0.003931510262191296, + 0.0034964033402502537, + 0.0024346420541405678, + 0.003285647602751851, + 0.002774012042209506, + 0.001413715654052794, + 0.002650920767337084, + 0.001600040472112596, + 0.0022810015361756086, + 0.0019585525151342154, + 0.002283989917486906, + 0.001964598661288619, + 0.004667668603360653, + 0.003182979766279459, + 0.0032327270600944757, + 0.0021323112305253744, + 0.007862191647291183, + 0.006658153608441353, + 0.0024173588026314974, + 0.0021143625490367413, + 0.01762942038476467, + 0.018897589296102524, + 0.02232610061764717, + 0.017185142263770103, + 0.017763396725058556, + 0.01384221576154232, + 0.016739170998334885, + 0.012196524068713188, + 0.004227925557643175, + 0.005319467280060053, + 0.003593316301703453, + 0.005011347122490406, + 0.00492350896820426, + 0.005770264193415642, + 0.0057433187030255795, + 0.00415345374494791, + 0.02294786088168621, + 0.060945846140384674, + 0.13452477753162384, + 0.027339160442352295, + 0.01438168715685606, + 0.015475744381546974, + 0.015376792289316654, + 0.05376768112182617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.023292383179068565 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, dynamic tracking shot follows a fully armored knight and his steed, the camera synchronizing them as they speed through a vast birch forest. The knight, clad in a full suit of meticulously polished, shimmering steel plate armor and a closed-faced helmet, his face obscured. A heavy crimson cape billows behind him at full speed, like a streak of flame. He leans forward, gripping the reins tightly, steadily guiding a powerful, muscular warhorse with a coat as sleek as obsidian. Golden afternoon sunlight penetrates the towering silvery-white birch canopy, casting a shifting, dappled light on the forest floor carpeted with golden and russet fallen leaves. With every powerful kick of the warhorse, its iron-shod hooves would suddenly lift the thick layer of dry fallen leaves on the ground backwards, forming a golden trail that spun, danced, and lingered behind them. The whole picture was full of a sense of speed and power. The rapid changes in light and shadow were intertwined with the flying fallen leaves, creating a tense and magnificent atmosphere of epic adventure.", + "seed": 3225, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_1/index_7729_seed_3225_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.3615400943905115, + "vlm_logprob": -0.3615400943905115, + "vlm_yes_no_margin": 1.6999999761581421, + "vlm_token_credit": -0.3615400943905115, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a full suit of polished steel plate armor with a closed-faced helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warhorse's coat black and sleek like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy crimson cape billow behind the knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the horse's iron-shod hooves kick up dry fallen leaves forming a spinning golden trail behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6313081979751587 + ], + "opposite_logprob_tokens": [ + -1.0063081979751587 + ], + "logprob_reward": -0.6313081979751587, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.6313081979751587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17751753330230713, + 0.11658624559640884, + 0.03447896987199783, + 0.06592758744955063, + 0.23899303376674652, + 0.06488053500652313, + 0.05668955296278, + 0.1605217307806015, + 0.026929710060358047, + 0.05747513100504875 + ], + "credit_map": [ + 0.016956208273768425, + 0.02864275686442852, + 0.03762880712747574, + 0.04544229805469513, + 0.0104362266138196, + 0.013308383524417877, + 0.011681666597723961, + 0.013421178795397282, + 0.018663765862584114, + 0.024796921759843826, + 0.018699923530220985, + 0.013782485388219357, + 0.014873539097607136, + 0.01224865484982729, + 0.00814132858067751, + 0.005379626527428627, + 0.004616109654307365, + 0.005860723089426756, + 0.005719471722841263, + 0.006267651449888945, + 0.0027109526563435793, + 0.0036314697936177254, + 0.0032101983670145273, + 0.0024623877834528685, + 0.011174827814102173, + 0.018366625532507896, + 0.008205450139939785, + 0.007717771455645561, + 0.007849513553082943, + 0.0047661857679486275, + 0.004740595817565918, + 0.0031066127121448517, + 0.05007083714008331, + 0.02492263913154602, + 0.06274739652872086, + 0.026287874206900597, + 0.01601494662463665, + 0.015337776392698288, + 0.022881127893924713, + 0.02073044143617153, + 0.009380330331623554, + 0.0054033189080655575, + 0.011607056483626366, + 0.007089699152857065, + 0.005508691072463989, + 0.006053504999727011, + 0.00894815195351839, + 0.010889774188399315, + 0.01039128564298153, + 0.006291206926107407, + 0.00791855901479721, + 0.006505254190415144, + 0.008580752648413181, + 0.007364969234913588, + 0.005339771043509245, + 0.004297748673707247, + 0.04805063456296921, + 0.005183618050068617, + 0.0684107318520546, + 0.016729800030589104, + 0.004109945148229599, + 0.0046143182553350925, + 0.0072385952807962894, + 0.006184074562042952, + 0.00421703327447176, + 0.0022837109863758087, + 0.0046523683704435825, + 0.002107168547809124, + 0.004586612805724144, + 0.005025928840041161, + 0.0021982628386467695, + 0.001858622650615871, + 0.007187427021563053, + 0.0027298221830278635, + 0.018187901005148888, + 0.004812099039554596, + 0.002334412420168519, + 0.004558457527309656, + 0.0022861212491989136, + 0.015378882177174091 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.6313081979751587 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a full suit of polished steel plate armor with a closed-faced helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warhorse's coat black and sleek like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy crimson cape billow behind the knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the horse's iron-shod hooves kick up dry fallen leaves forming a spinning golden trail behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9743051528930664 + ], + "opposite_logprob_tokens": [ + -0.4743051528930664 + ], + "logprob_reward": -0.9743051528930664, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9743051528930664, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.13256530463695526, + 0.07582029700279236, + 0.07381322979927063, + 0.037441279739141464, + 0.2154252529144287, + 0.14668500423431396, + 0.03462410345673561, + 0.06832180172204971, + 0.05697024613618851, + 0.15833349525928497 + ], + "credit_map": [ + 0.027801405638456345, + 0.012319725006818771, + 0.032221175730228424, + 0.012508782558143139, + 0.010874171741306782, + 0.011521969921886921, + 0.010216687805950642, + 0.015101398341357708, + 0.01035820972174406, + 0.00874034222215414, + 0.009864146821200848, + 0.010285735130310059, + 0.00815557036548853, + 0.01500315684825182, + 0.005732934456318617, + 0.007680202834308147, + 0.009035687893629074, + 0.006449232343584299, + 0.012047339230775833, + 0.007801642641425133, + 0.014040743932127953, + 0.008243235759437084, + 0.00963917002081871, + 0.006556183565407991, + 0.004716201219707727, + 0.0039465115405619144, + 0.0065423548221588135, + 0.003977385815232992, + 0.0037487628869712353, + 0.007925827987492085, + 0.0030572181567549706, + 0.003527018940076232, + 0.03368890658020973, + 0.016545942053198814, + 0.04847562685608864, + 0.01479988545179367, + 0.01842576265335083, + 0.04415673390030861, + 0.02157789282500744, + 0.017754504457116127, + 0.019752733409404755, + 0.012651746161282063, + 0.02180410362780094, + 0.02229914255440235, + 0.008556718938052654, + 0.021121695637702942, + 0.011656040325760841, + 0.028842821717262268, + 0.0069794077426195145, + 0.0033833326306194067, + 0.004317353945225477, + 0.004291253164410591, + 0.0032624106388539076, + 0.005073575768619776, + 0.0032910509034991264, + 0.004025717731565237, + 0.009341993369162083, + 0.00611833157017827, + 0.012191828340291977, + 0.013586809858679771, + 0.0034999919589608908, + 0.010671345517039299, + 0.005889162886887789, + 0.007022340781986713, + 0.009756934829056263, + 0.006477970629930496, + 0.012346821837127209, + 0.004246095195412636, + 0.003482166910544038, + 0.012974292039871216, + 0.0036971629597246647, + 0.003988799639046192, + 0.02199937216937542, + 0.005219792015850544, + 0.05270532891154289, + 0.010706493631005287, + 0.007069157902151346, + 0.014714994467794895, + 0.00540679506957531, + 0.04051154851913452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 5 + ], + "selected_reward": -0.9743051528930664 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a full suit of polished steel plate armor with a closed-faced helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warhorse's coat black and sleek like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy crimson cape billow behind the knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the horse's iron-shod hooves kick up dry fallen leaves forming a spinning golden trail behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04869372397661209 + ], + "opposite_logprob_tokens": [ + -3.0486936569213867 + ], + "logprob_reward": -0.04869372397661209, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04869372397661209, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38615134358406067, + 0.04200540855526924, + 0.09356164932250977, + 0.10363892465829849, + 0.07123193144798279, + 0.06161852180957794, + 0.03731097653508186, + 0.06344336271286011, + 0.07893376052379608, + 0.062104158103466034 + ], + "credit_map": [ + 0.070887990295887, + 0.023963123559951782, + 0.140919029712677, + 0.08549897372722626, + 0.018572833389043808, + 0.0170276015996933, + 0.016497917473316193, + 0.012783823534846306, + 0.005782321095466614, + 0.008500885218381882, + 0.006318572908639908, + 0.008060959167778492, + 0.002932298695668578, + 0.003061306895688176, + 0.003808863926678896, + 0.0035401966888457537, + 0.013710473664104939, + 0.006429752334952354, + 0.01845761016011238, + 0.010911841876804829, + 0.013024037703871727, + 0.007639730349183083, + 0.014233274385333061, + 0.009154926054179668, + 0.017762836068868637, + 0.027751753106713295, + 0.01410274114459753, + 0.013218361884355545, + 0.009985019452869892, + 0.009521427564322948, + 0.006905326619744301, + 0.004391442518681288, + 0.009269445203244686, + 0.007408349774777889, + 0.0073683978989720345, + 0.009012976661324501, + 0.01138685829937458, + 0.014572001062333584, + 0.0066179935820400715, + 0.005595908500254154, + 0.011192511767148972, + 0.0056876372545957565, + 0.010168151929974556, + 0.0072839404456317425, + 0.005210074596107006, + 0.01015862263739109, + 0.004786093253642321, + 0.0071314861997962, + 0.00604094285517931, + 0.004889949224889278, + 0.00507974810898304, + 0.0039303903467953205, + 0.005397627130150795, + 0.004943742882460356, + 0.0038106341380625963, + 0.0032179388217628, + 0.008327087387442589, + 0.01025985088199377, + 0.008751618675887585, + 0.009907334111630917, + 0.006497657857835293, + 0.0058824969455599785, + 0.007038666866719723, + 0.006778646260499954, + 0.012808708474040031, + 0.008481589145958424, + 0.015862025320529938, + 0.0050361426547169685, + 0.008374616503715515, + 0.01794644631445408, + 0.005644196644425392, + 0.004780035465955734, + 0.009165041148662567, + 0.0037108338437974453, + 0.01691080443561077, + 0.004892117343842983, + 0.003413202241063118, + 0.007283193990588188, + 0.0025347343180328608, + 0.014194225892424583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.04869372397661209 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a full suit of polished steel plate armor with a closed-faced helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warhorse's coat black and sleek like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy crimson cape billow behind the knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the horse's iron-shod hooves kick up dry fallen leaves forming a spinning golden trail behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704406678676605 + ], + "opposite_logprob_tokens": [ + -2.1270439624786377 + ], + "logprob_reward": -0.12704406678676605, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12704406678676605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3241858184337616, + 0.15021908283233643, + 0.031563885509967804, + 0.19638065993785858, + 0.15489450097084045, + 0.028221849352121353, + 0.014001565054059029, + 0.05402926355600357, + 0.01195533573627472, + 0.034548062831163406 + ], + "credit_map": [ + 0.10179479420185089, + 0.026366161182522774, + 0.11847424507141113, + 0.025735367089509964, + 0.011365187354385853, + 0.01396683044731617, + 0.016647273674607277, + 0.009835938923060894, + 0.023956548422574997, + 0.013555631972849369, + 0.014712725766003132, + 0.021808791905641556, + 0.017331795766949654, + 0.03365999087691307, + 0.010635132901370525, + 0.014558454975485802, + 0.005907195154577494, + 0.0021558268927037716, + 0.009604562073946, + 0.0036339738871902227, + 0.0027804607525467873, + 0.003069028491154313, + 0.0022013408597558737, + 0.00221149530261755, + 0.023012464866042137, + 0.02575112320482731, + 0.026249930262565613, + 0.014528762549161911, + 0.018792571499943733, + 0.05134058743715286, + 0.017322169616818428, + 0.019383031874895096, + 0.023713326081633568, + 0.010600226931273937, + 0.035621412098407745, + 0.00716365734115243, + 0.011116569861769676, + 0.036787036806344986, + 0.018572770059108734, + 0.011319483630359173, + 0.0041877334006130695, + 0.0025585643015801907, + 0.00415072264149785, + 0.0042049759067595005, + 0.002286805072799325, + 0.004370520822703838, + 0.002079447265714407, + 0.004383076913654804, + 0.003330477513372898, + 0.0014974644873291254, + 0.0020723044872283936, + 0.0009550842805765569, + 0.0010611813049763441, + 0.0021913752425462008, + 0.0014273766428232193, + 0.0014663004549220204, + 0.015691407024860382, + 0.0033159314189106226, + 0.01771332509815693, + 0.006297987420111895, + 0.0018822553101927042, + 0.0031036913860589266, + 0.0032213509548455477, + 0.002803314011543989, + 0.002012383658438921, + 0.0011966374004259706, + 0.0026151477359235287, + 0.0008630077354609966, + 0.0010600961977615952, + 0.002595582976937294, + 0.0008776998147368431, + 0.0007347797509282827, + 0.0047671059146523476, + 0.003254515118896961, + 0.00609493488445878, + 0.003250371664762497, + 0.005317444447427988, + 0.0038916473276913166, + 0.003406610805541277, + 0.0045654308050870895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.12704406678676605 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a full suit of polished steel plate armor with a closed-faced helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warhorse's coat black and sleek like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy crimson cape billow behind the knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the horse's iron-shod hooves kick up dry fallen leaves forming a spinning golden trail behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026349330320954323 + ], + "opposite_logprob_tokens": [ + -3.6513493061065674 + ], + "logprob_reward": -0.026349330320954323, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026349330320954323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6214602589607239, + 0.03522985801100731, + 0.022027669474482536, + 0.071824811398983, + 0.051357753574848175, + 0.05438201129436493, + 0.013546711765229702, + 0.03208671510219574, + 0.041318248957395554, + 0.05676596984267235 + ], + "credit_map": [ + 0.15819929540157318, + 0.039214447140693665, + 0.2399023324251175, + 0.10332594811916351, + 0.01242564432322979, + 0.029419435188174248, + 0.02494824305176735, + 0.014024863950908184, + 0.00465779472142458, + 0.005657919216901064, + 0.005464046727865934, + 0.005377247463911772, + 0.0029267205391079187, + 0.005081275478005409, + 0.002971351146697998, + 0.0030935024842619896, + 0.00372953899204731, + 0.0038477845955640078, + 0.00306069478392601, + 0.003352069528773427, + 0.001782535924576223, + 0.0019401866011321545, + 0.0022480878978967667, + 0.002066772198304534, + 0.01163728628307581, + 0.017911816015839577, + 0.00971743743866682, + 0.009562146849930286, + 0.007299699354916811, + 0.008652076125144958, + 0.004631455522030592, + 0.002412894507870078, + 0.01270874124020338, + 0.004035748541355133, + 0.015317016281187534, + 0.0037044277414679527, + 0.0036619447637349367, + 0.004565023351460695, + 0.0036073217634111643, + 0.003757530590519309, + 0.011477249674499035, + 0.0045088366605341434, + 0.008440830744802952, + 0.005388318095356226, + 0.0030726734548807144, + 0.008692000061273575, + 0.0037727465387433767, + 0.00902935303747654, + 0.002273150719702244, + 0.0014056832296773791, + 0.0019059974001720548, + 0.0017372346483170986, + 0.0018982102628797293, + 0.0017773532308638096, + 0.0013129854341968894, + 0.0012360966065898538, + 0.0046791755594313145, + 0.005334504414349794, + 0.006384535226970911, + 0.0040878355503082275, + 0.002370469970628619, + 0.003594261361286044, + 0.002548488089814782, + 0.003087444696575403, + 0.007240867242217064, + 0.003469740739092231, + 0.009519916027784348, + 0.0030700804200023413, + 0.003204895416274667, + 0.0100791584700346, + 0.0022763910237699747, + 0.002457199851050973, + 0.008202980272471905, + 0.004504683427512646, + 0.014434151351451874, + 0.005420411005616188, + 0.0034288321621716022, + 0.005481199361383915, + 0.003294078167527914, + 0.011999632231891155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.026349330320954323 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_10/index_tensor([8854, 8854])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_10/index_tensor([8854, 8854])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0cb7436ba9347bceb3c580aaf8a060de0c8b031f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_10/index_tensor([8854, 8854])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The scene begins in absolute darkness, then a static media close-up is violently illuminated by a massive, silent lightning bolt. In an instant, a brilliant, icy white light streaks across the void: a cluster of hard, perfectly spherical black fungi clinging tenaciously to a rotting, moss-covered log. Each sphere's carbonaceous surface, covered in a series of fine cracks, glitters with slick rainwater, its convex shapes catching the harsh light with sharp, mirrored points. Just as suddenly, the world plunges back into deep darkness. From this void, the camera begins a slow, deliberate downward tilt, transitioning into extreme close-up, carefully following a heavy, spherical raindrop as it traces a sticky path along the curved surface of one of the fungi. The transparent drop distorts the matte black texture beneath, its journey a mesmerizing micro-event following the raging, silent aftermath of the storm. The entire visual is captured with a stark, high-contrast digital look, characterized by piercing whites and deep, engrossing blacks, creating an atmosphere of tension, mystery, and intense concentration.", + "seed": 28608, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_10/index_8854_seed_28608_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07106259195134043, + "vlm_logprob": -0.07106259195134043, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.07106259195134043, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin in absolute darkness before being illuminated by a massive lightning bolt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hard, perfectly spherical black fungi visible clinging to a rotting, moss-covered log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fungi spheres exhibit fine cracks and a carbonaceous surface covered in slick rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow downward tilt transitioning into an extreme close-up of a heavy spherical raindrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a stark, high-contrast look with piercing whites and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28764909505844116 + ], + "opposite_logprob_tokens": [ + -1.412649154663086 + ], + "logprob_reward": -0.28764909505844116, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28764909505844116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2576717138290405, + 0.12452083081007004, + 0.029747162014245987, + 0.14712707698345184, + 0.057304635643959045, + 0.03464246913790703, + 0.032701801508665085, + 0.0657188817858696, + 0.15333078801631927, + 0.09723465144634247 + ], + "credit_map": [ + 0.03148419037461281, + 0.03107083961367607, + 0.03620331734418869, + 0.03010617382824421, + 0.047144971787929535, + 0.039805930107831955, + 0.020258238539099693, + 0.021598076447844505, + 0.020482713356614113, + 0.0075624240562319756, + 0.026120591908693314, + 0.016177261248230934, + 0.009218506515026093, + 0.02054928056895733, + 0.011560836806893349, + 0.012849211692810059, + 0.0031960925552994013, + 0.0037859282456338406, + 0.004397643264383078, + 0.0023108343593776226, + 0.005998624954372644, + 0.002564811846241355, + 0.004306003451347351, + 0.003187224268913269, + 0.02156699262559414, + 0.02170039340853691, + 0.016663808375597, + 0.016990235075354576, + 0.028764892369508743, + 0.017846504226326942, + 0.012576288543641567, + 0.011017960496246815, + 0.009305416606366634, + 0.008782249875366688, + 0.007319223135709763, + 0.006427397020161152, + 0.0041227941401302814, + 0.005985192954540253, + 0.008781750686466694, + 0.006580608896911144, + 0.005796424578875303, + 0.0049653807654976845, + 0.003963470458984375, + 0.0029715532436966896, + 0.0027362352702766657, + 0.003409142605960369, + 0.0047529092989861965, + 0.006047353148460388, + 0.003988098818808794, + 0.0034069290850311518, + 0.005974799394607544, + 0.002459528623148799, + 0.0054762582294642925, + 0.004272576421499252, + 0.0038317665457725525, + 0.0032918425276875496, + 0.0061323377303779125, + 0.00938131008297205, + 0.00782101321965456, + 0.005781589075922966, + 0.009034131653606892, + 0.009225111454725266, + 0.009950649924576283, + 0.008392742834985256, + 0.013125271536409855, + 0.016502920538187027, + 0.016922859475016594, + 0.03032003529369831, + 0.014545763842761517, + 0.01733432151377201, + 0.024507388472557068, + 0.020072221755981445, + 0.010358826257288456, + 0.008024054579436779, + 0.013892346061766148, + 0.02035835012793541, + 0.009007512591779232, + 0.011109571903944016, + 0.012752077542245388, + 0.011731908656656742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.28764909505844116 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin in absolute darkness before being illuminated by a massive lightning bolt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hard, perfectly spherical black fungi visible clinging to a rotting, moss-covered log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fungi spheres exhibit fine cracks and a carbonaceous surface covered in slick rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow downward tilt transitioning into an extreme close-up of a heavy spherical raindrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a stark, high-contrast look with piercing whites and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007642081007361412 + ], + "opposite_logprob_tokens": [ + -4.8826422691345215 + ], + "logprob_reward": -0.007642081007361412, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007642081007361412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13089965283870697, + 0.17809604108333588, + 0.119657963514328, + 0.049700263887643814, + 0.023907769471406937, + 0.01811801642179489, + 0.05360165238380432, + 0.08288116753101349, + 0.25407540798187256, + 0.08906206488609314 + ], + "credit_map": [ + 0.015941614285111427, + 0.018601013347506523, + 0.02064700238406658, + 0.017867786809802055, + 0.017848558723926544, + 0.022149253636598587, + 0.007874684408307076, + 0.009969743900001049, + 0.02909265272319317, + 0.01352012064307928, + 0.03718794509768486, + 0.017187615856528282, + 0.030553994700312614, + 0.01897694543004036, + 0.020624075084924698, + 0.010952706448733807, + 0.01167883351445198, + 0.018018072471022606, + 0.01813734509050846, + 0.008662845008075237, + 0.0260505061596632, + 0.011412277817726135, + 0.015591941773891449, + 0.010106141678988934, + 0.00844148825854063, + 0.0056228949688375, + 0.01262287050485611, + 0.005963402334600687, + 0.0046781678684055805, + 0.0037370899226516485, + 0.0052392734214663506, + 0.003395077306777239, + 0.004050551448017359, + 0.0030872474890202284, + 0.002149771898984909, + 0.002503179246559739, + 0.003018561750650406, + 0.0018893585074692965, + 0.0045805578120052814, + 0.0026285413186997175, + 0.003043701173737645, + 0.00222745630890131, + 0.0025409150402992964, + 0.0017961125122383237, + 0.001531822606921196, + 0.0018234531162306666, + 0.002075311727821827, + 0.0030792439356446266, + 0.006687191314995289, + 0.0039716544561088085, + 0.007359814364463091, + 0.006270973477512598, + 0.006673900876194239, + 0.00589338131248951, + 0.009606421925127506, + 0.007138312328606844, + 0.007309633307158947, + 0.010378596372902393, + 0.011346062645316124, + 0.008927296847105026, + 0.012409484945237637, + 0.009962854906916618, + 0.012698742561042309, + 0.009848492220044136, + 0.020718496292829514, + 0.023884054273366928, + 0.03215114399790764, + 0.04207979887723923, + 0.025142882019281387, + 0.02903834730386734, + 0.0462946742773056, + 0.03476601839065552, + 0.007310993038117886, + 0.005332023371011019, + 0.015161881223320961, + 0.019066793844103813, + 0.009196928702294827, + 0.008755712769925594, + 0.012036285363137722, + 0.012201450765132904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.007642081007361412 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin in absolute darkness before being illuminated by a massive lightning bolt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hard, perfectly spherical black fungi visible clinging to a rotting, moss-covered log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fungi spheres exhibit fine cracks and a carbonaceous surface covered in slick rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow downward tilt transitioning into an extreme close-up of a heavy spherical raindrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a stark, high-contrast look with piercing whites and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004642304033041 + ], + "opposite_logprob_tokens": [ + -5.379642486572266 + ], + "logprob_reward": -0.004642304033041, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004642304033041, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17902019619941711, + 0.06972146034240723, + 0.10087268054485321, + 0.07129555940628052, + 0.08079813420772552, + 0.03784080594778061, + 0.05814360827207565, + 0.0936465859413147, + 0.18404041230678558, + 0.1246204674243927 + ], + "credit_map": [ + 0.01900547370314598, + 0.03487011045217514, + 0.02369793690741062, + 0.02334296703338623, + 0.023643339052796364, + 0.02735142782330513, + 0.011935548856854439, + 0.01517340075224638, + 0.009656445123255253, + 0.004880990367382765, + 0.010537242516875267, + 0.007623138837516308, + 0.011305403895676136, + 0.008680415339767933, + 0.010348588228225708, + 0.006689239293336868, + 0.011316493153572083, + 0.012069792486727238, + 0.014095212332904339, + 0.007826754823327065, + 0.02491135708987713, + 0.009320614859461784, + 0.013534300960600376, + 0.00779816135764122, + 0.010270839557051659, + 0.008397149853408337, + 0.020076969638466835, + 0.00547060975804925, + 0.009698896668851376, + 0.005809144116938114, + 0.005199493374675512, + 0.006372466683387756, + 0.01293457206338644, + 0.01739226095378399, + 0.0072843534871935844, + 0.011225344613194466, + 0.008675264194607735, + 0.0062314365059137344, + 0.00829591415822506, + 0.00875899102538824, + 0.0053350054658949375, + 0.004933836869895458, + 0.004702983424067497, + 0.005640109069645405, + 0.0038553669583052397, + 0.004316616337746382, + 0.004834696650505066, + 0.004222196526825428, + 0.006999444682151079, + 0.0034553962759673595, + 0.00671928096562624, + 0.007397538051009178, + 0.0064225103706121445, + 0.007045677863061428, + 0.01171259768307209, + 0.008391167037189007, + 0.006613526958972216, + 0.014922289177775383, + 0.012785293161869049, + 0.009851588867604733, + 0.014120569452643394, + 0.01265969779342413, + 0.012613312341272831, + 0.010080313310027122, + 0.014104358851909637, + 0.017185721546411514, + 0.017657186836004257, + 0.031184891238808632, + 0.026378808543086052, + 0.02629297785460949, + 0.027540985494852066, + 0.02369549497961998, + 0.009868094697594643, + 0.0066870590671896935, + 0.01919710449874401, + 0.03058946691453457, + 0.012130508199334145, + 0.009870579466223717, + 0.018687205389142036, + 0.017590466886758804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.004642304033041 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin in absolute darkness before being illuminated by a massive lightning bolt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hard, perfectly spherical black fungi visible clinging to a rotting, moss-covered log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fungi spheres exhibit fine cracks and a carbonaceous surface covered in slick rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow downward tilt transitioning into an extreme close-up of a heavy spherical raindrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a stark, high-contrast look with piercing whites and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864376410841942 + ], + "opposite_logprob_tokens": [ + -3.0486438274383545 + ], + "logprob_reward": -0.04864376410841942, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864376410841942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19585295021533966, + 0.07740812748670578, + 0.10477835685014725, + 0.04007474333047867, + 0.04110479727387428, + 0.04321196302771568, + 0.04972962290048599, + 0.13703055679798126, + 0.2232218086719513, + 0.08758708089590073 + ], + "credit_map": [ + 0.024471424520015717, + 0.02935566008090973, + 0.037531089037656784, + 0.027457306161522865, + 0.01693442463874817, + 0.028515450656414032, + 0.01243431307375431, + 0.019153285771608353, + 0.011824599467217922, + 0.011366982012987137, + 0.01139337569475174, + 0.011783021502196789, + 0.008624682202935219, + 0.00844933744519949, + 0.00670403428375721, + 0.007262095808982849, + 0.010910160839557648, + 0.016441648826003075, + 0.015346885658800602, + 0.00732102757319808, + 0.021980710327625275, + 0.010467324405908585, + 0.014300552196800709, + 0.008010044693946838, + 0.005331358872354031, + 0.004823792725801468, + 0.006313034798949957, + 0.004944603890180588, + 0.004542792681604624, + 0.0037868223153054714, + 0.00652981037274003, + 0.003802528139203787, + 0.006079846993088722, + 0.007051038555800915, + 0.0047575621865689754, + 0.005965850315988064, + 0.0042203147895634174, + 0.00474570645019412, + 0.004309314768761396, + 0.003975163213908672, + 0.007057725451886654, + 0.0045045302249491215, + 0.004436017945408821, + 0.004442682024091482, + 0.004876796621829271, + 0.005069534759968519, + 0.005220332182943821, + 0.007604346610605717, + 0.005623076111078262, + 0.004572188016027212, + 0.008924962021410465, + 0.00643119914457202, + 0.005692030768841505, + 0.00578537629917264, + 0.005937105976045132, + 0.0067636859603226185, + 0.010720799677073956, + 0.0187167227268219, + 0.015872901305556297, + 0.016882644966244698, + 0.020911656320095062, + 0.016861479729413986, + 0.021190688014030457, + 0.01587366871535778, + 0.014744839631021023, + 0.016735956072807312, + 0.028230031952261925, + 0.0487939789891243, + 0.022419163957238197, + 0.026005132123827934, + 0.037431079894304276, + 0.028861621394753456, + 0.008071856573224068, + 0.006141593214124441, + 0.01310692261904478, + 0.014085812494158745, + 0.008473346941173077, + 0.013926000334322453, + 0.012329519726336002, + 0.011452030390501022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.04864376410841942 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin in absolute darkness before being illuminated by a massive lightning bolt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hard, perfectly spherical black fungi visible clinging to a rotting, moss-covered log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fungi spheres exhibit fine cracks and a carbonaceous surface covered in slick rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow downward tilt transitioning into an extreme close-up of a heavy spherical raindrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a stark, high-contrast look with piercing whites and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735715549439192 + ], + "opposite_logprob_tokens": [ + -5.006735801696777 + ], + "logprob_reward": -0.006735715549439192, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735715549439192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17685125768184662, + 0.07575538009405136, + 0.03874313086271286, + 0.12697839736938477, + 0.0932481586933136, + 0.061195842921733856, + 0.05880305543541908, + 0.12095870077610016, + 0.1450112760066986, + 0.10245480388402939 + ], + "credit_map": [ + 0.020083727315068245, + 0.02795523777604103, + 0.02958623506128788, + 0.019127121195197105, + 0.025507386773824692, + 0.023335125297307968, + 0.014705895446240902, + 0.016550524160265923, + 0.00889297854155302, + 0.007015753071755171, + 0.009299353696405888, + 0.013541280291974545, + 0.012955990619957447, + 0.01019846461713314, + 0.00894130952656269, + 0.004910241346806288, + 0.0035687233321368694, + 0.00432576285675168, + 0.005355439614504576, + 0.00284641538746655, + 0.010536321438848972, + 0.004307780880481005, + 0.005104019772261381, + 0.0026986703742295504, + 0.023413948714733124, + 0.013562777079641819, + 0.045149125158786774, + 0.007606798782944679, + 0.01020857971161604, + 0.005519102793186903, + 0.008789547719061375, + 0.012728534638881683, + 0.012056909501552582, + 0.01603689044713974, + 0.008373890072107315, + 0.010141519829630852, + 0.010969094932079315, + 0.00663349311798811, + 0.017873723059892654, + 0.011162635870277882, + 0.010657867416739464, + 0.007624264806509018, + 0.007581021171063185, + 0.005727883893996477, + 0.005077666603028774, + 0.0064479089342057705, + 0.007324259262531996, + 0.01075496431440115, + 0.006819284055382013, + 0.005365484859794378, + 0.009201970882713795, + 0.008362667635083199, + 0.007923411205410957, + 0.005650474224239588, + 0.008095331490039825, + 0.007384428288787603, + 0.009428368881344795, + 0.01638842560350895, + 0.017180481925606728, + 0.012989786453545094, + 0.017727579921483994, + 0.011931507848203182, + 0.02062831073999405, + 0.014684227295219898, + 0.011406703852117062, + 0.015077413059771061, + 0.018362099304795265, + 0.02948606014251709, + 0.016300415620207787, + 0.017092401161789894, + 0.02087407186627388, + 0.01641211472451687, + 0.007946610450744629, + 0.0054145990870893, + 0.012026621960103512, + 0.025339510291814804, + 0.008534234017133713, + 0.009743869304656982, + 0.01643948443233967, + 0.01700986921787262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.006735715549439192 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The scene begins in absolute darkness, then a static media close-up is violently illuminated by a massive, silent lightning bolt. In an instant, a brilliant, icy white light streaks across the void: a cluster of hard, perfectly spherical black fungi clinging tenaciously to a rotting, moss-covered log. Each sphere's carbonaceous surface, covered in a series of fine cracks, glitters with slick rainwater, its convex shapes catching the harsh light with sharp, mirrored points. Just as suddenly, the world plunges back into deep darkness. From this void, the camera begins a slow, deliberate downward tilt, transitioning into extreme close-up, carefully following a heavy, spherical raindrop as it traces a sticky path along the curved surface of one of the fungi. The transparent drop distorts the matte black texture beneath, its journey a mesmerizing micro-event following the raging, silent aftermath of the storm. The entire visual is captured with a stark, high-contrast digital look, characterized by piercing whites and deep, engrossing blacks, creating an atmosphere of tension, mystery, and intense concentration.", + "seed": 28609, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_10/index_8854_seed_28609_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08081903932616115, + "vlm_logprob": -0.08081903932616115, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.08081903932616115, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin in absolute darkness before being illuminated by a massive lightning bolt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hard, perfectly spherical black fungi visible clinging to a rotting, moss-covered log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fungi spheres exhibit fine cracks and a carbonaceous surface covered in slick rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow downward tilt transitioning into an extreme close-up of a heavy spherical raindrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a stark, high-contrast look with piercing whites and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3566720485687256 + ], + "opposite_logprob_tokens": [ + -1.2316720485687256 + ], + "logprob_reward": -0.3566720485687256, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3566720485687256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19649995863437653, + 0.027106260880827904, + 0.037127140909433365, + 0.045437999069690704, + 0.02652890607714653, + 0.3128071129322052, + 0.1219240128993988, + 0.16714388132095337, + 0.016866182908415794, + 0.04855852946639061 + ], + "credit_map": [ + 0.030524559319019318, + 0.01729690097272396, + 0.014825557358562946, + 0.011345900595188141, + 0.05572196841239929, + 0.03992600366473198, + 0.014037104323506355, + 0.01282196119427681, + 0.003505369182676077, + 0.0030428238678723574, + 0.0029992666095495224, + 0.0027278654742985964, + 0.0022025832440704107, + 0.005946677643805742, + 0.0030707952100783587, + 0.0036108798813074827, + 0.004574773367494345, + 0.003248699940741062, + 0.00473052030429244, + 0.0031937083695083857, + 0.006386337336152792, + 0.005862132180482149, + 0.004342751577496529, + 0.004788219928741455, + 0.005097732879221439, + 0.004052409436553717, + 0.006427470128983259, + 0.0037262411788105965, + 0.0024939794093370438, + 0.006385020446032286, + 0.01001698523759842, + 0.007238158956170082, + 0.002457595197483897, + 0.0035905775148421526, + 0.002320889150723815, + 0.0030388699378818274, + 0.002368916291743517, + 0.0035705233458429575, + 0.005211654584854841, + 0.003969881217926741, + 0.029276801273226738, + 0.03983958065509796, + 0.025947047397494316, + 0.030999811366200447, + 0.0387982577085495, + 0.04259578883647919, + 0.06348352134227753, + 0.041866280138492584, + 0.010077805258333683, + 0.011005972512066364, + 0.02263788878917694, + 0.009716400876641273, + 0.017307789996266365, + 0.02022503688931465, + 0.016171706840395927, + 0.014781415462493896, + 0.015815630555152893, + 0.013113862834870815, + 0.03625347465276718, + 0.018992580473423004, + 0.019014189019799232, + 0.024010706692934036, + 0.022933639585971832, + 0.0170097965747118, + 0.0017225459450855851, + 0.0016039515612646937, + 0.002482030540704727, + 0.0018441870342940092, + 0.002280531683936715, + 0.002007580129429698, + 0.002294554840773344, + 0.002630801172927022, + 0.004476232919842005, + 0.004884705878794193, + 0.00773152569308877, + 0.004934689495712519, + 0.006441508885473013, + 0.006246541161090136, + 0.007155763916671276, + 0.006687564309686422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.3566720485687256 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin in absolute darkness before being illuminated by a massive lightning bolt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hard, perfectly spherical black fungi visible clinging to a rotting, moss-covered log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fungi spheres exhibit fine cracks and a carbonaceous surface covered in slick rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow downward tilt transitioning into an extreme close-up of a heavy spherical raindrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a stark, high-contrast look with piercing whites and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00980218406766653 + ], + "opposite_logprob_tokens": [ + -4.634802341461182 + ], + "logprob_reward": -0.00980218406766653, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00980218406766653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06981421262025833, + 0.04244435206055641, + 0.05406045913696289, + 0.02941322512924671, + 0.0301984753459692, + 0.41952744126319885, + 0.13341866433620453, + 0.1517994999885559, + 0.03627634793519974, + 0.033047374337911606 + ], + "credit_map": [ + 0.009935419075191021, + 0.00984590221196413, + 0.007209694944322109, + 0.00571842584758997, + 0.014805137179791927, + 0.009140364825725555, + 0.006435944698750973, + 0.006723308004438877, + 0.004870704375207424, + 0.010546169243752956, + 0.003201159182935953, + 0.0053142160177230835, + 0.0029375662561506033, + 0.00582277774810791, + 0.005194170866161585, + 0.004557583015412092, + 0.007047804072499275, + 0.00327655510045588, + 0.0062318830750882626, + 0.003883408149704337, + 0.007014169357717037, + 0.006468304432928562, + 0.007786528207361698, + 0.012351798824965954, + 0.001892279600724578, + 0.001962513430044055, + 0.0031467373482882977, + 0.0032550545874983072, + 0.0019089211709797382, + 0.00373308383859694, + 0.008377316407859325, + 0.005137315485626459, + 0.0022652712650597095, + 0.0033860686235129833, + 0.002328544156625867, + 0.0036068467888981104, + 0.002697804942727089, + 0.004234792664647102, + 0.006706759799271822, + 0.004972382448613644, + 0.031694408506155014, + 0.045628711581230164, + 0.030818432569503784, + 0.03500533476471901, + 0.05505220592021942, + 0.061143193393945694, + 0.09704378992319107, + 0.06314131617546082, + 0.009899427182972431, + 0.014299760572612286, + 0.02266482077538967, + 0.011208911426365376, + 0.01756007969379425, + 0.021095214411616325, + 0.019286036491394043, + 0.01740439608693123, + 0.014282317832112312, + 0.012322631664574146, + 0.032781898975372314, + 0.017680762335658073, + 0.016101529821753502, + 0.022658061236143112, + 0.021383482962846756, + 0.014588802121579647, + 0.00353909726254642, + 0.0040160659700632095, + 0.007363282144069672, + 0.003013462061062455, + 0.004136219155043364, + 0.004665381275117397, + 0.003950406331568956, + 0.005592429544776678, + 0.003576994640752673, + 0.0058696074411273, + 0.004138626158237457, + 0.0042834035120904446, + 0.004281083587557077, + 0.003864030819386244, + 0.003728699404746294, + 0.003304925048723817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.00980218406766653 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin in absolute darkness before being illuminated by a massive lightning bolt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hard, perfectly spherical black fungi visible clinging to a rotting, moss-covered log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fungi spheres exhibit fine cracks and a carbonaceous surface covered in slick rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow downward tilt transitioning into an extreme close-up of a heavy spherical raindrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a stark, high-contrast look with piercing whites and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059562805108726025 + ], + "opposite_logprob_tokens": [ + -5.130956172943115 + ], + "logprob_reward": -0.0059562805108726025, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059562805108726025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23370835185050964, + 0.08841632306575775, + 0.07693703472614288, + 0.043503742665052414, + 0.0445307157933712, + 0.12143325805664062, + 0.0795304924249649, + 0.08992676436901093, + 0.08353962004184723, + 0.1384737193584442 + ], + "credit_map": [ + 0.034023456275463104, + 0.04234478622674942, + 0.024242915213108063, + 0.0248448196798563, + 0.028576726093888283, + 0.03439280763268471, + 0.0241059809923172, + 0.02117687463760376, + 0.011061671189963818, + 0.017205817624926567, + 0.010487489402294159, + 0.011763566173613071, + 0.006873734761029482, + 0.01321425475180149, + 0.008118300698697567, + 0.00969148613512516, + 0.011295831762254238, + 0.005951103754341602, + 0.009196002967655659, + 0.004421434365212917, + 0.011121625080704689, + 0.007672064937651157, + 0.008715739473700523, + 0.01856323331594467, + 0.004171437583863735, + 0.0038435545284301043, + 0.006288475356996059, + 0.004248448181897402, + 0.003181603504344821, + 0.005527022760361433, + 0.009858752600848675, + 0.006384446285665035, + 0.005841853097081184, + 0.005410332232713699, + 0.006011263467371464, + 0.006103265564888716, + 0.003811535192653537, + 0.006176905240863562, + 0.005131586920469999, + 0.006043971516191959, + 0.008891984820365906, + 0.011783203110098839, + 0.010858225636184216, + 0.013872183859348297, + 0.013126591220498085, + 0.01768314279615879, + 0.02607003040611744, + 0.019147902727127075, + 0.005494048353284597, + 0.009950998239219189, + 0.009608951397240162, + 0.008532607927918434, + 0.01389721967279911, + 0.011801357381045818, + 0.011373828165233135, + 0.008871480822563171, + 0.008056486025452614, + 0.008189760148525238, + 0.014814289286732674, + 0.010781481862068176, + 0.011012066155672073, + 0.013200229965150356, + 0.012647779658436775, + 0.01122466940432787, + 0.006574889179319143, + 0.006433167960494757, + 0.015712428838014603, + 0.008326214738190174, + 0.012047401629388332, + 0.012979169376194477, + 0.010719258338212967, + 0.010747088119387627, + 0.01045716367661953, + 0.017708880826830864, + 0.014364916831254959, + 0.016574528068304062, + 0.018413733690977097, + 0.02424691803753376, + 0.017527194693684578, + 0.01918039098381996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.0059562805108726025 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin in absolute darkness before being illuminated by a massive lightning bolt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hard, perfectly spherical black fungi visible clinging to a rotting, moss-covered log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fungi spheres exhibit fine cracks and a carbonaceous surface covered in slick rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow downward tilt transitioning into an extreme close-up of a heavy spherical raindrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a stark, high-contrast look with piercing whites and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020585225895047188 + ], + "opposite_logprob_tokens": [ + -3.895585298538208 + ], + "logprob_reward": -0.020585225895047188, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020585225895047188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0635695531964302, + 0.027903560549020767, + 0.055061567574739456, + 0.03486204892396927, + 0.026114264503121376, + 0.4285105764865875, + 0.14506669342517853, + 0.17906463146209717, + 0.009070486761629581, + 0.030776597559452057 + ], + "credit_map": [ + 0.01022940594702959, + 0.009758904576301575, + 0.00563729926943779, + 0.004407905042171478, + 0.014871770516037941, + 0.008193505927920341, + 0.005676747299730778, + 0.0047940220683813095, + 0.002966369967907667, + 0.005286866333335638, + 0.0030166206415742636, + 0.0029363769572228193, + 0.0019833091646432877, + 0.005664302036166191, + 0.0030464488081634045, + 0.0030032703652977943, + 0.0070624458603560925, + 0.0042605712078511715, + 0.006751664914190769, + 0.004757725168019533, + 0.009069402702152729, + 0.006331542041152716, + 0.006148118060082197, + 0.010680104605853558, + 0.003666390897706151, + 0.0029111679177731276, + 0.0038375742733478546, + 0.0034603262320160866, + 0.0019239906687289476, + 0.005065121687948704, + 0.00808267667889595, + 0.005914805456995964, + 0.0022794336546212435, + 0.0033751882147043943, + 0.002649639965966344, + 0.0026938102673739195, + 0.002227042568847537, + 0.0037871096283197403, + 0.0050729201175272465, + 0.004029121715575457, + 0.03773179277777672, + 0.05223240703344345, + 0.03308698907494545, + 0.040647003799676895, + 0.053084664046764374, + 0.05855490267276764, + 0.09288760274648666, + 0.06028524041175842, + 0.012377120554447174, + 0.014664888381958008, + 0.02808062918484211, + 0.012298714369535446, + 0.020558662712574005, + 0.021668804809451103, + 0.01984604448080063, + 0.015571853145956993, + 0.017529213801026344, + 0.013379979878664017, + 0.039715059101581573, + 0.019525492563843727, + 0.020876385271549225, + 0.025426199659705162, + 0.025421613827347755, + 0.017190687358379364, + 0.0006444295286200941, + 0.0005866747233085334, + 0.0015389781910926104, + 0.0009147447417490184, + 0.001418287749402225, + 0.0014725065557286143, + 0.0011786160757765174, + 0.0013162491377443075, + 0.004351391457021236, + 0.002691827481612563, + 0.004522429779171944, + 0.0037232020404189825, + 0.00425820704549551, + 0.003929875325411558, + 0.004547771066427231, + 0.002751898020505905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.020585225895047188 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin in absolute darkness before being illuminated by a massive lightning bolt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hard, perfectly spherical black fungi visible clinging to a rotting, moss-covered log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fungi spheres exhibit fine cracks and a carbonaceous surface covered in slick rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow downward tilt transitioning into an extreme close-up of a heavy spherical raindrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a stark, high-contrast look with piercing whites and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011079457588493824 + ], + "opposite_logprob_tokens": [ + -4.51107931137085 + ], + "logprob_reward": -0.011079457588493824, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011079457588493824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24318581819534302, + 0.05446891859173775, + 0.04194019362330437, + 0.10189705342054367, + 0.03553250804543495, + 0.21474412083625793, + 0.1178448274731636, + 0.07099801301956177, + 0.0386386439204216, + 0.08074987679719925 + ], + "credit_map": [ + 0.03723982349038124, + 0.029188698157668114, + 0.019673945382237434, + 0.016950508579611778, + 0.061370860785245895, + 0.04359244927763939, + 0.021411970257759094, + 0.013757559470832348, + 0.006743158213794231, + 0.013658926822245121, + 0.004930400289595127, + 0.0062919361516833305, + 0.0037858698051422834, + 0.00543162040412426, + 0.007868070155382156, + 0.005758934188634157, + 0.004740527831017971, + 0.0039155506528913975, + 0.006080766208469868, + 0.0032923375256359577, + 0.002768842037767172, + 0.007770483382046223, + 0.005530022084712982, + 0.007841664366424084, + 0.010474693030118942, + 0.00995370838791132, + 0.013034633360803127, + 0.006687050685286522, + 0.006608092226088047, + 0.014200921170413494, + 0.02360215038061142, + 0.017335807904601097, + 0.004656692035496235, + 0.002599612344056368, + 0.005061583127826452, + 0.0031588254496455193, + 0.0039492323994636536, + 0.007519115693867207, + 0.00346457795239985, + 0.005122868809849024, + 0.016213305294513702, + 0.019202962517738342, + 0.013064774684607983, + 0.015606214292347431, + 0.02666378952562809, + 0.04130321368575096, + 0.04772922024130821, + 0.03496062755584717, + 0.0073739392682909966, + 0.01504181232303381, + 0.013921195641160011, + 0.013140024617314339, + 0.019345059990882874, + 0.016115065664052963, + 0.02113501913845539, + 0.011772707104682922, + 0.005655127111822367, + 0.006688248831778765, + 0.013540861196815968, + 0.008616607636213303, + 0.010968009009957314, + 0.009704159572720528, + 0.008608526550233364, + 0.007216468453407288, + 0.0031972327269613743, + 0.0033719006460160017, + 0.0073644500225782394, + 0.0043528941459953785, + 0.0059682950377464294, + 0.003947126213461161, + 0.005379785783588886, + 0.005056958179920912, + 0.007989068515598774, + 0.010536633431911469, + 0.009957761503756046, + 0.007949847728013992, + 0.010780546814203262, + 0.012080827727913857, + 0.011809742078185081, + 0.009645449928939342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.011079457588493824 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The scene begins in absolute darkness, then a static media close-up is violently illuminated by a massive, silent lightning bolt. In an instant, a brilliant, icy white light streaks across the void: a cluster of hard, perfectly spherical black fungi clinging tenaciously to a rotting, moss-covered log. Each sphere's carbonaceous surface, covered in a series of fine cracks, glitters with slick rainwater, its convex shapes catching the harsh light with sharp, mirrored points. Just as suddenly, the world plunges back into deep darkness. From this void, the camera begins a slow, deliberate downward tilt, transitioning into extreme close-up, carefully following a heavy, spherical raindrop as it traces a sticky path along the curved surface of one of the fungi. The transparent drop distorts the matte black texture beneath, its journey a mesmerizing micro-event following the raging, silent aftermath of the storm. The entire visual is captured with a stark, high-contrast digital look, characterized by piercing whites and deep, engrossing blacks, creating an atmosphere of tension, mystery, and intense concentration.", + "seed": 28610, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_10/index_8854_seed_28610_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04334768075495958, + "vlm_logprob": -0.04334768075495958, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.04334768075495958, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin in absolute darkness before being illuminated by a massive lightning bolt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hard, perfectly spherical black fungi visible clinging to a rotting, moss-covered log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fungi spheres exhibit fine cracks and a carbonaceous surface covered in slick rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow downward tilt transitioning into an extreme close-up of a heavy spherical raindrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a stark, high-contrast look with piercing whites and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1354501098394394 + ], + "opposite_logprob_tokens": [ + -2.1354501247406006 + ], + "logprob_reward": -0.1354501098394394, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1354501098394394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07535554468631744, + 0.21298350393772125, + 0.10390754789113998, + 0.0388743095099926, + 0.03074895218014717, + 0.04936021938920021, + 0.05420071259140968, + 0.07333490252494812, + 0.17121465504169464, + 0.1900196671485901 + ], + "credit_map": [ + 0.007426188327372074, + 0.007038125768303871, + 0.015663200989365578, + 0.01190347783267498, + 0.004250948317348957, + 0.009662903845310211, + 0.010522904805839062, + 0.008887788280844688, + 0.021320756524801254, + 0.02410033345222473, + 0.019453732296824455, + 0.03126397356390953, + 0.029315628111362457, + 0.03840384632349014, + 0.03020334802567959, + 0.018921881914138794, + 0.011456267908215523, + 0.013368654996156693, + 0.010000761598348618, + 0.012625066563487053, + 0.008514239452779293, + 0.009276083670556545, + 0.027967918664216995, + 0.01069855410605669, + 0.002548405434936285, + 0.004724377766251564, + 0.004349313210695982, + 0.006914643105119467, + 0.0040557063184678555, + 0.003539104713127017, + 0.007690072990953922, + 0.005052683874964714, + 0.0024772074539214373, + 0.005518513731658459, + 0.0049106888473033905, + 0.005567577201873064, + 0.002050630748271942, + 0.0031752402428537607, + 0.003386060008779168, + 0.0036630339454859495, + 0.006208001170307398, + 0.0040320465341210365, + 0.007021890487521887, + 0.007145638111978769, + 0.00652372557669878, + 0.004442742560058832, + 0.009075315669178963, + 0.0049108583480119705, + 0.006187379825860262, + 0.005102722905576229, + 0.007588258478790522, + 0.009515711106359959, + 0.005602946970611811, + 0.006331598851829767, + 0.008265219628810883, + 0.005606873892247677, + 0.007168086711317301, + 0.0064700013026595116, + 0.010137963108718395, + 0.011703302152454853, + 0.010002323426306248, + 0.00589211517944932, + 0.015132809989154339, + 0.006828295532613993, + 0.01906469278037548, + 0.02195701375603676, + 0.027647078037261963, + 0.025486411526799202, + 0.013741137459874153, + 0.015410780906677246, + 0.021645348519086838, + 0.02626219019293785, + 0.011694510467350483, + 0.009027030318975449, + 0.04655799642205238, + 0.022676872089505196, + 0.01874300092458725, + 0.011770585551857948, + 0.023168228566646576, + 0.04638144001364708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 8 + ], + "selected_reward": -0.1354501098394394 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin in absolute darkness before being illuminated by a massive lightning bolt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hard, perfectly spherical black fungi visible clinging to a rotting, moss-covered log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fungi spheres exhibit fine cracks and a carbonaceous surface covered in slick rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow downward tilt transitioning into an extreme close-up of a heavy spherical raindrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a stark, high-contrast look with piercing whites and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014211754314601421 + ], + "opposite_logprob_tokens": [ + -4.264211654663086 + ], + "logprob_reward": -0.014211754314601421, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014211754314601421, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12629829347133636, + 0.034476980566978455, + 0.05027962848544121, + 0.0625104084610939, + 0.08477789908647537, + 0.06631729751825333, + 0.08893608301877975, + 0.3354516923427582, + 0.06201343238353729, + 0.08893831819295883 + ], + "credit_map": [ + 0.013960559852421284, + 0.009410278871655464, + 0.01981930248439312, + 0.01690075546503067, + 0.006299168802797794, + 0.014410669915378094, + 0.02469814009964466, + 0.02079942263662815, + 0.003788038156926632, + 0.0054364437237381935, + 0.004328591283410788, + 0.0041174027137458324, + 0.004722290672361851, + 0.004219620488584042, + 0.004983807913959026, + 0.0028807842172682285, + 0.004173032473772764, + 0.006819277070462704, + 0.0073342821560800076, + 0.01042885985225439, + 0.0035350474063307047, + 0.003616183064877987, + 0.00962686538696289, + 0.004746078979223967, + 0.004080102778971195, + 0.0071409004740417, + 0.011505287140607834, + 0.009495549835264683, + 0.004714767448604107, + 0.006573458667844534, + 0.009447623044252396, + 0.009552717208862305, + 0.006720265839248896, + 0.013426878489553928, + 0.011117655783891678, + 0.010429189540445805, + 0.009915116243064404, + 0.01084765512496233, + 0.012408891692757607, + 0.009912248700857162, + 0.00901282299309969, + 0.008661498315632343, + 0.008717376738786697, + 0.0068884119391441345, + 0.009716473519802094, + 0.006244582124054432, + 0.0094358054921031, + 0.007640328258275986, + 0.006966287270188332, + 0.0059449560940265656, + 0.013743766583502293, + 0.010527863167226315, + 0.017291761934757233, + 0.009807550348341465, + 0.016403762623667717, + 0.008250134065747261, + 0.027033856138586998, + 0.0267757810652256, + 0.0441863089799881, + 0.04664125293493271, + 0.04904690757393837, + 0.027299413457512856, + 0.08663555979728699, + 0.02783260866999626, + 0.007301593665033579, + 0.006845487747341394, + 0.010375084355473518, + 0.008738679811358452, + 0.005720996297895908, + 0.005385672207921743, + 0.007110212929546833, + 0.010535705834627151, + 0.011860697530210018, + 0.00791987031698227, + 0.016027675941586494, + 0.01223902590572834, + 0.008611993864178658, + 0.006887802854180336, + 0.009882953017950058, + 0.01550829503685236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.014211754314601421 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin in absolute darkness before being illuminated by a massive lightning bolt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hard, perfectly spherical black fungi visible clinging to a rotting, moss-covered log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fungi spheres exhibit fine cracks and a carbonaceous surface covered in slick rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow downward tilt transitioning into an extreme close-up of a heavy spherical raindrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a stark, high-contrast look with piercing whites and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008645604364573956 + ], + "opposite_logprob_tokens": [ + -4.758645534515381 + ], + "logprob_reward": -0.008645604364573956, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008645604364573956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09352756291627884, + 0.09018897265195847, + 0.045239586383104324, + 0.060749851167201996, + 0.05740213394165039, + 0.07678604871034622, + 0.06473079323768616, + 0.3072049021720886, + 0.05015658959746361, + 0.1540135145187378 + ], + "credit_map": [ + 0.00911234226077795, + 0.016279397532343864, + 0.011218911036849022, + 0.011561394669115543, + 0.005769542418420315, + 0.011359278112649918, + 0.015029092319309711, + 0.013197611086070538, + 0.009696080349385738, + 0.012197968550026417, + 0.009637841023504734, + 0.00809639785438776, + 0.013083133846521378, + 0.022271106019616127, + 0.00603802315890789, + 0.009168430231511593, + 0.003995150793343782, + 0.006708500906825066, + 0.007304004393517971, + 0.007786117494106293, + 0.004946952685713768, + 0.004312911070883274, + 0.005947401747107506, + 0.004238549619913101, + 0.004619569052010775, + 0.007920492440462112, + 0.01247350312769413, + 0.009390546940267086, + 0.00489745894446969, + 0.005258167628198862, + 0.007106873206794262, + 0.00908324308693409, + 0.003923023585230112, + 0.009825493209064007, + 0.008416296914219856, + 0.006154360715299845, + 0.007060544099658728, + 0.00800418108701706, + 0.007091027684509754, + 0.006927212700247765, + 0.014396464452147484, + 0.007937837392091751, + 0.010798864997923374, + 0.008066081441938877, + 0.006046512629836798, + 0.008977868594229221, + 0.009131439961493015, + 0.011430987156927586, + 0.005537346936762333, + 0.008409165777266026, + 0.009500140324234962, + 0.011314787901937962, + 0.007360960356891155, + 0.008445169776678085, + 0.007679966744035482, + 0.006483261939138174, + 0.024888917803764343, + 0.02601245790719986, + 0.041083935648202896, + 0.05034097284078598, + 0.0351703017950058, + 0.023538928478956223, + 0.0818648710846901, + 0.02430455945432186, + 0.004407055675983429, + 0.004460859578102827, + 0.008566776290535927, + 0.006741213612258434, + 0.004412610549479723, + 0.003350823651999235, + 0.007782197557389736, + 0.010435054078698158, + 0.013508765026926994, + 0.010244897566735744, + 0.02531338855624199, + 0.026472100988030434, + 0.01647387631237507, + 0.010653335601091385, + 0.024602718651294708, + 0.026744430884718895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.008645604364573956 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin in absolute darkness before being illuminated by a massive lightning bolt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hard, perfectly spherical black fungi visible clinging to a rotting, moss-covered log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fungi spheres exhibit fine cracks and a carbonaceous surface covered in slick rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow downward tilt transitioning into an extreme close-up of a heavy spherical raindrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a stark, high-contrast look with piercing whites and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048647284507751465 + ], + "opposite_logprob_tokens": [ + -3.048647403717041 + ], + "logprob_reward": -0.048647284507751465, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048647284507751465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.023742448538541794, + 0.046265337616205215, + 0.07749485969543457, + 0.03905896097421646, + 0.11593952029943466, + 0.145822212100029, + 0.05226602777838707, + 0.4037654995918274, + 0.05344385281205177, + 0.04220129922032356 + ], + "credit_map": [ + 0.002790398895740509, + 0.003610591869801283, + 0.0033498939592391253, + 0.0031719389371573925, + 0.0014526844024658203, + 0.003934183623641729, + 0.002252517966553569, + 0.0031802398152649403, + 0.005022416356950998, + 0.004789972677826881, + 0.002772530075162649, + 0.0038464097306132317, + 0.008170191198587418, + 0.014511963352560997, + 0.0036232913844287395, + 0.003528560046106577, + 0.010170713067054749, + 0.014082323759794235, + 0.008158133365213871, + 0.011121129617094994, + 0.004867300856858492, + 0.004596875049173832, + 0.017048664391040802, + 0.007449720520526171, + 0.0035094860941171646, + 0.005535951815545559, + 0.006301948335021734, + 0.005168678238987923, + 0.0030856402590870857, + 0.0033859543036669493, + 0.006602720357477665, + 0.005468579940497875, + 0.006584437098354101, + 0.015052633360028267, + 0.018023785203695297, + 0.02052033506333828, + 0.011875067837536335, + 0.01391909271478653, + 0.016131550073623657, + 0.013832628726959229, + 0.012871600687503815, + 0.019372256472706795, + 0.02055378258228302, + 0.018262425437569618, + 0.020571503788232803, + 0.015353923663496971, + 0.021738246083259583, + 0.01709846779704094, + 0.0037848714273422956, + 0.004511664155870676, + 0.0074198367074131966, + 0.005961359478533268, + 0.012268483638763428, + 0.005291321314871311, + 0.008198154158890247, + 0.004830338526517153, + 0.02947862632572651, + 0.033585503697395325, + 0.052381496876478195, + 0.053314872086048126, + 0.05958451330661774, + 0.03349435701966286, + 0.10895248502492905, + 0.03297366201877594, + 0.004039994906634092, + 0.004436513874679804, + 0.00696400785818696, + 0.007075376342982054, + 0.007624022662639618, + 0.005905290599912405, + 0.0120792081579566, + 0.005319436080753803, + 0.0033007182646542788, + 0.003068148158490658, + 0.0056917485781013966, + 0.008730084635317326, + 0.005283668637275696, + 0.0025005051866173744, + 0.0074626728892326355, + 0.006163754966109991 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.048647284507751465 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin in absolute darkness before being illuminated by a massive lightning bolt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hard, perfectly spherical black fungi visible clinging to a rotting, moss-covered log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fungi spheres exhibit fine cracks and a carbonaceous surface covered in slick rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow downward tilt transitioning into an extreme close-up of a heavy spherical raindrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a stark, high-contrast look with piercing whites and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009783650748431683 + ], + "opposite_logprob_tokens": [ + -4.634783744812012 + ], + "logprob_reward": -0.009783650748431683, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783650748431683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05601877346634865, + 0.11345111578702927, + 0.021415142342448235, + 0.02760343998670578, + 0.09335378557443619, + 0.09713107347488403, + 0.04660673812031746, + 0.0948001891374588, + 0.061122436076402664, + 0.3884972929954529 + ], + "credit_map": [ + 0.0070029255002737045, + 0.0041627949103713036, + 0.008821088820695877, + 0.010470351204276085, + 0.0029957948718219995, + 0.005479908082634211, + 0.006986464839428663, + 0.010099444538354874, + 0.012188148684799671, + 0.01121455430984497, + 0.008909113705158234, + 0.01266446989029646, + 0.01682019978761673, + 0.027546318247914314, + 0.014042404480278492, + 0.010065915063023567, + 0.0020720886532217264, + 0.003182041458785534, + 0.002505965530872345, + 0.003419801825657487, + 0.0015198285691440105, + 0.0022687078453600407, + 0.003638077527284622, + 0.0028086318634450436, + 0.002582746557891369, + 0.004403182305395603, + 0.002972942078486085, + 0.0032047757413238287, + 0.002893459051847458, + 0.002454617992043495, + 0.004944595042616129, + 0.0041471198201179504, + 0.006298450753092766, + 0.02256103605031967, + 0.01964908093214035, + 0.009835377335548401, + 0.009809491224586964, + 0.008658596314489841, + 0.0074250586330890656, + 0.009116693399846554, + 0.016078747808933258, + 0.01067063957452774, + 0.014688354916870594, + 0.015093278139829636, + 0.008367041125893593, + 0.0085325688123703, + 0.012738330289721489, + 0.010962104424834251, + 0.005100032780319452, + 0.0063554951921105385, + 0.006362760439515114, + 0.0072310506366193295, + 0.005809511058032513, + 0.005926704965531826, + 0.00555502949282527, + 0.004266153555363417, + 0.008786424994468689, + 0.008882489055395126, + 0.015215498395264149, + 0.01531134732067585, + 0.011841930449008942, + 0.012786700390279293, + 0.013294699601829052, + 0.008681096136569977, + 0.004337629768997431, + 0.006350519601255655, + 0.009125004522502422, + 0.010903717949986458, + 0.006159202195703983, + 0.003608879866078496, + 0.011566157452762127, + 0.009071326814591885, + 0.021863652393221855, + 0.018845997750759125, + 0.10822878032922745, + 0.047688115388154984, + 0.022232891991734505, + 0.022889098152518272, + 0.03753805533051491, + 0.10921069234609604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 5 + ], + "selected_reward": -0.009783650748431683 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The scene begins in absolute darkness, then a static media close-up is violently illuminated by a massive, silent lightning bolt. In an instant, a brilliant, icy white light streaks across the void: a cluster of hard, perfectly spherical black fungi clinging tenaciously to a rotting, moss-covered log. Each sphere's carbonaceous surface, covered in a series of fine cracks, glitters with slick rainwater, its convex shapes catching the harsh light with sharp, mirrored points. Just as suddenly, the world plunges back into deep darkness. From this void, the camera begins a slow, deliberate downward tilt, transitioning into extreme close-up, carefully following a heavy, spherical raindrop as it traces a sticky path along the curved surface of one of the fungi. The transparent drop distorts the matte black texture beneath, its journey a mesmerizing micro-event following the raging, silent aftermath of the storm. The entire visual is captured with a stark, high-contrast digital look, characterized by piercing whites and deep, engrossing blacks, creating an atmosphere of tension, mystery, and intense concentration.", + "seed": 28611, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_10/index_8854_seed_28611_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.033347731735557315, + "vlm_logprob": -0.033347731735557315, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.033347731735557315, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin in absolute darkness before being illuminated by a massive lightning bolt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hard, perfectly spherical black fungi visible clinging to a rotting, moss-covered log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fungi spheres exhibit fine cracks and a carbonaceous surface covered in slick rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow downward tilt transitioning into an extreme close-up of a heavy spherical raindrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a stark, high-contrast look with piercing whites and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11810342222452164 + ], + "opposite_logprob_tokens": [ + -2.243103504180908 + ], + "logprob_reward": -0.11810342222452164, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11810342222452164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2628454864025116, + 0.2250908464193344, + 0.1293831467628479, + 0.03860357776284218, + 0.03377540409564972, + 0.05911394953727722, + 0.10187577456235886, + 0.01480118278414011, + 0.08163846284151077, + 0.052872177213430405 + ], + "credit_map": [ + 0.022017860785126686, + 0.021063100546598434, + 0.011292045935988426, + 0.03070889785885811, + 0.04766135290265083, + 0.08736847341060638, + 0.025845550000667572, + 0.0168882105499506, + 0.00995771586894989, + 0.06217486038804054, + 0.011632001958787441, + 0.007604178041219711, + 0.007949957624077797, + 0.09187574684619904, + 0.009451430290937424, + 0.024444958195090294, + 0.019622420892119408, + 0.015063798055052757, + 0.01453491486608982, + 0.011436502449214458, + 0.011654503643512726, + 0.0251631923019886, + 0.019126057624816895, + 0.012781751342117786, + 0.005404244177043438, + 0.004213511012494564, + 0.005260489881038666, + 0.0033241224009543657, + 0.004793508443981409, + 0.003300718730315566, + 0.007922696880996227, + 0.00438428670167923, + 0.0035074318293482065, + 0.00312480004504323, + 0.005080163478851318, + 0.0048125069588422775, + 0.0034856272395700216, + 0.0034216654021292925, + 0.005815660580992699, + 0.004527549725025892, + 0.003872252069413662, + 0.006494973320513964, + 0.00881577841937542, + 0.009200873784720898, + 0.007812763564288616, + 0.0073507241904735565, + 0.007900433614850044, + 0.007666146382689476, + 0.007690961007028818, + 0.00820149201899767, + 0.013415886089205742, + 0.01362979132682085, + 0.015772422775626183, + 0.012607752345502377, + 0.015456108376383781, + 0.015101355500519276, + 0.001086479052901268, + 0.0009664130047895014, + 0.001514220959506929, + 0.0016513362061232328, + 0.0016432128613814712, + 0.002207469893619418, + 0.003296099603176117, + 0.00243595102801919, + 0.004664340987801552, + 0.009524615481495857, + 0.01661946438252926, + 0.00558509724214673, + 0.00850437581539154, + 0.02154073677957058, + 0.006574366707354784, + 0.008625462651252747, + 0.004788051825016737, + 0.009679156355559826, + 0.006545839831233025, + 0.0052077313885092735, + 0.007705319672822952, + 0.006959586404263973, + 0.0071357237175107, + 0.00485076941549778 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.11810342222452164 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin in absolute darkness before being illuminated by a massive lightning bolt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hard, perfectly spherical black fungi visible clinging to a rotting, moss-covered log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fungi spheres exhibit fine cracks and a carbonaceous surface covered in slick rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow downward tilt transitioning into an extreme close-up of a heavy spherical raindrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a stark, high-contrast look with piercing whites and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004108322784304619 + ], + "opposite_logprob_tokens": [ + -5.504108428955078 + ], + "logprob_reward": -0.004108322784304619, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004108322784304619, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2568480372428894, + 0.11609910428524017, + 0.047530945390462875, + 0.07240647822618484, + 0.0750449150800705, + 0.04452326148748398, + 0.16800172626972198, + 0.046392183750867844, + 0.12152745574712753, + 0.051625896245241165 + ], + "credit_map": [ + 0.01222224161028862, + 0.017445189878344536, + 0.008751711808145046, + 0.027057629078626633, + 0.05809657648205757, + 0.0996258333325386, + 0.019754497334361076, + 0.013894365169107914, + 0.009697443805634975, + 0.038161978125572205, + 0.006550231482833624, + 0.005333233159035444, + 0.007516459096223116, + 0.03560784459114075, + 0.007168824318796396, + 0.006063092965632677, + 0.007287196349352598, + 0.007215479854494333, + 0.005034857429563999, + 0.005002264864742756, + 0.004083412233740091, + 0.008352484554052353, + 0.004745978396385908, + 0.005809272639453411, + 0.009058544412255287, + 0.009772825986146927, + 0.009610895998775959, + 0.005945342127233744, + 0.010395240969955921, + 0.007436216808855534, + 0.012828481383621693, + 0.007358931470662355, + 0.0064647002145648, + 0.013378288596868515, + 0.009057953953742981, + 0.00787661038339138, + 0.011919273994863033, + 0.00627440121024847, + 0.010779823176562786, + 0.009293865412473679, + 0.0033348428551107645, + 0.003941004630178213, + 0.006406336557120085, + 0.00846710056066513, + 0.006320426240563393, + 0.004718701355159283, + 0.005160659551620483, + 0.0061741904355585575, + 0.009854314848780632, + 0.01189806405454874, + 0.0190909281373024, + 0.022335907444357872, + 0.029486611485481262, + 0.014702894724905491, + 0.03369971737265587, + 0.02693329192698002, + 0.0023657612036913633, + 0.001749914838001132, + 0.013545495457947254, + 0.0037886169739067554, + 0.002567172748968005, + 0.0038994201458990574, + 0.004953024908900261, + 0.01352277584373951, + 0.007742455694824457, + 0.004868440795689821, + 0.02941095270216465, + 0.010044441558420658, + 0.010157899931073189, + 0.0395948551595211, + 0.008728433400392532, + 0.010979973711073399, + 0.004536197520792484, + 0.008595782332122326, + 0.007072009611874819, + 0.0056153652258217335, + 0.008129526861011982, + 0.0077896141447126865, + 0.004773755092173815, + 0.00511364359408617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.004108322784304619 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin in absolute darkness before being illuminated by a massive lightning bolt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hard, perfectly spherical black fungi visible clinging to a rotting, moss-covered log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fungi spheres exhibit fine cracks and a carbonaceous surface covered in slick rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow downward tilt transitioning into an extreme close-up of a heavy spherical raindrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a stark, high-contrast look with piercing whites and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004099775105714798 + ], + "opposite_logprob_tokens": [ + -5.5040998458862305 + ], + "logprob_reward": -0.004099775105714798, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004099775105714798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10424981266260147, + 0.22884759306907654, + 0.10089538246393204, + 0.027823662385344505, + 0.044639334082603455, + 0.05083731189370155, + 0.2023032307624817, + 0.08140113949775696, + 0.11715060472488403, + 0.0418519601225853 + ], + "credit_map": [ + 0.011946944519877434, + 0.00836250651627779, + 0.007439049892127514, + 0.010971221141517162, + 0.018548356369137764, + 0.027535729110240936, + 0.009739909321069717, + 0.009706095792353153, + 0.012230120599269867, + 0.07932234555482864, + 0.008736270479857922, + 0.005588674917817116, + 0.006053940858691931, + 0.09302579611539841, + 0.007259569596499205, + 0.0166308730840683, + 0.014366153627634048, + 0.012057959102094173, + 0.011948681436479092, + 0.008207211270928383, + 0.010904069989919662, + 0.02059694565832615, + 0.0126417838037014, + 0.010172574780881405, + 0.0036168780643492937, + 0.004360392224043608, + 0.0033070610370486975, + 0.0020638322457671165, + 0.004123887047171593, + 0.002370130503550172, + 0.004674204159528017, + 0.0033072757069021463, + 0.004784632008522749, + 0.006229428108781576, + 0.006188166793435812, + 0.005895308684557676, + 0.005652730818837881, + 0.003898768685758114, + 0.006413714028894901, + 0.005576587747782469, + 0.0040688347071409225, + 0.005645718891173601, + 0.0056236241944134235, + 0.007185879163444042, + 0.008214525878429413, + 0.0067995754070580006, + 0.006840390153229237, + 0.006458762567490339, + 0.01529885083436966, + 0.014868775382637978, + 0.027644233778119087, + 0.025949811562895775, + 0.02905784361064434, + 0.021043024957180023, + 0.037223074585199356, + 0.031217610463500023, + 0.003896791720762849, + 0.002532018581405282, + 0.025704050436615944, + 0.0065674311481416225, + 0.0037523205392062664, + 0.006453696638345718, + 0.005971643142402172, + 0.02652319334447384, + 0.007063592318445444, + 0.013120354153215885, + 0.02129378356039524, + 0.009379721246659756, + 0.01499871350824833, + 0.030429022386670113, + 0.0101510901004076, + 0.010714324191212654, + 0.0049666049890220165, + 0.007495471276342869, + 0.00536946440115571, + 0.003277420997619629, + 0.005659799557179213, + 0.006007061339914799, + 0.004837898537516594, + 0.004238237626850605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 8 + ], + "selected_reward": -0.004099775105714798 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin in absolute darkness before being illuminated by a massive lightning bolt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hard, perfectly spherical black fungi visible clinging to a rotting, moss-covered log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fungi spheres exhibit fine cracks and a carbonaceous surface covered in slick rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow downward tilt transitioning into an extreme close-up of a heavy spherical raindrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a stark, high-contrast look with piercing whites and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369118645787239 + ], + "opposite_logprob_tokens": [ + -3.408691167831421 + ], + "logprob_reward": -0.03369118645787239, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369118645787239, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14547255635261536, + 0.18595284223556519, + 0.12231901288032532, + 0.059745870530605316, + 0.060713473707437515, + 0.037589557468891144, + 0.10305573791265488, + 0.03704480826854706, + 0.1517825424671173, + 0.09632354974746704 + ], + "credit_map": [ + 0.015582333318889141, + 0.010797582566738129, + 0.009729603305459023, + 0.014015180058777332, + 0.02535366453230381, + 0.04889920353889465, + 0.011861618608236313, + 0.009233381599187851, + 0.008587229065597057, + 0.061045777052640915, + 0.0069372523576021194, + 0.007469677831977606, + 0.00506102479994297, + 0.07488476485013962, + 0.007336466107517481, + 0.014630643650889397, + 0.01696617528796196, + 0.014853891916573048, + 0.01155758649110794, + 0.011493818834424019, + 0.013932785950601101, + 0.024072572588920593, + 0.014523793011903763, + 0.014918394386768341, + 0.00795113854110241, + 0.007287737913429737, + 0.006405670195817947, + 0.00569923548027873, + 0.008763324469327927, + 0.005272772163152695, + 0.013073230162262917, + 0.005292762536555529, + 0.006012657191604376, + 0.006848995108157396, + 0.008537180721759796, + 0.008739995770156384, + 0.005878576543182135, + 0.007135608699172735, + 0.009479694999754429, + 0.008080761879682541, + 0.003965611103922129, + 0.0033357408829033375, + 0.005084419157356024, + 0.005207989830523729, + 0.004876349121332169, + 0.0045930310152471066, + 0.005641711875796318, + 0.004884704947471619, + 0.00571867823600769, + 0.008232848718762398, + 0.010013246908783913, + 0.016312843188643456, + 0.01860330067574978, + 0.012076767161488533, + 0.015937283635139465, + 0.01616078056395054, + 0.003567052772268653, + 0.003219024511054158, + 0.004082134459167719, + 0.003976964857429266, + 0.004628906492143869, + 0.004739756695926189, + 0.008305457420647144, + 0.004525511525571346, + 0.012030508369207382, + 0.02049706131219864, + 0.020226120948791504, + 0.015996068716049194, + 0.020649267360568047, + 0.030231773853302002, + 0.014712309464812279, + 0.01743943989276886, + 0.009920921176671982, + 0.013497306033968925, + 0.011735898442566395, + 0.010194353759288788, + 0.017654428258538246, + 0.012590684927999973, + 0.010621496476233006, + 0.010108455084264278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.03369118645787239 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin in absolute darkness before being illuminated by a massive lightning bolt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hard, perfectly spherical black fungi visible clinging to a rotting, moss-covered log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fungi spheres exhibit fine cracks and a carbonaceous surface covered in slick rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow downward tilt transitioning into an extreme close-up of a heavy spherical raindrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a stark, high-contrast look with piercing whites and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735952105373144 + ], + "opposite_logprob_tokens": [ + -5.006735801696777 + ], + "logprob_reward": -0.006735952105373144, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735952105373144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12697000801563263, + 0.3280370533466339, + 0.06872495263814926, + 0.03467968851327896, + 0.023767217993736267, + 0.07495740801095963, + 0.08491697907447815, + 0.045760173350572586, + 0.1402142196893692, + 0.07197228074073792 + ], + "credit_map": [ + 0.021513819694519043, + 0.015997502952814102, + 0.009246746078133583, + 0.02110847644507885, + 0.006507865618914366, + 0.015608379617333412, + 0.020039472728967667, + 0.016947751864790916, + 0.017964337021112442, + 0.09964808076620102, + 0.014510911889374256, + 0.011873180978000164, + 0.011412215419113636, + 0.13161346316337585, + 0.012715788558125496, + 0.028299055993556976, + 0.009386424906551838, + 0.008279305882751942, + 0.005816928576678038, + 0.005660598166286945, + 0.009964647702872753, + 0.011415539309382439, + 0.010032929480075836, + 0.008168582804501057, + 0.004961819853633642, + 0.0033298921771347523, + 0.004764782264828682, + 0.002785150660201907, + 0.005085361655801535, + 0.0035494130570441484, + 0.006548276636749506, + 0.003654991975054145, + 0.0020241255406290293, + 0.0021860196720808744, + 0.0036154137924313545, + 0.00302277528680861, + 0.0034745344892144203, + 0.003106211544945836, + 0.0033828201703727245, + 0.0029553158674389124, + 0.005236580967903137, + 0.007616578135639429, + 0.010272104293107986, + 0.011564887128770351, + 0.011309966444969177, + 0.007107103243470192, + 0.012523029930889606, + 0.009327158331871033, + 0.00681992806494236, + 0.006564499344676733, + 0.010014915838837624, + 0.012866944074630737, + 0.008244230411946774, + 0.012754078023135662, + 0.01435947511345148, + 0.013292904943227768, + 0.0037256882060319185, + 0.002192473504692316, + 0.010847567580640316, + 0.005261668004095554, + 0.002659039804711938, + 0.004005044233053923, + 0.0062928106635808945, + 0.010775882750749588, + 0.011505703441798687, + 0.015019191429018974, + 0.025932736694812775, + 0.01172001101076603, + 0.016592606902122498, + 0.029770685359835625, + 0.013668997213244438, + 0.016004296019673347, + 0.006542247254401445, + 0.01084137987345457, + 0.008331457152962685, + 0.00940957386046648, + 0.010409337468445301, + 0.008709488436579704, + 0.011960521340370178, + 0.0057682772167027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.006735952105373144 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_11/index_tensor([3044, 3044])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_11/index_tensor([3044, 3044])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..67cb2fc9ea91cbff1644a926a3411d5dd90288ae --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_11/index_tensor([3044, 3044])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, low-angle, extreme wide shot captures a vast port scene, rendered entirely in a stark black and white ink style that evokes the minimalist grandeur of traditional Su-Mei painting. Towering gantry cranes, their complex structures reduced to bold, powerful calligraphy of pure black ink, loom over a vast, serene ocean. Anchored at the water's edge is a massive container ship, its massive hull rendered as a single, solid, and featureless block of the deepest black, its immense mass anchoring the composition. The sea is a placid surface composed of diluted gray currents, seamlessly blending with the misty, white void that envelops the background and sky. A long, narrow gangway cuts a sharp diagonal line from the unseen dock toward the ship, above which a row of figures slowly and deliberately ascends. Each figure is a small, blurry blob of black ink, their individual identity completely abstracted as they steadily ascend, their forms gradually blurring and dissolving into the ethereal white mist near the top of the gangway, creating a somber and contemplative atmosphere of quiet, anonymous descent into the unknown.", + "seed": 20186, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_11/index_3044_seed_20186_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.17299201465211808, + "vlm_logprob": -0.17299201465211808, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.17299201465211808, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast port with towering gantry cranes and a massive container ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a stark black and white ink style resembling Su-Mei painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container ship's hull depicted as a single, solid, featureless block of deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sea depicted as a placid surface of diluted gray currents blending with a white misty background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small, blurry black ink figures shown ascending a long, narrow gangway that cuts a sharp diagonal line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8288331031799316 + ], + "opposite_logprob_tokens": [ + -0.5788331031799316 + ], + "logprob_reward": -0.8288331031799316, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8288331031799316, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2956365942955017, + 0.13785484433174133, + 0.12039262801408768, + 0.06951429694890976, + 0.05240605026483536, + 0.06396561115980148, + 0.03146204724907875, + 0.02820999175310135, + 0.16170455515384674, + 0.038853369653224945 + ], + "credit_map": [ + 0.04812188819050789, + 0.04326990991830826, + 0.02560385689139366, + 0.04163053631782532, + 0.022273855283856392, + 0.030925458297133446, + 0.04634528607130051, + 0.037465792149305344, + 0.029943203553557396, + 0.025301173329353333, + 0.011857756413519382, + 0.016148600727319717, + 0.008180654607713223, + 0.011235486716032028, + 0.019132308661937714, + 0.016055649146437645, + 0.023825613781809807, + 0.023114586248993874, + 0.012707609683275223, + 0.013705057092010975, + 0.006745365913957357, + 0.00874756183475256, + 0.020016703754663467, + 0.011530139483511448, + 0.011286379769444466, + 0.011208254843950272, + 0.008234759792685509, + 0.013706296682357788, + 0.003776878584176302, + 0.007193155586719513, + 0.00731071038171649, + 0.006797862239181995, + 0.007404832635074854, + 0.007764259818941355, + 0.0035868147388100624, + 0.006669436581432819, + 0.005204651039093733, + 0.0070047276094555855, + 0.007141457870602608, + 0.007629869505763054, + 0.010133814997971058, + 0.009898283518850803, + 0.007572590839117765, + 0.009336604736745358, + 0.004642010200768709, + 0.004168078303337097, + 0.009719491936266422, + 0.00849473848938942, + 0.0046240887604653835, + 0.004173867404460907, + 0.003289673011749983, + 0.0034500316251069307, + 0.0023917737416923046, + 0.0037342854775488377, + 0.006214244291186333, + 0.003584082005545497, + 0.003928717691451311, + 0.004256089683622122, + 0.002514271531254053, + 0.0033857219386845827, + 0.0020451275631785393, + 0.00299322628416121, + 0.00372441322542727, + 0.0053624254651367664, + 0.018888304010033607, + 0.01879681646823883, + 0.009709862992167473, + 0.011112780310213566, + 0.02269923873245716, + 0.028589092195034027, + 0.01752706989645958, + 0.03438138961791992, + 0.006571759469807148, + 0.0059021636843681335, + 0.0042894273065030575, + 0.00481006084010005, + 0.0026699055451899767, + 0.002992834197357297, + 0.006633738987147808, + 0.004983481485396624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.8288331031799316 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast port with towering gantry cranes and a massive container ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a stark black and white ink style resembling Su-Mei painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container ship's hull depicted as a single, solid, featureless block of deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sea depicted as a placid surface of diluted gray currents blending with a white misty background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small, blurry black ink figures shown ascending a long, narrow gangway that cuts a sharp diagonal line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014202704653143883 + ], + "opposite_logprob_tokens": [ + -4.26420259475708 + ], + "logprob_reward": -0.014202704653143883, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202704653143883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19896700978279114, + 0.13592255115509033, + 0.1490478813648224, + 0.1634332239627838, + 0.0509733110666275, + 0.07364223897457123, + 0.08008483052253723, + 0.06427869945764542, + 0.04885678365826607, + 0.034793462604284286 + ], + "credit_map": [ + 0.04369983449578285, + 0.03878859803080559, + 0.021433817222714424, + 0.02223697304725647, + 0.011994443833827972, + 0.013703975826501846, + 0.025404488667845726, + 0.021704867482185364, + 0.04169873148202896, + 0.03613080084323883, + 0.01132765132933855, + 0.01348535344004631, + 0.005597441922873259, + 0.006696332711726427, + 0.01226525753736496, + 0.00872097723186016, + 0.02387850172817707, + 0.024198779836297035, + 0.011972911655902863, + 0.018205221742391586, + 0.011615990661084652, + 0.01217229850590229, + 0.03375430032610893, + 0.01324987132102251, + 0.042206380516290665, + 0.029183927923440933, + 0.010889163240790367, + 0.01914564147591591, + 0.01132984273135662, + 0.011679806746542454, + 0.021102424710989, + 0.01789604127407074, + 0.010277523659169674, + 0.007930648513138294, + 0.0049484544433653355, + 0.004576747305691242, + 0.004763250704854727, + 0.004744336474686861, + 0.006913617718964815, + 0.006818730384111404, + 0.011446291580796242, + 0.011389199644327164, + 0.010815267451107502, + 0.009886366315186024, + 0.0047509027644991875, + 0.005671576131135225, + 0.012072390876710415, + 0.007610241416841745, + 0.01306605152785778, + 0.012191271409392357, + 0.013240823522210121, + 0.010602065362036228, + 0.004492877516895533, + 0.004893228877335787, + 0.010655694641172886, + 0.01094281766563654, + 0.009016895666718483, + 0.00985860638320446, + 0.01054278016090393, + 0.009231515228748322, + 0.00403736112639308, + 0.005244286730885506, + 0.008152090013027191, + 0.008195167407393456, + 0.006085698492825031, + 0.006301463581621647, + 0.005846074782311916, + 0.005024985875934362, + 0.005162730347365141, + 0.005618279799818993, + 0.006334718782454729, + 0.008482835255563259, + 0.00523820286616683, + 0.004548533353954554, + 0.003299651201814413, + 0.003152797929942608, + 0.003480442799627781, + 0.003664029063656926, + 0.006631176453083754, + 0.004778633825480938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.014202704653143883 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast port with towering gantry cranes and a massive container ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a stark black and white ink style resembling Su-Mei painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container ship's hull depicted as a single, solid, featureless block of deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sea depicted as a placid surface of diluted gray currents blending with a white misty background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small, blurry black ink figures shown ascending a long, narrow gangway that cuts a sharp diagonal line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052548618987202644 + ], + "opposite_logprob_tokens": [ + -5.255254745483398 + ], + "logprob_reward": -0.0052548618987202644, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052548618987202644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15179234743118286, + 0.08822613954544067, + 0.22720010578632355, + 0.09867656975984573, + 0.1317075937986374, + 0.1211981549859047, + 0.03682313486933708, + 0.05001575127243996, + 0.03444144129753113, + 0.05991876497864723 + ], + "credit_map": [ + 0.021300503984093666, + 0.018024049699306488, + 0.009801068343222141, + 0.017778122797608376, + 0.015889329835772514, + 0.02316409908235073, + 0.020731113851070404, + 0.025104064494371414, + 0.015676507726311684, + 0.015158329159021378, + 0.005831618327647448, + 0.0077177779749035835, + 0.010361732915043831, + 0.00878249853849411, + 0.011805204674601555, + 0.012892469763755798, + 0.054613225162029266, + 0.0516473725438118, + 0.01083222683519125, + 0.017643185332417488, + 0.021086890250444412, + 0.02392098680138588, + 0.026654494926333427, + 0.0208017248660326, + 0.020601682364940643, + 0.017100157216191292, + 0.007978481240570545, + 0.01061188243329525, + 0.008534988388419151, + 0.012621317990124226, + 0.01012794766575098, + 0.011100110597908497, + 0.024869464337825775, + 0.024387789890170097, + 0.006775094661861658, + 0.009812957607209682, + 0.011892382986843586, + 0.017296645790338516, + 0.020807189866900444, + 0.015866057947278023, + 0.01938946358859539, + 0.01941770128905773, + 0.007269213907420635, + 0.009888717904686928, + 0.008635045029222965, + 0.016984887421131134, + 0.018658511340618134, + 0.02095460705459118, + 0.007538316771388054, + 0.006527833640575409, + 0.003138132393360138, + 0.003160207299515605, + 0.0030765559058636427, + 0.003499714657664299, + 0.004188988823443651, + 0.005693383980542421, + 0.00822269544005394, + 0.007290700916200876, + 0.004030108451843262, + 0.004296244587749243, + 0.004999736789613962, + 0.006661009509116411, + 0.0054578399285674095, + 0.009057416580617428, + 0.005173766985535622, + 0.004700231831520796, + 0.002800681395456195, + 0.0038507019635289907, + 0.0035601453855633736, + 0.00361749529838562, + 0.006335495505481958, + 0.004402924329042435, + 0.007319901138544083, + 0.007770933210849762, + 0.003996964544057846, + 0.005156899802386761, + 0.007143130525946617, + 0.009030657820403576, + 0.008752923458814621, + 0.010747353546321392 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.0052548618987202644 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast port with towering gantry cranes and a massive container ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a stark black and white ink style resembling Su-Mei painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container ship's hull depicted as a single, solid, featureless block of deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sea depicted as a placid surface of diluted gray currents blending with a white misty background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small, blurry black ink figures shown ascending a long, narrow gangway that cuts a sharp diagonal line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024839758407324553 + ], + "opposite_logprob_tokens": [ + -6.00248384475708 + ], + "logprob_reward": -0.0024839758407324553, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024839758407324553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16546441614627838, + 0.14189718663692474, + 0.08426094055175781, + 0.23853442072868347, + 0.05697682499885559, + 0.050411924719810486, + 0.10101398080587387, + 0.06532604247331619, + 0.058068860322237015, + 0.03804547339677811 + ], + "credit_map": [ + 0.02127854898571968, + 0.02268975041806698, + 0.014161047525703907, + 0.026700912043452263, + 0.016382239758968353, + 0.019481662660837173, + 0.02459573931992054, + 0.020174510776996613, + 0.04038244113326073, + 0.034172043204307556, + 0.011401203460991383, + 0.013233935460448265, + 0.006215489469468594, + 0.008990870788693428, + 0.016807623207569122, + 0.010693567804992199, + 0.022408658638596535, + 0.020425647497177124, + 0.008048782125115395, + 0.008999032899737358, + 0.004592784680426121, + 0.00468191783875227, + 0.00815342552959919, + 0.006950685288757086, + 0.03992290049791336, + 0.03789506480097771, + 0.02707710862159729, + 0.04473681002855301, + 0.014055510051548481, + 0.021321527659893036, + 0.028094850480556488, + 0.025430621579289436, + 0.011182998307049274, + 0.008540927432477474, + 0.004729763604700565, + 0.005072163417935371, + 0.0044429246336221695, + 0.00502428226172924, + 0.009543148800730705, + 0.008440612815320492, + 0.009959734044969082, + 0.008196363225579262, + 0.005466125905513763, + 0.005918402224779129, + 0.0037072990089654922, + 0.0034882929176092148, + 0.007671466562896967, + 0.006004239898175001, + 0.016474643722176552, + 0.015917761251330376, + 0.007780039682984352, + 0.006964880973100662, + 0.009898676536977291, + 0.006877669598907232, + 0.024091510102152824, + 0.01300878543406725, + 0.01393562089651823, + 0.013679725117981434, + 0.006784823257476091, + 0.004857457242906094, + 0.004808853846043348, + 0.0034541343338787556, + 0.010405177250504494, + 0.007400245871394873, + 0.008874081075191498, + 0.007968449033796787, + 0.003811942646279931, + 0.006059688050299883, + 0.006260087713599205, + 0.005778121296316385, + 0.01024593971669674, + 0.009070546366274357, + 0.006429383065551519, + 0.006752431392669678, + 0.004648106172680855, + 0.003522621700540185, + 0.002233800245448947, + 0.0034602240193635225, + 0.0058018844574689865, + 0.005197019781917334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.0024839758407324553 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast port with towering gantry cranes and a massive container ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a stark black and white ink style resembling Su-Mei painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container ship's hull depicted as a single, solid, featureless block of deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sea depicted as a placid surface of diluted gray currents blending with a white misty background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small, blurry black ink figures shown ascending a long, narrow gangway that cuts a sharp diagonal line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014185427688062191 + ], + "opposite_logprob_tokens": [ + -4.264185428619385 + ], + "logprob_reward": -0.014185427688062191, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014185427688062191, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30407947301864624, + 0.11816712468862534, + 0.09624020010232925, + 0.08599377423524857, + 0.046575915068387985, + 0.1602409929037094, + 0.0480777807533741, + 0.04647822305560112, + 0.037318695336580276, + 0.05682777240872383 + ], + "credit_map": [ + 0.045832034200429916, + 0.04273132607340813, + 0.04533836245536804, + 0.06160550191998482, + 0.024698816239833832, + 0.020947838202118874, + 0.03624492138624191, + 0.02668069861829281, + 0.024812351912260056, + 0.020907655358314514, + 0.014205828309059143, + 0.015612448565661907, + 0.007101526949554682, + 0.00775382574647665, + 0.01691448874771595, + 0.010859014466404915, + 0.026121065020561218, + 0.02466643787920475, + 0.006135636940598488, + 0.008144175633788109, + 0.006089404225349426, + 0.0070875054225325584, + 0.010718400590121746, + 0.007277580909430981, + 0.014489990659058094, + 0.01375096570700407, + 0.008633780293166637, + 0.014180835336446762, + 0.005427103955298662, + 0.006864987313747406, + 0.013872601091861725, + 0.008773518726229668, + 0.007607098203152418, + 0.00684146024286747, + 0.005264206323772669, + 0.005491912830621004, + 0.003415244398638606, + 0.006213358603417873, + 0.005998500157147646, + 0.005744141526520252, + 0.026461008936166763, + 0.02433006837964058, + 0.026685042306780815, + 0.021910300478339195, + 0.011399864219129086, + 0.010090192779898643, + 0.028060205280780792, + 0.011304317973554134, + 0.007686261553317308, + 0.007220996078103781, + 0.006665767170488834, + 0.005087033845484257, + 0.00394141860306263, + 0.005456522572785616, + 0.006695401854813099, + 0.005324384663254023, + 0.008080249652266502, + 0.0076916515827178955, + 0.0034585092216730118, + 0.0029261093586683273, + 0.004864920862019062, + 0.004655596800148487, + 0.007663368713110685, + 0.007137821055948734, + 0.006504612509161234, + 0.005377006251364946, + 0.004381542094051838, + 0.004463891498744488, + 0.0035429352428764105, + 0.002790753496810794, + 0.005172402132302523, + 0.005085553973913193, + 0.0080359922721982, + 0.007789819501340389, + 0.0072652315720915794, + 0.007140189874917269, + 0.005870988592505455, + 0.004296049475669861, + 0.00968884490430355, + 0.006740660406649113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.014185427688062191 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, extreme wide shot captures a vast port scene, rendered entirely in a stark black and white ink style that evokes the minimalist grandeur of traditional Su-Mei painting. Towering gantry cranes, their complex structures reduced to bold, powerful calligraphy of pure black ink, loom over a vast, serene ocean. Anchored at the water's edge is a massive container ship, its massive hull rendered as a single, solid, and featureless block of the deepest black, its immense mass anchoring the composition. The sea is a placid surface composed of diluted gray currents, seamlessly blending with the misty, white void that envelops the background and sky. A long, narrow gangway cuts a sharp diagonal line from the unseen dock toward the ship, above which a row of figures slowly and deliberately ascends. Each figure is a small, blurry blob of black ink, their individual identity completely abstracted as they steadily ascend, their forms gradually blurring and dissolving into the ethereal white mist near the top of the gangway, creating a somber and contemplative atmosphere of quiet, anonymous descent into the unknown.", + "seed": 20187, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_11/index_3044_seed_20187_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04532134784385562, + "vlm_logprob": -0.04532134784385562, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.04532134784385562, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast port with towering gantry cranes and a massive container ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a stark black and white ink style resembling Su-Mei painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container ship's hull depicted as a single, solid, featureless block of deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sea depicted as a placid surface of diluted gray currents blending with a white misty background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small, blurry black ink figures shown ascending a long, narrow gangway that cuts a sharp diagonal line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16743142902851105 + ], + "opposite_logprob_tokens": [ + -1.9174314737319946 + ], + "logprob_reward": -0.16743142902851105, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16743142902851105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2959096431732178, + 0.17761386930942535, + 0.04562494903802872, + 0.04866588115692139, + 0.03912283480167389, + 0.14180628955364227, + 0.0639113038778305, + 0.02123395726084709, + 0.04657332971692085, + 0.11953789740800858 + ], + "credit_map": [ + 0.03228902816772461, + 0.046997830271720886, + 0.02423113025724888, + 0.03194640949368477, + 0.05666998028755188, + 0.037188250571489334, + 0.025085538625717163, + 0.041501473635435104, + 0.025830479338765144, + 0.030127082020044327, + 0.0119624650105834, + 0.022028174251317978, + 0.025943497195839882, + 0.018043125048279762, + 0.01804465241730213, + 0.02563438192009926, + 0.005059640388935804, + 0.006523653399199247, + 0.0052377087995409966, + 0.004797866567969322, + 0.0059331683441996574, + 0.00794949010014534, + 0.004619000479578972, + 0.00550442049279809, + 0.006558084860444069, + 0.0063900938257575035, + 0.005493112839758396, + 0.004305876791477203, + 0.005978836212307215, + 0.005658063571900129, + 0.006838465575128794, + 0.007443343289196491, + 0.0039079077541828156, + 0.006395262200385332, + 0.0035586345475167036, + 0.005111621227115393, + 0.005597697105258703, + 0.004674908239394426, + 0.0041900924406945705, + 0.005686706863343716, + 0.02116602286696434, + 0.02245500683784485, + 0.012482278048992157, + 0.012492571957409382, + 0.020832637324929237, + 0.01817971281707287, + 0.017008695751428604, + 0.01718936115503311, + 0.010234730318188667, + 0.008860901929438114, + 0.005201715975999832, + 0.0055189975537359715, + 0.007206980604678392, + 0.007149037439376116, + 0.007633726578205824, + 0.012105216272175312, + 0.003885807702317834, + 0.0031271744519472122, + 0.002357736462727189, + 0.0023049511946737766, + 0.0024915016256272793, + 0.002511425642296672, + 0.0023309392854571342, + 0.0022244229912757874, + 0.007376511115580797, + 0.006478980183601379, + 0.0034107412211596966, + 0.003996321931481361, + 0.006558153312653303, + 0.006437279284000397, + 0.004856443498283625, + 0.007458899635821581, + 0.0158233679831028, + 0.01941472478210926, + 0.009369557723402977, + 0.010772552341222763, + 0.021212836727499962, + 0.016590749844908714, + 0.012514306232333183, + 0.013839798048138618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.16743142902851105 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast port with towering gantry cranes and a massive container ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a stark black and white ink style resembling Su-Mei painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container ship's hull depicted as a single, solid, featureless block of deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sea depicted as a placid surface of diluted gray currents blending with a white misty background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small, blurry black ink figures shown ascending a long, narrow gangway that cuts a sharp diagonal line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028348281048238277 + ], + "opposite_logprob_tokens": [ + -5.877834796905518 + ], + "logprob_reward": -0.0028348281048238277, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028348281048238277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17774531245231628, + 0.4761994183063507, + 0.025999698787927628, + 0.042984481900930405, + 0.08361261337995529, + 0.040915243327617645, + 0.06158706918358803, + 0.020824257284402847, + 0.027578039094805717, + 0.042553845793008804 + ], + "credit_map": [ + 0.010070235468447208, + 0.02378137782216072, + 0.020493533462285995, + 0.019088393077254295, + 0.0255904458463192, + 0.018330121412873268, + 0.022135984152555466, + 0.038255222141742706, + 0.031128663569688797, + 0.06953204423189163, + 0.0387771911919117, + 0.04015311971306801, + 0.04192008078098297, + 0.03582248464226723, + 0.07595184445381165, + 0.1429140269756317, + 0.003080913797020912, + 0.004015984944999218, + 0.001944509451277554, + 0.0022564290557056665, + 0.003483701031655073, + 0.003548188367858529, + 0.0030497002881020308, + 0.004620273131877184, + 0.004426854662597179, + 0.006272951606661081, + 0.004438979085534811, + 0.003271462395787239, + 0.006219340953975916, + 0.005727704614400864, + 0.0055372747592628, + 0.007089917082339525, + 0.009051494300365448, + 0.011347340419888496, + 0.008476345799863338, + 0.008368263021111488, + 0.01454011257737875, + 0.009895958937704563, + 0.008493037894368172, + 0.013440068811178207, + 0.003955162595957518, + 0.005819352809339762, + 0.005157927516847849, + 0.004268698859959841, + 0.005650329403579235, + 0.004913975019007921, + 0.006096824537962675, + 0.005052979104220867, + 0.007946320809423923, + 0.009578549303114414, + 0.005148847121745348, + 0.005421365145593882, + 0.00939147174358368, + 0.008100736886262894, + 0.0069138407707214355, + 0.0090859429910779, + 0.0021318825893104076, + 0.0025633268523961306, + 0.002238793531432748, + 0.0030402704142034054, + 0.0023559655528515577, + 0.0020423252135515213, + 0.002777400892227888, + 0.0036742924712598324, + 0.0031527469400316477, + 0.003462704364210367, + 0.00258458754979074, + 0.0026464511174708605, + 0.0031571316067129374, + 0.003679189132526517, + 0.004015651997178793, + 0.0048795780166983604, + 0.007449951488524675, + 0.0064447675831615925, + 0.003769604954868555, + 0.003509594826027751, + 0.005877012386918068, + 0.0047448077239096165, + 0.0042382641695439816, + 0.0065198494121432304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.0028348281048238277 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast port with towering gantry cranes and a massive container ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a stark black and white ink style resembling Su-Mei painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container ship's hull depicted as a single, solid, featureless block of deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sea depicted as a placid surface of diluted gray currents blending with a white misty background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small, blurry black ink figures shown ascending a long, narrow gangway that cuts a sharp diagonal line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004100605845451355 + ], + "opposite_logprob_tokens": [ + -5.504100799560547 + ], + "logprob_reward": -0.004100605845451355, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004100605845451355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11025787889957428, + 0.10122400522232056, + 0.056703224778175354, + 0.052202336490154266, + 0.10495199263095856, + 0.1486283391714096, + 0.23725609481334686, + 0.13819482922554016, + 0.04046393930912018, + 0.0101173659786582 + ], + "credit_map": [ + 0.008213914930820465, + 0.01394159160554409, + 0.00897104013711214, + 0.010116493329405785, + 0.011476214043796062, + 0.010512649081647396, + 0.011839115060865879, + 0.03518687188625336, + 0.00600361218675971, + 0.014069471508264542, + 0.008870433084666729, + 0.009416566230356693, + 0.011312991380691528, + 0.010189341381192207, + 0.015102093107998371, + 0.026259493082761765, + 0.0057874685153365135, + 0.008731165900826454, + 0.003905694931745529, + 0.004672464914619923, + 0.008450353518128395, + 0.0083902757614851, + 0.005222535692155361, + 0.011543270200490952, + 0.007748065982013941, + 0.0072585782036185265, + 0.004819183144718409, + 0.006089023780077696, + 0.006747275125235319, + 0.006357367616146803, + 0.007194011937826872, + 0.005988833960145712, + 0.011568130925297737, + 0.017291676253080368, + 0.009698485024273396, + 0.00860192347317934, + 0.013672378845512867, + 0.010083918459713459, + 0.012521674856543541, + 0.021513806656003, + 0.012375261634588242, + 0.01880020834505558, + 0.015004787594079971, + 0.020578818395733833, + 0.019230058416724205, + 0.018310433253645897, + 0.01806670054793358, + 0.02626209147274494, + 0.023778216913342476, + 0.03195720538496971, + 0.017353801056742668, + 0.01717587374150753, + 0.015259955078363419, + 0.024637140333652496, + 0.03891543298959732, + 0.06817849725484848, + 0.019202440977096558, + 0.018635639920830727, + 0.015360849909484386, + 0.01616033725440502, + 0.01523591112345457, + 0.013053864240646362, + 0.01766865886747837, + 0.022877130657434464, + 0.006526875775307417, + 0.005425822921097279, + 0.0037891273386776447, + 0.004012581426650286, + 0.003628491424024105, + 0.003865495789796114, + 0.006715056486427784, + 0.0065004900097846985, + 0.0013429174432531, + 0.0013275118544697762, + 0.0011346015380695462, + 0.0010634985519573092, + 0.001407983130775392, + 0.0010586363496258855, + 0.001347146462649107, + 0.0014350720448419452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.004100605845451355 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast port with towering gantry cranes and a massive container ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a stark black and white ink style resembling Su-Mei painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container ship's hull depicted as a single, solid, featureless block of deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sea depicted as a placid surface of diluted gray currents blending with a white misty background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small, blurry black ink figures shown ascending a long, narrow gangway that cuts a sharp diagonal line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003610759973526001 + ], + "opposite_logprob_tokens": [ + -5.628610610961914 + ], + "logprob_reward": -0.003610759973526001, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003610759973526001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28762996196746826, + 0.1236817017197609, + 0.06556601077318192, + 0.08528407663106918, + 0.1085960865020752, + 0.07312964648008347, + 0.061137355864048004, + 0.037385400384664536, + 0.085896797478199, + 0.07169289886951447 + ], + "credit_map": [ + 0.028963332995772362, + 0.041284915059804916, + 0.029362041503190994, + 0.024013467133045197, + 0.018633564934134483, + 0.022284697741270065, + 0.0356476753950119, + 0.08744029700756073, + 0.01456127967685461, + 0.02150907926261425, + 0.008625035174190998, + 0.011655375361442566, + 0.01999974250793457, + 0.013122153468430042, + 0.012683835811913013, + 0.02152520976960659, + 0.008465232327580452, + 0.00929795578122139, + 0.007020477671176195, + 0.0072109634056687355, + 0.007516696583479643, + 0.006441593170166016, + 0.008707067929208279, + 0.010906030423939228, + 0.009258641861379147, + 0.014187577180564404, + 0.008401819504797459, + 0.008455141447484493, + 0.010952414013445377, + 0.009587117470800877, + 0.01048686821013689, + 0.013954507187008858, + 0.007674379739910364, + 0.018407471477985382, + 0.010065692476928234, + 0.008154948242008686, + 0.01457170583307743, + 0.010845609940588474, + 0.010131435468792915, + 0.02874484844505787, + 0.013692736625671387, + 0.010769001208245754, + 0.007926181890070438, + 0.006944520398974419, + 0.008611937053501606, + 0.0086836451664567, + 0.007367383688688278, + 0.009134244173765182, + 0.0066712708212435246, + 0.008306450210511684, + 0.005610733758658171, + 0.005381216295063496, + 0.008185283280909061, + 0.009414703585207462, + 0.008354587480425835, + 0.009213119745254517, + 0.005828827619552612, + 0.005172690376639366, + 0.00391644099727273, + 0.003772524418309331, + 0.0045583597384393215, + 0.00409416388720274, + 0.004206436220556498, + 0.005835960619151592, + 0.014635004103183746, + 0.011236503720283508, + 0.007698747329413891, + 0.008012932725250721, + 0.008005931973457336, + 0.009053601883351803, + 0.013723259791731834, + 0.013530823402106762, + 0.010630742646753788, + 0.01157142873853445, + 0.005432935897260904, + 0.0056845215149223804, + 0.013270013965666294, + 0.00936640240252018, + 0.007012701127678156, + 0.008724159561097622 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.003610759973526001 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast port with towering gantry cranes and a massive container ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a stark black and white ink style resembling Su-Mei painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container ship's hull depicted as a single, solid, featureless block of deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sea depicted as a placid surface of diluted gray currents blending with a white misty background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small, blurry black ink figures shown ascending a long, narrow gangway that cuts a sharp diagonal line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048629116266965866 + ], + "opposite_logprob_tokens": [ + -3.04862904548645 + ], + "logprob_reward": -0.048629116266965866, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048629116266965866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3896426260471344, + 0.2036348581314087, + 0.07002298533916473, + 0.05898869037628174, + 0.07095356285572052, + 0.03985085338354111, + 0.06927317380905151, + 0.021364731714129448, + 0.041798125952482224, + 0.034470390528440475 + ], + "credit_map": [ + 0.034398939460515976, + 0.05415003001689911, + 0.026034953072667122, + 0.028490709140896797, + 0.02658817730844021, + 0.03613441810011864, + 0.047841429710388184, + 0.1360040009021759, + 0.022071776911616325, + 0.03252461180090904, + 0.018777864053845406, + 0.01741676777601242, + 0.017749009653925896, + 0.017152272164821625, + 0.02751624956727028, + 0.05042629688978195, + 0.009259777143597603, + 0.010228720493614674, + 0.005911062937229872, + 0.007229943294078112, + 0.011328689754009247, + 0.00959401112049818, + 0.008000502362847328, + 0.008470278233289719, + 0.008432331494987011, + 0.008086921647191048, + 0.005669726990163326, + 0.006515686400234699, + 0.008058196865022182, + 0.006478575058281422, + 0.007193977478891611, + 0.008553272113204002, + 0.006073507480323315, + 0.009659289382398129, + 0.007569052278995514, + 0.01002485677599907, + 0.008997135795652866, + 0.012466900050640106, + 0.007918259128928185, + 0.008244560100138187, + 0.003037606133148074, + 0.005964509677141905, + 0.0037597855553030968, + 0.004569540731608868, + 0.007051339838653803, + 0.0057703410275280476, + 0.005307449493557215, + 0.004390278831124306, + 0.011865516193211079, + 0.009686291217803955, + 0.005795632489025593, + 0.006857738830149174, + 0.005131466779857874, + 0.0078011141158640385, + 0.007436595391482115, + 0.014698822982609272, + 0.0020085677970200777, + 0.00291297840885818, + 0.002038385486230254, + 0.0024152144324034452, + 0.0037865866906940937, + 0.002730301581323147, + 0.002787274308502674, + 0.002685422310605645, + 0.003326167119666934, + 0.0045736790634691715, + 0.0042128004133701324, + 0.004449034109711647, + 0.0035079254303127527, + 0.003958531655371189, + 0.007642875891178846, + 0.010127107612788677, + 0.004228430334478617, + 0.00521953497081995, + 0.003453798359259963, + 0.0035600222181528807, + 0.004037552047520876, + 0.004665559623390436, + 0.004446268547326326, + 0.004859221167862415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.048629116266965866 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, extreme wide shot captures a vast port scene, rendered entirely in a stark black and white ink style that evokes the minimalist grandeur of traditional Su-Mei painting. Towering gantry cranes, their complex structures reduced to bold, powerful calligraphy of pure black ink, loom over a vast, serene ocean. Anchored at the water's edge is a massive container ship, its massive hull rendered as a single, solid, and featureless block of the deepest black, its immense mass anchoring the composition. The sea is a placid surface composed of diluted gray currents, seamlessly blending with the misty, white void that envelops the background and sky. A long, narrow gangway cuts a sharp diagonal line from the unseen dock toward the ship, above which a row of figures slowly and deliberately ascends. Each figure is a small, blurry blob of black ink, their individual identity completely abstracted as they steadily ascend, their forms gradually blurring and dissolving into the ethereal white mist near the top of the gangway, creating a somber and contemplative atmosphere of quiet, anonymous descent into the unknown.", + "seed": 20188, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_11/index_3044_seed_20188_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03586299458984286, + "vlm_logprob": -0.03586299458984286, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.03586299458984286, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast port with towering gantry cranes and a massive container ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a stark black and white ink style resembling Su-Mei painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container ship's hull depicted as a single, solid, featureless block of deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sea depicted as a placid surface of diluted gray currents blending with a white misty background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small, blurry black ink figures shown ascending a long, narrow gangway that cuts a sharp diagonal line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0947285071015358 + ], + "opposite_logprob_tokens": [ + -2.469728469848633 + ], + "logprob_reward": -0.0947285071015358, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0947285071015358, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3819429278373718, + 0.007934821769595146, + 0.5689586400985718, + 0.026585226878523827, + 0.008049044758081436, + 0.0016559191280975938, + 0.001860075630247593, + 0.0014566838508471847, + 0.001363867544569075, + 0.0001928635610966012 + ], + "credit_map": [ + 0.0014939401298761368, + 0.0029022886883467436, + 0.0013088604900985956, + 0.001160976360552013, + 0.18452511727809906, + 0.18380428850650787, + 0.002773105166852474, + 0.003974323160946369, + 0.0007388967787846923, + 0.0005541769205592573, + 0.0009594222065061331, + 0.0007461868808604777, + 0.0011408708523958921, + 0.0014421491650864482, + 0.0013343337923288345, + 0.0010187842417508364, + 0.002433937741443515, + 0.005602039862424135, + 0.0022196194622665644, + 0.0016870563849806786, + 0.2759247124195099, + 0.27441245317459106, + 0.0031874768901616335, + 0.0034912677947431803, + 0.0003512096300255507, + 0.0003041607851628214, + 0.0002774350286927074, + 0.0003850110515486449, + 0.011838288977742195, + 0.01189506147056818, + 0.000834541511721909, + 0.000699515629094094, + 0.000716270471457392, + 0.0005336934700608253, + 0.0007604787242598832, + 0.0010268292389810085, + 0.0013458298053592443, + 0.0014354068553075194, + 0.0013666662853211164, + 0.0008638694416731596, + 0.00011128118785563856, + 9.159612818621099e-05, + 0.00012898756540380418, + 0.0001330802624579519, + 0.0002487625170033425, + 0.0003064917982555926, + 0.00038075089105404913, + 0.00025496858870610595, + 0.00020170363131910563, + 0.0001600715913809836, + 0.00015020629507489502, + 0.00017866639245767146, + 0.0003451126394793391, + 0.00031916407169774175, + 0.00024761390523053706, + 0.0002575370599515736, + 0.0001740901352604851, + 0.0001433031284250319, + 0.00010871539416257292, + 0.00014680267486255616, + 0.00027156391297467053, + 0.00027575663989409804, + 0.00016943008813541383, + 0.00016702160064596683, + 7.445314986398444e-05, + 6.761043186997995e-05, + 6.219486385816708e-05, + 0.00012875364336650819, + 0.00012112359399907291, + 0.00023683578183408827, + 0.0003400716814212501, + 0.0003328243037685752, + 1.1277921657892875e-05, + 1.2709209840977564e-05, + 9.02922147361096e-06, + 1.5606363376718946e-05, + 3.253133763791993e-05, + 3.85775238100905e-05, + 2.7467473046272062e-05, + 4.5664481149287894e-05 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.0947285071015358 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast port with towering gantry cranes and a massive container ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a stark black and white ink style resembling Su-Mei painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container ship's hull depicted as a single, solid, featureless block of deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sea depicted as a placid surface of diluted gray currents blending with a white misty background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small, blurry black ink figures shown ascending a long, narrow gangway that cuts a sharp diagonal line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004649067297577858 + ], + "opposite_logprob_tokens": [ + -5.3796491622924805 + ], + "logprob_reward": -0.004649067297577858, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004649067297577858, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7827946543693542, + 0.013497795909643173, + 0.1316709667444229, + 0.044128768146038055, + 0.0030091789085417986, + 0.003986364230513573, + 0.0015623151557520032, + 0.001444468041881919, + 0.016478676348924637, + 0.0014267589431256056 + ], + "credit_map": [ + 0.00295778620056808, + 0.006031420081853867, + 0.002612380776554346, + 0.0023707624059170485, + 0.3784957826137543, + 0.376784086227417, + 0.0055455700494349, + 0.007996882311999798, + 0.001091057900339365, + 0.0007022028439678252, + 0.0010136949131265283, + 0.001146326307207346, + 0.002518898341804743, + 0.0023722369223833084, + 0.002552362624555826, + 0.002101015532389283, + 0.000705810496583581, + 0.001341724069789052, + 0.000703389523550868, + 0.0007473956211470068, + 0.062485069036483765, + 0.06224570423364639, + 0.0015642468351870775, + 0.0018776346696540713, + 0.0007741291774436831, + 0.0006126314983703196, + 0.0007086578407324851, + 0.000797468179371208, + 0.019882509484887123, + 0.019874725490808487, + 0.0006723643164150417, + 0.0008062828565016389, + 0.00017784640658646822, + 0.000296023121336475, + 0.0002545704774092883, + 0.00022955355234444141, + 0.000584208988584578, + 0.0005450242897495627, + 0.0005361326038837433, + 0.00038581923581659794, + 0.0003317056689411402, + 0.00029162922874093056, + 0.0003427744086366147, + 0.0005364279495552182, + 0.000706397695466876, + 0.0005862549296580255, + 0.0006281849928200245, + 0.0005629893857985735, + 0.00014250236563384533, + 9.124388452619314e-05, + 0.0001510591246187687, + 0.00017879056395031512, + 0.00016105611575767398, + 0.00023299336316995323, + 0.0002954346127808094, + 0.00030923515441827476, + 0.0001237636461155489, + 0.00011581833678064868, + 0.00010678456601453945, + 0.00016143967513926327, + 0.0002493876963853836, + 0.0002479481336195022, + 0.0002548929478507489, + 0.00018443305452819914, + 0.0007264536106958985, + 0.0005768770352005959, + 0.0005948222824372351, + 0.0010467388201504946, + 0.0012816416565328836, + 0.00288076582364738, + 0.004726319573819637, + 0.004645057022571564, + 0.00022018120216671377, + 0.00011183141759829596, + 0.00011015860945917666, + 9.836994286160916e-05, + 0.0001825364597607404, + 0.00024964584736153483, + 0.00019688688917085528, + 0.00025714864023029804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.004649067297577858 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast port with towering gantry cranes and a massive container ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a stark black and white ink style resembling Su-Mei painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container ship's hull depicted as a single, solid, featureless block of deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sea depicted as a placid surface of diluted gray currents blending with a white misty background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small, blurry black ink figures shown ascending a long, narrow gangway that cuts a sharp diagonal line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029796818271279335 + ], + "opposite_logprob_tokens": [ + -3.529796838760376 + ], + "logprob_reward": -0.029796818271279335, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029796818271279335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.684186577796936, + 0.010902591049671173, + 0.2662660777568817, + 0.02018950693309307, + 0.00425583915784955, + 0.003990377299487591, + 0.002862421562895179, + 0.001753598335199058, + 0.004606137052178383, + 0.0009869415080174804 + ], + "credit_map": [ + 0.002525079995393753, + 0.005405721720308065, + 0.0023461191449314356, + 0.0024063088931143284, + 0.3303787410259247, + 0.3287860155105591, + 0.005197110585868359, + 0.007141441106796265, + 0.000544697162695229, + 0.0005986570031382143, + 0.0008031337638385594, + 0.0007742697489447892, + 0.00234470097348094, + 0.0024990064557641745, + 0.0016142751555889845, + 0.001723850960843265, + 0.0014412811724469066, + 0.002666585613042116, + 0.0011954002548009157, + 0.0010861322516575456, + 0.12854629755020142, + 0.12763044238090515, + 0.00180136039853096, + 0.0018985922215506434, + 0.0005951967905275524, + 0.0003757876402232796, + 0.00047878758050501347, + 0.00047446522512473166, + 0.00838560238480568, + 0.008540813811123371, + 0.0007186036673374474, + 0.0006202516960911453, + 0.0001785510394256562, + 0.0002712249697651714, + 0.0001938051136676222, + 0.00036486235330812633, + 0.0007068496197462082, + 0.0007181080291047692, + 0.001076634624041617, + 0.0007458035252057016, + 0.00022064046061132103, + 0.0002805791737046093, + 0.00017433938046451658, + 0.0004579296801239252, + 0.0007911290158517659, + 0.000708040373865515, + 0.0008145268075168133, + 0.0005431921454146504, + 0.00019098439952358603, + 0.00013743290037382394, + 0.00023753210552968085, + 0.00029261360759846866, + 0.0003464433830231428, + 0.0004264176532160491, + 0.0006595849408768117, + 0.0005714125582017004, + 0.00016903206415008754, + 0.00015639116463717073, + 0.0001381048932671547, + 0.00021461519645527005, + 0.00028538619517348707, + 0.0003143006470054388, + 0.00023157612304203212, + 0.0002441919641569257, + 0.00018520263256505132, + 0.00015031182556413114, + 0.0001407863019267097, + 0.0003386275202501565, + 0.0003487550711724907, + 0.0008016856154426932, + 0.001346486620604992, + 0.001294281450100243, + 0.0001222902355948463, + 7.818463927833363e-05, + 7.41591866244562e-05, + 7.115628250176087e-05, + 9.635814058128744e-05, + 0.00018555625865701586, + 0.00014826816914137453, + 0.0002109686320181936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.029796818271279335 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast port with towering gantry cranes and a massive container ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a stark black and white ink style resembling Su-Mei painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container ship's hull depicted as a single, solid, featureless block of deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sea depicted as a placid surface of diluted gray currents blending with a white misty background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small, blurry black ink figures shown ascending a long, narrow gangway that cuts a sharp diagonal line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001510670525021851 + ], + "opposite_logprob_tokens": [ + -6.5015106201171875 + ], + "logprob_reward": -0.001510670525021851, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001510670525021851, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6436010003089905, + 0.0183224156498909, + 0.31512826681137085, + 0.009362471289932728, + 0.0034794246312230825, + 0.003270356683060527, + 0.002220657654106617, + 0.001966979820281267, + 0.0019933206494897604, + 0.0006551378173753619 + ], + "credit_map": [ + 0.002483429154381156, + 0.005008847918361425, + 0.0021973608527332544, + 0.0019595923367887735, + 0.31101682782173157, + 0.3095499873161316, + 0.004498066380620003, + 0.0068868896923959255, + 0.0005460408865474164, + 0.0002908048627432436, + 0.0006123704370111227, + 0.00042201566975563765, + 0.007542278151959181, + 0.007586108520627022, + 0.0006958147278055549, + 0.0006269833538681269, + 0.0014107126044109464, + 0.00306955655105412, + 0.0012538123410195112, + 0.0010259628761559725, + 0.152359738945961, + 0.151593878865242, + 0.0021982237230986357, + 0.0022163651883602142, + 0.0004445291415322572, + 0.00029086106223985553, + 0.0004020529449917376, + 0.00042746629333123565, + 0.0032803916838020086, + 0.0033496003597974777, + 0.000556917570065707, + 0.0006106520886532962, + 0.0002009821037063375, + 0.00020859397773165256, + 0.0002882599947042763, + 0.00037044877535663545, + 0.0005380038637667894, + 0.0005767581169493496, + 0.0006632317090407014, + 0.0006331459735520184, + 0.00014766839740332216, + 0.0001753729156916961, + 0.00015917005657684058, + 0.00027451125788502395, + 0.000463873177068308, + 0.0006059137522242963, + 0.0007862505735829473, + 0.0006575961597263813, + 0.00018322747200727463, + 0.0001470956631237641, + 0.00015234238526318222, + 0.00029850832652300596, + 0.00036160257877781987, + 0.0003281839599367231, + 0.00037300141411833465, + 0.00037669597077183425, + 0.0001663639268372208, + 0.00013447222590912133, + 0.00013712228974327445, + 0.00021155475405976176, + 0.0002040094550466165, + 0.00033782317768782377, + 0.00034985985257662833, + 0.00042577425483614206, + 9.104828495765105e-05, + 8.301374327857047e-05, + 7.742799789411947e-05, + 0.0001630078040761873, + 0.0002120400604326278, + 0.0003132854762952775, + 0.000560861430130899, + 0.0004926358233205974, + 5.472144766827114e-05, + 4.9060752644436434e-05, + 4.822141272597946e-05, + 5.547979162656702e-05, + 7.409428508253768e-05, + 0.00011637363058980554, + 0.00011737106251530349, + 0.00013981542724650353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.001510670525021851 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast port with towering gantry cranes and a massive container ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a stark black and white ink style resembling Su-Mei painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container ship's hull depicted as a single, solid, featureless block of deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sea depicted as a placid surface of diluted gray currents blending with a white misty background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small, blurry black ink figures shown ascending a long, narrow gangway that cuts a sharp diagonal line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04862990975379944 + ], + "opposite_logprob_tokens": [ + -3.0486299991607666 + ], + "logprob_reward": -0.04862990975379944, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04862990975379944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.975547194480896, + 0.00406312569975853, + 0.01342454832047224, + 0.0022189216688275337, + 0.0014237661380320787, + 0.0006312786717899144, + 0.0007155546918511391, + 0.0006754316855221987, + 0.0010146411368623376, + 0.000285475718555972 + ], + "credit_map": [ + 0.0037575357127934694, + 0.007497481536120176, + 0.003293975256383419, + 0.0029696538113057613, + 0.4716036915779114, + 0.4695180058479309, + 0.006791772786527872, + 0.010115215554833412, + 0.00010481352364877239, + 7.489482959499583e-05, + 0.0001022576616378501, + 0.00011536447709659114, + 0.0017053891206160188, + 0.0016845116624608636, + 0.0001354372507194057, + 0.00014045735588297248, + 7.720432040514424e-05, + 0.00013993783795740455, + 0.00011204833572264761, + 8.985956810647622e-05, + 0.006365605629980564, + 0.0063219028525054455, + 0.00015248714771587402, + 0.0001655044761719182, + 3.9977985579753295e-05, + 3.054381886613555e-05, + 3.752506017917767e-05, + 3.9930389903020114e-05, + 0.0009630277054384351, + 0.0009840804850682616, + 6.652561569353566e-05, + 5.731082637794316e-05, + 0.00010485998791409656, + 7.635894871782511e-05, + 0.0001247962936758995, + 0.00017549873155076057, + 0.00018595735309645534, + 0.00020946876611560583, + 0.00029737973818555474, + 0.00024944645701907575, + 4.361234823591076e-05, + 3.5973149351775646e-05, + 5.074282307759859e-05, + 6.310064782155678e-05, + 9.436040272703394e-05, + 0.00011404515680624172, + 0.00013975804904475808, + 8.968618203653023e-05, + 4.50948100478854e-05, + 3.715278216986917e-05, + 4.1816783777903765e-05, + 7.475771417375654e-05, + 0.00010057322651846334, + 0.00011262564657954499, + 0.00014361519424710423, + 0.00015991863620001823, + 7.86937162047252e-05, + 5.501624764292501e-05, + 4.436700328369625e-05, + 5.926109224674292e-05, + 9.433470404474065e-05, + 0.00012250294093973935, + 0.00011362915392965078, + 0.00010762691090349108, + 4.316044214647263e-05, + 3.941941031371243e-05, + 3.3398820960428566e-05, + 8.329188131028786e-05, + 9.823839354794472e-05, + 0.00017564749578014016, + 0.00027494042296893895, + 0.0002665444044396281, + 3.6602090403903276e-05, + 2.3200618670671247e-05, + 2.098481490975246e-05, + 2.0286212020437233e-05, + 3.602063952712342e-05, + 4.960863952874206e-05, + 4.0674425690667704e-05, + 5.8098285080632195e-05 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.04862990975379944 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, extreme wide shot captures a vast port scene, rendered entirely in a stark black and white ink style that evokes the minimalist grandeur of traditional Su-Mei painting. Towering gantry cranes, their complex structures reduced to bold, powerful calligraphy of pure black ink, loom over a vast, serene ocean. Anchored at the water's edge is a massive container ship, its massive hull rendered as a single, solid, and featureless block of the deepest black, its immense mass anchoring the composition. The sea is a placid surface composed of diluted gray currents, seamlessly blending with the misty, white void that envelops the background and sky. A long, narrow gangway cuts a sharp diagonal line from the unseen dock toward the ship, above which a row of figures slowly and deliberately ascends. Each figure is a small, blurry blob of black ink, their individual identity completely abstracted as they steadily ascend, their forms gradually blurring and dissolving into the ethereal white mist near the top of the gangway, creating a somber and contemplative atmosphere of quiet, anonymous descent into the unknown.", + "seed": 20189, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_11/index_3044_seed_20189_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09508895436301827, + "vlm_logprob": -0.09508895436301827, + "vlm_yes_no_margin": 4.149999988079071, + "vlm_token_credit": -0.09508895436301827, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast port with towering gantry cranes and a massive container ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a stark black and white ink style resembling Su-Mei painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container ship's hull depicted as a single, solid, featureless block of deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sea depicted as a placid surface of diluted gray currents blending with a white misty background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small, blurry black ink figures shown ascending a long, narrow gangway that cuts a sharp diagonal line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.43478304147720337 + ], + "opposite_logprob_tokens": [ + -1.0597829818725586 + ], + "logprob_reward": -0.43478304147720337, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.43478304147720337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3468921482563019, + 0.2176794409751892, + 0.05646449699997902, + 0.1663685292005539, + 0.048977334052324295, + 0.039362359791994095, + 0.03330552950501442, + 0.025179067626595497, + 0.052540864795446396, + 0.013230272568762302 + ], + "credit_map": [ + 0.027015477418899536, + 0.058455225080251694, + 0.03185763210058212, + 0.05071519315242767, + 0.03812342509627342, + 0.05085686221718788, + 0.04844944179058075, + 0.041418880224227905, + 0.026529736816883087, + 0.032429393380880356, + 0.025647390633821487, + 0.024257272481918335, + 0.02095237746834755, + 0.029151955619454384, + 0.032351814210414886, + 0.026359498500823975, + 0.00753316143527627, + 0.009849393740296364, + 0.005046759266406298, + 0.006788762751966715, + 0.006246975623071194, + 0.008274651132524014, + 0.00596965616568923, + 0.006755134556442499, + 0.018432699143886566, + 0.02798987179994583, + 0.012478617019951344, + 0.022388454526662827, + 0.018825167790055275, + 0.025324638932943344, + 0.015512123703956604, + 0.025416947901248932, + 0.0073554376140236855, + 0.007015698589384556, + 0.004698272328823805, + 0.006631120573729277, + 0.005473885219544172, + 0.006714674178510904, + 0.0042313034646213055, + 0.006856939755380154, + 0.003539593890309334, + 0.005472081247717142, + 0.003521648468449712, + 0.005564241670072079, + 0.0037581685464829206, + 0.005989486817270517, + 0.005146104842424393, + 0.006371036171913147, + 0.0045873080380260944, + 0.005427214317023754, + 0.00291601219214499, + 0.004269586876034737, + 0.003512986935675144, + 0.004973221570253372, + 0.002993353409692645, + 0.0046258484944701195, + 0.002101185265928507, + 0.004032352473586798, + 0.0018849929329007864, + 0.003765638917684555, + 0.002650628099218011, + 0.003852690104395151, + 0.0030163894407451153, + 0.003875190392136574, + 0.0049906400963664055, + 0.00777829997241497, + 0.005856821313500404, + 0.007325449958443642, + 0.00533330300822854, + 0.007851309143006802, + 0.004523861687630415, + 0.008881178684532642, + 0.0012356856605038047, + 0.0023516768123954535, + 0.0008282251073978841, + 0.0012894320534542203, + 0.0016099397325888276, + 0.0023171924985945225, + 0.0015524547779932618, + 0.002045665867626667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.43478304147720337 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast port with towering gantry cranes and a massive container ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a stark black and white ink style resembling Su-Mei painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container ship's hull depicted as a single, solid, featureless block of deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sea depicted as a placid surface of diluted gray currents blending with a white misty background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small, blurry black ink figures shown ascending a long, narrow gangway that cuts a sharp diagonal line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003204688895493746 + ], + "opposite_logprob_tokens": [ + -5.753204822540283 + ], + "logprob_reward": -0.003204688895493746, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003204688895493746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39683282375335693, + 0.1099400669336319, + 0.037018515169620514, + 0.06288638710975647, + 0.06175268441438675, + 0.05582990497350693, + 0.1331857442855835, + 0.02428666688501835, + 0.0606759637594223, + 0.0575912743806839 + ], + "credit_map": [ + 0.032296326011419296, + 0.05274005979299545, + 0.09660425782203674, + 0.03968527913093567, + 0.03483893722295761, + 0.04663821682333946, + 0.031537555158138275, + 0.062492188066244125, + 0.009926003403961658, + 0.014801966957747936, + 0.023222440853714943, + 0.012503384612500668, + 0.009927199222147465, + 0.013805839233100414, + 0.008925286121666431, + 0.01682795025408268, + 0.004468642175197601, + 0.006549840327352285, + 0.002871705451980233, + 0.003759060986340046, + 0.00343795376829803, + 0.005984565243124962, + 0.0038430679123848677, + 0.006103677209466696, + 0.006449784152209759, + 0.009248754009604454, + 0.011342265643179417, + 0.00621348200365901, + 0.004596417304128408, + 0.009177259169518948, + 0.005721596069633961, + 0.010136829689145088, + 0.006313881371170282, + 0.00877537950873375, + 0.010865597985684872, + 0.0066587417386472225, + 0.0042650834657251835, + 0.008209869265556335, + 0.006425834260880947, + 0.01023829635232687, + 0.007086306344717741, + 0.008798626251518726, + 0.006479873787611723, + 0.0052878232672810555, + 0.006494063884019852, + 0.008542811498045921, + 0.004931289702653885, + 0.008209114894270897, + 0.017056047916412354, + 0.021495122462511063, + 0.01485666073858738, + 0.011958758346736431, + 0.012153123505413532, + 0.021107135340571404, + 0.009792093187570572, + 0.02476680278778076, + 0.0026676326524466276, + 0.004700284916907549, + 0.001947686425410211, + 0.0024292466696351767, + 0.002780705923214555, + 0.004696502350270748, + 0.002081615850329399, + 0.0029829922132194042, + 0.007901489734649658, + 0.01128113828599453, + 0.006107357330620289, + 0.0055408598855137825, + 0.0049430434592068195, + 0.010345440357923508, + 0.004938397090882063, + 0.009618238545954227, + 0.006784520577639341, + 0.011424146592617035, + 0.005490356124937534, + 0.0038734611589461565, + 0.004441617522388697, + 0.009631789289414883, + 0.004032990895211697, + 0.011912394315004349 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.003204688895493746 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast port with towering gantry cranes and a massive container ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a stark black and white ink style resembling Su-Mei painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container ship's hull depicted as a single, solid, featureless block of deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sea depicted as a placid surface of diluted gray currents blending with a white misty background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small, blurry black ink figures shown ascending a long, narrow gangway that cuts a sharp diagonal line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008641350083053112 + ], + "opposite_logprob_tokens": [ + -4.758641242980957 + ], + "logprob_reward": -0.008641350083053112, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008641350083053112, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21685171127319336, + 0.21105563640594482, + 0.1372101753950119, + 0.047167472541332245, + 0.07478504627943039, + 0.05594958737492561, + 0.038094662129879, + 0.04921542480587959, + 0.06516172736883163, + 0.10450860857963562 + ], + "credit_map": [ + 0.016045687720179558, + 0.030558383092284203, + 0.026635009795427322, + 0.025815878063440323, + 0.018118346109986305, + 0.034929968416690826, + 0.025656556710600853, + 0.03909188136458397, + 0.019248858094215393, + 0.03750403970479965, + 0.022759893909096718, + 0.020058313384652138, + 0.02532130479812622, + 0.03459447994828224, + 0.02198197692632675, + 0.02958676405251026, + 0.011414538137614727, + 0.020385032519698143, + 0.011824732646346092, + 0.014833131805062294, + 0.01862688921391964, + 0.022725792601704597, + 0.016962917521595955, + 0.020437143743038177, + 0.00430777482688427, + 0.007641556207090616, + 0.004558331798762083, + 0.0055544935166835785, + 0.00502531323581934, + 0.007514145225286484, + 0.0052701509557664394, + 0.007295710034668446, + 0.006734019611030817, + 0.009375283494591713, + 0.006521240808069706, + 0.00994971115142107, + 0.008139939047396183, + 0.011169138364493847, + 0.011305933818221092, + 0.0115897785872221, + 0.00636232690885663, + 0.007506715599447489, + 0.004517263732850552, + 0.007100156508386135, + 0.006798166316002607, + 0.009142215363681316, + 0.006412320304661989, + 0.00811042170971632, + 0.005702091380953789, + 0.005325042176991701, + 0.0035209255293011665, + 0.0033757290802896023, + 0.0033529086504131556, + 0.006686768960207701, + 0.0033947851043194532, + 0.006736410781741142, + 0.00484002660959959, + 0.009794321842491627, + 0.0038282633759081364, + 0.004162605851888657, + 0.0059725623577833176, + 0.009033284150063992, + 0.005836800206452608, + 0.005747558083385229, + 0.003463463857769966, + 0.007024307269603014, + 0.005912997294217348, + 0.008956940844655037, + 0.007236397359520197, + 0.009712855331599712, + 0.012040645815432072, + 0.010814116336405277, + 0.0075773620046675205, + 0.014761003665626049, + 0.005940655712038279, + 0.00915362685918808, + 0.009803642518818378, + 0.017706656828522682, + 0.016354547813534737, + 0.02321111597120762 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.008641350083053112 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast port with towering gantry cranes and a massive container ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a stark black and white ink style resembling Su-Mei painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container ship's hull depicted as a single, solid, featureless block of deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sea depicted as a placid surface of diluted gray currents blending with a white misty background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small, blurry black ink figures shown ascending a long, narrow gangway that cuts a sharp diagonal line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00248540285974741 + ], + "opposite_logprob_tokens": [ + -6.002485275268555 + ], + "logprob_reward": -0.00248540285974741, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.00248540285974741, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3097946345806122, + 0.13898201286792755, + 0.08351738005876541, + 0.09167131036520004, + 0.06357461214065552, + 0.0810747966170311, + 0.04164627566933632, + 0.06434508413076401, + 0.062403466552495956, + 0.06299033761024475 + ], + "credit_map": [ + 0.022936107590794563, + 0.03620601445436478, + 0.044840820133686066, + 0.036508336663246155, + 0.03244668245315552, + 0.043820612132549286, + 0.036417849361896515, + 0.05661822482943535, + 0.010840167291462421, + 0.017893705517053604, + 0.021137872710824013, + 0.014392011798918247, + 0.017448395490646362, + 0.022489355877041817, + 0.015045477077364922, + 0.019735045731067657, + 0.007424593903124332, + 0.010446637868881226, + 0.008621768094599247, + 0.007420036010444164, + 0.011411715298891068, + 0.012958669103682041, + 0.011879323981702328, + 0.01335464883595705, + 0.007161523215472698, + 0.012342635542154312, + 0.006520729977637529, + 0.011104948818683624, + 0.011411021463572979, + 0.015007480047643185, + 0.011609204113483429, + 0.016513776034116745, + 0.00813966989517212, + 0.010517575778067112, + 0.005102983675897121, + 0.005681244656443596, + 0.007192936725914478, + 0.009698115289211273, + 0.006754446309059858, + 0.01048764493316412, + 0.013569124974310398, + 0.012931851670145988, + 0.0073767416179180145, + 0.007910843938589096, + 0.009241116233170033, + 0.011922251433134079, + 0.00690108397975564, + 0.011221791617572308, + 0.005087180063128471, + 0.006560209207236767, + 0.00509680574759841, + 0.004334162920713425, + 0.0041120098903775215, + 0.006541996728628874, + 0.002949435031041503, + 0.006964483764022589, + 0.006439072545617819, + 0.013053684495389462, + 0.004253311548382044, + 0.006205721292644739, + 0.006210269406437874, + 0.0108633516356349, + 0.007477107457816601, + 0.00984257273375988, + 0.006096556317061186, + 0.012106833048164845, + 0.004957590252161026, + 0.007188402581959963, + 0.005739372223615646, + 0.011510168202221394, + 0.005433837883174419, + 0.009370715357363224, + 0.004911152180284262, + 0.013132219202816486, + 0.0043776314705610275, + 0.0056543899700045586, + 0.006320571061223745, + 0.010841980576515198, + 0.008370582945644855, + 0.009381813928484917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.00248540285974741 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast port with towering gantry cranes and a massive container ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a stark black and white ink style resembling Su-Mei painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container ship's hull depicted as a single, solid, featureless block of deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sea depicted as a placid surface of diluted gray currents blending with a white misty background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small, blurry black ink figures shown ascending a long, narrow gangway that cuts a sharp diagonal line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026330288499593735 + ], + "opposite_logprob_tokens": [ + -3.6513302326202393 + ], + "logprob_reward": -0.026330288499593735, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026330288499593735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41072511672973633, + 0.13563013076782227, + 0.05927325412631035, + 0.08120191842317581, + 0.07219751924276352, + 0.05994333699345589, + 0.05569090694189072, + 0.035947900265455246, + 0.05190234258770943, + 0.03748755529522896 + ], + "credit_map": [ + 0.044388677924871445, + 0.050446562469005585, + 0.08059091866016388, + 0.033170126378536224, + 0.03656570985913277, + 0.053415149450302124, + 0.046894852072000504, + 0.0652531161904335, + 0.01615392416715622, + 0.02789166383445263, + 0.01185009628534317, + 0.01364374253898859, + 0.012024945579469204, + 0.02294314280152321, + 0.010495832189917564, + 0.02062677964568138, + 0.006830176338553429, + 0.011325298808515072, + 0.007357455790042877, + 0.005303007084876299, + 0.006512015592306852, + 0.007916740141808987, + 0.007048868574202061, + 0.006979693658649921, + 0.007992014288902283, + 0.013206282630562782, + 0.008358165621757507, + 0.010584615170955658, + 0.008037823252379894, + 0.011871027760207653, + 0.007551474496722221, + 0.01360052265226841, + 0.009264416992664337, + 0.008936313912272453, + 0.009392966516315937, + 0.007896851748228073, + 0.006914226338267326, + 0.00998225249350071, + 0.008513838052749634, + 0.011296645738184452, + 0.006530182436108589, + 0.008176698349416256, + 0.005105038173496723, + 0.007699965965002775, + 0.006540692877024412, + 0.009747756645083427, + 0.0057033030316233635, + 0.010439705103635788, + 0.006413973402231932, + 0.009990100748836994, + 0.00517891813069582, + 0.006042611785233021, + 0.004925296641886234, + 0.009462971240282059, + 0.004760520067065954, + 0.008916518650949001, + 0.004981882870197296, + 0.006038484629243612, + 0.0032480694353580475, + 0.003921739757061005, + 0.0037473118863999844, + 0.006504566874355078, + 0.00261400593444705, + 0.004891841672360897, + 0.005016591865569353, + 0.01178770698606968, + 0.004133432172238827, + 0.0051293279975652695, + 0.004839171655476093, + 0.009624387137591839, + 0.0046958294697105885, + 0.006675897166132927, + 0.0044583845883607864, + 0.006245293188840151, + 0.0034803152084350586, + 0.003150625852867961, + 0.0030693712178617716, + 0.006291908677667379, + 0.005161499138921499, + 0.00563016114756465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.026330288499593735 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_12/index_tensor([5516, 5516])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_12/index_tensor([5516, 5516])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a412bf8fe34c65e021c0408903d39044eab95dd5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_12/index_tensor([5516, 5516])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "wide shot, initially static on a busy, turn-of-the-century train platform, begins its steady rightward movement, employing subtle handheld jerks, synchronized with the rhythmic click of wheels on the track. The central subject is a majestic, old-world steam train, its locomotive a behemoth of polished black steel and gleaming brass fittings, powerfully expelling a thick plume of white steam that catches the afternoon sun. The carriages, painted a deep, rich burgundy with delicate gold leaf pinstripes, begin to glide forward, their massive drive wheels turning with a deliberate, powerful rhythm. As the camera pans alongside, the bustling platform is revealed in detail: ornate cast-iron columns support a vast vaulted glass ceiling, through which the warm golden light of the afternoon sun filters, piercing the hazy air. Crowds of people dressed in period costume—men in tweed jackets and bowler hats, women in long skirts and elegant hats—wave goodbye, their faces a mixture of excitement and poignancy. The low-angled sunlight casts long shadows across the wet cobblestone platform, and the train’s metallic surface gleams, lending the entire scene an air of nostalgia and grandeur, a cinematic ode to a bygone age of travel, charged with the energy of a departure as the train steadily gathers speed and moves away from the camera’s field of view.", + "seed": 7486, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_12/index_5516_seed_7486_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03882674118503928, + "vlm_logprob": -0.03882674118503928, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.03882674118503928, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large steam train with a black steel locomotive and burgundy carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a thick plume of white steam that catches the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crowds of people in period costumes (tweed jackets, bowler hats, long skirts) waving goodbye on the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a busy train platform with ornate cast-iron columns and a vaulted glass ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then move steadily to the right with subtle handheld jerks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018475115299224854 + ], + "opposite_logprob_tokens": [ + -4.39347505569458 + ], + "logprob_reward": -0.018475115299224854, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.018475115299224854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1295294314622879, + 0.3214157223701477, + 0.06602250784635544, + 0.09558144956827164, + 0.04478674381971359, + 0.03553503006696701, + 0.20892244577407837, + 0.05239285156130791, + 0.023033903911709785, + 0.022779913619160652 + ], + "credit_map": [ + 0.013150792568922043, + 0.011569640599191189, + 0.01719902642071247, + 0.021506009623408318, + 0.016556184738874435, + 0.01665840670466423, + 0.020017500966787338, + 0.01287186797708273, + 0.03287532925605774, + 0.046011339873075485, + 0.023004522547125816, + 0.049967896193265915, + 0.033344946801662445, + 0.05041145160794258, + 0.026976825669407845, + 0.05882341042160988, + 0.00604407861828804, + 0.010670707561075687, + 0.006952039897441864, + 0.005688998382538557, + 0.0061116605065763, + 0.008450159803032875, + 0.009410052560269833, + 0.01269481424242258, + 0.012102981097996235, + 0.016921866685152054, + 0.005614453926682472, + 0.007881445810198784, + 0.009542020969092846, + 0.020191920921206474, + 0.010294229723513126, + 0.013032526709139347, + 0.003846112871542573, + 0.005676766391843557, + 0.005043432582169771, + 0.006023881956934929, + 0.0038933425676077604, + 0.006257401313632727, + 0.007206382695585489, + 0.006839423440396786, + 0.003212888026610017, + 0.005245862063020468, + 0.0030968808569014072, + 0.004239206202328205, + 0.0025035287253558636, + 0.005124786403030157, + 0.006639573723077774, + 0.005472302436828613, + 0.01444252859801054, + 0.04185906797647476, + 0.008497055619955063, + 0.024193089455366135, + 0.00927130039781332, + 0.013990907929837704, + 0.027701880782842636, + 0.06896660476922989, + 0.005004059989005327, + 0.009630868211388588, + 0.0033024728763848543, + 0.003984978422522545, + 0.0026786732487380505, + 0.005409648641943932, + 0.007923895493149757, + 0.014458253048360348, + 0.002041162922978401, + 0.0032408868428319693, + 0.002684494014829397, + 0.0027149261441081762, + 0.0018483569147065282, + 0.0036964856553822756, + 0.0037427276838570833, + 0.0030648650135844946, + 0.0026775829028338194, + 0.0027873851358890533, + 0.003736352315172553, + 0.0030894961673766375, + 0.001301532844081521, + 0.0020315460860729218, + 0.004547790624201298, + 0.0026082294061779976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.018475115299224854 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large steam train with a black steel locomotive and burgundy carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a thick plume of white steam that catches the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crowds of people in period costumes (tweed jackets, bowler hats, long skirts) waving goodbye on the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a busy train platform with ornate cast-iron columns and a vaulted glass ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then move steadily to the right with subtle handheld jerks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026354670524597168 + ], + "opposite_logprob_tokens": [ + -3.6513547897338867 + ], + "logprob_reward": -0.026354670524597168, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026354670524597168, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20544764399528503, + 0.3660280406475067, + 0.03786279633641243, + 0.142835795879364, + 0.04820394888520241, + 0.025478705763816833, + 0.10168354958295822, + 0.028781065717339516, + 0.026431957259774208, + 0.017246508970856667 + ], + "credit_map": [ + 0.015512816607952118, + 0.022093217819929123, + 0.01953219063580036, + 0.034715551882982254, + 0.027995342388749123, + 0.03239738196134567, + 0.029567448422312737, + 0.02363368682563305, + 0.027707938104867935, + 0.07238193601369858, + 0.03180789202451706, + 0.05728127807378769, + 0.03941032662987709, + 0.048781394958496094, + 0.04082794860005379, + 0.04782930016517639, + 0.004928139969706535, + 0.0035669959615916014, + 0.0036248494870960712, + 0.004970930051058531, + 0.00462480029091239, + 0.004729779437184334, + 0.005858594086021185, + 0.005558702163398266, + 0.01530240848660469, + 0.027015917003154755, + 0.006837231107056141, + 0.011949464678764343, + 0.013649160973727703, + 0.03297701105475426, + 0.016211429610848427, + 0.018893158063292503, + 0.005331662483513355, + 0.00751524930819869, + 0.004623353481292725, + 0.006719436030834913, + 0.005108882673084736, + 0.005898758769035339, + 0.006594995968043804, + 0.006411605514585972, + 0.003060647053644061, + 0.002893239026889205, + 0.002791706472635269, + 0.004865043330937624, + 0.0015584119828417897, + 0.0030570784583687782, + 0.004047565162181854, + 0.0032050120644271374, + 0.008891931734979153, + 0.011420346796512604, + 0.00816295575350523, + 0.026402585208415985, + 0.004986381623893976, + 0.007324216887354851, + 0.014736508950591087, + 0.019758611917495728, + 0.0032591361086815596, + 0.0037588998675346375, + 0.0018862157594412565, + 0.002710909815505147, + 0.0019384059123694897, + 0.004007306881248951, + 0.004998419899493456, + 0.006221767980605364, + 0.002139393240213394, + 0.0033498441334813833, + 0.0022697437088936567, + 0.00346032134257257, + 0.0020114630460739136, + 0.004902878776192665, + 0.004154873080551624, + 0.004143437836319208, + 0.0016856063157320023, + 0.002519947476685047, + 0.0023493692278862, + 0.0025650449097156525, + 0.0015938419383019209, + 0.0020352595020085573, + 0.0025392358656972647, + 0.001958201639354229 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.026354670524597168 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large steam train with a black steel locomotive and burgundy carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a thick plume of white steam that catches the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crowds of people in period costumes (tweed jackets, bowler hats, long skirts) waving goodbye on the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a busy train platform with ornate cast-iron columns and a vaulted glass ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then move steadily to the right with subtle handheld jerks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032045841217041 + ], + "opposite_logprob_tokens": [ + -2.350320339202881 + ], + "logprob_reward": -0.10032045841217041, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032045841217041, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19222156703472137, + 0.2008058875799179, + 0.07449711114168167, + 0.0845172330737114, + 0.14482207596302032, + 0.036359332501888275, + 0.12723158299922943, + 0.06898427754640579, + 0.03126140311360359, + 0.03929945454001427 + ], + "credit_map": [ + 0.020549312233924866, + 0.021397704258561134, + 0.01776071824133396, + 0.024919703602790833, + 0.024651197716593742, + 0.029395321384072304, + 0.026705754920840263, + 0.026841850951313972, + 0.020809179171919823, + 0.02324545383453369, + 0.032242100685834885, + 0.03802685812115669, + 0.011994614265859127, + 0.013680018484592438, + 0.019497808068990707, + 0.04130987077951431, + 0.014050036668777466, + 0.008529994636774063, + 0.008318167179822922, + 0.008865593001246452, + 0.005761959124356508, + 0.010068366304039955, + 0.008023271337151527, + 0.010879726149141788, + 0.009418826550245285, + 0.011320117861032486, + 0.006445269100368023, + 0.009661747142672539, + 0.008316273801028728, + 0.015200473368167877, + 0.01091640256345272, + 0.013238124549388885, + 0.018035225570201874, + 0.0198034830391407, + 0.0216580368578434, + 0.017891615629196167, + 0.013745622709393501, + 0.017279041931033134, + 0.017964031547307968, + 0.018445026129484177, + 0.004782994277775288, + 0.003456755541265011, + 0.002869904972612858, + 0.004025149159133434, + 0.0027435440570116043, + 0.008019710890948772, + 0.005684514995664358, + 0.004776757210493088, + 0.008189080283045769, + 0.018246108666062355, + 0.00699631916359067, + 0.02781575545668602, + 0.006147855892777443, + 0.011552550829946995, + 0.018457988277077675, + 0.029825933277606964, + 0.00633449899032712, + 0.01212079543620348, + 0.004387150052934885, + 0.00874018669128418, + 0.00363656273111701, + 0.006545037962496281, + 0.009716938249766827, + 0.017503106966614723, + 0.0038410548586398363, + 0.003613553009927273, + 0.0024668111000210047, + 0.005066012963652611, + 0.0020315612200647593, + 0.0044540055096149445, + 0.005805309861898422, + 0.0039830924943089485, + 0.004356990102678537, + 0.0037405528128147125, + 0.004484133794903755, + 0.0090300552546978, + 0.0024351293686777353, + 0.0046743471175432205, + 0.006535274907946587, + 0.004042970947921276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.10032045841217041 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large steam train with a black steel locomotive and burgundy carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a thick plume of white steam that catches the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crowds of people in period costumes (tweed jackets, bowler hats, long skirts) waving goodbye on the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a busy train platform with ornate cast-iron columns and a vaulted glass ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then move steadily to the right with subtle handheld jerks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059431265108287334 + ], + "opposite_logprob_tokens": [ + -5.130943298339844 + ], + "logprob_reward": -0.0059431265108287334, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059431265108287334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20134517550468445, + 0.09613299369812012, + 0.06188029795885086, + 0.08229261636734009, + 0.153056800365448, + 0.07285550236701965, + 0.12625475227832794, + 0.050730206072330475, + 0.13251252472400665, + 0.02293914370238781 + ], + "credit_map": [ + 0.021336862817406654, + 0.026021311059594154, + 0.025332240387797356, + 0.023359503597021103, + 0.026881670579314232, + 0.027255576103925705, + 0.027559489011764526, + 0.02359851449728012, + 0.01260201446712017, + 0.012578773312270641, + 0.009563791565597057, + 0.01207799557596445, + 0.010553011670708656, + 0.013216371648013592, + 0.010476959869265556, + 0.01506406906992197, + 0.010849419049918652, + 0.007141111884266138, + 0.0086691128090024, + 0.00806129164993763, + 0.004700252786278725, + 0.005689139477908611, + 0.008439648896455765, + 0.008330321870744228, + 0.008432479575276375, + 0.00973515398800373, + 0.00602653156965971, + 0.012224365957081318, + 0.007222577929496765, + 0.014502262696623802, + 0.012486081570386887, + 0.011663158424198627, + 0.015015390701591969, + 0.027605818584561348, + 0.0136070242151618, + 0.01551100704818964, + 0.017209352925419807, + 0.016562245786190033, + 0.019199324771761894, + 0.02834663726389408, + 0.007262577768415213, + 0.009501245804131031, + 0.005738735198974609, + 0.009284532628953457, + 0.004157283343374729, + 0.011042362079024315, + 0.01418154127895832, + 0.011687221936881542, + 0.010369318537414074, + 0.02172834239900112, + 0.005723992362618446, + 0.010453961789608002, + 0.008498065173625946, + 0.014827795326709747, + 0.016352174803614616, + 0.03830110281705856, + 0.005707668140530586, + 0.005337513983249664, + 0.005239384714514017, + 0.009915103204548359, + 0.0028706968296319246, + 0.005672065075486898, + 0.008674729615449905, + 0.007313045207411051, + 0.011098042130470276, + 0.019304512068629265, + 0.01245600450783968, + 0.014913849532604218, + 0.00998444203287363, + 0.017124712467193604, + 0.027840789407491684, + 0.019790170714259148, + 0.001628798432648182, + 0.0020442926324903965, + 0.0029724445194005966, + 0.0037929448299109936, + 0.0013521169312298298, + 0.001987607916817069, + 0.0052787018939852715, + 0.003882236545905471 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.0059431265108287334 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large steam train with a black steel locomotive and burgundy carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a thick plume of white steam that catches the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crowds of people in period costumes (tweed jackets, bowler hats, long skirts) waving goodbye on the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a busy train platform with ornate cast-iron columns and a vaulted glass ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then move steadily to the right with subtle handheld jerks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043040335178375244 + ], + "opposite_logprob_tokens": [ + -3.1680402755737305 + ], + "logprob_reward": -0.043040335178375244, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043040335178375244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18304157257080078, + 0.15055952966213226, + 0.12829551100730896, + 0.05357234179973602, + 0.12066664546728134, + 0.053737983107566833, + 0.11809401214122772, + 0.04956602305173874, + 0.06534335017204285, + 0.07712315768003464 + ], + "credit_map": [ + 0.016528969630599022, + 0.02038094401359558, + 0.01864980347454548, + 0.028659189119935036, + 0.021188469603657722, + 0.033876996487379074, + 0.02220745012164116, + 0.021549727767705917, + 0.01664021797478199, + 0.022226089611649513, + 0.01689774915575981, + 0.024421637877821922, + 0.012503654696047306, + 0.013407665304839611, + 0.014601033180952072, + 0.02986145205795765, + 0.020863544195890427, + 0.015094883739948273, + 0.014725088141858578, + 0.014970818534493446, + 0.012938540428876877, + 0.012846829369664192, + 0.019802089780569077, + 0.017053699120879173, + 0.004068139009177685, + 0.006072562653571367, + 0.0034208374563604593, + 0.008376111276447773, + 0.005008045583963394, + 0.006623518653213978, + 0.011278470978140831, + 0.008724648505449295, + 0.008780031464993954, + 0.013911801390349865, + 0.01430731825530529, + 0.01999570243060589, + 0.01024439837783575, + 0.017178073525428772, + 0.018966583535075188, + 0.017282720655202866, + 0.005730221047997475, + 0.005180757492780685, + 0.005922624375671148, + 0.007792276795953512, + 0.003481796942651272, + 0.008039400912821293, + 0.009504664689302444, + 0.008086233399808407, + 0.009904295206069946, + 0.020069841295480728, + 0.008076300844550133, + 0.013434521853923798, + 0.005600104108452797, + 0.010781243443489075, + 0.0185188390314579, + 0.03170885145664215, + 0.0045722792856395245, + 0.005167623981833458, + 0.003915198612958193, + 0.008156764321029186, + 0.0032100060489028692, + 0.005245175678282976, + 0.01138521172106266, + 0.007913758046925068, + 0.006945523899048567, + 0.008660749532282352, + 0.005976174026727676, + 0.009844450280070305, + 0.005330935120582581, + 0.007906744256615639, + 0.011799721047282219, + 0.008879046887159348, + 0.007538029924035072, + 0.006263867020606995, + 0.008947721682488918, + 0.01842651516199112, + 0.004357323981821537, + 0.009374463930726051, + 0.013680826872587204, + 0.008534400723874569 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.043040335178375244 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide shot, initially static on a busy, turn-of-the-century train platform, begins its steady rightward movement, employing subtle handheld jerks, synchronized with the rhythmic click of wheels on the track. The central subject is a majestic, old-world steam train, its locomotive a behemoth of polished black steel and gleaming brass fittings, powerfully expelling a thick plume of white steam that catches the afternoon sun. The carriages, painted a deep, rich burgundy with delicate gold leaf pinstripes, begin to glide forward, their massive drive wheels turning with a deliberate, powerful rhythm. As the camera pans alongside, the bustling platform is revealed in detail: ornate cast-iron columns support a vast vaulted glass ceiling, through which the warm golden light of the afternoon sun filters, piercing the hazy air. Crowds of people dressed in period costume—men in tweed jackets and bowler hats, women in long skirts and elegant hats—wave goodbye, their faces a mixture of excitement and poignancy. The low-angled sunlight casts long shadows across the wet cobblestone platform, and the train’s metallic surface gleams, lending the entire scene an air of nostalgia and grandeur, a cinematic ode to a bygone age of travel, charged with the energy of a departure as the train steadily gathers speed and moves away from the camera’s field of view.", + "seed": 7487, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_12/index_5516_seed_7487_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.037089562276378274, + "vlm_logprob": -0.037089562276378274, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.037089562276378274, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large steam train with a black steel locomotive and burgundy carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a thick plume of white steam that catches the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crowds of people in period costumes (tweed jackets, bowler hats, long skirts) waving goodbye on the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a busy train platform with ornate cast-iron columns and a vaulted glass ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then move steadily to the right with subtle handheld jerks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01739390194416046 + ], + "opposite_logprob_tokens": [ + -4.892394065856934 + ], + "logprob_reward": -0.01739390194416046, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.01739390194416046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42199549078941345, + 0.06457027792930603, + 0.04024948179721832, + 0.03352855145931244, + 0.06788694858551025, + 0.06635895371437073, + 0.10473986715078354, + 0.07258626073598862, + 0.06766896694898605, + 0.06041523069143295 + ], + "credit_map": [ + 0.01692858710885048, + 0.06173025816679001, + 0.020342348143458366, + 0.05787057802081108, + 0.04127819091081619, + 0.12722787261009216, + 0.059625834226608276, + 0.03699180483818054, + 0.003071667393669486, + 0.0045282673090696335, + 0.004946264438331127, + 0.006208800710737705, + 0.003302485914900899, + 0.019290750846266747, + 0.017477184534072876, + 0.005744858179241419, + 0.0033048486802726984, + 0.004697395488619804, + 0.008625665679574013, + 0.005182787775993347, + 0.003105051815509796, + 0.00571056641638279, + 0.0046714735217392445, + 0.004951694048941135, + 0.003874828340485692, + 0.0036425734870135784, + 0.0053721158765256405, + 0.004159577190876007, + 0.003048995742574334, + 0.0037959080655127764, + 0.004372585564851761, + 0.005261964630335569, + 0.0036309529095888138, + 0.009749592281877995, + 0.005438518710434437, + 0.011057112365961075, + 0.009767950512468815, + 0.0052107940427958965, + 0.011668970808386803, + 0.011363052763044834, + 0.005007804837077856, + 0.007317684590816498, + 0.005666966084390879, + 0.0069582462310791016, + 0.009879572317004204, + 0.007835273630917072, + 0.012902114540338516, + 0.010791296139359474, + 0.004669028799980879, + 0.008301670663058758, + 0.005254615098237991, + 0.008405188098549843, + 0.01049698144197464, + 0.02536061219871044, + 0.013619815930724144, + 0.02863195538520813, + 0.007333947811275721, + 0.00545841921120882, + 0.012189204804599285, + 0.005933975800871849, + 0.007692316547036171, + 0.010146288201212883, + 0.014244272373616695, + 0.009587837383151054, + 0.0030750867445021868, + 0.005408071912825108, + 0.005546632222831249, + 0.010023035109043121, + 0.010518203489482403, + 0.011649934574961662, + 0.010906945914030075, + 0.01054105069488287, + 0.0034091670531779528, + 0.004387326538562775, + 0.004732063505798578, + 0.007175304926931858, + 0.00576942041516304, + 0.011385060846805573, + 0.012628916651010513, + 0.010927969589829445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.01739390194416046 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large steam train with a black steel locomotive and burgundy carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a thick plume of white steam that catches the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crowds of people in period costumes (tweed jackets, bowler hats, long skirts) waving goodbye on the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a busy train platform with ornate cast-iron columns and a vaulted glass ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then move steadily to the right with subtle handheld jerks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496659502387047 + ], + "opposite_logprob_tokens": [ + -2.929966688156128 + ], + "logprob_reward": -0.05496659502387047, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496659502387047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40707042813301086, + 0.12239707261323929, + 0.0582146979868412, + 0.04292719438672066, + 0.046993523836135864, + 0.08799513429403305, + 0.06393680721521378, + 0.06206104904413223, + 0.054030660539865494, + 0.054373450577259064 + ], + "credit_map": [ + 0.015334847383201122, + 0.05738811567425728, + 0.019726619124412537, + 0.056051500141620636, + 0.045599669218063354, + 0.11876168847084045, + 0.06467878818511963, + 0.029529187828302383, + 0.003846302628517151, + 0.010557067580521107, + 0.006388263311237097, + 0.010243580676615238, + 0.007718465756624937, + 0.03986949101090431, + 0.03397533297538757, + 0.009798574261367321, + 0.003952397033572197, + 0.005042659118771553, + 0.007557143457233906, + 0.008259682916104794, + 0.004592964891344309, + 0.009844264946877956, + 0.010178985074162483, + 0.008786601945757866, + 0.002635994693264365, + 0.005959897302091122, + 0.003559431526809931, + 0.006631269585341215, + 0.00438654376193881, + 0.006564630661159754, + 0.007115415297448635, + 0.00607400992885232, + 0.0016535954782739282, + 0.006034740246832371, + 0.0034697167575359344, + 0.0053094495087862015, + 0.007945327088236809, + 0.004281179513782263, + 0.009567835368216038, + 0.008731679990887642, + 0.007073193322867155, + 0.009336947463452816, + 0.007236195262521505, + 0.00900787953287363, + 0.014015914872288704, + 0.01054985448718071, + 0.017655275762081146, + 0.013119877316057682, + 0.0020682867616415024, + 0.003634745953604579, + 0.0038004748057574034, + 0.0067980410531163216, + 0.004499423783272505, + 0.01755312643945217, + 0.007548951543867588, + 0.01803375594317913, + 0.005961963441222906, + 0.0049598622135818005, + 0.010998900048434734, + 0.009859471581876278, + 0.0050195856019854546, + 0.006973362993448973, + 0.011018740013241768, + 0.007269164081662893, + 0.002961333841085434, + 0.003991630859673023, + 0.004798967391252518, + 0.0077238138765096664, + 0.006710011977702379, + 0.009010526351630688, + 0.010745558887720108, + 0.00808881688863039, + 0.003001794684678316, + 0.003941723611205816, + 0.0038597658276557922, + 0.005609611514955759, + 0.005813250783830881, + 0.011366887018084526, + 0.010116638615727425, + 0.010663778521120548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.05496659502387047 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large steam train with a black steel locomotive and burgundy carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a thick plume of white steam that catches the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crowds of people in period costumes (tweed jackets, bowler hats, long skirts) waving goodbye on the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a busy train platform with ornate cast-iron columns and a vaulted glass ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then move steadily to the right with subtle handheld jerks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054965466260910034 + ], + "opposite_logprob_tokens": [ + -2.9299654960632324 + ], + "logprob_reward": -0.054965466260910034, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054965466260910034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3071502149105072, + 0.08239158242940903, + 0.06361270695924759, + 0.09989228844642639, + 0.026812348514795303, + 0.11195986717939377, + 0.1108555942773819, + 0.04410264268517494, + 0.09285111725330353, + 0.06037161126732826 + ], + "credit_map": [ + 0.010105532594025135, + 0.02506149187684059, + 0.013491020537912846, + 0.032774727791547775, + 0.022742150351405144, + 0.10139232873916626, + 0.08067137002944946, + 0.020911600440740585, + 0.007940906099975109, + 0.011557255871593952, + 0.012197880074381828, + 0.009886056184768677, + 0.009248326532542706, + 0.013263027183711529, + 0.008451133035123348, + 0.009847001172602177, + 0.004076797980815172, + 0.006095465738326311, + 0.006901780609041452, + 0.011236531659960747, + 0.0048844898119568825, + 0.010918274521827698, + 0.010105200111865997, + 0.009394166991114616, + 0.008514322340488434, + 0.011601482518017292, + 0.02173418365418911, + 0.013345148414373398, + 0.0071339854039251804, + 0.009808343835175037, + 0.012867145240306854, + 0.014887678436934948, + 0.00196273741312325, + 0.003342170501127839, + 0.002202731790021062, + 0.0057538896799087524, + 0.0026118389796465635, + 0.0022833463735878468, + 0.004887260030955076, + 0.0037683749105781317, + 0.009142312221229076, + 0.013092301785945892, + 0.01569017581641674, + 0.013860994018614292, + 0.012207435443997383, + 0.010556735098361969, + 0.019770881161093712, + 0.017639031633734703, + 0.003565416904166341, + 0.0055249095894396305, + 0.00597345270216465, + 0.012748234905302525, + 0.0095591489225626, + 0.026486625894904137, + 0.01906990446150303, + 0.027927905321121216, + 0.0015226437244564295, + 0.0024531204253435135, + 0.00253099063411355, + 0.005617466289550066, + 0.0025799674913287163, + 0.011549328453838825, + 0.006738096009939909, + 0.011111031286418438, + 0.0062381429597735405, + 0.006554813589900732, + 0.012214308604598045, + 0.01282836776226759, + 0.008174273185431957, + 0.016727019101381302, + 0.012600453570485115, + 0.017513737082481384, + 0.0030981122981756926, + 0.004461119417101145, + 0.004906312562525272, + 0.007352418266236782, + 0.006268129218369722, + 0.011543989181518555, + 0.011414503678679466, + 0.011327025480568409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.054965466260910034 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large steam train with a black steel locomotive and burgundy carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a thick plume of white steam that catches the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crowds of people in period costumes (tweed jackets, bowler hats, long skirts) waving goodbye on the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a busy train platform with ornate cast-iron columns and a vaulted glass ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then move steadily to the right with subtle handheld jerks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031887658406049013 + ], + "opposite_logprob_tokens": [ + -5.753188610076904 + ], + "logprob_reward": -0.0031887658406049013, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031887658406049013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.322722464799881, + 0.037099599838256836, + 0.03540036454796791, + 0.12985394895076752, + 0.05990317463874817, + 0.06583740562200546, + 0.13397164642810822, + 0.06075645238161087, + 0.1067899838089943, + 0.04766504094004631 + ], + "credit_map": [ + 0.014978424645960331, + 0.05374879017472267, + 0.016778601333498955, + 0.05824875459074974, + 0.037543535232543945, + 0.06737165153026581, + 0.047786515206098557, + 0.026266159489750862, + 0.0024560908786952496, + 0.006458264775574207, + 0.0038510868325829506, + 0.004666468594223261, + 0.0043350327759981155, + 0.0062559135258197784, + 0.004745135549455881, + 0.004331601317971945, + 0.002475908724591136, + 0.0035634140949696302, + 0.005727774929255247, + 0.0046339030377566814, + 0.002740651136264205, + 0.0062241582199931145, + 0.005259252618998289, + 0.004775296896696091, + 0.006686881184577942, + 0.016820305958390236, + 0.008859872817993164, + 0.015218394808471203, + 0.013592223636806011, + 0.021584955975413322, + 0.02659737505018711, + 0.020493920892477036, + 0.0030605311039835215, + 0.006424408871680498, + 0.0032701233867555857, + 0.008707761764526367, + 0.00891221035271883, + 0.005346319172531366, + 0.01359147485345602, + 0.01059033628553152, + 0.0058564795181155205, + 0.009794121608138084, + 0.007593516726046801, + 0.006063966546207666, + 0.008607245050370693, + 0.007088049780577421, + 0.010581555776298046, + 0.010252467356622219, + 0.006611271295696497, + 0.01004826556891203, + 0.007898863404989243, + 0.011586172506213188, + 0.01229169126600027, + 0.03323055058717728, + 0.016800493001937866, + 0.035504329949617386, + 0.005066452082246542, + 0.004872642923146486, + 0.006639149505645037, + 0.005997765343636274, + 0.0055238716304302216, + 0.01132210437208414, + 0.010538150556385517, + 0.010796314105391502, + 0.01104880589991808, + 0.008549762889742851, + 0.021517541259527206, + 0.010407024063169956, + 0.0074584693647921085, + 0.017078161239624023, + 0.01582854986190796, + 0.014901652000844479, + 0.0035548056475818157, + 0.004422832280397415, + 0.0027494749519973993, + 0.005075865425169468, + 0.003224290208891034, + 0.007409298326820135, + 0.012185142375528812, + 0.009043324738740921 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.0031887658406049013 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large steam train with a black steel locomotive and burgundy carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a thick plume of white steam that catches the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crowds of people in period costumes (tweed jackets, bowler hats, long skirts) waving goodbye on the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a busy train platform with ornate cast-iron columns and a vaulted glass ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then move steadily to the right with subtle handheld jerks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054933082312345505 + ], + "opposite_logprob_tokens": [ + -2.9299330711364746 + ], + "logprob_reward": -0.054933082312345505, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054933082312345505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2255849391222, + 0.07151772826910019, + 0.04924877732992172, + 0.025144224986433983, + 0.09909729659557343, + 0.059154827147722244, + 0.3053082823753357, + 0.0425829142332077, + 0.041080813854932785, + 0.08128023892641068 + ], + "credit_map": [ + 0.010318499058485031, + 0.020397931337356567, + 0.013133094646036625, + 0.035295356065034866, + 0.025160402059555054, + 0.03577549010515213, + 0.06565096229314804, + 0.019853191450238228, + 0.005615793168544769, + 0.007789932657033205, + 0.005240897182375193, + 0.00831677671521902, + 0.006899869069457054, + 0.01486825942993164, + 0.016491975635290146, + 0.006294222082942724, + 0.0033849047031253576, + 0.006603047251701355, + 0.005420723930001259, + 0.005881393328309059, + 0.004987104795873165, + 0.009570576250553131, + 0.007272304967045784, + 0.006128722336143255, + 0.0017779669724404812, + 0.0039323377422988415, + 0.0031966278329491615, + 0.002864752896130085, + 0.0030508122872561216, + 0.003390092868357897, + 0.0035720649175345898, + 0.0033595687709748745, + 0.00487237423658371, + 0.011536980979144573, + 0.004155892413109541, + 0.01034998707473278, + 0.01624516025185585, + 0.009246328845620155, + 0.02314136177301407, + 0.019549215212464333, + 0.0035696816630661488, + 0.00710271904245019, + 0.003538910299539566, + 0.0062624127604067326, + 0.009187591262161732, + 0.006570369470864534, + 0.011788287200033665, + 0.011134854517877102, + 0.012518017552793026, + 0.02035897597670555, + 0.013003069907426834, + 0.02314196527004242, + 0.030725732445716858, + 0.07587146759033203, + 0.0458284430205822, + 0.0838605985045433, + 0.003523011226207018, + 0.003847651882097125, + 0.00457049161195755, + 0.004230643156915903, + 0.008256576023995876, + 0.006017072591930628, + 0.0069008031859993935, + 0.005236668977886438, + 0.00415010703727603, + 0.002465376164764166, + 0.007039471995085478, + 0.004381808452308178, + 0.0043807257898151875, + 0.005659487098455429, + 0.0070275613106787205, + 0.0059762741439044476, + 0.004246888216584921, + 0.00627163564786315, + 0.005783453118056059, + 0.008543640375137329, + 0.00834721326828003, + 0.01640813797712326, + 0.018516598269343376, + 0.013162673451006413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.054933082312345505 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide shot, initially static on a busy, turn-of-the-century train platform, begins its steady rightward movement, employing subtle handheld jerks, synchronized with the rhythmic click of wheels on the track. The central subject is a majestic, old-world steam train, its locomotive a behemoth of polished black steel and gleaming brass fittings, powerfully expelling a thick plume of white steam that catches the afternoon sun. The carriages, painted a deep, rich burgundy with delicate gold leaf pinstripes, begin to glide forward, their massive drive wheels turning with a deliberate, powerful rhythm. As the camera pans alongside, the bustling platform is revealed in detail: ornate cast-iron columns support a vast vaulted glass ceiling, through which the warm golden light of the afternoon sun filters, piercing the hazy air. Crowds of people dressed in period costume—men in tweed jackets and bowler hats, women in long skirts and elegant hats—wave goodbye, their faces a mixture of excitement and poignancy. The low-angled sunlight casts long shadows across the wet cobblestone platform, and the train’s metallic surface gleams, lending the entire scene an air of nostalgia and grandeur, a cinematic ode to a bygone age of travel, charged with the energy of a departure as the train steadily gathers speed and moves away from the camera’s field of view.", + "seed": 7488, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_12/index_5516_seed_7488_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03211029721423984, + "vlm_logprob": -0.03211029721423984, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.03211029721423984, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large steam train with a black steel locomotive and burgundy carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a thick plume of white steam that catches the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crowds of people in period costumes (tweed jackets, bowler hats, long skirts) waving goodbye on the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a busy train platform with ornate cast-iron columns and a vaulted glass ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then move steadily to the right with subtle handheld jerks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029186030849814415 + ], + "opposite_logprob_tokens": [ + -3.9041860103607178 + ], + "logprob_reward": -0.029186030849814415, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.029186030849814415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3158946633338928, + 0.18707877397537231, + 0.08870786428451538, + 0.025352634489536285, + 0.018394675105810165, + 0.17712728679180145, + 0.02794882468879223, + 0.0641750767827034, + 0.0569675974547863, + 0.03835267201066017 + ], + "credit_map": [ + 0.022278008982539177, + 0.133261576294899, + 0.015759257599711418, + 0.02068406343460083, + 0.018881095573306084, + 0.006822425872087479, + 0.08574458211660385, + 0.012463604100048542, + 0.022826848551630974, + 0.022257588803768158, + 0.019508736208081245, + 0.02671489492058754, + 0.039780184626579285, + 0.008547537960112095, + 0.02956809289753437, + 0.017874889075756073, + 0.00562889501452446, + 0.0037448150105774403, + 0.0038683279417455196, + 0.003278266405686736, + 0.0035024811513721943, + 0.0367414727807045, + 0.004226706922054291, + 0.02771688625216484, + 0.0027186458464711905, + 0.003304561832919717, + 0.0034480479080229998, + 0.0028514754958450794, + 0.0036632975097745657, + 0.003239801386371255, + 0.003042499767616391, + 0.003084302879869938, + 0.0016419547609984875, + 0.0012782784178853035, + 0.0023511280305683613, + 0.003277405397966504, + 0.0021564208436757326, + 0.00283538899384439, + 0.0025361659936606884, + 0.0023179324343800545, + 0.014266002923250198, + 0.035827044397592545, + 0.012049810029566288, + 0.014698358252644539, + 0.019760163500905037, + 0.026101136580109596, + 0.02684972621500492, + 0.02757502719759941, + 0.0022620123345404863, + 0.0021737965289503336, + 0.001827221130952239, + 0.0032772172708064318, + 0.0031162400264292955, + 0.006748591084033251, + 0.0035866848193109035, + 0.004957057535648346, + 0.005752038210630417, + 0.0051992516964674, + 0.007271364331245422, + 0.009488247334957123, + 0.006664242595434189, + 0.008759796619415283, + 0.011504950933158398, + 0.009535177610814571, + 0.00554272485896945, + 0.005008463282138109, + 0.004829684272408485, + 0.004804019350558519, + 0.012271231040358543, + 0.008329498581588268, + 0.009206038899719715, + 0.006975934375077486, + 0.0031614457257092, + 0.0016154495533555746, + 0.0029644095338881016, + 0.002961446065455675, + 0.0045942566357553005, + 0.010608799755573273, + 0.005204909946769476, + 0.007241951301693916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.029186030849814415 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large steam train with a black steel locomotive and burgundy carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a thick plume of white steam that catches the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crowds of people in period costumes (tweed jackets, bowler hats, long skirts) waving goodbye on the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a busy train platform with ornate cast-iron columns and a vaulted glass ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then move steadily to the right with subtle handheld jerks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018199829384684563 + ], + "opposite_logprob_tokens": [ + -4.018199920654297 + ], + "logprob_reward": -0.018199829384684563, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018199829384684563, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17869561910629272, + 0.024836400523781776, + 0.07061237841844559, + 0.04195188358426094, + 0.05244653671979904, + 0.38123321533203125, + 0.09597465395927429, + 0.008845299482345581, + 0.09809386730194092, + 0.04731004685163498 + ], + "credit_map": [ + 0.011889236979186535, + 0.07534255087375641, + 0.0074351197108626366, + 0.011434211395680904, + 0.010473676957190037, + 0.003852645866572857, + 0.050663962960243225, + 0.007604236248880625, + 0.002873797668144107, + 0.002915129764005542, + 0.0027233297005295753, + 0.003631414147093892, + 0.005281737074255943, + 0.0014134636148810387, + 0.0037546143867075443, + 0.0022429178934544325, + 0.004895293153822422, + 0.0030982340686023235, + 0.00282871350646019, + 0.002493243897333741, + 0.0029209048952907324, + 0.028383605182170868, + 0.0027527548372745514, + 0.02323964610695839, + 0.004656077362596989, + 0.005030895583331585, + 0.003214015159755945, + 0.005232286173850298, + 0.0050872862339019775, + 0.006250476464629173, + 0.0070761023089289665, + 0.005404750816524029, + 0.006232284940779209, + 0.004373036790639162, + 0.004548539873212576, + 0.0059980908408761024, + 0.004616462159901857, + 0.012227438390254974, + 0.005289234220981598, + 0.009161454625427723, + 0.03152595832943916, + 0.07920952886343002, + 0.025645356625318527, + 0.029410215094685555, + 0.039125069975852966, + 0.055157825350761414, + 0.06031818315386772, + 0.06084110960364342, + 0.005877074785530567, + 0.0072525846771895885, + 0.007635932881385088, + 0.012378125451505184, + 0.015178082510828972, + 0.007760080508887768, + 0.021537985652685165, + 0.018354792147874832, + 0.0012086144415661693, + 0.0007001832709647715, + 0.0007699260604567826, + 0.0009788613533601165, + 0.001256742631085217, + 0.0013716381508857012, + 0.0014495899667963386, + 0.0011097443057224154, + 0.008691055700182915, + 0.007081417366862297, + 0.007973680272698402, + 0.009934051893651485, + 0.02247159741818905, + 0.015155496075749397, + 0.013817843049764633, + 0.01296873576939106, + 0.0031243155244737864, + 0.0015169933903962374, + 0.003988980315625668, + 0.003925395663827658, + 0.0038735431153327227, + 0.015011647716164589, + 0.005338805261999369, + 0.01053037028759718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.018199829384684563 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large steam train with a black steel locomotive and burgundy carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a thick plume of white steam that catches the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crowds of people in period costumes (tweed jackets, bowler hats, long skirts) waving goodbye on the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a busy train platform with ornate cast-iron columns and a vaulted glass ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then move steadily to the right with subtle handheld jerks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204884499311447 + ], + "opposite_logprob_tokens": [ + -2.81204891204834 + ], + "logprob_reward": -0.06204884499311447, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204884499311447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40180227160453796, + 0.04005543142557144, + 0.05615504831075668, + 0.03668413683772087, + 0.05820583552122116, + 0.07906761020421982, + 0.06246529892086983, + 0.029796890914440155, + 0.15267597138881683, + 0.08309146761894226 + ], + "credit_map": [ + 0.026656176894903183, + 0.16752822697162628, + 0.021560899913311005, + 0.028815606608986855, + 0.023843444883823395, + 0.01019863411784172, + 0.10590493679046631, + 0.017294352874159813, + 0.006455604452639818, + 0.003576101502403617, + 0.006199696101248264, + 0.00473392428830266, + 0.0060984669253230095, + 0.0024381480179727077, + 0.005796352867037058, + 0.0047571356408298016, + 0.0036909794434905052, + 0.00279972143471241, + 0.0030513284727931023, + 0.002941643353551626, + 0.0022103607188910246, + 0.021953970193862915, + 0.0029686472844332457, + 0.016538400202989578, + 0.004419974982738495, + 0.004493950866162777, + 0.005569808650761843, + 0.003420149441808462, + 0.004165317863225937, + 0.004504598677158356, + 0.005464869551360607, + 0.004645470064133406, + 0.004660761449486017, + 0.0040434119291603565, + 0.004394766874611378, + 0.0061970544047653675, + 0.004926069173961878, + 0.01745728589594364, + 0.006184525787830353, + 0.010341959074139595, + 0.006014679092913866, + 0.009245602414011955, + 0.007865287363529205, + 0.00932817067950964, + 0.0076007298193871975, + 0.01321189384907484, + 0.014287494122982025, + 0.011513752862811089, + 0.004516845569014549, + 0.004316067323088646, + 0.005018108990043402, + 0.006918136030435562, + 0.006992391310632229, + 0.014688108116388321, + 0.008288743905723095, + 0.011726893484592438, + 0.003044038312509656, + 0.002651715651154518, + 0.0032681627199053764, + 0.0038101477548480034, + 0.0031050636898726225, + 0.004321616142988205, + 0.005443787667900324, + 0.004152359906584024, + 0.014573143795132637, + 0.011621099896728992, + 0.01375608891248703, + 0.016132008284330368, + 0.03498508408665657, + 0.021184207871556282, + 0.021288061514496803, + 0.01913628727197647, + 0.005237505305558443, + 0.002504661213606596, + 0.006472157314419746, + 0.004102123901247978, + 0.007703843526542187, + 0.02946310304105282, + 0.007715976797044277, + 0.019892092794179916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.06204884499311447 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large steam train with a black steel locomotive and burgundy carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a thick plume of white steam that catches the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crowds of people in period costumes (tweed jackets, bowler hats, long skirts) waving goodbye on the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a busy train platform with ornate cast-iron columns and a vaulted glass ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then move steadily to the right with subtle handheld jerks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024851649068295956 + ], + "opposite_logprob_tokens": [ + -6.002485275268555 + ], + "logprob_reward": -0.0024851649068295956, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024851649068295956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32739630341529846, + 0.06602738797664642, + 0.02875376120209694, + 0.037779927253723145, + 0.05650041252374649, + 0.1353749930858612, + 0.045919016003608704, + 0.04491346701979637, + 0.11078496277332306, + 0.1465497612953186 + ], + "credit_map": [ + 0.030121076852083206, + 0.12207690626382828, + 0.02264479547739029, + 0.021846668794751167, + 0.02473626099526882, + 0.010523306205868721, + 0.07803595811128616, + 0.017411334440112114, + 0.01260373368859291, + 0.009841016493737698, + 0.006363420281559229, + 0.006804607808589935, + 0.012511254288256168, + 0.003376840613782406, + 0.009737771935760975, + 0.004788744263350964, + 0.0037552809808403254, + 0.002589296782389283, + 0.0036894658114761114, + 0.003680381691083312, + 0.002735150745138526, + 0.005082170478999615, + 0.003227759152650833, + 0.00399425346404314, + 0.0038253427483141422, + 0.004004940390586853, + 0.004553414881229401, + 0.005233167205005884, + 0.003580948803573847, + 0.005713772960007191, + 0.0056100101210176945, + 0.00525833060964942, + 0.005062337499111891, + 0.0052594211883842945, + 0.005574188195168972, + 0.006136800162494183, + 0.005099776666611433, + 0.010767636820673943, + 0.010385875590145588, + 0.008214371278882027, + 0.007658893242478371, + 0.013211946934461594, + 0.011537300422787666, + 0.01662760227918625, + 0.014798272401094437, + 0.018965154886245728, + 0.034254852682352066, + 0.0183209627866745, + 0.004570852033793926, + 0.004444503225386143, + 0.003647860139608383, + 0.004251622129231691, + 0.0042398665100336075, + 0.010910279117524624, + 0.0053641702979803085, + 0.008489861153066158, + 0.005230841692537069, + 0.0028591405134648085, + 0.003920834977179766, + 0.0029771323315799236, + 0.00552917318418622, + 0.012792474590241909, + 0.004665341228246689, + 0.006938527803868055, + 0.01947992667555809, + 0.007121120113879442, + 0.011283728294074535, + 0.01081760786473751, + 0.015553251840174198, + 0.015896370634436607, + 0.015483914874494076, + 0.015149036422371864, + 0.012208972126245499, + 0.004534522537142038, + 0.01352949719876051, + 0.00996785331517458, + 0.014547668397426605, + 0.041666772216558456, + 0.020564725622534752, + 0.029529744759202003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.0024851649068295956 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large steam train with a black steel locomotive and burgundy carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a thick plume of white steam that catches the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crowds of people in period costumes (tweed jackets, bowler hats, long skirts) waving goodbye on the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a busy train platform with ornate cast-iron columns and a vaulted glass ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then move steadily to the right with subtle handheld jerks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048631615936756134 + ], + "opposite_logprob_tokens": [ + -3.0486316680908203 + ], + "logprob_reward": -0.048631615936756134, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048631615936756134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1521114706993103, + 0.10389173030853271, + 0.07496935874223709, + 0.05443451926112175, + 0.0711469054222107, + 0.19057632982730865, + 0.06960292905569077, + 0.10382259637117386, + 0.1320340931415558, + 0.0474100336432457 + ], + "credit_map": [ + 0.01994558610022068, + 0.037884775549173355, + 0.01794414222240448, + 0.018056223168969154, + 0.01172545738518238, + 0.006334404926747084, + 0.029195385053753853, + 0.011025507934391499, + 0.011535853147506714, + 0.013380622491240501, + 0.009936941787600517, + 0.01490921899676323, + 0.018208198249340057, + 0.005975442472845316, + 0.016302520409226418, + 0.013642935082316399, + 0.004414261784404516, + 0.004837674088776112, + 0.006304916925728321, + 0.00491023063659668, + 0.003167244140058756, + 0.023717431351542473, + 0.004447875078767538, + 0.02316972054541111, + 0.013361720368266106, + 0.006708064582198858, + 0.004887488670647144, + 0.00469745509326458, + 0.006179298739880323, + 0.006250023376196623, + 0.005130269099026918, + 0.007220200262963772, + 0.009535125456750393, + 0.008004747331142426, + 0.005386429373174906, + 0.009066257625818253, + 0.007364332675933838, + 0.012304972857236862, + 0.008208708837628365, + 0.011276323348283768, + 0.014592762105166912, + 0.031127845868468285, + 0.014444227330386639, + 0.025136630982160568, + 0.021937789395451546, + 0.030949527397751808, + 0.02592022344470024, + 0.026467330753803253, + 0.005811373703181744, + 0.008045043796300888, + 0.005322916898876429, + 0.008860547095537186, + 0.007905123755335808, + 0.0058510988019406796, + 0.015312581323087215, + 0.012494242750108242, + 0.013235816732048988, + 0.008074376732110977, + 0.008492494933307171, + 0.009970433078706264, + 0.01357181929051876, + 0.024183183908462524, + 0.011308867484331131, + 0.014985606074333191, + 0.018372535705566406, + 0.00917050801217556, + 0.015339153818786144, + 0.015094570815563202, + 0.023007992655038834, + 0.01584720052778721, + 0.019422447308897972, + 0.015779683366417885, + 0.006715931463986635, + 0.0022979113273322582, + 0.005051343236118555, + 0.005152921658009291, + 0.005512551870197058, + 0.006884599104523659, + 0.008596250787377357, + 0.007198523264378309 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.048631615936756134 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide shot, initially static on a busy, turn-of-the-century train platform, begins its steady rightward movement, employing subtle handheld jerks, synchronized with the rhythmic click of wheels on the track. The central subject is a majestic, old-world steam train, its locomotive a behemoth of polished black steel and gleaming brass fittings, powerfully expelling a thick plume of white steam that catches the afternoon sun. The carriages, painted a deep, rich burgundy with delicate gold leaf pinstripes, begin to glide forward, their massive drive wheels turning with a deliberate, powerful rhythm. As the camera pans alongside, the bustling platform is revealed in detail: ornate cast-iron columns support a vast vaulted glass ceiling, through which the warm golden light of the afternoon sun filters, piercing the hazy air. Crowds of people dressed in period costume—men in tweed jackets and bowler hats, women in long skirts and elegant hats—wave goodbye, their faces a mixture of excitement and poignancy. The low-angled sunlight casts long shadows across the wet cobblestone platform, and the train’s metallic surface gleams, lending the entire scene an air of nostalgia and grandeur, a cinematic ode to a bygone age of travel, charged with the energy of a departure as the train steadily gathers speed and moves away from the camera’s field of view.", + "seed": 7489, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_12/index_5516_seed_7489_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03922134866006673, + "vlm_logprob": -0.03922134866006673, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.03922134866006673, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large steam train with a black steel locomotive and burgundy carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a thick plume of white steam that catches the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crowds of people in period costumes (tweed jackets, bowler hats, long skirts) waving goodbye on the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a busy train platform with ornate cast-iron columns and a vaulted glass ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then move steadily to the right with subtle handheld jerks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04765801131725311 + ], + "opposite_logprob_tokens": [ + -3.1726579666137695 + ], + "logprob_reward": -0.04765801131725311, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04765801131725311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07958117127418518, + 0.0417943000793457, + 0.013531139120459557, + 0.06405740231275558, + 0.05284092202782631, + 0.2831679880619049, + 0.08031227439641953, + 0.21968486905097961, + 0.07223277539014816, + 0.09279721230268478 + ], + "credit_map": [ + 0.0064347004517912865, + 0.00984250195324421, + 0.00821342971175909, + 0.01908264122903347, + 0.00794286374002695, + 0.006166907958686352, + 0.013580569066107273, + 0.008317559026181698, + 0.002313922857865691, + 0.00218534329906106, + 0.005630256608128548, + 0.009489012882113457, + 0.007764650508761406, + 0.004433951806277037, + 0.005124314222484827, + 0.004852847661823034, + 0.0010096852201968431, + 0.001307485974393785, + 0.0019417782314121723, + 0.0024533092509955168, + 0.001605415716767311, + 0.0012753652408719063, + 0.0024196854792535305, + 0.001518413657322526, + 0.004067122470587492, + 0.016151253134012222, + 0.010367629118263721, + 0.005872924346476793, + 0.006516478024423122, + 0.008431544527411461, + 0.00612909160554409, + 0.006521354429423809, + 0.004177098628133535, + 0.012614475563168526, + 0.01167275570333004, + 0.007002789527177811, + 0.004896658007055521, + 0.003553050337359309, + 0.004494226071983576, + 0.00442986935377121, + 0.027237698435783386, + 0.024306325241923332, + 0.07008768618106842, + 0.05091583728790283, + 0.020811166614294052, + 0.03268936648964882, + 0.03944001346826553, + 0.01767990179359913, + 0.00501268170773983, + 0.0034562130458652973, + 0.025499466806650162, + 0.008995750918984413, + 0.005273716989904642, + 0.005911463871598244, + 0.007252620067447424, + 0.0189103614538908, + 0.017446361482143402, + 0.05779222026467323, + 0.0718797892332077, + 0.01844232901930809, + 0.013530079275369644, + 0.009858982637524605, + 0.019414497539401054, + 0.011320612393319607, + 0.0077512371353805065, + 0.014353311620652676, + 0.006528649013489485, + 0.008494348265230656, + 0.011174771003425121, + 0.00584307499229908, + 0.011133791878819466, + 0.0069535868242383, + 0.01225243415683508, + 0.015147598460316658, + 0.018160857260227203, + 0.009817726910114288, + 0.00784272886812687, + 0.005894030444324017, + 0.016331182792782784, + 0.00735065434128046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 9 + ], + "selected_reward": -0.04765801131725311 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large steam train with a black steel locomotive and burgundy carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a thick plume of white steam that catches the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crowds of people in period costumes (tweed jackets, bowler hats, long skirts) waving goodbye on the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a busy train platform with ornate cast-iron columns and a vaulted glass ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then move steadily to the right with subtle handheld jerks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003393024206161 + ], + "opposite_logprob_tokens": [ + -2.6950340270996094 + ], + "logprob_reward": -0.07003393024206161, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003393024206161, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07304547727108002, + 0.057609930634498596, + 0.023437688127160072, + 0.24070517718791962, + 0.057294633239507675, + 0.044163066893815994, + 0.2375786453485489, + 0.09405338019132614, + 0.12140820175409317, + 0.05070377141237259 + ], + "credit_map": [ + 0.007561012636870146, + 0.013708328828215599, + 0.008143449202179909, + 0.010252129286527634, + 0.0069838594645261765, + 0.006276423577219248, + 0.010194726288318634, + 0.009925552643835545, + 0.003790464950725436, + 0.002571849385276437, + 0.008930576965212822, + 0.01279330626130104, + 0.012475743889808655, + 0.00664566783234477, + 0.005928113590925932, + 0.004474204033613205, + 0.002131971763446927, + 0.0038508472498506308, + 0.002443610457703471, + 0.004295337479561567, + 0.0021281589288264513, + 0.002014966681599617, + 0.003777761710807681, + 0.0027950352523475885, + 0.014398890547454357, + 0.07536603510379791, + 0.060664065182209015, + 0.02277331054210663, + 0.021362420171499252, + 0.021975703537464142, + 0.007478948682546616, + 0.01668579876422882, + 0.004867846146225929, + 0.012704489752650261, + 0.01218724250793457, + 0.005417290609329939, + 0.006284921895712614, + 0.0059025585651397705, + 0.0053987340070307255, + 0.004531554877758026, + 0.004522147122770548, + 0.0034196197520941496, + 0.0120071517303586, + 0.006354056764394045, + 0.003996073268353939, + 0.005880475044250488, + 0.00530501501634717, + 0.0026785312220454216, + 0.026367001235485077, + 0.01678181253373623, + 0.07676959782838821, + 0.028526555746793747, + 0.016714274883270264, + 0.01834656111896038, + 0.026959983631968498, + 0.027112871408462524, + 0.007348516955971718, + 0.023955848067998886, + 0.030168935656547546, + 0.008240200579166412, + 0.006127003580331802, + 0.004599386360496283, + 0.008558000437915325, + 0.005055489484220743, + 0.012224942445755005, + 0.02001476101577282, + 0.011772786267101765, + 0.01483947690576315, + 0.022499719634652138, + 0.010723224841058254, + 0.01743096113204956, + 0.011902335099875927, + 0.0070755318738520145, + 0.00809869822114706, + 0.009172230958938599, + 0.005416783969849348, + 0.0045884293504059315, + 0.0036242937203496695, + 0.008619347587227821, + 0.004108457360416651 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 8 + ], + "selected_reward": -0.07003393024206161 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large steam train with a black steel locomotive and burgundy carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a thick plume of white steam that catches the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crowds of people in period costumes (tweed jackets, bowler hats, long skirts) waving goodbye on the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a busy train platform with ornate cast-iron columns and a vaulted glass ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then move steadily to the right with subtle handheld jerks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014198473654687405 + ], + "opposite_logprob_tokens": [ + -4.264198303222656 + ], + "logprob_reward": -0.014198473654687405, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014198473654687405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06900270283222198, + 0.11890286952257156, + 0.0343707837164402, + 0.0891406312584877, + 0.022159449756145477, + 0.024580109864473343, + 0.3664999008178711, + 0.18496331572532654, + 0.026270609349012375, + 0.06410962343215942 + ], + "credit_map": [ + 0.007811496499925852, + 0.017846673727035522, + 0.007951234467327595, + 0.010274801403284073, + 0.0052593848668038845, + 0.0058111101388931274, + 0.008345399983227253, + 0.005702603608369827, + 0.006586594507098198, + 0.006096661556512117, + 0.01615511253476143, + 0.03690924495458603, + 0.02096281200647354, + 0.008707656525075436, + 0.013375372625887394, + 0.010109427385032177, + 0.0025770608335733414, + 0.0031242158729583025, + 0.005648710764944553, + 0.0064181974157691, + 0.0037808348424732685, + 0.0031862331088632345, + 0.005086018703877926, + 0.004549515433609486, + 0.005383665673434734, + 0.02829025872051716, + 0.022070124745368958, + 0.008324411697685719, + 0.007896640338003635, + 0.008183149620890617, + 0.0029793663416057825, + 0.006013023667037487, + 0.002248605713248253, + 0.005102715454995632, + 0.003096903208643198, + 0.0019030161201953888, + 0.003199670696631074, + 0.0024506929330527782, + 0.0021986723877489567, + 0.001959175104275346, + 0.0024544401094317436, + 0.001793275703676045, + 0.003620499512180686, + 0.005667476914823055, + 0.0018841029377654195, + 0.003185543231666088, + 0.003753374330699444, + 0.0022213980555534363, + 0.04757307842373848, + 0.02492203377187252, + 0.11030780524015427, + 0.044539377093315125, + 0.022740179672837257, + 0.029000021517276764, + 0.04478291794657707, + 0.042634516954422, + 0.014982356689870358, + 0.04738860949873924, + 0.06251127272844315, + 0.015007245354354382, + 0.011223971843719482, + 0.007865994237363338, + 0.016647549346089363, + 0.00933633279055357, + 0.0019224431598559022, + 0.002435058355331421, + 0.00364897632971406, + 0.005469920579344034, + 0.004110621754080057, + 0.0026626382023096085, + 0.0039310092106461525, + 0.0020899460650980473, + 0.0074921101331710815, + 0.011923141777515411, + 0.01255069114267826, + 0.006382368039339781, + 0.005144392140209675, + 0.004061051178723574, + 0.011361250653862953, + 0.005194623023271561 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 1 + ], + "selected_reward": -0.014198473654687405 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large steam train with a black steel locomotive and burgundy carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a thick plume of white steam that catches the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crowds of people in period costumes (tweed jackets, bowler hats, long skirts) waving goodbye on the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a busy train platform with ornate cast-iron columns and a vaulted glass ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then move steadily to the right with subtle handheld jerks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021942604798823595 + ], + "opposite_logprob_tokens": [ + -6.127194404602051 + ], + "logprob_reward": -0.0021942604798823595, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021942604798823595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09463027864694595, + 0.09109412878751755, + 0.08220702409744263, + 0.06358834356069565, + 0.038565944880247116, + 0.02056969329714775, + 0.3148069381713867, + 0.11583756655454636, + 0.03694840148091316, + 0.1417517066001892 + ], + "credit_map": [ + 0.007597232237458229, + 0.010901806876063347, + 0.013866272754967213, + 0.024398593232035637, + 0.0071098096668720245, + 0.0063429223373532295, + 0.016536718234419823, + 0.007876931689679623, + 0.003960594069212675, + 0.003950965125113726, + 0.006431827321648598, + 0.02476302720606327, + 0.007821761071681976, + 0.0067458064295351505, + 0.009830574505031109, + 0.027589570730924606, + 0.0061985501088202, + 0.008203456178307533, + 0.014829098246991634, + 0.009068027138710022, + 0.011125108227133751, + 0.009215123951435089, + 0.012328295968472958, + 0.011239366605877876, + 0.004193203058093786, + 0.018240252509713173, + 0.01277216523885727, + 0.005735478363931179, + 0.005499755032360554, + 0.00596691993996501, + 0.005228810478001833, + 0.005951759871095419, + 0.00375201809220016, + 0.009473158977925777, + 0.007868441753089428, + 0.003402674337849021, + 0.005317720118910074, + 0.0035979559179395437, + 0.0026397837791591883, + 0.002514192368835211, + 0.0020924489945173264, + 0.0018512271344661713, + 0.002209160942584276, + 0.0031358140986412764, + 0.0023851303849369287, + 0.0032046139240264893, + 0.003260832978412509, + 0.0024304650723934174, + 0.03720354661345482, + 0.02042141743004322, + 0.10467570275068283, + 0.038302935659885406, + 0.018654780462384224, + 0.024128077551722527, + 0.035897642374038696, + 0.03552282601594925, + 0.0100014116615057, + 0.02731773816049099, + 0.03960506245493889, + 0.009870612993836403, + 0.006894541438668966, + 0.005888839717954397, + 0.01133397314697504, + 0.0049253967590630054, + 0.003005190519616008, + 0.0069860368967056274, + 0.005687213968485594, + 0.003426153911277652, + 0.00511161470785737, + 0.0033430554904043674, + 0.0058198245242238045, + 0.0035693116951733828, + 0.01765790395438671, + 0.021444177255034447, + 0.026247624307870865, + 0.015062930062413216, + 0.011854158714413643, + 0.010035107843577862, + 0.02683209255337715, + 0.012617703527212143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.0021942604798823595 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a large steam train with a black steel locomotive and burgundy carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a thick plume of white steam that catches the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crowds of people in period costumes (tweed jackets, bowler hats, long skirts) waving goodbye on the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a busy train platform with ornate cast-iron columns and a vaulted glass ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then move steadily to the right with subtle handheld jerks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06202206760644913 + ], + "opposite_logprob_tokens": [ + -2.8120219707489014 + ], + "logprob_reward": -0.06202206760644913, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06202206760644913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0730418711900711, + 0.10841186344623566, + 0.03473199903964996, + 0.10997597873210907, + 0.026992687955498695, + 0.08687695115804672, + 0.09121128916740417, + 0.3653281033039093, + 0.039346762001514435, + 0.0640825405716896 + ], + "credit_map": [ + 0.006874700076878071, + 0.008539029397070408, + 0.010091608390212059, + 0.010976455174386501, + 0.006423037964850664, + 0.005860523786395788, + 0.01247511524707079, + 0.01180140022188425, + 0.007323491387069225, + 0.005764322821050882, + 0.01092479471117258, + 0.008866626769304276, + 0.02556513622403145, + 0.010328894481062889, + 0.011794636026024818, + 0.027843955904245377, + 0.0023316829465329647, + 0.003793510841205716, + 0.004313499666750431, + 0.00882985070347786, + 0.0024778267834335566, + 0.002450820989906788, + 0.006244589574635029, + 0.0042902156710624695, + 0.00703877629712224, + 0.03277371823787689, + 0.025325313210487366, + 0.010355000384151936, + 0.009077398106455803, + 0.010635136626660824, + 0.005905230063945055, + 0.008865393698215485, + 0.0036809321027249098, + 0.005164378322660923, + 0.002490853890776634, + 0.0028524312656372786, + 0.00394643098115921, + 0.0033353513572365046, + 0.0028003195766359568, + 0.002721987897530198, + 0.00882005412131548, + 0.007205584552139044, + 0.013875880278646946, + 0.01860685460269451, + 0.006622172426432371, + 0.009247750975191593, + 0.013811872340738773, + 0.008686776272952557, + 0.008189753629267216, + 0.005238314624875784, + 0.0321974977850914, + 0.010445446707308292, + 0.006446166895329952, + 0.007072444539517164, + 0.009166841395199299, + 0.012454815208911896, + 0.02952703833580017, + 0.08734661340713501, + 0.12662996351718903, + 0.03033226728439331, + 0.024985654279589653, + 0.015152218751609325, + 0.03317049890756607, + 0.018183834850788116, + 0.004547331482172012, + 0.00491287698969245, + 0.0037111539859324694, + 0.0024921856820583344, + 0.008134027943015099, + 0.005213940516114235, + 0.005444757640361786, + 0.00489048520103097, + 0.007628305349498987, + 0.009571308270096779, + 0.01409440953284502, + 0.006016851402819157, + 0.005417567677795887, + 0.003975820727646351, + 0.011794437654316425, + 0.005583834834396839 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 1 + ], + "selected_reward": -0.06202206760644913 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_13/index_tensor([777, 777])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_13/index_tensor([777, 777])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e1a8ec6dce8369609820ca1be47fe94584c2f255 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_13/index_tensor([777, 777])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, medium-long shot opens on a regal woman standing resolutely on a mossy rock ledge within a lush, vibrant canyon, drawn in the classic Miyazaki manga style. She has long, flowing black hair, partially secured by an ornate hosta, and her expression is one of calm, focused determination. She wears a sumptuous red silk robe embroidered with golden cloud motifs, which flutters gently in the breeze. The canyon around her is a masterpiece of soft, hand-painted aesthetic, with towering green cliffs, ancient, gnarled trees, and soft golden light filtering down. In her hands, she powerfully grasps a massive fan crafted from dark lacquered wood and heavy silk, painted with an ink dragon. She decisively swings the fan in an elegant yet incredibly powerful arc. As the swing is complete, a devastating mountain torrent is magically unleashed, erupting forward in a torrent. In stark contrast to the stylized world, the chaotic movement of the water is rendered in surreal, smooth detail, captured at a high frame rate. The camera dramatically zooms out to wide shots, revealing the sheer scale of the flood as it surges through the canyon, its tumbling, photorealistic waves and spray crashing against the stylized cliffs, creating a stunning and surreal fusion of artistic styles.", + "seed": 27060, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_13/index_777_seed_27060_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.017965144850313662, + "vlm_logprob": -0.017965144850313662, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.017965144850313662, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a regal woman with long, flowing black hair wearing a red silk robe with golden cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman standing on a mossy rock ledge in a lush canyon while holding a massive dark lacquered fan with an ink dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing the fan in an arc that magically unleashes a torrent of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water rendered in surreal, smooth, photorealistic detail with tumbling waves and spray, contrasting with the stylized world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera dramatically zoom out to wide shots revealing the scale of the flood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008686495013535023 + ], + "opposite_logprob_tokens": [ + -5.008686542510986 + ], + "logprob_reward": -0.008686495013535023, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.008686495013535023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10086897760629654, + 0.09065502882003784, + 0.02909410558640957, + 0.2146088182926178, + 0.01983713172376156, + 0.13159246742725372, + 0.020631853491067886, + 0.1923346072435379, + 0.10948480665683746, + 0.09089220315217972 + ], + "credit_map": [ + 0.007551395334303379, + 0.009796900674700737, + 0.020331740379333496, + 0.008784803561866283, + 0.012022552080452442, + 0.015856215730309486, + 0.007743756286799908, + 0.018781613558530807, + 0.00962448213249445, + 0.010051778517663479, + 0.009768684394657612, + 0.01745406538248062, + 0.009726433083415031, + 0.010482092387974262, + 0.013090232387185097, + 0.010457254946231842, + 0.0022115057799965143, + 0.0022446149960160255, + 0.007831749506294727, + 0.0022388091310858727, + 0.0028338097035884857, + 0.0023564002476632595, + 0.0018345313146710396, + 0.0075426846742630005, + 0.01328244712203741, + 0.008545216172933578, + 0.07890986651182175, + 0.009609418921172619, + 0.006369088310748339, + 0.010876973159611225, + 0.008475997485220432, + 0.07853979617357254, + 0.0017951548798009753, + 0.0015723153483122587, + 0.0049964431673288345, + 0.0016800242010504007, + 0.0018497499404475093, + 0.0015360619872808456, + 0.0019561005756258965, + 0.0044512818567454815, + 0.015041724778711796, + 0.00926443375647068, + 0.022763805463910103, + 0.0072369761765003204, + 0.011414249427616596, + 0.03587806224822998, + 0.00799034908413887, + 0.02200286090373993, + 0.004788179881870747, + 0.002478974871337414, + 0.003712574951350689, + 0.0015469184145331383, + 0.0021341999527066946, + 0.002470883075147867, + 0.0016135210171341896, + 0.0018866003956645727, + 0.010498607531189919, + 0.006833495572209358, + 0.02745499648153782, + 0.014990719966590405, + 0.006792132742702961, + 0.02901257574558258, + 0.021363480016589165, + 0.075388602912426, + 0.008777232840657234, + 0.008500095456838608, + 0.01207906473428011, + 0.01147883664816618, + 0.005823321174830198, + 0.008664899505674839, + 0.0260288305580616, + 0.028132520616054535, + 0.006047025788575411, + 0.005140936002135277, + 0.02014014683663845, + 0.012508636340498924, + 0.004871911834925413, + 0.012111859396100044, + 0.014306805096566677, + 0.015764879062771797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 5 + ], + "selected_reward": -0.008686495013535023 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a regal woman with long, flowing black hair wearing a red silk robe with golden cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman standing on a mossy rock ledge in a lush canyon while holding a massive dark lacquered fan with an ink dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing the fan in an arc that magically unleashes a torrent of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water rendered in surreal, smooth, photorealistic detail with tumbling waves and spray, contrasting with the stylized world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera dramatically zoom out to wide shots revealing the scale of the flood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004104642663151026 + ], + "opposite_logprob_tokens": [ + -5.5041046142578125 + ], + "logprob_reward": -0.004104642663151026, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004104642663151026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07268358021974564, + 0.05258406326174736, + 0.04594682529568672, + 0.08571923524141312, + 0.014236566610634327, + 0.29762956500053406, + 0.0150665994733572, + 0.10187077522277832, + 0.10638782382011414, + 0.20787496864795685 + ], + "credit_map": [ + 0.00430566631257534, + 0.005314679816365242, + 0.021642232313752174, + 0.00488070584833622, + 0.006799765396863222, + 0.006963247433304787, + 0.0040307361632585526, + 0.018746549263596535, + 0.0051061129197478294, + 0.004448390565812588, + 0.009102226234972477, + 0.007221778854727745, + 0.005270075984299183, + 0.005437231156975031, + 0.005760268308222294, + 0.010237975977361202, + 0.0037899615708738565, + 0.00518064433708787, + 0.009433813393115997, + 0.004367250949144363, + 0.00420001894235611, + 0.004555186722427607, + 0.003403358394280076, + 0.011016596108675003, + 0.005423899739980698, + 0.002755350200459361, + 0.03218984603881836, + 0.003224872052669525, + 0.0022401874884963036, + 0.004731879569590092, + 0.002751304768025875, + 0.03240189701318741, + 0.0011057429946959019, + 0.0009249876602552831, + 0.004447885788977146, + 0.000738494738470763, + 0.0011930727632716298, + 0.001036847592331469, + 0.0006434825481846929, + 0.004146052524447441, + 0.029067400842905045, + 0.020235033705830574, + 0.05284770950675011, + 0.018221575766801834, + 0.030067093670368195, + 0.08412408828735352, + 0.018170759081840515, + 0.044895902276039124, + 0.0031653637997806072, + 0.0018686658004298806, + 0.002459322102367878, + 0.0013071297435089946, + 0.0013558685313910246, + 0.0016437676968052983, + 0.0014455661876127124, + 0.0018209157278761268, + 0.006202156189829111, + 0.004569892305880785, + 0.010288777761161327, + 0.010199313051998615, + 0.005014356225728989, + 0.019044747576117516, + 0.013523804023861885, + 0.033027730882167816, + 0.009572148323059082, + 0.01330749224871397, + 0.013685026206076145, + 0.01011702511459589, + 0.007301179692149162, + 0.009819654747843742, + 0.017461275681853294, + 0.02512402832508087, + 0.004605661612004042, + 0.006230067927390337, + 0.009521282278001308, + 0.08783963322639465, + 0.004612749442458153, + 0.005899243056774139, + 0.07855477184057236, + 0.010611547157168388 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 8 + ], + "selected_reward": -0.004104642663151026 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a regal woman with long, flowing black hair wearing a red silk robe with golden cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman standing on a mossy rock ledge in a lush canyon while holding a massive dark lacquered fan with an ink dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing the fan in an arc that magically unleashes a torrent of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water rendered in surreal, smooth, photorealistic detail with tumbling waves and spray, contrasting with the stylized world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera dramatically zoom out to wide shots revealing the scale of the flood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015160269103944302 + ], + "opposite_logprob_tokens": [ + -6.501515865325928 + ], + "logprob_reward": -0.0015160269103944302, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015160269103944302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.023284170776605606, + 0.04865100234746933, + 0.006360032595694065, + 0.047203417867422104, + 0.013673299923539162, + 0.05104059353470802, + 0.007236850447952747, + 0.3602018356323242, + 0.08054473251104355, + 0.3618040680885315 + ], + "credit_map": [ + 0.0011959196999669075, + 0.001901863026432693, + 0.006183433346450329, + 0.0019929041154682636, + 0.002057525096461177, + 0.0024865830782800913, + 0.0016717351973056793, + 0.005794205237179995, + 0.004231806378811598, + 0.004088700748980045, + 0.012296288274228573, + 0.006469373125582933, + 0.0032427299302071333, + 0.003047357313334942, + 0.004776878748089075, + 0.010497868061065674, + 0.0008207701612263918, + 0.0007137741195037961, + 0.0009468505159020424, + 0.0009510040981695056, + 0.0007190468022599816, + 0.0007570151938125491, + 0.0007295070099644363, + 0.0007220649858936667, + 0.002894765930250287, + 0.002050639595836401, + 0.01666026934981346, + 0.002322303829714656, + 0.0018390138866379857, + 0.002576499478891492, + 0.00200343388132751, + 0.016856495290994644, + 0.0010221361881121993, + 0.0009958326118066907, + 0.004320974927395582, + 0.0009558431920595467, + 0.0009552484843879938, + 0.0009148461977019906, + 0.0008796254987828434, + 0.003628793405368924, + 0.005097554996609688, + 0.00334493862465024, + 0.009080838412046432, + 0.003323398996144533, + 0.004920589737594128, + 0.014598635025322437, + 0.002991205547004938, + 0.007683426607400179, + 0.0013093942543491721, + 0.0008947595488280058, + 0.0010113557800650597, + 0.0007624659920111299, + 0.0007577883661724627, + 0.0008003414259292185, + 0.000643143430352211, + 0.001057601417414844, + 0.020123010501265526, + 0.011867805384099483, + 0.06490211933851242, + 0.024549733847379684, + 0.010805514641106129, + 0.048344794660806656, + 0.03589886054396629, + 0.14370998740196228, + 0.007380529772490263, + 0.006811191793531179, + 0.014231676235795021, + 0.01182159036397934, + 0.005274460185319185, + 0.007047432009130716, + 0.010042238049209118, + 0.01793561317026615, + 0.011068597435951233, + 0.012258836068212986, + 0.019082428887486458, + 0.14912129938602448, + 0.008246039971709251, + 0.017381701618433, + 0.12190347164869308, + 0.022741694003343582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.0015160269103944302 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a regal woman with long, flowing black hair wearing a red silk robe with golden cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman standing on a mossy rock ledge in a lush canyon while holding a massive dark lacquered fan with an ink dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing the fan in an arc that magically unleashes a torrent of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water rendered in surreal, smooth, photorealistic detail with tumbling waves and spray, contrasting with the stylized world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera dramatically zoom out to wide shots revealing the scale of the flood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054945945739746094 + ], + "opposite_logprob_tokens": [ + -2.929945945739746 + ], + "logprob_reward": -0.054945945739746094, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054945945739746094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03425147384405136, + 0.01989421248435974, + 0.020304666832089424, + 0.021301353350281715, + 0.017380526289343834, + 0.06025373563170433, + 0.011077756993472576, + 0.4185464680194855, + 0.12518605589866638, + 0.2718037962913513 + ], + "credit_map": [ + 0.001872550928965211, + 0.002172840293496847, + 0.010419085621833801, + 0.0016565948026254773, + 0.003469982650130987, + 0.0034237049985677004, + 0.001199622405692935, + 0.010037092491984367, + 0.00215862481854856, + 0.002618200611323118, + 0.002214939333498478, + 0.003429784905165434, + 0.002069789217785001, + 0.0023832968436181545, + 0.0026920472737401724, + 0.002327529713511467, + 0.0022767221089452505, + 0.0017475467175245285, + 0.00373080070130527, + 0.0016264539444819093, + 0.0016472178976982832, + 0.003709414741024375, + 0.0017956816591322422, + 0.0037708294112235308, + 0.0014630514197051525, + 0.0008522699354216456, + 0.006962301209568977, + 0.0010166731663048267, + 0.0010747620835900307, + 0.001598106580786407, + 0.001006584265269339, + 0.007327605038881302, + 0.0012175953015685081, + 0.0011088413884863257, + 0.005563345272094011, + 0.001274701440706849, + 0.001125702983699739, + 0.0012238973286002874, + 0.001151075353845954, + 0.004715367220342159, + 0.006284559145569801, + 0.004202337935566902, + 0.010257333517074585, + 0.0035602485295385122, + 0.006135566625744104, + 0.017339786514639854, + 0.003601574804633856, + 0.008872328326106071, + 0.0010871542617678642, + 0.0009509792434982955, + 0.002244445728138089, + 0.0012489910004660487, + 0.0007574902265332639, + 0.0011982630239799619, + 0.001135580358095467, + 0.002454852918162942, + 0.022899897769093513, + 0.015650460496544838, + 0.06554372608661652, + 0.030804045498371124, + 0.014367355965077877, + 0.06005174294114113, + 0.044240836054086685, + 0.16498838365077972, + 0.01201893575489521, + 0.009459910914301872, + 0.015079942531883717, + 0.016664305701851845, + 0.007824692875146866, + 0.011838948354125023, + 0.02549774944782257, + 0.026801561936736107, + 0.006031031720340252, + 0.008366253226995468, + 0.015972748398780823, + 0.10851550847291946, + 0.006857539992779493, + 0.005774368066340685, + 0.10675397515296936, + 0.01353236474096775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.054945945739746094 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a regal woman with long, flowing black hair wearing a red silk robe with golden cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman standing on a mossy rock ledge in a lush canyon while holding a massive dark lacquered fan with an ink dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing the fan in an arc that magically unleashes a torrent of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water rendered in surreal, smooth, photorealistic detail with tumbling waves and spray, contrasting with the stylized world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera dramatically zoom out to wide shots revealing the scale of the flood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020572613924741745 + ], + "opposite_logprob_tokens": [ + -3.8955726623535156 + ], + "logprob_reward": -0.020572613924741745, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020572613924741745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02573070488870144, + 0.01971004717051983, + 0.014699500985443592, + 0.07716537266969681, + 0.0309336856007576, + 0.10780646651983261, + 0.038957253098487854, + 0.034359436482191086, + 0.188438281416893, + 0.46219927072525024 + ], + "credit_map": [ + 0.0023146397434175014, + 0.0032539081294089556, + 0.0022261266130954027, + 0.00390847772359848, + 0.0028463162016123533, + 0.004677874967455864, + 0.0032616569660604, + 0.003241701750084758, + 0.0012753934133797884, + 0.0011304590152576566, + 0.006082208827137947, + 0.0013409804087132215, + 0.0012801560806110501, + 0.0014472219627350569, + 0.0012857833644375205, + 0.005867843050509691, + 0.0019835354760289192, + 0.001219450612552464, + 0.002628485206514597, + 0.0011731814593076706, + 0.001323367585428059, + 0.0025821782182902098, + 0.001398308202624321, + 0.002390993991866708, + 0.004594574682414532, + 0.003888869658112526, + 0.02709071896970272, + 0.0037958500906825066, + 0.0032570473849773407, + 0.003911399282515049, + 0.0035462621599435806, + 0.02708064392209053, + 0.002807889599353075, + 0.0023736669681966305, + 0.009797684848308563, + 0.002119258278980851, + 0.0022010072134435177, + 0.002453902969136834, + 0.0014616450062021613, + 0.007718629203736782, + 0.01084217894822359, + 0.007638425100594759, + 0.01913980208337307, + 0.006230039522051811, + 0.010115782730281353, + 0.030493799597024918, + 0.006492936052381992, + 0.016853485256433487, + 0.005411967169493437, + 0.003985122311860323, + 0.007617779541760683, + 0.003610015381127596, + 0.0035256550181657076, + 0.004734475165605545, + 0.0029284765478223562, + 0.007143760100007057, + 0.004691042471677065, + 0.004197569098323584, + 0.006669575348496437, + 0.0032412821892648935, + 0.0026584623847156763, + 0.0042144120670855045, + 0.00343874329701066, + 0.005248344969004393, + 0.014182033948600292, + 0.018812378868460655, + 0.03047116845846176, + 0.03360590711236, + 0.015041419304907322, + 0.020681118592619896, + 0.027617471292614937, + 0.028026778250932693, + 0.03499554470181465, + 0.02366527169942856, + 0.13924816250801086, + 0.07082207500934601, + 0.02123347483575344, + 0.044426802545785904, + 0.06810995936393738, + 0.05969797819852829 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.020572613924741745 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium-long shot opens on a regal woman standing resolutely on a mossy rock ledge within a lush, vibrant canyon, drawn in the classic Miyazaki manga style. She has long, flowing black hair, partially secured by an ornate hosta, and her expression is one of calm, focused determination. She wears a sumptuous red silk robe embroidered with golden cloud motifs, which flutters gently in the breeze. The canyon around her is a masterpiece of soft, hand-painted aesthetic, with towering green cliffs, ancient, gnarled trees, and soft golden light filtering down. In her hands, she powerfully grasps a massive fan crafted from dark lacquered wood and heavy silk, painted with an ink dragon. She decisively swings the fan in an elegant yet incredibly powerful arc. As the swing is complete, a devastating mountain torrent is magically unleashed, erupting forward in a torrent. In stark contrast to the stylized world, the chaotic movement of the water is rendered in surreal, smooth detail, captured at a high frame rate. The camera dramatically zooms out to wide shots, revealing the sheer scale of the flood as it surges through the canyon, its tumbling, photorealistic waves and spray crashing against the stylized cliffs, creating a stunning and surreal fusion of artistic styles.", + "seed": 27061, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_13/index_777_seed_27061_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07730777610559017, + "vlm_logprob": -0.07730777610559017, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.07730777610559017, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a regal woman with long, flowing black hair wearing a red silk robe with golden cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman standing on a mossy rock ledge in a lush canyon while holding a massive dark lacquered fan with an ink dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing the fan in an arc that magically unleashes a torrent of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water rendered in surreal, smooth, photorealistic detail with tumbling waves and spray, contrasting with the stylized world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera dramatically zoom out to wide shots revealing the scale of the flood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011476687155663967 + ], + "opposite_logprob_tokens": [ + -4.761476516723633 + ], + "logprob_reward": -0.011476687155663967, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.011476687155663967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3484928607940674, + 0.05703805014491081, + 0.03936076909303665, + 0.027407117187976837, + 0.05121171101927757, + 0.04407953843474388, + 0.049009185284376144, + 0.0666707307100296, + 0.06619566679000854, + 0.25053441524505615 + ], + "credit_map": [ + 0.02776065655052662, + 0.01928665302693844, + 0.10556789487600327, + 0.01757570169866085, + 0.03028884343802929, + 0.09305678308010101, + 0.02627239003777504, + 0.028683919459581375, + 0.008150235749781132, + 0.0040411329828202724, + 0.011108801700174809, + 0.004102861043065786, + 0.005081550218164921, + 0.00822416227310896, + 0.006071052048355341, + 0.010258253663778305, + 0.0034118120092898607, + 0.004063955042511225, + 0.006547445431351662, + 0.0060114189982414246, + 0.004337500315159559, + 0.0044595482759177685, + 0.004471289459615946, + 0.00605779979377985, + 0.003114269347861409, + 0.0023425391409546137, + 0.003944307565689087, + 0.0033383413683623075, + 0.003240062855184078, + 0.002974843140691519, + 0.003417770378291607, + 0.005034984555095434, + 0.003792344592511654, + 0.011632916517555714, + 0.006197729613631964, + 0.006346921436488628, + 0.005319863557815552, + 0.005828763358294964, + 0.005665756296366453, + 0.0064274161122739315, + 0.004147723317146301, + 0.006872293073683977, + 0.004350207280367613, + 0.004137110896408558, + 0.0061025177128612995, + 0.005240154452621937, + 0.0064218114130198956, + 0.006807717960327864, + 0.008163523860275745, + 0.0028856354765594006, + 0.007512023206800222, + 0.002401133766397834, + 0.004191431682556868, + 0.004976307973265648, + 0.002789014484733343, + 0.016090115532279015, + 0.005109694320708513, + 0.00907458458095789, + 0.0086385328322649, + 0.009887325577437878, + 0.008665858767926693, + 0.005184757988899946, + 0.010756676085293293, + 0.009353306144475937, + 0.004767776932567358, + 0.011343137361109257, + 0.00340301264077425, + 0.010120614431798458, + 0.008505495265126228, + 0.004085086286067963, + 0.009776532649993896, + 0.014194011688232422, + 0.018424848094582558, + 0.03125300258398056, + 0.03089446760714054, + 0.04203140363097191, + 0.027446633204817772, + 0.02834637090563774, + 0.036939628422260284, + 0.03519805148243904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.011476687155663967 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a regal woman with long, flowing black hair wearing a red silk robe with golden cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman standing on a mossy rock ledge in a lush canyon while holding a massive dark lacquered fan with an ink dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing the fan in an arc that magically unleashes a torrent of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water rendered in surreal, smooth, photorealistic detail with tumbling waves and spray, contrasting with the stylized world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera dramatically zoom out to wide shots revealing the scale of the flood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032078970689326525 + ], + "opposite_logprob_tokens": [ + -5.753207683563232 + ], + "logprob_reward": -0.0032078970689326525, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032078970689326525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33140891790390015, + 0.037511713802814484, + 0.051392972469329834, + 0.05345411226153374, + 0.039710257202386856, + 0.028766367584466934, + 0.03219810873270035, + 0.06478717178106308, + 0.14880013465881348, + 0.21197019517421722 + ], + "credit_map": [ + 0.02707502245903015, + 0.013700304552912712, + 0.10448811203241348, + 0.013056722469627857, + 0.027639418840408325, + 0.09912874549627304, + 0.019085975363850594, + 0.027234641835093498, + 0.004668662324547768, + 0.002676765201613307, + 0.006075217854231596, + 0.0028279805555939674, + 0.003749176859855652, + 0.005494816694408655, + 0.004339190665632486, + 0.007679907139390707, + 0.004241688176989555, + 0.0050378101877868176, + 0.009051958099007607, + 0.007850328460335732, + 0.004832442384213209, + 0.0074141058139503, + 0.00568124046549201, + 0.00728340120986104, + 0.006333575583994389, + 0.0037879471201449633, + 0.01211934257298708, + 0.004782536532729864, + 0.003946990706026554, + 0.00781824067234993, + 0.005605223588645458, + 0.009060259908437729, + 0.00283858273178339, + 0.0076546259224414825, + 0.005617175716906786, + 0.005031498149037361, + 0.0034546786919236183, + 0.004780507180839777, + 0.004336567595601082, + 0.00599662633612752, + 0.0028754037339240313, + 0.003547241445630789, + 0.0032404630910605192, + 0.002683207392692566, + 0.004230462480336428, + 0.0035023645032197237, + 0.004170663189142942, + 0.004516563378274441, + 0.00524485670030117, + 0.004169817548245192, + 0.003909217659384012, + 0.0031256596557796, + 0.003574524773284793, + 0.0030244318768382072, + 0.003806320484727621, + 0.00534328306093812, + 0.00638892175629735, + 0.008459917269647121, + 0.0064432695508003235, + 0.009637776762247086, + 0.007419979199767113, + 0.005343849305063486, + 0.010590446181595325, + 0.010503011755645275, + 0.010740755125880241, + 0.021995868533849716, + 0.009939493611454964, + 0.025044528767466545, + 0.01571590267121792, + 0.009369989857077599, + 0.024627333506941795, + 0.03136627376079559, + 0.015594534575939178, + 0.02910866215825081, + 0.025922704488039017, + 0.03646618500351906, + 0.021294238045811653, + 0.01957198604941368, + 0.033044829964637756, + 0.030967075377702713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.0032078970689326525 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a regal woman with long, flowing black hair wearing a red silk robe with golden cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman standing on a mossy rock ledge in a lush canyon while holding a massive dark lacquered fan with an ink dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing the fan in an arc that magically unleashes a torrent of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water rendered in surreal, smooth, photorealistic detail with tumbling waves and spray, contrasting with the stylized world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera dramatically zoom out to wide shots revealing the scale of the flood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001515431678853929 + ], + "opposite_logprob_tokens": [ + -6.5015153884887695 + ], + "logprob_reward": -0.001515431678853929, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001515431678853929, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27279335260391235, + 0.021115608513355255, + 0.031750258058309555, + 0.02263595350086689, + 0.031273793429136276, + 0.02754882164299488, + 0.02452520839869976, + 0.08198484778404236, + 0.21634319424629211, + 0.27002885937690735 + ], + "credit_map": [ + 0.02279222197830677, + 0.009391603991389275, + 0.0915796086192131, + 0.00936095416545868, + 0.020482858642935753, + 0.09368462860584259, + 0.00893859751522541, + 0.016562912613153458, + 0.0022383262403309345, + 0.0013838979648426175, + 0.0029657536651939154, + 0.0015535004204139113, + 0.0018016559770330787, + 0.003409550990909338, + 0.0025519051123410463, + 0.005211020354181528, + 0.003939199261367321, + 0.0025255586951971054, + 0.004973825067281723, + 0.0035220561549067497, + 0.0031984997913241386, + 0.005040476564317942, + 0.00293255178257823, + 0.0056180949322879314, + 0.0028004120104014874, + 0.0022665751166641712, + 0.0027749354485422373, + 0.0028986171819269657, + 0.002323537366464734, + 0.0029809465631842613, + 0.0023903793189674616, + 0.004200551193207502, + 0.002846129471436143, + 0.0069563668221235275, + 0.0033306621480733156, + 0.004755441565066576, + 0.00331862922757864, + 0.003290099324658513, + 0.0037502688355743885, + 0.003026197897270322, + 0.002791787963360548, + 0.003957085777074099, + 0.003001992590725422, + 0.0026904556434601545, + 0.0034244479611516, + 0.0038646848406642675, + 0.0036381392274051905, + 0.004180226940661669, + 0.003641138318926096, + 0.0018055251566693187, + 0.0031334052328020334, + 0.0023384978994727135, + 0.0022629108279943466, + 0.0026433810126036406, + 0.0030296905897557735, + 0.005670661572366953, + 0.004795353394001722, + 0.008386525325477123, + 0.007071438245475292, + 0.013502965681254864, + 0.007215150166302919, + 0.0051685539074242115, + 0.019023127853870392, + 0.016821738332509995, + 0.015149295330047607, + 0.04242202267050743, + 0.012475647032260895, + 0.031016897410154343, + 0.026340624317526817, + 0.012489194981753826, + 0.03450681269168854, + 0.04194271191954613, + 0.019258668646216393, + 0.0406239777803421, + 0.030548999086022377, + 0.05056458339095116, + 0.025937454774975777, + 0.02181338705122471, + 0.04658757522702217, + 0.034694232046604156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.001515431678853929 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a regal woman with long, flowing black hair wearing a red silk robe with golden cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman standing on a mossy rock ledge in a lush canyon while holding a massive dark lacquered fan with an ink dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing the fan in an arc that magically unleashes a torrent of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water rendered in surreal, smooth, photorealistic detail with tumbling waves and spray, contrasting with the stylized world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera dramatically zoom out to wide shots revealing the scale of the flood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813219726085663 + ], + "opposite_logprob_tokens": [ + -1.4063220024108887 + ], + "logprob_reward": -0.2813219726085663, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813219726085663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2175188809633255, + 0.06838375329971313, + 0.026337243616580963, + 0.026177261024713516, + 0.0414748415350914, + 0.013841859064996243, + 0.020961161702871323, + 0.057644300162792206, + 0.16640232503414154, + 0.361258327960968 + ], + "credit_map": [ + 0.01799982786178589, + 0.007108132820576429, + 0.0785624161362648, + 0.00747938547283411, + 0.015279445797204971, + 0.06772097200155258, + 0.011281891725957394, + 0.012086797505617142, + 0.005929253064095974, + 0.0026397756300866604, + 0.02489337883889675, + 0.003780838567763567, + 0.002776371082291007, + 0.021033646538853645, + 0.0027914494276046753, + 0.004539041314274073, + 0.003933176398277283, + 0.00317953759804368, + 0.004468295257538557, + 0.0023249643854796886, + 0.0023549997713416815, + 0.0037422708701342344, + 0.0019995172042399645, + 0.004334482830017805, + 0.003157982137054205, + 0.0018590273102745414, + 0.004917293321341276, + 0.0024840242695063353, + 0.0029467293061316013, + 0.0038805645890533924, + 0.0023823704104870558, + 0.0045492686331272125, + 0.0035249809734523296, + 0.009518817998468876, + 0.0048385970294475555, + 0.005227983929216862, + 0.0048449295572936535, + 0.004793383181095123, + 0.004385334439575672, + 0.004340809769928455, + 0.001137253362685442, + 0.0020230314694344997, + 0.0018091145902872086, + 0.0015320759266614914, + 0.0020580573473125696, + 0.0016875098226591945, + 0.0019884060602635145, + 0.001606410020031035, + 0.0021857048850506544, + 0.0015075954142957926, + 0.004313374403864145, + 0.0016828514635562897, + 0.0020815811585634947, + 0.0024311873130500317, + 0.0020396201871335506, + 0.004719245713204145, + 0.004120150581002235, + 0.006551108323037624, + 0.005434602964669466, + 0.005559336859732866, + 0.00784611888229847, + 0.005149255506694317, + 0.011580074205994606, + 0.011403651908040047, + 0.010628049261868, + 0.03154163807630539, + 0.009512417949736118, + 0.02352428436279297, + 0.023548642173409462, + 0.008237890899181366, + 0.026601474732160568, + 0.03280792757868767, + 0.026067212224006653, + 0.03898713365197182, + 0.04331440106034279, + 0.058835603296756744, + 0.042659468948841095, + 0.044190891087055206, + 0.053099244832992554, + 0.054104357957839966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.2813219726085663 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a regal woman with long, flowing black hair wearing a red silk robe with golden cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman standing on a mossy rock ledge in a lush canyon while holding a massive dark lacquered fan with an ink dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing the fan in an arc that magically unleashes a torrent of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water rendered in surreal, smooth, photorealistic detail with tumbling waves and spray, contrasting with the stylized world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera dramatically zoom out to wide shots revealing the scale of the flood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08901689201593399 + ], + "opposite_logprob_tokens": [ + -2.464016914367676 + ], + "logprob_reward": -0.08901689201593399, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901689201593399, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052152786403894424, + 0.08985109627246857, + 0.03209517523646355, + 0.026164952665567398, + 0.03340453654527664, + 0.06448991596698761, + 0.06956995278596878, + 0.11230529844760895, + 0.40007221698760986, + 0.1198941245675087 + ], + "credit_map": [ + 0.007260635960847139, + 0.005757195875048637, + 0.01029955968260765, + 0.004161630291491747, + 0.005719506181776524, + 0.005925922188907862, + 0.0064618587493896484, + 0.006566478405147791, + 0.007284026592969894, + 0.0036888231988996267, + 0.024993088096380234, + 0.004975450690835714, + 0.004036924801766872, + 0.03193717077374458, + 0.0038758590817451477, + 0.009059754200279713, + 0.005488067865371704, + 0.0023664541076868773, + 0.004113196860998869, + 0.0029272560495883226, + 0.00359322433359921, + 0.004222559276968241, + 0.0031166200060397387, + 0.006267794873565435, + 0.0030233615543693304, + 0.002175957430154085, + 0.004764624871313572, + 0.002547649433836341, + 0.002736323745921254, + 0.004053956363350153, + 0.002873658202588558, + 0.003989420831203461, + 0.003551450790837407, + 0.00634972658008337, + 0.004539936780929565, + 0.004753456450998783, + 0.00338851404376328, + 0.0038411389105021954, + 0.003494380507618189, + 0.0034859320148825645, + 0.0051648057997226715, + 0.010472942143678665, + 0.006240073591470718, + 0.007011894136667252, + 0.007709558587521315, + 0.008816836401820183, + 0.008527170866727829, + 0.010546629317104816, + 0.010363629087805748, + 0.005240745842456818, + 0.016678236424922943, + 0.004438991192728281, + 0.005443623289465904, + 0.005911875516176224, + 0.006362071260809898, + 0.015130780637264252, + 0.008907228708267212, + 0.015887800604104996, + 0.005488501396030188, + 0.013633407652378082, + 0.021953364834189415, + 0.006879314314574003, + 0.021324297413229942, + 0.018231384456157684, + 0.023139886558055878, + 0.05096099153161049, + 0.024152087047696114, + 0.07255906611680984, + 0.037037089467048645, + 0.02191285602748394, + 0.0732492133975029, + 0.09706103801727295, + 0.01048812735825777, + 0.021603859961032867, + 0.013038490898907185, + 0.01700783520936966, + 0.012600740417838097, + 0.009272929280996323, + 0.02233624830842018, + 0.013545883819460869 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.08901689201593399 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium-long shot opens on a regal woman standing resolutely on a mossy rock ledge within a lush, vibrant canyon, drawn in the classic Miyazaki manga style. She has long, flowing black hair, partially secured by an ornate hosta, and her expression is one of calm, focused determination. She wears a sumptuous red silk robe embroidered with golden cloud motifs, which flutters gently in the breeze. The canyon around her is a masterpiece of soft, hand-painted aesthetic, with towering green cliffs, ancient, gnarled trees, and soft golden light filtering down. In her hands, she powerfully grasps a massive fan crafted from dark lacquered wood and heavy silk, painted with an ink dragon. She decisively swings the fan in an elegant yet incredibly powerful arc. As the swing is complete, a devastating mountain torrent is magically unleashed, erupting forward in a torrent. In stark contrast to the stylized world, the chaotic movement of the water is rendered in surreal, smooth detail, captured at a high frame rate. The camera dramatically zooms out to wide shots, revealing the sheer scale of the flood as it surges through the canyon, its tumbling, photorealistic waves and spray crashing against the stylized cliffs, creating a stunning and surreal fusion of artistic styles.", + "seed": 27062, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_13/index_777_seed_27062_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08325402536429465, + "vlm_logprob": -0.08325402536429465, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.08325402536429465, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a regal woman with long, flowing black hair wearing a red silk robe with golden cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman standing on a mossy rock ledge in a lush canyon while holding a massive dark lacquered fan with an ink dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing the fan in an arc that magically unleashes a torrent of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water rendered in surreal, smooth, photorealistic detail with tumbling waves and spray, contrasting with the stylized world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera dramatically zoom out to wide shots revealing the scale of the flood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008452828973531723 + ], + "opposite_logprob_tokens": [ + -5.008452892303467 + ], + "logprob_reward": -0.008452828973531723, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.008452828973531723, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27251768112182617, + 0.056329451501369476, + 0.033100426197052, + 0.02458943985402584, + 0.03678683564066887, + 0.09408629685640335, + 0.04809890314936638, + 0.2305528223514557, + 0.12589623034000397, + 0.0780419111251831 + ], + "credit_map": [ + 0.042182765901088715, + 0.028990766033530235, + 0.04233141243457794, + 0.029353976249694824, + 0.037414100021123886, + 0.032159313559532166, + 0.03272783011198044, + 0.027357541024684906, + 0.008065101690590382, + 0.0039321016520261765, + 0.00968052539974451, + 0.006267559248954058, + 0.005521444138139486, + 0.007307192776352167, + 0.008560114540159702, + 0.006995412986725569, + 0.004088414367288351, + 0.0022908744867891073, + 0.0058487835340201855, + 0.003625139594078064, + 0.0032251798547804356, + 0.004490906838327646, + 0.005129022058099508, + 0.004402106627821922, + 0.0026926579885184765, + 0.0018135977443307638, + 0.004486825782805681, + 0.0029323252383619547, + 0.0024476987309753895, + 0.0037325252778828144, + 0.0035424414090812206, + 0.002941368380561471, + 0.0051725623197853565, + 0.0025736819952726364, + 0.005571022629737854, + 0.003894321620464325, + 0.0031527928076684475, + 0.005384746938943863, + 0.003457564627751708, + 0.007580143865197897, + 0.01328191626816988, + 0.013599161058664322, + 0.011972436681389809, + 0.009596612304449081, + 0.009887654334306717, + 0.014912603422999382, + 0.010480350814759731, + 0.010355561040341854, + 0.007464221213012934, + 0.005160330329090357, + 0.005893859080970287, + 0.004624924622476101, + 0.007314203307032585, + 0.00511244498193264, + 0.007057331968098879, + 0.0054715899750590324, + 0.015299303457140923, + 0.02877167798578739, + 0.028406362980604172, + 0.02360142022371292, + 0.035926222801208496, + 0.02491012215614319, + 0.034273967146873474, + 0.03936371952295303, + 0.012622076086699963, + 0.012923422269523144, + 0.015803415328264236, + 0.019163992255926132, + 0.01997007057070732, + 0.012245032005012035, + 0.01679410971701145, + 0.01637411303818226, + 0.010696441866457462, + 0.008001088164746761, + 0.006754362490028143, + 0.013872655108571053, + 0.009126557037234306, + 0.007543652784079313, + 0.010461331345140934, + 0.011585819534957409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.008452828973531723 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a regal woman with long, flowing black hair wearing a red silk robe with golden cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman standing on a mossy rock ledge in a lush canyon while holding a massive dark lacquered fan with an ink dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing the fan in an arc that magically unleashes a torrent of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water rendered in surreal, smooth, photorealistic detail with tumbling waves and spray, contrasting with the stylized world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera dramatically zoom out to wide shots revealing the scale of the flood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031989850103855133 + ], + "opposite_logprob_tokens": [ + -5.753199100494385 + ], + "logprob_reward": -0.0031989850103855133, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031989850103855133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1726541519165039, + 0.058737270534038544, + 0.03972658887505531, + 0.022639574483036995, + 0.025049852207303047, + 0.03830978646874428, + 0.08181669563055038, + 0.2916863262653351, + 0.11002460867166519, + 0.15935511887073517 + ], + "credit_map": [ + 0.030560513958334923, + 0.016517778858542442, + 0.03272480517625809, + 0.01512244250625372, + 0.018184268847107887, + 0.022294364869594574, + 0.015612487681210041, + 0.02163749746978283, + 0.008774611167609692, + 0.0049139587208628654, + 0.008710592053830624, + 0.00644828611984849, + 0.0063279299065470695, + 0.006854391191154718, + 0.008848821744322777, + 0.007858679629862309, + 0.006275798659771681, + 0.004114461597055197, + 0.0049986341036856174, + 0.004369507078081369, + 0.00508230971172452, + 0.004286599811166525, + 0.006258350796997547, + 0.00434092478826642, + 0.003798677586019039, + 0.0018032985972240567, + 0.0032126023434102535, + 0.002036199439316988, + 0.002126034116372466, + 0.0034713975619524717, + 0.0026515007484704256, + 0.0035398630425333977, + 0.0035572557244449854, + 0.0026004298124462366, + 0.003152363235130906, + 0.002765842480584979, + 0.0035214785020798445, + 0.0029190147761255503, + 0.0034777005203068256, + 0.003055766923353076, + 0.005767675116658211, + 0.005101692862808704, + 0.00663605285808444, + 0.003574100788682699, + 0.003145248629152775, + 0.005371761508285999, + 0.0032826915849000216, + 0.005430564284324646, + 0.013495532795786858, + 0.007976192981004715, + 0.012808418832719326, + 0.010684036649763584, + 0.00656861113384366, + 0.01009452436119318, + 0.009365161880850792, + 0.010824211873114109, + 0.019357983022928238, + 0.03165478631854057, + 0.036030933260917664, + 0.03637704998254776, + 0.041204486042261124, + 0.03309343010187149, + 0.043549809604883194, + 0.05041782185435295, + 0.00975991040468216, + 0.01731215789914131, + 0.013007991015911102, + 0.014577380381524563, + 0.01532684825360775, + 0.009221788495779037, + 0.014953378587961197, + 0.015865158289670944, + 0.01539652794599533, + 0.013559612445533276, + 0.010233287699520588, + 0.027015114203095436, + 0.023461351171135902, + 0.021599620580673218, + 0.021014289930462837, + 0.027075311169028282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.0031989850103855133 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a regal woman with long, flowing black hair wearing a red silk robe with golden cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman standing on a mossy rock ledge in a lush canyon while holding a massive dark lacquered fan with an ink dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing the fan in an arc that magically unleashes a torrent of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water rendered in surreal, smooth, photorealistic detail with tumbling waves and spray, contrasting with the stylized world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera dramatically zoom out to wide shots revealing the scale of the flood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021975908894091845 + ], + "opposite_logprob_tokens": [ + -6.127197742462158 + ], + "logprob_reward": -0.0021975908894091845, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021975908894091845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07626751065254211, + 0.022629953920841217, + 0.018066242337226868, + 0.01612890511751175, + 0.01616887003183365, + 0.03056355193257332, + 0.038602955639362335, + 0.32479411363601685, + 0.2851446866989136, + 0.17163316905498505 + ], + "credit_map": [ + 0.012593910098075867, + 0.007781302090734243, + 0.013599464669823647, + 0.0064480057917535305, + 0.007214107550680637, + 0.010593962855637074, + 0.007583905011415482, + 0.010452856309711933, + 0.0034467328805476427, + 0.0016604968113824725, + 0.003871601540595293, + 0.002256651408970356, + 0.001711102551780641, + 0.004175212234258652, + 0.0026810248382389545, + 0.0028271309565752745, + 0.0027152153197675943, + 0.0012622285867109895, + 0.002492882078513503, + 0.001703336602076888, + 0.0015912232920527458, + 0.0029661860316991806, + 0.0021237139590084553, + 0.0032114561181515455, + 0.002220297697931528, + 0.0014822172233834863, + 0.002068679314106703, + 0.0013826685026288033, + 0.001982800429686904, + 0.0028473944403231144, + 0.002114448929205537, + 0.002030400326475501, + 0.0022721029818058014, + 0.0011132495710626245, + 0.0024113168474286795, + 0.0014380804495885968, + 0.0014246756909415126, + 0.002514795633032918, + 0.001927022822201252, + 0.003067625919356942, + 0.004594018217176199, + 0.00303481868468225, + 0.005957372952252626, + 0.0022557643242180347, + 0.002100385259836912, + 0.005884788930416107, + 0.0020255271811038256, + 0.004710874054580927, + 0.005914774257689714, + 0.004573899321258068, + 0.003731531323865056, + 0.0055334740318357944, + 0.0048597813583910465, + 0.004406420513987541, + 0.00435684947296977, + 0.005226225592195988, + 0.029322601854801178, + 0.036045774817466736, + 0.04589582979679108, + 0.04709707945585251, + 0.04156269133090973, + 0.032076433300971985, + 0.05055108293890953, + 0.0422426201403141, + 0.023692328482866287, + 0.026839056983590126, + 0.02406172640621662, + 0.03582261875271797, + 0.04860169440507889, + 0.03543976694345474, + 0.04147038236260414, + 0.0492171086370945, + 0.015549733303487301, + 0.01763300411403179, + 0.011465616524219513, + 0.02065448649227619, + 0.02401064708828926, + 0.02643159218132496, + 0.02548946626484394, + 0.03039860911667347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.0021975908894091845 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a regal woman with long, flowing black hair wearing a red silk robe with golden cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman standing on a mossy rock ledge in a lush canyon while holding a massive dark lacquered fan with an ink dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing the fan in an arc that magically unleashes a torrent of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water rendered in surreal, smooth, photorealistic detail with tumbling waves and spray, contrasting with the stylized world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera dramatically zoom out to wide shots revealing the scale of the flood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31341397762298584 + ], + "opposite_logprob_tokens": [ + -1.3134139776229858 + ], + "logprob_reward": -0.31341397762298584, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31341397762298584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12645214796066284, + 0.02870607003569603, + 0.028621405363082886, + 0.03381852060556412, + 0.02042226865887642, + 0.07845254987478256, + 0.060339659452438354, + 0.3225909173488617, + 0.1360025256872177, + 0.164593905210495 + ], + "credit_map": [ + 0.0198261309415102, + 0.011065049096941948, + 0.016615966334939003, + 0.016114311292767525, + 0.01823645643889904, + 0.014329586178064346, + 0.01640170067548752, + 0.013862963765859604, + 0.00510193919762969, + 0.0024370199535042048, + 0.003713645739480853, + 0.0035060627851635218, + 0.003574744565412402, + 0.003899156814441085, + 0.0031654739286750555, + 0.0033080310095101595, + 0.005823767744004726, + 0.003030957654118538, + 0.0036706558894366026, + 0.003183765336871147, + 0.0038809585385024548, + 0.0032226303592324257, + 0.003062774892896414, + 0.0027458970434963703, + 0.009112942032516003, + 0.002605125540867448, + 0.005860773380845785, + 0.002383063081651926, + 0.002533189719542861, + 0.004567834082990885, + 0.0032350029796361923, + 0.003520591650158167, + 0.003352813422679901, + 0.0017648739740252495, + 0.0036345331463962793, + 0.002036529593169689, + 0.0013384580379351974, + 0.0029437527991831303, + 0.002371781039983034, + 0.0029795279260724783, + 0.014450491406023502, + 0.005453110206872225, + 0.016311530023813248, + 0.0060594952665269375, + 0.004657126497477293, + 0.014304202981293201, + 0.004697774071246386, + 0.01251881942152977, + 0.010058860294520855, + 0.0073745897971093655, + 0.007094286847859621, + 0.007590561173856258, + 0.005210247356444597, + 0.0068916757591068745, + 0.008274824358522892, + 0.007844616658985615, + 0.02597549930214882, + 0.03132176026701927, + 0.04581611230969429, + 0.04652578756213188, + 0.05038396641612053, + 0.03746140003204346, + 0.04735952988266945, + 0.03774688020348549, + 0.01198386587202549, + 0.01360055897384882, + 0.018959244713187218, + 0.027470842003822327, + 0.014201780781149864, + 0.009545255452394485, + 0.019867727532982826, + 0.02037326991558075, + 0.02351810596883297, + 0.016332875937223434, + 0.010350427590310574, + 0.03447616472840309, + 0.015613283962011337, + 0.010882444679737091, + 0.025144996121525764, + 0.028275609016418457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.31341397762298584 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a regal woman with long, flowing black hair wearing a red silk robe with golden cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman standing on a mossy rock ledge in a lush canyon while holding a massive dark lacquered fan with an ink dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing the fan in an arc that magically unleashes a torrent of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water rendered in surreal, smooth, photorealistic detail with tumbling waves and spray, contrasting with the stylized world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera dramatically zoom out to wide shots revealing the scale of the flood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08900674432516098 + ], + "opposite_logprob_tokens": [ + -2.4640066623687744 + ], + "logprob_reward": -0.08900674432516098, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08900674432516098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05964783951640129, + 0.01848217472434044, + 0.022095423191785812, + 0.01609771139919758, + 0.02312609553337097, + 0.024685796350240707, + 0.03015221655368805, + 0.24055390059947968, + 0.38451555371284485, + 0.180643230676651 + ], + "credit_map": [ + 0.01405241433531046, + 0.005015128292143345, + 0.007933374494314194, + 0.005646114703267813, + 0.004645944107323885, + 0.007963458076119423, + 0.0069819167256355286, + 0.00740948598831892, + 0.0032954211346805096, + 0.0013443889329209924, + 0.0024118483997881413, + 0.0015954936388880014, + 0.0017210581572726369, + 0.002882827538996935, + 0.002153765643015504, + 0.0030773726757615805, + 0.003269660985097289, + 0.0019052679417654872, + 0.003009002422913909, + 0.0019179665250703692, + 0.0029161295387893915, + 0.002806194359436631, + 0.003022013697773218, + 0.0032491872552782297, + 0.00275040534324944, + 0.0015011562500149012, + 0.001832044217735529, + 0.0016696070088073611, + 0.0022003338672220707, + 0.002129322150722146, + 0.002146635204553604, + 0.0018682068912312388, + 0.003413876285776496, + 0.002411758992820978, + 0.003286750055849552, + 0.0021587132941931486, + 0.0032556969672441483, + 0.0028301558922976255, + 0.0027271106373518705, + 0.0030420348048210144, + 0.0037718929816037416, + 0.004363957792520523, + 0.003523602383211255, + 0.002626195549964905, + 0.001823383616283536, + 0.0032731478568166494, + 0.0017125606536865234, + 0.0035910550504922867, + 0.004445401951670647, + 0.003355293069034815, + 0.0038898952770978212, + 0.003276222152635455, + 0.003345020581036806, + 0.003286088816821575, + 0.004540293477475643, + 0.004014002159237862, + 0.020009422674775124, + 0.02159958705306053, + 0.03817650303244591, + 0.03161516413092613, + 0.037987180054187775, + 0.021958302706480026, + 0.04368983209133148, + 0.025517908856272697, + 0.03645694628357887, + 0.05165133252739906, + 0.047685399651527405, + 0.06085733324289322, + 0.042564064264297485, + 0.03683535382151604, + 0.05295291915535927, + 0.0555121973156929, + 0.02118033729493618, + 0.01942162588238716, + 0.013401363976299763, + 0.04140708968043327, + 0.018464818596839905, + 0.010344046168029308, + 0.028356824070215225, + 0.028067121282219887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.08900674432516098 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium-long shot opens on a regal woman standing resolutely on a mossy rock ledge within a lush, vibrant canyon, drawn in the classic Miyazaki manga style. She has long, flowing black hair, partially secured by an ornate hosta, and her expression is one of calm, focused determination. She wears a sumptuous red silk robe embroidered with golden cloud motifs, which flutters gently in the breeze. The canyon around her is a masterpiece of soft, hand-painted aesthetic, with towering green cliffs, ancient, gnarled trees, and soft golden light filtering down. In her hands, she powerfully grasps a massive fan crafted from dark lacquered wood and heavy silk, painted with an ink dragon. She decisively swings the fan in an elegant yet incredibly powerful arc. As the swing is complete, a devastating mountain torrent is magically unleashed, erupting forward in a torrent. In stark contrast to the stylized world, the chaotic movement of the water is rendered in surreal, smooth detail, captured at a high frame rate. The camera dramatically zooms out to wide shots, revealing the sheer scale of the flood as it surges through the canyon, its tumbling, photorealistic waves and spray crashing against the stylized cliffs, creating a stunning and surreal fusion of artistic styles.", + "seed": 27063, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_13/index_777_seed_27063_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10160396359860897, + "vlm_logprob": -0.10160396359860897, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.10160396359860897, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a regal woman with long, flowing black hair wearing a red silk robe with golden cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman standing on a mossy rock ledge in a lush canyon while holding a massive dark lacquered fan with an ink dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing the fan in an arc that magically unleashes a torrent of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water rendered in surreal, smooth, photorealistic detail with tumbling waves and spray, contrasting with the stylized world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera dramatically zoom out to wide shots revealing the scale of the flood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05575273558497429 + ], + "opposite_logprob_tokens": [ + -2.930752754211426 + ], + "logprob_reward": -0.05575273558497429, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05575273558497429, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29732510447502136, + 0.08474797755479813, + 0.059180352836847305, + 0.02929033152759075, + 0.18324889242649078, + 0.15434779226779938, + 0.03456319868564606, + 0.06798328459262848, + 0.051467981189489365, + 0.037845101207494736 + ], + "credit_map": [ + 0.024670906364917755, + 0.025112338364124298, + 0.062158845365047455, + 0.017318500205874443, + 0.05882329121232033, + 0.05179251730442047, + 0.03327520191669464, + 0.024173500016331673, + 0.008167925290763378, + 0.008216138929128647, + 0.013986045494675636, + 0.006046721711754799, + 0.018945250660181046, + 0.011858134530484676, + 0.010656211525201797, + 0.006871551275253296, + 0.00636274041607976, + 0.006751082371920347, + 0.009094785898923874, + 0.0065396553836762905, + 0.00985868088901043, + 0.007691431324928999, + 0.006144995801150799, + 0.006736983545124531, + 0.003430198645219207, + 0.00214388151653111, + 0.004763936158269644, + 0.002444551093503833, + 0.005251217633485794, + 0.004566107410937548, + 0.003229678375646472, + 0.003460758598521352, + 0.02112276665866375, + 0.01407468132674694, + 0.02638888731598854, + 0.02377408929169178, + 0.03633207827806473, + 0.017461085692048073, + 0.026134159415960312, + 0.01796114444732666, + 0.01911081187427044, + 0.011797096580266953, + 0.02778201922774315, + 0.014377348124980927, + 0.024832578375935555, + 0.014487311244010925, + 0.029369037598371506, + 0.012591595761477947, + 0.005029977764934301, + 0.0032991047482937574, + 0.00562649592757225, + 0.005232428200542927, + 0.004394186660647392, + 0.0029356633313000202, + 0.0044137416407465935, + 0.003631601808592677, + 0.010975396260619164, + 0.005736297462135553, + 0.008512799628078938, + 0.0072778137400746346, + 0.010440913029015064, + 0.008265442214906216, + 0.009662625379860401, + 0.007112002000212669, + 0.005223041865974665, + 0.004720502067357302, + 0.008733125403523445, + 0.010232759639620781, + 0.0043556238524615765, + 0.0035756779834628105, + 0.008958213031291962, + 0.005669036414474249, + 0.00452828872948885, + 0.003724474459886551, + 0.005406402517110109, + 0.006792236585170031, + 0.0031977728940546513, + 0.004013527650386095, + 0.0059553394094109535, + 0.004227059427648783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.05575273558497429 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a regal woman with long, flowing black hair wearing a red silk robe with golden cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman standing on a mossy rock ledge in a lush canyon while holding a massive dark lacquered fan with an ink dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing the fan in an arc that magically unleashes a torrent of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water rendered in surreal, smooth, photorealistic detail with tumbling waves and spray, contrasting with the stylized world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera dramatically zoom out to wide shots revealing the scale of the flood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011087238788604736 + ], + "opposite_logprob_tokens": [ + -4.511087417602539 + ], + "logprob_reward": -0.011087238788604736, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011087238788604736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17512834072113037, + 0.07261152565479279, + 0.11324132978916168, + 0.03915198892354965, + 0.17240875959396362, + 0.06281814724206924, + 0.05294831097126007, + 0.09897075593471527, + 0.079733707010746, + 0.1329871267080307 + ], + "credit_map": [ + 0.014975124970078468, + 0.009839557111263275, + 0.053529128432273865, + 0.008598357439041138, + 0.021332696080207825, + 0.044279661029577255, + 0.011540896259248257, + 0.011032918468117714, + 0.006164131686091423, + 0.005889301188290119, + 0.013462991453707218, + 0.005402218550443649, + 0.016587207093834877, + 0.009926902130246162, + 0.008549868129193783, + 0.0066289049573242664, + 0.009378485381603241, + 0.012852467596530914, + 0.015616563148796558, + 0.010493610054254532, + 0.015548761934041977, + 0.01176133006811142, + 0.02023562230169773, + 0.01735449582338333, + 0.0037464469205588102, + 0.0034745768643915653, + 0.005921343341469765, + 0.003720921464264393, + 0.0055572036653757095, + 0.005166491027921438, + 0.006465618032962084, + 0.005099390167742968, + 0.015557745471596718, + 0.036804910749197006, + 0.04410822317004204, + 0.01568281464278698, + 0.01961284503340721, + 0.014026208780705929, + 0.01706732250750065, + 0.009548686444759369, + 0.010162449441850185, + 0.004600207321345806, + 0.01228630542755127, + 0.005481410305947065, + 0.007948171347379684, + 0.007398503366857767, + 0.009626248851418495, + 0.005314852111041546, + 0.012714642100036144, + 0.003663667244836688, + 0.010679804719984531, + 0.005432457663118839, + 0.005801379680633545, + 0.006514144130051136, + 0.0032167977187782526, + 0.004925422836095095, + 0.015704108402132988, + 0.010011259466409683, + 0.011606714688241482, + 0.010361257940530777, + 0.01491138618439436, + 0.012860102578997612, + 0.013746583834290504, + 0.009769346565008163, + 0.007689386140555143, + 0.007043711841106415, + 0.012518944218754768, + 0.016129743307828903, + 0.0060817161574959755, + 0.006977743934839964, + 0.014918333850800991, + 0.008374132215976715, + 0.019562089815735817, + 0.010210251435637474, + 0.014901177957654, + 0.02903694100677967, + 0.00924903154373169, + 0.011848492547869682, + 0.022852271795272827, + 0.015326866880059242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.011087238788604736 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a regal woman with long, flowing black hair wearing a red silk robe with golden cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman standing on a mossy rock ledge in a lush canyon while holding a massive dark lacquered fan with an ink dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing the fan in an arc that magically unleashes a torrent of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water rendered in surreal, smooth, photorealistic detail with tumbling waves and spray, contrasting with the stylized world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera dramatically zoom out to wide shots revealing the scale of the flood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011074976995587349 + ], + "opposite_logprob_tokens": [ + -4.511075019836426 + ], + "logprob_reward": -0.011074976995587349, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011074976995587349, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1271742284297943, + 0.024309782311320305, + 0.06239161267876625, + 0.025469033047556877, + 0.11739031225442886, + 0.053667038679122925, + 0.036592014133930206, + 0.18542829155921936, + 0.26891058683395386, + 0.09866712242364883 + ], + "credit_map": [ + 0.025167230516672134, + 0.00738384947180748, + 0.0382649302482605, + 0.005975618027150631, + 0.008881578221917152, + 0.027007028460502625, + 0.006614691577851772, + 0.007879307493567467, + 0.005437381565570831, + 0.0015376487281173468, + 0.005226554814726114, + 0.0018341997638344765, + 0.0023105149157345295, + 0.003851446323096752, + 0.002019514562562108, + 0.00209252187050879, + 0.009463080205023289, + 0.004664377309381962, + 0.009751340374350548, + 0.0054646749049425125, + 0.007966522127389908, + 0.008689137175679207, + 0.009035451337695122, + 0.007357030175626278, + 0.0032264741603285074, + 0.0020432923920452595, + 0.0043400987051427364, + 0.0024712735321372747, + 0.0037345660384744406, + 0.0028466652147471905, + 0.0043627130798995495, + 0.002443948294967413, + 0.013097613118588924, + 0.03083719126880169, + 0.03588631749153137, + 0.009703319519758224, + 0.00852510891854763, + 0.008421423844993114, + 0.005058522801846266, + 0.005860813893377781, + 0.010486187413334846, + 0.0043670362792909145, + 0.010031643323600292, + 0.003738138824701309, + 0.007953507825732231, + 0.006578623317182064, + 0.006362501531839371, + 0.004149401560425758, + 0.004301164299249649, + 0.002359547885134816, + 0.00511100422590971, + 0.004402507096529007, + 0.009016796015202999, + 0.0034280919935554266, + 0.004889370407909155, + 0.0030835303477942944, + 0.02588590979576111, + 0.015393147245049477, + 0.03403191268444061, + 0.019833553582429886, + 0.021759290248155594, + 0.018105633556842804, + 0.027499530464410782, + 0.022919323295354843, + 0.029382940381765366, + 0.021320374682545662, + 0.04663672298192978, + 0.049370672553777695, + 0.01767694763839245, + 0.027270635589957237, + 0.05026237666606903, + 0.026989910751581192, + 0.015371748246252537, + 0.0076930830255150795, + 0.010741464793682098, + 0.02196473814547062, + 0.008008384145796299, + 0.008404148742556572, + 0.016334906220436096, + 0.010148645378649235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.011074976995587349 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a regal woman with long, flowing black hair wearing a red silk robe with golden cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman standing on a mossy rock ledge in a lush canyon while holding a massive dark lacquered fan with an ink dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing the fan in an arc that magically unleashes a torrent of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water rendered in surreal, smooth, photorealistic detail with tumbling waves and spray, contrasting with the stylized world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera dramatically zoom out to wide shots revealing the scale of the flood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870547115802765 + ], + "opposite_logprob_tokens": [ + -1.137054681777954 + ], + "logprob_reward": -0.3870547115802765, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870547115802765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1220787987112999, + 0.053487032651901245, + 0.03710908815264702, + 0.02612166665494442, + 0.09069056808948517, + 0.1558263599872589, + 0.05088212713599205, + 0.1088746041059494, + 0.1534111499786377, + 0.20151863992214203 + ], + "credit_map": [ + 0.012034246698021889, + 0.009736021980643272, + 0.02566356025636196, + 0.008167257532477379, + 0.02277659997344017, + 0.02135522849857807, + 0.012351448647677898, + 0.009994431398808956, + 0.006069627124816179, + 0.004835460800677538, + 0.008179323747754097, + 0.0038175401277840137, + 0.012847591191530228, + 0.00738478871062398, + 0.006065680645406246, + 0.004287015181034803, + 0.004581372253596783, + 0.00354372919537127, + 0.009314950555562973, + 0.0028268159367144108, + 0.003352982457727194, + 0.008173488080501556, + 0.0024791378527879715, + 0.0028366127517074347, + 0.0035538210067898035, + 0.001819461234845221, + 0.0064711361192166805, + 0.0022336840629577637, + 0.0028873742558062077, + 0.0036801837850362062, + 0.002639955375343561, + 0.0028360504657030106, + 0.0132563766092062, + 0.011736751534044743, + 0.016614748165011406, + 0.0072799851186573505, + 0.01214047521352768, + 0.010997897945344448, + 0.010483582504093647, + 0.008180754259228706, + 0.04285123571753502, + 0.019231809303164482, + 0.03033456765115261, + 0.011283015832304955, + 0.01140056736767292, + 0.02118990197777748, + 0.010361104272305965, + 0.009174163453280926, + 0.009203347377479076, + 0.0035390309058129787, + 0.007974924519658089, + 0.00787363015115261, + 0.00563577376306057, + 0.005464004818350077, + 0.005799152422696352, + 0.005392263177782297, + 0.017960581928491592, + 0.0143728731200099, + 0.015273407101631165, + 0.013341784477233887, + 0.011041142046451569, + 0.011125999502837658, + 0.014786449261009693, + 0.010972362942993641, + 0.01603507809340954, + 0.016190623864531517, + 0.026942115277051926, + 0.023678507655858994, + 0.012899529188871384, + 0.018541255965828896, + 0.02427339181303978, + 0.014850654639303684, + 0.03714344650506973, + 0.014570745639503002, + 0.02665604278445244, + 0.030871059745550156, + 0.017156356945633888, + 0.01719285175204277, + 0.032371215522289276, + 0.02555692009627819 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.3870547115802765 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a regal woman with long, flowing black hair wearing a red silk robe with golden cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman standing on a mossy rock ledge in a lush canyon while holding a massive dark lacquered fan with an ink dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing the fan in an arc that magically unleashes a torrent of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water rendered in surreal, smooth, photorealistic detail with tumbling waves and spray, contrasting with the stylized world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera dramatically zoom out to wide shots revealing the scale of the flood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305015504360199 + ], + "opposite_logprob_tokens": [ + -3.1680500507354736 + ], + "logprob_reward": -0.04305015504360199, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305015504360199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1485675722360611, + 0.03599894419312477, + 0.024311425164341927, + 0.016767852008342743, + 0.0734698548913002, + 0.08040798455476761, + 0.07005956023931503, + 0.205365851521492, + 0.3040420711040497, + 0.04100881144404411 + ], + "credit_map": [ + 0.02317618392407894, + 0.006717076525092125, + 0.04798160493373871, + 0.006436589639633894, + 0.00976283848285675, + 0.03921608254313469, + 0.006939523387700319, + 0.008337678387761116, + 0.008895326405763626, + 0.0025381888262927532, + 0.006800679489970207, + 0.002996679162606597, + 0.004569806158542633, + 0.004025784321129322, + 0.0027887143660336733, + 0.003383767558261752, + 0.004322780296206474, + 0.0020569872576743364, + 0.004762278404086828, + 0.0020689480006694794, + 0.003696340834721923, + 0.003248572815209627, + 0.0023481803946197033, + 0.0018073359970003366, + 0.002721946220844984, + 0.0015557246515527368, + 0.0028673363849520683, + 0.0017526202136650681, + 0.0020217765122652054, + 0.0026176366955041885, + 0.001692265155725181, + 0.0015385464066639543, + 0.007981722243130207, + 0.0203346386551857, + 0.02199091576039791, + 0.0063294400461018085, + 0.004807265009731054, + 0.005907529033720493, + 0.003220946993678808, + 0.0028973903972655535, + 0.013526839204132557, + 0.011803056113421917, + 0.01730397902429104, + 0.006559515837579966, + 0.008379636332392693, + 0.00987274944782257, + 0.007165278773754835, + 0.005796932149678469, + 0.013057594187557697, + 0.0044630528427660465, + 0.012272554449737072, + 0.010310080833733082, + 0.008103646337985992, + 0.007706078700721264, + 0.0069707101210951805, + 0.007175840903073549, + 0.035190656781196594, + 0.016768310219049454, + 0.024041976779699326, + 0.02180132456123829, + 0.024975012987852097, + 0.020605698227882385, + 0.03577742725610733, + 0.026205454021692276, + 0.039908502250909805, + 0.027970219030976295, + 0.03300312161445618, + 0.039019498974084854, + 0.027400897815823555, + 0.04535888880491257, + 0.05243583768606186, + 0.03894511237740517, + 0.007317546755075455, + 0.003907863982021809, + 0.004755882080644369, + 0.006929805502295494, + 0.004040551371872425, + 0.003092131344601512, + 0.007146692834794521, + 0.0038183375727385283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.04305015504360199 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_14/index_tensor([5028, 5028])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_14/index_tensor([5028, 5028])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0969f33caf74e36a18557548bbce85e4ba170d63 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_14/index_tensor([5028, 5028])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic tracking shot, filmed from a low angle, follows a majestic, massive Chinese-style house ship as it plows through the surging waves of a vermilion-colored lake with unstoppable force. Constructed of dark lacquered wood, its surface reflects an eerie red light. Its multi-layered structure, with its soaring eaves and angled corners, is adorned with rows of crimson silk lanterns, swaying wildly under the violent voyage. A massive, intricately carved dragon head, its eyes inlaid with jade, fiercely cleaves the surging crimson waves, stirring up a sea of blood-red spray. The entire world is enveloped by the distant, crimson skyline where the water meets the sky, casting long, twisting shadows and creating an atmosphere of both grandeur and unease. Suddenly, the frame flashes violently with a blinding white light, the camera's focus momentarily losing focus from the clear, intricate carvings of the ship's prow and rapidly switching to the blazing, blood-red dividing line of the distant skyline, reducing the massive house ship to a majestic, blurred silhouette in the foreground.", + "seed": 21591, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_14/index_5028_seed_21591_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08285220265388489, + "vlm_logprob": -0.08285220265388489, + "vlm_yes_no_margin": 2.525, + "vlm_token_credit": -0.08285220265388489, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Chinese-style house ship constructed of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through a vermilion-colored lake with surging waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship feature a massive, intricately carved dragon head with jade-inlaid eyes cleaving the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle with a dynamic tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden flash of blinding white light causing the camera to lose focus and blur the ship into a silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10279823839664459 + ], + "opposite_logprob_tokens": [ + -2.4777982234954834 + ], + "logprob_reward": -0.10279823839664459, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.10279823839664459, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1119486391544342, + 0.19426798820495605, + 0.07606836408376694, + 0.0635286495089531, + 0.08337676525115967, + 0.07740828394889832, + 0.12810787558555603, + 0.1266680657863617, + 0.06603141129016876, + 0.07259397953748703 + ], + "credit_map": [ + 0.012160014361143112, + 0.011757479049265385, + 0.010614794678986073, + 0.014536863192915916, + 0.013284415006637573, + 0.016186216846108437, + 0.017535846680402756, + 0.01587299257516861, + 0.018215792253613472, + 0.03413905203342438, + 0.014102119021117687, + 0.0397818461060524, + 0.025478526949882507, + 0.033190011978149414, + 0.014355994760990143, + 0.015004639513790607, + 0.008347304537892342, + 0.011881452053785324, + 0.008197186514735222, + 0.012394496239721775, + 0.008548379875719547, + 0.012446954846382141, + 0.007295213174074888, + 0.0069573684595525265, + 0.006286303512752056, + 0.011493610218167305, + 0.005395472981035709, + 0.0074766590259969234, + 0.010479452088475227, + 0.010298311710357666, + 0.006280859932303429, + 0.005817976780235767, + 0.006503724027425051, + 0.013631150126457214, + 0.00661842804402113, + 0.013310245238244534, + 0.017101505771279335, + 0.010219050571322441, + 0.006435029208660126, + 0.009557632729411125, + 0.006641884800046682, + 0.01486766804009676, + 0.005875063594430685, + 0.007785845082253218, + 0.01859404891729355, + 0.010225250385701656, + 0.00701122498139739, + 0.006407290697097778, + 0.017130577936768532, + 0.014666346833109856, + 0.013684880919754505, + 0.013502374291419983, + 0.015317384153604507, + 0.028254400938749313, + 0.013924802653491497, + 0.01162710040807724, + 0.010752101428806782, + 0.021207008510828018, + 0.013798902742564678, + 0.011178296990692616, + 0.02021603100001812, + 0.022907748818397522, + 0.010875106789171696, + 0.015732865780591965, + 0.004266242031008005, + 0.0050885602831840515, + 0.009093848057091236, + 0.00787496566772461, + 0.012632995843887329, + 0.01078485045582056, + 0.008634318597614765, + 0.007655624765902758, + 0.006287247873842716, + 0.008453539572656155, + 0.013394584879279137, + 0.011874777264893055, + 0.00686058821156621, + 0.0053445715457201, + 0.009537129662930965, + 0.010841540060937405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 7 + ], + "selected_reward": -0.10279823839664459 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Chinese-style house ship constructed of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through a vermilion-colored lake with surging waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship feature a massive, intricately carved dragon head with jade-inlaid eyes cleaving the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle with a dynamic tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden flash of blinding white light causing the camera to lose focus and blur the ship into a silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07901059091091156 + ], + "opposite_logprob_tokens": [ + -2.579010486602783 + ], + "logprob_reward": -0.07901059091091156, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07901059091091156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.043502599000930786, + 0.30014291405677795, + 0.0472433902323246, + 0.07623589783906937, + 0.09976430982351303, + 0.18638168275356293, + 0.09583526104688644, + 0.03856514394283295, + 0.040502727031707764, + 0.07182600349187851 + ], + "credit_map": [ + 0.006530719343572855, + 0.006614092271775007, + 0.004567904397845268, + 0.004249335266649723, + 0.006020704749971628, + 0.006095969118177891, + 0.004441258497536182, + 0.0049826158210635185, + 0.04655544087290764, + 0.04074208438396454, + 0.020871903747320175, + 0.03622961416840553, + 0.03886702284216881, + 0.038085710257291794, + 0.05080002173781395, + 0.02799113839864731, + 0.005604398902505636, + 0.008366902358829975, + 0.004320792853832245, + 0.0046905698254704475, + 0.00634919386357069, + 0.008351561613380909, + 0.004402168095111847, + 0.005157808773219585, + 0.009425719268620014, + 0.007271160371601582, + 0.00597094651311636, + 0.0057355971075594425, + 0.0066900611855089664, + 0.01769637130200863, + 0.00551786320284009, + 0.017928190529346466, + 0.013443433679640293, + 0.011609095148742199, + 0.010568049736320972, + 0.010764245875179768, + 0.01651356741786003, + 0.014343378134071827, + 0.011072440072894096, + 0.011450104415416718, + 0.021156519651412964, + 0.01361621730029583, + 0.02631165273487568, + 0.01264219731092453, + 0.020720267668366432, + 0.026714274659752846, + 0.02431664615869522, + 0.04090392217040062, + 0.015787286683917046, + 0.01061019767075777, + 0.00734001025557518, + 0.005581159144639969, + 0.010902955196797848, + 0.025965528562664986, + 0.00957128033041954, + 0.010076852515339851, + 0.0036115595139563084, + 0.006339048966765404, + 0.004342548083513975, + 0.003255000337958336, + 0.007619327399879694, + 0.005450019147247076, + 0.0037546460516750813, + 0.004192999564111233, + 0.0027736590709537268, + 0.003946415148675442, + 0.006022915244102478, + 0.003909236751496792, + 0.00789104774594307, + 0.006977649871259928, + 0.004284137859940529, + 0.0046976651065051556, + 0.006477531511336565, + 0.007534883916378021, + 0.011999531649053097, + 0.010596135631203651, + 0.0076284403912723064, + 0.00823354721069336, + 0.0077869039960205555, + 0.011569038033485413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 4 + ], + "selected_reward": -0.07901059091091156 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Chinese-style house ship constructed of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through a vermilion-colored lake with surging waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship feature a massive, intricately carved dragon head with jade-inlaid eyes cleaving the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle with a dynamic tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden flash of blinding white light causing the camera to lose focus and blur the ship into a silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308920353651047 + ], + "opposite_logprob_tokens": [ + -3.1680891513824463 + ], + "logprob_reward": -0.04308920353651047, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308920353651047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08489318192005157, + 0.40387704968452454, + 0.06950397789478302, + 0.05294369161128998, + 0.07241269201040268, + 0.08211372047662735, + 0.04078055918216705, + 0.04871676117181778, + 0.06165260449051857, + 0.08310579508543015 + ], + "credit_map": [ + 0.010491978377103806, + 0.015914443880319595, + 0.007462795358151197, + 0.008566576987504959, + 0.01565217599272728, + 0.010661734268069267, + 0.008447859436273575, + 0.00769561342895031, + 0.057293299585580826, + 0.05004784092307091, + 0.029327360913157463, + 0.05731218308210373, + 0.043335847556591034, + 0.06654149293899536, + 0.06173056364059448, + 0.03828846663236618, + 0.007900387980043888, + 0.011747965589165688, + 0.006952004972845316, + 0.009417690336704254, + 0.013438469730317593, + 0.008150451816618443, + 0.006643760483711958, + 0.005253247916698456, + 0.006860595196485519, + 0.009482219815254211, + 0.004242297727614641, + 0.005419299006462097, + 0.006660458631813526, + 0.005584500730037689, + 0.003932470455765724, + 0.010761848650872707, + 0.009286808781325817, + 0.009144304320216179, + 0.005420332308858633, + 0.012208822183310986, + 0.013404859229922295, + 0.01079605333507061, + 0.004705511499196291, + 0.00744600361213088, + 0.007737132720649242, + 0.013318302109837532, + 0.008702665567398071, + 0.00611530477181077, + 0.017981482669711113, + 0.009544097818434238, + 0.007430174853652716, + 0.01128456275910139, + 0.005798468831926584, + 0.006392303854227066, + 0.0034790770150721073, + 0.0030700580682605505, + 0.007284300867468119, + 0.005482184235006571, + 0.004158069379627705, + 0.0051161013543605804, + 0.004993063863366842, + 0.006204779259860516, + 0.005892112851142883, + 0.004502537194639444, + 0.009096384048461914, + 0.007041680160909891, + 0.004524271469563246, + 0.006461933255195618, + 0.004470419604331255, + 0.005598294548690319, + 0.00811114814132452, + 0.00584777956828475, + 0.01390781532973051, + 0.010311945341527462, + 0.0058838739059865475, + 0.007521326187998056, + 0.006524343974888325, + 0.005442611873149872, + 0.025626245886087418, + 0.01123293861746788, + 0.008010431192815304, + 0.004286985844373703, + 0.009312525391578674, + 0.01266971230506897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.04308920353651047 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Chinese-style house ship constructed of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through a vermilion-colored lake with surging waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship feature a massive, intricately carved dragon head with jade-inlaid eyes cleaving the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle with a dynamic tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden flash of blinding white light causing the camera to lose focus and blur the ship into a silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032973438501358 + ], + "opposite_logprob_tokens": [ + -2.350329637527466 + ], + "logprob_reward": -0.10032973438501358, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032973438501358, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052115295082330704, + 0.37207087874412537, + 0.07393700629472733, + 0.08407313376665115, + 0.10653501749038696, + 0.051145341247320175, + 0.05103101581335068, + 0.07999502867460251, + 0.0473455935716629, + 0.08175166696310043 + ], + "credit_map": [ + 0.004731701221317053, + 0.01052102167159319, + 0.005154521204531193, + 0.004704582504928112, + 0.008431212045252323, + 0.006172036286443472, + 0.0061462754383683205, + 0.006253943778574467, + 0.058506328612565994, + 0.04470372945070267, + 0.02725602686405182, + 0.0423143208026886, + 0.04402878135442734, + 0.051972489804029465, + 0.06523516774177551, + 0.03805401176214218, + 0.006386118475347757, + 0.016308840364217758, + 0.006632370874285698, + 0.007105924654752016, + 0.011210008524358273, + 0.010211859829723835, + 0.007247572299093008, + 0.008834308944642544, + 0.009671944193542004, + 0.00752503564581275, + 0.005911587737500668, + 0.0058605424128472805, + 0.007199784740805626, + 0.01871274970471859, + 0.006156169343739748, + 0.023035326972603798, + 0.020082933828234673, + 0.01328170858323574, + 0.011897677555680275, + 0.010786854662001133, + 0.013537922874093056, + 0.011503809131681919, + 0.008095097728073597, + 0.017349014058709145, + 0.004348481073975563, + 0.008035263977944851, + 0.0043354332447052, + 0.004805110860615969, + 0.012245770543813705, + 0.006568800192326307, + 0.004636094905436039, + 0.006170385051518679, + 0.008070371113717556, + 0.007206639740616083, + 0.0071305762976408005, + 0.005480517633259296, + 0.006964190863072872, + 0.006028761155903339, + 0.0047158789820969105, + 0.005434080492705107, + 0.008232435211539268, + 0.016190016642212868, + 0.007765854708850384, + 0.005569735541939735, + 0.014847049489617348, + 0.012570812366902828, + 0.0072302622720599174, + 0.007588865235447884, + 0.003129862016066909, + 0.003586879000067711, + 0.007210833951830864, + 0.006900772452354431, + 0.008456934243440628, + 0.005863179452717304, + 0.006329694762825966, + 0.005867436062544584, + 0.006334966979920864, + 0.006482037715613842, + 0.019780145958065987, + 0.015338337980210781, + 0.007934859953820705, + 0.006268236320465803, + 0.012045284733176231, + 0.007567797787487507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 3 + ], + "selected_reward": -0.10032973438501358 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Chinese-style house ship constructed of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through a vermilion-colored lake with surging waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship feature a massive, intricately carved dragon head with jade-inlaid eyes cleaving the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle with a dynamic tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden flash of blinding white light causing the camera to lose focus and blur the ship into a silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903324604034424 + ], + "opposite_logprob_tokens": [ + -2.4640331268310547 + ], + "logprob_reward": -0.08903324604034424, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903324604034424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07419493049383163, + 0.16173851490020752, + 0.04552942141890526, + 0.050487007945775986, + 0.04944440349936485, + 0.08471609652042389, + 0.14830631017684937, + 0.11084344238042831, + 0.0652630403637886, + 0.20947684347629547 + ], + "credit_map": [ + 0.008081914857029915, + 0.013204301707446575, + 0.007923553697764874, + 0.00872780755162239, + 0.011078990995883942, + 0.009432720951735973, + 0.008265957236289978, + 0.007479681633412838, + 0.024833228439092636, + 0.019979221746325493, + 0.012510756030678749, + 0.01809476688504219, + 0.018831340596079826, + 0.0227859765291214, + 0.027585091069340706, + 0.017118128016591072, + 0.005048674531280994, + 0.007152525708079338, + 0.004217361565679312, + 0.005959739442914724, + 0.005136718507856131, + 0.007990432903170586, + 0.0035922457464039326, + 0.006431726273149252, + 0.0055186026729643345, + 0.004479122813791037, + 0.0035487846471369267, + 0.0028605591505765915, + 0.00499896751716733, + 0.010479490272700787, + 0.0037797889672219753, + 0.014821693301200867, + 0.005285640247166157, + 0.005114100873470306, + 0.0042122104205191135, + 0.005111538805067539, + 0.00845724530518055, + 0.00717840064316988, + 0.006835028529167175, + 0.007250238675624132, + 0.008636672981083393, + 0.012191317975521088, + 0.008354462683200836, + 0.005026858299970627, + 0.015220475383102894, + 0.011885972693562508, + 0.007994921877980232, + 0.015405407175421715, + 0.018185779452323914, + 0.01694638654589653, + 0.02110137604176998, + 0.016195936128497124, + 0.012469198554754257, + 0.02664436772465706, + 0.019725201651453972, + 0.017038071528077126, + 0.013651711866259575, + 0.016526097431778908, + 0.014132938347756863, + 0.01188638899475336, + 0.017635619267821312, + 0.014606872573494911, + 0.009869186207652092, + 0.012534632347524166, + 0.005288675893098116, + 0.005277186166495085, + 0.009377487003803253, + 0.007750175893306732, + 0.012631107121706009, + 0.008268731646239758, + 0.008689736016094685, + 0.00797993317246437, + 0.020154422149062157, + 0.023254459723830223, + 0.026258379220962524, + 0.0302706491202116, + 0.035714343190193176, + 0.021504323929548264, + 0.027035731822252274, + 0.025284523144364357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 6 + ], + "selected_reward": -0.08903324604034424 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic tracking shot, filmed from a low angle, follows a majestic, massive Chinese-style house ship as it plows through the surging waves of a vermilion-colored lake with unstoppable force. Constructed of dark lacquered wood, its surface reflects an eerie red light. Its multi-layered structure, with its soaring eaves and angled corners, is adorned with rows of crimson silk lanterns, swaying wildly under the violent voyage. A massive, intricately carved dragon head, its eyes inlaid with jade, fiercely cleaves the surging crimson waves, stirring up a sea of blood-red spray. The entire world is enveloped by the distant, crimson skyline where the water meets the sky, casting long, twisting shadows and creating an atmosphere of both grandeur and unease. Suddenly, the frame flashes violently with a blinding white light, the camera's focus momentarily losing focus from the clear, intricate carvings of the ship's prow and rapidly switching to the blazing, blood-red dividing line of the distant skyline, reducing the massive house ship to a majestic, blurred silhouette in the foreground.", + "seed": 21592, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_14/index_5028_seed_21592_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1953825056552887, + "vlm_logprob": -0.1953825056552887, + "vlm_yes_no_margin": 1.725, + "vlm_token_credit": -0.1953825056552887, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Chinese-style house ship constructed of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through a vermilion-colored lake with surging waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship feature a massive, intricately carved dragon head with jade-inlaid eyes cleaving the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle with a dynamic tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden flash of blinding white light causing the camera to lose focus and blur the ship into a silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2663421034812927 + ], + "opposite_logprob_tokens": [ + -1.5163421630859375 + ], + "logprob_reward": -0.2663421034812927, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2663421034812927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3284193277359009, + 0.05482002720236778, + 0.18846511840820312, + 0.032887428998947144, + 0.038957953453063965, + 0.053642500191926956, + 0.06184769049286842, + 0.12864096462726593, + 0.03867816552519798, + 0.07364078611135483 + ], + "credit_map": [ + 0.10946866869926453, + 0.012724011205136776, + 0.12596145272254944, + 0.028723476454615593, + 0.014090723358094692, + 0.014571387320756912, + 0.01231409888714552, + 0.010565516538918018, + 0.009900032542645931, + 0.003541636746376753, + 0.009229850955307484, + 0.0064186761155724525, + 0.004989751148968935, + 0.008027694188058376, + 0.004374717362225056, + 0.008337666280567646, + 0.026550715789198875, + 0.019627057015895844, + 0.01945439726114273, + 0.022406674921512604, + 0.023778222501277924, + 0.02974020689725876, + 0.019017400220036507, + 0.027890438213944435, + 0.004660922102630138, + 0.003648672718554735, + 0.006392841227352619, + 0.005838979501277208, + 0.0028510852716863155, + 0.003168829483911395, + 0.0024932059459388256, + 0.003832894144579768, + 0.0041543589904904366, + 0.005194866564124823, + 0.0037043753545731306, + 0.005192060023546219, + 0.00695171020925045, + 0.005171620287001133, + 0.004347827285528183, + 0.004241136834025383, + 0.008460544049739838, + 0.006229815073311329, + 0.005681236274540424, + 0.011397138237953186, + 0.005705730058252811, + 0.006147695239633322, + 0.005485814996063709, + 0.004534527659416199, + 0.00705752382054925, + 0.007655289489775896, + 0.006521012634038925, + 0.00668860599398613, + 0.009643488563597202, + 0.009671183302998543, + 0.007924781180918217, + 0.006685808766633272, + 0.009832143783569336, + 0.019543403759598732, + 0.01231292076408863, + 0.015496725216507912, + 0.03240249305963516, + 0.015474462881684303, + 0.014362791553139687, + 0.009216013364493847, + 0.003143810434266925, + 0.006499855779111385, + 0.004956526681780815, + 0.0062968311831355095, + 0.003808489302173257, + 0.004954100586473942, + 0.0028595749754458666, + 0.006158974952995777, + 0.00771329365670681, + 0.023772194981575012, + 0.005880186799913645, + 0.009978371672332287, + 0.00883463304489851, + 0.009211311116814613, + 0.00418873131275177, + 0.004062063992023468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.2663421034812927 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Chinese-style house ship constructed of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through a vermilion-colored lake with surging waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship feature a massive, intricately carved dragon head with jade-inlaid eyes cleaving the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle with a dynamic tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden flash of blinding white light causing the camera to lose focus and blur the ship into a silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16042090952396393 + ], + "opposite_logprob_tokens": [ + -1.9104208946228027 + ], + "logprob_reward": -0.16042090952396393, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16042090952396393, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24001479148864746, + 0.030560122802853584, + 0.16738708317279816, + 0.027544531971216202, + 0.04020609334111214, + 0.10138702392578125, + 0.1483982354402542, + 0.07392693310976028, + 0.11318208277225494, + 0.05739313364028931 + ], + "credit_map": [ + 0.07263582199811935, + 0.009504170157015324, + 0.08333855122327805, + 0.024899939075112343, + 0.013373678550124168, + 0.015987170860171318, + 0.011003425344824791, + 0.009272044524550438, + 0.003196877194568515, + 0.00414287019520998, + 0.0046153892762959, + 0.0038459086790680885, + 0.00392170250415802, + 0.005133902654051781, + 0.002595200203359127, + 0.0031082716304808855, + 0.02547789365053177, + 0.014348589815199375, + 0.0233062282204628, + 0.017471933737397194, + 0.019781695678830147, + 0.02721867337822914, + 0.015974942594766617, + 0.023807132616639137, + 0.004073046613484621, + 0.0015643280930817127, + 0.007511452306061983, + 0.00552450492978096, + 0.00180678756441921, + 0.001766679692082107, + 0.002826387993991375, + 0.0024713431484997272, + 0.004285621456801891, + 0.008606305345892906, + 0.0053490810096263885, + 0.005501673091202974, + 0.007254892494529486, + 0.003316184040158987, + 0.0036594236735254526, + 0.0022329119965434074, + 0.01975206285715103, + 0.007041462231427431, + 0.009684860706329346, + 0.024866869673132896, + 0.006409996189177036, + 0.01300396304577589, + 0.011415348388254642, + 0.009212457574903965, + 0.019208138808608055, + 0.021211709827184677, + 0.015364809893071651, + 0.01499862689524889, + 0.0252070352435112, + 0.026559725403785706, + 0.012593204155564308, + 0.013254997320473194, + 0.0065306005999445915, + 0.013952999375760555, + 0.007266254164278507, + 0.011648768559098244, + 0.013173297047615051, + 0.006973646115511656, + 0.009769011288881302, + 0.004612360615283251, + 0.009797795675694942, + 0.015595844946801662, + 0.011502875946462154, + 0.01978820376098156, + 0.010927924886345863, + 0.01572633720934391, + 0.009421681985259056, + 0.020421411842107773, + 0.0062767378985881805, + 0.011027425527572632, + 0.0056082578375935555, + 0.008074482902884483, + 0.00634314538910985, + 0.008230642415583134, + 0.0041600377298891544, + 0.007672404404729605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.16042090952396393 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Chinese-style house ship constructed of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through a vermilion-colored lake with surging waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship feature a massive, intricately carved dragon head with jade-inlaid eyes cleaving the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle with a dynamic tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden flash of blinding white light causing the camera to lose focus and blur the ship into a silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2814430892467499 + ], + "opposite_logprob_tokens": [ + -1.4064431190490723 + ], + "logprob_reward": -0.2814430892467499, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2814430892467499, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0994981899857521, + 0.06622064113616943, + 0.15916277468204498, + 0.07053132355213165, + 0.09779024869203568, + 0.13589602708816528, + 0.05505106970667839, + 0.08192401379346848, + 0.08925076574087143, + 0.14467495679855347 + ], + "credit_map": [ + 0.01617683656513691, + 0.00613780040293932, + 0.023010537028312683, + 0.011792934499680996, + 0.010907992720603943, + 0.014032196253538132, + 0.008994804695248604, + 0.008445091545581818, + 0.009288634173572063, + 0.004718326497823, + 0.010132926516234875, + 0.0067408219911158085, + 0.00794055126607418, + 0.014286765828728676, + 0.004877360537648201, + 0.008235256187617779, + 0.022367307916283607, + 0.014416009187698364, + 0.04315883293747902, + 0.020516211166977882, + 0.01625453308224678, + 0.015879672020673752, + 0.014128219336271286, + 0.012441977858543396, + 0.011696187779307365, + 0.003977954853326082, + 0.018438588827848434, + 0.013355245813727379, + 0.0043042623437941074, + 0.004064065404236317, + 0.007954459637403488, + 0.006740558426827192, + 0.010115229524672031, + 0.017698703333735466, + 0.0117917126044631, + 0.012190159410238266, + 0.018059344962239265, + 0.009350867010653019, + 0.011505144648253918, + 0.007079083938151598, + 0.015447447076439857, + 0.02262619324028492, + 0.012915725819766521, + 0.026063209399580956, + 0.02162429317831993, + 0.010604795999825, + 0.015033585950732231, + 0.011580782011151314, + 0.006635300349444151, + 0.007506458554416895, + 0.005402702372521162, + 0.006168590392917395, + 0.00763142853975296, + 0.009201827459037304, + 0.006821094546467066, + 0.005683669820427895, + 0.0070928470231592655, + 0.016697581857442856, + 0.010254603810608387, + 0.01075070258229971, + 0.012914116494357586, + 0.00920139905065298, + 0.007848975248634815, + 0.007163783069700003, + 0.008110192604362965, + 0.013353375717997551, + 0.009177882224321365, + 0.010892070829868317, + 0.015907708555459976, + 0.013570215553045273, + 0.007237640209496021, + 0.011001686565577984, + 0.01622067764401436, + 0.04841720312833786, + 0.009721864946186543, + 0.017899228259921074, + 0.020508531481027603, + 0.017315171658992767, + 0.005609381012618542, + 0.008982889354228973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 5 + ], + "selected_reward": -0.2814430892467499 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Chinese-style house ship constructed of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through a vermilion-colored lake with surging waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship feature a massive, intricately carved dragon head with jade-inlaid eyes cleaving the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle with a dynamic tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden flash of blinding white light causing the camera to lose focus and blur the ship into a silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22562441229820251 + ], + "opposite_logprob_tokens": [ + -1.600624442100525 + ], + "logprob_reward": -0.22562441229820251, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22562441229820251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10476415604352951, + 0.031187841668725014, + 0.05729842558503151, + 0.031222514808177948, + 0.05163683742284775, + 0.10217045247554779, + 0.43488988280296326, + 0.08260117471218109, + 0.03542596101760864, + 0.06880278885364532 + ], + "credit_map": [ + 0.021210353821516037, + 0.006939463317394257, + 0.02641216665506363, + 0.016228141263127327, + 0.011382922530174255, + 0.00814398005604744, + 0.007511342875659466, + 0.00693578040227294, + 0.003563993377611041, + 0.004309709649533033, + 0.004738282412290573, + 0.003737136023119092, + 0.004834966734051704, + 0.004733020439743996, + 0.002769301412627101, + 0.0025014260318130255, + 0.009086426347494125, + 0.004985312931239605, + 0.013862820342183113, + 0.007011313457041979, + 0.006596579682081938, + 0.005551255773752928, + 0.004748463165014982, + 0.005456246435642242, + 0.0041299862787127495, + 0.003358701942488551, + 0.005701873451471329, + 0.004294790793210268, + 0.0037052424158900976, + 0.0028554322198033333, + 0.0031853429973125458, + 0.003991143312305212, + 0.0049920110031962395, + 0.00770120695233345, + 0.006634199526160955, + 0.00639232387766242, + 0.009204701520502567, + 0.005646673031151295, + 0.006996368523687124, + 0.004069345537573099, + 0.020414767786860466, + 0.00820357445627451, + 0.010163538157939911, + 0.022946955636143684, + 0.008192500099539757, + 0.014554472640156746, + 0.010787442326545715, + 0.006907195318490267, + 0.056732140481472015, + 0.060631994158029556, + 0.047804057598114014, + 0.04505065083503723, + 0.07261037826538086, + 0.07833454757928848, + 0.035770293325185776, + 0.03795577213168144, + 0.005844940897077322, + 0.014279881492257118, + 0.009168000891804695, + 0.011701828800141811, + 0.013694250024855137, + 0.008336610160768032, + 0.012169376946985722, + 0.007406276185065508, + 0.0028698823880404234, + 0.007164475042372942, + 0.0035707871429622173, + 0.005627893842756748, + 0.0032752405386418104, + 0.004902542568743229, + 0.002628156216815114, + 0.0053869751282036304, + 0.007213898003101349, + 0.022782716900110245, + 0.004696494434028864, + 0.009491492994129658, + 0.009163492359220982, + 0.00845292303711176, + 0.0032717559952288866, + 0.003730006283149123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.22562441229820251 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Chinese-style house ship constructed of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through a vermilion-colored lake with surging waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship feature a massive, intricately carved dragon head with jade-inlaid eyes cleaving the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle with a dynamic tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden flash of blinding white light causing the camera to lose focus and blur the ship into a silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043082013726234436 + ], + "opposite_logprob_tokens": [ + -3.1680819988250732 + ], + "logprob_reward": -0.043082013726234436, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043082013726234436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38133737444877625, + 0.018304772675037384, + 0.01613861881196499, + 0.011624189093708992, + 0.007658206392079592, + 0.010643663816154003, + 0.045337263494729996, + 0.015227564610540867, + 0.026624787598848343, + 0.4671035706996918 + ], + "credit_map": [ + 0.13087375462055206, + 0.011778977699577808, + 0.15058386325836182, + 0.035012152045965195, + 0.014217120595276356, + 0.01453353464603424, + 0.013369784690439701, + 0.010968194343149662, + 0.0029445562977343798, + 0.0012337876250967383, + 0.0032295123673975468, + 0.002380541991442442, + 0.0018356163054704666, + 0.0021185502409934998, + 0.001832950976677239, + 0.0027292564045637846, + 0.001993434736505151, + 0.0018311400199308991, + 0.0018223338993266225, + 0.001906412886455655, + 0.002443099394440651, + 0.0022843971382826567, + 0.0016891363775357604, + 0.0021686647087335587, + 0.0016248547472059727, + 0.0010501551441848278, + 0.0024324425030499697, + 0.0019513178849592805, + 0.0010809819214046001, + 0.0009188628755509853, + 0.001215464319102466, + 0.001350110280327499, + 0.0007981498492881656, + 0.0011351389111950994, + 0.0014363770605996251, + 0.0009741269168443978, + 0.0012945214984938502, + 0.0007161846151575446, + 0.0007300038123503327, + 0.00057370332069695, + 0.0013694833032786846, + 0.0008201253949664533, + 0.0018132665427401662, + 0.0015530590899288654, + 0.0012034060200676322, + 0.0009742037509568036, + 0.0010971201118081808, + 0.0018129990203306079, + 0.0054351226426661015, + 0.006146017462015152, + 0.005563597194850445, + 0.004379490856081247, + 0.007905447855591774, + 0.007338196504861116, + 0.003990248776972294, + 0.004579138942062855, + 0.0016863871132954955, + 0.002292327117174864, + 0.0021029608324170113, + 0.0018936099950224161, + 0.0026042007375508547, + 0.0017725087236613035, + 0.0011566561879590154, + 0.0017189139034599066, + 0.0024216773454099894, + 0.004175075329840183, + 0.0032760121393948793, + 0.003514478914439678, + 0.0019521467620506883, + 0.00323148793540895, + 0.0038825625088065863, + 0.004171346314251423, + 0.03722590208053589, + 0.13740627467632294, + 0.041742246598005295, + 0.07089216262102127, + 0.05706657096743584, + 0.05406704545021057, + 0.0359731987118721, + 0.03273018077015877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.043082013726234436 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic tracking shot, filmed from a low angle, follows a majestic, massive Chinese-style house ship as it plows through the surging waves of a vermilion-colored lake with unstoppable force. Constructed of dark lacquered wood, its surface reflects an eerie red light. Its multi-layered structure, with its soaring eaves and angled corners, is adorned with rows of crimson silk lanterns, swaying wildly under the violent voyage. A massive, intricately carved dragon head, its eyes inlaid with jade, fiercely cleaves the surging crimson waves, stirring up a sea of blood-red spray. The entire world is enveloped by the distant, crimson skyline where the water meets the sky, casting long, twisting shadows and creating an atmosphere of both grandeur and unease. Suddenly, the frame flashes violently with a blinding white light, the camera's focus momentarily losing focus from the clear, intricate carvings of the ship's prow and rapidly switching to the blazing, blood-red dividing line of the distant skyline, reducing the massive house ship to a majestic, blurred silhouette in the foreground.", + "seed": 21593, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_14/index_5028_seed_21593_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.12100395411252976, + "vlm_logprob": -0.12100395411252976, + "vlm_yes_no_margin": 2.2, + "vlm_token_credit": -0.12100395411252976, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Chinese-style house ship constructed of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through a vermilion-colored lake with surging waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship feature a massive, intricately carved dragon head with jade-inlaid eyes cleaving the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle with a dynamic tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden flash of blinding white light causing the camera to lose focus and blur the ship into a silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11242132633924484 + ], + "opposite_logprob_tokens": [ + -2.3624212741851807 + ], + "logprob_reward": -0.11242132633924484, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.11242132633924484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1626000702381134, + 0.10455449670553207, + 0.12148675322532654, + 0.14264953136444092, + 0.15720193088054657, + 0.06646811962127686, + 0.03830639272928238, + 0.08014674484729767, + 0.050075337290763855, + 0.07651061564683914 + ], + "credit_map": [ + 0.01952761597931385, + 0.011293523013591766, + 0.02811860293149948, + 0.020373554900288582, + 0.021809717640280724, + 0.01347669679671526, + 0.019076896831393242, + 0.028923453763127327, + 0.01654893346130848, + 0.006259465124458075, + 0.019902942702174187, + 0.01152542419731617, + 0.012480451725423336, + 0.014120254665613174, + 0.01057168934494257, + 0.013145336881279945, + 0.011446196585893631, + 0.015413003973662853, + 0.013284904882311821, + 0.014789757318794727, + 0.020713843405246735, + 0.016799896955490112, + 0.014594492502510548, + 0.01444465946406126, + 0.01392258983105421, + 0.014138754457235336, + 0.01856561377644539, + 0.01795906573534012, + 0.020689185708761215, + 0.015752850100398064, + 0.016033614054322243, + 0.02558785118162632, + 0.020124703645706177, + 0.012383690103888512, + 0.015567620284855366, + 0.01790565811097622, + 0.023084914311766624, + 0.01908884570002556, + 0.017822593450546265, + 0.03122391179203987, + 0.005622320808470249, + 0.009345995262265205, + 0.009379773400723934, + 0.007876629009842873, + 0.007925201207399368, + 0.008713333867490292, + 0.006928582675755024, + 0.01067628338932991, + 0.0038459727074950933, + 0.005148556083440781, + 0.0065639447420835495, + 0.003396151587367058, + 0.003676909487694502, + 0.00422837445512414, + 0.0047463420778512955, + 0.0067001404240727425, + 0.009991311468183994, + 0.011001039296388626, + 0.00960532110184431, + 0.008549286983907223, + 0.007293239235877991, + 0.009373495355248451, + 0.007545777130872011, + 0.016787273809313774, + 0.004618924576789141, + 0.008099588565528393, + 0.008939807303249836, + 0.007063641212880611, + 0.007961155846714973, + 0.005490032024681568, + 0.0032398952171206474, + 0.004662294406443834, + 0.006425705272704363, + 0.014866932295262814, + 0.011404873803257942, + 0.005794168915599585, + 0.009189230389893055, + 0.010931014083325863, + 0.004517055582255125, + 0.01338162925094366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.11242132633924484 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Chinese-style house ship constructed of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through a vermilion-colored lake with surging waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship feature a massive, intricately carved dragon head with jade-inlaid eyes cleaving the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle with a dynamic tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden flash of blinding white light causing the camera to lose focus and blur the ship into a silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521592676639557 + ], + "opposite_logprob_tokens": [ + -1.5021592378616333 + ], + "logprob_reward": -0.2521592676639557, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521592676639557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12939253449440002, + 0.04856845736503601, + 0.11386382579803467, + 0.08783066272735596, + 0.04196382313966751, + 0.13544732332229614, + 0.06895025819540024, + 0.20659151673316956, + 0.06629788130521774, + 0.10109373927116394 + ], + "credit_map": [ + 0.019831163808703423, + 0.009783253073692322, + 0.03375174105167389, + 0.009520133025944233, + 0.011300693266093731, + 0.010565000586211681, + 0.01237057987600565, + 0.02226995676755905, + 0.006668921560049057, + 0.003485216526314616, + 0.008501355536282063, + 0.005709535907953978, + 0.007077493704855442, + 0.004757948685437441, + 0.006049055606126785, + 0.006318921688944101, + 0.010864113457500935, + 0.015978695824742317, + 0.009004698134958744, + 0.01278229895979166, + 0.02186351828277111, + 0.0158541277050972, + 0.012594512663781643, + 0.014921844005584717, + 0.008531182073056698, + 0.009180820547044277, + 0.009472081437706947, + 0.008894994854927063, + 0.018057798966765404, + 0.010291662067174911, + 0.008432787843048573, + 0.01496933214366436, + 0.007674174848943949, + 0.003566386178135872, + 0.004156550858169794, + 0.0034533319994807243, + 0.004549421835690737, + 0.0059261629357934, + 0.005032246466726065, + 0.007605543360114098, + 0.011769075877964497, + 0.013458659872412682, + 0.023555131629109383, + 0.016771860420703888, + 0.02245393767952919, + 0.01607825979590416, + 0.012041370384395123, + 0.019319025799632072, + 0.006075333803892136, + 0.011163975112140179, + 0.012985587120056152, + 0.00664951466023922, + 0.004079914186149836, + 0.008602259680628777, + 0.007921905256807804, + 0.011471759527921677, + 0.025736365467309952, + 0.029336925595998764, + 0.03201000764966011, + 0.022784868255257607, + 0.019944828003644943, + 0.02959204651415348, + 0.022257434204220772, + 0.024929016828536987, + 0.00846187025308609, + 0.00995075423270464, + 0.010132122784852982, + 0.007915487512946129, + 0.0107253547757864, + 0.008801155723631382, + 0.003940141759812832, + 0.006370985880494118, + 0.00888965092599392, + 0.013568109832704067, + 0.014070534147322178, + 0.005811300594359636, + 0.01120090577751398, + 0.02020382322371006, + 0.0059980652295053005, + 0.021351339295506477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.2521592676639557 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Chinese-style house ship constructed of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through a vermilion-colored lake with surging waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship feature a massive, intricately carved dragon head with jade-inlaid eyes cleaving the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle with a dynamic tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden flash of blinding white light causing the camera to lose focus and blur the ship into a silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07007917016744614 + ], + "opposite_logprob_tokens": [ + -2.6950790882110596 + ], + "logprob_reward": -0.07007917016744614, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07007917016744614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.163918137550354, + 0.07026037573814392, + 0.1330566108226776, + 0.15946492552757263, + 0.11176923662424088, + 0.17821258306503296, + 0.021405629813671112, + 0.06513184309005737, + 0.04993397742509842, + 0.04684668034315109 + ], + "credit_map": [ + 0.02332528866827488, + 0.012154839001595974, + 0.033630792051553726, + 0.01790565811097622, + 0.017807964235544205, + 0.01184720080345869, + 0.016303401440382004, + 0.030942996963858604, + 0.011195837520062923, + 0.005181380081921816, + 0.011642714031040668, + 0.0074042570777237415, + 0.011277023702859879, + 0.006871698424220085, + 0.007260971702635288, + 0.009426498785614967, + 0.012178171426057816, + 0.017412154003977776, + 0.013029956258833408, + 0.015211196616292, + 0.023903558030724525, + 0.021591978147625923, + 0.014997410587966442, + 0.01473218947649002, + 0.014674659818410873, + 0.01579304225742817, + 0.019091255962848663, + 0.018943892791867256, + 0.029426859691739082, + 0.017503608018159866, + 0.016950072720646858, + 0.027081534266471863, + 0.013604626059532166, + 0.009797115810215473, + 0.011293954215943813, + 0.011866559274494648, + 0.018228936940431595, + 0.013371823355555534, + 0.012663491070270538, + 0.020942723378539085, + 0.01690586656332016, + 0.017376748844981194, + 0.02738511562347412, + 0.02189202606678009, + 0.024019263684749603, + 0.02112666890025139, + 0.018756819888949394, + 0.030750080943107605, + 0.0017516250954940915, + 0.0040352423675358295, + 0.0029665473848581314, + 0.0018715764163061976, + 0.0023762022610753775, + 0.002376927062869072, + 0.0022309471387416124, + 0.0037965618539601564, + 0.0077851139940321445, + 0.011205279268324375, + 0.00816960446536541, + 0.0048181903548538685, + 0.007906155660748482, + 0.008968104608356953, + 0.006144424434751272, + 0.010134970769286156, + 0.0050664134323596954, + 0.006132157519459724, + 0.009929818101227283, + 0.007691926322877407, + 0.008440227247774601, + 0.006334593519568443, + 0.003248770022764802, + 0.0030900691635906696, + 0.004381669219583273, + 0.009843560867011547, + 0.00579475611448288, + 0.003642096184194088, + 0.00628126785159111, + 0.006341999862343073, + 0.00268738460727036, + 0.00787394493818283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.07007917016744614 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Chinese-style house ship constructed of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through a vermilion-colored lake with surging waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship feature a massive, intricately carved dragon head with jade-inlaid eyes cleaving the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle with a dynamic tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden flash of blinding white light causing the camera to lose focus and blur the ship into a silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033285617828369 + ], + "opposite_logprob_tokens": [ + -2.350332736968994 + ], + "logprob_reward": -0.10033285617828369, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033285617828369, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060114528983831406, + 0.042511504143476486, + 0.09085879474878311, + 0.045127496123313904, + 0.061024103313684464, + 0.12056002020835876, + 0.07492711395025253, + 0.19849532842636108, + 0.261487752199173, + 0.04489323869347572 + ], + "credit_map": [ + 0.009461215697228909, + 0.005191188771277666, + 0.012473825365304947, + 0.005435499362647533, + 0.007086440455168486, + 0.0046362485736608505, + 0.0057510267943143845, + 0.010079089552164078, + 0.005220786202698946, + 0.004621576052159071, + 0.004156775772571564, + 0.0058338213711977005, + 0.004775532986968756, + 0.00641682930290699, + 0.004726924933493137, + 0.006759262643754482, + 0.010105361230671406, + 0.010668368078768253, + 0.012618317268788815, + 0.008523860946297646, + 0.013424252159893513, + 0.015120481140911579, + 0.00933042075484991, + 0.011067740619182587, + 0.005156985484063625, + 0.005397253669798374, + 0.005582297686487436, + 0.0046102688647806644, + 0.006531445309519768, + 0.00570418406277895, + 0.005236027296632528, + 0.006909036543220282, + 0.010158536955714226, + 0.004904328845441341, + 0.005673946347087622, + 0.0053280009888112545, + 0.008313552476465702, + 0.007591671776026487, + 0.006744294427335262, + 0.01230977289378643, + 0.013954132795333862, + 0.017404699698090553, + 0.014238948002457619, + 0.012250307947397232, + 0.016466911882162094, + 0.01576356217265129, + 0.01575286127626896, + 0.014728608541190624, + 0.006219587288796902, + 0.014132480137050152, + 0.011751274578273296, + 0.006615958176553249, + 0.011200004257261753, + 0.00872455071657896, + 0.007522342726588249, + 0.008760929107666016, + 0.026775049045681953, + 0.027815107256174088, + 0.02987215295433998, + 0.0197991281747818, + 0.02469247207045555, + 0.024415213614702225, + 0.02193828485906124, + 0.02318793535232544, + 0.03395616263151169, + 0.04524312540888786, + 0.040783096104860306, + 0.03493410721421242, + 0.03775009140372276, + 0.027798227965831757, + 0.016944682225584984, + 0.024078289046883583, + 0.003446992952376604, + 0.007010978646576405, + 0.009559569880366325, + 0.004057671874761581, + 0.003962726332247257, + 0.008396309800446033, + 0.002338964259251952, + 0.006120030302554369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.10033285617828369 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Chinese-style house ship constructed of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through a vermilion-colored lake with surging waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship feature a massive, intricately carved dragon head with jade-inlaid eyes cleaving the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle with a dynamic tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden flash of blinding white light causing the camera to lose focus and blur the ship into a silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002715021371841 + ], + "opposite_logprob_tokens": [ + -2.6950271129608154 + ], + "logprob_reward": -0.07002715021371841, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002715021371841, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057920679450035095, + 0.05087234079837799, + 0.058701884001493454, + 0.06125576049089432, + 0.03369646891951561, + 0.05231177806854248, + 0.037060029804706573, + 0.12429416179656982, + 0.19691970944404602, + 0.3269672393798828 + ], + "credit_map": [ + 0.008286046795547009, + 0.004764414858072996, + 0.010288383811712265, + 0.007782417815178633, + 0.009056751616299152, + 0.004663768690079451, + 0.006156235933303833, + 0.006922653876245022, + 0.008444329723715782, + 0.004153296817094088, + 0.007702638395130634, + 0.00390789145603776, + 0.004140326753258705, + 0.006505823694169521, + 0.0043680183589458466, + 0.011650013737380505, + 0.004971827380359173, + 0.0062362742610275745, + 0.007139736786484718, + 0.006957986392080784, + 0.00902333203703165, + 0.008770240470767021, + 0.007366885431110859, + 0.008235595189034939, + 0.00526631623506546, + 0.006242543924599886, + 0.007089356891810894, + 0.0074673001654446125, + 0.011782430112361908, + 0.007103694602847099, + 0.00643426738679409, + 0.009869848378002644, + 0.004189653322100639, + 0.003871629247441888, + 0.003424032824113965, + 0.0037084012292325497, + 0.004670259542763233, + 0.004670681431889534, + 0.004828699864447117, + 0.004333110060542822, + 0.005796125158667564, + 0.005823350511491299, + 0.007140118628740311, + 0.006159527227282524, + 0.005424829199910164, + 0.006351830437779427, + 0.007150919642299414, + 0.008465074934065342, + 0.00308905984275043, + 0.007040620781481266, + 0.00630147336050868, + 0.003597342874854803, + 0.003926658537238836, + 0.004423626232892275, + 0.00391772435978055, + 0.004763522185385227, + 0.01573374681174755, + 0.01853152923285961, + 0.020822674036026, + 0.013109700754284859, + 0.011451367288827896, + 0.01860497146844864, + 0.012835506349802017, + 0.013204656541347504, + 0.0286153145134449, + 0.02970224618911743, + 0.018550971522927284, + 0.024015549570322037, + 0.023746710270643234, + 0.01975180208683014, + 0.020413558930158615, + 0.032123543322086334, + 0.02838377095758915, + 0.04786938801407814, + 0.03613948076963425, + 0.029736598953604698, + 0.03910352289676666, + 0.05272742360830307, + 0.02094360999763012, + 0.07206343114376068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.07002715021371841 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic tracking shot, filmed from a low angle, follows a majestic, massive Chinese-style house ship as it plows through the surging waves of a vermilion-colored lake with unstoppable force. Constructed of dark lacquered wood, its surface reflects an eerie red light. Its multi-layered structure, with its soaring eaves and angled corners, is adorned with rows of crimson silk lanterns, swaying wildly under the violent voyage. A massive, intricately carved dragon head, its eyes inlaid with jade, fiercely cleaves the surging crimson waves, stirring up a sea of blood-red spray. The entire world is enveloped by the distant, crimson skyline where the water meets the sky, casting long, twisting shadows and creating an atmosphere of both grandeur and unease. Suddenly, the frame flashes violently with a blinding white light, the camera's focus momentarily losing focus from the clear, intricate carvings of the ship's prow and rapidly switching to the blazing, blood-red dividing line of the distant skyline, reducing the massive house ship to a majestic, blurred silhouette in the foreground.", + "seed": 21594, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_14/index_5028_seed_21594_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.38489196095615624, + "vlm_logprob": -0.38489196095615624, + "vlm_yes_no_margin": 1.475000011920929, + "vlm_token_credit": -0.38489196095615624, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Chinese-style house ship constructed of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through a vermilion-colored lake with surging waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship feature a massive, intricately carved dragon head with jade-inlaid eyes cleaving the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle with a dynamic tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden flash of blinding white light causing the camera to lose focus and blur the ship into a silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3535913825035095 + ], + "opposite_logprob_tokens": [ + -1.2285914421081543 + ], + "logprob_reward": -0.3535913825035095, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.3535913825035095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06254608184099197, + 0.02683127671480179, + 0.008542574942111969, + 0.5074806213378906, + 0.15991023182868958, + 0.024028705433011055, + 0.18204858899116516, + 0.016977248713374138, + 0.007613127585500479, + 0.004021576140075922 + ], + "credit_map": [ + 0.0062611899338662624, + 0.0068487259559333324, + 0.010275526903569698, + 0.0066609312780201435, + 0.007978624664247036, + 0.007969439961016178, + 0.010685707442462444, + 0.005865932907909155, + 0.0025130959693342447, + 0.0050922567024827, + 0.003781905397772789, + 0.0038211399223655462, + 0.0030578337609767914, + 0.0033524848986417055, + 0.0022411812096834183, + 0.0029713779222220182, + 0.0010842111660167575, + 0.0006984129431657493, + 0.001089140772819519, + 0.0014964353758841753, + 0.001142684486694634, + 0.0010393806733191013, + 0.0011788582196459174, + 0.0008134516538120806, + 0.07001108676195145, + 0.10171528905630112, + 0.041767194867134094, + 0.03157495707273483, + 0.04301566258072853, + 0.13445641100406647, + 0.03339028358459473, + 0.05154970660805702, + 0.00828262884169817, + 0.011347897350788116, + 0.0472756028175354, + 0.007105821277946234, + 0.0606951080262661, + 0.012124521657824516, + 0.0064334385097026825, + 0.006645224988460541, + 0.002315630903467536, + 0.0018697091145440936, + 0.003089646343141794, + 0.004546244628727436, + 0.002072023693472147, + 0.002922376850619912, + 0.0029265820048749447, + 0.004286493174731731, + 0.018846334889531136, + 0.08081866055727005, + 0.0226298701018095, + 0.012451838701963425, + 0.020378611981868744, + 0.016076145693659782, + 0.006144059356302023, + 0.004703063517808914, + 0.0011271609691902995, + 0.00194451748393476, + 0.003506060689687729, + 0.0016208316665142775, + 0.003815627424046397, + 0.001752153504639864, + 0.0009573335410095751, + 0.0022535633761435747, + 0.0005095619708299637, + 0.0004608658782672137, + 0.0023911092430353165, + 0.0008642952307127416, + 0.000348961737472564, + 0.0004971255548298359, + 0.00041484140092507005, + 0.0021263661328703165, + 0.00044894457096233964, + 0.0004668945912271738, + 0.0004214642394799739, + 0.0006485021440312266, + 0.0005955232190899551, + 0.0003741213877219707, + 0.0005510274786502123, + 0.000515098450705409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 4 + ], + "selected_reward": -0.3535913825035095 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Chinese-style house ship constructed of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through a vermilion-colored lake with surging waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship feature a massive, intricately carved dragon head with jade-inlaid eyes cleaving the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle with a dynamic tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden flash of blinding white light causing the camera to lose focus and blur the ship into a silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0539300441741943 + ], + "opposite_logprob_tokens": [ + -0.42893001437187195 + ], + "logprob_reward": -1.0539300441741943, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0539300441741943, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.058974649757146835, + 0.030009828507900238, + 0.016278285533189774, + 0.27486568689346313, + 0.1993982046842575, + 0.04415171593427658, + 0.31257879734039307, + 0.022176658734679222, + 0.013595779426395893, + 0.02797034941613674 + ], + "credit_map": [ + 0.006533984560519457, + 0.005890661384910345, + 0.00576190697029233, + 0.005669008009135723, + 0.007378254551440477, + 0.008166404440999031, + 0.012804674915969372, + 0.006769754458218813, + 0.004963818471878767, + 0.003002127166837454, + 0.0038122504483908415, + 0.00374065269716084, + 0.002726026577875018, + 0.0034820062573999166, + 0.005522487219423056, + 0.002760459901764989, + 0.00271057290956378, + 0.001813527662307024, + 0.0017318807076662779, + 0.0019060513004660606, + 0.0021158596500754356, + 0.0019313718657940626, + 0.0017391402507200837, + 0.0023298822343349457, + 0.037059884518384933, + 0.05372738093137741, + 0.024147313088178635, + 0.016377724707126617, + 0.023814445361495018, + 0.07227878272533417, + 0.018766000866889954, + 0.028694139793515205, + 0.009083484299480915, + 0.013707703910768032, + 0.0650019496679306, + 0.006132880225777626, + 0.08439721912145615, + 0.0121071245521307, + 0.005387998651713133, + 0.003579851007089019, + 0.004000782500952482, + 0.008181517012417316, + 0.0059251971542835236, + 0.003439757041633129, + 0.012519867159426212, + 0.003477306105196476, + 0.0031324047595262527, + 0.003474884433671832, + 0.03217461332678795, + 0.1368241012096405, + 0.0396510474383831, + 0.022062577307224274, + 0.0338456891477108, + 0.02790898270905018, + 0.011215467005968094, + 0.0088962996378541, + 0.001743054948747158, + 0.0013273406075313687, + 0.007786729373037815, + 0.0028368334751576185, + 0.0013093650341033936, + 0.0018900433788076043, + 0.002011524513363838, + 0.003271768568083644, + 0.002197670517489314, + 0.0015300051309168339, + 0.0012746629072353244, + 0.0014902242692187428, + 0.0008179136202670634, + 0.0012020316207781434, + 0.002547510666772723, + 0.0025357608683407307, + 0.0015650358982384205, + 0.005971524398773909, + 0.001854308182373643, + 0.001120837521739304, + 0.010203695856034756, + 0.0026433668099343777, + 0.0026940081734210253, + 0.0019175736233592033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 4 + ], + "selected_reward": -1.0539300441741943 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Chinese-style house ship constructed of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through a vermilion-colored lake with surging waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship feature a massive, intricately carved dragon head with jade-inlaid eyes cleaving the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle with a dynamic tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden flash of blinding white light causing the camera to lose focus and blur the ship into a silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2814776301383972 + ], + "opposite_logprob_tokens": [ + -1.406477689743042 + ], + "logprob_reward": -0.2814776301383972, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2814776301383972, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02076003886759281, + 0.011665540747344494, + 0.005783709231764078, + 0.6405201554298401, + 0.21357989311218262, + 0.008782454766333103, + 0.0614485926926136, + 0.009699593298137188, + 0.004595357459038496, + 0.023164687678217888 + ], + "credit_map": [ + 0.002279480919241905, + 0.002412131056189537, + 0.0026382473297417164, + 0.0029575861990451813, + 0.002328367903828621, + 0.002453617751598358, + 0.0029612628277391195, + 0.0027293439488857985, + 0.0008928995812311769, + 0.0019835548009723425, + 0.0016120296204462647, + 0.0019526694668456912, + 0.0012614764273166656, + 0.0013612824259325862, + 0.001104707713238895, + 0.0014969210606068373, + 0.0007545749540440738, + 0.000709290208760649, + 0.0006381238345056772, + 0.0007073162705637515, + 0.0007514394819736481, + 0.0006484801997430623, + 0.0006894886610098183, + 0.0008849955047480762, + 0.08963295817375183, + 0.13012368977069855, + 0.05177540332078934, + 0.0381060428917408, + 0.05539446696639061, + 0.16777554154396057, + 0.04246010631322861, + 0.0652519091963768, + 0.008969230577349663, + 0.015158572234213352, + 0.06939873099327087, + 0.00569903152063489, + 0.09193123877048492, + 0.013805967755615711, + 0.004923922009766102, + 0.003693197388201952, + 0.0008516312227584422, + 0.0011255365097895265, + 0.0015475961845368147, + 0.0009296597563661635, + 0.0016802852042019367, + 0.0007431760313920677, + 0.0008695553406141698, + 0.0010350145166739821, + 0.006107133347541094, + 0.026623360812664032, + 0.008090289309620857, + 0.004582221154123545, + 0.00664796307682991, + 0.005393289029598236, + 0.002265270333737135, + 0.0017390623688697815, + 0.0008203252800740302, + 0.000822868081741035, + 0.003059219103306532, + 0.0012168125249445438, + 0.0010941354557871819, + 0.0007566712447442114, + 0.0007619742536917329, + 0.0011675863061100245, + 0.0006906030466780066, + 0.000399984885007143, + 0.0008910639444366097, + 0.0005240031168796122, + 0.0002576610422693193, + 0.0002400106895947829, + 0.0007035636808723211, + 0.0008884670096449554, + 0.0013004201464354992, + 0.00504427682608366, + 0.0017305169021710753, + 0.0006122093182057142, + 0.008222965523600578, + 0.0021041017025709152, + 0.0012993860291317105, + 0.002850812394171953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 6 + ], + "selected_reward": -0.2814776301383972 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Chinese-style house ship constructed of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through a vermilion-colored lake with surging waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship feature a massive, intricately carved dragon head with jade-inlaid eyes cleaving the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle with a dynamic tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden flash of blinding white light causing the camera to lose focus and blur the ship into a silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22566446661949158 + ], + "opposite_logprob_tokens": [ + -1.600664496421814 + ], + "logprob_reward": -0.22566446661949158, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22566446661949158, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03647920489311218, + 0.015015313401818275, + 0.014655248261988163, + 0.37003886699676514, + 0.21973657608032227, + 0.03305165097117424, + 0.18002398312091827, + 0.00961841270327568, + 0.01272604614496231, + 0.10865468531847 + ], + "credit_map": [ + 0.004026688169687986, + 0.0036003771238029003, + 0.006429580505937338, + 0.004832430277019739, + 0.0033895515371114016, + 0.004093058407306671, + 0.006668386980891228, + 0.0034391325898468494, + 0.0017650470836088061, + 0.002443930832669139, + 0.0020507886074483395, + 0.0023523650597780943, + 0.0016557701164856553, + 0.0019235223298892379, + 0.0014346534153446555, + 0.001389236655086279, + 0.0020571935456246138, + 0.0013545193942263722, + 0.0020386595278978348, + 0.0019360908772796392, + 0.0016723677981644869, + 0.0017349886475130916, + 0.002030400326475501, + 0.00183102919254452, + 0.05292166769504547, + 0.0760732963681221, + 0.029148893430829048, + 0.022272147238254547, + 0.03120470605790615, + 0.097364142537117, + 0.023768361657857895, + 0.037285659462213516, + 0.008516197092831135, + 0.016059868037700653, + 0.07096561789512634, + 0.006279875990003347, + 0.09416249394416809, + 0.013340141624212265, + 0.0055799782276153564, + 0.004832425620406866, + 0.0028739720582962036, + 0.004224499687552452, + 0.00486421026289463, + 0.004123666323721409, + 0.006638555787503719, + 0.002831987803801894, + 0.0026579515542834997, + 0.0048368084244430065, + 0.01877063699066639, + 0.0795016959309578, + 0.022933393716812134, + 0.012369439005851746, + 0.0200528372079134, + 0.01574460230767727, + 0.005745632573962212, + 0.00490575609728694, + 0.001041362527757883, + 0.0009524262859486043, + 0.0015731572639197111, + 0.0016662649577483535, + 0.001265913131646812, + 0.00115023588296026, + 0.0009481701999902725, + 0.0010208835592493415, + 0.0006270533194765449, + 0.001252459711395204, + 0.002615654841065407, + 0.002235118532553315, + 0.0006328758900053799, + 0.0005180799635127187, + 0.00223584845662117, + 0.0026089565362781286, + 0.005770683754235506, + 0.02264356054365635, + 0.009192383848130703, + 0.002968460088595748, + 0.03649727255105972, + 0.009654986672103405, + 0.005188400391489267, + 0.016738947480916977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 6 + ], + "selected_reward": -0.22566446661949158 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Chinese-style house ship constructed of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through a vermilion-colored lake with surging waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship feature a massive, intricately carved dragon head with jade-inlaid eyes cleaving the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle with a dynamic tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden flash of blinding white light causing the camera to lose focus and blur the ship into a silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009796281345188618 + ], + "opposite_logprob_tokens": [ + -4.634796142578125 + ], + "logprob_reward": -0.009796281345188618, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009796281345188618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02542753331363201, + 0.013285764493048191, + 0.010604789480566978, + 0.09414640069007874, + 0.30743539333343506, + 0.02584550529718399, + 0.043981537222862244, + 0.0896439254283905, + 0.059408482164144516, + 0.33022069931030273 + ], + "credit_map": [ + 0.0023248731158673763, + 0.002766877179965377, + 0.0033463293220847845, + 0.0030928219202905893, + 0.002998840995132923, + 0.00271951244212687, + 0.005715173669159412, + 0.0024631021078675985, + 0.0019859643653035164, + 0.0018417818937450647, + 0.0010453517315909266, + 0.001360355643555522, + 0.0017874327022582293, + 0.0017448808066546917, + 0.0022712803911417723, + 0.001248714979737997, + 0.0017885720590129495, + 0.000875593104865402, + 0.001193926320411265, + 0.0013403570046648383, + 0.001327832811512053, + 0.0016654535429552197, + 0.0011596212862059474, + 0.0012534325942397118, + 0.011991238221526146, + 0.01831902749836445, + 0.006985146086663008, + 0.00857547577470541, + 0.008651284500956535, + 0.02219879999756813, + 0.008648539893329144, + 0.00877687893807888, + 0.012343263253569603, + 0.020331617444753647, + 0.10195749998092651, + 0.010994954966008663, + 0.12951889634132385, + 0.017292335629463196, + 0.007876422256231308, + 0.007120396941900253, + 0.0035401601344347, + 0.003493261756375432, + 0.003307691775262356, + 0.0033717178739607334, + 0.0038586521986871958, + 0.002332317177206278, + 0.0027152341790497303, + 0.0032264674082398415, + 0.005103017203509808, + 0.012855150736868382, + 0.003929430153220892, + 0.004330669529736042, + 0.005537114106118679, + 0.004822662100195885, + 0.004025089554488659, + 0.003378405002877116, + 0.006875046994537115, + 0.007361059542745352, + 0.030102331191301346, + 0.0101589635014534, + 0.010602299124002457, + 0.005697192158550024, + 0.006925378926098347, + 0.011921640485525131, + 0.004595378413796425, + 0.004796095658093691, + 0.012352165766060352, + 0.00576747627928853, + 0.005378721747547388, + 0.004154991824179888, + 0.005382094997912645, + 0.01698155514895916, + 0.01617891900241375, + 0.07036411017179489, + 0.02636944316327572, + 0.009752858430147171, + 0.1144116073846817, + 0.029929086565971375, + 0.016007427126169205, + 0.04720719903707504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 3 + ], + "selected_reward": -0.009796281345188618 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_15/index_tensor([1674, 1674])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_15/index_tensor([1674, 1674])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b5fb24958e6342382bb2696d0468028893ec4c31 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_15/index_tensor([1674, 1674])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic tracking shot, shot from the side, perfectly follows an armored character at a smooth, high-frame rate as he gallops across a world composed entirely of vast pixel blocks. The character wears a full suit of plate armor composed of crisp silver pixels, a helmet with a T-shaped visor, and a cape of red pixels billowing dramatically behind him at high speed. He leans over, gripping the reins tightly, guiding a sturdy steed made of brown and tan pixels, its four blocky legs racing with a smooth, powerful looping animation. They cross a bridge woven from thick green and brown pixelated vines, spanning a bottomless digital abyss populated by slow-moving, occasionally glitchy purple and black data cubes. With each thump of the horse's hooves, the vine bridge sways slightly, its tiny pixelated leaves vibrating. In the distant background, a majestic mountain range, also made of blocks, rises beneath a cyan pixelated sky dotted with perfectly square white clouds. The entire scene is bathed in a bright, even digital light, with few shadows, highlighting the saturated and vibrant colors, all of which together create an intense, heart-pounding sense of nostalgic gaming retro charm.", + "seed": 30699, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_15/index_1674_seed_30699_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1667013904079795, + "vlm_logprob": -0.1667013904079795, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.1667013904079795, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an armored character wearing a full suit of plate armor with a T-shaped visor helmet and a red pixelated cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character galloping on a sturdy steed made of brown and tan pixels with four blocky legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are they crossing a bridge woven from thick green and brown pixelated vines spanning a bottomless digital abyss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature a blocky mountain range beneath a cyan pixelated sky with square white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side, performing a smooth dynamic tracking shot that follows the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1004558801651001 + ], + "opposite_logprob_tokens": [ + -2.3504557609558105 + ], + "logprob_reward": -0.1004558801651001, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1004558801651001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09332052618265152, + 0.14263080060482025, + 0.13725057244300842, + 0.13510023057460785, + 0.2693850100040436, + 0.04870682954788208, + 0.02909034863114357, + 0.03923729807138443, + 0.06943851709365845, + 0.03583983704447746 + ], + "credit_map": [ + 0.013762477785348892, + 0.018177609890699387, + 0.008976886048913002, + 0.011794060468673706, + 0.011830323375761509, + 0.01044705230742693, + 0.008383196778595448, + 0.00994891207665205, + 0.017799004912376404, + 0.024904556572437286, + 0.020037740468978882, + 0.017778968438506126, + 0.015118563547730446, + 0.01365039125084877, + 0.013137300498783588, + 0.020204277709126472, + 0.015546099282801151, + 0.01831679791212082, + 0.010726351290941238, + 0.01856977678835392, + 0.021364808082580566, + 0.018844900652766228, + 0.01386906486004591, + 0.02001277543604374, + 0.026909297332167625, + 0.022257845848798752, + 0.008763817138969898, + 0.013791756704449654, + 0.02312929928302765, + 0.019077684730291367, + 0.009638580493628979, + 0.011531949043273926, + 0.018419956788420677, + 0.042272187769412994, + 0.029702741652727127, + 0.04488282650709152, + 0.04063376784324646, + 0.032937441021203995, + 0.03148161992430687, + 0.029054487124085426, + 0.006552267819643021, + 0.004215194843709469, + 0.006796039175242186, + 0.007265687920153141, + 0.006577835418283939, + 0.00601607421413064, + 0.004691604524850845, + 0.006592127960175276, + 0.0032924285624176264, + 0.0037612447049468756, + 0.0025324516464024782, + 0.0036646646913141012, + 0.003621812677010894, + 0.002888323972001672, + 0.0040969992987811565, + 0.005232424475252628, + 0.006773697212338448, + 0.004352917429059744, + 0.003844448132440448, + 0.005839960183948278, + 0.0033150215167552233, + 0.00427035940811038, + 0.006506709381937981, + 0.004334188532084227, + 0.011872735805809498, + 0.011309324763715267, + 0.00588875962421298, + 0.01013348437845707, + 0.007351771928369999, + 0.008718370459973812, + 0.007931474596261978, + 0.006232596468180418, + 0.004425640683621168, + 0.004606297239661217, + 0.007477169390767813, + 0.003490116447210312, + 0.0033700321801006794, + 0.005173101555556059, + 0.002894693287089467, + 0.004402790684252977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.1004558801651001 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an armored character wearing a full suit of plate armor with a T-shaped visor helmet and a red pixelated cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character galloping on a sturdy steed made of brown and tan pixels with four blocky legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are they crossing a bridge woven from thick green and brown pixelated vines spanning a bottomless digital abyss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature a blocky mountain range beneath a cyan pixelated sky with square white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side, performing a smooth dynamic tracking shot that follows the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004645389039069414 + ], + "opposite_logprob_tokens": [ + -5.379645347595215 + ], + "logprob_reward": -0.004645389039069414, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004645389039069414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11888325214385986, + 0.13581326603889465, + 0.11191638559103012, + 0.15473392605781555, + 0.1550668329000473, + 0.05602354556322098, + 0.0902610570192337, + 0.019354622811079025, + 0.10166682302951813, + 0.05628025904297829 + ], + "credit_map": [ + 0.015311099588871002, + 0.020000295713543892, + 0.011685935780405998, + 0.022656328976154327, + 0.010960862971842289, + 0.013826800510287285, + 0.009859778918325901, + 0.014582154341042042, + 0.02057168073952198, + 0.021363338455557823, + 0.01359323225915432, + 0.02379399724304676, + 0.02005433477461338, + 0.012198149226605892, + 0.00902101956307888, + 0.015217514708638191, + 0.009694979526102543, + 0.01311546377837658, + 0.017741385847330093, + 0.015210043638944626, + 0.018433915451169014, + 0.012978535145521164, + 0.012658197432756424, + 0.012083875946700573, + 0.02603672258555889, + 0.029414953663945198, + 0.010810923762619495, + 0.01969504915177822, + 0.02368270792067051, + 0.020931335166096687, + 0.012013604864478111, + 0.012148636393249035, + 0.011753513477742672, + 0.01831027679145336, + 0.015542079694569111, + 0.018955837935209274, + 0.02157767117023468, + 0.01585245691239834, + 0.02938382513821125, + 0.023691166192293167, + 0.007216587197035551, + 0.006485436111688614, + 0.007352378685027361, + 0.007292907685041428, + 0.007526667322963476, + 0.006631627678871155, + 0.00545136583968997, + 0.008066580630838871, + 0.00830124132335186, + 0.005787769332528114, + 0.009797537699341774, + 0.009461942128837109, + 0.0066838571801781654, + 0.0069338236935436726, + 0.026368392631411552, + 0.01692650280892849, + 0.003691630670800805, + 0.0019206083379685879, + 0.0016110021388158202, + 0.00340632232837379, + 0.001421260298229754, + 0.002102409489452839, + 0.002680443925783038, + 0.002520945854485035, + 0.015655381605029106, + 0.017727579921483994, + 0.007560298778116703, + 0.014172880910336971, + 0.014597589150071144, + 0.012408536858856678, + 0.010219003073871136, + 0.009325550869107246, + 0.01256146002560854, + 0.004095953889191151, + 0.009901518933475018, + 0.003618782851845026, + 0.0035038215573877096, + 0.013532847166061401, + 0.0032000478822737932, + 0.005865829531103373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 1 + ], + "selected_reward": -0.004645389039069414 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an armored character wearing a full suit of plate armor with a T-shaped visor helmet and a red pixelated cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character galloping on a sturdy steed made of brown and tan pixels with four blocky legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are they crossing a bridge woven from thick green and brown pixelated vines spanning a bottomless digital abyss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature a blocky mountain range beneath a cyan pixelated sky with square white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side, performing a smooth dynamic tracking shot that follows the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933753490447998 + ], + "opposite_logprob_tokens": [ + -0.6933753490447998 + ], + "logprob_reward": -0.6933753490447998, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933753490447998, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08861182630062103, + 0.10046429932117462, + 0.1982308179140091, + 0.17125922441482544, + 0.09850521385669708, + 0.08218572288751602, + 0.06426578015089035, + 0.08835203945636749, + 0.04434093087911606, + 0.06378413736820221 + ], + "credit_map": [ + 0.014698617160320282, + 0.013840921223163605, + 0.006730612833052874, + 0.009733905084431171, + 0.009791393764317036, + 0.01213053334504366, + 0.006542068440467119, + 0.015143778175115585, + 0.01381729543209076, + 0.013791549019515514, + 0.010242111049592495, + 0.01749061793088913, + 0.012436378747224808, + 0.009531711228191853, + 0.011288285255432129, + 0.01186634786427021, + 0.02074611559510231, + 0.013262483291327953, + 0.054787710309028625, + 0.039691466838121414, + 0.013676444999873638, + 0.01943778246641159, + 0.023648245260119438, + 0.01298056822270155, + 0.022252008318901062, + 0.028649387881159782, + 0.011902890168130398, + 0.017765384167432785, + 0.033807408064603806, + 0.023679908365011215, + 0.016078967601060867, + 0.0171232670545578, + 0.010339541360735893, + 0.010040904395282269, + 0.02004344016313553, + 0.010973852127790451, + 0.008883786387741566, + 0.010634269565343857, + 0.009407451376318932, + 0.018181970342993736, + 0.008558117784559727, + 0.0058812545612454414, + 0.01172998733818531, + 0.010058949701488018, + 0.007258288096636534, + 0.007709349039942026, + 0.018970735371112823, + 0.012019037269055843, + 0.005709242541342974, + 0.004767994862049818, + 0.006264047231525183, + 0.007324904203414917, + 0.004703518934547901, + 0.0051843333058059216, + 0.01812107115983963, + 0.012190665118396282, + 0.013299686834216118, + 0.010556661523878574, + 0.00866653025150299, + 0.013435597531497478, + 0.010543501935899258, + 0.009068600833415985, + 0.012744542211294174, + 0.010036913678050041, + 0.007966194301843643, + 0.004927098751068115, + 0.004213846754282713, + 0.005323735065758228, + 0.0032932008616626263, + 0.008424900472164154, + 0.004126780666410923, + 0.0060651772655546665, + 0.00786532461643219, + 0.0035072702448815107, + 0.017048731446266174, + 0.004623617045581341, + 0.0034839091822504997, + 0.00939889345318079, + 0.005572441499680281, + 0.012283950112760067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.6933753490447998 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an armored character wearing a full suit of plate armor with a T-shaped visor helmet and a red pixelated cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character galloping on a sturdy steed made of brown and tan pixels with four blocky legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are they crossing a bridge woven from thick green and brown pixelated vines spanning a bottomless digital abyss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature a blocky mountain range beneath a cyan pixelated sky with square white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side, performing a smooth dynamic tracking shot that follows the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005247865337878466 + ], + "opposite_logprob_tokens": [ + -5.255248069763184 + ], + "logprob_reward": -0.005247865337878466, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005247865337878466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07857991755008698, + 0.10144596546888351, + 0.1044774278998375, + 0.1877179592847824, + 0.16353753209114075, + 0.09030918776988983, + 0.1750025749206543, + 0.030920060351490974, + 0.03952052816748619, + 0.028488896787166595 + ], + "credit_map": [ + 0.011883259750902653, + 0.009713328443467617, + 0.007182685192674398, + 0.012491444125771523, + 0.007070585153996944, + 0.014612693339586258, + 0.006983645260334015, + 0.008642277680337429, + 0.01843193918466568, + 0.013305269181728363, + 0.0069351885467767715, + 0.008081969805061817, + 0.026380157098174095, + 0.01293128076940775, + 0.008927629329264164, + 0.006452531553804874, + 0.011160342022776604, + 0.01201252918690443, + 0.01730479672551155, + 0.012665401212871075, + 0.013393182307481766, + 0.010210368782281876, + 0.013725029304623604, + 0.01400577649474144, + 0.025551598519086838, + 0.03111252188682556, + 0.011946813203394413, + 0.01913277991116047, + 0.03979933261871338, + 0.026457445695996284, + 0.01525923889130354, + 0.018458232283592224, + 0.0133228013291955, + 0.024018380790948868, + 0.013471124693751335, + 0.025190141052007675, + 0.026383988559246063, + 0.019761748611927032, + 0.021598251536488533, + 0.01979108899831772, + 0.00901673175394535, + 0.007725732866674662, + 0.014156661927700043, + 0.010174567811191082, + 0.010701818391680717, + 0.009288398548960686, + 0.016488362103700638, + 0.012756912037730217, + 0.018191270530223846, + 0.013083813712000847, + 0.01832825504243374, + 0.018368694931268692, + 0.014043048955500126, + 0.013557067140936852, + 0.04707653820514679, + 0.03235388919711113, + 0.00676612788811326, + 0.004102709237486124, + 0.002669825917109847, + 0.0036723751109093428, + 0.003709006356075406, + 0.0038170835468918085, + 0.0030624379869550467, + 0.003120492910966277, + 0.007612084504216909, + 0.003947400487959385, + 0.004093564581125975, + 0.0036294180899858475, + 0.0036541412118822336, + 0.008312687277793884, + 0.0034767298493534327, + 0.004794502630829811, + 0.004904889967292547, + 0.0017782108625397086, + 0.006038087885826826, + 0.002001637127250433, + 0.0016617003129795194, + 0.005493277683854103, + 0.002173496875911951, + 0.004437596071511507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 4 + ], + "selected_reward": -0.005247865337878466 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an armored character wearing a full suit of plate armor with a T-shaped visor helmet and a red pixelated cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character galloping on a sturdy steed made of brown and tan pixels with four blocky legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are they crossing a bridge woven from thick green and brown pixelated vines spanning a bottomless digital abyss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature a blocky mountain range beneath a cyan pixelated sky with square white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side, performing a smooth dynamic tracking shot that follows the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978246845304966 + ], + "opposite_logprob_tokens": [ + -3.52978253364563 + ], + "logprob_reward": -0.02978246845304966, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978246845304966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0954088419675827, + 0.21416734158992767, + 0.12665994465351105, + 0.09613706916570663, + 0.07433860749006271, + 0.06378486007452011, + 0.034936871379613876, + 0.036544445902109146, + 0.18374572694301605, + 0.07427632808685303 + ], + "credit_map": [ + 0.011431309394538403, + 0.014388546347618103, + 0.010584399104118347, + 0.022169629111886024, + 0.009334202855825424, + 0.009584854356944561, + 0.009947972372174263, + 0.007967927493155003, + 0.04087493568658829, + 0.02675003744661808, + 0.017389027401804924, + 0.01696365885436535, + 0.057449452579021454, + 0.023037785664200783, + 0.017680810764431953, + 0.014021613635122776, + 0.012044109404087067, + 0.011402216739952564, + 0.030437245965003967, + 0.02124948799610138, + 0.013686468824744225, + 0.011920933611690998, + 0.015269745141267776, + 0.01064973697066307, + 0.017146067693829536, + 0.014210296794772148, + 0.006344572640955448, + 0.00987200252711773, + 0.017798837274312973, + 0.013623286038637161, + 0.00802079401910305, + 0.009121212176978588, + 0.005772894248366356, + 0.009587360545992851, + 0.012397373095154762, + 0.008900845423340797, + 0.010583871975541115, + 0.007916898466646671, + 0.012534241192042828, + 0.0066451276652514935, + 0.010642731562256813, + 0.006694456096738577, + 0.009859025478363037, + 0.009686005301773548, + 0.006370898336172104, + 0.007526440545916557, + 0.006439670920372009, + 0.006565631367266178, + 0.004810829646885395, + 0.0062882183119654655, + 0.003155749524012208, + 0.004967136774212122, + 0.003544809762388468, + 0.003877605777233839, + 0.004026386421173811, + 0.004266134928911924, + 0.005521442275494337, + 0.004769410006701946, + 0.0027047947514802217, + 0.004118552431464195, + 0.004209526814520359, + 0.004213858395814896, + 0.007348514627665281, + 0.0036583475302904844, + 0.0355367474257946, + 0.021221816539764404, + 0.013670950196683407, + 0.02346600592136383, + 0.01871301420032978, + 0.03831426799297333, + 0.017531942576169968, + 0.015290986746549606, + 0.017904508858919144, + 0.009320731274783611, + 0.0053983889520168304, + 0.008845208212733269, + 0.006032819859683514, + 0.01541195996105671, + 0.006558436434715986, + 0.004804277326911688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 2 + ], + "selected_reward": -0.02978246845304966 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic tracking shot, shot from the side, perfectly follows an armored character at a smooth, high-frame rate as he gallops across a world composed entirely of vast pixel blocks. The character wears a full suit of plate armor composed of crisp silver pixels, a helmet with a T-shaped visor, and a cape of red pixels billowing dramatically behind him at high speed. He leans over, gripping the reins tightly, guiding a sturdy steed made of brown and tan pixels, its four blocky legs racing with a smooth, powerful looping animation. They cross a bridge woven from thick green and brown pixelated vines, spanning a bottomless digital abyss populated by slow-moving, occasionally glitchy purple and black data cubes. With each thump of the horse's hooves, the vine bridge sways slightly, its tiny pixelated leaves vibrating. In the distant background, a majestic mountain range, also made of blocks, rises beneath a cyan pixelated sky dotted with perfectly square white clouds. The entire scene is bathed in a bright, even digital light, with few shadows, highlighting the saturated and vibrant colors, all of which together create an intense, heart-pounding sense of nostalgic gaming retro charm.", + "seed": 30700, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_15/index_1674_seed_30700_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06823996687307954, + "vlm_logprob": -0.06823996687307954, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.06823996687307954, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an armored character wearing a full suit of plate armor with a T-shaped visor helmet and a red pixelated cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character galloping on a sturdy steed made of brown and tan pixels with four blocky legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are they crossing a bridge woven from thick green and brown pixelated vines spanning a bottomless digital abyss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature a blocky mountain range beneath a cyan pixelated sky with square white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side, performing a smooth dynamic tracking shot that follows the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18017491698265076 + ], + "opposite_logprob_tokens": [ + -1.8051749467849731 + ], + "logprob_reward": -0.18017491698265076, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18017491698265076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17932027578353882, + 0.12036813050508499, + 0.13047298789024353, + 0.10953658074140549, + 0.034244686365127563, + 0.05584323778748512, + 0.12152193486690521, + 0.16200080513954163, + 0.06127285212278366, + 0.02541852369904518 + ], + "credit_map": [ + 0.025764217600226402, + 0.02172311395406723, + 0.024720024317502975, + 0.02017994411289692, + 0.027100443840026855, + 0.020396696403622627, + 0.024920174852013588, + 0.014515654183924198, + 0.010706684552133083, + 0.013553845696151257, + 0.016237983480095863, + 0.011988461017608643, + 0.012753177434206009, + 0.01548752374947071, + 0.023972634226083755, + 0.015667812898755074, + 0.008485944010317326, + 0.022017253562808037, + 0.013016284443438053, + 0.00974899623543024, + 0.013374995440244675, + 0.024369999766349792, + 0.031100383028388023, + 0.008359125815331936, + 0.009027899242937565, + 0.009038052521646023, + 0.014079170301556587, + 0.018574465066194534, + 0.016079220920801163, + 0.01624460518360138, + 0.017774797976017, + 0.008718369528651237, + 0.0037086722441017628, + 0.003233067225664854, + 0.006443379446864128, + 0.0059891026467084885, + 0.002961327088996768, + 0.0033611529506742954, + 0.005505452863872051, + 0.00304253282956779, + 0.005565629340708256, + 0.0056403623893857, + 0.007496951147913933, + 0.005013630725443363, + 0.007748323027044535, + 0.0058739762753248215, + 0.014412108808755875, + 0.004092253744602203, + 0.011451410129666328, + 0.01105305552482605, + 0.01609872281551361, + 0.01095917634665966, + 0.01835382916033268, + 0.010695600882172585, + 0.034675247967243195, + 0.008234900422394276, + 0.020472552627325058, + 0.017149483785033226, + 0.02482621558010578, + 0.02447105571627617, + 0.02370661497116089, + 0.01708165369927883, + 0.02120232954621315, + 0.01309089083224535, + 0.00832251738756895, + 0.008996767923235893, + 0.010739590041339397, + 0.010259083472192287, + 0.0042058383114635944, + 0.004748647566884756, + 0.008355763740837574, + 0.005644645541906357, + 0.00394104840233922, + 0.0030925655737519264, + 0.0033998640719801188, + 0.003776873229071498, + 0.002553334692493081, + 0.0038656459655612707, + 0.003094068495556712, + 0.0016951223369687796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.18017491698265076 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an armored character wearing a full suit of plate armor with a T-shaped visor helmet and a red pixelated cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character galloping on a sturdy steed made of brown and tan pixels with four blocky legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are they crossing a bridge woven from thick green and brown pixelated vines spanning a bottomless digital abyss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature a blocky mountain range beneath a cyan pixelated sky with square white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side, performing a smooth dynamic tracking shot that follows the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024947968777269125 + ], + "opposite_logprob_tokens": [ + -6.002494812011719 + ], + "logprob_reward": -0.0024947968777269125, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024947968777269125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19212687015533447, + 0.15826697647571564, + 0.09480910003185272, + 0.10454873740673065, + 0.06049906089901924, + 0.13100597262382507, + 0.09389421343803406, + 0.061867885291576385, + 0.05115504190325737, + 0.05182606726884842 + ], + "credit_map": [ + 0.026096409186720848, + 0.02229478769004345, + 0.033070772886276245, + 0.02825472690165043, + 0.019767245277762413, + 0.014418654143810272, + 0.02415448985993862, + 0.024069784209132195, + 0.012945876456797123, + 0.018495159223675728, + 0.01930057629942894, + 0.023163780570030212, + 0.019145453348755836, + 0.02125612646341324, + 0.026405902579426765, + 0.01755410060286522, + 0.006861434783786535, + 0.01613009162247181, + 0.009558221325278282, + 0.008038733154535294, + 0.009010043926537037, + 0.022246239706873894, + 0.016938403248786926, + 0.006025934591889381, + 0.009186373092234135, + 0.009662335738539696, + 0.02258959785103798, + 0.011465732008218765, + 0.014689596369862556, + 0.010062281973659992, + 0.017367936670780182, + 0.00952488835901022, + 0.007043940015137196, + 0.00539223849773407, + 0.012213150970637798, + 0.009750060737133026, + 0.007257489021867514, + 0.00575582031160593, + 0.008419102057814598, + 0.0046672578901052475, + 0.010360934771597385, + 0.012993136420845985, + 0.025657258927822113, + 0.01986626721918583, + 0.016386570408940315, + 0.012631284072995186, + 0.01707010716199875, + 0.016040418297052383, + 0.013294011354446411, + 0.009883477352559566, + 0.013672892935574055, + 0.017041323706507683, + 0.008509672246873379, + 0.009107058867812157, + 0.015801824629306793, + 0.0065839518792927265, + 0.007125892676413059, + 0.006900845095515251, + 0.007714131381362677, + 0.007429804652929306, + 0.012169279158115387, + 0.006276065483689308, + 0.008539916947484016, + 0.005711948033422232, + 0.006966305430978537, + 0.007014721166342497, + 0.009883101098239422, + 0.007559793069958687, + 0.004144401755183935, + 0.004134573508054018, + 0.007648943457752466, + 0.0038032045122236013, + 0.006855309009552002, + 0.008106707595288754, + 0.006509503815323114, + 0.007031349930912256, + 0.007546719163656235, + 0.006039891391992569, + 0.006352329161018133, + 0.0033842548727989197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.0024947968777269125 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an armored character wearing a full suit of plate armor with a T-shaped visor helmet and a red pixelated cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character galloping on a sturdy steed made of brown and tan pixels with four blocky legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are they crossing a bridge woven from thick green and brown pixelated vines spanning a bottomless digital abyss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature a blocky mountain range beneath a cyan pixelated sky with square white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side, performing a smooth dynamic tracking shot that follows the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703923881053925 + ], + "opposite_logprob_tokens": [ + -2.1270391941070557 + ], + "logprob_reward": -0.12703923881053925, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12703923881053925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13012975454330444, + 0.15323422849178314, + 0.0768902376294136, + 0.13524706661701202, + 0.04385695233941078, + 0.08242159336805344, + 0.21118959784507751, + 0.056710511445999146, + 0.07121201604604721, + 0.039107974618673325 + ], + "credit_map": [ + 0.01948590949177742, + 0.016240602359175682, + 0.017172235995531082, + 0.013717768713831902, + 0.020134015008807182, + 0.014299124479293823, + 0.01948426105082035, + 0.009595842100679874, + 0.01267218217253685, + 0.016702182590961456, + 0.023160353302955627, + 0.023464912548661232, + 0.012913431972265244, + 0.020099766552448273, + 0.01870008371770382, + 0.02552131935954094, + 0.01262619998306036, + 0.0069761076010763645, + 0.007975062355399132, + 0.006594387348741293, + 0.008910422213375568, + 0.01791967637836933, + 0.011811121366918087, + 0.004077265039086342, + 0.016928240656852722, + 0.01207438763231039, + 0.028369763866066933, + 0.020538823679089546, + 0.01147066056728363, + 0.019994238391518593, + 0.014203845523297787, + 0.011667128652334213, + 0.006735695526003838, + 0.005698327906429768, + 0.007243780884891748, + 0.0066302623599767685, + 0.003958849236369133, + 0.00530487485229969, + 0.005468628369271755, + 0.002816538093611598, + 0.0069580781273543835, + 0.008136607706546783, + 0.012147639878094196, + 0.01129940990358591, + 0.011479869484901428, + 0.008026848547160625, + 0.017502032220363617, + 0.006871113553643227, + 0.030909132212400436, + 0.024521486833691597, + 0.026823490858078003, + 0.038703326135873795, + 0.014993391931056976, + 0.02136402018368244, + 0.03719666600227356, + 0.016678113490343094, + 0.007096572779119015, + 0.007782819680869579, + 0.006505508907139301, + 0.006939209997653961, + 0.008097011595964432, + 0.006680903024971485, + 0.007203833200037479, + 0.006404652260243893, + 0.010809993371367455, + 0.014265638776123524, + 0.01006239652633667, + 0.009915070608258247, + 0.005476649384945631, + 0.0077421092428267, + 0.007250707130879164, + 0.005689458455890417, + 0.005201897583901882, + 0.0062144179828464985, + 0.004590075463056564, + 0.007748854346573353, + 0.0028566881082952023, + 0.004151524510234594, + 0.005507617257535458, + 0.002836903790012002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 3 + ], + "selected_reward": -0.12703923881053925 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an armored character wearing a full suit of plate armor with a T-shaped visor helmet and a red pixelated cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character galloping on a sturdy steed made of brown and tan pixels with four blocky legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are they crossing a bridge woven from thick green and brown pixelated vines spanning a bottomless digital abyss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature a blocky mountain range beneath a cyan pixelated sky with square white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side, performing a smooth dynamic tracking shot that follows the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017102628480643034 + ], + "opposite_logprob_tokens": [ + -6.376710414886475 + ], + "logprob_reward": -0.0017102628480643034, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017102628480643034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.077031709253788, + 0.09097325056791306, + 0.047032516449689865, + 0.09216243773698807, + 0.0647551417350769, + 0.07675809413194656, + 0.10802686959505081, + 0.02895689196884632, + 0.060233984142541885, + 0.3540691137313843 + ], + "credit_map": [ + 0.015382899902760983, + 0.00918227806687355, + 0.010768062435090542, + 0.009841475635766983, + 0.007182693108916283, + 0.007471454329788685, + 0.009542817249894142, + 0.007660027593374252, + 0.011752592399716377, + 0.009486107155680656, + 0.012783974409103394, + 0.011556586250662804, + 0.00813640933483839, + 0.011756072752177715, + 0.01255692820996046, + 0.01294457633048296, + 0.004706284496933222, + 0.005570086184889078, + 0.007768387906253338, + 0.005504687316715717, + 0.004183722659945488, + 0.006664701737463474, + 0.007822779007256031, + 0.0048118652775883675, + 0.011967713013291359, + 0.007586891297250986, + 0.019689520820975304, + 0.012556562200188637, + 0.009449128061532974, + 0.013134936802089214, + 0.009688829071819782, + 0.008088856004178524, + 0.006179725285619497, + 0.008748013526201248, + 0.009946059435606003, + 0.00882480014115572, + 0.008641372434794903, + 0.006061135325580835, + 0.01068404782563448, + 0.005669984966516495, + 0.0072753396816551685, + 0.007375522516667843, + 0.01479806937277317, + 0.012070250697433949, + 0.00805641245096922, + 0.007387143559753895, + 0.010667691007256508, + 0.009127665311098099, + 0.012311388738453388, + 0.012950673699378967, + 0.016043758019804955, + 0.02458943985402584, + 0.008812620304524899, + 0.009965290315449238, + 0.014747148379683495, + 0.00860655028373003, + 0.003575124079361558, + 0.004155752249062061, + 0.004859232343733311, + 0.003957727458328009, + 0.0033070005010813475, + 0.0029061781242489815, + 0.0037419931031763554, + 0.0024538857396692038, + 0.007550699636340141, + 0.008059389889240265, + 0.009619057178497314, + 0.010675475932657719, + 0.003745348658412695, + 0.005224942695349455, + 0.008944150991737843, + 0.0064149112440645695, + 0.041276950389146805, + 0.056298382580280304, + 0.038439031690359116, + 0.04951493814587593, + 0.056479182094335556, + 0.03752816095948219, + 0.05439231917262077, + 0.020140159875154495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 3 + ], + "selected_reward": -0.0017102628480643034 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an armored character wearing a full suit of plate armor with a T-shaped visor helmet and a red pixelated cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character galloping on a sturdy steed made of brown and tan pixels with four blocky legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are they crossing a bridge woven from thick green and brown pixelated vines spanning a bottomless digital abyss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature a blocky mountain range beneath a cyan pixelated sky with square white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side, performing a smooth dynamic tracking shot that follows the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029780618846416473 + ], + "opposite_logprob_tokens": [ + -3.529780626296997 + ], + "logprob_reward": -0.029780618846416473, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029780618846416473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09203134477138519, + 0.10176759213209152, + 0.083119697868824, + 0.10363557934761047, + 0.11325685679912567, + 0.22069180011749268, + 0.17515243589878082, + 0.04221950098872185, + 0.038971010595560074, + 0.029154181480407715 + ], + "credit_map": [ + 0.010306285694241524, + 0.009830869734287262, + 0.01606159657239914, + 0.009641366079449654, + 0.011800026521086693, + 0.007122480776160955, + 0.013484520837664604, + 0.013784203678369522, + 0.011115404777228832, + 0.011855117045342922, + 0.012292379513382912, + 0.014887605793774128, + 0.011714416556060314, + 0.009663970209658146, + 0.021279538050293922, + 0.008959157392382622, + 0.006253944709897041, + 0.0191600751131773, + 0.008771720342338085, + 0.009022661484777927, + 0.008192182518541813, + 0.015838872641324997, + 0.010108990594744682, + 0.0057712518610060215, + 0.00925939530134201, + 0.008091478608548641, + 0.019359130412340164, + 0.02109704725444317, + 0.009674478322267532, + 0.010457933880388737, + 0.01494376640766859, + 0.01075234916061163, + 0.01308359019458294, + 0.014877918176352978, + 0.015832729637622833, + 0.013179196044802666, + 0.016005342826247215, + 0.011011525988578796, + 0.01993327960371971, + 0.009333268739283085, + 0.019800571724772453, + 0.019461942836642265, + 0.040002088993787766, + 0.03301897272467613, + 0.027153104543685913, + 0.02026279829442501, + 0.036494720727205276, + 0.024497592821717262, + 0.018280433490872383, + 0.019022654742002487, + 0.02819666638970375, + 0.030231447890400887, + 0.015032408758997917, + 0.016137853264808655, + 0.030339714139699936, + 0.017911264672875404, + 0.005579216405749321, + 0.005256867967545986, + 0.00532913813367486, + 0.005552632734179497, + 0.00808163546025753, + 0.0045921471901237965, + 0.005095477681607008, + 0.0027323877438902855, + 0.005908129271119833, + 0.005367170553654432, + 0.005704739596694708, + 0.004592590034008026, + 0.0031964799854904413, + 0.004896755795925856, + 0.006662182509899139, + 0.0026429626159369946, + 0.0035415347665548325, + 0.005076480098068714, + 0.002940706443041563, + 0.004243427887558937, + 0.0042329574935138226, + 0.0025825107004493475, + 0.004834062419831753, + 0.0017025013221427798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.029780618846416473 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic tracking shot, shot from the side, perfectly follows an armored character at a smooth, high-frame rate as he gallops across a world composed entirely of vast pixel blocks. The character wears a full suit of plate armor composed of crisp silver pixels, a helmet with a T-shaped visor, and a cape of red pixels billowing dramatically behind him at high speed. He leans over, gripping the reins tightly, guiding a sturdy steed made of brown and tan pixels, its four blocky legs racing with a smooth, powerful looping animation. They cross a bridge woven from thick green and brown pixelated vines, spanning a bottomless digital abyss populated by slow-moving, occasionally glitchy purple and black data cubes. With each thump of the horse's hooves, the vine bridge sways slightly, its tiny pixelated leaves vibrating. In the distant background, a majestic mountain range, also made of blocks, rises beneath a cyan pixelated sky dotted with perfectly square white clouds. The entire scene is bathed in a bright, even digital light, with few shadows, highlighting the saturated and vibrant colors, all of which together create an intense, heart-pounding sense of nostalgic gaming retro charm.", + "seed": 30701, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_15/index_1674_seed_30701_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03661206462420523, + "vlm_logprob": -0.03661206462420523, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.03661206462420523, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an armored character wearing a full suit of plate armor with a T-shaped visor helmet and a red pixelated cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character galloping on a sturdy steed made of brown and tan pixels with four blocky legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are they crossing a bridge woven from thick green and brown pixelated vines spanning a bottomless digital abyss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature a blocky mountain range beneath a cyan pixelated sky with square white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side, performing a smooth dynamic tracking shot that follows the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08930650353431702 + ], + "opposite_logprob_tokens": [ + -2.464306592941284 + ], + "logprob_reward": -0.08930650353431702, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08930650353431702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31895652413368225, + 0.13356588780879974, + 0.0996163859963417, + 0.10166666656732559, + 0.10656704008579254, + 0.06189151108264923, + 0.04707881435751915, + 0.04923616722226143, + 0.06251955032348633, + 0.0189013984054327 + ], + "credit_map": [ + 0.0512271448969841, + 0.025344407185912132, + 0.04422091320157051, + 0.03179559111595154, + 0.0329226553440094, + 0.038297999650239944, + 0.0626627504825592, + 0.03248504176735878, + 0.010122348554432392, + 0.011708752252161503, + 0.016177533194422722, + 0.013609586283564568, + 0.01677335426211357, + 0.014655990526080132, + 0.022249076515436172, + 0.028269249945878983, + 0.025474173948168755, + 0.011859958060085773, + 0.00716150039806962, + 0.007401750423014164, + 0.006558798719197512, + 0.02599584497511387, + 0.00931538175791502, + 0.005848981440067291, + 0.01590353064239025, + 0.023888809606432915, + 0.009116769768297672, + 0.009055418893694878, + 0.00890031922608614, + 0.013924279250204563, + 0.01232206355780363, + 0.008555466309189796, + 0.011599426157772541, + 0.008546697907149792, + 0.021482698619365692, + 0.010044284164905548, + 0.014368923380970955, + 0.008253990672528744, + 0.014926314353942871, + 0.017344702035188675, + 0.00952958408743143, + 0.006782134063541889, + 0.006472777109593153, + 0.008299591951072216, + 0.008826172910630703, + 0.007472493685781956, + 0.007243132218718529, + 0.007265627384185791, + 0.004602083005011082, + 0.0035630222409963608, + 0.006913154385983944, + 0.0062370579689741135, + 0.005794412922114134, + 0.004031122662127018, + 0.010982125997543335, + 0.004955836106091738, + 0.007626750972121954, + 0.00456351088359952, + 0.004937706049531698, + 0.008486123755574226, + 0.004951733164489269, + 0.004969669971615076, + 0.007159790955483913, + 0.00654088519513607, + 0.005061648786067963, + 0.008270801045000553, + 0.008494983427226543, + 0.01382554229348898, + 0.007060860749334097, + 0.005683206487447023, + 0.008196129463613033, + 0.005926382727921009, + 0.0025459567550569773, + 0.0025718759279698133, + 0.0016800680896267295, + 0.0023141733836382627, + 0.002468529623001814, + 0.0017116833478212357, + 0.0033356635831296444, + 0.002273449208587408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.08930650353431702 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an armored character wearing a full suit of plate armor with a T-shaped visor helmet and a red pixelated cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character galloping on a sturdy steed made of brown and tan pixels with four blocky legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are they crossing a bridge woven from thick green and brown pixelated vines spanning a bottomless digital abyss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature a blocky mountain range beneath a cyan pixelated sky with square white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side, performing a smooth dynamic tracking shot that follows the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031987475231289864 + ], + "opposite_logprob_tokens": [ + -5.753198623657227 + ], + "logprob_reward": -0.0031987475231289864, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031987475231289864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2034977823495865, + 0.1293168067932129, + 0.13086554408073425, + 0.10859011113643646, + 0.14968241751194, + 0.0710238441824913, + 0.05949821323156357, + 0.054126281291246414, + 0.06150835379958153, + 0.031890708953142166 + ], + "credit_map": [ + 0.03138802945613861, + 0.027523193508386612, + 0.02034180611371994, + 0.01994236558675766, + 0.03148816153407097, + 0.02698702923953533, + 0.03160521015524864, + 0.014221975579857826, + 0.012756089679896832, + 0.015829844400286674, + 0.012158608995378017, + 0.014146864414215088, + 0.01740822196006775, + 0.014026246964931488, + 0.016741203144192696, + 0.026249712333083153, + 0.01944500021636486, + 0.011164187453687191, + 0.014870964922010899, + 0.015918679535388947, + 0.012200319208204746, + 0.019414860755205154, + 0.02673579752445221, + 0.0111157251521945, + 0.017120059579610825, + 0.02360885962843895, + 0.010341010056436062, + 0.010146318934857845, + 0.009176929481327534, + 0.014603833667933941, + 0.011858397163450718, + 0.011734697967767715, + 0.016986308619379997, + 0.012975472025573254, + 0.025545654818415642, + 0.013222468085587025, + 0.018845606595277786, + 0.01755058392882347, + 0.026264972984790802, + 0.018291344866156578, + 0.01076821330934763, + 0.007738006766885519, + 0.006035029422491789, + 0.009004147723317146, + 0.010462476871907711, + 0.008577176369726658, + 0.009691652841866016, + 0.008747134357690811, + 0.007187653332948685, + 0.004029083997011185, + 0.009156852029263973, + 0.009276408702135086, + 0.004678372759371996, + 0.00712747173383832, + 0.01080352533608675, + 0.007238836959004402, + 0.007212901953607798, + 0.006373405922204256, + 0.005368953570723534, + 0.010651580058038235, + 0.0051024798303842545, + 0.005388661287724972, + 0.007599264848977327, + 0.006429028231650591, + 0.005863504949957132, + 0.0048130834475159645, + 0.008622005581855774, + 0.017448272556066513, + 0.0040869396179914474, + 0.005778161808848381, + 0.010131210088729858, + 0.004765169695019722, + 0.004361222963780165, + 0.006958420854061842, + 0.004173279739916325, + 0.003128659212961793, + 0.0026208399794995785, + 0.0034400431904941797, + 0.003189715789631009, + 0.004018524661660194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.0031987475231289864 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an armored character wearing a full suit of plate armor with a T-shaped visor helmet and a red pixelated cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character galloping on a sturdy steed made of brown and tan pixels with four blocky legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are they crossing a bridge woven from thick green and brown pixelated vines spanning a bottomless digital abyss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature a blocky mountain range beneath a cyan pixelated sky with square white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side, performing a smooth dynamic tracking shot that follows the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062034618109464645 + ], + "opposite_logprob_tokens": [ + -2.8120346069335938 + ], + "logprob_reward": -0.062034618109464645, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062034618109464645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17598560452461243, + 0.11516398191452026, + 0.07741716504096985, + 0.16521258652210236, + 0.11928299069404602, + 0.11461414396762848, + 0.0565776489675045, + 0.08077362179756165, + 0.06067405268549919, + 0.034298188984394073 + ], + "credit_map": [ + 0.03492691367864609, + 0.020667554810643196, + 0.02192237786948681, + 0.01697276346385479, + 0.013414795510470867, + 0.02529021166265011, + 0.0229386854916811, + 0.019852299243211746, + 0.014596497640013695, + 0.016007166355848312, + 0.014927084557712078, + 0.01329843420535326, + 0.009954788722097874, + 0.01278572715818882, + 0.01632780022919178, + 0.01726647838950157, + 0.009118480607867241, + 0.0060591124929487705, + 0.009688523598015308, + 0.009473582729697227, + 0.006637991406023502, + 0.008450629189610481, + 0.017811061814427376, + 0.010177780874073505, + 0.025488385930657387, + 0.0440097451210022, + 0.012546936050057411, + 0.013987221755087376, + 0.017137696966528893, + 0.02221245877444744, + 0.015028219670057297, + 0.014801926910877228, + 0.017342448234558105, + 0.011506015434861183, + 0.010456510819494724, + 0.008515195921063423, + 0.010122932493686676, + 0.03055952489376068, + 0.015099280513823032, + 0.01568109169602394, + 0.017782708629965782, + 0.01053353026509285, + 0.013661950826644897, + 0.016909530386328697, + 0.012095658108592033, + 0.013457232154905796, + 0.011842367239296436, + 0.01833116076886654, + 0.005648743361234665, + 0.003932417836040258, + 0.011886286549270153, + 0.010422592051327229, + 0.003872714703902602, + 0.006635292898863554, + 0.007823175750672817, + 0.006356425117701292, + 0.012728694826364517, + 0.014533285982906818, + 0.006498737260699272, + 0.009821983985602856, + 0.009960695169866085, + 0.01177208125591278, + 0.00970668438822031, + 0.005751457065343857, + 0.006229954306036234, + 0.0071043772622942924, + 0.009753534570336342, + 0.009154975414276123, + 0.00379645312204957, + 0.00678302301093936, + 0.006415972486138344, + 0.01143576018512249, + 0.004172816872596741, + 0.004150609020143747, + 0.0042202528566122055, + 0.005713001824915409, + 0.0029074870981276035, + 0.003510856069624424, + 0.004836997948586941, + 0.004786166828125715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.062034618109464645 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an armored character wearing a full suit of plate armor with a T-shaped visor helmet and a red pixelated cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character galloping on a sturdy steed made of brown and tan pixels with four blocky legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are they crossing a bridge woven from thick green and brown pixelated vines spanning a bottomless digital abyss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature a blocky mountain range beneath a cyan pixelated sky with square white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side, performing a smooth dynamic tracking shot that follows the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021929519716650248 + ], + "opposite_logprob_tokens": [ + -6.127192974090576 + ], + "logprob_reward": -0.0021929519716650248, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021929519716650248, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1947806477546692, + 0.24358126521110535, + 0.0841258317232132, + 0.07384015619754791, + 0.13564276695251465, + 0.08185946941375732, + 0.04440785199403763, + 0.09063147753477097, + 0.02444743551313877, + 0.02668311633169651 + ], + "credit_map": [ + 0.028262848034501076, + 0.01806696131825447, + 0.03167695924639702, + 0.01878293603658676, + 0.019186921417713165, + 0.02294113300740719, + 0.029375528916716576, + 0.026487350463867188, + 0.027374768629670143, + 0.014271014370024204, + 0.06295868009328842, + 0.04071103036403656, + 0.013458519242703915, + 0.02830926701426506, + 0.02832820825278759, + 0.02816975675523281, + 0.009319793432950974, + 0.0060265990905463696, + 0.024043964222073555, + 0.011696594767272472, + 0.005206984933465719, + 0.00930686853826046, + 0.011441965587437153, + 0.0070830597542226315, + 0.00841996818780899, + 0.010101520456373692, + 0.007478108163923025, + 0.00729141291230917, + 0.008211003616452217, + 0.009107030928134918, + 0.016501057893037796, + 0.006730054039508104, + 0.015189345926046371, + 0.008391247130930424, + 0.03559058532118797, + 0.01312323473393917, + 0.011249409057199955, + 0.012423819862306118, + 0.019966427236795425, + 0.019708700478076935, + 0.01077213604003191, + 0.007164039183408022, + 0.007340865209698677, + 0.009811360388994217, + 0.009337667375802994, + 0.012965945526957512, + 0.011345582082867622, + 0.01312187034636736, + 0.004679149016737938, + 0.003127528354525566, + 0.009001013822853565, + 0.009122205898165703, + 0.0028944823425263166, + 0.005591926630586386, + 0.005482625216245651, + 0.004508918151259422, + 0.010513752698898315, + 0.008401823230087757, + 0.013080197386443615, + 0.02386763133108616, + 0.006187599617987871, + 0.008991806767880917, + 0.01063639298081398, + 0.008952274918556213, + 0.0035748735535889864, + 0.0023157962132245302, + 0.00292007508687675, + 0.00417521083727479, + 0.0023752343840897083, + 0.0029346614610403776, + 0.0036405837163329124, + 0.002510999795049429, + 0.003676073392853141, + 0.004612003453075886, + 0.0036277847830206156, + 0.003041601739823818, + 0.002935988362878561, + 0.0027348455041646957, + 0.0037064768839627504, + 0.0023483410477638245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.0021929519716650248 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an armored character wearing a full suit of plate armor with a T-shaped visor helmet and a red pixelated cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character galloping on a sturdy steed made of brown and tan pixels with four blocky legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are they crossing a bridge woven from thick green and brown pixelated vines spanning a bottomless digital abyss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature a blocky mountain range beneath a cyan pixelated sky with square white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side, performing a smooth dynamic tracking shot that follows the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026327501982450485 + ], + "opposite_logprob_tokens": [ + -3.651327610015869 + ], + "logprob_reward": -0.026327501982450485, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026327501982450485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2128978967666626, + 0.10838388651609421, + 0.08579502254724503, + 0.08931512385606766, + 0.22271540760993958, + 0.09484551101922989, + 0.042574796825647354, + 0.06924249976873398, + 0.05513921007514, + 0.019090648740530014 + ], + "credit_map": [ + 0.03055133856832981, + 0.02281162329018116, + 0.03984814137220383, + 0.015766866505146027, + 0.022479958832263947, + 0.022423770278692245, + 0.026752134785056114, + 0.032264068722724915, + 0.01282689068466425, + 0.007218259386718273, + 0.0268831979483366, + 0.009232614189386368, + 0.013868613168597221, + 0.0126310670748353, + 0.014768855646252632, + 0.010954393073916435, + 0.013328125700354576, + 0.006450943183153868, + 0.008752184920012951, + 0.009527930989861488, + 0.007707335986196995, + 0.01377285085618496, + 0.019829414784908295, + 0.006426239386200905, + 0.015269932337105274, + 0.015341078862547874, + 0.00844131875783205, + 0.009152587503194809, + 0.008504637517035007, + 0.015599880367517471, + 0.008710576221346855, + 0.008295117877423763, + 0.026940573006868362, + 0.017437860369682312, + 0.047091998159885406, + 0.0165623277425766, + 0.02349909394979477, + 0.03504820168018341, + 0.031094811856746674, + 0.025040550157427788, + 0.012585225515067577, + 0.007843195460736752, + 0.006011605728417635, + 0.015508851036429405, + 0.0126700010150671, + 0.011970741674304008, + 0.015440271236002445, + 0.012815612368285656, + 0.00571456179022789, + 0.003460316453129053, + 0.006024802569299936, + 0.00399750703945756, + 0.004700647667050362, + 0.007310271728783846, + 0.005752803757786751, + 0.005613887216895819, + 0.007845253683626652, + 0.007239622063934803, + 0.009215676225721836, + 0.014655585400760174, + 0.007450007367879152, + 0.006518938113003969, + 0.008923298679292202, + 0.00739411823451519, + 0.007364295888692141, + 0.006262788083404303, + 0.0065279812552034855, + 0.00915075745433569, + 0.007123821880668402, + 0.005829626228660345, + 0.007854912430047989, + 0.005025027319788933, + 0.002195164794102311, + 0.0021905971225351095, + 0.003220819868147373, + 0.003067970974370837, + 0.0016468982212245464, + 0.002437668852508068, + 0.0018860650015994906, + 0.0024454647209495306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.026327501982450485 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic tracking shot, shot from the side, perfectly follows an armored character at a smooth, high-frame rate as he gallops across a world composed entirely of vast pixel blocks. The character wears a full suit of plate armor composed of crisp silver pixels, a helmet with a T-shaped visor, and a cape of red pixels billowing dramatically behind him at high speed. He leans over, gripping the reins tightly, guiding a sturdy steed made of brown and tan pixels, its four blocky legs racing with a smooth, powerful looping animation. They cross a bridge woven from thick green and brown pixelated vines, spanning a bottomless digital abyss populated by slow-moving, occasionally glitchy purple and black data cubes. With each thump of the horse's hooves, the vine bridge sways slightly, its tiny pixelated leaves vibrating. In the distant background, a majestic mountain range, also made of blocks, rises beneath a cyan pixelated sky dotted with perfectly square white clouds. The entire scene is bathed in a bright, even digital light, with few shadows, highlighting the saturated and vibrant colors, all of which together create an intense, heart-pounding sense of nostalgic gaming retro charm.", + "seed": 30702, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_15/index_1674_seed_30702_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.045978601463139056, + "vlm_logprob": -0.045978601463139056, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.045978601463139056, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an armored character wearing a full suit of plate armor with a T-shaped visor helmet and a red pixelated cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character galloping on a sturdy steed made of brown and tan pixels with four blocky legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are they crossing a bridge woven from thick green and brown pixelated vines spanning a bottomless digital abyss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature a blocky mountain range beneath a cyan pixelated sky with square white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side, performing a smooth dynamic tracking shot that follows the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11324768513441086 + ], + "opposite_logprob_tokens": [ + -2.2382476329803467 + ], + "logprob_reward": -0.11324768513441086, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11324768513441086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2980292737483978, + 0.045184943825006485, + 0.29668787121772766, + 0.0751432478427887, + 0.04917343333363533, + 0.09120745956897736, + 0.028250878676772118, + 0.0333753377199173, + 0.02775481343269348, + 0.0551927424967289 + ], + "credit_map": [ + 0.030630625784397125, + 0.050539519637823105, + 0.01696690544486046, + 0.040533293038606644, + 0.04105557128787041, + 0.04348333179950714, + 0.04874790832400322, + 0.026072099804878235, + 0.008480443619191647, + 0.003784378059208393, + 0.00518812844529748, + 0.005864899605512619, + 0.005138305947184563, + 0.0071341232396662235, + 0.005723004695028067, + 0.0038716578856110573, + 0.03883399814367294, + 0.03819079324603081, + 0.02459201216697693, + 0.04976843670010567, + 0.042501095682382584, + 0.03801662474870682, + 0.03883052617311478, + 0.025954391807317734, + 0.011730988509953022, + 0.005495916586369276, + 0.006842190399765968, + 0.00777657562866807, + 0.008143949322402477, + 0.01965189166367054, + 0.009514064528048038, + 0.005987668409943581, + 0.007052750326693058, + 0.007232391741126776, + 0.004242908209562302, + 0.006965601816773415, + 0.006579988170415163, + 0.008051127195358276, + 0.004018212202936411, + 0.005030452739447355, + 0.009364848025143147, + 0.010146079584956169, + 0.01127154566347599, + 0.02098257839679718, + 0.00806337594985962, + 0.010747024789452553, + 0.010006112046539783, + 0.010625895112752914, + 0.004399660974740982, + 0.003334016539156437, + 0.0032567086163908243, + 0.002914063399657607, + 0.003070787526667118, + 0.003987329080700874, + 0.0028354651294648647, + 0.004452846944332123, + 0.004483473487198353, + 0.004217862617224455, + 0.004053929820656776, + 0.0042643421329557896, + 0.004787484183907509, + 0.004130234941840172, + 0.002975278999656439, + 0.004462728276848793, + 0.0026375213637948036, + 0.002656213939189911, + 0.004229368641972542, + 0.004755823872983456, + 0.003230369882658124, + 0.002787078497931361, + 0.002963480306789279, + 0.004494958091527224, + 0.005519516300410032, + 0.006829395890235901, + 0.0056661502458155155, + 0.007017333991825581, + 0.008504076860845089, + 0.006596632767468691, + 0.008818056434392929, + 0.006241577211767435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.11324768513441086 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an armored character wearing a full suit of plate armor with a T-shaped visor helmet and a red pixelated cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character galloping on a sturdy steed made of brown and tan pixels with four blocky legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are they crossing a bridge woven from thick green and brown pixelated vines spanning a bottomless digital abyss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature a blocky mountain range beneath a cyan pixelated sky with square white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side, performing a smooth dynamic tracking shot that follows the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059547401033341885 + ], + "opposite_logprob_tokens": [ + -5.130954742431641 + ], + "logprob_reward": -0.0059547401033341885, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059547401033341885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19330069422721863, + 0.03142745792865753, + 0.10632108896970749, + 0.048020925372838974, + 0.05818807706236839, + 0.041757818311452866, + 0.15405893325805664, + 0.04962605610489845, + 0.05111001059412956, + 0.26618891954421997 + ], + "credit_map": [ + 0.020270418375730515, + 0.022822601720690727, + 0.01625729724764824, + 0.0280934888869524, + 0.025591574609279633, + 0.027667535468935966, + 0.033240050077438354, + 0.019357729703187943, + 0.0043658665381371975, + 0.003803219646215439, + 0.0026953660417348146, + 0.003306278493255377, + 0.0035927556455135345, + 0.004136157687753439, + 0.0035506426356732845, + 0.005977172404527664, + 0.01206468790769577, + 0.010226805694401264, + 0.010459446348249912, + 0.016782887279987335, + 0.015414555557072163, + 0.014263903722167015, + 0.013078920543193817, + 0.014029888436198235, + 0.00781896710395813, + 0.0033282411750406027, + 0.003153675002977252, + 0.005000733770430088, + 0.00487843481823802, + 0.012849409133195877, + 0.007622710429131985, + 0.0033687546383589506, + 0.009047029539942741, + 0.006946923211216927, + 0.005029851105064154, + 0.0067819105461239815, + 0.007038426119834185, + 0.01449546031653881, + 0.003866807324811816, + 0.0049816700629889965, + 0.005723270121961832, + 0.0035601588897407055, + 0.005395138636231422, + 0.007594272959977388, + 0.00341902207583189, + 0.007190444506704807, + 0.0038918862119317055, + 0.004983626306056976, + 0.025685183703899384, + 0.012563440948724747, + 0.02383807674050331, + 0.015351111069321632, + 0.02036629058420658, + 0.022903528064489365, + 0.011132225394248962, + 0.02221907489001751, + 0.008805453777313232, + 0.004569278564304113, + 0.007132853846997023, + 0.0067263273522257805, + 0.005274479743093252, + 0.00596380652859807, + 0.003923092968761921, + 0.007230763323605061, + 0.005606299266219139, + 0.004854325670748949, + 0.0070334733463823795, + 0.010195116512477398, + 0.004695597570389509, + 0.004991484340280294, + 0.005559139419347048, + 0.008174573071300983, + 0.03135311231017113, + 0.026756085455417633, + 0.029167164117097855, + 0.04159542918205261, + 0.02850758284330368, + 0.028947781771421432, + 0.05364708602428436, + 0.026214664801955223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.0059547401033341885 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an armored character wearing a full suit of plate armor with a T-shaped visor helmet and a red pixelated cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character galloping on a sturdy steed made of brown and tan pixels with four blocky legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are they crossing a bridge woven from thick green and brown pixelated vines spanning a bottomless digital abyss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature a blocky mountain range beneath a cyan pixelated sky with square white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side, performing a smooth dynamic tracking shot that follows the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897060364484787 + ], + "opposite_logprob_tokens": [ + -2.5789706707000732 + ], + "logprob_reward": -0.07897060364484787, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897060364484787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2038663625717163, + 0.0724121481180191, + 0.27273452281951904, + 0.09519871324300766, + 0.07994641363620758, + 0.0648636668920517, + 0.07420768588781357, + 0.10134206712245941, + 0.02046949602663517, + 0.014958872459828854 + ], + "credit_map": [ + 0.02501276135444641, + 0.027027957141399384, + 0.02157408557832241, + 0.021561747416853905, + 0.028051303699612617, + 0.030784446746110916, + 0.025096669793128967, + 0.024757392704486847, + 0.016534239053726196, + 0.00538397952914238, + 0.006755540147423744, + 0.008970470167696476, + 0.0092705013230443, + 0.013952984474599361, + 0.0055142841301858425, + 0.006030142307281494, + 0.03223415091633797, + 0.02323015034198761, + 0.031542543321847916, + 0.05141850933432579, + 0.03193294629454613, + 0.03093518316745758, + 0.044059161096811295, + 0.02738189324736595, + 0.010601883754134178, + 0.007839085534214973, + 0.007795803714543581, + 0.011267552152276039, + 0.00808777380734682, + 0.015080848708748817, + 0.026234539225697517, + 0.008291225880384445, + 0.012824971228837967, + 0.00602429173886776, + 0.006933828350156546, + 0.009390744380652905, + 0.007947773672640324, + 0.023497169837355614, + 0.005858635995537043, + 0.007468997500836849, + 0.009197568520903587, + 0.008552971296012402, + 0.007745405659079552, + 0.00968717597424984, + 0.008390688337385654, + 0.008379848673939705, + 0.005445373244583607, + 0.00746463518589735, + 0.011419788002967834, + 0.0059468382969498634, + 0.013263517990708351, + 0.008134241215884686, + 0.009211603552103043, + 0.009532241150736809, + 0.005614410620182753, + 0.011085044592618942, + 0.012895048595964909, + 0.01047965232282877, + 0.011376545764505863, + 0.012590890750288963, + 0.01996786892414093, + 0.012449796311557293, + 0.009330683387815952, + 0.012251579202711582, + 0.0037018596194684505, + 0.00216916436329484, + 0.0024000569246709347, + 0.0029476911295205355, + 0.002416192786768079, + 0.002827693708240986, + 0.001913099898956716, + 0.002093737479299307, + 0.0017937864176928997, + 0.0013679753756150603, + 0.0024070239160209894, + 0.0018248872365802526, + 0.0013610764872282743, + 0.0015366134466603398, + 0.0020861164666712284, + 0.0025813935790210962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.07897060364484787 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an armored character wearing a full suit of plate armor with a T-shaped visor helmet and a red pixelated cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character galloping on a sturdy steed made of brown and tan pixels with four blocky legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are they crossing a bridge woven from thick green and brown pixelated vines spanning a bottomless digital abyss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature a blocky mountain range beneath a cyan pixelated sky with square white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side, performing a smooth dynamic tracking shot that follows the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019364668987691402 + ], + "opposite_logprob_tokens": [ + -6.251936435699463 + ], + "logprob_reward": -0.0019364668987691402, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019364668987691402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11502374708652496, + 0.16414418816566467, + 0.12170463800430298, + 0.09555064141750336, + 0.16573946177959442, + 0.05616433173418045, + 0.09749700129032135, + 0.041966743767261505, + 0.08848348259925842, + 0.05372576043009758 + ], + "credit_map": [ + 0.010885193012654781, + 0.011826518923044205, + 0.007470182608813047, + 0.025790246203541756, + 0.01360333152115345, + 0.011390923522412777, + 0.022741302847862244, + 0.011316048912703991, + 0.05038361996412277, + 0.012929209508001804, + 0.00964730978012085, + 0.013839089311659336, + 0.019864538684487343, + 0.03672248125076294, + 0.009155288338661194, + 0.011602646671235561, + 0.013179824687540531, + 0.006816352251917124, + 0.014093069359660149, + 0.03152865171432495, + 0.00942225567996502, + 0.014077037572860718, + 0.018570179119706154, + 0.014017270877957344, + 0.011186308227479458, + 0.008825447410345078, + 0.0081432294100523, + 0.010649650357663631, + 0.010554778389632702, + 0.013599194586277008, + 0.024713978171348572, + 0.007878054864704609, + 0.02477913163602352, + 0.012858019210398197, + 0.018109098076820374, + 0.022713057696819305, + 0.013910578563809395, + 0.04381972551345825, + 0.012928443029522896, + 0.01662142015993595, + 0.007779473904520273, + 0.005551678594201803, + 0.0072391824796795845, + 0.00914825964719057, + 0.006676303222775459, + 0.0075303902849555016, + 0.00531011912971735, + 0.006928924936801195, + 0.013819100335240364, + 0.0074498881585896015, + 0.019284013658761978, + 0.010173492133617401, + 0.011112523265182972, + 0.012623423710465431, + 0.007266344036906958, + 0.01576821319758892, + 0.006057029590010643, + 0.006234648171812296, + 0.004410014487802982, + 0.0064664869569242, + 0.005597140174359083, + 0.0048741744831204414, + 0.004022770561277866, + 0.0043044788762927055, + 0.014972470700740814, + 0.007070179563015699, + 0.009611573070287704, + 0.01804392784833908, + 0.008202685043215752, + 0.011407149955630302, + 0.00937836617231369, + 0.009797129780054092, + 0.007587721571326256, + 0.004144053906202316, + 0.006706243380904198, + 0.008305246010422707, + 0.004620740190148354, + 0.005963925272226334, + 0.006241850089281797, + 0.010155981406569481 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.0019364668987691402 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an armored character wearing a full suit of plate armor with a T-shaped visor helmet and a red pixelated cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character galloping on a sturdy steed made of brown and tan pixels with four blocky legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are they crossing a bridge woven from thick green and brown pixelated vines spanning a bottomless digital abyss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature a blocky mountain range beneath a cyan pixelated sky with square white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side, performing a smooth dynamic tracking shot that follows the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978351153433323 + ], + "opposite_logprob_tokens": [ + -3.5297834873199463 + ], + "logprob_reward": -0.02978351153433323, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978351153433323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15264156460762024, + 0.17245568335056305, + 0.1231059581041336, + 0.047557611018419266, + 0.14377257227897644, + 0.05168665200471878, + 0.13397538661956787, + 0.05200366675853729, + 0.05281030759215355, + 0.06999053806066513 + ], + "credit_map": [ + 0.017373334616422653, + 0.020550522953271866, + 0.015076962299644947, + 0.015877125784754753, + 0.017878493294119835, + 0.019417913630604744, + 0.026845822110772133, + 0.019621403887867928, + 0.020847151055932045, + 0.011692196130752563, + 0.02362293191254139, + 0.016263632103800774, + 0.02130098082125187, + 0.025222046300768852, + 0.01790151558816433, + 0.035605233162641525, + 0.016097232699394226, + 0.012978446669876575, + 0.008704027161002159, + 0.025584835559129715, + 0.015437893569469452, + 0.020663687959313393, + 0.013457571156322956, + 0.0101822754368186, + 0.006902189459651709, + 0.005100014619529247, + 0.005363140720874071, + 0.004883126821368933, + 0.006169525906443596, + 0.010139747522771358, + 0.004994343034923077, + 0.004005527123808861, + 0.02225649356842041, + 0.011407042853534222, + 0.01584639959037304, + 0.020076913759112358, + 0.011499888263642788, + 0.03895013406872749, + 0.009899498894810677, + 0.013836224563419819, + 0.007486931513994932, + 0.008211747743189335, + 0.006192642729729414, + 0.005874212831258774, + 0.00745048001408577, + 0.006631897296756506, + 0.0046493555419147015, + 0.005189390853047371, + 0.023013142868876457, + 0.008046962320804596, + 0.01886506751179695, + 0.017087941989302635, + 0.011995427310466766, + 0.023697176948189735, + 0.009692907333374023, + 0.02157677337527275, + 0.010421277023851871, + 0.005403237417340279, + 0.0072992113418877125, + 0.005256807431578636, + 0.006243527866899967, + 0.006942877545952797, + 0.004543187562376261, + 0.005893549416214228, + 0.0054915836080908775, + 0.005066134501248598, + 0.008277902379631996, + 0.008844004012644291, + 0.005256169475615025, + 0.004732949193567038, + 0.005217518657445908, + 0.009924052283167839, + 0.006419557146728039, + 0.007562225218862295, + 0.00822931993752718, + 0.010854769498109818, + 0.008906085975468159, + 0.00781773030757904, + 0.010227063670754433, + 0.009973792359232903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.02978351153433323 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_2/index_tensor([2666, 2666])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_2/index_tensor([2666, 2666])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f6459dfe0628975d2e25cae6cffa9839831093de --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_2/index_tensor([2666, 2666])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "fixed, overhead bird's-eye view, with a near-Godlike perspective, calmly observes an elderly man with a sickly face. He lies on his back on a bed covered in slightly wrinkled off-white cotton sheets. His sparse, graying hair is matted to a sweat-soaked forehead, deep wrinkles knotted tightly together with pain. He wears a pale, light-blue striped hospital gown, one button undone at the collar, revealing the shrunken skin below his collarbone. An old-fashioned lamp on the bedside table emits the only warm yellow light, casting long, soft shadows in the dim room, accentuating his sunken eye sockets and cheekbones. Suddenly, a violent cough seizes him. He struggles to tilt his head to one side, his body instinctively hunching over. His thin shoulders heave violently and spasmodically with each deep, hoarse cough, and the dark gray blanket covering his chest rises and falls with it. This painful process is slow and difficult, and the entire picture presents a gloomy oil painting texture under the dim light, creating a heartbreaking atmosphere of tranquility and depression as a life struggles in illness.", + "seed": 31483, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_2/index_2666_seed_31483_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.34904237017035483, + "vlm_logprob": -0.34904237017035483, + "vlm_yes_no_margin": 1.8, + "vlm_token_credit": -0.34904237017035483, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed overhead bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elderly man wearing a pale light-blue striped hospital gown with one undone button?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by a warm yellow old-fashioned lamp casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a violent coughing action with heaving shoulders and a rising/falling blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a gloomy oil painting texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9200172424316406 + ], + "opposite_logprob_tokens": [ + -0.5450172424316406 + ], + "logprob_reward": -0.9200172424316406, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.9200172424316406, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3842180669307709, + 0.05193063244223595, + 0.06469228863716125, + 0.1252809315919876, + 0.0404944084584713, + 0.04094955697655678, + 0.018405456095933914, + 0.04057583212852478, + 0.08906302601099014, + 0.14438973367214203 + ], + "credit_map": [ + 0.02508392371237278, + 0.07118602842092514, + 0.027834586799144745, + 0.04200748726725578, + 0.08934124559164047, + 0.043664880096912384, + 0.046686042100191116, + 0.0384138785302639, + 0.0038128583692014217, + 0.00814125221222639, + 0.002954037394374609, + 0.005898868199437857, + 0.012296668253839016, + 0.007216167636215687, + 0.006858950946480036, + 0.004751826636493206, + 0.003874157089740038, + 0.01361132226884365, + 0.004196522291749716, + 0.0070265596732497215, + 0.01145758293569088, + 0.008296716026961803, + 0.008359720930457115, + 0.007869706489145756, + 0.010194873437285423, + 0.02861599065363407, + 0.009655523113906384, + 0.01013769581913948, + 0.023372484371066093, + 0.014994243159890175, + 0.014696630649268627, + 0.013613485731184483, + 0.0030135191045701504, + 0.00811123289167881, + 0.0025867014192044735, + 0.00415733503177762, + 0.006243051495403051, + 0.005834127310663462, + 0.006085710600018501, + 0.0044627320021390915, + 0.005397783126682043, + 0.006505362223833799, + 0.005403431132435799, + 0.004749023821204901, + 0.006036264356225729, + 0.004370403941720724, + 0.004332306794822216, + 0.00415498111397028, + 0.001338692964054644, + 0.0030569778755307198, + 0.0009653259767219424, + 0.0022562220692634583, + 0.00402821134775877, + 0.0027615067083388567, + 0.0022903308272361755, + 0.001708187977783382, + 0.002508286852389574, + 0.006891670636832714, + 0.0019488456891849637, + 0.005222859792411327, + 0.007261907681822777, + 0.006282901391386986, + 0.004389422945678234, + 0.0060699391178786755, + 0.005659619811922312, + 0.012825564481317997, + 0.006549247540533543, + 0.011598879471421242, + 0.018804626539349556, + 0.012007040902972221, + 0.012603899464011192, + 0.009014143608510494, + 0.008117363788187504, + 0.027353636920452118, + 0.005124107003211975, + 0.012241381220519543, + 0.02638090029358864, + 0.037770744413137436, + 0.016804365441203117, + 0.010597246699035168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.9200172424316406 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed overhead bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elderly man wearing a pale light-blue striped hospital gown with one undone button?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by a warm yellow old-fashioned lamp casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a violent coughing action with heaving shoulders and a rising/falling blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a gloomy oil painting texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033737171441316605 + ], + "opposite_logprob_tokens": [ + -3.4087371826171875 + ], + "logprob_reward": -0.033737171441316605, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033737171441316605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21062229573726654, + 0.16832266747951508, + 0.04063491150736809, + 0.11039992421865463, + 0.04615927115082741, + 0.06277219206094742, + 0.04468788206577301, + 0.03884506970643997, + 0.10309900343418121, + 0.17445681989192963 + ], + "credit_map": [ + 0.0151048069819808, + 0.033889152109622955, + 0.016721097752451897, + 0.02876320108771324, + 0.045010317116975784, + 0.023438362404704094, + 0.02482948638498783, + 0.022865885868668556, + 0.010848723351955414, + 0.01955241523683071, + 0.010997931472957134, + 0.02074681967496872, + 0.022342413663864136, + 0.02735489048063755, + 0.03479670360684395, + 0.02168276160955429, + 0.002763202181085944, + 0.005366484168916941, + 0.0022473190911114216, + 0.004201071336865425, + 0.008703876286745071, + 0.007597044575959444, + 0.00588624645024538, + 0.003869669744744897, + 0.0076996637508273125, + 0.02167881280183792, + 0.006788893602788448, + 0.009194033220410347, + 0.025876056402921677, + 0.013872167095541954, + 0.013866356573998928, + 0.011423936113715172, + 0.0025224704295396805, + 0.0066492692567408085, + 0.003000503173097968, + 0.006380817387253046, + 0.007834415882825851, + 0.0054043615236878395, + 0.009323266334831715, + 0.005044163670390844, + 0.005339403636753559, + 0.010104836896061897, + 0.006418440490961075, + 0.009110856801271439, + 0.008259625174105167, + 0.00621101213619113, + 0.00998776312917471, + 0.007340256590396166, + 0.0026749265380203724, + 0.005963609088212252, + 0.002752245170995593, + 0.006952962838113308, + 0.009048745967447758, + 0.004773312248289585, + 0.006960789207369089, + 0.005561289377510548, + 0.0022155330516397953, + 0.005360806360840797, + 0.002153577283024788, + 0.004524391610175371, + 0.005328748840838671, + 0.008184780366718769, + 0.004161972552537918, + 0.006915264762938023, + 0.006328326649963856, + 0.013730263337492943, + 0.007226716727018356, + 0.01339644379913807, + 0.018406370654702187, + 0.014291517436504364, + 0.017055559903383255, + 0.01266380026936531, + 0.007961476221680641, + 0.021707337349653244, + 0.007427385542541742, + 0.021335434168577194, + 0.03530532866716385, + 0.031088754534721375, + 0.028105836361646652, + 0.02152526192367077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.033737171441316605 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed overhead bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elderly man wearing a pale light-blue striped hospital gown with one undone button?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by a warm yellow old-fashioned lamp casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a violent coughing action with heaving shoulders and a rising/falling blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a gloomy oil painting texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308703541755676 + ], + "opposite_logprob_tokens": [ + -3.1680870056152344 + ], + "logprob_reward": -0.04308703541755676, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308703541755676, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2173396646976471, + 0.09442242980003357, + 0.04573075845837593, + 0.058195337653160095, + 0.10143302381038666, + 0.19434593617916107, + 0.016706908121705055, + 0.12186282873153687, + 0.05289429426193237, + 0.09706874936819077 + ], + "credit_map": [ + 0.020073560997843742, + 0.04023104906082153, + 0.01272414717823267, + 0.02071533165872097, + 0.048910271376371384, + 0.02596990577876568, + 0.029016094282269478, + 0.019699307158589363, + 0.00889180600643158, + 0.013618698343634605, + 0.007042660843580961, + 0.010903974995017052, + 0.01321103423833847, + 0.015260831452906132, + 0.015623433515429497, + 0.00987000111490488, + 0.004534128587692976, + 0.007475271821022034, + 0.004794664215296507, + 0.005796638783067465, + 0.00889134593307972, + 0.004854096565395594, + 0.004737592302262783, + 0.004647024441510439, + 0.004950713366270065, + 0.010683446191251278, + 0.005109712481498718, + 0.005365177057683468, + 0.01243759598582983, + 0.00609408738091588, + 0.006319941021502018, + 0.007234666030853987, + 0.010913079604506493, + 0.019423874095082283, + 0.009437772445380688, + 0.009877695702016354, + 0.01405238825827837, + 0.015194229781627655, + 0.011904053390026093, + 0.010629936121404171, + 0.023823902010917664, + 0.0348447822034359, + 0.019065510481595993, + 0.020773282274603844, + 0.031609632074832916, + 0.021817918866872787, + 0.023410627618432045, + 0.01900029554963112, + 0.0018919749418273568, + 0.002291630720719695, + 0.0013632432091981173, + 0.0019920372869819403, + 0.0032219230197370052, + 0.0020777208264917135, + 0.0020973612554371357, + 0.0017710180254653096, + 0.011808693408966064, + 0.0137619124725461, + 0.010873541235923767, + 0.01157242152839899, + 0.014170748181641102, + 0.027748573571443558, + 0.010183876380324364, + 0.021743066608905792, + 0.0027209597174078226, + 0.0061818077228963375, + 0.003604827681556344, + 0.009876945056021214, + 0.010837390087544918, + 0.005452489480376244, + 0.007529190741479397, + 0.006690684240311384, + 0.0057953293435275555, + 0.015216633677482605, + 0.004247490782290697, + 0.01135468389838934, + 0.013038786128163338, + 0.0209891926497221, + 0.01724429801106453, + 0.009182345122098923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.04308703541755676 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed overhead bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elderly man wearing a pale light-blue striped hospital gown with one undone button?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by a warm yellow old-fashioned lamp casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a violent coughing action with heaving shoulders and a rising/falling blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a gloomy oil painting texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933957934379578 + ], + "opposite_logprob_tokens": [ + -0.6933957934379578 + ], + "logprob_reward": -0.6933957934379578, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933957934379578, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08503205329179764, + 0.031647689640522, + 0.041089460253715515, + 0.06501781940460205, + 0.03029579296708107, + 0.029436852782964706, + 0.02236292138695717, + 0.06937136501073837, + 0.11477349698543549, + 0.5109724402427673 + ], + "credit_map": [ + 0.007753088604658842, + 0.014752713963389397, + 0.006048541981726885, + 0.009860554710030556, + 0.01598278619349003, + 0.009717262350022793, + 0.012790325097739697, + 0.008126803673803806, + 0.002383205574005842, + 0.005257010459899902, + 0.0025845409836620092, + 0.0032792442943900824, + 0.00538058253005147, + 0.0036762231029570103, + 0.005157717503607273, + 0.0039291721768677235, + 0.002145341131836176, + 0.007272213231772184, + 0.0022546027321368456, + 0.004213239066302776, + 0.008607495576143265, + 0.006262879818677902, + 0.005988989025354385, + 0.004344705492258072, + 0.0034699381794780493, + 0.010789060033857822, + 0.0033617608714848757, + 0.005883020348846912, + 0.015486782416701317, + 0.008750124834477901, + 0.010646600276231766, + 0.006630541756749153, + 0.002558118663728237, + 0.005854475777596235, + 0.0017886952264234424, + 0.002797748427838087, + 0.004356858320534229, + 0.005548691377043724, + 0.004583209287375212, + 0.002808002522215247, + 0.0020133035723119974, + 0.0040517887100577354, + 0.0016424014465883374, + 0.0040368554182350636, + 0.006398603320121765, + 0.0038747875951230526, + 0.004119770135730505, + 0.003299349220469594, + 0.0018275880720466375, + 0.003540518693625927, + 0.0012026962358504534, + 0.003047824138775468, + 0.00457421038299799, + 0.002583414316177368, + 0.0026809026021510363, + 0.0029057725332677364, + 0.004285210743546486, + 0.01001760270446539, + 0.0033906351309269667, + 0.008289042860269547, + 0.012700001709163189, + 0.011746601201593876, + 0.010804078541696072, + 0.008138203993439674, + 0.006232432555407286, + 0.012652778998017311, + 0.007645590230822563, + 0.012281435541808605, + 0.024662425741553307, + 0.014059069566428661, + 0.022801270708441734, + 0.014438522048294544, + 0.027728933840990067, + 0.07605300098657608, + 0.018672365695238113, + 0.04493970423936844, + 0.11854816228151321, + 0.11947369575500488, + 0.06581166386604309, + 0.039745036512613297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.6933957934379578 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed overhead bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elderly man wearing a pale light-blue striped hospital gown with one undone button?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by a warm yellow old-fashioned lamp casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a violent coughing action with heaving shoulders and a rising/falling blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a gloomy oil painting texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497460812330246 + ], + "opposite_logprob_tokens": [ + -2.9299745559692383 + ], + "logprob_reward": -0.05497460812330246, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497460812330246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27692896127700806, + 0.09698645025491714, + 0.047608017921447754, + 0.06390141695737839, + 0.08299451321363449, + 0.08421343564987183, + 0.05137695372104645, + 0.07830842584371567, + 0.08959093689918518, + 0.12809081375598907 + ], + "credit_map": [ + 0.02666572853922844, + 0.03621650114655495, + 0.017926938831806183, + 0.0336260125041008, + 0.06246249005198479, + 0.042960431426763535, + 0.03374508395791054, + 0.023325800895690918, + 0.007726409938186407, + 0.01271040178835392, + 0.008094257675111294, + 0.011707383207976818, + 0.016076352447271347, + 0.013688127510249615, + 0.014181933365762234, + 0.012801592238247395, + 0.003313426161184907, + 0.005020336247980595, + 0.002718732226639986, + 0.005781741812825203, + 0.008483384735882282, + 0.006396077573299408, + 0.007879051379859447, + 0.008015266619622707, + 0.003911083098500967, + 0.01026806142181158, + 0.003795393742620945, + 0.006218079011887312, + 0.014129646122455597, + 0.009605901315808296, + 0.009472418576478958, + 0.006500840187072754, + 0.009703048504889011, + 0.0162974800914526, + 0.008120955899357796, + 0.008394746109843254, + 0.011023684404790401, + 0.009594389237463474, + 0.009925885125994682, + 0.009934338741004467, + 0.01338426023721695, + 0.01354141067713499, + 0.010206667706370354, + 0.008505625650286674, + 0.011979988776147366, + 0.009737079963088036, + 0.0089930584654212, + 0.00786534883081913, + 0.0034632605966180563, + 0.007361497264355421, + 0.00281026610173285, + 0.004184473305940628, + 0.010053739883005619, + 0.00895619299262762, + 0.008041427470743656, + 0.006506102625280619, + 0.004815735854208469, + 0.007296126335859299, + 0.0037750178016722202, + 0.006137109361588955, + 0.008119099773466587, + 0.023710496723651886, + 0.006801960524171591, + 0.017652885988354683, + 0.00516302278265357, + 0.010553425177931786, + 0.0048350319266319275, + 0.01484711468219757, + 0.018991392105817795, + 0.012617682106792927, + 0.012258774600923061, + 0.010324503295123577, + 0.006590374279767275, + 0.015597287565469742, + 0.0059019215404987335, + 0.015701230615377426, + 0.027087748050689697, + 0.016531743109226227, + 0.02869899943470955, + 0.01198152918368578 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.05497460812330246 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "fixed, overhead bird's-eye view, with a near-Godlike perspective, calmly observes an elderly man with a sickly face. He lies on his back on a bed covered in slightly wrinkled off-white cotton sheets. His sparse, graying hair is matted to a sweat-soaked forehead, deep wrinkles knotted tightly together with pain. He wears a pale, light-blue striped hospital gown, one button undone at the collar, revealing the shrunken skin below his collarbone. An old-fashioned lamp on the bedside table emits the only warm yellow light, casting long, soft shadows in the dim room, accentuating his sunken eye sockets and cheekbones. Suddenly, a violent cough seizes him. He struggles to tilt his head to one side, his body instinctively hunching over. His thin shoulders heave violently and spasmodically with each deep, hoarse cough, and the dark gray blanket covering his chest rises and falls with it. This painful process is slow and difficult, and the entire picture presents a gloomy oil painting texture under the dim light, creating a heartbreaking atmosphere of tranquility and depression as a life struggles in illness.", + "seed": 31484, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_2/index_2666_seed_31484_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.4758619487285614, + "vlm_logprob": -0.4758619487285614, + "vlm_yes_no_margin": 1.225, + "vlm_token_credit": -0.4758619487285614, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed overhead bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elderly man wearing a pale light-blue striped hospital gown with one undone button?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by a warm yellow old-fashioned lamp casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a violent coughing action with heaving shoulders and a rising/falling blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a gloomy oil painting texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.6238306760787964 + ], + "opposite_logprob_tokens": [ + -0.2488306611776352 + ], + "logprob_reward": -1.6238306760787964, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.6238306760787964, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.26359835267066956, + 0.23311780393123627, + 0.03238758072257042, + 0.04755142331123352, + 0.02918376214802265, + 0.042100779712200165, + 0.10438110679388046, + 0.08218565583229065, + 0.028788894414901733, + 0.136704683303833 + ], + "credit_map": [ + 0.022930536419153214, + 0.03267514705657959, + 0.03730520233511925, + 0.022821081802248955, + 0.0547800287604332, + 0.043230846524238586, + 0.019716929644346237, + 0.03013860061764717, + 0.027365418151021004, + 0.022806843742728233, + 0.04640448838472366, + 0.039788953959941864, + 0.02736993506550789, + 0.023848408833146095, + 0.02365974150598049, + 0.021874014288187027, + 0.003408556105569005, + 0.0029202026780694723, + 0.00769354123622179, + 0.006335032172501087, + 0.0055567240342497826, + 0.002013334073126316, + 0.002820318564772606, + 0.0016398740699514747, + 0.0031481320038437843, + 0.004922198131680489, + 0.009447349235415459, + 0.003367823548614979, + 0.011671819724142551, + 0.0038026380352675915, + 0.003709284821525216, + 0.007482178974896669, + 0.001890838029794395, + 0.003260574070736766, + 0.004430573899298906, + 0.00639604264870286, + 0.004436921328306198, + 0.0020461175590753555, + 0.004998252261430025, + 0.0017244394402951002, + 0.0035530489403754473, + 0.006671121343970299, + 0.006226952653378248, + 0.006304859183728695, + 0.0045361327938735485, + 0.004392087925225496, + 0.005077472887933254, + 0.005339102353900671, + 0.008495391346514225, + 0.009922905825078487, + 0.012914731167256832, + 0.010397651232779026, + 0.015215952880680561, + 0.012894093990325928, + 0.018749454990029335, + 0.015790920704603195, + 0.00982417818158865, + 0.009268327616155148, + 0.01462172158062458, + 0.00861577782779932, + 0.007960994727909565, + 0.008278586901724339, + 0.011128270998597145, + 0.012487801723182201, + 0.0026352866552770138, + 0.0023870838340371847, + 0.006298106163740158, + 0.0030152443796396255, + 0.005489159841090441, + 0.0023452932946383953, + 0.0037650945596396923, + 0.002853624988347292, + 0.011829756200313568, + 0.014736229553818703, + 0.03858475014567375, + 0.02046249434351921, + 0.014926089905202389, + 0.011586503125727177, + 0.011503973975777626, + 0.013074883259832859 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -1.6238306760787964 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed overhead bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elderly man wearing a pale light-blue striped hospital gown with one undone button?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by a warm yellow old-fashioned lamp casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a violent coughing action with heaving shoulders and a rising/falling blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a gloomy oil painting texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07905707508325577 + ], + "opposite_logprob_tokens": [ + -2.579056978225708 + ], + "logprob_reward": -0.07905707508325577, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07905707508325577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18715840578079224, + 0.16809740662574768, + 0.08940871804952621, + 0.10261444002389908, + 0.019701674580574036, + 0.09623599797487259, + 0.07374908775091171, + 0.08882789313793182, + 0.0489724762737751, + 0.12523393332958221 + ], + "credit_map": [ + 0.02923566661775112, + 0.015955036506056786, + 0.0362437404692173, + 0.02420101687312126, + 0.02129787765443325, + 0.024168170988559723, + 0.01489999983459711, + 0.021156907081604004, + 0.014154982753098011, + 0.016324389725923538, + 0.02837638556957245, + 0.023438401520252228, + 0.034560445696115494, + 0.016359996050596237, + 0.0163241159170866, + 0.01855868473649025, + 0.008415968157351017, + 0.009885800071060658, + 0.013322587125003338, + 0.010720368474721909, + 0.02433631755411625, + 0.007524682674556971, + 0.00873969029635191, + 0.0064633130095899105, + 0.010743762366473675, + 0.007491645403206348, + 0.02427455224096775, + 0.010282160714268684, + 0.018356189131736755, + 0.007885864935815334, + 0.010189776308834553, + 0.013390501961112022, + 0.0017891050083562732, + 0.0021201493218541145, + 0.004623753018677235, + 0.0036077562253922224, + 0.0024204084184020758, + 0.0014429243747144938, + 0.0023142672143876553, + 0.0013833108823746443, + 0.008736135438084602, + 0.010521798394620419, + 0.016337081789970398, + 0.01506138127297163, + 0.00842681247740984, + 0.008841229602694511, + 0.012422407045960426, + 0.015889154747128487, + 0.006917062681168318, + 0.00723619107156992, + 0.008136561140418053, + 0.007360703777521849, + 0.009964984841644764, + 0.008385094814002514, + 0.0132438438013196, + 0.012504654005169868, + 0.006648165639489889, + 0.009076490066945553, + 0.011356155388057232, + 0.013284390792250633, + 0.012066084891557693, + 0.007087122183293104, + 0.01877943053841591, + 0.010530059225857258, + 0.005306233651936054, + 0.003945653326809406, + 0.011910537257790565, + 0.005768453702330589, + 0.005446610972285271, + 0.004825402982532978, + 0.005215580575168133, + 0.006554008461534977, + 0.011453329585492611, + 0.015076725743710995, + 0.036812834441661835, + 0.019302982836961746, + 0.013227459974586964, + 0.009640135802328587, + 0.009185188449919224, + 0.010535284876823425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.07905707508325577 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed overhead bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elderly man wearing a pale light-blue striped hospital gown with one undone button?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by a warm yellow old-fashioned lamp casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a violent coughing action with heaving shoulders and a rising/falling blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a gloomy oil painting texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11299964040517807 + ], + "opposite_logprob_tokens": [ + -2.237999677658081 + ], + "logprob_reward": -0.11299964040517807, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11299964040517807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1970827281475067, + 0.22201845049858093, + 0.09248832613229752, + 0.06866715848445892, + 0.0663541853427887, + 0.1916375458240509, + 0.03521183133125305, + 0.03621377423405647, + 0.04600619152188301, + 0.04431978613138199 + ], + "credit_map": [ + 0.024149717763066292, + 0.019323060289025307, + 0.047027502208948135, + 0.022317973896861076, + 0.023607753217220306, + 0.020890023559331894, + 0.015459371730685234, + 0.02430734597146511, + 0.0208085048943758, + 0.018448490649461746, + 0.06186024472117424, + 0.050740133970975876, + 0.02394997514784336, + 0.015303891152143478, + 0.016665618866682053, + 0.014241605065762997, + 0.010510606691241264, + 0.009815525263547897, + 0.010919347405433655, + 0.015197432599961758, + 0.017669586464762688, + 0.008329997770488262, + 0.012114552780985832, + 0.007931290194392204, + 0.006853173486888409, + 0.005371183156967163, + 0.018404342234134674, + 0.004482842516154051, + 0.0067631253041327, + 0.007915789261460304, + 0.00516471965238452, + 0.013711987063288689, + 0.007197736296802759, + 0.0069185891188681126, + 0.012376242317259312, + 0.014545779675245285, + 0.009384033270180225, + 0.0039422595873475075, + 0.008334535174071789, + 0.003655017586424947, + 0.03409216180443764, + 0.018703347072005272, + 0.03052130527794361, + 0.02347518689930439, + 0.02108648233115673, + 0.02401266060769558, + 0.019361544400453568, + 0.020384876057505608, + 0.0036794233601540327, + 0.0032380218617618084, + 0.005219134967774153, + 0.007170418743044138, + 0.0036104528699070215, + 0.0025430687237530947, + 0.005459646228700876, + 0.004291667137295008, + 0.0040051196701824665, + 0.004364228807389736, + 0.004582575988024473, + 0.005321062169969082, + 0.0040342495776712894, + 0.003747600596398115, + 0.004190429113805294, + 0.005968512035906315, + 0.004734819754958153, + 0.00435435539111495, + 0.009022120386362076, + 0.006151462905108929, + 0.0061010075733065605, + 0.0034488141536712646, + 0.006575009319931269, + 0.005618607625365257, + 0.0035357223823666573, + 0.004757342394441366, + 0.012306625954806805, + 0.00716241542249918, + 0.004111199639737606, + 0.003170890035107732, + 0.003826918313279748, + 0.005448676645755768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.11299964040517807 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed overhead bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elderly man wearing a pale light-blue striped hospital gown with one undone button?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by a warm yellow old-fashioned lamp casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a violent coughing action with heaving shoulders and a rising/falling blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a gloomy oil painting texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4743555188179016 + ], + "opposite_logprob_tokens": [ + -0.9743555188179016 + ], + "logprob_reward": -0.4743555188179016, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4743555188179016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10393641889095306, + 0.11758877336978912, + 0.09073886275291443, + 0.04680366814136505, + 0.032034993171691895, + 0.05720687285065651, + 0.07548690587282181, + 0.048332579433918, + 0.28177252411842346, + 0.14609843492507935 + ], + "credit_map": [ + 0.011553955264389515, + 0.011244569905102253, + 0.013172617182135582, + 0.011648832820355892, + 0.018488626927137375, + 0.014823364093899727, + 0.010949925519526005, + 0.012054519727826118, + 0.01266741193830967, + 0.011379608884453773, + 0.03426378592848778, + 0.02260868065059185, + 0.012159842066466808, + 0.006766428705304861, + 0.008953663520514965, + 0.008789343759417534, + 0.009067745879292488, + 0.010636596940457821, + 0.014406685717403889, + 0.016887880861759186, + 0.01627974584698677, + 0.00775938481092453, + 0.008314560167491436, + 0.007386261131614447, + 0.003642265684902668, + 0.004349674563854933, + 0.005917748436331749, + 0.005418851971626282, + 0.007464045193046331, + 0.004891238175332546, + 0.006655809935182333, + 0.008464038372039795, + 0.0024631142150610685, + 0.004701637662947178, + 0.003997195977717638, + 0.005999204237014055, + 0.0038349474780261517, + 0.004180904943495989, + 0.004052218981087208, + 0.0028057703748345375, + 0.005612215492874384, + 0.005411303602159023, + 0.009065733291208744, + 0.009711910970509052, + 0.005284812301397324, + 0.005427319556474686, + 0.0072807976976037025, + 0.009412781335413456, + 0.006044586189091206, + 0.009193969890475273, + 0.009324606508016586, + 0.008312004618346691, + 0.014268221333622932, + 0.0062013124115765095, + 0.010367793962359428, + 0.011774406768381596, + 0.006486475933343172, + 0.005811687558889389, + 0.004432088229805231, + 0.006759924348443747, + 0.007201268337666988, + 0.004730310756713152, + 0.006759254727512598, + 0.006151567678898573, + 0.027125973254442215, + 0.022068733349442482, + 0.06233132630586624, + 0.028987083584070206, + 0.03914506733417511, + 0.021346284076571465, + 0.04298049584031105, + 0.037787552922964096, + 0.014398992992937565, + 0.011996213346719742, + 0.03809487819671631, + 0.024140615016222, + 0.01547520887106657, + 0.011866453103721142, + 0.015420847572386265, + 0.014705224893987179 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 1 + ], + "selected_reward": -0.4743555188179016 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed overhead bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elderly man wearing a pale light-blue striped hospital gown with one undone button?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by a warm yellow old-fashioned lamp casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a violent coughing action with heaving shoulders and a rising/falling blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a gloomy oil painting texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906683325767517 + ], + "opposite_logprob_tokens": [ + -2.464066743850708 + ], + "logprob_reward": -0.08906683325767517, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906683325767517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1077519878745079, + 0.1856059730052948, + 0.05507681146264076, + 0.05706629157066345, + 0.05734243616461754, + 0.05840540677309036, + 0.1419471949338913, + 0.07741262763738632, + 0.13643351197242737, + 0.12295781821012497 + ], + "credit_map": [ + 0.010154607705771923, + 0.012584539130330086, + 0.013539732433855534, + 0.010834257118403912, + 0.014450118876993656, + 0.020685333758592606, + 0.01304155495017767, + 0.012461841106414795, + 0.024994760751724243, + 0.014351822435855865, + 0.05189700052142143, + 0.034940265119075775, + 0.018821697682142258, + 0.010909431613981724, + 0.01529700681567192, + 0.01439400389790535, + 0.004262225702404976, + 0.006597435567528009, + 0.007488789968192577, + 0.011042522266507149, + 0.009754505008459091, + 0.004991624038666487, + 0.006603775080293417, + 0.004335931967943907, + 0.005159107502549887, + 0.0044040372595191, + 0.01629350334405899, + 0.004394532646983862, + 0.008414586074650288, + 0.004538221750408411, + 0.003924572840332985, + 0.009937729686498642, + 0.0052777440287172794, + 0.0050208610482513905, + 0.011923720128834248, + 0.014072451740503311, + 0.008078587241470814, + 0.003148773917928338, + 0.00650424649938941, + 0.00331605295650661, + 0.004616956226527691, + 0.007597061805427074, + 0.010499897412955761, + 0.008822636678814888, + 0.006310417782515287, + 0.004895740654319525, + 0.007824852131307125, + 0.007837846875190735, + 0.012144125998020172, + 0.010571743361651897, + 0.016969431191682816, + 0.014989851973950863, + 0.020919889211654663, + 0.014641883783042431, + 0.024460475891828537, + 0.027249781414866447, + 0.008076313883066177, + 0.008921160362660885, + 0.013354972004890442, + 0.00876855943351984, + 0.008098098449409008, + 0.008298483677208424, + 0.010384420864284039, + 0.011510612443089485, + 0.015251905657351017, + 0.013994021341204643, + 0.026618001982569695, + 0.01770319603383541, + 0.015103394165635109, + 0.00897795706987381, + 0.022339049726724625, + 0.01644597388803959, + 0.008754263631999493, + 0.01663133129477501, + 0.02292964607477188, + 0.016224650666117668, + 0.01882023550570011, + 0.015346619300544262, + 0.013278604485094547, + 0.010972476564347744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 8 + ], + "selected_reward": -0.08906683325767517 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fixed, overhead bird's-eye view, with a near-Godlike perspective, calmly observes an elderly man with a sickly face. He lies on his back on a bed covered in slightly wrinkled off-white cotton sheets. His sparse, graying hair is matted to a sweat-soaked forehead, deep wrinkles knotted tightly together with pain. He wears a pale, light-blue striped hospital gown, one button undone at the collar, revealing the shrunken skin below his collarbone. An old-fashioned lamp on the bedside table emits the only warm yellow light, casting long, soft shadows in the dim room, accentuating his sunken eye sockets and cheekbones. Suddenly, a violent cough seizes him. He struggles to tilt his head to one side, his body instinctively hunching over. His thin shoulders heave violently and spasmodically with each deep, hoarse cough, and the dark gray blanket covering his chest rises and falls with it. This painful process is slow and difficult, and the entire picture presents a gloomy oil painting texture under the dim light, creating a heartbreaking atmosphere of tranquility and depression as a life struggles in illness.", + "seed": 31485, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_2/index_2666_seed_31485_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.3628164786845446, + "vlm_logprob": -0.3628164786845446, + "vlm_yes_no_margin": 1.95, + "vlm_token_credit": -0.3628164786845446, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed overhead bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elderly man wearing a pale light-blue striped hospital gown with one undone button?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by a warm yellow old-fashioned lamp casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a violent coughing action with heaving shoulders and a rising/falling blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a gloomy oil painting texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4267221689224243 + ], + "opposite_logprob_tokens": [ + -0.3017221689224243 + ], + "logprob_reward": -1.4267221689224243, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4267221689224243, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.21373522281646729, + 0.03813624754548073, + 0.08597491681575775, + 0.04384628310799599, + 0.24298381805419922, + 0.027775580063462257, + 0.11243534088134766, + 0.15508486330509186, + 0.030308173969388008, + 0.049719568341970444 + ], + "credit_map": [ + 0.02705339901149273, + 0.048978012055158615, + 0.03127533197402954, + 0.035444896668195724, + 0.018908686935901642, + 0.015615626238286495, + 0.021211622282862663, + 0.015247632749378681, + 0.00355329317972064, + 0.008428128436207771, + 0.00429559824988246, + 0.005952123552560806, + 0.004166612401604652, + 0.0026864157989621162, + 0.006642437539994717, + 0.0024116358254104853, + 0.010298763401806355, + 0.02721598744392395, + 0.01284380629658699, + 0.01564779132604599, + 0.006825389806181192, + 0.003504850436002016, + 0.0062730335630476475, + 0.003365288022905588, + 0.004393380600959063, + 0.00959012284874916, + 0.005734554957598448, + 0.009016810916364193, + 0.003320215968415141, + 0.004789593163877726, + 0.004570687189698219, + 0.0024309144355356693, + 0.02375822514295578, + 0.07586484402418137, + 0.032647326588630676, + 0.044220175594091415, + 0.025555580854415894, + 0.007740973494946957, + 0.021669447422027588, + 0.011527221649885178, + 0.0030989369843155146, + 0.0055997190065681934, + 0.005638370756059885, + 0.003532527247443795, + 0.0020295563153922558, + 0.002864073496311903, + 0.003223532810807228, + 0.0017888641450554132, + 0.011234774254262447, + 0.032744210213422775, + 0.01496552862226963, + 0.020167944952845573, + 0.009621928445994854, + 0.005616571754217148, + 0.013549398630857468, + 0.004534977488219738, + 0.019096270203590393, + 0.028856201097369194, + 0.020653389394283295, + 0.026352284476161003, + 0.01845667138695717, + 0.009381343610584736, + 0.020780133083462715, + 0.011508554220199585, + 0.0037353301886469126, + 0.004785856697708368, + 0.004869828931987286, + 0.004975078161805868, + 0.0031094872392714024, + 0.002419018652290106, + 0.004390105605125427, + 0.002023466629907489, + 0.004542735870927572, + 0.010438657365739346, + 0.0050408183597028255, + 0.011072064749896526, + 0.004701434168964624, + 0.002949988469481468, + 0.008310656994581223, + 0.002663208404555917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -1.4267221689224243 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed overhead bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elderly man wearing a pale light-blue striped hospital gown with one undone button?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by a warm yellow old-fashioned lamp casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a violent coughing action with heaving shoulders and a rising/falling blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a gloomy oil painting texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029846109449863434 + ], + "opposite_logprob_tokens": [ + -3.52984619140625 + ], + "logprob_reward": -0.029846109449863434, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029846109449863434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16800715029239655, + 0.14693613350391388, + 0.09817332774400711, + 0.05134059116244316, + 0.1576819121837616, + 0.05615059286355972, + 0.0451839417219162, + 0.14038971066474915, + 0.08837500214576721, + 0.04776158183813095 + ], + "credit_map": [ + 0.018429234623908997, + 0.04348240792751312, + 0.022100156173110008, + 0.029800089076161385, + 0.013480599969625473, + 0.010450739413499832, + 0.019225897267460823, + 0.011038043536245823, + 0.012295608408749104, + 0.04417400807142258, + 0.019666269421577454, + 0.02393566258251667, + 0.01919158361852169, + 0.006264066323637962, + 0.016258906573057175, + 0.00515004713088274, + 0.007572692818939686, + 0.0146027822047472, + 0.01374773308634758, + 0.020762808620929718, + 0.011184931732714176, + 0.007481161970645189, + 0.01679116301238537, + 0.006030072458088398, + 0.005311166401952505, + 0.010394638404250145, + 0.006880265660583973, + 0.009696374647319317, + 0.0031749636400491, + 0.00829792208969593, + 0.0034575408790260553, + 0.004127729218453169, + 0.015264703892171383, + 0.05091142654418945, + 0.02106776088476181, + 0.028837233781814575, + 0.015244563110172749, + 0.0055328854359686375, + 0.014096928760409355, + 0.0067264242097735405, + 0.005590404383838177, + 0.010044089518487453, + 0.009454406797885895, + 0.010760720819234848, + 0.004530009347945452, + 0.00452681677415967, + 0.007423970382660627, + 0.0038201792631298304, + 0.00462790671736002, + 0.006291513796895742, + 0.00703597255051136, + 0.007812554016709328, + 0.0037480820901691914, + 0.00490910280495882, + 0.0070415702648460865, + 0.003717245766893029, + 0.019622959196567535, + 0.023350896313786507, + 0.02393924444913864, + 0.021062206476926804, + 0.0154371727257967, + 0.007237257435917854, + 0.01955747790634632, + 0.010182511061429977, + 0.010040828958153725, + 0.009255812503397465, + 0.014120751060545444, + 0.014583704993128777, + 0.010890894569456577, + 0.0070777470245957375, + 0.0166199691593647, + 0.0057853031903505325, + 0.003960337955504656, + 0.010017133317887783, + 0.005206631030887365, + 0.007437714841216803, + 0.004973898641765118, + 0.0027825715951621532, + 0.007720538880676031, + 0.005662758834660053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.029846109449863434 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed overhead bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elderly man wearing a pale light-blue striped hospital gown with one undone button?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by a warm yellow old-fashioned lamp casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a violent coughing action with heaving shoulders and a rising/falling blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a gloomy oil painting texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026367442682385445 + ], + "opposite_logprob_tokens": [ + -3.651367425918579 + ], + "logprob_reward": -0.026367442682385445, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026367442682385445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18487033247947693, + 0.10156535357236862, + 0.21714292466640472, + 0.05648535490036011, + 0.16299836337566376, + 0.1047077625989914, + 0.02428957261145115, + 0.0675574466586113, + 0.05011165514588356, + 0.03027118556201458 + ], + "credit_map": [ + 0.022849055007100105, + 0.04778635874390602, + 0.024623751640319824, + 0.035200417041778564, + 0.014313573017716408, + 0.011417304165661335, + 0.020050862804055214, + 0.008629034273326397, + 0.010998489335179329, + 0.01416304986923933, + 0.012065136805176735, + 0.015066132880747318, + 0.010521994903683662, + 0.013206246308982372, + 0.017388178035616875, + 0.008156132884323597, + 0.021101685240864754, + 0.03233768418431282, + 0.031332604587078094, + 0.025540247559547424, + 0.018016371876001358, + 0.040951523929834366, + 0.023791184648871422, + 0.02407163567841053, + 0.006410940084606409, + 0.010849629528820515, + 0.007511454168707132, + 0.008492616005241871, + 0.0038124960847198963, + 0.009305844083428383, + 0.0050890701822936535, + 0.0050133089534938335, + 0.01677495241165161, + 0.04987744614481926, + 0.021995658054947853, + 0.028752921149134636, + 0.017314748838543892, + 0.00572598772123456, + 0.014791279099881649, + 0.007765376940369606, + 0.00810480397194624, + 0.014752632938325405, + 0.022421006113290787, + 0.011622385121881962, + 0.007831758819520473, + 0.01949228346347809, + 0.010574744082987309, + 0.009908157400786877, + 0.0022206956055015326, + 0.006063453387469053, + 0.0034241885878145695, + 0.004414859693497419, + 0.002044558059424162, + 0.0016279022675007582, + 0.003064435441046953, + 0.0014294834109023213, + 0.007360558491200209, + 0.013566330075263977, + 0.007713534403592348, + 0.010923691093921661, + 0.00814330019056797, + 0.003650538856163621, + 0.010358274914324284, + 0.0058412267826497555, + 0.006376820616424084, + 0.00767012033611536, + 0.008009436540305614, + 0.006721792742609978, + 0.005776169244199991, + 0.0061507211066782475, + 0.005885932594537735, + 0.00352066894993186, + 0.0031051100231707096, + 0.00684930756688118, + 0.002964566694572568, + 0.005298716016113758, + 0.0026183882728219032, + 0.00206653936766088, + 0.005238743964582682, + 0.002129816450178623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.026367442682385445 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed overhead bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elderly man wearing a pale light-blue striped hospital gown with one undone button?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by a warm yellow old-fashioned lamp casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a violent coughing action with heaving shoulders and a rising/falling blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a gloomy oil painting texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521490752696991 + ], + "opposite_logprob_tokens": [ + -1.5021491050720215 + ], + "logprob_reward": -0.2521490752696991, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521490752696991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09625913202762604, + 0.05786961689591408, + 0.045603927224874496, + 0.07682722806930542, + 0.09864448010921478, + 0.03514386713504791, + 0.12257414311170578, + 0.2123533934354782, + 0.13637717068195343, + 0.11834696680307388 + ], + "credit_map": [ + 0.011780057102441788, + 0.021805815398693085, + 0.012535428628325462, + 0.016351331025362015, + 0.008752675727009773, + 0.0077189733274281025, + 0.009883731603622437, + 0.007431119680404663, + 0.00529737863689661, + 0.014377111569046974, + 0.00823540985584259, + 0.010226164944469929, + 0.005476993042975664, + 0.002838226966559887, + 0.008790765888988972, + 0.002627566223964095, + 0.0044403704814612865, + 0.009494701400399208, + 0.008433961309492588, + 0.008866768330335617, + 0.0025794850662350655, + 0.002993501489982009, + 0.005532828625291586, + 0.003262309590354562, + 0.014219156466424465, + 0.012302216142416, + 0.009977667592465878, + 0.01078052632510662, + 0.006958612706512213, + 0.009082036092877388, + 0.007916457019746304, + 0.0055905599147081375, + 0.009220249950885773, + 0.028128471225500107, + 0.01323295384645462, + 0.020441092550754547, + 0.0074703749269247055, + 0.006165741011500359, + 0.009144280105829239, + 0.004841317422688007, + 0.00464343698695302, + 0.006762462668120861, + 0.004745629616081715, + 0.00417344132438302, + 0.003930138424038887, + 0.0032769357785582542, + 0.004140443168580532, + 0.0034713796339929104, + 0.014355628751218319, + 0.028260717168450356, + 0.016000336036086082, + 0.023819521069526672, + 0.010823756456375122, + 0.005906176753342152, + 0.01846269704401493, + 0.0049453130923211575, + 0.02883175201714039, + 0.0382138155400753, + 0.04188736900687218, + 0.026416298002004623, + 0.02113158442080021, + 0.014019745402038097, + 0.02407388761639595, + 0.01777895726263523, + 0.02048313245177269, + 0.026677289977669716, + 0.02301313914358616, + 0.025793934240937233, + 0.01103332918137312, + 0.008056879974901676, + 0.014711679890751839, + 0.006607796065509319, + 0.010749859735369682, + 0.03557095676660538, + 0.013633041642606258, + 0.021031754091382027, + 0.01061251014471054, + 0.005316315684467554, + 0.01406053639948368, + 0.007371997460722923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.2521490752696991 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed overhead bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elderly man wearing a pale light-blue striped hospital gown with one undone button?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by a warm yellow old-fashioned lamp casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a violent coughing action with heaving shoulders and a rising/falling blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a gloomy oil painting texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899759709835052 + ], + "opposite_logprob_tokens": [ + -2.5789976119995117 + ], + "logprob_reward": -0.07899759709835052, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899759709835052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21050484478473663, + 0.1407039612531662, + 0.11456151306629181, + 0.04317677021026611, + 0.05378029868006706, + 0.06302547454833984, + 0.053384799510240555, + 0.11345230787992477, + 0.09497180581092834, + 0.1124381572008133 + ], + "credit_map": [ + 0.022574985399842262, + 0.04825318604707718, + 0.02649780735373497, + 0.03899519890546799, + 0.01950501650571823, + 0.012036332860589027, + 0.03330284729599953, + 0.009339481592178345, + 0.012646503746509552, + 0.029478007927536964, + 0.01838882826268673, + 0.017010876908898354, + 0.015433091670274734, + 0.019943388178944588, + 0.01640024408698082, + 0.011403029784560204, + 0.010258941911160946, + 0.021049633622169495, + 0.01707289181649685, + 0.01652476005256176, + 0.01058690994977951, + 0.012404130771756172, + 0.014571740292012691, + 0.012092508375644684, + 0.0048269787803292274, + 0.009711396880447865, + 0.005594510585069656, + 0.00910794548690319, + 0.003735882695764303, + 0.002540412824600935, + 0.005601482931524515, + 0.0020581611897796392, + 0.005132313817739487, + 0.013371695764362812, + 0.006479725707322359, + 0.010646057315170765, + 0.005968938581645489, + 0.0030196961015462875, + 0.006771629210561514, + 0.002390240551903844, + 0.006335098296403885, + 0.011453851126134396, + 0.008863620460033417, + 0.010484553873538971, + 0.006817718502134085, + 0.006473645567893982, + 0.008352785371243954, + 0.004244204144924879, + 0.005343486089259386, + 0.014465359970927238, + 0.007431599777191877, + 0.008642690256237984, + 0.005357637535780668, + 0.0032032025046646595, + 0.00562273059040308, + 0.0033180888276547194, + 0.014140346087515354, + 0.024199990555644035, + 0.021575259044766426, + 0.01571340672671795, + 0.01049437839537859, + 0.005525131244212389, + 0.012694769538939, + 0.009109029546380043, + 0.019762307405471802, + 0.024278409779071808, + 0.012452623806893826, + 0.011639186181128025, + 0.005760133732110262, + 0.007264412473887205, + 0.008447141386568546, + 0.0053675915114581585, + 0.011508133262395859, + 0.031861986964941025, + 0.012881254777312279, + 0.017783712595701218, + 0.012591242790222168, + 0.005223721265792847, + 0.013544746674597263, + 0.007043363992124796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.07899759709835052 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fixed, overhead bird's-eye view, with a near-Godlike perspective, calmly observes an elderly man with a sickly face. He lies on his back on a bed covered in slightly wrinkled off-white cotton sheets. His sparse, graying hair is matted to a sweat-soaked forehead, deep wrinkles knotted tightly together with pain. He wears a pale, light-blue striped hospital gown, one button undone at the collar, revealing the shrunken skin below his collarbone. An old-fashioned lamp on the bedside table emits the only warm yellow light, casting long, soft shadows in the dim room, accentuating his sunken eye sockets and cheekbones. Suddenly, a violent cough seizes him. He struggles to tilt his head to one side, his body instinctively hunching over. His thin shoulders heave violently and spasmodically with each deep, hoarse cough, and the dark gray blanket covering his chest rises and falls with it. This painful process is slow and difficult, and the entire picture presents a gloomy oil painting texture under the dim light, creating a heartbreaking atmosphere of tranquility and depression as a life struggles in illness.", + "seed": 31486, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_2/index_2666_seed_31486_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.3566759143024683, + "vlm_logprob": -0.3566759143024683, + "vlm_yes_no_margin": 2.0, + "vlm_token_credit": -0.3566759143024683, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed overhead bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elderly man wearing a pale light-blue striped hospital gown with one undone button?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by a warm yellow old-fashioned lamp casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a violent coughing action with heaving shoulders and a rising/falling blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a gloomy oil painting texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.000037431716919 + ], + "opposite_logprob_tokens": [ + -0.500037431716919 + ], + "logprob_reward": -1.000037431716919, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -1.000037431716919, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.24912579357624054, + 0.08828651905059814, + 0.09423364698886871, + 0.04249020293354988, + 0.06595506519079208, + 0.041175976395606995, + 0.20394855737686157, + 0.05166449397802353, + 0.08950860053300858, + 0.07361113280057907 + ], + "credit_map": [ + 0.01766226813197136, + 0.03249998018145561, + 0.03912283852696419, + 0.03792896121740341, + 0.06408904492855072, + 0.02132641337811947, + 0.01754332147538662, + 0.01895299181342125, + 0.0062536573968827724, + 0.010350611060857773, + 0.013416104950010777, + 0.01773640885949135, + 0.014671796932816505, + 0.009857283905148506, + 0.007403564639389515, + 0.008597096428275108, + 0.005445703398436308, + 0.015418047085404396, + 0.014059762470424175, + 0.012802544049918652, + 0.01711905375123024, + 0.005165602080523968, + 0.012803653255105019, + 0.011419280432164669, + 0.002946557244285941, + 0.006828992627561092, + 0.00708848936483264, + 0.010835839435458183, + 0.006894151214510202, + 0.00209426018409431, + 0.002968721091747284, + 0.0028331950306892395, + 0.005015373229980469, + 0.006745511200278997, + 0.010313062928617, + 0.012811682187020779, + 0.009775289334356785, + 0.0032043408136814833, + 0.008476166054606438, + 0.009613639675080776, + 0.002737835980951786, + 0.005923904944211245, + 0.004803450778126717, + 0.006937203463166952, + 0.00629214383661747, + 0.002414730377495289, + 0.00718563050031662, + 0.004881083499640226, + 0.015833133831620216, + 0.03849788382649422, + 0.02706504985690117, + 0.04538325220346451, + 0.024760883301496506, + 0.008698300458490849, + 0.024333201348781586, + 0.01937686651945114, + 0.00386344688013196, + 0.005333447828888893, + 0.008371631614863873, + 0.01421311590820551, + 0.006122885271906853, + 0.0031880210153758526, + 0.0060692476108670235, + 0.004502701107412577, + 0.006799515802413225, + 0.013172016479074955, + 0.01165910717099905, + 0.01600138656795025, + 0.012300348840653896, + 0.0034120373893529177, + 0.01389277447015047, + 0.012271421030163765, + 0.007011808454990387, + 0.01329409796744585, + 0.009597696363925934, + 0.009800698608160019, + 0.011384177021682262, + 0.007458401843905449, + 0.007894082926213741, + 0.0071701668202877045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -1.000037431716919 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed overhead bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elderly man wearing a pale light-blue striped hospital gown with one undone button?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by a warm yellow old-fashioned lamp casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a violent coughing action with heaving shoulders and a rising/falling blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a gloomy oil painting texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029844488948583603 + ], + "opposite_logprob_tokens": [ + -3.5298445224761963 + ], + "logprob_reward": -0.029844488948583603, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029844488948583603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1755833774805069, + 0.03791398927569389, + 0.055308982729911804, + 0.04896114766597748, + 0.15008756518363953, + 0.0970870703458786, + 0.050086554139852524, + 0.16957242786884308, + 0.12944592535495758, + 0.0859530046582222 + ], + "credit_map": [ + 0.010148273780941963, + 0.021033523604273796, + 0.025610139593482018, + 0.027749737724661827, + 0.03556255251169205, + 0.019367486238479614, + 0.01677642948925495, + 0.01933521218597889, + 0.00263647991232574, + 0.0039027268067002296, + 0.005383566953241825, + 0.00490497425198555, + 0.006264481693506241, + 0.004376139026135206, + 0.005230216775089502, + 0.005215398967266083, + 0.003707207040861249, + 0.01229849737137556, + 0.007677169516682625, + 0.005252514965832233, + 0.01114729419350624, + 0.005089023150503635, + 0.0051175691187381744, + 0.005019703414291143, + 0.0034521035850048065, + 0.003997117280960083, + 0.007252092473208904, + 0.009672434069216251, + 0.005287021864205599, + 0.006721746176481247, + 0.005818773992359638, + 0.00675985449925065, + 0.008390603587031364, + 0.01953580416738987, + 0.022893037647008896, + 0.020245278254151344, + 0.039843544363975525, + 0.010699625127017498, + 0.014227492734789848, + 0.014252173714339733, + 0.007793063297867775, + 0.019813602790236473, + 0.01152730081230402, + 0.01846352033317089, + 0.013601916842162609, + 0.0053347451612353325, + 0.01092490553855896, + 0.00962800346314907, + 0.003320968011394143, + 0.009851794689893723, + 0.007123397663235664, + 0.011010967195034027, + 0.004711777437478304, + 0.0036500743590295315, + 0.005137775093317032, + 0.0052797989919781685, + 0.013721868395805359, + 0.020456403493881226, + 0.023682691156864166, + 0.03860737010836601, + 0.026056459173560143, + 0.012262696400284767, + 0.01708780601620674, + 0.017697112634778023, + 0.009331604465842247, + 0.022271983325481415, + 0.018872536718845367, + 0.015439670532941818, + 0.016119103878736496, + 0.014847229234874249, + 0.01772809959948063, + 0.014835678040981293, + 0.006190090905874968, + 0.01115344651043415, + 0.012197231873869896, + 0.023876899853348732, + 0.0101554486900568, + 0.005349710118025541, + 0.010296176187694073, + 0.006733989343047142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.029844488948583603 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed overhead bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elderly man wearing a pale light-blue striped hospital gown with one undone button?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by a warm yellow old-fashioned lamp casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a violent coughing action with heaving shoulders and a rising/falling blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a gloomy oil painting texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026363959535956383 + ], + "opposite_logprob_tokens": [ + -3.6513638496398926 + ], + "logprob_reward": -0.026363959535956383, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026363959535956383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28560638427734375, + 0.03921445086598396, + 0.30850672721862793, + 0.07711109519004822, + 0.03670012205839157, + 0.05254094675183296, + 0.09733439981937408, + 0.029880613088607788, + 0.04207222908735275, + 0.031032970175147057 + ], + "credit_map": [ + 0.023172233253717422, + 0.0413186177611351, + 0.044577211141586304, + 0.041806526482105255, + 0.05478014424443245, + 0.028669334948062897, + 0.0225975401699543, + 0.02868478372693062, + 0.002810129662975669, + 0.003771441522985697, + 0.005348545499145985, + 0.006232949439436197, + 0.005834411364048719, + 0.0053654806688427925, + 0.004113693255931139, + 0.005737798288464546, + 0.02812500298023224, + 0.038163717836141586, + 0.046593308448791504, + 0.04529108107089996, + 0.050391219556331635, + 0.023591268807649612, + 0.03586363419890404, + 0.040487512946128845, + 0.004465281963348389, + 0.009812646545469761, + 0.011610761284828186, + 0.019713282585144043, + 0.014899841509759426, + 0.004737078677862883, + 0.005431400611996651, + 0.00644080713391304, + 0.002127480460330844, + 0.00484737753868103, + 0.005438338965177536, + 0.007116819266229868, + 0.006043968256562948, + 0.00303178234025836, + 0.00400493573397398, + 0.004089421592652798, + 0.004131725989282131, + 0.007329816929996014, + 0.006508903112262487, + 0.00689115934073925, + 0.009175935760140419, + 0.0033371057361364365, + 0.008175916038453579, + 0.0069903843104839325, + 0.006994516588747501, + 0.018163615837693214, + 0.013166962191462517, + 0.023368731141090393, + 0.011341935023665428, + 0.0039001351688057184, + 0.011542405933141708, + 0.008856099098920822, + 0.002746549202129245, + 0.004802686627954245, + 0.0040585314854979515, + 0.005869924556463957, + 0.004161997232586145, + 0.0023799550253897905, + 0.0029233512468636036, + 0.0029376160819083452, + 0.0031503685750067234, + 0.009221990592777729, + 0.005289559718221426, + 0.00599898723885417, + 0.005580716300755739, + 0.0025528459809720516, + 0.005203368607908487, + 0.005074389278888702, + 0.003468915121629834, + 0.0043351356871426105, + 0.004416617099195719, + 0.005070049781352282, + 0.005094749387353659, + 0.0017911905888468027, + 0.003340141847729683, + 0.003516170894727111 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.026363959535956383 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed overhead bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elderly man wearing a pale light-blue striped hospital gown with one undone button?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by a warm yellow old-fashioned lamp casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a violent coughing action with heaving shoulders and a rising/falling blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a gloomy oil painting texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6934211850166321 + ], + "opposite_logprob_tokens": [ + -0.6934211850166321 + ], + "logprob_reward": -0.6934211850166321, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6934211850166321, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1919124871492386, + 0.08372728526592255, + 0.07091758400201797, + 0.04637376591563225, + 0.11940446496009827, + 0.08042135834693909, + 0.09260796010494232, + 0.06597244739532471, + 0.15585824847221375, + 0.0928044468164444 + ], + "credit_map": [ + 0.013324744999408722, + 0.020754029974341393, + 0.03175241872668266, + 0.022311614826321602, + 0.04317650943994522, + 0.013124175369739532, + 0.02157587744295597, + 0.025893110781908035, + 0.004086013417690992, + 0.009370126761496067, + 0.01326612289994955, + 0.011438093148171902, + 0.018605230376124382, + 0.011000926606357098, + 0.007547183893620968, + 0.0084135876968503, + 0.007735048420727253, + 0.00909114908427, + 0.01010878849774599, + 0.009173495694994926, + 0.008889570832252502, + 0.007371148094534874, + 0.007408257573843002, + 0.011140130460262299, + 0.0038330364041030407, + 0.004611424636095762, + 0.007172349374741316, + 0.010050659999251366, + 0.006531719584017992, + 0.004867242649197578, + 0.004232526756823063, + 0.005074805114418268, + 0.008441072888672352, + 0.018552256748080254, + 0.017201444134116173, + 0.026263181120157242, + 0.0184094849973917, + 0.00455035176128149, + 0.01374155655503273, + 0.01224511954933405, + 0.005678166635334492, + 0.015296995639801025, + 0.009371771477162838, + 0.010308351367712021, + 0.01730598881840706, + 0.005028214305639267, + 0.009686584584414959, + 0.0077452827244997025, + 0.005702603608369827, + 0.013090928085148335, + 0.013455080799758434, + 0.018519306555390358, + 0.01119981613010168, + 0.008154517970979214, + 0.01059094350785017, + 0.011894759722054005, + 0.005178235005587339, + 0.011397214606404305, + 0.010086030699312687, + 0.017314447090029716, + 0.006108314264565706, + 0.00392012158408761, + 0.00718745356425643, + 0.004780631046742201, + 0.011308650486171246, + 0.03304189071059227, + 0.02069919928908348, + 0.020654872059822083, + 0.021833809092640877, + 0.00737996818497777, + 0.020471371710300446, + 0.020468493923544884, + 0.008487601764500141, + 0.01951710321009159, + 0.011895556934177876, + 0.013223482295870781, + 0.01445816271007061, + 0.010909110307693481, + 0.007655837573111057, + 0.006657588295638561 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.6934211850166321 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed overhead bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elderly man wearing a pale light-blue striped hospital gown with one undone button?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by a warm yellow old-fashioned lamp casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a violent coughing action with heaving shoulders and a rising/falling blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a gloomy oil painting texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371250629425049 + ], + "opposite_logprob_tokens": [ + -3.408712387084961 + ], + "logprob_reward": -0.03371250629425049, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371250629425049, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21311123669147491, + 0.1647852659225464, + 0.14369583129882812, + 0.02466636523604393, + 0.07670172303915024, + 0.053056519478559494, + 0.15617479383945465, + 0.04998357966542244, + 0.05661044269800186, + 0.06121425703167915 + ], + "credit_map": [ + 0.015368464402854443, + 0.022590935230255127, + 0.034745924174785614, + 0.04264014586806297, + 0.0414002425968647, + 0.018970182165503502, + 0.01896822266280651, + 0.018427114933729172, + 0.00993196852505207, + 0.01997801847755909, + 0.02644243836402893, + 0.029372207820415497, + 0.03856555372476578, + 0.011430141516029835, + 0.013618802651762962, + 0.015446135774254799, + 0.010179419070482254, + 0.0268907081335783, + 0.02088860794901848, + 0.019887885078787804, + 0.02644507773220539, + 0.009646246209740639, + 0.014225845225155354, + 0.01553204096853733, + 0.0018636890454217792, + 0.0029549289029091597, + 0.004055648576468229, + 0.006592245306819677, + 0.0036630788818001747, + 0.0011481355177238584, + 0.002238508081063628, + 0.0021501306910067797, + 0.005663864780217409, + 0.01242606621235609, + 0.010736402124166489, + 0.012839934788644314, + 0.013020284473896027, + 0.004554434679448605, + 0.008875427767634392, + 0.008585301227867603, + 0.003803790779784322, + 0.009742695838212967, + 0.005941695999354124, + 0.007131104823201895, + 0.009077050723135471, + 0.004385239444673061, + 0.007072825450450182, + 0.00590211758390069, + 0.011943185701966286, + 0.02786361426115036, + 0.020474886521697044, + 0.031875647604465485, + 0.02017248421907425, + 0.0075271171517670155, + 0.02026417851448059, + 0.016053684055805206, + 0.005348600912839174, + 0.006420486606657505, + 0.007178317289799452, + 0.014641806483268738, + 0.004479280207306147, + 0.0028795055113732815, + 0.004773966036736965, + 0.004261615686118603, + 0.0038936580531299114, + 0.007837611250579357, + 0.008354675956070423, + 0.006686460226774216, + 0.009628432802855968, + 0.00561421737074852, + 0.008591940626502037, + 0.006003447342664003, + 0.005770623683929443, + 0.010095023550093174, + 0.00833347998559475, + 0.008364790119230747, + 0.008312768302857876, + 0.004575864411890507, + 0.009710674174129963, + 0.006051034200936556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.03371250629425049 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "No", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_3/index_tensor([2821, 2821])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_3/index_tensor([2821, 2821])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..87282ed4868e61e32a99e175ab90d8d7edef3c33 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_3/index_tensor([2821, 2821])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Static, low-angle, wide-angle shots rendered in ultra-detailed 8K resolution capture the moment of catastrophic failure using the visual language of ultra-high-speed photography. The CN Tower's iconic, needle-like concrete structure stands at the center, silhouetted against a desolate, perfectly clear blue sky. Bright, unforgiving sunlight casts sharp, geometric shadows across its surface. As the main observation deck begins to collapse, the action begins in extreme slow motion; spiderwebs on the concrete facade split open instantly, before massive chunks weighing several tons violently break away, revealing twisted, dark steel and internal framing beneath. Shards of glass from the panoramic windows explode outward like crystalline mist, catching the sunlight as they begin their graceful yet terrifying descent. As the initial collapse unleashes a vast, billowing plume of fine gray dust that blooms outward from the rupture point, its particles individually visible against the blue background, every microsecond of the structural damage is rendered with incredible force. The scene is a surreal, silent devastation, a beautiful yet terrifying physical ballet where falling debris appears to be suspended in mid-air, creating an atmosphere of both awe-inspiring and clinical catastrophe.", + "seed": 24114, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_3/index_2821_seed_24114_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10603959967847913, + "vlm_logprob": -0.10603959967847913, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.10603959967847913, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the CN Tower's concrete structure located at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle with a wide-angle lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action of the observation deck collapsing and chunks breaking away rendered in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, unforgiving sunlight and sharp geometric shadows visible on the tower's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast, billowing plume of fine gray dust visible blooming outward from the rupture point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03500589728355408 + ], + "opposite_logprob_tokens": [ + -3.660005807876587 + ], + "logprob_reward": -0.03500589728355408, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03500589728355408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07139061391353607, + 0.05842410773038864, + 0.03155919909477234, + 0.3981545567512512, + 0.0672527328133583, + 0.07334349304437637, + 0.03906329348683357, + 0.060525305569171906, + 0.16638629138469696, + 0.033900365233421326 + ], + "credit_map": [ + 0.006685952190309763, + 0.0034663244150578976, + 0.02744317427277565, + 0.002783220959827304, + 0.004317165818065405, + 0.009774213656783104, + 0.0017983500147238374, + 0.015122216194868088, + 0.003537589218467474, + 0.004980659112334251, + 0.003960015252232552, + 0.00423457333818078, + 0.0053795683197677135, + 0.01856422983109951, + 0.0036395604256540537, + 0.014127913862466812, + 0.0020421564113348722, + 0.0033015781082212925, + 0.002825336065143347, + 0.0022493053693324327, + 0.002891535870730877, + 0.008314646780490875, + 0.0016217472730204463, + 0.008312893100082874, + 0.020822864025831223, + 0.010194290429353714, + 0.15661589801311493, + 0.012611576355993748, + 0.015838170424103737, + 0.16547076404094696, + 0.005983669310808182, + 0.010617364197969437, + 0.00820670835673809, + 0.011050364002585411, + 0.013220075517892838, + 0.006049858406186104, + 0.00658989604562521, + 0.007831561379134655, + 0.006861269939690828, + 0.00744299590587616, + 0.018992090597748756, + 0.007670875638723373, + 0.017567763105034828, + 0.0045872642658650875, + 0.00638708146288991, + 0.007030006032437086, + 0.005445278249680996, + 0.005663132760673761, + 0.005075917113572359, + 0.009143512696027756, + 0.004511093720793724, + 0.003301287069916725, + 0.0054837861098349094, + 0.0038563937414437532, + 0.0033748026471585035, + 0.004316502716392279, + 0.016215551644563675, + 0.006390016060322523, + 0.011932753957808018, + 0.003804141655564308, + 0.004682585131376982, + 0.007330648601055145, + 0.005230308510363102, + 0.004939299076795578, + 0.04349464550614357, + 0.01189008355140686, + 0.01893973909318447, + 0.012991934083402157, + 0.01844838075339794, + 0.03720372915267944, + 0.01267770305275917, + 0.010740076191723347, + 0.005163731053471565, + 0.0038503576070070267, + 0.003932597581297159, + 0.005569823551923037, + 0.003550645662471652, + 0.003073776373639703, + 0.006080462131649256, + 0.002678972203284502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 5 + ], + "selected_reward": -0.03500589728355408 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the CN Tower's concrete structure located at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle with a wide-angle lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action of the observation deck collapsing and chunks breaking away rendered in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, unforgiving sunlight and sharp geometric shadows visible on the tower's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast, billowing plume of fine gray dust visible blooming outward from the rupture point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07006238400936127 + ], + "opposite_logprob_tokens": [ + -2.6950623989105225 + ], + "logprob_reward": -0.07006238400936127, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07006238400936127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3404718339443207, + 0.04943307116627693, + 0.1397181749343872, + 0.06212783232331276, + 0.05338766798377037, + 0.0208116602152586, + 0.046646181493997574, + 0.2266560196876526, + 0.04286104068160057, + 0.01788647659122944 + ], + "credit_map": [ + 0.045721378177404404, + 0.014278535731136799, + 0.10994182527065277, + 0.011103362776339054, + 0.027136025950312614, + 0.07446068525314331, + 0.007326841354370117, + 0.0505031943321228, + 0.0030889343470335007, + 0.002637490862980485, + 0.011666503734886646, + 0.0019322660518810153, + 0.0043649617582559586, + 0.0188432764261961, + 0.0019586728885769844, + 0.00494096614420414, + 0.008150269277393818, + 0.004629738163203001, + 0.015079377219080925, + 0.0026408601552248, + 0.005283552687615156, + 0.05761466920375824, + 0.004348834045231342, + 0.041970886290073395, + 0.005946284160017967, + 0.004313088487833738, + 0.021156733855605125, + 0.0027206435333937407, + 0.0033165195491164923, + 0.017561566084623337, + 0.002484151627868414, + 0.004628849681466818, + 0.006240260321646929, + 0.014266951940953732, + 0.007947811856865883, + 0.004660641774535179, + 0.005620312411338091, + 0.006097343750298023, + 0.004072473384439945, + 0.004481874871999025, + 0.004993499256670475, + 0.0022039050236344337, + 0.001874527195468545, + 0.001007946440950036, + 0.0023075321223586798, + 0.004888528026640415, + 0.001387649797834456, + 0.0021480750292539597, + 0.007069034036248922, + 0.004781457595527172, + 0.010295618325471878, + 0.004597490653395653, + 0.005462028086185455, + 0.004205630626529455, + 0.004519197624176741, + 0.005715727340430021, + 0.09042494744062424, + 0.007628193590790033, + 0.08092532306909561, + 0.0068277232348918915, + 0.010291050188243389, + 0.021484263241291046, + 0.004704921972006559, + 0.0043696132488548756, + 0.011844118125736713, + 0.003935770597308874, + 0.005066273268312216, + 0.0031106597743928432, + 0.004451839253306389, + 0.008575626648962498, + 0.0026254041586071253, + 0.003251352347433567, + 0.003921476658433676, + 0.0017061938997358084, + 0.0017244082409888506, + 0.0018561125034466386, + 0.0028862138278782368, + 0.0029112230986356735, + 0.0013627036241814494, + 0.0015181468334048986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.07006238400936127 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the CN Tower's concrete structure located at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle with a wide-angle lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action of the observation deck collapsing and chunks breaking away rendered in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, unforgiving sunlight and sharp geometric shadows visible on the tower's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast, billowing plume of fine gray dust visible blooming outward from the rupture point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14278195798397064 + ], + "opposite_logprob_tokens": [ + -2.017781972885132 + ], + "logprob_reward": -0.14278195798397064, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14278195798397064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20118944346904755, + 0.006535226479172707, + 0.012586119584739208, + 0.14454008638858795, + 0.09226720780134201, + 0.0673794373869896, + 0.10077524185180664, + 0.3141055703163147, + 0.026204530149698257, + 0.03441712260246277 + ], + "credit_map": [ + 0.025195922702550888, + 0.007922562770545483, + 0.06804610043764114, + 0.005719552282243967, + 0.016303209587931633, + 0.04429593309760094, + 0.00402116309851408, + 0.029684992507100105, + 0.0006873589009046555, + 0.001040215720422566, + 0.000879968807566911, + 0.0005937143578194082, + 0.0012662195367738605, + 0.0009052950772456825, + 0.0006626868853345513, + 0.0004997669602744281, + 0.000795316940639168, + 0.0007437950116582215, + 0.001395577797666192, + 0.0004472880973480642, + 0.0006150330300442874, + 0.0036502787843346596, + 0.0004437353927642107, + 0.004495094530284405, + 0.00735344598069787, + 0.004427063278853893, + 0.0542592890560627, + 0.004716037772595882, + 0.006010922137647867, + 0.05801377445459366, + 0.0035167699679732323, + 0.006242772098630667, + 0.012112786062061787, + 0.013628128916025162, + 0.02146937884390354, + 0.0073329065926373005, + 0.00894348043948412, + 0.011529254727065563, + 0.008142419159412384, + 0.009108856320381165, + 0.013389229774475098, + 0.008639865554869175, + 0.008823241107165813, + 0.004818051587790251, + 0.007776438724249601, + 0.007909056730568409, + 0.005829197820276022, + 0.010194355621933937, + 0.019017687067389488, + 0.02043812908232212, + 0.023047935217618942, + 0.007474015932530165, + 0.010807950049638748, + 0.007407395169138908, + 0.005576647352427244, + 0.007005489896982908, + 0.12607818841934204, + 0.010901295579969883, + 0.11199542880058289, + 0.009022120386362076, + 0.013776934705674648, + 0.03047368675470352, + 0.005741722881793976, + 0.006116207223385572, + 0.005720941349864006, + 0.0020433026365935802, + 0.0028716030064970255, + 0.003173677483573556, + 0.0029724673368036747, + 0.0045344093814492226, + 0.0028811951633542776, + 0.0020069312304258347, + 0.004654020071029663, + 0.0060756332240998745, + 0.003904381301254034, + 0.0060730366967618465, + 0.004240711219608784, + 0.0027714467141777277, + 0.004758552182465792, + 0.0019393411930650473 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.14278195798397064 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the CN Tower's concrete structure located at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle with a wide-angle lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action of the observation deck collapsing and chunks breaking away rendered in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, unforgiving sunlight and sharp geometric shadows visible on the tower's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast, billowing plume of fine gray dust visible blooming outward from the rupture point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813030779361725 + ], + "opposite_logprob_tokens": [ + -1.40630304813385 + ], + "logprob_reward": -0.2813030779361725, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813030779361725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11765167117118835, + 0.015768317505717278, + 0.14207996428012848, + 0.15352889895439148, + 0.06358803808689117, + 0.1293393224477768, + 0.06506738066673279, + 0.11180609464645386, + 0.10436450690031052, + 0.09680573642253876 + ], + "credit_map": [ + 0.01637231558561325, + 0.005053468979895115, + 0.03430923819541931, + 0.0037858132272958755, + 0.011516189202666283, + 0.02904437854886055, + 0.0027178397867828608, + 0.014852428808808327, + 0.0016398372827097774, + 0.001392496982589364, + 0.0036267719697207212, + 0.0008835500339046121, + 0.001828533597290516, + 0.003393532708287239, + 0.0011382974917069077, + 0.001865297555923462, + 0.009325658902525902, + 0.004899794701486826, + 0.05384035408496857, + 0.0045302254147827625, + 0.005769993644207716, + 0.01143772155046463, + 0.0030158618465065956, + 0.049260351806879044, + 0.008153031580150127, + 0.003342229640111327, + 0.06061401963233948, + 0.004773473832756281, + 0.0056552463211119175, + 0.06511928141117096, + 0.0020492342300713062, + 0.003822374390438199, + 0.008277221582829952, + 0.009588655084371567, + 0.01468516606837511, + 0.0048363772220909595, + 0.006539735943078995, + 0.008065621368587017, + 0.005042838864028454, + 0.006552424747496843, + 0.019496755674481392, + 0.013282724656164646, + 0.02337205596268177, + 0.012074504978954792, + 0.010240132920444012, + 0.022205496206879616, + 0.016025017946958542, + 0.012642639689147472, + 0.013163626194000244, + 0.007982039824128151, + 0.01730126515030861, + 0.005355179775506258, + 0.006959064863622189, + 0.0047470577992498875, + 0.004350379109382629, + 0.005208766553550959, + 0.042386505752801895, + 0.0055039869621396065, + 0.03730490803718567, + 0.004530634265393019, + 0.005806274712085724, + 0.010316061787307262, + 0.00297571555711329, + 0.0029820126947015524, + 0.026337474584579468, + 0.007130365818738937, + 0.012020128779113293, + 0.0067774392664432526, + 0.013744521886110306, + 0.022354688495397568, + 0.007786869537085295, + 0.008213027380406857, + 0.02130875550210476, + 0.010457698255777359, + 0.010821628384292126, + 0.009502771310508251, + 0.012948513962328434, + 0.012705155648291111, + 0.009419429115951061, + 0.009641784243285656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 5 + ], + "selected_reward": -0.2813030779361725 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the CN Tower's concrete structure located at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle with a wide-angle lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action of the observation deck collapsing and chunks breaking away rendered in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, unforgiving sunlight and sharp geometric shadows visible on the tower's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast, billowing plume of fine gray dust visible blooming outward from the rupture point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010446811793372035 + ], + "opposite_logprob_tokens": [ + -6.876044750213623 + ], + "logprob_reward": -0.0010446811793372035, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010446811793372035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.024365274235606194, + 0.014735819771885872, + 0.04796228185296059, + 0.018817121163010597, + 0.08896738290786743, + 0.1100786104798317, + 0.054515279829502106, + 0.4006637930870056, + 0.1826409101486206, + 0.05725358799099922 + ], + "credit_map": [ + 0.00249618012458086, + 0.0018979887245222926, + 0.007879582233726978, + 0.0012135719880461693, + 0.0018172204727306962, + 0.0036635291762650013, + 0.0009010094217956066, + 0.0044961911626160145, + 0.0013416598085314035, + 0.0020961787085980177, + 0.0016596016939729452, + 0.0012914658291265368, + 0.0020932217594236135, + 0.0026376955211162567, + 0.0019559140782803297, + 0.0016600809758529067, + 0.0027901374269276857, + 0.001486239256337285, + 0.0194038487970829, + 0.0018443419830873609, + 0.002016575774177909, + 0.007392614148557186, + 0.0010184068232774734, + 0.012010117992758751, + 0.0014778545591980219, + 0.0013066898100078106, + 0.004881983157247305, + 0.0011600109282881021, + 0.0011759442277252674, + 0.006036968436092138, + 0.0016105244867503643, + 0.0011671464890241623, + 0.01154252514243126, + 0.013084440492093563, + 0.020762139931321144, + 0.006843484006822109, + 0.008671388030052185, + 0.011016342788934708, + 0.0073358165100216866, + 0.009711243212223053, + 0.0349494069814682, + 0.005652512889355421, + 0.011928566731512547, + 0.006800561212003231, + 0.00687236525118351, + 0.027691271156072617, + 0.008856366388499737, + 0.007327558938413858, + 0.012346521019935608, + 0.010633271187543869, + 0.012393570505082607, + 0.002981747267767787, + 0.005362330470234156, + 0.004389800596982241, + 0.00269951275549829, + 0.0037085292860865593, + 0.1599665880203247, + 0.014762411825358868, + 0.13772329688072205, + 0.011974591761827469, + 0.01872951351106167, + 0.04077136144042015, + 0.008379527367651463, + 0.008356522768735886, + 0.04747960716485977, + 0.013313799165189266, + 0.02112489752471447, + 0.011052170768380165, + 0.021757125854492188, + 0.044795408844947815, + 0.011567666195333004, + 0.011550221592187881, + 0.014173964038491249, + 0.00326435430906713, + 0.005462334491312504, + 0.004347893409430981, + 0.008091880939900875, + 0.011057519353926182, + 0.004634446930140257, + 0.006221195217221975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.0010446811793372035 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Static, low-angle, wide-angle shots rendered in ultra-detailed 8K resolution capture the moment of catastrophic failure using the visual language of ultra-high-speed photography. The CN Tower's iconic, needle-like concrete structure stands at the center, silhouetted against a desolate, perfectly clear blue sky. Bright, unforgiving sunlight casts sharp, geometric shadows across its surface. As the main observation deck begins to collapse, the action begins in extreme slow motion; spiderwebs on the concrete facade split open instantly, before massive chunks weighing several tons violently break away, revealing twisted, dark steel and internal framing beneath. Shards of glass from the panoramic windows explode outward like crystalline mist, catching the sunlight as they begin their graceful yet terrifying descent. As the initial collapse unleashes a vast, billowing plume of fine gray dust that blooms outward from the rupture point, its particles individually visible against the blue background, every microsecond of the structural damage is rendered with incredible force. The scene is a surreal, silent devastation, a beautiful yet terrifying physical ballet where falling debris appears to be suspended in mid-air, creating an atmosphere of both awe-inspiring and clinical catastrophe.", + "seed": 24115, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_3/index_2821_seed_24115_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.10712598732206971, + "vlm_logprob": -0.10712598732206971, + "vlm_yes_no_margin": 3.049999976158142, + "vlm_token_credit": -0.10712598732206971, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the CN Tower's concrete structure located at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle with a wide-angle lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action of the observation deck collapsing and chunks breaking away rendered in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, unforgiving sunlight and sharp geometric shadows visible on the tower's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast, billowing plume of fine gray dust visible blooming outward from the rupture point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04221339151263237 + ], + "opposite_logprob_tokens": [ + -3.2922134399414062 + ], + "logprob_reward": -0.04221339151263237, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04221339151263237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2971934378147125, + 0.08127565681934357, + 0.09943757206201553, + 0.08271301537752151, + 0.047752805054187775, + 0.06174934282898903, + 0.09691791236400604, + 0.10634614527225494, + 0.03959084302186966, + 0.08702327311038971 + ], + "credit_map": [ + 0.014292384497821331, + 0.009124727919697762, + 0.01396659854799509, + 0.0073418463580310345, + 0.012044291943311691, + 0.10988164693117142, + 0.008748569525778294, + 0.12179336696863174, + 0.008875240571796894, + 0.0032610977068543434, + 0.029467029497027397, + 0.0038553171325474977, + 0.004942832048982382, + 0.009805317968130112, + 0.002851904137060046, + 0.018216921016573906, + 0.005289089400321245, + 0.004465354140847921, + 0.00547422468662262, + 0.002891422947868705, + 0.005246603861451149, + 0.037418607622385025, + 0.003338366746902466, + 0.03531390056014061, + 0.009374392218887806, + 0.00813333597034216, + 0.012497751973569393, + 0.006698810029774904, + 0.012540196068584919, + 0.01482168585062027, + 0.008567798882722855, + 0.01007904950529337, + 0.009953567758202553, + 0.004317577462643385, + 0.003325263038277626, + 0.0037672617472708225, + 0.004888184834271669, + 0.01125597395002842, + 0.0040268623270094395, + 0.006218112073838711, + 0.009574376046657562, + 0.004748435225337744, + 0.006411328911781311, + 0.005917290225625038, + 0.006167948246002197, + 0.006133201532065868, + 0.012268857099115849, + 0.010527903214097023, + 0.015660366043448448, + 0.016545334830880165, + 0.00952241662889719, + 0.0076032523065805435, + 0.011613432317972183, + 0.010990060865879059, + 0.01109534502029419, + 0.013887707144021988, + 0.01518288068473339, + 0.010957643389701843, + 0.011295578442513943, + 0.009049120359122753, + 0.009063494391739368, + 0.018435075879096985, + 0.01304419245570898, + 0.01931815966963768, + 0.005531274247914553, + 0.007215247489511967, + 0.003468266921117902, + 0.004285798873752356, + 0.006022615358233452, + 0.0036035366356372833, + 0.00475595286116004, + 0.004708150867372751, + 0.013574306853115559, + 0.00846065767109394, + 0.012029051780700684, + 0.008194518275558949, + 0.012437037192285061, + 0.014989452436566353, + 0.00481184758245945, + 0.012526399455964565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.04221339151263237 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the CN Tower's concrete structure located at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle with a wide-angle lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action of the observation deck collapsing and chunks breaking away rendered in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, unforgiving sunlight and sharp geometric shadows visible on the tower's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast, billowing plume of fine gray dust visible blooming outward from the rupture point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12707409262657166 + ], + "opposite_logprob_tokens": [ + -2.1270740032196045 + ], + "logprob_reward": -0.12707409262657166, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12707409262657166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05239725112915039, + 0.07557687908411026, + 0.05459416285157204, + 0.022882284596562386, + 0.687358021736145, + 0.03286176919937134, + 0.022782349959015846, + 0.029402287676930428, + 0.0045774453319609165, + 0.017567573115229607 + ], + "credit_map": [ + 0.009927577339112759, + 0.003940206486731768, + 0.005914047826081514, + 0.003665923373773694, + 0.003701112000271678, + 0.0107994070276618, + 0.0034660750534385443, + 0.010982902720570564, + 0.0036673310678452253, + 0.0032624313607811928, + 0.003063700394704938, + 0.0019007317023351789, + 0.002717470983043313, + 0.029836691915988922, + 0.0017921366961672902, + 0.02933637425303459, + 0.0023505541030317545, + 0.0017813666490837932, + 0.006199561059474945, + 0.0012925976188853383, + 0.0019030093681067228, + 0.022292394191026688, + 0.0013462136266753078, + 0.017428463324904442, + 0.0026505892165005207, + 0.0024183194618672132, + 0.003580985125154257, + 0.0017533382633700967, + 0.003637105692178011, + 0.0035525879357010126, + 0.0026215766556560993, + 0.0026677825953811407, + 0.2459639012813568, + 0.02190140262246132, + 0.030878422781825066, + 0.013975040055811405, + 0.01657223142683506, + 0.2685684263706207, + 0.02204469032585621, + 0.06745393574237823, + 0.005329970736056566, + 0.0028499511536210775, + 0.003587688785046339, + 0.0020421738736331463, + 0.004628216847777367, + 0.006652690004557371, + 0.0021702914964407682, + 0.005600784905254841, + 0.0038735477719455957, + 0.0038622331339865923, + 0.002515878528356552, + 0.0018390514887869358, + 0.0026359912008047104, + 0.0026037809439003468, + 0.002312257420271635, + 0.0031396092381328344, + 0.004164792597293854, + 0.002215607324615121, + 0.003647912060841918, + 0.002578779123723507, + 0.0026135039515793324, + 0.004893539473414421, + 0.004160650540143251, + 0.005127502139657736, + 0.0005814388860017061, + 0.0006569449906237423, + 0.0005063796415925026, + 0.000494443520437926, + 0.0005289869732223451, + 0.0005395193002186716, + 0.00039935679524205625, + 0.0008703753701411188, + 0.0024955517146736383, + 0.003110426478087902, + 0.0018064610194414854, + 0.0018268132116645575, + 0.003351419698446989, + 0.0022338780108839273, + 0.0010312737431377172, + 0.0017117481911554933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.12707409262657166 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the CN Tower's concrete structure located at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle with a wide-angle lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action of the observation deck collapsing and chunks breaking away rendered in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, unforgiving sunlight and sharp geometric shadows visible on the tower's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast, billowing plume of fine gray dust visible blooming outward from the rupture point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520752549171448 + ], + "opposite_logprob_tokens": [ + -1.5020751953125 + ], + "logprob_reward": -0.2520752549171448, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520752549171448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24524398148059845, + 0.0909590870141983, + 0.018798811361193657, + 0.015880079939961433, + 0.17921561002731323, + 0.08706668764352798, + 0.1396947056055069, + 0.12136887013912201, + 0.06230069324374199, + 0.03947147727012634 + ], + "credit_map": [ + 0.00916481763124466, + 0.006040958222001791, + 0.00931314192712307, + 0.00386290461756289, + 0.0053844088688492775, + 0.10146903991699219, + 0.004186184611171484, + 0.10582251846790314, + 0.0057685296051204205, + 0.003389570629224181, + 0.017688386142253876, + 0.002516541862860322, + 0.0026920505333691835, + 0.03561229258775711, + 0.0021524711046367884, + 0.021139245480298996, + 0.0016139011131599545, + 0.0009214626043103635, + 0.004595864564180374, + 0.000671586487442255, + 0.0011787025723606348, + 0.00649647880345583, + 0.0006807210738770664, + 0.0026400939095765352, + 0.0018977494910359383, + 0.0011233651312068105, + 0.002614468103274703, + 0.0010537017369642854, + 0.0016262888675555587, + 0.003980654291808605, + 0.0011403049575164914, + 0.0024435478262603283, + 0.0622810460627079, + 0.008148779161274433, + 0.009433458559215069, + 0.004907486494630575, + 0.006459400523453951, + 0.06564954668283463, + 0.0074323746375739574, + 0.014903513714671135, + 0.015110304579138756, + 0.006612621713429689, + 0.007565218955278397, + 0.0064099938608706, + 0.009203244000673294, + 0.019217034801840782, + 0.008724971674382687, + 0.014223298989236355, + 0.01997688226401806, + 0.021427111700177193, + 0.013667456805706024, + 0.012761145830154419, + 0.01792532578110695, + 0.01720605604350567, + 0.017173152416944504, + 0.019557572901248932, + 0.01800289936363697, + 0.013201974332332611, + 0.009568721055984497, + 0.013641209341585636, + 0.01478036493062973, + 0.013993247412145138, + 0.015411842614412308, + 0.022768603637814522, + 0.00817854329943657, + 0.006628993432968855, + 0.007895620539784431, + 0.008181121200323105, + 0.00775162735953927, + 0.008292712271213531, + 0.005883386824280024, + 0.00948868878185749, + 0.005740209482610226, + 0.004592305514961481, + 0.005591512192040682, + 0.004064311273396015, + 0.004881097003817558, + 0.005537759978324175, + 0.003022687043994665, + 0.00604159664362669 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.2520752549171448 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the CN Tower's concrete structure located at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle with a wide-angle lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action of the observation deck collapsing and chunks breaking away rendered in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, unforgiving sunlight and sharp geometric shadows visible on the tower's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast, billowing plume of fine gray dust visible blooming outward from the rupture point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292830109596252 + ], + "opposite_logprob_tokens": [ + -2.2379283905029297 + ], + "logprob_reward": -0.11292830109596252, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292830109596252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15587982535362244, + 0.133558988571167, + 0.09029226005077362, + 0.04624538496136665, + 0.3465661406517029, + 0.07635105401277542, + 0.06648387759923935, + 0.019960960373282433, + 0.017678091302514076, + 0.04698345437645912 + ], + "credit_map": [ + 0.009362987242639065, + 0.004371361806988716, + 0.010503952391445637, + 0.00418234383687377, + 0.0047860643826425076, + 0.055681776255369186, + 0.004009311553090811, + 0.06298200786113739, + 0.007175212260335684, + 0.003464329522103071, + 0.015973253175616264, + 0.002086048247292638, + 0.003750792471691966, + 0.054943978786468506, + 0.001894496032036841, + 0.044270869344472885, + 0.004160708747804165, + 0.0020468460861593485, + 0.010796352289617062, + 0.0014541895361617208, + 0.0026284626219421625, + 0.03788164630532265, + 0.0013614825438708067, + 0.029962560161948204, + 0.002416331321001053, + 0.0017150043277069926, + 0.0031337228137999773, + 0.0013365817721933126, + 0.0027322922833263874, + 0.01765618287026882, + 0.0013548398856073618, + 0.01590043120086193, + 0.1286008208990097, + 0.011438462883234024, + 0.017261138185858727, + 0.006964823696762323, + 0.008432387374341488, + 0.13308881223201752, + 0.010658982209861279, + 0.030120719224214554, + 0.014351943507790565, + 0.005829377565532923, + 0.008194179274141788, + 0.005551543552428484, + 0.010019653476774693, + 0.01589009538292885, + 0.006932642310857773, + 0.009581616148352623, + 0.009383214637637138, + 0.010430270805954933, + 0.005026482976973057, + 0.00690804049372673, + 0.008691847324371338, + 0.007867418229579926, + 0.009498910047113895, + 0.008677685633301735, + 0.0027660762425512075, + 0.0023181908763945103, + 0.0018172968411818147, + 0.003519325517117977, + 0.0028016383294016123, + 0.001586238038726151, + 0.0022529156412929296, + 0.0028992786537855864, + 0.0026188637129962444, + 0.0025070314295589924, + 0.0016823785845190287, + 0.00181921839248389, + 0.001757472869940102, + 0.001437021535821259, + 0.0016937173204496503, + 0.00416238559409976, + 0.006578637287020683, + 0.006975528784096241, + 0.005815497599542141, + 0.005034281872212887, + 0.006527463439851999, + 0.00829076673835516, + 0.003257407806813717, + 0.004503866657614708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.11292830109596252 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the CN Tower's concrete structure located at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle with a wide-angle lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action of the observation deck collapsing and chunks breaking away rendered in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, unforgiving sunlight and sharp geometric shadows visible on the tower's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast, billowing plume of fine gray dust visible blooming outward from the rupture point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013388964580371976 + ], + "opposite_logprob_tokens": [ + -6.626338958740234 + ], + "logprob_reward": -0.0013388964580371976, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013388964580371976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13290806114673615, + 0.09378381818532944, + 0.0471060574054718, + 0.05361348018050194, + 0.08820094913244247, + 0.050975501537323, + 0.22469072043895721, + 0.18350714445114136, + 0.06684069335460663, + 0.058373548090457916 + ], + "credit_map": [ + 0.00731833279132843, + 0.004007947165518999, + 0.010850588791072369, + 0.003416783642023802, + 0.004238000605255365, + 0.05423808470368385, + 0.0032508487347513437, + 0.04558747634291649, + 0.006389108952134848, + 0.004832406993955374, + 0.01465432345867157, + 0.003470177762210369, + 0.007675889879465103, + 0.03204388916492462, + 0.0037255436182022095, + 0.02099248208105564, + 0.003640163689851761, + 0.0034376971889287233, + 0.003720901207998395, + 0.0021750573068857193, + 0.004502179566770792, + 0.014428389258682728, + 0.0023859154898673296, + 0.012815749272704124, + 0.006235739681869745, + 0.004638319835066795, + 0.0061725955456495285, + 0.0033229971304535866, + 0.006867213174700737, + 0.012490360997617245, + 0.004246354103088379, + 0.009639899246394634, + 0.02957516349852085, + 0.004275901708751917, + 0.005306333769112825, + 0.0034669847227633, + 0.003140656277537346, + 0.031013255938887596, + 0.0037963143549859524, + 0.0076263402588665485, + 0.00907158013433218, + 0.004636548925191164, + 0.008996150456368923, + 0.0051312497816979885, + 0.004680798854678869, + 0.0067141009494662285, + 0.004977623466402292, + 0.006767451763153076, + 0.03972514346241951, + 0.042420923709869385, + 0.01852363906800747, + 0.01827303320169449, + 0.026914702728390694, + 0.025822028517723083, + 0.02382943034172058, + 0.029181823134422302, + 0.02614954486489296, + 0.02396112121641636, + 0.017630353569984436, + 0.022355467081069946, + 0.023943433538079262, + 0.018214212730526924, + 0.02537367306649685, + 0.025879334658384323, + 0.009889624081552029, + 0.00740528479218483, + 0.007391044404357672, + 0.009175945073366165, + 0.006064019165933132, + 0.008710671216249466, + 0.007210778538137674, + 0.010993325151503086, + 0.008422202430665493, + 0.00817486084997654, + 0.007720563560724258, + 0.005005369428545237, + 0.007963869720697403, + 0.007251404691487551, + 0.004383825231343508, + 0.00945145171135664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.0013388964580371976 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Static, low-angle, wide-angle shots rendered in ultra-detailed 8K resolution capture the moment of catastrophic failure using the visual language of ultra-high-speed photography. The CN Tower's iconic, needle-like concrete structure stands at the center, silhouetted against a desolate, perfectly clear blue sky. Bright, unforgiving sunlight casts sharp, geometric shadows across its surface. As the main observation deck begins to collapse, the action begins in extreme slow motion; spiderwebs on the concrete facade split open instantly, before massive chunks weighing several tons violently break away, revealing twisted, dark steel and internal framing beneath. Shards of glass from the panoramic windows explode outward like crystalline mist, catching the sunlight as they begin their graceful yet terrifying descent. As the initial collapse unleashes a vast, billowing plume of fine gray dust that blooms outward from the rupture point, its particles individually visible against the blue background, every microsecond of the structural damage is rendered with incredible force. The scene is a surreal, silent devastation, a beautiful yet terrifying physical ballet where falling debris appears to be suspended in mid-air, creating an atmosphere of both awe-inspiring and clinical catastrophe.", + "seed": 24116, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_3/index_2821_seed_24116_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.030519904103130102, + "vlm_logprob": -0.030519904103130102, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.030519904103130102, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the CN Tower's concrete structure located at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle with a wide-angle lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action of the observation deck collapsing and chunks breaking away rendered in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, unforgiving sunlight and sharp geometric shadows visible on the tower's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast, billowing plume of fine gray dust visible blooming outward from the rupture point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03743552789092064 + ], + "opposite_logprob_tokens": [ + -3.537435531616211 + ], + "logprob_reward": -0.03743552789092064, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03743552789092064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02209613099694252, + 0.027860458940267563, + 0.07197316735982895, + 0.7754250168800354, + 0.018691876903176308, + 0.008993932977318764, + 0.017618268728256226, + 0.04469256103038788, + 0.007022288162261248, + 0.005626343656331301 + ], + "credit_map": [ + 0.002967167180031538, + 0.003081920789554715, + 0.003297189949080348, + 0.0012151460396125913, + 0.003476570826023817, + 0.0021418288815766573, + 0.0027749096043407917, + 0.003141398774459958, + 0.0028317365795373917, + 0.0035337849985808134, + 0.005869769025593996, + 0.0012233786983415484, + 0.005173479672521353, + 0.003486799541860819, + 0.002085629152134061, + 0.0036558820866048336, + 0.00808640941977501, + 0.0009074946865439415, + 0.027726223692297935, + 0.002110518515110016, + 0.0018044878961518407, + 0.00546076986938715, + 0.0013176443753764033, + 0.024559618905186653, + 0.2797829508781433, + 0.01732872985303402, + 0.10200794041156769, + 0.010911563411355019, + 0.015388122759759426, + 0.05109477415680885, + 0.020970623940229416, + 0.2779403328895569, + 0.003095621010288596, + 0.001507086562924087, + 0.0028442172333598137, + 0.0037001878954470158, + 0.001188729889690876, + 0.0019832218531519175, + 0.0016884371871128678, + 0.0026843745727092028, + 0.0013973559252917767, + 0.0007034505251795053, + 0.0015846772585064173, + 0.0013716513058170676, + 0.0008214166737161577, + 0.0013111321022734046, + 0.000940381025429815, + 0.0008638677536509931, + 0.002434835070744157, + 0.0015197054017335176, + 0.0033692040015012026, + 0.002162372227758169, + 0.0016168580623343587, + 0.0020264030899852514, + 0.0021038963459432125, + 0.0023849960416555405, + 0.009415618143975735, + 0.003633793443441391, + 0.01174343004822731, + 0.003639577655121684, + 0.002943573985248804, + 0.004776385147124529, + 0.004327767062932253, + 0.004212416708469391, + 0.001186314388178289, + 0.0006846536998637021, + 0.0011161691509187222, + 0.0005770494462922215, + 0.0005011541652493179, + 0.000989220687188208, + 0.000737409049179405, + 0.0012303179828450084, + 0.0008502971613779664, + 0.0008559227571822703, + 0.0010674847289919853, + 0.000532803765963763, + 0.00048018485540524125, + 0.0005851239548064768, + 0.0006845184252597392, + 0.0005700078327208757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 7 + ], + "selected_reward": -0.03743552789092064 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the CN Tower's concrete structure located at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle with a wide-angle lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action of the observation deck collapsing and chunks breaking away rendered in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, unforgiving sunlight and sharp geometric shadows visible on the tower's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast, billowing plume of fine gray dust visible blooming outward from the rupture point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023316025733947754 + ], + "opposite_logprob_tokens": [ + -3.773315906524658 + ], + "logprob_reward": -0.023316025733947754, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023316025733947754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04418332129716873, + 0.09914092719554901, + 0.29294317960739136, + 0.1979568600654602, + 0.07675015926361084, + 0.05515066906809807, + 0.025259921327233315, + 0.16617025434970856, + 0.00489892577752471, + 0.037545762956142426 + ], + "credit_map": [ + 0.007912610657513142, + 0.0034171126317232847, + 0.00820209551602602, + 0.0019969921559095383, + 0.00522333150729537, + 0.004509510938078165, + 0.0053434656001627445, + 0.007578201126307249, + 0.011905526742339134, + 0.0037401553709059954, + 0.033768180757761, + 0.0031909560784697533, + 0.0039755613543093204, + 0.011196318082511425, + 0.0027845739386975765, + 0.028579652309417725, + 0.03326621651649475, + 0.004693883471190929, + 0.1062922328710556, + 0.010216981172561646, + 0.0091684116050601, + 0.025631193071603775, + 0.006023203954100609, + 0.09765105694532394, + 0.07086144387722015, + 0.004566890187561512, + 0.025925280526280403, + 0.0035477476194500923, + 0.004631149582564831, + 0.013487225398421288, + 0.005747046787291765, + 0.06919008493423462, + 0.012855344451963902, + 0.004812533967196941, + 0.010165393352508545, + 0.009148240089416504, + 0.006082671228796244, + 0.013574372045695782, + 0.006950153969228268, + 0.013161451555788517, + 0.0087123429402709, + 0.003524705534800887, + 0.011155456304550171, + 0.006896487437188625, + 0.004848498851060867, + 0.009138958528637886, + 0.005090050399303436, + 0.005784173496067524, + 0.003528136294335127, + 0.002128349617123604, + 0.0035117738880217075, + 0.003117291023954749, + 0.002267630770802498, + 0.003126702504232526, + 0.0035372753627598286, + 0.004042761400341988, + 0.03547302633523941, + 0.012758037075400352, + 0.044368159025907516, + 0.010976092889904976, + 0.013888141140341759, + 0.01704542338848114, + 0.012310048565268517, + 0.019351325929164886, + 0.0007167548174038529, + 0.0005556555115617812, + 0.0009764955029822886, + 0.0004268235061317682, + 0.0005115507519803941, + 0.0005362483789213002, + 0.0004993877955712378, + 0.0006760095711797476, + 0.008174708113074303, + 0.0026169726625084877, + 0.012161185033619404, + 0.002228779485449195, + 0.002197194378823042, + 0.0033614449203014374, + 0.002969984896481037, + 0.003835489973425865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 7 + ], + "selected_reward": -0.023316025733947754 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the CN Tower's concrete structure located at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle with a wide-angle lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action of the observation deck collapsing and chunks breaking away rendered in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, unforgiving sunlight and sharp geometric shadows visible on the tower's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast, billowing plume of fine gray dust visible blooming outward from the rupture point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369637206196785 + ], + "opposite_logprob_tokens": [ + -3.408696413040161 + ], + "logprob_reward": -0.03369637206196785, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369637206196785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.026839856058359146, + 0.031365904957056046, + 0.05800097808241844, + 0.5588801503181458, + 0.12492522597312927, + 0.04826584830880165, + 0.03983839228749275, + 0.060814931988716125, + 0.009430383332073689, + 0.04163831099867821 + ], + "credit_map": [ + 0.0034642911050468683, + 0.0016035826411098242, + 0.007378158625215292, + 0.0008930503390729427, + 0.00261533516459167, + 0.003570270724594593, + 0.0018117167055606842, + 0.005503450054675341, + 0.0036563840694725513, + 0.0010404683416709304, + 0.011734512634575367, + 0.00104525126516819, + 0.0017928228480741382, + 0.0037647446151822805, + 0.001101624802686274, + 0.007230096496641636, + 0.006879521533846855, + 0.0007547456189058721, + 0.022398509085178375, + 0.0016403275076299906, + 0.0011645444901660085, + 0.00418101018294692, + 0.0015337655786424875, + 0.019448552280664444, + 0.2009042650461197, + 0.01267811469733715, + 0.07496695220470428, + 0.007495658937841654, + 0.01134548056870699, + 0.036915868520736694, + 0.015177755616605282, + 0.19939608871936798, + 0.018983466550707817, + 0.011370612308382988, + 0.016916243359446526, + 0.012902731075882912, + 0.010252478532493114, + 0.02033628709614277, + 0.010702034458518028, + 0.023461375385522842, + 0.006666284054517746, + 0.00413806177675724, + 0.00900912843644619, + 0.0066636670380830765, + 0.004364303778856993, + 0.006185591220855713, + 0.005223437678068876, + 0.0060153743252158165, + 0.007027869578450918, + 0.0031860515009611845, + 0.010831756517291069, + 0.003083988558501005, + 0.002795639680698514, + 0.004181096330285072, + 0.003282654331997037, + 0.005449335090816021, + 0.012513287365436554, + 0.004713005851954222, + 0.015865236520767212, + 0.004669472109526396, + 0.005626793950796127, + 0.00584818422794342, + 0.004788558464497328, + 0.0067903995513916016, + 0.0013238179963082075, + 0.0009633703739382327, + 0.0012417137622833252, + 0.0010934242745861411, + 0.0009782593697309494, + 0.001195823773741722, + 0.0011951529886573553, + 0.0014388212002813816, + 0.00577699625864625, + 0.004613129422068596, + 0.009383132681250572, + 0.0027586144860833883, + 0.004266621544957161, + 0.007244637701660395, + 0.003323222044855356, + 0.004271956160664558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 7 + ], + "selected_reward": -0.03369637206196785 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the CN Tower's concrete structure located at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle with a wide-angle lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action of the observation deck collapsing and chunks breaking away rendered in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, unforgiving sunlight and sharp geometric shadows visible on the tower's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast, billowing plume of fine gray dust visible blooming outward from the rupture point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495474860072136 + ], + "opposite_logprob_tokens": [ + -2.929954767227173 + ], + "logprob_reward": -0.05495474860072136, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495474860072136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.020092083141207695, + 0.007718768902122974, + 0.05802598595619202, + 0.7887148261070251, + 0.07147322595119476, + 0.016052057966589928, + 0.005807774607092142, + 0.016977554187178612, + 0.007051689550280571, + 0.008086035959422588 + ], + "credit_map": [ + 0.003097346518188715, + 0.001291113207116723, + 0.0052493480034172535, + 0.0007627690210938454, + 0.001567444996908307, + 0.0021873083896934986, + 0.0018949565710499883, + 0.004041797481477261, + 0.0014563041040673852, + 0.0008397864294238389, + 0.001539682038128376, + 0.00033595491549931467, + 0.0007712715305387974, + 0.0009498489089310169, + 0.0005513879586942494, + 0.0012745331041514874, + 0.007549639325588942, + 0.0006630553980357945, + 0.023114366456866264, + 0.001520671765320003, + 0.001129348878748715, + 0.003957376815378666, + 0.0009137004963122308, + 0.01917782798409462, + 0.2871406674385071, + 0.017586980015039444, + 0.10288335382938385, + 0.009678401052951813, + 0.016586652025580406, + 0.0518764965236187, + 0.021721966564655304, + 0.281240314245224, + 0.01090068556368351, + 0.005489058792591095, + 0.00967766810208559, + 0.012196398340165615, + 0.006600743159651756, + 0.01140719372779131, + 0.00722265150398016, + 0.007978828623890877, + 0.0026298929005861282, + 0.0008222395554184914, + 0.004014706239104271, + 0.0017663161270320415, + 0.001321305986493826, + 0.002857050858438015, + 0.0011205290211364627, + 0.0015200175112113357, + 0.000954190269112587, + 0.00037866554339416325, + 0.0012967676157131791, + 0.0005431780591607094, + 0.00046277546789497137, + 0.0009324711281806231, + 0.0006424316088669002, + 0.000597295060288161, + 0.0035191066563129425, + 0.0013004993088543415, + 0.004603228531777859, + 0.0010601795511320233, + 0.001607862883247435, + 0.0018225929234176874, + 0.0012139437021687627, + 0.0018501403974369168, + 0.001021484611555934, + 0.0007112970924936235, + 0.0012346754083409905, + 0.00088614592095837, + 0.0006908536888659, + 0.0008890306344255805, + 0.0006421953439712524, + 0.0009760066168382764, + 0.0009595511364750564, + 0.0009482118766754866, + 0.0010930205462500453, + 0.0008102875435724854, + 0.0007007752428762615, + 0.001008278806693852, + 0.0015951604582369328, + 0.000970749999396503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 2 + ], + "selected_reward": -0.05495474860072136 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the CN Tower's concrete structure located at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle with a wide-angle lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action of the observation deck collapsing and chunks breaking away rendered in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, unforgiving sunlight and sharp geometric shadows visible on the tower's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast, billowing plume of fine gray dust visible blooming outward from the rupture point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003196846228092909 + ], + "opposite_logprob_tokens": [ + -5.753196716308594 + ], + "logprob_reward": -0.003196846228092909, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003196846228092909, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03182542324066162, + 0.013619081117212772, + 0.037407491356134415, + 0.3629600405693054, + 0.15986578166484833, + 0.02572111040353775, + 0.014102251268923283, + 0.09762609004974365, + 0.05131210386753082, + 0.20556066930294037 + ], + "credit_map": [ + 0.0037822462618350983, + 0.00198077573440969, + 0.009382367134094238, + 0.0012656948529183865, + 0.0025420156307518482, + 0.003943633753806353, + 0.002845189766958356, + 0.00608349684625864, + 0.0015420515555888414, + 0.0016931263962760568, + 0.0028684467542916536, + 0.0006491352687589824, + 0.00208125333301723, + 0.0016084567178040743, + 0.0012207016116008162, + 0.001955909188836813, + 0.004302831366658211, + 0.001068137469701469, + 0.012642089277505875, + 0.0015354995848610997, + 0.001972937723621726, + 0.0038352850824594498, + 0.0016035293228924274, + 0.010447178035974503, + 0.1299608051776886, + 0.008628674782812595, + 0.04508783668279648, + 0.005281285848468542, + 0.01073632575571537, + 0.024159135296940804, + 0.011672344990074635, + 0.12743361294269562, + 0.025714244693517685, + 0.011605808511376381, + 0.0216950923204422, + 0.022990986704826355, + 0.015400445088744164, + 0.025964662432670593, + 0.015898458659648895, + 0.020596060901880264, + 0.004035669844597578, + 0.0019587173592299223, + 0.004855625331401825, + 0.003307848470285535, + 0.0024203869979828596, + 0.0037850707303732634, + 0.0027352790348231792, + 0.0026225107721984386, + 0.002110671252012253, + 0.001000485266558826, + 0.001853189547546208, + 0.00180726102553308, + 0.0014824155950918794, + 0.0015516811981797218, + 0.0024388530291616917, + 0.001857692957855761, + 0.020160451531410217, + 0.008296559564769268, + 0.02476688101887703, + 0.00774776516482234, + 0.006829360034316778, + 0.01040528155863285, + 0.009842358529567719, + 0.009577425196766853, + 0.007441929541528225, + 0.005898460745811462, + 0.0076679652556777, + 0.006219721864908934, + 0.005354659631848335, + 0.005161851178854704, + 0.005598392337560654, + 0.007969118654727936, + 0.04837958514690399, + 0.015089086256921291, + 0.07163257896900177, + 0.015700863674283028, + 0.00793148297816515, + 0.012804304249584675, + 0.012742456048727036, + 0.02128029800951481 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 4 + ], + "selected_reward": -0.003196846228092909 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Static, low-angle, wide-angle shots rendered in ultra-detailed 8K resolution capture the moment of catastrophic failure using the visual language of ultra-high-speed photography. The CN Tower's iconic, needle-like concrete structure stands at the center, silhouetted against a desolate, perfectly clear blue sky. Bright, unforgiving sunlight casts sharp, geometric shadows across its surface. As the main observation deck begins to collapse, the action begins in extreme slow motion; spiderwebs on the concrete facade split open instantly, before massive chunks weighing several tons violently break away, revealing twisted, dark steel and internal framing beneath. Shards of glass from the panoramic windows explode outward like crystalline mist, catching the sunlight as they begin their graceful yet terrifying descent. As the initial collapse unleashes a vast, billowing plume of fine gray dust that blooms outward from the rupture point, its particles individually visible against the blue background, every microsecond of the structural damage is rendered with incredible force. The scene is a surreal, silent devastation, a beautiful yet terrifying physical ballet where falling debris appears to be suspended in mid-air, creating an atmosphere of both awe-inspiring and clinical catastrophe.", + "seed": 24117, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_3/index_2821_seed_24117_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.14970235195942222, + "vlm_logprob": -0.14970235195942222, + "vlm_yes_no_margin": 2.775, + "vlm_token_credit": -0.14970235195942222, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the CN Tower's concrete structure located at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle with a wide-angle lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action of the observation deck collapsing and chunks breaking away rendered in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, unforgiving sunlight and sharp geometric shadows visible on the tower's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast, billowing plume of fine gray dust visible blooming outward from the rupture point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09565985947847366 + ], + "opposite_logprob_tokens": [ + -2.4706599712371826 + ], + "logprob_reward": -0.09565985947847366, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09565985947847366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09762103855609894, + 0.03265053778886795, + 0.7281621694564819, + 0.010298334062099457, + 0.025813555344939232, + 0.031104525551199913, + 0.02075193077325821, + 0.03133663535118103, + 0.012463896535336971, + 0.0097974743694067 + ], + "credit_map": [ + 0.004816571716219187, + 0.0042306287214159966, + 0.026640094816684723, + 0.0026113667991012335, + 0.005505992099642754, + 0.04051939398050308, + 0.0020284049678593874, + 0.011268584989011288, + 0.002186223864555359, + 0.002349947579205036, + 0.0068593514151871204, + 0.0011993980733677745, + 0.003098204731941223, + 0.012008060701191425, + 0.0013296781107783318, + 0.003619670867919922, + 0.02560509741306305, + 0.0208200104534626, + 0.2702088952064514, + 0.016561634838581085, + 0.017507009208202362, + 0.3218643367290497, + 0.00492336368188262, + 0.050671789795160294, + 0.0010658379178494215, + 0.0006229434511624277, + 0.0035344238858669996, + 0.00035276066046208143, + 0.0007780278101563454, + 0.0021846280433237553, + 0.0004897422040812671, + 0.001269969972781837, + 0.002340826438739896, + 0.002823206828907132, + 0.005017319694161415, + 0.0016546164406463504, + 0.004677127115428448, + 0.003658614819869399, + 0.002672183094546199, + 0.0029696603305637836, + 0.0032670011278241873, + 0.0028531928546726704, + 0.003217649180442095, + 0.003291867906227708, + 0.004384634084999561, + 0.0038159224204719067, + 0.005035605281591415, + 0.005238651297986507, + 0.0027030447963625193, + 0.0018418715335428715, + 0.0021510517690330744, + 0.002262095920741558, + 0.003576091956347227, + 0.0027234929148107767, + 0.0026267096400260925, + 0.002867573406547308, + 0.004941068589687347, + 0.0024296913761645555, + 0.005129462573677301, + 0.0025595990009605885, + 0.002970040077343583, + 0.005838045850396156, + 0.004083259031176567, + 0.0033854718785732985, + 0.0019107917323708534, + 0.0011249777162447572, + 0.002040449297055602, + 0.001504689920693636, + 0.001291526947170496, + 0.0017375369789078832, + 0.0016095249447971582, + 0.0012443988816812634, + 0.0012707236455753446, + 0.0009196046739816666, + 0.0015177878085523844, + 0.0007856937008909881, + 0.0009047836647368968, + 0.002375092124566436, + 0.0007763868197798729, + 0.0012474018149077892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.09565985947847366 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the CN Tower's concrete structure located at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle with a wide-angle lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action of the observation deck collapsing and chunks breaking away rendered in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, unforgiving sunlight and sharp geometric shadows visible on the tower's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast, billowing plume of fine gray dust visible blooming outward from the rupture point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295311152935028 + ], + "opposite_logprob_tokens": [ + -2.2379531860351562 + ], + "logprob_reward": -0.11295311152935028, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295311152935028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07257357239723206, + 0.0446789525449276, + 0.7789067625999451, + 0.004858530592173338, + 0.022997962310910225, + 0.024699732661247253, + 0.014266320504248142, + 0.018632477149367332, + 0.011085698381066322, + 0.00729992613196373 + ], + "credit_map": [ + 0.005498597864061594, + 0.003859366988763213, + 0.012310415506362915, + 0.0018990354146808386, + 0.004324625711888075, + 0.02889285422861576, + 0.0020072816405445337, + 0.013781403191387653, + 0.004124842584133148, + 0.0026231473311781883, + 0.011992358602583408, + 0.0012496463023126125, + 0.0029073255136609077, + 0.015286502428352833, + 0.0012083695037290454, + 0.005286762956529856, + 0.024435173720121384, + 0.029505809769034386, + 0.19474643468856812, + 0.013710699044167995, + 0.021323520690202713, + 0.3576239049434662, + 0.00580718694254756, + 0.1317541003227234, + 0.0005151506629772484, + 0.000320437247864902, + 0.0011952348286285996, + 0.00021201166964601725, + 0.0005100473645143211, + 0.00135340029373765, + 0.0002857089857570827, + 0.0004665402229875326, + 0.0022992442827671766, + 0.0019183560507372022, + 0.006020787172019482, + 0.0008929242612794042, + 0.003104693256318569, + 0.004911950323730707, + 0.0015730687882751226, + 0.002276939805597067, + 0.0024769564624875784, + 0.0024933400563895702, + 0.0032318823505192995, + 0.002106096362695098, + 0.004194957669824362, + 0.003954107407480478, + 0.0023603662848472595, + 0.0038820290938019753, + 0.0021977575961500406, + 0.001328262034803629, + 0.002362934872508049, + 0.001521794474683702, + 0.0010657863458618522, + 0.0028758004773408175, + 0.0013194072525948286, + 0.0015945796621963382, + 0.002950064605101943, + 0.0014149305643513799, + 0.0029018342029303312, + 0.00152576866094023, + 0.001801930833607912, + 0.0035892841406166553, + 0.0023029104340821505, + 0.002145756036043167, + 0.002694826340302825, + 0.000907129084225744, + 0.0031257071532309055, + 0.000621659797616303, + 0.00078500562813133, + 0.00150162517093122, + 0.0006492946995422244, + 0.0008004521369002759, + 0.0009022090816870332, + 0.0006692165043205023, + 0.0008749014814384282, + 0.0007467584800906479, + 0.0007395356660708785, + 0.0012628826079890132, + 0.0010535677429288626, + 0.0010508549166843295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.11295311152935028 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the CN Tower's concrete structure located at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle with a wide-angle lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action of the observation deck collapsing and chunks breaking away rendered in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, unforgiving sunlight and sharp geometric shadows visible on the tower's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast, billowing plume of fine gray dust visible blooming outward from the rupture point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31341615319252014 + ], + "opposite_logprob_tokens": [ + -1.3134161233901978 + ], + "logprob_reward": -0.31341615319252014, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31341615319252014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2308771163225174, + 0.061531681567430496, + 0.2365589737892151, + 0.01686924509704113, + 0.21439185738563538, + 0.08146002888679504, + 0.07279754430055618, + 0.03593680262565613, + 0.025459259748458862, + 0.024117451161146164 + ], + "credit_map": [ + 0.008354565128684044, + 0.007555514574050903, + 0.07946090400218964, + 0.005083883181214333, + 0.005987393204122782, + 0.10314769297838211, + 0.0022611012682318687, + 0.019026048481464386, + 0.002819663845002651, + 0.0020933758933097124, + 0.025703957304358482, + 0.0015200794441625476, + 0.001819893135689199, + 0.02437533251941204, + 0.0008391700685024261, + 0.0023602107539772987, + 0.015400216914713383, + 0.012925839982926846, + 0.10548189282417297, + 0.005846371408551931, + 0.004625482950359583, + 0.043930865824222565, + 0.0021441634744405746, + 0.04620414972305298, + 0.0013611434260383248, + 0.001095264800824225, + 0.0058261435478925705, + 0.0006923885084688663, + 0.0013780994340777397, + 0.0027490740176290274, + 0.001064186915755272, + 0.0027029444463551044, + 0.020332515239715576, + 0.017776593565940857, + 0.0661931186914444, + 0.00987681932747364, + 0.023557402193546295, + 0.04597754403948784, + 0.014664041809737682, + 0.016013825312256813, + 0.007768379058688879, + 0.008747548796236515, + 0.011641056276857853, + 0.00769767677411437, + 0.014633500017225742, + 0.009861762635409832, + 0.011552701704204082, + 0.009557404555380344, + 0.008810526691377163, + 0.00629717530682683, + 0.007523665204644203, + 0.008584357798099518, + 0.013237206265330315, + 0.009644809179008007, + 0.009260273538529873, + 0.009439535439014435, + 0.0053599001839756966, + 0.003459318308159709, + 0.0046920920722186565, + 0.0040901582688093185, + 0.004516061395406723, + 0.0036459872499108315, + 0.0048031797632575035, + 0.0053701079450547695, + 0.004118888173252344, + 0.0028958693146705627, + 0.0036666535306721926, + 0.0032875798642635345, + 0.0029821645002812147, + 0.0031812649685889482, + 0.002249432960525155, + 0.0030774057377129793, + 0.002542846603319049, + 0.0022708065807819366, + 0.0033020051196217537, + 0.0020288340747356415, + 0.002461640164256096, + 0.004710297100245953, + 0.002900088671594858, + 0.0039009354077279568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.31341615319252014 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the CN Tower's concrete structure located at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle with a wide-angle lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action of the observation deck collapsing and chunks breaking away rendered in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, unforgiving sunlight and sharp geometric shadows visible on the tower's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast, billowing plume of fine gray dust visible blooming outward from the rupture point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22555942833423615 + ], + "opposite_logprob_tokens": [ + -1.6005594730377197 + ], + "logprob_reward": -0.22555942833423615, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22555942833423615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22218844294548035, + 0.04298010468482971, + 0.3778580129146576, + 0.10379339754581451, + 0.0833471268415451, + 0.06365351378917694, + 0.046283092349767685, + 0.03223239257931709, + 0.015247009694576263, + 0.012417013756930828 + ], + "credit_map": [ + 0.011135670356452465, + 0.007657800801098347, + 0.08593081682920456, + 0.006556840613484383, + 0.00756274675950408, + 0.09047440439462662, + 0.0039477054961025715, + 0.008922433480620384, + 0.006597311235964298, + 0.0037214565090835094, + 0.00431912811473012, + 0.0019687104504555464, + 0.004302464425563812, + 0.010418149642646313, + 0.001920158858411014, + 0.009732723236083984, + 0.012582080438733101, + 0.01547056995332241, + 0.08568485081195831, + 0.0065176826901733875, + 0.010430926457047462, + 0.17350278794765472, + 0.0031292338389903307, + 0.07053983211517334, + 0.01162166427820921, + 0.004372854251414537, + 0.03361505642533302, + 0.002547105774283409, + 0.0042814817279577255, + 0.03455442190170288, + 0.0023790658451616764, + 0.01042174268513918, + 0.008999100886285305, + 0.004686735570430756, + 0.022940633818507195, + 0.0027426169253885746, + 0.006968332454562187, + 0.02342156693339348, + 0.0045489841140806675, + 0.009039158001542091, + 0.006007407326251268, + 0.009181834757328033, + 0.006547651253640652, + 0.005316433496773243, + 0.014014041051268578, + 0.007862476631999016, + 0.0063074869103729725, + 0.00841617863625288, + 0.007724044378846884, + 0.0038786940276622772, + 0.006595308892428875, + 0.004547392018139362, + 0.004028887953609228, + 0.007618426810950041, + 0.005530333612114191, + 0.006360001862049103, + 0.005239996127784252, + 0.002894120989367366, + 0.004188841674476862, + 0.003001326462253928, + 0.0037855489645153284, + 0.004705385770648718, + 0.0039778719656169415, + 0.0044392989948391914, + 0.003512765746563673, + 0.0013850473333150148, + 0.004259525798261166, + 0.000871578580699861, + 0.0011415646877139807, + 0.001734002958983183, + 0.0010323411552235484, + 0.0013101823860779405, + 0.0014376355102285743, + 0.0014285894576460123, + 0.0018068047938868403, + 0.001261636964045465, + 0.0011633401736617088, + 0.0022700391709804535, + 0.001183205284178257, + 0.0018657614709809422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.22555942833423615 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the CN Tower's concrete structure located at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle with a wide-angle lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action of the observation deck collapsing and chunks breaking away rendered in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, unforgiving sunlight and sharp geometric shadows visible on the tower's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast, billowing plume of fine gray dust visible blooming outward from the rupture point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009232072625309229 + ], + "opposite_logprob_tokens": [ + -7.000923156738281 + ], + "logprob_reward": -0.0009232072625309229, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009232072625309229, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09660528600215912, + 0.07996362447738647, + 0.324156790971756, + 0.026661206036806107, + 0.12873953580856323, + 0.07169056683778763, + 0.12649483978748322, + 0.07228313386440277, + 0.04531800001859665, + 0.028086965903639793 + ], + "credit_map": [ + 0.006110460963100195, + 0.0041649192571640015, + 0.03538224473595619, + 0.002518969587981701, + 0.005077027250081301, + 0.03739515691995621, + 0.0016656330553814769, + 0.004290876444429159, + 0.0041353702545166016, + 0.002595740370452404, + 0.02812325954437256, + 0.0018678625347092748, + 0.002661024685949087, + 0.0335690900683403, + 0.0009378043469041586, + 0.006073485594242811, + 0.01705734245479107, + 0.012576376087963581, + 0.14407242834568024, + 0.008352064527571201, + 0.006288120988756418, + 0.10847833752632141, + 0.0024912054650485516, + 0.024840932339429855, + 0.003347279503941536, + 0.0015813532518222928, + 0.009297243319451809, + 0.0008983895531855524, + 0.0014140649000182748, + 0.007484525442123413, + 0.0008737391326576471, + 0.001764611923135817, + 0.011805900372564793, + 0.011441132985055447, + 0.034596722573041916, + 0.006973868701606989, + 0.016600310802459717, + 0.024373454973101616, + 0.010986853390932083, + 0.011961291544139385, + 0.0062064784578979015, + 0.007369106635451317, + 0.01150473766028881, + 0.005728859454393387, + 0.011390362866222858, + 0.009959949180483818, + 0.009355503134429455, + 0.010175572708249092, + 0.016337096691131592, + 0.010433252900838852, + 0.0179614070802927, + 0.016334541141986847, + 0.014352252706885338, + 0.022014252841472626, + 0.014254390262067318, + 0.014807669445872307, + 0.012682551518082619, + 0.006532527506351471, + 0.011071993038058281, + 0.007122654467821121, + 0.006516290828585625, + 0.010665408335626125, + 0.00953430775552988, + 0.00815741065889597, + 0.007608626037836075, + 0.00477548036724329, + 0.0077483500353991985, + 0.005626594182103872, + 0.004475942812860012, + 0.004666841588914394, + 0.005206303671002388, + 0.005209868308156729, + 0.003155898302793503, + 0.002464325400069356, + 0.0038888382259756327, + 0.002350952010601759, + 0.0028842277824878693, + 0.005533630959689617, + 0.004440767224878073, + 0.003368327161297202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 6 + ], + "selected_reward": -0.0009232072625309229 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_4/index_tensor([9315, 9315])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_4/index_tensor([9315, 9315])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..326ddbac4671d8814e73a5b7f4835ecabfe1df34 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_4/index_tensor([9315, 9315])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a signature bullet-time shot, the camera pans in an extremely slow arc around the subject from a slightly upward angle, freezing the world in a surreal moment. A young fisherman, composed entirely of 8-bit pixels, wears a simple blue pixelated headscarf and a plain dark brown sleeveless jacket, standing barefoot on the wet shore of a lake composed of square blocks of pixels. He leans back slightly at the spectacle before him, his hands tightly gripping a crude, equally pixelated fishing rod. His eyes, composed of a few simple dots, reveal utter concentration and awe. Beneath him, the rippling lake surface, in extreme slow motion, ripples in concentric circles of cyan and dark blue pixels, as if time has frozen. And in the distance, a landscape painting composed of grayscale pixels unfolds in the background, its distinct peaks and crests clearly defined, a perfect fusion of retro F-style gaming and classical Eastern aesthetics. The absolute focal point of the image, a vast water crown composed of translucent white and light blue pixels, rises majestically from the center of the lake at a speed almost imperceptible to the naked eye, surrounded by countless suspended pixelated droplets, shimmering with light. The entire scene is bathed in a soft, evenly diffused light, devoid of clear shadows, further emphasizing the flatness of the image and the unique artistic style, creating an awe-inspiring atmosphere that is both serene and full of epic fantasy.", + "seed": 149, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_4/index_9315_seed_149_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04619148718193174, + "vlm_logprob": -0.04619148718193174, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.04619148718193174, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young fisherman composed entirely of 8-bit pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning in a slow arc around the subject from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lake surface composed of square blocks of pixels with rippling concentric circles in cyan and dark blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast water crown composed of translucent white and light blue pixels rising from the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a landscape painting composed of grayscale pixels with distinct peaks and crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18143229186534882 + ], + "opposite_logprob_tokens": [ + -1.8064322471618652 + ], + "logprob_reward": -0.18143229186534882, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18143229186534882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0941842570900917, + 0.06224329397082329, + 0.05313616618514061, + 0.06059771776199341, + 0.09856116026639938, + 0.12164289504289627, + 0.13194309175014496, + 0.19417628645896912, + 0.10183165222406387, + 0.08168350160121918 + ], + "credit_map": [ + 0.01565427891910076, + 0.014539712108671665, + 0.008122165687382221, + 0.008458984084427357, + 0.015150864608585835, + 0.01380750723183155, + 0.00946320965886116, + 0.008987528271973133, + 0.008973568677902222, + 0.008115667849779129, + 0.006545944605022669, + 0.007391972932964563, + 0.009258708916604519, + 0.0078616663813591, + 0.0071967714466154575, + 0.006898990832269192, + 0.009449413977563381, + 0.006015880033373833, + 0.006395487580448389, + 0.006375566590577364, + 0.005674630403518677, + 0.006179893855005503, + 0.005406706593930721, + 0.007638582494109869, + 0.011515679769217968, + 0.006849335506558418, + 0.00539774913340807, + 0.0062094940803945065, + 0.007732288911938667, + 0.011571264825761318, + 0.00595344603061676, + 0.005368456244468689, + 0.012515398673713207, + 0.011577596887946129, + 0.013724520802497864, + 0.008823266252875328, + 0.011361847631633282, + 0.018887143582105637, + 0.010578389279544353, + 0.011092985048890114, + 0.015434067696332932, + 0.04552530124783516, + 0.0034809259232133627, + 0.003951797727495432, + 0.03234988451004028, + 0.009489048272371292, + 0.005859572906047106, + 0.005552281625568867, + 0.01668805629014969, + 0.029568595811724663, + 0.020406460389494896, + 0.012173795141279697, + 0.01677132584154606, + 0.015727104619145393, + 0.00856789667159319, + 0.012039833702147007, + 0.022299079224467278, + 0.02790486440062523, + 0.019292164593935013, + 0.014280443079769611, + 0.028452051803469658, + 0.05162272974848747, + 0.010720856487751007, + 0.019604073837399483, + 0.023017585277557373, + 0.009299497120082378, + 0.009335017763078213, + 0.006785865407437086, + 0.012849605642259121, + 0.022969244047999382, + 0.008238609880208969, + 0.00933621171861887, + 0.00984896719455719, + 0.008880375884473324, + 0.005700485315173864, + 0.005332826171070337, + 0.01069013960659504, + 0.01960652694106102, + 0.006085382308810949, + 0.015538793057203293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.18143229186534882 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young fisherman composed entirely of 8-bit pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning in a slow arc around the subject from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lake surface composed of square blocks of pixels with rippling concentric circles in cyan and dark blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast water crown composed of translucent white and light blue pixels rising from the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a landscape painting composed of grayscale pixels with distinct peaks and crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370155766606331 + ], + "opposite_logprob_tokens": [ + -3.4087016582489014 + ], + "logprob_reward": -0.03370155766606331, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370155766606331, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07820211350917816, + 0.034197673201560974, + 0.028962867334485054, + 0.033840734511613846, + 0.04452374577522278, + 0.41952043771743774, + 0.06774908304214478, + 0.19861072301864624, + 0.04934408888220787, + 0.04504847526550293 + ], + "credit_map": [ + 0.012987751513719559, + 0.010705060325562954, + 0.010323965921998024, + 0.008260942064225674, + 0.00830140057951212, + 0.0125697897747159, + 0.005054894834756851, + 0.009998321533203125, + 0.004452280234545469, + 0.005191443022340536, + 0.0034260160755366087, + 0.0032877076882869005, + 0.005530022084712982, + 0.0049827201291918755, + 0.004129186272621155, + 0.003198302583768964, + 0.004610893316566944, + 0.0032547730952501297, + 0.0038180220872163773, + 0.002513042651116848, + 0.003834168426692486, + 0.0036799025256186724, + 0.002749546431005001, + 0.0045025241561234, + 0.004482951480895281, + 0.005163799971342087, + 0.0030989216174930334, + 0.004478038754314184, + 0.0042349365539848804, + 0.005667590070515871, + 0.0033516325056552887, + 0.003362868446856737, + 0.0054124402813613415, + 0.007268089335411787, + 0.00569729320704937, + 0.004838388878852129, + 0.007733616046607494, + 0.005519647616893053, + 0.004233697894960642, + 0.003820579731836915, + 0.04687613993883133, + 0.16783854365348816, + 0.009855660609900951, + 0.013490661978721619, + 0.1269008070230484, + 0.025963125750422478, + 0.01685330457985401, + 0.01174222957342863, + 0.008089447394013405, + 0.010647531598806381, + 0.01123097538948059, + 0.006071164738386869, + 0.0073829470202326775, + 0.011006000451743603, + 0.006753536872565746, + 0.006567490287125111, + 0.020845793187618256, + 0.026956770569086075, + 0.019136136397719383, + 0.01434493251144886, + 0.032599564641714096, + 0.0540158748626709, + 0.010896466672420502, + 0.019815202802419662, + 0.007069248706102371, + 0.0071245660074055195, + 0.0035479941871017218, + 0.0041259885765612125, + 0.006601192522794008, + 0.01082433108240366, + 0.0051781716756522655, + 0.004872605670243502, + 0.0038350033573806286, + 0.0036657939199358225, + 0.002758225193247199, + 0.0028413005638867617, + 0.008977502584457397, + 0.009571141563355923, + 0.0035291635431349277, + 0.009870348498225212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.03370155766606331 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young fisherman composed entirely of 8-bit pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning in a slow arc around the subject from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lake surface composed of square blocks of pixels with rippling concentric circles in cyan and dark blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast water crown composed of translucent white and light blue pixels rising from the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a landscape painting composed of grayscale pixels with distinct peaks and crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019407500512897968 + ], + "opposite_logprob_tokens": [ + -6.251940727233887 + ], + "logprob_reward": -0.0019407500512897968, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019407500512897968, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09177370369434357, + 0.09198667854070663, + 0.14701253175735474, + 0.10873503237962723, + 0.06647402048110962, + 0.06385602802038193, + 0.07590756565332413, + 0.20456191897392273, + 0.08588072657585144, + 0.06381180137395859 + ], + "credit_map": [ + 0.01365018542855978, + 0.01542186550796032, + 0.009701719507575035, + 0.008813424967229366, + 0.011591045185923576, + 0.014407450333237648, + 0.008622216992080212, + 0.009565802291035652, + 0.013805479742586613, + 0.012031110934913158, + 0.009445849806070328, + 0.01101981196552515, + 0.017489921301603317, + 0.009667355567216873, + 0.009700654074549675, + 0.008826506324112415, + 0.022476403042674065, + 0.01609109155833721, + 0.019263137131929398, + 0.015470528975129128, + 0.016132356598973274, + 0.019807878881692886, + 0.015050075016915798, + 0.0227210633456707, + 0.0183130893856287, + 0.011485453695058823, + 0.010532134212553501, + 0.015341360121965408, + 0.015435215085744858, + 0.015589503571391106, + 0.01154550351202488, + 0.010492774657905102, + 0.008375206030905247, + 0.008944587782025337, + 0.007170447148382664, + 0.005934262648224831, + 0.0071353609673678875, + 0.012899021618068218, + 0.00906835775822401, + 0.006946788169443607, + 0.007376286666840315, + 0.007859157398343086, + 0.00979585014283657, + 0.007820476777851582, + 0.010748528875410557, + 0.006711644120514393, + 0.006699575111269951, + 0.006844513118267059, + 0.009151741862297058, + 0.01935841701924801, + 0.007812767289578915, + 0.006304274313151836, + 0.010264286771416664, + 0.01179811917245388, + 0.005893049295991659, + 0.005324915517121553, + 0.023331617936491966, + 0.029296189546585083, + 0.020359717309474945, + 0.016935423016548157, + 0.028096994385123253, + 0.047357626259326935, + 0.013439666479825974, + 0.025744693353772163, + 0.012286718003451824, + 0.011261848732829094, + 0.007520182989537716, + 0.008377058431506157, + 0.009832593612372875, + 0.0169553030282259, + 0.01010245643556118, + 0.009544574655592442, + 0.008620676584541798, + 0.007831398397684097, + 0.003809525165706873, + 0.004127772059291601, + 0.011219744570553303, + 0.017519760876893997, + 0.006051071919500828, + 0.004631852265447378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 3 + ], + "selected_reward": -0.0019407500512897968 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young fisherman composed entirely of 8-bit pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning in a slow arc around the subject from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lake surface composed of square blocks of pixels with rippling concentric circles in cyan and dark blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast water crown composed of translucent white and light blue pixels rising from the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a landscape painting composed of grayscale pixels with distinct peaks and crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005248695611953735 + ], + "opposite_logprob_tokens": [ + -5.255248546600342 + ], + "logprob_reward": -0.005248695611953735, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005248695611953735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0555393248796463, + 0.05777933821082115, + 0.07119688391685486, + 0.07204906642436981, + 0.06985089182853699, + 0.22345159947872162, + 0.058831967413425446, + 0.11728885024785995, + 0.0648229643702507, + 0.20918908715248108 + ], + "credit_map": [ + 0.007353276014328003, + 0.009998071938753128, + 0.006565714254975319, + 0.00560886412858963, + 0.007212976459413767, + 0.007361022289842367, + 0.0039095343090593815, + 0.007529862690716982, + 0.008145779371261597, + 0.00727076455950737, + 0.005593504756689072, + 0.0052093821577727795, + 0.011419379152357578, + 0.007989165373146534, + 0.007533667143434286, + 0.004617698490619659, + 0.009932409971952438, + 0.007739407941699028, + 0.007648833561688662, + 0.00652092881500721, + 0.010371033102273941, + 0.01216228399425745, + 0.008039777167141438, + 0.008782208897173405, + 0.009265556000173092, + 0.008287070319056511, + 0.007223828230053186, + 0.008012264966964722, + 0.010110176168382168, + 0.012596026994287968, + 0.0096252067014575, + 0.006928938906639814, + 0.007788643706589937, + 0.008557663299143314, + 0.005046017933636904, + 0.006937408819794655, + 0.014768857508897781, + 0.010354998521506786, + 0.011286653578281403, + 0.005110646598041058, + 0.032297659665346146, + 0.05186663940548897, + 0.01540567260235548, + 0.011266861110925674, + 0.050942666828632355, + 0.021723920479416847, + 0.02152525633573532, + 0.01842292584478855, + 0.009214507415890694, + 0.009717206470668316, + 0.007734977174550295, + 0.0044424571096897125, + 0.008604233153164387, + 0.010647078976035118, + 0.00507930014282465, + 0.00339220161549747, + 0.01390680018812418, + 0.012399119324982166, + 0.009364879690110683, + 0.009976186789572239, + 0.01647227071225643, + 0.024455418810248375, + 0.01404582429677248, + 0.016668355092406273, + 0.011124187149107456, + 0.007584527134895325, + 0.005849176086485386, + 0.005686861462891102, + 0.007354564964771271, + 0.013356727547943592, + 0.007434715051203966, + 0.006432207301259041, + 0.01904682070016861, + 0.018982915207743645, + 0.010909712873399258, + 0.021116578951478004, + 0.04147797077894211, + 0.040830645710229874, + 0.02552422322332859, + 0.03130021691322327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 7 + ], + "selected_reward": -0.005248695611953735 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young fisherman composed entirely of 8-bit pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning in a slow arc around the subject from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lake surface composed of square blocks of pixels with rippling concentric circles in cyan and dark blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast water crown composed of translucent white and light blue pixels rising from the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a landscape painting composed of grayscale pixels with distinct peaks and crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008634140715003014 + ], + "opposite_logprob_tokens": [ + -4.758634090423584 + ], + "logprob_reward": -0.008634140715003014, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008634140715003014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051747582852840424, + 0.07201163470745087, + 0.036722030490636826, + 0.21303319931030273, + 0.12254398316144943, + 0.07987506687641144, + 0.10207415372133255, + 0.12005729973316193, + 0.124495729804039, + 0.07743928581476212 + ], + "credit_map": [ + 0.007305116858333349, + 0.009021297097206116, + 0.007020746823400259, + 0.006986591499298811, + 0.00597032904624939, + 0.005582777783274651, + 0.003578115487471223, + 0.006282612215727568, + 0.009076427668333054, + 0.008050831966102123, + 0.005952708423137665, + 0.007803598418831825, + 0.01518066693097353, + 0.011283027939498425, + 0.008592226542532444, + 0.006072148215025663, + 0.004638100974261761, + 0.0046812244690954685, + 0.003747959854081273, + 0.0030296798795461655, + 0.00763370655477047, + 0.005542653612792492, + 0.003618147922679782, + 0.0038305576890707016, + 0.03058527037501335, + 0.021382777020335197, + 0.013714869506657124, + 0.018084503710269928, + 0.051614612340927124, + 0.033338915556669235, + 0.025036685168743134, + 0.019275587052106857, + 0.014566057361662388, + 0.01573253609240055, + 0.01246989518404007, + 0.010087418369948864, + 0.023432543501257896, + 0.019246036186814308, + 0.015438301488757133, + 0.011571197770535946, + 0.00829937495291233, + 0.028275933116674423, + 0.0025288143660873175, + 0.003802809864282608, + 0.02132541686296463, + 0.007108947727829218, + 0.004645952023565769, + 0.003887822385877371, + 0.012172318994998932, + 0.014355520717799664, + 0.0066736009903252125, + 0.007439436390995979, + 0.014284429140388966, + 0.02821119874715805, + 0.010173533111810684, + 0.008764121681451797, + 0.013412670232355595, + 0.015084769576787949, + 0.011079874821007252, + 0.009750298224389553, + 0.01704487018287182, + 0.0293833389878273, + 0.008452908135950565, + 0.01584857702255249, + 0.020736243575811386, + 0.015227927826344967, + 0.016329461708664894, + 0.009935739450156689, + 0.014612860977649689, + 0.022538606077432632, + 0.013231866993010044, + 0.011883030645549297, + 0.01075610239058733, + 0.0071934545412659645, + 0.00424023624509573, + 0.004441121593117714, + 0.01612347923219204, + 0.01669960282742977, + 0.005333649460226297, + 0.012651641853153706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 4 + ], + "selected_reward": -0.008634140715003014 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a signature bullet-time shot, the camera pans in an extremely slow arc around the subject from a slightly upward angle, freezing the world in a surreal moment. A young fisherman, composed entirely of 8-bit pixels, wears a simple blue pixelated headscarf and a plain dark brown sleeveless jacket, standing barefoot on the wet shore of a lake composed of square blocks of pixels. He leans back slightly at the spectacle before him, his hands tightly gripping a crude, equally pixelated fishing rod. His eyes, composed of a few simple dots, reveal utter concentration and awe. Beneath him, the rippling lake surface, in extreme slow motion, ripples in concentric circles of cyan and dark blue pixels, as if time has frozen. And in the distance, a landscape painting composed of grayscale pixels unfolds in the background, its distinct peaks and crests clearly defined, a perfect fusion of retro F-style gaming and classical Eastern aesthetics. The absolute focal point of the image, a vast water crown composed of translucent white and light blue pixels, rises majestically from the center of the lake at a speed almost imperceptible to the naked eye, surrounded by countless suspended pixelated droplets, shimmering with light. The entire scene is bathed in a soft, evenly diffused light, devoid of clear shadows, further emphasizing the flatness of the image and the unique artistic style, creating an awe-inspiring atmosphere that is both serene and full of epic fantasy.", + "seed": 150, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_4/index_9315_seed_150_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1546300676651299, + "vlm_logprob": -0.1546300676651299, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.1546300676651299, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young fisherman composed entirely of 8-bit pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning in a slow arc around the subject from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lake surface composed of square blocks of pixels with rippling concentric circles in cyan and dark blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast water crown composed of translucent white and light blue pixels rising from the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a landscape painting composed of grayscale pixels with distinct peaks and crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5793074369430542 + ], + "opposite_logprob_tokens": [ + -0.8293074369430542 + ], + "logprob_reward": -0.5793074369430542, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5793074369430542, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26740288734436035, + 0.03160344436764717, + 0.047942280769348145, + 0.026284655556082726, + 0.1298406571149826, + 0.05256398022174835, + 0.06308365613222122, + 0.11215216666460037, + 0.19397996366024017, + 0.07514633238315582 + ], + "credit_map": [ + 0.039629168808460236, + 0.03591160848736763, + 0.02960873953998089, + 0.045911043882369995, + 0.022125350311398506, + 0.02551078051328659, + 0.0402342714369297, + 0.02847188524901867, + 0.004037769045680761, + 0.00447507668286562, + 0.004355777520686388, + 0.003577592084184289, + 0.005776107776910067, + 0.0034263611305505037, + 0.0029524986166507006, + 0.003002259647473693, + 0.005796331446617842, + 0.00825891736894846, + 0.00604510260745883, + 0.0050919875502586365, + 0.00464293360710144, + 0.006548445206135511, + 0.006569425109773874, + 0.004989133682101965, + 0.003990904428064823, + 0.00349825294688344, + 0.0038570421747863293, + 0.0019847212824970484, + 0.004105791449546814, + 0.0018393611535429955, + 0.004621072206646204, + 0.002387508051469922, + 0.021390466019511223, + 0.024467086419463158, + 0.01399140153080225, + 0.009121350012719631, + 0.018322492018342018, + 0.012060821987688541, + 0.018626168370246887, + 0.011860860511660576, + 0.005678043235093355, + 0.006926005240529776, + 0.004031518939882517, + 0.005982591770589352, + 0.008915050886571407, + 0.0044846804812550545, + 0.009520151652395725, + 0.007025933358818293, + 0.008770243264734745, + 0.007683353964239359, + 0.01369186770170927, + 0.006732082460075617, + 0.0062544262036681175, + 0.004997425712645054, + 0.008086484856903553, + 0.006867765914648771, + 0.015121771022677422, + 0.029206549748778343, + 0.011192391626536846, + 0.011060443706810474, + 0.014114891178905964, + 0.009900321252644062, + 0.010721876285970211, + 0.010833914391696453, + 0.022336237132549286, + 0.07030093669891357, + 0.010672514326870441, + 0.01644655503332615, + 0.023552309721708298, + 0.013891667127609253, + 0.022042402997612953, + 0.01473733689635992, + 0.010662256740033627, + 0.011798825114965439, + 0.006005570758134127, + 0.00641498202458024, + 0.015657959505915642, + 0.008935386314988136, + 0.009079255163669586, + 0.006592093035578728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.5793074369430542 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young fisherman composed entirely of 8-bit pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning in a slow arc around the subject from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lake surface composed of square blocks of pixels with rippling concentric circles in cyan and dark blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast water crown composed of translucent white and light blue pixels rising from the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a landscape painting composed of grayscale pixels with distinct peaks and crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899142801761627 + ], + "opposite_logprob_tokens": [ + -2.578991413116455 + ], + "logprob_reward": -0.07899142801761627, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899142801761627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0583459734916687, + 0.05211968347430229, + 0.13174141943454742, + 0.24923419952392578, + 0.08964525908231735, + 0.07882566004991531, + 0.04702857881784439, + 0.08534456789493561, + 0.1119585707783699, + 0.09575611352920532 + ], + "credit_map": [ + 0.00807760190218687, + 0.007799200247973204, + 0.009701021946966648, + 0.007051871158182621, + 0.006348197348415852, + 0.006281845737248659, + 0.006260107737034559, + 0.0068261269479990005, + 0.005845427047461271, + 0.00692765275016427, + 0.00804151501506567, + 0.00930541567504406, + 0.006739283446222544, + 0.004285569302737713, + 0.005305042490363121, + 0.005669776350259781, + 0.023458024486899376, + 0.026977716013789177, + 0.017305821180343628, + 0.008405101485550404, + 0.01489439606666565, + 0.012073013000190258, + 0.020033694803714752, + 0.008593647740781307, + 0.043671026825904846, + 0.04079660400748253, + 0.027257321402430534, + 0.018808843567967415, + 0.03683024272322655, + 0.028461791574954987, + 0.03418400138616562, + 0.0192243754863739, + 0.008523065596818924, + 0.00794430635869503, + 0.008898231200873852, + 0.008016714826226234, + 0.0145456213504076, + 0.009027467109262943, + 0.018664022907614708, + 0.014025832526385784, + 0.010507717728614807, + 0.009236241690814495, + 0.006491645704954863, + 0.007422047667205334, + 0.012070955708622932, + 0.005930619314312935, + 0.016380051150918007, + 0.01078637782484293, + 0.006705079227685928, + 0.008772186003625393, + 0.0034697328228503466, + 0.004081951454281807, + 0.005157042294740677, + 0.005098570138216019, + 0.00868974905461073, + 0.0050542703829705715, + 0.012364658527076244, + 0.030621085315942764, + 0.005830747075378895, + 0.007042739074677229, + 0.008667897433042526, + 0.007787207141518593, + 0.006350679788738489, + 0.006679553538560867, + 0.01306818425655365, + 0.034709759056568146, + 0.006189779844135046, + 0.009417014196515083, + 0.015446381643414497, + 0.008073553442955017, + 0.01602856256067753, + 0.0090253334492445, + 0.015169973485171795, + 0.017036905512213707, + 0.0074140592478215694, + 0.009872890077531338, + 0.013258741237223148, + 0.013321621343493462, + 0.008678928948938847, + 0.011002987623214722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 8 + ], + "selected_reward": -0.07899142801761627 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young fisherman composed entirely of 8-bit pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning in a slow arc around the subject from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lake surface composed of square blocks of pixels with rippling concentric circles in cyan and dark blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast water crown composed of translucent white and light blue pixels rising from the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a landscape painting composed of grayscale pixels with distinct peaks and crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058546058833599 + ], + "opposite_logprob_tokens": [ + -3.895585536956787 + ], + "logprob_reward": -0.02058546058833599, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058546058833599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21725717186927795, + 0.05837063491344452, + 0.05498513951897621, + 0.20693610608577728, + 0.10509097576141357, + 0.0671808198094368, + 0.04190054535865784, + 0.12424863874912262, + 0.08120150119066238, + 0.04282848909497261 + ], + "credit_map": [ + 0.024661201983690262, + 0.028468595817685127, + 0.019914431497454643, + 0.04503303021192551, + 0.02123369090259075, + 0.020213961601257324, + 0.03700825944542885, + 0.020723991096019745, + 0.008631370961666107, + 0.007634127978235483, + 0.010095411911606789, + 0.007549882400780916, + 0.009128604084253311, + 0.005232718773186207, + 0.005067423451691866, + 0.0050310916267335415, + 0.008287318050861359, + 0.00872714165598154, + 0.006459210999310017, + 0.004627407528460026, + 0.009737993590533733, + 0.006193863227963448, + 0.006535863503813744, + 0.004416335374116898, + 0.014610652811825275, + 0.045444779098033905, + 0.00882645882666111, + 0.01137643400579691, + 0.03287500888109207, + 0.016032151877880096, + 0.06494175642728806, + 0.012828848324716091, + 0.01557459868490696, + 0.015055998228490353, + 0.008632459677755833, + 0.009362154640257359, + 0.014864632859826088, + 0.009680955670773983, + 0.021155426278710365, + 0.010764744132757187, + 0.007907619699835777, + 0.007033077068626881, + 0.008806503377854824, + 0.008410974405705929, + 0.009058197028934956, + 0.0067057074047625065, + 0.009977483190596104, + 0.00928125437349081, + 0.005922728683799505, + 0.006310745142400265, + 0.009469466283917427, + 0.0036830264143645763, + 0.004115365911275148, + 0.004044872242957354, + 0.0041954778134822845, + 0.0041588605381548405, + 0.013932354748249054, + 0.028609035536646843, + 0.008470202796161175, + 0.010241124778985977, + 0.016909349709749222, + 0.011217018589377403, + 0.022879047319293022, + 0.011990499682724476, + 0.007513505406677723, + 0.021511845290660858, + 0.006441691890358925, + 0.009906399063766003, + 0.00865590013563633, + 0.005595795810222626, + 0.011924448423087597, + 0.009651907719671726, + 0.005363006144762039, + 0.008719997480511665, + 0.0037576048634946346, + 0.004334607627242804, + 0.005665652919560671, + 0.006749877706170082, + 0.004114854149520397, + 0.00412288261577487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.02058546058833599 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young fisherman composed entirely of 8-bit pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning in a slow arc around the subject from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lake surface composed of square blocks of pixels with rippling concentric circles in cyan and dark blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast water crown composed of translucent white and light blue pixels rising from the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a landscape painting composed of grayscale pixels with distinct peaks and crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005248576868325472 + ], + "opposite_logprob_tokens": [ + -5.255248546600342 + ], + "logprob_reward": -0.005248576868325472, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005248576868325472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16381777822971344, + 0.06012985110282898, + 0.08914494514465332, + 0.18920116126537323, + 0.07481969892978668, + 0.12519609928131104, + 0.07356120645999908, + 0.04865685850381851, + 0.079792819917202, + 0.0956796333193779 + ], + "credit_map": [ + 0.017668820917606354, + 0.020309243351221085, + 0.015896210446953773, + 0.033856675028800964, + 0.01631605252623558, + 0.017635608091950417, + 0.02562721259891987, + 0.016507942229509354, + 0.009335861541330814, + 0.009449580684304237, + 0.006454786751419306, + 0.009221667423844337, + 0.00966784916818142, + 0.00678494805470109, + 0.004449718631803989, + 0.004765441175550222, + 0.008418167941272259, + 0.0199118759483099, + 0.008155984804034233, + 0.008587769232690334, + 0.007118933368474245, + 0.006778184324502945, + 0.023746943101286888, + 0.0064270878210663795, + 0.029966052621603012, + 0.038032326847314835, + 0.022005600854754448, + 0.013724753633141518, + 0.0178392231464386, + 0.01861020363867283, + 0.03497232124209404, + 0.014050669968128204, + 0.008078672923147678, + 0.007354008965194225, + 0.006044653244316578, + 0.005891996435821056, + 0.016966089606285095, + 0.007244422100484371, + 0.013891467824578285, + 0.00934839528053999, + 0.01752488501369953, + 0.017169330269098282, + 0.013342485763132572, + 0.013149200938642025, + 0.016654126346111298, + 0.015161457471549511, + 0.017875272780656815, + 0.014319338835775852, + 0.009249945171177387, + 0.013468471355736256, + 0.005036140326410532, + 0.007720527704805136, + 0.009345946833491325, + 0.008175788447260857, + 0.013303153216838837, + 0.007261232938617468, + 0.00592006603255868, + 0.0114979799836874, + 0.0028698635287582874, + 0.004387576133012772, + 0.00625263387337327, + 0.004372928291559219, + 0.008408913388848305, + 0.004946892615407705, + 0.009886122308671474, + 0.02796776406466961, + 0.004877515137195587, + 0.0070215072482824326, + 0.009336821734905243, + 0.006110398098826408, + 0.008402276784181595, + 0.006190415937453508, + 0.013184425421059132, + 0.01681636832654476, + 0.007375749759376049, + 0.009665322490036488, + 0.01763840764760971, + 0.008990686386823654, + 0.011615857481956482, + 0.010392810218036175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.005248576868325472 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young fisherman composed entirely of 8-bit pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning in a slow arc around the subject from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lake surface composed of square blocks of pixels with rippling concentric circles in cyan and dark blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast water crown composed of translucent white and light blue pixels rising from the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a landscape painting composed of grayscale pixels with distinct peaks and crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08901743590831757 + ], + "opposite_logprob_tokens": [ + -2.464017391204834 + ], + "logprob_reward": -0.08901743590831757, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901743590831757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2715184688568115, + 0.05129196122288704, + 0.1404367834329605, + 0.09492944926023483, + 0.05033319443464279, + 0.04089532792568207, + 0.031221000477671623, + 0.09323149174451828, + 0.12575379014015198, + 0.1003885343670845 + ], + "credit_map": [ + 0.02943260222673416, + 0.036174144595861435, + 0.026591351255774498, + 0.05791928246617317, + 0.028023719787597656, + 0.02486753836274147, + 0.042930785566568375, + 0.025579027831554413, + 0.005818250123411417, + 0.006438190583139658, + 0.006673946976661682, + 0.006036227103322744, + 0.010531802661716938, + 0.006771682295948267, + 0.005209256894886494, + 0.0038126036524772644, + 0.01694304682314396, + 0.03319286182522774, + 0.0116830850020051, + 0.010119888931512833, + 0.013057909905910492, + 0.01262570172548294, + 0.03321705386042595, + 0.009597234427928925, + 0.01368595939129591, + 0.01744474656879902, + 0.007146215066313744, + 0.0071912300772964954, + 0.014236479066312313, + 0.009242958389222622, + 0.018386419862508774, + 0.007595438975840807, + 0.00920030102133751, + 0.00946058426052332, + 0.004739418625831604, + 0.004150883294641972, + 0.0071532889269292355, + 0.004792810417711735, + 0.0066370838321745396, + 0.004198824521154165, + 0.007515305187553167, + 0.006012427620589733, + 0.005754982586950064, + 0.003787255845963955, + 0.005216686986386776, + 0.003412805963307619, + 0.005727811250835657, + 0.0034680527169257402, + 0.003991198725998402, + 0.004755333997309208, + 0.003674770938232541, + 0.0028295672964304686, + 0.005315426737070084, + 0.0033944016322493553, + 0.004334484692662954, + 0.002925816923379898, + 0.00853997003287077, + 0.011599875055253506, + 0.008560790680348873, + 0.007809871807694435, + 0.016146117821335793, + 0.007303970400243998, + 0.022932492196559906, + 0.010338407009840012, + 0.015541990287601948, + 0.038718126714229584, + 0.009215343743562698, + 0.012518229894340038, + 0.014387042261660099, + 0.008288189768791199, + 0.015303480438888073, + 0.011781400069594383, + 0.01294595655053854, + 0.01770857721567154, + 0.009024763479828835, + 0.011094559915363789, + 0.01671506278216839, + 0.00731020700186491, + 0.014619702473282814, + 0.010969704017043114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.08901743590831757 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a signature bullet-time shot, the camera pans in an extremely slow arc around the subject from a slightly upward angle, freezing the world in a surreal moment. A young fisherman, composed entirely of 8-bit pixels, wears a simple blue pixelated headscarf and a plain dark brown sleeveless jacket, standing barefoot on the wet shore of a lake composed of square blocks of pixels. He leans back slightly at the spectacle before him, his hands tightly gripping a crude, equally pixelated fishing rod. His eyes, composed of a few simple dots, reveal utter concentration and awe. Beneath him, the rippling lake surface, in extreme slow motion, ripples in concentric circles of cyan and dark blue pixels, as if time has frozen. And in the distance, a landscape painting composed of grayscale pixels unfolds in the background, its distinct peaks and crests clearly defined, a perfect fusion of retro F-style gaming and classical Eastern aesthetics. The absolute focal point of the image, a vast water crown composed of translucent white and light blue pixels, rises majestically from the center of the lake at a speed almost imperceptible to the naked eye, surrounded by countless suspended pixelated droplets, shimmering with light. The entire scene is bathed in a soft, evenly diffused light, devoid of clear shadows, further emphasizing the flatness of the image and the unique artistic style, creating an awe-inspiring atmosphere that is both serene and full of epic fantasy.", + "seed": 151, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_4/index_9315_seed_151_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02400817535817623, + "vlm_logprob": -0.02400817535817623, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.02400817535817623, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young fisherman composed entirely of 8-bit pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning in a slow arc around the subject from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lake surface composed of square blocks of pixels with rippling concentric circles in cyan and dark blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast water crown composed of translucent white and light blue pixels rising from the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a landscape painting composed of grayscale pixels with distinct peaks and crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025235570967197418 + ], + "opposite_logprob_tokens": [ + -3.775235652923584 + ], + "logprob_reward": -0.025235570967197418, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.025235570967197418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22359681129455566, + 0.06203784793615341, + 0.06969794631004333, + 0.06777174770832062, + 0.14563582837581635, + 0.10713590681552887, + 0.10009261965751648, + 0.10818025469779968, + 0.06150888279080391, + 0.05434219166636467 + ], + "credit_map": [ + 0.03966079652309418, + 0.020984847098588943, + 0.02542962320148945, + 0.020855287089943886, + 0.023966435343027115, + 0.04762091115117073, + 0.010795715264976025, + 0.034283194690942764, + 0.011144613847136497, + 0.006094411946833134, + 0.004730524495244026, + 0.006054813042283058, + 0.010268849320709705, + 0.014087899588048458, + 0.004024145659059286, + 0.005632587242871523, + 0.010305619798600674, + 0.012367114424705505, + 0.006653934717178345, + 0.012586558237671852, + 0.009430209174752235, + 0.00886757206171751, + 0.004149687942117453, + 0.0053372494876384735, + 0.01627196930348873, + 0.007861408405005932, + 0.00371379591524601, + 0.0038600333500653505, + 0.007990519516170025, + 0.01900872215628624, + 0.005308567546308041, + 0.003756728023290634, + 0.04281594604253769, + 0.007083339616656303, + 0.010177000425755978, + 0.0049985246732831, + 0.04969293251633644, + 0.015737654641270638, + 0.006120437290519476, + 0.009010006673634052, + 0.015081575140357018, + 0.014889752492308617, + 0.013555705547332764, + 0.007637473288923502, + 0.015407121740281582, + 0.01376811508089304, + 0.012258423492312431, + 0.014537735842168331, + 0.012480230070650578, + 0.018769126385450363, + 0.011242440901696682, + 0.012997850775718689, + 0.01545424573123455, + 0.010525532066822052, + 0.00861022062599659, + 0.010012973099946976, + 0.016381651163101196, + 0.012387264519929886, + 0.025139084085822105, + 0.008080575615167618, + 0.014370727352797985, + 0.008528301492333412, + 0.008923879824578762, + 0.014368772506713867, + 0.008632505312561989, + 0.00550357811152935, + 0.009387912228703499, + 0.007785145193338394, + 0.007054193411022425, + 0.00782967172563076, + 0.008345462381839752, + 0.006970415823161602, + 0.006906543858349323, + 0.007251636125147343, + 0.008814716711640358, + 0.005124118644744158, + 0.006053050514310598, + 0.008919733576476574, + 0.003549323184415698, + 0.007723072078078985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.025235570967197418 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young fisherman composed entirely of 8-bit pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning in a slow arc around the subject from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lake surface composed of square blocks of pixels with rippling concentric circles in cyan and dark blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast water crown composed of translucent white and light blue pixels rising from the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a landscape painting composed of grayscale pixels with distinct peaks and crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898183912038803 + ], + "opposite_logprob_tokens": [ + -2.578981876373291 + ], + "logprob_reward": -0.07898183912038803, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898183912038803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08624426275491714, + 0.09085827320814133, + 0.03464847803115845, + 0.02471131645143032, + 0.1196400374174118, + 0.19666436314582825, + 0.17183689773082733, + 0.0874875858426094, + 0.14273491501808167, + 0.04517387971282005 + ], + "credit_map": [ + 0.012802837416529655, + 0.007576275151222944, + 0.006318851839751005, + 0.007612165063619614, + 0.021602807566523552, + 0.012762550264596939, + 0.007889547385275364, + 0.009679228998720646, + 0.010355676524341106, + 0.006219544447958469, + 0.015460084192454815, + 0.004454692360013723, + 0.011353754438459873, + 0.012593786232173443, + 0.006209042388945818, + 0.024211691692471504, + 0.0053663114085793495, + 0.0039462558925151825, + 0.004598353058099747, + 0.002916865050792694, + 0.004724171943962574, + 0.005872458219528198, + 0.0024406227748841047, + 0.004783439915627241, + 0.003389290301129222, + 0.00372106465511024, + 0.00356604834087193, + 0.002397495321929455, + 0.00337873469106853, + 0.0026416045147925615, + 0.0022724575828760862, + 0.003344620345160365, + 0.013505624607205391, + 0.008997646160423756, + 0.00782717950642109, + 0.03523220121860504, + 0.02610265091061592, + 0.010809550061821938, + 0.008372864685952663, + 0.008792323991656303, + 0.022835060954093933, + 0.023836366832256317, + 0.01903831586241722, + 0.01705775409936905, + 0.03555631637573242, + 0.017033066600561142, + 0.036754608154296875, + 0.024552876129746437, + 0.029275327920913696, + 0.021246738731861115, + 0.014424949884414673, + 0.013934231363236904, + 0.023913860321044922, + 0.021120429039001465, + 0.02872457727789879, + 0.019196778535842896, + 0.012791622430086136, + 0.011751890182495117, + 0.012803846038877964, + 0.006290574092417955, + 0.014615085907280445, + 0.01508278213441372, + 0.005787397734820843, + 0.008364378474652767, + 0.02336948551237583, + 0.018297648057341576, + 0.03131534159183502, + 0.010907195508480072, + 0.018597770482301712, + 0.0153582114726305, + 0.01185691263526678, + 0.013032340444624424, + 0.007078682538121939, + 0.004137021955102682, + 0.010636302642524242, + 0.0032264438923448324, + 0.004816113505512476, + 0.0058409119956195354, + 0.0032879470381885767, + 0.006150456145405769 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 8 + ], + "selected_reward": -0.07898183912038803 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young fisherman composed entirely of 8-bit pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning in a slow arc around the subject from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lake surface composed of square blocks of pixels with rippling concentric circles in cyan and dark blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast water crown composed of translucent white and light blue pixels rising from the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a landscape painting composed of grayscale pixels with distinct peaks and crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005254980642348528 + ], + "opposite_logprob_tokens": [ + -5.255254745483398 + ], + "logprob_reward": -0.005254980642348528, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005254980642348528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14137180149555206, + 0.06556985527276993, + 0.04216436296701431, + 0.05751967802643776, + 0.10443534702062607, + 0.21646662056446075, + 0.058685269206762314, + 0.2393346130847931, + 0.04240082949399948, + 0.032051656395196915 + ], + "credit_map": [ + 0.019511409103870392, + 0.02165810950100422, + 0.019143322482705116, + 0.015384175814688206, + 0.019015898928046227, + 0.018857654184103012, + 0.010572603903710842, + 0.017228620126843452, + 0.01146077923476696, + 0.005245371256023645, + 0.004939772188663483, + 0.005082862917333841, + 0.009134265594184399, + 0.012673473916947842, + 0.007169501855969429, + 0.009863835759460926, + 0.0054938835091888905, + 0.006139695644378662, + 0.0041368259117007256, + 0.004571075085550547, + 0.007802940439432859, + 0.005570041015744209, + 0.004056078847497702, + 0.00439382204785943, + 0.012315806932747364, + 0.005328791681677103, + 0.007915946654975414, + 0.0038499587681144476, + 0.006562248338013887, + 0.011669525876641273, + 0.003645153483375907, + 0.006232252344489098, + 0.01480069849640131, + 0.01458485797047615, + 0.009457840584218502, + 0.009557671844959259, + 0.01736532337963581, + 0.01832052506506443, + 0.008885380811989307, + 0.01146305538713932, + 0.03291945904493332, + 0.02432204596698284, + 0.041487980633974075, + 0.01651545986533165, + 0.028260499238967896, + 0.023812301456928253, + 0.01854194886982441, + 0.03060692548751831, + 0.009609730914235115, + 0.007731229066848755, + 0.004763656295835972, + 0.0076324474066495895, + 0.008424491621553898, + 0.008177481591701508, + 0.006565034855157137, + 0.00578119745478034, + 0.02787252888083458, + 0.024028368294239044, + 0.04001590982079506, + 0.0199214369058609, + 0.03621499985456467, + 0.034582801163196564, + 0.02578045055270195, + 0.030918119475245476, + 0.0062382216565310955, + 0.007862081751227379, + 0.005272686947137117, + 0.004957820288836956, + 0.005688051227480173, + 0.004429282620549202, + 0.004966273903846741, + 0.002986411564052105, + 0.005246141459792852, + 0.008101352490484715, + 0.0023622922599315643, + 0.0026123926509171724, + 0.0045897443778812885, + 0.003625298384577036, + 0.003099017310887575, + 0.002415415830910206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.005254980642348528 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young fisherman composed entirely of 8-bit pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning in a slow arc around the subject from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lake surface composed of square blocks of pixels with rippling concentric circles in cyan and dark blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast water crown composed of translucent white and light blue pixels rising from the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a landscape painting composed of grayscale pixels with distinct peaks and crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019374187104403973 + ], + "opposite_logprob_tokens": [ + -6.251937389373779 + ], + "logprob_reward": -0.0019374187104403973, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019374187104403973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23829826712608337, + 0.10322558879852295, + 0.10196886211633682, + 0.10127890110015869, + 0.20315559208393097, + 0.07364553958177567, + 0.057505037635564804, + 0.04493333026766777, + 0.035267774015665054, + 0.04072103649377823 + ], + "credit_map": [ + 0.03750808537006378, + 0.021422872319817543, + 0.01501146238297224, + 0.01831384189426899, + 0.06611866503953934, + 0.03353898972272873, + 0.026239143684506416, + 0.020145215094089508, + 0.013524439185857773, + 0.005307411774992943, + 0.014348633587360382, + 0.008114482276141644, + 0.009681514464318752, + 0.01625896990299225, + 0.008378265425562859, + 0.027611883357167244, + 0.014621112495660782, + 0.008983319625258446, + 0.010157999582588673, + 0.006270003039389849, + 0.022450963035225868, + 0.01435056608170271, + 0.011549131013453007, + 0.01358577236533165, + 0.01882169395685196, + 0.010906726121902466, + 0.0085652656853199, + 0.006270320154726505, + 0.01729574054479599, + 0.020288584753870964, + 0.009974479675292969, + 0.00915609486401081, + 0.024557549506425858, + 0.013898544944822788, + 0.014142274856567383, + 0.06301883608102798, + 0.03804805874824524, + 0.018674151971936226, + 0.01638655923306942, + 0.014429640024900436, + 0.009454349987208843, + 0.008260668255388737, + 0.008671291172504425, + 0.005462562199681997, + 0.013494369573891163, + 0.012188306078314781, + 0.005986352916806936, + 0.010127647779881954, + 0.007104598451405764, + 0.008001569658517838, + 0.004277796950191259, + 0.006470204330980778, + 0.009639358147978783, + 0.0077460226602852345, + 0.008325797505676746, + 0.005939692724496126, + 0.003908513579517603, + 0.00703911017626524, + 0.0033100550062954426, + 0.005765559617429972, + 0.007031201384961605, + 0.005799849517643452, + 0.00813867524266243, + 0.003940367605537176, + 0.004095117561519146, + 0.0038133624475449324, + 0.004487925209105015, + 0.0049523850902915, + 0.004062286112457514, + 0.00295091699808836, + 0.00748251099139452, + 0.0034232717007398605, + 0.006561571732163429, + 0.0092609953135252, + 0.0036136358976364136, + 0.002797273686155677, + 0.005234575364738703, + 0.004887450952082872, + 0.004548666067421436, + 0.003816870041191578 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.0019374187104403973 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young fisherman composed entirely of 8-bit pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning in a slow arc around the subject from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lake surface composed of square blocks of pixels with rippling concentric circles in cyan and dark blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast water crown composed of translucent white and light blue pixels rising from the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a landscape painting composed of grayscale pixels with distinct peaks and crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008631067350506783 + ], + "opposite_logprob_tokens": [ + -4.758631229400635 + ], + "logprob_reward": -0.008631067350506783, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008631067350506783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08909787237644196, + 0.06546717882156372, + 0.02797541208565235, + 0.016072465106844902, + 0.1475161463022232, + 0.31082385778427124, + 0.1203528419137001, + 0.10102761536836624, + 0.06893353909254074, + 0.052733104676008224 + ], + "credit_map": [ + 0.012646356597542763, + 0.0108266556635499, + 0.008864745497703552, + 0.009090950712561607, + 0.019429657608270645, + 0.014361448585987091, + 0.006653067655861378, + 0.007224993780255318, + 0.008131002075970173, + 0.0042549134232103825, + 0.009462915360927582, + 0.004228753037750721, + 0.006150068249553442, + 0.009656572714447975, + 0.00609453534707427, + 0.017488418146967888, + 0.005020538344979286, + 0.0030241855420172215, + 0.0027926552575081587, + 0.002213242929428816, + 0.005508351139724255, + 0.003689988050609827, + 0.0019680564291775227, + 0.003758395556360483, + 0.002652121242135763, + 0.0018957393476739526, + 0.0025148296263068914, + 0.0017052326584234834, + 0.0023574538063257933, + 0.0017640021396800876, + 0.0015285792760550976, + 0.0016545072430744767, + 0.02938215062022209, + 0.010302946902811527, + 0.009738944470882416, + 0.037124153226614, + 0.022538835182785988, + 0.017710264772176743, + 0.008847789838910103, + 0.011871067807078362, + 0.0374908410012722, + 0.029783088713884354, + 0.030179498717188835, + 0.02570972964167595, + 0.05678357556462288, + 0.03980904072523117, + 0.050091296434402466, + 0.04097678139805794, + 0.022424643859267235, + 0.014404810033738613, + 0.011360826902091503, + 0.008742272853851318, + 0.019086889922618866, + 0.017156362533569336, + 0.011495614424347878, + 0.015681417658925056, + 0.013615391217172146, + 0.01062772423028946, + 0.021563341841101646, + 0.0070321811363101006, + 0.014305414631962776, + 0.01456222590059042, + 0.006587917450815439, + 0.01273341104388237, + 0.009502002038061619, + 0.006244536489248276, + 0.009232758544385433, + 0.006948454771190882, + 0.008371866308152676, + 0.012596176005899906, + 0.010014048777520657, + 0.006023695692420006, + 0.010159461759030819, + 0.007274507079273462, + 0.0064096758142113686, + 0.004126464948058128, + 0.008145195432007313, + 0.010266479104757309, + 0.002860153093934059, + 0.003491166979074478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.008631067350506783 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a signature bullet-time shot, the camera pans in an extremely slow arc around the subject from a slightly upward angle, freezing the world in a surreal moment. A young fisherman, composed entirely of 8-bit pixels, wears a simple blue pixelated headscarf and a plain dark brown sleeveless jacket, standing barefoot on the wet shore of a lake composed of square blocks of pixels. He leans back slightly at the spectacle before him, his hands tightly gripping a crude, equally pixelated fishing rod. His eyes, composed of a few simple dots, reveal utter concentration and awe. Beneath him, the rippling lake surface, in extreme slow motion, ripples in concentric circles of cyan and dark blue pixels, as if time has frozen. And in the distance, a landscape painting composed of grayscale pixels unfolds in the background, its distinct peaks and crests clearly defined, a perfect fusion of retro F-style gaming and classical Eastern aesthetics. The absolute focal point of the image, a vast water crown composed of translucent white and light blue pixels, rises majestically from the center of the lake at a speed almost imperceptible to the naked eye, surrounded by countless suspended pixelated droplets, shimmering with light. The entire scene is bathed in a soft, evenly diffused light, devoid of clear shadows, further emphasizing the flatness of the image and the unique artistic style, creating an awe-inspiring atmosphere that is both serene and full of epic fantasy.", + "seed": 152, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_4/index_9315_seed_152_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.14180989377200603, + "vlm_logprob": -0.14180989377200603, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.14180989377200603, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young fisherman composed entirely of 8-bit pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning in a slow arc around the subject from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lake surface composed of square blocks of pixels with rippling concentric circles in cyan and dark blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast water crown composed of translucent white and light blue pixels rising from the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a landscape painting composed of grayscale pixels with distinct peaks and crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08296442031860352 + ], + "opposite_logprob_tokens": [ + -2.5829644203186035 + ], + "logprob_reward": -0.08296442031860352, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08296442031860352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1739499419927597, + 0.12109272181987762, + 0.07101032882928848, + 0.11532764136791229, + 0.24913442134857178, + 0.08412651717662811, + 0.04678051546216011, + 0.0341760478913784, + 0.03556763380765915, + 0.06883420050144196 + ], + "credit_map": [ + 0.013624066486954689, + 0.02660258300602436, + 0.018030552193522453, + 0.01670147106051445, + 0.021751442924141884, + 0.03341587632894516, + 0.026980554684996605, + 0.016843395307660103, + 0.008340772241353989, + 0.012866051867604256, + 0.009374653920531273, + 0.011841855011880398, + 0.016521336510777473, + 0.016152940690517426, + 0.02446536161005497, + 0.021529758349061012, + 0.007410458754748106, + 0.012298958376049995, + 0.010327481664717197, + 0.00820060446858406, + 0.00725594675168395, + 0.014667630195617676, + 0.0057137892581522465, + 0.005135455168783665, + 0.014447152614593506, + 0.020785773172974586, + 0.012819120660424232, + 0.013653104193508625, + 0.011585434898734093, + 0.021941035985946655, + 0.009974999353289604, + 0.010121013969182968, + 0.018820999190211296, + 0.029829077422618866, + 0.016805453225970268, + 0.03679409623146057, + 0.02804143726825714, + 0.04911302030086517, + 0.0337456651031971, + 0.03598468378186226, + 0.009383314289152622, + 0.011956263333559036, + 0.008076601661741734, + 0.007816753350198269, + 0.008863327093422413, + 0.01570410653948784, + 0.013798690401017666, + 0.008527466095983982, + 0.004677336197346449, + 0.005379412788897753, + 0.003689745906740427, + 0.004151812754571438, + 0.004353680182248354, + 0.007281886879354715, + 0.011543367058038712, + 0.005703272297978401, + 0.0032510810997337103, + 0.0043668244034051895, + 0.0030250141862779856, + 0.003540544304996729, + 0.005726622883230448, + 0.006615121383219957, + 0.004094342235475779, + 0.003556497162207961, + 0.0017736110603436828, + 0.0061392271891236305, + 0.0020158193074166775, + 0.0031451231334358454, + 0.0034520195331424475, + 0.007851953618228436, + 0.006935214623808861, + 0.00425466476008296, + 0.00701477425172925, + 0.007501460146158934, + 0.006813045125454664, + 0.006171870045363903, + 0.00750643340870738, + 0.010740336962044239, + 0.011308813467621803, + 0.011777467094361782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.08296442031860352 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young fisherman composed entirely of 8-bit pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning in a slow arc around the subject from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lake surface composed of square blocks of pixels with rippling concentric circles in cyan and dark blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast water crown composed of translucent white and light blue pixels rising from the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a landscape painting composed of grayscale pixels with distinct peaks and crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233173370361328 + ], + "opposite_logprob_tokens": [ + -0.8983173370361328 + ], + "logprob_reward": -0.5233173370361328, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233173370361328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4436880946159363, + 0.058820437639951706, + 0.05071709305047989, + 0.046817000955343246, + 0.056920126080513, + 0.07473805546760559, + 0.10038866847753525, + 0.04829986020922661, + 0.0600801520049572, + 0.059530556201934814 + ], + "credit_map": [ + 0.03423251584172249, + 0.08291199058294296, + 0.028314916417002678, + 0.04551343619823456, + 0.05377347022294998, + 0.1154780238866806, + 0.050571370869874954, + 0.03289235383272171, + 0.004658325109630823, + 0.008186946623027325, + 0.004035327583551407, + 0.00495578208938241, + 0.006595678627490997, + 0.011553315445780754, + 0.012245520018041134, + 0.0065895430743694305, + 0.004522969014942646, + 0.007545709144324064, + 0.007752429693937302, + 0.005563946906477213, + 0.006004232447594404, + 0.006996415555477142, + 0.006913357879966497, + 0.005418032873421907, + 0.006199862342327833, + 0.005536012351512909, + 0.004826339427381754, + 0.006417455617338419, + 0.00569845549762249, + 0.009576080366969109, + 0.004499438218772411, + 0.004063356202095747, + 0.005623689386993647, + 0.009855378419160843, + 0.005217631813138723, + 0.006627602968364954, + 0.006142479833215475, + 0.01358044147491455, + 0.005164044443517923, + 0.004708858206868172, + 0.009529792703688145, + 0.007477785460650921, + 0.0074351755902171135, + 0.007268026936799288, + 0.013806323520839214, + 0.014618395827710629, + 0.007154425140470266, + 0.007448134012520313, + 0.015066801570355892, + 0.009892493486404419, + 0.010495483875274658, + 0.008792269043624401, + 0.014301965944468975, + 0.018246306106448174, + 0.010990530252456665, + 0.012602821923792362, + 0.0037918146699666977, + 0.00462005939334631, + 0.0030651164706796408, + 0.004332184791564941, + 0.009054357185959816, + 0.008708473294973373, + 0.006197903770953417, + 0.00852995179593563, + 0.0071457973681390285, + 0.009584066458046436, + 0.00537529680877924, + 0.006197944283485413, + 0.006924366112798452, + 0.010812342166900635, + 0.008086987771093845, + 0.005953351501375437, + 0.004214343149214983, + 0.006186508107930422, + 0.00487905228510499, + 0.004157640505582094, + 0.007388180587440729, + 0.008935132995247841, + 0.015835361555218697, + 0.007934335619211197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.5233173370361328 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young fisherman composed entirely of 8-bit pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning in a slow arc around the subject from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lake surface composed of square blocks of pixels with rippling concentric circles in cyan and dark blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast water crown composed of translucent white and light blue pixels rising from the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a landscape painting composed of grayscale pixels with distinct peaks and crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011083701625466347 + ], + "opposite_logprob_tokens": [ + -4.511083602905273 + ], + "logprob_reward": -0.011083701625466347, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011083701625466347, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.217542365193367, + 0.06068294495344162, + 0.05636320635676384, + 0.06162821501493454, + 0.10824387520551682, + 0.0813789889216423, + 0.07772990316152573, + 0.2342323362827301, + 0.048648178577423096, + 0.053550027310848236 + ], + "credit_map": [ + 0.013577241450548172, + 0.028960853815078735, + 0.01847682148218155, + 0.023659538477659225, + 0.024266289547085762, + 0.045207429677248, + 0.035502053797245026, + 0.027892127633094788, + 0.004780111368745565, + 0.008046303875744343, + 0.005176916718482971, + 0.006201514042913914, + 0.007841765880584717, + 0.009395799599587917, + 0.009800917468965054, + 0.009439609944820404, + 0.004526739940047264, + 0.006825676187872887, + 0.006409148685634136, + 0.005686737596988678, + 0.009591437876224518, + 0.006770034786313772, + 0.00874180905520916, + 0.007811625022441149, + 0.003958144225180149, + 0.0061010438948869705, + 0.0028823758475482464, + 0.006753048859536648, + 0.011906411498785019, + 0.01155828032642603, + 0.009226197376847267, + 0.009242711588740349, + 0.011752022430300713, + 0.012758072465658188, + 0.01098588202148676, + 0.014508426189422607, + 0.01217235904186964, + 0.022690871730446815, + 0.011287241242825985, + 0.012089001014828682, + 0.005969902966171503, + 0.008970251306891441, + 0.004548892844468355, + 0.008870074525475502, + 0.013139131478965282, + 0.019564183428883553, + 0.01181481871753931, + 0.008501734584569931, + 0.0047410777769982815, + 0.0072091612964868546, + 0.004624313209205866, + 0.007732521742582321, + 0.010997667908668518, + 0.009564788080751896, + 0.020470671355724335, + 0.012389695271849632, + 0.0151986014097929, + 0.03381834924221039, + 0.014212162233889103, + 0.018421992659568787, + 0.036177270114421844, + 0.047237664461135864, + 0.03778140991926193, + 0.03138488903641701, + 0.002543424256145954, + 0.007500595413148403, + 0.0024182007182389498, + 0.005323946475982666, + 0.004916749428957701, + 0.009822538122534752, + 0.009181976318359375, + 0.006940743885934353, + 0.0033561033196747303, + 0.004448410123586655, + 0.004029818344861269, + 0.005169361364096403, + 0.005693134386092424, + 0.009722309187054634, + 0.013787254691123962, + 0.007343634031713009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.011083701625466347 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young fisherman composed entirely of 8-bit pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning in a slow arc around the subject from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lake surface composed of square blocks of pixels with rippling concentric circles in cyan and dark blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast water crown composed of translucent white and light blue pixels rising from the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a landscape painting composed of grayscale pixels with distinct peaks and crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430445596575737 + ], + "opposite_logprob_tokens": [ + -3.1680445671081543 + ], + "logprob_reward": -0.0430445596575737, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430445596575737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20840409398078918, + 0.05085043981671333, + 0.07526037842035294, + 0.07603617012500763, + 0.1539694368839264, + 0.0998634621500969, + 0.07176470756530762, + 0.09545613825321198, + 0.05828122794628143, + 0.11011393368244171 + ], + "credit_map": [ + 0.013766073621809483, + 0.02329131029546261, + 0.014336924999952316, + 0.02446308732032776, + 0.024900559335947037, + 0.035840436816215515, + 0.04073198512196541, + 0.031073734164237976, + 0.0060741836205124855, + 0.00787478219717741, + 0.004991110879927874, + 0.005069749429821968, + 0.006054997444152832, + 0.009946400299668312, + 0.006631793919950724, + 0.004207420162856579, + 0.005501586012542248, + 0.00846517737954855, + 0.005766882561147213, + 0.006986348424106836, + 0.011114748194813728, + 0.015131048858165741, + 0.013465572148561478, + 0.008829016238451004, + 0.00552884628996253, + 0.011176954954862595, + 0.005506337620317936, + 0.0066378177143633366, + 0.013597306795418262, + 0.01630043238401413, + 0.008212982676923275, + 0.009075495414435863, + 0.016175217926502228, + 0.014320802874863148, + 0.01798287220299244, + 0.022559067234396935, + 0.015840962529182434, + 0.034281160682439804, + 0.017838306725025177, + 0.014971050433814526, + 0.006284994538873434, + 0.01104586012661457, + 0.004520770162343979, + 0.008870676159858704, + 0.016979830339550972, + 0.01991361193358898, + 0.018804941326379776, + 0.013442774303257465, + 0.005361981224268675, + 0.011455770581960678, + 0.004918013233691454, + 0.007893621921539307, + 0.0064149703830480576, + 0.011342527344822884, + 0.016863353550434113, + 0.007514472119510174, + 0.006677589844912291, + 0.011007609777152538, + 0.0061432309448719025, + 0.009321597404778004, + 0.011193587444722652, + 0.016806386411190033, + 0.017836350947618484, + 0.016469787806272507, + 0.0037294887006282806, + 0.005230329930782318, + 0.0033720694482326508, + 0.006783123593777418, + 0.007103921379894018, + 0.008642120286822319, + 0.011524925008416176, + 0.011895245872437954, + 0.00525609590113163, + 0.006862304173409939, + 0.0049963416531682014, + 0.009355937130749226, + 0.01308835856616497, + 0.019446194171905518, + 0.021833419799804688, + 0.029275277629494667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.0430445596575737 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young fisherman composed entirely of 8-bit pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning in a slow arc around the subject from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lake surface composed of square blocks of pixels with rippling concentric circles in cyan and dark blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast water crown composed of translucent white and light blue pixels rising from the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a landscape painting composed of grayscale pixels with distinct peaks and crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0486394502222538 + ], + "opposite_logprob_tokens": [ + -3.0486395359039307 + ], + "logprob_reward": -0.0486394502222538, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0486394502222538, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17037075757980347, + 0.09866691380739212, + 0.05242929607629776, + 0.1042286604642868, + 0.08048216253519058, + 0.0818982869386673, + 0.05299801379442215, + 0.07632482796907425, + 0.044105786830186844, + 0.23849526047706604 + ], + "credit_map": [ + 0.017226409167051315, + 0.027278581634163857, + 0.015540806576609612, + 0.019694529473781586, + 0.01727510243654251, + 0.0328584648668766, + 0.023913465440273285, + 0.016583403572440147, + 0.005901432130485773, + 0.011732690967619419, + 0.007676570676267147, + 0.010509810410439968, + 0.011464040726423264, + 0.018654853105545044, + 0.016359670087695122, + 0.016367845237255096, + 0.0035827509127557278, + 0.006733718793839216, + 0.00391374621540308, + 0.004472027532756329, + 0.007741904351860285, + 0.012465914711356163, + 0.007882366888225079, + 0.005636867601424456, + 0.007526481989771128, + 0.01522012334316969, + 0.005561992060393095, + 0.010957663878798485, + 0.011374080553650856, + 0.025348186492919922, + 0.016422906890511513, + 0.011817234568297863, + 0.006042000837624073, + 0.006854567211121321, + 0.005778808146715164, + 0.007184601854532957, + 0.010136873461306095, + 0.01378999650478363, + 0.01731972210109234, + 0.013375590555369854, + 0.007933109998703003, + 0.008450274355709553, + 0.006331592798233032, + 0.007596853654831648, + 0.010197225958108902, + 0.015169232152402401, + 0.01645500585436821, + 0.00976499728858471, + 0.0032787167001515627, + 0.0076568336226046085, + 0.0031257166992872953, + 0.004619287792593241, + 0.005910404957830906, + 0.010012423619627953, + 0.012666645459830761, + 0.0057279858738183975, + 0.005508258938789368, + 0.006560466252267361, + 0.004898622166365385, + 0.005491958931088448, + 0.010374327190220356, + 0.010971352458000183, + 0.014567234553396702, + 0.017952609807252884, + 0.0025616490747779608, + 0.006213309243321419, + 0.002177994465455413, + 0.00452388683333993, + 0.004137860145419836, + 0.010050795041024685, + 0.007684290874749422, + 0.0067560020834207535, + 0.012459415011107922, + 0.032699063420295715, + 0.020349010825157166, + 0.016600560396909714, + 0.026405856013298035, + 0.05777961388230324, + 0.043389782309532166, + 0.02881196327507496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.0486394502222538 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_5/index_tensor([9887, 9887])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_5/index_tensor([9887, 9887])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..927a525947d1e13aa0396b4a2aa5f7430383a368 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_5/index_tensor([9887, 9887])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, extreme close-up captures this moment of suspended tension, focusing on a metal coin, its thin, reedy rim impossibly balanced. Made of a slightly worn, silvery metal, the coin catches the soft, aimless sunlight filtering through the car window, casting a thin, bright beam of light against the dark, textured surface of the vehicle's dashboard. Subtle, almost imperceptible vibrations from the car's engine ripple through the dashboard, causing the coin to begin its erratic oscillation. The light on its rim flickers erratically as it sways back and forth, its latent energy poised to be unleashed. Suddenly, it succumbs to gravity, falling to the ground with a sharp, decisive motion, and instantly erupting into a rapid spin. The coin's surface transforms into a dazzling, hypnotic blur of reflected light, a shimmering silver disk humming against the textured surface. The rotation gradually slows, the frantic whirring becoming a slower, rhythmic sway, until, after one final, languid rotation, the coin comes to a complete stillness, lying flat and motionless. The camera’s tight focus now reveals the worn, smooth relief of the coin’s surface, its rounded shape now stable and fully visible, bringing the scene from a state of instability to a quietly dynamic resolution.", + "seed": 3950, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_5/index_9887_seed_3950_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.2461428977549076, + "vlm_logprob": -0.2461428977549076, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.2461428977549076, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a metal coin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin made of slightly worn, silvery metal with a thin, reedy rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show soft sunlight filtering through a car window onto a dark, textured dashboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the coin fall to the ground and spin rapidly before slowing to a complete stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tightly focused on the coin's surface details throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012898819521069527 + ], + "opposite_logprob_tokens": [ + -4.887898921966553 + ], + "logprob_reward": -0.012898819521069527, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.012898819521069527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1980033963918686, + 0.1257956176996231, + 0.2476232796907425, + 0.061685074120759964, + 0.18015354871749878, + 0.028652368113398552, + 0.044659268110990524, + 0.04794766381382942, + 0.030278421938419342, + 0.03520135208964348 + ], + "credit_map": [ + 0.019403841346502304, + 0.03151686117053032, + 0.02799135632812977, + 0.04217895492911339, + 0.02844715490937233, + 0.010917752049863338, + 0.02452254481613636, + 0.01302492804825306, + 0.012014053761959076, + 0.010487205348908901, + 0.014579764567315578, + 0.02194567769765854, + 0.014992954209446907, + 0.013982590287923813, + 0.018577957525849342, + 0.01921541802585125, + 0.028812488541007042, + 0.022579971700906754, + 0.032264746725559235, + 0.03613404557108879, + 0.027617264539003372, + 0.03371739387512207, + 0.04728764295578003, + 0.01920972391963005, + 0.006920153740793467, + 0.006744051817804575, + 0.006236023735255003, + 0.010568363592028618, + 0.007497046142816544, + 0.007320280186831951, + 0.009524393826723099, + 0.006874762009829283, + 0.019115054979920387, + 0.019053276628255844, + 0.019587744027376175, + 0.030739804729819298, + 0.023271365091204643, + 0.020052624866366386, + 0.025076329708099365, + 0.02325734682381153, + 0.003143398789688945, + 0.0026903818361461163, + 0.002917755860835314, + 0.0051495470106601715, + 0.00400770315900445, + 0.0038148777093738317, + 0.004167591221630573, + 0.0027611134573817253, + 0.005834019277244806, + 0.0039774393662810326, + 0.005003230646252632, + 0.009598994627594948, + 0.003908946178853512, + 0.005494986195117235, + 0.005870925262570381, + 0.004970721900463104, + 0.004972315393388271, + 0.0034675574861466885, + 0.005665105767548084, + 0.01281120628118515, + 0.004002667032182217, + 0.004984287079423666, + 0.007093873340636492, + 0.0049506486393511295, + 0.0029051259625703096, + 0.0025304188020527363, + 0.0037413742393255234, + 0.007545240223407745, + 0.0026353769935667515, + 0.002558516338467598, + 0.004598280414938927, + 0.0037640873342752457, + 0.004037370905280113, + 0.004512212239205837, + 0.004292014986276627, + 0.006555596832185984, + 0.0025694346986711025, + 0.004281135741621256, + 0.005620882846415043, + 0.003332704771310091 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.012898819521069527 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a metal coin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin made of slightly worn, silvery metal with a thin, reedy rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show soft sunlight filtering through a car window onto a dark, textured dashboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the coin fall to the ground and spin rapidly before slowing to a complete stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tightly focused on the coin's surface details throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0206286683678627 + ], + "opposite_logprob_tokens": [ + -3.8956286907196045 + ], + "logprob_reward": -0.0206286683678627, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.0206286683678627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3521775007247925, + 0.1129763051867485, + 0.1193540170788765, + 0.10517711937427521, + 0.059738386422395706, + 0.043166954070329666, + 0.03800550848245621, + 0.0468747541308403, + 0.053716398775577545, + 0.06881296634674072 + ], + "credit_map": [ + 0.03679719939827919, + 0.043111663311719894, + 0.04923997446894646, + 0.06548058241605759, + 0.06034661829471588, + 0.0294465571641922, + 0.03981652483344078, + 0.027938392013311386, + 0.01228845864534378, + 0.011581914499402046, + 0.013842303305864334, + 0.020458504557609558, + 0.011156965047121048, + 0.012663722969591618, + 0.019032644107937813, + 0.011951794847846031, + 0.014182817190885544, + 0.012038854882121086, + 0.015900298953056335, + 0.015681466087698936, + 0.012345820665359497, + 0.016783438622951508, + 0.022160915657877922, + 0.010260415263473988, + 0.012554838322103024, + 0.01336906012147665, + 0.012148448266088963, + 0.014380548149347305, + 0.009931032545864582, + 0.014830457977950573, + 0.016561359167099, + 0.011401394382119179, + 0.00644255056977272, + 0.007922224700450897, + 0.00628137169405818, + 0.007415300235152245, + 0.00799302477389574, + 0.0084625743329525, + 0.008252873085439205, + 0.00696847029030323, + 0.0048736101016402245, + 0.004843016620725393, + 0.0045460923574864864, + 0.007442132569849491, + 0.0050010476261377335, + 0.005121367052197456, + 0.0065167974680662155, + 0.004822894930839539, + 0.0035206875763833523, + 0.004626999143511057, + 0.0036764205433428288, + 0.005445756949484348, + 0.004882865119725466, + 0.00526693370193243, + 0.0063783577643334866, + 0.004207491409033537, + 0.004318293649703264, + 0.003937899135053158, + 0.00576895521953702, + 0.012316195294260979, + 0.003752400167286396, + 0.004361387342214584, + 0.006357940379530191, + 0.0060616848058998585, + 0.004832920618355274, + 0.004423678852617741, + 0.006274545565247536, + 0.013955636881291866, + 0.004572531674057245, + 0.006450590677559376, + 0.0070803225971758366, + 0.006126180291175842, + 0.007667646277695894, + 0.007803508546203375, + 0.008317451924085617, + 0.014780113473534584, + 0.004687744192779064, + 0.0086726825684309, + 0.009419561363756657, + 0.007464263588190079 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0206286683678627 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a metal coin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin made of slightly worn, silvery metal with a thin, reedy rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show soft sunlight filtering through a car window onto a dark, textured dashboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the coin fall to the ground and spin rapidly before slowing to a complete stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tightly focused on the coin's surface details throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02637348137795925 + ], + "opposite_logprob_tokens": [ + -3.6513733863830566 + ], + "logprob_reward": -0.02637348137795925, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02637348137795925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1556004285812378, + 0.2878083884716034, + 0.058070629835128784, + 0.1440112590789795, + 0.09170667082071304, + 0.05030917376279831, + 0.06256767362356186, + 0.048908524215221405, + 0.0776764526963234, + 0.023340806365013123 + ], + "credit_map": [ + 0.022129451856017113, + 0.01788334548473358, + 0.021115658804774284, + 0.025835571810603142, + 0.025086909532546997, + 0.01346284057945013, + 0.01886541210114956, + 0.011221233755350113, + 0.021725116297602654, + 0.028202038258314133, + 0.028150824829936028, + 0.03984849900007248, + 0.03504178673028946, + 0.03177438676357269, + 0.06592079997062683, + 0.03714493662118912, + 0.004956628195941448, + 0.005823937244713306, + 0.007284774444997311, + 0.008578629232943058, + 0.006326226983219385, + 0.008105738088488579, + 0.011188212782144547, + 0.0058064828626811504, + 0.019916174933314323, + 0.025212906301021576, + 0.01616615056991577, + 0.017957033589482307, + 0.016509709879755974, + 0.017518069595098495, + 0.01601817086338997, + 0.014713042415678501, + 0.009069359861314297, + 0.01053400058299303, + 0.010339444503188133, + 0.01961054466664791, + 0.008732127957046032, + 0.00875468272715807, + 0.013726999051868916, + 0.010939509607851505, + 0.006242275703698397, + 0.005172611679881811, + 0.005773329641669989, + 0.010844009928405285, + 0.0035805024672299623, + 0.00437965290620923, + 0.008850131183862686, + 0.005466657690703869, + 0.007297835312783718, + 0.00804729200899601, + 0.006364880129694939, + 0.010257530957460403, + 0.006695238407701254, + 0.00846927985548973, + 0.008950002491474152, + 0.006485617719590664, + 0.004985501058399677, + 0.004520429763942957, + 0.0060255941934883595, + 0.012843160890042782, + 0.0033403276465833187, + 0.004522108472883701, + 0.00660064397379756, + 0.006070754956454039, + 0.007551579270511866, + 0.008056244812905788, + 0.008569931611418724, + 0.018704621121287346, + 0.004720158409327269, + 0.008618577383458614, + 0.012081890366971493, + 0.009373456239700317, + 0.0023694485425949097, + 0.0029256220441311598, + 0.0025021762121468782, + 0.004409788176417351, + 0.0018386669689789414, + 0.002514903200790286, + 0.0038275381084531546, + 0.00295266299508512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.02637348137795925 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a metal coin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin made of slightly worn, silvery metal with a thin, reedy rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show soft sunlight filtering through a car window onto a dark, textured dashboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the coin fall to the ground and spin rapidly before slowing to a complete stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tightly focused on the coin's surface details throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1370928287506104 + ], + "opposite_logprob_tokens": [ + -0.38709282875061035 + ], + "logprob_reward": -1.1370928287506104, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1370928287506104, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.14177511632442474, + 0.26512181758880615, + 0.1420820951461792, + 0.11129484325647354, + 0.12141405045986176, + 0.022436970844864845, + 0.03642364591360092, + 0.0945659726858139, + 0.05236095190048218, + 0.012524539604783058 + ], + "credit_map": [ + 0.01702866144478321, + 0.02439025044441223, + 0.01623690500855446, + 0.025372184813022614, + 0.0193698238581419, + 0.009054237976670265, + 0.014899412170052528, + 0.015423636883497238, + 0.022876638919115067, + 0.029492121189832687, + 0.02803131379187107, + 0.04114113375544548, + 0.030253607779741287, + 0.023884210735559464, + 0.046052057296037674, + 0.043390724807977676, + 0.01354338601231575, + 0.013866967521607876, + 0.01888621598482132, + 0.021634528413414955, + 0.019328244030475616, + 0.024099668487906456, + 0.015885891392827034, + 0.014837191440165043, + 0.012523704208433628, + 0.013408897444605827, + 0.013662599958479404, + 0.018879296258091927, + 0.00907824095338583, + 0.01417918037623167, + 0.01953127793967724, + 0.010031637735664845, + 0.010690733790397644, + 0.013458583503961563, + 0.011721963994204998, + 0.021346643567085266, + 0.015498444437980652, + 0.013075301423668861, + 0.01875564642250538, + 0.01686673052608967, + 0.0027765214908868074, + 0.002120926510542631, + 0.0026829151902347803, + 0.004711988382041454, + 0.002105779480189085, + 0.0026155717205256224, + 0.0029595254454761744, + 0.0024637409951537848, + 0.0035968946758657694, + 0.004602792672812939, + 0.003711204044520855, + 0.005583032499998808, + 0.004584503825753927, + 0.0044179935939610004, + 0.0056647793389856815, + 0.004262447357177734, + 0.010022644884884357, + 0.006420821417123079, + 0.011799720115959644, + 0.026993805542588234, + 0.007506328169256449, + 0.009861310012638569, + 0.013036930933594704, + 0.008924413472414017, + 0.006581902038305998, + 0.0063246628269553185, + 0.006367107853293419, + 0.011058653704822063, + 0.003305447520688176, + 0.006766463164240122, + 0.006352622527629137, + 0.005604089703410864, + 0.0011586560867726803, + 0.0016851926920935512, + 0.0013350853696465492, + 0.002048602094873786, + 0.0015262948581948876, + 0.0017345737433061004, + 0.0017844079993665218, + 0.0012517266441136599 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -1.1370928287506104 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a metal coin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin made of slightly worn, silvery metal with a thin, reedy rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show soft sunlight filtering through a car window onto a dark, textured dashboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the coin fall to the ground and spin rapidly before slowing to a complete stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tightly focused on the coin's surface details throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372069075703621 + ], + "opposite_logprob_tokens": [ + -3.4087207317352295 + ], + "logprob_reward": -0.03372069075703621, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372069075703621, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1519043892621994, + 0.14639824628829956, + 0.2138998806476593, + 0.07065096497535706, + 0.14055848121643066, + 0.061362095177173615, + 0.059446871280670166, + 0.048456352204084396, + 0.037938348948955536, + 0.06938439607620239 + ], + "credit_map": [ + 0.021518396213650703, + 0.02321012318134308, + 0.01645437255501747, + 0.018269775435328484, + 0.02473059855401516, + 0.011653640307486057, + 0.020258810371160507, + 0.015808679163455963, + 0.012696747668087482, + 0.01160021498799324, + 0.017316246405243874, + 0.025165461003780365, + 0.014422880485653877, + 0.018548404797911644, + 0.023732254281640053, + 0.022916018962860107, + 0.02369684725999832, + 0.01914103887975216, + 0.02548753283917904, + 0.032459646463394165, + 0.023158324882388115, + 0.02811228111386299, + 0.04377785697579384, + 0.018066350370645523, + 0.009782622568309307, + 0.009652873501181602, + 0.00981153268367052, + 0.008745863102376461, + 0.005038857460021973, + 0.009605162777006626, + 0.01120740920305252, + 0.006806649267673492, + 0.017218448221683502, + 0.01633044146001339, + 0.012921790592372417, + 0.01890173926949501, + 0.02296440117061138, + 0.022191312164068222, + 0.016054678708314896, + 0.01397568080574274, + 0.007470841985195875, + 0.005822138395160437, + 0.006074937991797924, + 0.010590539313852787, + 0.0052796038798987865, + 0.0070409020408988, + 0.011555907316505909, + 0.007527222856879234, + 0.006300116889178753, + 0.005330377724021673, + 0.007085006218403578, + 0.01414477825164795, + 0.006278316956013441, + 0.006420456804335117, + 0.008253799751400948, + 0.005634022410959005, + 0.005884543992578983, + 0.005942835472524166, + 0.004809246864169836, + 0.006705807987600565, + 0.006713175680488348, + 0.008020861074328423, + 0.005342594347894192, + 0.005037287715822458, + 0.0036928392946720123, + 0.004101413767784834, + 0.004245285876095295, + 0.006458702962845564, + 0.004835478961467743, + 0.004731890745460987, + 0.005517072509974241, + 0.004355662502348423, + 0.009680173359811306, + 0.010202636942267418, + 0.007843353785574436, + 0.009869306348264217, + 0.006416555494070053, + 0.010463973507285118, + 0.007779579143971205, + 0.007128810044378042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.03372069075703621 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up captures this moment of suspended tension, focusing on a metal coin, its thin, reedy rim impossibly balanced. Made of a slightly worn, silvery metal, the coin catches the soft, aimless sunlight filtering through the car window, casting a thin, bright beam of light against the dark, textured surface of the vehicle's dashboard. Subtle, almost imperceptible vibrations from the car's engine ripple through the dashboard, causing the coin to begin its erratic oscillation. The light on its rim flickers erratically as it sways back and forth, its latent energy poised to be unleashed. Suddenly, it succumbs to gravity, falling to the ground with a sharp, decisive motion, and instantly erupting into a rapid spin. The coin's surface transforms into a dazzling, hypnotic blur of reflected light, a shimmering silver disk humming against the textured surface. The rotation gradually slows, the frantic whirring becoming a slower, rhythmic sway, until, after one final, languid rotation, the coin comes to a complete stillness, lying flat and motionless. The camera’s tight focus now reveals the worn, smooth relief of the coin’s surface, its rounded shape now stable and fully visible, bringing the scene from a state of instability to a quietly dynamic resolution.", + "seed": 3951, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_5/index_9887_seed_3951_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11931390818208457, + "vlm_logprob": -0.11931390818208457, + "vlm_yes_no_margin": 3.0999999940395355, + "vlm_token_credit": -0.11931390818208457, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a metal coin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin made of slightly worn, silvery metal with a thin, reedy rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show soft sunlight filtering through a car window onto a dark, textured dashboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the coin fall to the ground and spin rapidly before slowing to a complete stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tightly focused on the coin's surface details throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010067871771752834 + ], + "opposite_logprob_tokens": [ + -5.135067939758301 + ], + "logprob_reward": -0.010067871771752834, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.010067871771752834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2533262073993683, + 0.08967086672782898, + 0.05595692619681358, + 0.14224445819854736, + 0.06186281144618988, + 0.1145804300904274, + 0.14859266579151154, + 0.06611447036266327, + 0.03606167435646057, + 0.03158942237496376 + ], + "credit_map": [ + 0.021771583706140518, + 0.0612151175737381, + 0.016040699556469917, + 0.0178326815366745, + 0.03146037086844444, + 0.03470408916473389, + 0.03512978181242943, + 0.03517190366983414, + 0.01023182924836874, + 0.009930712170898914, + 0.0079638147726655, + 0.010035677812993526, + 0.013919515535235405, + 0.013848381116986275, + 0.010631893761456013, + 0.01310905534774065, + 0.0053187282755970955, + 0.006912509445101023, + 0.0033104459289461374, + 0.007199668325483799, + 0.009457806125283241, + 0.00976442638784647, + 0.00800260528922081, + 0.005990740843117237, + 0.008694708347320557, + 0.03064202517271042, + 0.00911460816860199, + 0.008519768714904785, + 0.05010002478957176, + 0.01506663765758276, + 0.010359585285186768, + 0.00974710937589407, + 0.006135586183518171, + 0.012242994271218777, + 0.006445631850510836, + 0.005166006740182638, + 0.0181540846824646, + 0.004383524879813194, + 0.004328558221459389, + 0.005006427876651287, + 0.0043192896991968155, + 0.026697849854826927, + 0.0022017876617610455, + 0.0039602541364729404, + 0.05142384395003319, + 0.01603805646300316, + 0.00617225980386138, + 0.00376709527336061, + 0.0076005044393241405, + 0.02023501694202423, + 0.004796430002897978, + 0.005629161838442087, + 0.058585625141859055, + 0.021410644054412842, + 0.009623776189982891, + 0.020711518824100494, + 0.0073297396302223206, + 0.007306233048439026, + 0.009150984697043896, + 0.009032031521201134, + 0.011681687086820602, + 0.006794564891606569, + 0.0066524227149784565, + 0.008166814222931862, + 0.0035359642934054136, + 0.008609184063971043, + 0.00210984842851758, + 0.0029260669834911823, + 0.005267741624265909, + 0.004521039314568043, + 0.003138010622933507, + 0.005953824147582054, + 0.001660635694861412, + 0.0073133669793605804, + 0.0016743912128731608, + 0.0019084911327809095, + 0.010149667970836163, + 0.002995881484821439, + 0.0019198090303689241, + 0.003967178054153919 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.010067871771752834 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a metal coin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin made of slightly worn, silvery metal with a thin, reedy rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show soft sunlight filtering through a car window onto a dark, textured dashboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the coin fall to the ground and spin rapidly before slowing to a complete stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tightly focused on the coin's surface details throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03374316543340683 + ], + "opposite_logprob_tokens": [ + -3.408743143081665 + ], + "logprob_reward": -0.03374316543340683, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03374316543340683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32570162415504456, + 0.12468955665826797, + 0.04802922531962395, + 0.06339576840400696, + 0.10584111511707306, + 0.10677657276391983, + 0.0674324780702591, + 0.0659196600317955, + 0.06830981373786926, + 0.023904167115688324 + ], + "credit_map": [ + 0.026470400393009186, + 0.08524274826049805, + 0.014381937682628632, + 0.02156790718436241, + 0.05624149739742279, + 0.051661062985658646, + 0.03164263442158699, + 0.03849343955516815, + 0.011521752923727036, + 0.012032978236675262, + 0.008657545782625675, + 0.014134682714939117, + 0.022414881736040115, + 0.020621703937649727, + 0.014083489775657654, + 0.021222516894340515, + 0.00601053424179554, + 0.004994363058358431, + 0.0021978833246976137, + 0.004111551214009523, + 0.006751831155270338, + 0.006789090111851692, + 0.007234708871692419, + 0.009939267300069332, + 0.007615828420966864, + 0.00812057126313448, + 0.00959219690412283, + 0.007299262098968029, + 0.011722265742719173, + 0.00813106819987297, + 0.0044695534743368626, + 0.006445019040256739, + 0.011220847256481647, + 0.019137326627969742, + 0.012646881863474846, + 0.010490648448467255, + 0.02913704328238964, + 0.006922748405486345, + 0.008893251419067383, + 0.007392365951091051, + 0.014318480156362057, + 0.008499841205775738, + 0.004357051569968462, + 0.006868070922791958, + 0.04131714254617691, + 0.01727369800209999, + 0.00779248820617795, + 0.006349798757582903, + 0.00485272565856576, + 0.0040608798153698444, + 0.003058217465877533, + 0.004134636837989092, + 0.0218682661652565, + 0.011818249709904194, + 0.005506063345819712, + 0.012133440002799034, + 0.005237935110926628, + 0.021795935928821564, + 0.0021384544670581818, + 0.0027427654713392258, + 0.00491291843354702, + 0.0049838596023619175, + 0.0043801795691251755, + 0.019727610051631927, + 0.004921832587569952, + 0.014093938283622265, + 0.003691131714731455, + 0.004049005452543497, + 0.008742057718336582, + 0.006608061026781797, + 0.006233204156160355, + 0.019970586523413658, + 0.00156898214481771, + 0.0054508401080966, + 0.0010463972575962543, + 0.0015998847084119916, + 0.007730148732662201, + 0.0023418853525072336, + 0.0015615031588822603, + 0.0026045243721455336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.03374316543340683 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a metal coin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin made of slightly worn, silvery metal with a thin, reedy rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show soft sunlight filtering through a car window onto a dark, textured dashboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the coin fall to the ground and spin rapidly before slowing to a complete stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tightly focused on the coin's surface details throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981787547469139 + ], + "opposite_logprob_tokens": [ + -3.529817819595337 + ], + "logprob_reward": -0.02981787547469139, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981787547469139, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.132942333817482, + 0.034121073782444, + 0.03263598307967186, + 0.03671281039714813, + 0.04813506826758385, + 0.5238078832626343, + 0.12229545414447784, + 0.027503933757543564, + 0.02444528602063656, + 0.017400190234184265 + ], + "credit_map": [ + 0.011619086377322674, + 0.022313013672828674, + 0.007396943401545286, + 0.007505560293793678, + 0.031847771257162094, + 0.02301272377371788, + 0.01779783144593239, + 0.011449404060840607, + 0.00329425442032516, + 0.003894599387422204, + 0.0022852118127048016, + 0.00352354533970356, + 0.005825468804687262, + 0.005095412954688072, + 0.004781703930348158, + 0.00542087946087122, + 0.00331281335093081, + 0.003025757148861885, + 0.00193586980458349, + 0.003207453293725848, + 0.00431084306910634, + 0.004831482190638781, + 0.004532686900347471, + 0.007479075808078051, + 0.0030634645372629166, + 0.004558860789984465, + 0.0025213591288775206, + 0.00386179075576365, + 0.008278018794953823, + 0.00424701115116477, + 0.005329990293830633, + 0.004852315876632929, + 0.004561221227049828, + 0.009086028672754765, + 0.005819089710712433, + 0.0040383110754191875, + 0.014654613099992275, + 0.0036446067970246077, + 0.003561519319191575, + 0.002769678831100464, + 0.022018996998667717, + 0.12559393048286438, + 0.009181835688650608, + 0.015373327769339085, + 0.24200008809566498, + 0.07566039264202118, + 0.021224183961749077, + 0.012755083851516247, + 0.009994310326874256, + 0.02807682380080223, + 0.010473779402673244, + 0.012468129396438599, + 0.0225770715624094, + 0.011110592633485794, + 0.01267918199300766, + 0.01491557713598013, + 0.0028692097403109074, + 0.010329841636121273, + 0.0015857249964028597, + 0.0014800784410908818, + 0.003561027580872178, + 0.0022498599719256163, + 0.0017214404651895165, + 0.0037067504599690437, + 0.00233739847317338, + 0.00461760675534606, + 0.0021386058069765568, + 0.0023044792469590902, + 0.0029006979893893003, + 0.0026257848367094994, + 0.0022369602229446173, + 0.005283753387629986, + 0.0016111882869154215, + 0.004134821239858866, + 0.001996927661821246, + 0.002003660425543785, + 0.003128419630229473, + 0.0014144639717414975, + 0.0017985804006457329, + 0.0013121276861056685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.02981787547469139 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a metal coin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin made of slightly worn, silvery metal with a thin, reedy rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show soft sunlight filtering through a car window onto a dark, textured dashboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the coin fall to the ground and spin rapidly before slowing to a complete stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tightly focused on the coin's surface details throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47427108883857727 + ], + "opposite_logprob_tokens": [ + -0.9742710590362549 + ], + "logprob_reward": -0.47427108883857727, + "yes_no_margin_reward": 0.4999999701976776, + "token_credit_reward": -0.47427108883857727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07161235809326172, + 0.0497514046728611, + 0.0251923818141222, + 0.05333548039197922, + 0.03074084222316742, + 0.6762872338294983, + 0.031426385045051575, + 0.013331119902431965, + 0.023435229435563087, + 0.024887608364224434 + ], + "credit_map": [ + 0.004985163453966379, + 0.01578768528997898, + 0.003226711880415678, + 0.004790941718965769, + 0.016301514580845833, + 0.012048491276800632, + 0.0065934485755860806, + 0.007878406904637814, + 0.005106883589178324, + 0.00513501139357686, + 0.001970672747120261, + 0.004786469042301178, + 0.009707548655569553, + 0.00910568330436945, + 0.004923103842884302, + 0.009016032330691814, + 0.003330521984025836, + 0.0021889638155698776, + 0.0018111575627699494, + 0.0026928973384201527, + 0.002943199360743165, + 0.0036991473753005266, + 0.004234530963003635, + 0.004291963297873735, + 0.006366870366036892, + 0.007644069846719503, + 0.007455980405211449, + 0.0058158207684755325, + 0.009081660769879818, + 0.006725454702973366, + 0.004505047108978033, + 0.005740576423704624, + 0.003171845804899931, + 0.004895012825727463, + 0.0025021368637681007, + 0.003438635729253292, + 0.006062607746571302, + 0.0026043918915092945, + 0.004061790648847818, + 0.004004421643912792, + 0.02334209717810154, + 0.18781830370426178, + 0.009380845353007317, + 0.01952945627272129, + 0.3009149432182312, + 0.08896850794553757, + 0.03072737157344818, + 0.015605646185576916, + 0.0025853365659713745, + 0.003567527048289776, + 0.0025789858773350716, + 0.0030242290813475847, + 0.009618584997951984, + 0.0041367425583302975, + 0.0023491971660405397, + 0.0035657782573252916, + 0.0010958209168165922, + 0.003506458830088377, + 0.0008956189849413931, + 0.0009657896007411182, + 0.0014453585026785731, + 0.0011869051959365606, + 0.001262792618945241, + 0.0029723751358687878, + 0.002285040682181716, + 0.0042592547833919525, + 0.002333175390958786, + 0.002201042603701353, + 0.0033936118707060814, + 0.0032354409340769053, + 0.0019372513052076101, + 0.0037904134951531887, + 0.0021120065357536077, + 0.00899291131645441, + 0.0013475241139531136, + 0.0015560899628326297, + 0.006960878148674965, + 0.001478554680943489, + 0.0014949800679460168, + 0.0009446623735129833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.47427108883857727 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a metal coin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin made of slightly worn, silvery metal with a thin, reedy rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show soft sunlight filtering through a car window onto a dark, textured dashboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the coin fall to the ground and spin rapidly before slowing to a complete stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tightly focused on the coin's surface details throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048669539391994476 + ], + "opposite_logprob_tokens": [ + -3.0486695766448975 + ], + "logprob_reward": -0.048669539391994476, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048669539391994476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09488017112016678, + 0.022708985954523087, + 0.03345627337694168, + 0.026111746206879616, + 0.01581708714365959, + 0.6888709664344788, + 0.07574928551912308, + 0.017638742923736572, + 0.012341213412582874, + 0.012425526976585388 + ], + "credit_map": [ + 0.006041284184902906, + 0.029737258329987526, + 0.003365399083122611, + 0.00582604855298996, + 0.017289217561483383, + 0.015487496741116047, + 0.007589157670736313, + 0.009544305503368378, + 0.0026633937377482653, + 0.0026179866399616003, + 0.0015507758362218738, + 0.002697812393307686, + 0.003752458607777953, + 0.003939545247703791, + 0.0022536979522556067, + 0.0032333151903003454, + 0.004426808096468449, + 0.0036215013824403286, + 0.0016834429698064923, + 0.00273844413459301, + 0.0038398527540266514, + 0.005266990512609482, + 0.0053841732442379, + 0.006495062727481127, + 0.002472355030477047, + 0.0037231328897178173, + 0.0031881448812782764, + 0.002314035315066576, + 0.006907033268362284, + 0.002579910447821021, + 0.0020326043013483286, + 0.0028945328667759895, + 0.001696408842690289, + 0.0027808400336652994, + 0.001619296963326633, + 0.001596634043380618, + 0.004005866590887308, + 0.0011124302400276065, + 0.001643551280722022, + 0.0013620591489598155, + 0.029218275099992752, + 0.1669681966304779, + 0.012111728079617023, + 0.019999850541353226, + 0.3186938166618347, + 0.09884625673294067, + 0.025995735079050064, + 0.017037099227309227, + 0.003116405801847577, + 0.010467289946973324, + 0.002408742206171155, + 0.0027383703272789717, + 0.028586771339178085, + 0.009875845164060593, + 0.005306118633598089, + 0.013249742798507214, + 0.0016364871989935637, + 0.0017909331945702434, + 0.0018171494593843818, + 0.0018189632100984454, + 0.003876138012856245, + 0.001902144867926836, + 0.0018490506336092949, + 0.0029478773940354586, + 0.0010992774041369557, + 0.0031478579621762037, + 0.0006764933350495994, + 0.0009634027373977005, + 0.0013170363381505013, + 0.0011261567706242204, + 0.0012416248209774494, + 0.002769363345578313, + 0.0008084552828222513, + 0.003632064675912261, + 0.0005326633690856397, + 0.000616198405623436, + 0.004363640211522579, + 0.0011697049485519528, + 0.0006855370011180639, + 0.0006172623252496123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.048669539391994476 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up captures this moment of suspended tension, focusing on a metal coin, its thin, reedy rim impossibly balanced. Made of a slightly worn, silvery metal, the coin catches the soft, aimless sunlight filtering through the car window, casting a thin, bright beam of light against the dark, textured surface of the vehicle's dashboard. Subtle, almost imperceptible vibrations from the car's engine ripple through the dashboard, causing the coin to begin its erratic oscillation. The light on its rim flickers erratically as it sways back and forth, its latent energy poised to be unleashed. Suddenly, it succumbs to gravity, falling to the ground with a sharp, decisive motion, and instantly erupting into a rapid spin. The coin's surface transforms into a dazzling, hypnotic blur of reflected light, a shimmering silver disk humming against the textured surface. The rotation gradually slows, the frantic whirring becoming a slower, rhythmic sway, until, after one final, languid rotation, the coin comes to a complete stillness, lying flat and motionless. The camera’s tight focus now reveals the worn, smooth relief of the coin’s surface, its rounded shape now stable and fully visible, bringing the scene from a state of instability to a quietly dynamic resolution.", + "seed": 3952, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_5/index_9887_seed_3952_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.14392641168087722, + "vlm_logprob": -0.14392641168087722, + "vlm_yes_no_margin": 3.324999952316284, + "vlm_token_credit": -0.14392641168087722, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a metal coin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin made of slightly worn, silvery metal with a thin, reedy rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show soft sunlight filtering through a car window onto a dark, textured dashboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the coin fall to the ground and spin rapidly before slowing to a complete stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tightly focused on the coin's surface details throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009901338256895542 + ], + "opposite_logprob_tokens": [ + -5.384901523590088 + ], + "logprob_reward": -0.009901338256895542, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.009901338256895542, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09697574377059937, + 0.11101894080638885, + 0.1192270815372467, + 0.11432554572820663, + 0.05483158677816391, + 0.15196643769741058, + 0.12153912335634232, + 0.03698538988828659, + 0.05789508298039436, + 0.1352350115776062 + ], + "credit_map": [ + 0.011303468607366085, + 0.009041506797075272, + 0.009129729121923447, + 0.024782974272966385, + 0.009091793559491634, + 0.013029311783611774, + 0.011704501695930958, + 0.008892460726201534, + 0.012015817686915398, + 0.014461357146501541, + 0.011627450585365295, + 0.011775558814406395, + 0.017244303598999977, + 0.013515275903046131, + 0.019235052168369293, + 0.011144128628075123, + 0.014551373198628426, + 0.017092255875468254, + 0.010190949775278568, + 0.00858943723142147, + 0.020435629412531853, + 0.01321443822234869, + 0.022638779133558273, + 0.012514225207269192, + 0.00990197155624628, + 0.0335313156247139, + 0.006956322584301233, + 0.004266906995326281, + 0.041955042630434036, + 0.007126929238438606, + 0.004577029030770063, + 0.0060100313276052475, + 0.00981244258582592, + 0.006889098323881626, + 0.004075611941516399, + 0.004833161365240812, + 0.004476605914533138, + 0.006302256137132645, + 0.008493887260556221, + 0.00994852464646101, + 0.02488700859248638, + 0.02639508806169033, + 0.013293042778968811, + 0.014590131118893623, + 0.01775338687002659, + 0.017143575474619865, + 0.019846903160214424, + 0.01805729977786541, + 0.019508130848407745, + 0.013946321792900562, + 0.011313136667013168, + 0.009092682041227818, + 0.013808955438435078, + 0.02449120208621025, + 0.012033744715154171, + 0.017344942316412926, + 0.004681657999753952, + 0.004790164530277252, + 0.003933017607778311, + 0.004137503914535046, + 0.004178938921540976, + 0.004938472528010607, + 0.005295835435390472, + 0.00502979988232255, + 0.009991822764277458, + 0.006952331867069006, + 0.005881421267986298, + 0.004819132387638092, + 0.006500313989818096, + 0.009377566166222095, + 0.006471764296293259, + 0.007900730706751347, + 0.019490724429488182, + 0.03084472008049488, + 0.01303893979638815, + 0.007363729644566774, + 0.00837882049381733, + 0.009808734059333801, + 0.014976908452808857, + 0.031332436949014664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 6 + ], + "selected_reward": -0.009901338256895542 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a metal coin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin made of slightly worn, silvery metal with a thin, reedy rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show soft sunlight filtering through a car window onto a dark, textured dashboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the coin fall to the ground and spin rapidly before slowing to a complete stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tightly focused on the coin's surface details throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020615706220269203 + ], + "opposite_logprob_tokens": [ + -3.895615816116333 + ], + "logprob_reward": -0.020615706220269203, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020615706220269203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2392246574163437, + 0.14266186952590942, + 0.1518053114414215, + 0.07786142081022263, + 0.050239499658346176, + 0.06506291776895523, + 0.06796121597290039, + 0.04541698098182678, + 0.05880656838417053, + 0.10095955431461334 + ], + "credit_map": [ + 0.027049392461776733, + 0.028451090678572655, + 0.02151576615869999, + 0.05754062533378601, + 0.025930846109986305, + 0.026159318163990974, + 0.02826087363064289, + 0.024316731840372086, + 0.020747803151607513, + 0.01808841899037361, + 0.017746280878782272, + 0.010715212672948837, + 0.021747702732682228, + 0.02271556854248047, + 0.013036772608757019, + 0.01786411926150322, + 0.019781067967414856, + 0.026493733748793602, + 0.01129318866878748, + 0.009779556654393673, + 0.022175326943397522, + 0.014040032401680946, + 0.029536020010709763, + 0.01870638318359852, + 0.007845074869692326, + 0.022832179442048073, + 0.007212466560304165, + 0.0033574262633919716, + 0.01582704298198223, + 0.005994458682835102, + 0.008783727884292603, + 0.006009042728692293, + 0.009958825074136257, + 0.007489894982427359, + 0.004555860999971628, + 0.00397670641541481, + 0.006181572563946247, + 0.006905939429998398, + 0.0057764514349401, + 0.005394250154495239, + 0.010281730443239212, + 0.010456337593495846, + 0.007037360221147537, + 0.00712882773950696, + 0.005742842797189951, + 0.00751461461186409, + 0.00892071146517992, + 0.007980492897331715, + 0.011619583703577518, + 0.009418429806828499, + 0.00520929554477334, + 0.005303364247083664, + 0.006556044798344374, + 0.012252682819962502, + 0.007201099302619696, + 0.01040071714669466, + 0.006869837641716003, + 0.00614941818639636, + 0.005422867368906736, + 0.0032431676518172026, + 0.003707987256348133, + 0.005690355319529772, + 0.005954418331384659, + 0.008378926664590836, + 0.011680806055665016, + 0.007140891160815954, + 0.01055222935974598, + 0.005079853348433971, + 0.004605855327099562, + 0.008160057477653027, + 0.005398291628807783, + 0.006188583560287952, + 0.012204750441014767, + 0.028066249564290047, + 0.007308309897780418, + 0.005450243130326271, + 0.006757641211152077, + 0.007043775636702776, + 0.008588656783103943, + 0.025539927184581757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.020615706220269203 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a metal coin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin made of slightly worn, silvery metal with a thin, reedy rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show soft sunlight filtering through a car window onto a dark, textured dashboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the coin fall to the ground and spin rapidly before slowing to a complete stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tightly focused on the coin's surface details throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018200531601905823 + ], + "opposite_logprob_tokens": [ + -4.018200397491455 + ], + "logprob_reward": -0.018200531601905823, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018200531601905823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16840091347694397, + 0.16016051173210144, + 0.10561402887105942, + 0.20554743707180023, + 0.09918931871652603, + 0.059863973408937454, + 0.039647676050662994, + 0.06931308656930923, + 0.055683642625808716, + 0.0365794412791729 + ], + "credit_map": [ + 0.017593270167708397, + 0.023186838254332542, + 0.017277684062719345, + 0.03697025030851364, + 0.018518520519137383, + 0.020994704216718674, + 0.018223922699689865, + 0.015635721385478973, + 0.020669467747211456, + 0.01819971576333046, + 0.016130831092596054, + 0.01721879094839096, + 0.023175952956080437, + 0.027094272896647453, + 0.021244347095489502, + 0.016427138820290565, + 0.01396680623292923, + 0.012305382639169693, + 0.011974833905696869, + 0.009995241649448872, + 0.009503964334726334, + 0.010441116988658905, + 0.019888147711753845, + 0.017538532614707947, + 0.015122600831091404, + 0.06414461135864258, + 0.012571103870868683, + 0.008799776434898376, + 0.07103312760591507, + 0.011503234505653381, + 0.01053525134921074, + 0.01183775532990694, + 0.01646278239786625, + 0.013537066988646984, + 0.007036743685603142, + 0.009396540001034737, + 0.014009020291268826, + 0.012174728326499462, + 0.012760904617607594, + 0.013811536133289337, + 0.009396829642355442, + 0.007483487017452717, + 0.006373594049364328, + 0.005179069470614195, + 0.007409334182739258, + 0.008693214505910873, + 0.006819224916398525, + 0.008509216830134392, + 0.00763322040438652, + 0.0060322037898004055, + 0.003604308934882283, + 0.00368771655485034, + 0.0034772963263094425, + 0.0063817305490374565, + 0.0034529510885477066, + 0.005378244444727898, + 0.01681653968989849, + 0.01224302314221859, + 0.00842689722776413, + 0.004230697173625231, + 0.006368143018335104, + 0.009181405417621136, + 0.004489724058657885, + 0.0075566526502370834, + 0.0075285970233380795, + 0.011632734909653664, + 0.004986196756362915, + 0.003487125737592578, + 0.005570727400481701, + 0.006891252938657999, + 0.003659409238025546, + 0.011927601881325245, + 0.005666542332619429, + 0.004614261444658041, + 0.005373657215386629, + 0.0032921384554356337, + 0.003103205468505621, + 0.005054184701293707, + 0.00517949229106307, + 0.004295958671718836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.018200531601905823 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a metal coin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin made of slightly worn, silvery metal with a thin, reedy rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show soft sunlight filtering through a car window onto a dark, textured dashboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the coin fall to the ground and spin rapidly before slowing to a complete stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tightly focused on the coin's surface details throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328042149543762 + ], + "opposite_logprob_tokens": [ + -0.7578042149543762 + ], + "logprob_reward": -0.6328042149543762, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328042149543762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17727446556091309, + 0.14749369025230408, + 0.15137481689453125, + 0.13267314434051514, + 0.07782547175884247, + 0.11708372831344604, + 0.09545581042766571, + 0.032430678606033325, + 0.05111590772867203, + 0.017272241413593292 + ], + "credit_map": [ + 0.02179194614291191, + 0.016518792137503624, + 0.016227180138230324, + 0.03479650989174843, + 0.016787752509117126, + 0.03276042640209198, + 0.017045224085450172, + 0.021346647292375565, + 0.017581705003976822, + 0.020343896001577377, + 0.014289157465100288, + 0.011964188888669014, + 0.023311223834753036, + 0.02545938827097416, + 0.01891951821744442, + 0.015624609775841236, + 0.018409719690680504, + 0.02115342579782009, + 0.014403050765395164, + 0.012525227852165699, + 0.021556111052632332, + 0.014064247719943523, + 0.031101010739803314, + 0.01816203072667122, + 0.01206178031861782, + 0.03556537255644798, + 0.009165367111563683, + 0.00733081204816699, + 0.04484454169869423, + 0.009453393518924713, + 0.006588639225810766, + 0.007663242518901825, + 0.011700830422341824, + 0.009345305152237415, + 0.007120051886886358, + 0.01162642240524292, + 0.007586646359413862, + 0.006410214118659496, + 0.011623225174844265, + 0.012412777170538902, + 0.015143242664635181, + 0.016791297122836113, + 0.013343393802642822, + 0.01237679086625576, + 0.011286083608865738, + 0.011752313934266567, + 0.021575238555669785, + 0.014815361239016056, + 0.016612691804766655, + 0.011504738591611385, + 0.007840468548238277, + 0.007525420282036066, + 0.00823213905096054, + 0.019271837547421455, + 0.00909245852380991, + 0.015376063995063305, + 0.00494245532900095, + 0.0038427645340561867, + 0.003815487027168274, + 0.003318725386634469, + 0.0033942516893148422, + 0.0046780784614384174, + 0.004255334846675396, + 0.004183582495898008, + 0.01023432519286871, + 0.00577560206875205, + 0.006036772392690182, + 0.004205356817692518, + 0.005644527729600668, + 0.007360332645475864, + 0.00542413629591465, + 0.006434853654354811, + 0.0028135369066148996, + 0.0022053811699151993, + 0.0013810533564537764, + 0.0019905592780560255, + 0.0015344723360612988, + 0.002699334640055895, + 0.0025582979433238506, + 0.0020896049682050943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.6328042149543762 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a metal coin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin made of slightly worn, silvery metal with a thin, reedy rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show soft sunlight filtering through a car window onto a dark, textured dashboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the coin fall to the ground and spin rapidly before slowing to a complete stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tightly focused on the coin's surface details throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038110267370939255 + ], + "opposite_logprob_tokens": [ + -3.2881102561950684 + ], + "logprob_reward": -0.038110267370939255, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038110267370939255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23312512040138245, + 0.096766896545887, + 0.11845486611127853, + 0.0903039276599884, + 0.07695559412240982, + 0.06635502725839615, + 0.08147861808538437, + 0.03254666551947594, + 0.060966167598962784, + 0.14304712414741516 + ], + "credit_map": [ + 0.020299352705478668, + 0.028384758159518242, + 0.02704789489507675, + 0.07004944235086441, + 0.020309317857027054, + 0.01994824968278408, + 0.03082004003226757, + 0.016266070306301117, + 0.012972414493560791, + 0.014503926038742065, + 0.011651174165308475, + 0.010946347378194332, + 0.01219185907393694, + 0.012931236997246742, + 0.010770519264042377, + 0.010799424722790718, + 0.010688047856092453, + 0.01725505106151104, + 0.01022220216691494, + 0.009485790506005287, + 0.015795961022377014, + 0.015950998291373253, + 0.02380654588341713, + 0.015250268392264843, + 0.010012817569077015, + 0.027775464579463005, + 0.011142559349536896, + 0.004331214819103479, + 0.015959324315190315, + 0.007302763406187296, + 0.007376492023468018, + 0.006403292529284954, + 0.009355754591524601, + 0.00843594130128622, + 0.010800585150718689, + 0.010770842432975769, + 0.008339609950780869, + 0.00961691327393055, + 0.01130523718893528, + 0.008330713957548141, + 0.008558460511267185, + 0.00866654608398676, + 0.008104494772851467, + 0.008189757354557514, + 0.0067183044739067554, + 0.006909650284796953, + 0.010529858060181141, + 0.008677948266267776, + 0.01449175737798214, + 0.011583765037357807, + 0.0063097490929067135, + 0.0062978859059512615, + 0.007421284448355436, + 0.015016108751296997, + 0.008390110917389393, + 0.011967957951128483, + 0.007426959928125143, + 0.004087199456989765, + 0.003653963329270482, + 0.0023806157987564802, + 0.00329232239164412, + 0.0049980212934315205, + 0.0036008774768561125, + 0.0031067095696926117, + 0.007734342943876982, + 0.008595552295446396, + 0.007216322235763073, + 0.006451047025620937, + 0.005808225367218256, + 0.008107868954539299, + 0.009643693454563618, + 0.007409119047224522, + 0.033504508435726166, + 0.015523957088589668, + 0.021968558430671692, + 0.011811326257884502, + 0.012701991014182568, + 0.012179703451693058, + 0.016661347821354866, + 0.018695734441280365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.038110267370939255 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up captures this moment of suspended tension, focusing on a metal coin, its thin, reedy rim impossibly balanced. Made of a slightly worn, silvery metal, the coin catches the soft, aimless sunlight filtering through the car window, casting a thin, bright beam of light against the dark, textured surface of the vehicle's dashboard. Subtle, almost imperceptible vibrations from the car's engine ripple through the dashboard, causing the coin to begin its erratic oscillation. The light on its rim flickers erratically as it sways back and forth, its latent energy poised to be unleashed. Suddenly, it succumbs to gravity, falling to the ground with a sharp, decisive motion, and instantly erupting into a rapid spin. The coin's surface transforms into a dazzling, hypnotic blur of reflected light, a shimmering silver disk humming against the textured surface. The rotation gradually slows, the frantic whirring becoming a slower, rhythmic sway, until, after one final, languid rotation, the coin comes to a complete stillness, lying flat and motionless. The camera’s tight focus now reveals the worn, smooth relief of the coin’s surface, its rounded shape now stable and fully visible, bringing the scene from a state of instability to a quietly dynamic resolution.", + "seed": 3953, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_5/index_9887_seed_3953_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1947171464562416, + "vlm_logprob": -0.1947171464562416, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.1947171464562416, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a metal coin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin made of slightly worn, silvery metal with a thin, reedy rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show soft sunlight filtering through a car window onto a dark, textured dashboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the coin fall to the ground and spin rapidly before slowing to a complete stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tightly focused on the coin's surface details throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009363899007439613 + ], + "opposite_logprob_tokens": [ + -5.509364128112793 + ], + "logprob_reward": -0.009363899007439613, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.009363899007439613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12317562103271484, + 0.057382065802812576, + 0.1600414216518402, + 0.0642641931772232, + 0.28808724880218506, + 0.14828811585903168, + 0.054877396672964096, + 0.04130861163139343, + 0.025630053132772446, + 0.03694527596235275 + ], + "credit_map": [ + 0.01291934959590435, + 0.012273837812244892, + 0.010521805845201015, + 0.016793090850114822, + 0.027590692043304443, + 0.01491830125451088, + 0.017469538375735283, + 0.010688996873795986, + 0.005764822009950876, + 0.0069990926422178745, + 0.003915937151759863, + 0.005459080450236797, + 0.015764474868774414, + 0.0055186934769153595, + 0.007504948414862156, + 0.006455014459788799, + 0.009735007770359516, + 0.012216942384839058, + 0.007773812394589186, + 0.0258167777210474, + 0.031528569757938385, + 0.014164097607135773, + 0.03810834139585495, + 0.020697882398962975, + 0.0052683050744235516, + 0.00534772826358676, + 0.0034843480680137873, + 0.00562217365950346, + 0.012752410024404526, + 0.016317620873451233, + 0.0072638229466974735, + 0.008207780309021473, + 0.02174454741179943, + 0.021229485049843788, + 0.017640607431530952, + 0.03126668930053711, + 0.030512575060129166, + 0.0869387686252594, + 0.02301214449107647, + 0.055742427706718445, + 0.02544918656349182, + 0.011479003354907036, + 0.01738675683736801, + 0.02824636921286583, + 0.015995293855667114, + 0.012753834947943687, + 0.019606810063123703, + 0.01737086847424507, + 0.004787631798535585, + 0.004372791852802038, + 0.0073998840525746346, + 0.011078119277954102, + 0.005096720531582832, + 0.006095384247601032, + 0.009064494632184505, + 0.006982375402003527, + 0.0030768418218940496, + 0.0030959725845605135, + 0.0042693130671978, + 0.008090370334684849, + 0.006846597883850336, + 0.004845879971981049, + 0.0062799775041639805, + 0.004803658928722143, + 0.0020440807566046715, + 0.00177576276473701, + 0.0033153227996081114, + 0.004601054824888706, + 0.0037891832180321217, + 0.0032099084928631783, + 0.003571063280105591, + 0.003323676995933056, + 0.00262949476018548, + 0.0027858729008585215, + 0.004522325005382299, + 0.007191944867372513, + 0.00576800387352705, + 0.005239308346062899, + 0.0043417601846158504, + 0.004466564394533634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 5 + ], + "selected_reward": -0.009363899007439613 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a metal coin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin made of slightly worn, silvery metal with a thin, reedy rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show soft sunlight filtering through a car window onto a dark, textured dashboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the coin fall to the ground and spin rapidly before slowing to a complete stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tightly focused on the coin's surface details throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016098985448479652 + ], + "opposite_logprob_tokens": [ + -4.141098976135254 + ], + "logprob_reward": -0.016098985448479652, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016098985448479652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1550711989402771, + 0.055412329733371735, + 0.160424143075943, + 0.0865098312497139, + 0.2747555077075958, + 0.055192407220602036, + 0.060826703906059265, + 0.059570640325546265, + 0.03970995172858238, + 0.05252726376056671 + ], + "credit_map": [ + 0.019466806203126907, + 0.015476260334253311, + 0.008300972171127796, + 0.022602025419473648, + 0.027797261252999306, + 0.019620060920715332, + 0.02088460884988308, + 0.020923200994729996, + 0.0051243421621620655, + 0.0055503156036138535, + 0.0037034826818853617, + 0.012287493795156479, + 0.009786388836801052, + 0.006266587879508734, + 0.006671891547739506, + 0.0060218265280127525, + 0.01605590060353279, + 0.02358594909310341, + 0.010269450955092907, + 0.017496738582849503, + 0.03062678873538971, + 0.022309141233563423, + 0.021972164511680603, + 0.01810801401734352, + 0.007937506772577763, + 0.008250030688941479, + 0.007660815492272377, + 0.008569037541747093, + 0.01391030102968216, + 0.01786782033741474, + 0.007717181462794542, + 0.014597137458622456, + 0.025582848116755486, + 0.0257889237254858, + 0.021812131628394127, + 0.03822711110115051, + 0.035802848637104034, + 0.05873914435505867, + 0.031694624572992325, + 0.03710789978504181, + 0.006299513857811689, + 0.004780177492648363, + 0.0069792685098946095, + 0.010062528774142265, + 0.005366716533899307, + 0.005751396995037794, + 0.008906801231205463, + 0.007046000100672245, + 0.004584830719977617, + 0.004500651732087135, + 0.0076439292170107365, + 0.01451495848596096, + 0.006757983006536961, + 0.005363159347325563, + 0.010338544845581055, + 0.007122644688934088, + 0.007517315447330475, + 0.005624182987958193, + 0.00783548690378666, + 0.012640089727938175, + 0.0069903903640806675, + 0.004996296484023333, + 0.007081926800310612, + 0.006884952541440725, + 0.004002797417342663, + 0.003295209491625428, + 0.00559475552290678, + 0.007597844582051039, + 0.003414955921471119, + 0.004780175164341927, + 0.00640166224911809, + 0.0046225497499108315, + 0.005830755457282066, + 0.005231971852481365, + 0.006990905851125717, + 0.009605949744582176, + 0.006924969144165516, + 0.005068143829703331, + 0.006905854679644108, + 0.00596871180459857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.016098985448479652 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a metal coin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin made of slightly worn, silvery metal with a thin, reedy rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show soft sunlight filtering through a car window onto a dark, textured dashboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the coin fall to the ground and spin rapidly before slowing to a complete stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tightly focused on the coin's surface details throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016081858426332474 + ], + "opposite_logprob_tokens": [ + -4.141081809997559 + ], + "logprob_reward": -0.016081858426332474, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016081858426332474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14028680324554443, + 0.06282413750886917, + 0.11561421304941177, + 0.11324862390756607, + 0.09992960840463638, + 0.048311710357666016, + 0.13577809929847717, + 0.07693769037723541, + 0.10207346826791763, + 0.10499567538499832 + ], + "credit_map": [ + 0.010969824157655239, + 0.017582077533006668, + 0.01074607390910387, + 0.012480504810810089, + 0.033958423882722855, + 0.01670951396226883, + 0.02108559012413025, + 0.01675478182733059, + 0.006419516168534756, + 0.007387850433588028, + 0.004171940963715315, + 0.009804126806557178, + 0.01397177204489708, + 0.0064275613985955715, + 0.009100358933210373, + 0.00554101075977087, + 0.009059672243893147, + 0.013872107490897179, + 0.008264276199042797, + 0.017217038199305534, + 0.016808032989501953, + 0.012816107831895351, + 0.016784263774752617, + 0.020792704075574875, + 0.013204026967287064, + 0.012580051086843014, + 0.010394537821412086, + 0.012351125478744507, + 0.014643331058323383, + 0.020389486104249954, + 0.015727870166301727, + 0.013958184979856014, + 0.008145736530423164, + 0.009224342182278633, + 0.0076079415157437325, + 0.011824511922895908, + 0.013743466697633266, + 0.02245953492820263, + 0.01150013878941536, + 0.015423932112753391, + 0.006170290056616068, + 0.005076924338936806, + 0.005140502471476793, + 0.007730712648481131, + 0.007429907098412514, + 0.004743663594126701, + 0.006001724395900965, + 0.006017984822392464, + 0.014451497234404087, + 0.012413711287081242, + 0.018993783742189407, + 0.016975507140159607, + 0.02784089557826519, + 0.02026796154677868, + 0.010741112753748894, + 0.014093621633946896, + 0.00895720161497593, + 0.00805815402418375, + 0.010717101395130157, + 0.013047323562204838, + 0.00999185349792242, + 0.009975140914320946, + 0.00823238492012024, + 0.007958522066473961, + 0.010615834966301918, + 0.00848157238215208, + 0.014677035622298717, + 0.013815776444971561, + 0.019570088014006615, + 0.014783577993512154, + 0.008380270563066006, + 0.011749300174415112, + 0.008689803071320057, + 0.007881921716034412, + 0.012194186449050903, + 0.01742255687713623, + 0.017466353252530098, + 0.014386409893631935, + 0.013787142001092434, + 0.013167294673621655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.016081858426332474 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a metal coin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin made of slightly worn, silvery metal with a thin, reedy rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show soft sunlight filtering through a car window onto a dark, textured dashboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the coin fall to the ground and spin rapidly before slowing to a complete stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tightly focused on the coin's surface details throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983302116394043 + ], + "opposite_logprob_tokens": [ + -0.5233302116394043 + ], + "logprob_reward": -0.8983302116394043, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983302116394043, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.18060363829135895, + 0.06662699580192566, + 0.22697791457176208, + 0.18693332374095917, + 0.13550244271755219, + 0.03960220888257027, + 0.06208668276667595, + 0.039847780019044876, + 0.018812382593750954, + 0.04300662502646446 + ], + "credit_map": [ + 0.02914985455572605, + 0.01732366532087326, + 0.011806854046881199, + 0.035050198435783386, + 0.019040947780013084, + 0.018276333808898926, + 0.019512787461280823, + 0.03044300712645054, + 0.007305222563445568, + 0.006942628882825375, + 0.006710612680763006, + 0.012330496683716774, + 0.011803270317614079, + 0.0059660268016159534, + 0.008858688175678253, + 0.006710058078169823, + 0.015272566117346287, + 0.032903581857681274, + 0.013449196703732014, + 0.025047939270734787, + 0.04325556010007858, + 0.030643479898571968, + 0.037717971950769424, + 0.028687622398138046, + 0.014105888083577156, + 0.021732520312070847, + 0.010580888949334621, + 0.01335922721773386, + 0.028186194598674774, + 0.05285469442605972, + 0.016259148716926575, + 0.029854780063033104, + 0.011140057817101479, + 0.008906901814043522, + 0.01611941121518612, + 0.027476884424686432, + 0.012190941721200943, + 0.017505431547760963, + 0.02107144519686699, + 0.02109137363731861, + 0.006642249412834644, + 0.0028260955587029457, + 0.0050092292949557304, + 0.005845523905009031, + 0.0054441760294139385, + 0.004187591839581728, + 0.005023541860282421, + 0.004623807966709137, + 0.004817803855985403, + 0.006033986806869507, + 0.006178840529173613, + 0.011206565424799919, + 0.006281303707510233, + 0.007545718923211098, + 0.01139778271317482, + 0.008624682202935219, + 0.0037060570903122425, + 0.0029883764218539, + 0.0039055850356817245, + 0.0053699868731200695, + 0.00829926785081625, + 0.0055347406305372715, + 0.005470401607453823, + 0.004573365207761526, + 0.0016795555129647255, + 0.0016756867989897728, + 0.0023994522634893656, + 0.00393280852586031, + 0.0023532023187726736, + 0.0017172537045553327, + 0.002879013307392597, + 0.002175411907956004, + 0.004252181854099035, + 0.003611363237723708, + 0.006933450233191252, + 0.006661455612629652, + 0.00506384763866663, + 0.005837148986756802, + 0.005226531066000462, + 0.005420649889856577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.8983302116394043 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a metal coin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin made of slightly worn, silvery metal with a thin, reedy rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show soft sunlight filtering through a car window onto a dark, textured dashboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the coin fall to the ground and spin rapidly before slowing to a complete stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tightly focused on the coin's surface details throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033710777759552 + ], + "opposite_logprob_tokens": [ + -3.4087107181549072 + ], + "logprob_reward": -0.033710777759552, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033710777759552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0899813175201416, + 0.02526404894888401, + 0.11017702519893646, + 0.04674296826124191, + 0.4509083926677704, + 0.11344564706087112, + 0.056493453681468964, + 0.03993357717990875, + 0.030189471319317818, + 0.036864180117845535 + ], + "credit_map": [ + 0.010365121066570282, + 0.008729858323931694, + 0.007782935630530119, + 0.014164604246616364, + 0.017351148650050163, + 0.011319180950522423, + 0.010829620994627476, + 0.009438838809728622, + 0.003216494806110859, + 0.0026141873095184565, + 0.0029356754384934902, + 0.003091828664764762, + 0.004211776424199343, + 0.0026756420265883207, + 0.003761725965887308, + 0.002756717149168253, + 0.012782448902726173, + 0.015969328582286835, + 0.0069624632596969604, + 0.012138331308960915, + 0.016617508605122566, + 0.012650539167225361, + 0.016680216416716576, + 0.016376186162233353, + 0.004605513997375965, + 0.005612746346741915, + 0.004179546609520912, + 0.005419732071459293, + 0.008167289197444916, + 0.006104285828769207, + 0.00665227510035038, + 0.006001576315611601, + 0.032712485641241074, + 0.02954692766070366, + 0.02986203506588936, + 0.05076409876346588, + 0.044898759573698044, + 0.14131835103034973, + 0.03050772100687027, + 0.09129802137613297, + 0.01708368770778179, + 0.009914364665746689, + 0.011944335885345936, + 0.022147467359900475, + 0.011613823473453522, + 0.010526670143008232, + 0.0162806436419487, + 0.013934648595750332, + 0.004676487762480974, + 0.004947157576680183, + 0.006571386009454727, + 0.011068676598370075, + 0.006261117290705442, + 0.0050940425135195255, + 0.010871159844100475, + 0.007003419101238251, + 0.005067114252597094, + 0.002832980826497078, + 0.004324266221374273, + 0.007987385615706444, + 0.005216896068304777, + 0.004015164449810982, + 0.005712540354579687, + 0.0047772289253771305, + 0.0025699634570628405, + 0.0021941375453025103, + 0.004155693110078573, + 0.005096101667732, + 0.0056704310700297356, + 0.0033514131791889668, + 0.003927210811525583, + 0.0032245186157524586, + 0.003570395056158304, + 0.0036110959481447935, + 0.0056416988372802734, + 0.0035224587190896273, + 0.007231697905808687, + 0.006079540122300386, + 0.0030511284712702036, + 0.00415616063401103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.033710777759552 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_6/index_tensor([9415, 9415])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_6/index_tensor([9415, 9415])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3ad721b850467e4e3bbfe9910aadd05f792cb183 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_6/index_tensor([9415, 9415])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, cinematic shot captures a massive, multi-jointed robotic arm as it rotates majestically and deliberately within the vast hall of an atmospheric reactor. The arm's flawless, polished chrome surface acts like a distorted mirror, reflecting the eerie, pulsating azure glow of the submerged reactor core deep within the circular fissure below. The camera remains perfectly still, emphasizing the machinery's immense scale and contrasting starkly with the savage concrete architecture of the silent chamber. A fine mist hangs in the air, capturing and amplifying the sharp, volumetric, harsh white light that cuts through from unseen fixtures high in the cavernous ceiling. As the arm moves with an unstoppable, silent grace, the blue light twists and slides across its metallic contours, creating a liquid-like effect. The scene, dominated by a cool palette of deep blue, metallic silver, and shadowy black, creates a powerful, awe-inspiring atmosphere of energized and futuristic power, a powerful and awe-inspiring force under control.", + "seed": 19874, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_6/index_9415_seed_19874_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.4787610800471157, + "vlm_logprob": -0.4787610800471157, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.4787610800471157, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive, multi-jointed robotic arm visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robotic arm's surface polished chrome that reflects light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the arm rotating within a vast hall containing a submerged reactor core with an eerie, pulsating azure glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool palette of deep blue, metallic silver, and shadowy black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010977590456604958 + ], + "opposite_logprob_tokens": [ + -4.635977745056152 + ], + "logprob_reward": -0.010977590456604958, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.010977590456604958, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21764244139194489, + 0.03843442350625992, + 0.13078491389751434, + 0.08602087199687958, + 0.18135596811771393, + 0.038819100707769394, + 0.11107566207647324, + 0.048018235713243484, + 0.10545224696397781, + 0.042396221309900284 + ], + "credit_map": [ + 0.015273339115083218, + 0.05367961525917053, + 0.02493945322930813, + 0.0250422153621912, + 0.029747290536761284, + 0.033430177718400955, + 0.0202785711735487, + 0.015251762233674526, + 0.003647377947345376, + 0.003519548336043954, + 0.0062650335021317005, + 0.005017608404159546, + 0.007140993606299162, + 0.0055915359407663345, + 0.004704764578491449, + 0.002547553740441799, + 0.01431766152381897, + 0.018048202618956566, + 0.01695745624601841, + 0.017059268429875374, + 0.026881730183959007, + 0.0192189309746027, + 0.008244411088526249, + 0.010057235136628151, + 0.006902533117681742, + 0.011630844324827194, + 0.008659775368869305, + 0.011874829418957233, + 0.016767436638474464, + 0.015498948283493519, + 0.009102587588131428, + 0.005583912134170532, + 0.011962269432842731, + 0.014898682944476604, + 0.021716073155403137, + 0.027187099680304527, + 0.037110697478055954, + 0.026546252891421318, + 0.019097305834293365, + 0.0228375606238842, + 0.005299573764204979, + 0.004057667218148708, + 0.004875119309872389, + 0.004287231247872114, + 0.006576435640454292, + 0.005243998020887375, + 0.005427052732557058, + 0.0030520218424499035, + 0.01365035679191351, + 0.012476759031414986, + 0.011450902558863163, + 0.011558127589523792, + 0.021737588569521904, + 0.01562032476067543, + 0.012711994349956512, + 0.011869588866829872, + 0.006903200410306454, + 0.0035037032794207335, + 0.006916316691786051, + 0.005033697932958603, + 0.006664195563644171, + 0.006024827715009451, + 0.006329658441245556, + 0.006642633583396673, + 0.012814704328775406, + 0.010438574478030205, + 0.014092201367020607, + 0.009739869274199009, + 0.01970626786351204, + 0.013684365898370743, + 0.0140764769166708, + 0.010899772867560387, + 0.004834968131035566, + 0.004427407402545214, + 0.006687547080218792, + 0.004667320288717747, + 0.00604622345417738, + 0.005448420997709036, + 0.00589258037507534, + 0.004391748458147049 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.010977590456604958 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive, multi-jointed robotic arm visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robotic arm's surface polished chrome that reflects light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the arm rotating within a vast hall containing a submerged reactor core with an eerie, pulsating azure glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool palette of deep blue, metallic silver, and shadowy black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023301001638174057 + ], + "opposite_logprob_tokens": [ + -3.773300886154175 + ], + "logprob_reward": -0.023301001638174057, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023301001638174057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5912110805511475, + 0.027711043134331703, + 0.10255884379148483, + 0.04977553337812424, + 0.035524625331163406, + 0.022662993520498276, + 0.048142123967409134, + 0.0322638601064682, + 0.04548332840204239, + 0.04466657713055611 + ], + "credit_map": [ + 0.04124247282743454, + 0.2022278606891632, + 0.03812694922089577, + 0.06489792466163635, + 0.049684926867485046, + 0.13024799525737762, + 0.0386810377240181, + 0.02610187605023384, + 0.0036963329184800386, + 0.003671418409794569, + 0.004121738485991955, + 0.0038539888337254524, + 0.0031732171773910522, + 0.003401530208066106, + 0.003522800747305155, + 0.0022700168192386627, + 0.010178216733038425, + 0.01432894729077816, + 0.015279872342944145, + 0.0165692251175642, + 0.01432789396494627, + 0.012708197347819805, + 0.009312476962804794, + 0.009853999130427837, + 0.0027958168648183346, + 0.004701520781964064, + 0.0035509061999619007, + 0.003863667603582144, + 0.013724667020142078, + 0.012033449485898018, + 0.00582285039126873, + 0.003282650373876095, + 0.0024122612085193396, + 0.0032661682926118374, + 0.004322578199207783, + 0.005945570766925812, + 0.0066318875178694725, + 0.004559433553367853, + 0.004061063751578331, + 0.004325659014284611, + 0.003218054538592696, + 0.0031483375933021307, + 0.002328215865418315, + 0.0021231668069958687, + 0.003397334599867463, + 0.0036561123561114073, + 0.0029816145543009043, + 0.0018101538298651576, + 0.0048061697743833065, + 0.004740444477647543, + 0.006259697489440441, + 0.0060914987698197365, + 0.00890885666012764, + 0.006635443773120642, + 0.006261166650801897, + 0.004438844975084066, + 0.0034703391138464212, + 0.0020650539081543684, + 0.0042629665695130825, + 0.0034663695842027664, + 0.004779978189617395, + 0.003928229212760925, + 0.005433647893369198, + 0.004857271909713745, + 0.004354238044470549, + 0.00390452379360795, + 0.0071544102393090725, + 0.004194275941699743, + 0.0070991492830216885, + 0.005786550231277943, + 0.005655635613948107, + 0.007334541529417038, + 0.004052415024489164, + 0.003444909118115902, + 0.009729016572237015, + 0.006076019257307053, + 0.005587772931903601, + 0.005101706832647324, + 0.005309720523655415, + 0.005365012213587761 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.023301001638174057 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive, multi-jointed robotic arm visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robotic arm's surface polished chrome that reflects light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the arm rotating within a vast hall containing a submerged reactor core with an eerie, pulsating azure glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool palette of deep blue, metallic silver, and shadowy black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006746016908437014 + ], + "opposite_logprob_tokens": [ + -5.0067458152771 + ], + "logprob_reward": -0.006746016908437014, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006746016908437014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3521890640258789, + 0.10247980058193207, + 0.19425925612449646, + 0.08977963030338287, + 0.03961219638586044, + 0.0339837372303009, + 0.07302307337522507, + 0.03677023947238922, + 0.03826506435871124, + 0.03963789716362953 + ], + "credit_map": [ + 0.02607370726764202, + 0.1264217346906662, + 0.018839478492736816, + 0.035356126725673676, + 0.02885568141937256, + 0.0784355103969574, + 0.023717956617474556, + 0.014488905668258667, + 0.011789849027991295, + 0.0133738424628973, + 0.013962669298052788, + 0.013607960194349289, + 0.014160234481096268, + 0.01219681091606617, + 0.014222423546016216, + 0.009166011586785316, + 0.01510316226631403, + 0.018714675679802895, + 0.023919004946947098, + 0.016593042761087418, + 0.05712801590561867, + 0.03796767815947533, + 0.01400308683514595, + 0.010830596089363098, + 0.007685219869017601, + 0.011284143663942814, + 0.01092210877686739, + 0.014468144625425339, + 0.014274378307163715, + 0.012241031974554062, + 0.011379559524357319, + 0.00752505986019969, + 0.0031815862748771906, + 0.004287557676434517, + 0.005146574229001999, + 0.005363135132938623, + 0.006526289042085409, + 0.0049895127303898335, + 0.004226135555654764, + 0.0058914124965667725, + 0.004403334576636553, + 0.003542051650583744, + 0.004364327061921358, + 0.004142405465245247, + 0.005172452423721552, + 0.004384211264550686, + 0.00510253245010972, + 0.0028724242001771927, + 0.008547001518309116, + 0.007604834157973528, + 0.007754528895020485, + 0.008151117712259293, + 0.016331428661942482, + 0.010935798287391663, + 0.0072247860953211784, + 0.006473580840975046, + 0.003220856189727783, + 0.002889945637434721, + 0.004469436593353748, + 0.0064971656538546085, + 0.005218446720391512, + 0.004259584471583366, + 0.007107304409146309, + 0.003107501659542322, + 0.0036387646105140448, + 0.004746030550450087, + 0.0049970922991633415, + 0.002592037431895733, + 0.009036674164235592, + 0.005408747587352991, + 0.0038723573088645935, + 0.003973361570388079, + 0.004735518712550402, + 0.00570059847086668, + 0.004454503301531076, + 0.005172867327928543, + 0.006179630756378174, + 0.004327050410211086, + 0.0051041883416473866, + 0.00396354403346777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.006746016908437014 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive, multi-jointed robotic arm visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robotic arm's surface polished chrome that reflects light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the arm rotating within a vast hall containing a submerged reactor core with an eerie, pulsating azure glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool palette of deep blue, metallic silver, and shadowy black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.350281000137329 + ], + "opposite_logprob_tokens": [ + -0.10028109699487686 + ], + "logprob_reward": -2.350281000137329, + "yes_no_margin_reward": -2.25, + "token_credit_reward": -2.350281000137329, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.14616286754608154, + 0.14231662452220917, + 0.08358971029520035, + 0.0669102892279625, + 0.07917897403240204, + 0.08165894448757172, + 0.15426456928253174, + 0.06634548306465149, + 0.059612736105918884, + 0.11995981633663177 + ], + "credit_map": [ + 0.011263647116720676, + 0.02171460911631584, + 0.02082490175962448, + 0.015137302689254284, + 0.028169099241495132, + 0.018968656659126282, + 0.01779451221227646, + 0.01229014154523611, + 0.015028541907668114, + 0.019457828253507614, + 0.02332422323524952, + 0.016800636425614357, + 0.015585207380354404, + 0.01700967364013195, + 0.022496933117508888, + 0.012613588012754917, + 0.006457291543483734, + 0.012029281817376614, + 0.010656895115971565, + 0.008036324754357338, + 0.01804700493812561, + 0.01289945375174284, + 0.008069016970694065, + 0.007394448388367891, + 0.005547170527279377, + 0.011879799887537956, + 0.007509013172239065, + 0.009747523814439774, + 0.007364077493548393, + 0.0077759320847690105, + 0.011218679137527943, + 0.005868104752153158, + 0.006847070995718241, + 0.009411067701876163, + 0.009259087964892387, + 0.01629311591386795, + 0.010565197095274925, + 0.010901893489062786, + 0.010256835259497166, + 0.005644713528454304, + 0.008460265584290028, + 0.009849317371845245, + 0.009557118639349937, + 0.010005377233028412, + 0.011013864539563656, + 0.01163559965789318, + 0.014550231397151947, + 0.006587173789739609, + 0.020687101408839226, + 0.011560636572539806, + 0.01534327957779169, + 0.019544709473848343, + 0.03398464247584343, + 0.02221338264644146, + 0.017504820600152016, + 0.01342599093914032, + 0.009651875123381615, + 0.005848812870681286, + 0.0059058330953121185, + 0.008859602734446526, + 0.0081415306776762, + 0.008223582059144974, + 0.013078431598842144, + 0.0066358125768601894, + 0.006750874221324921, + 0.008858945220708847, + 0.0050144437700510025, + 0.0069737727753818035, + 0.011327078565955162, + 0.0074028256349265575, + 0.006807228084653616, + 0.006477572023868561, + 0.01322594378143549, + 0.012325050309300423, + 0.018058113753795624, + 0.016455935314297676, + 0.016939207911491394, + 0.010705134831368923, + 0.016259869560599327, + 0.015990566462278366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -2.350281000137329 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive, multi-jointed robotic arm visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robotic arm's surface polished chrome that reflects light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the arm rotating within a vast hall containing a submerged reactor core with an eerie, pulsating azure glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool palette of deep blue, metallic silver, and shadowy black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002499791095033288 + ], + "opposite_logprob_tokens": [ + -6.002499580383301 + ], + "logprob_reward": -0.002499791095033288, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002499791095033288, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12423045933246613, + 0.1349528580904007, + 0.1335209608078003, + 0.06819306313991547, + 0.09078648686408997, + 0.08582029491662979, + 0.12408298999071121, + 0.07601439952850342, + 0.043389786034822464, + 0.11900877952575684 + ], + "credit_map": [ + 0.010667528957128525, + 0.02708348073065281, + 0.016666976734995842, + 0.013189290650188923, + 0.017788071185350418, + 0.015751948580145836, + 0.013169006444513798, + 0.009914131835103035, + 0.016252143308520317, + 0.015596562065184116, + 0.023582812398672104, + 0.016983874142169952, + 0.01646854728460312, + 0.017128843814134598, + 0.01778334379196167, + 0.011156721040606499, + 0.01288721151649952, + 0.014389317482709885, + 0.020672762766480446, + 0.017063651233911514, + 0.029169075191020966, + 0.020341958850622177, + 0.008390652947127819, + 0.010606322437524796, + 0.0072399829514324665, + 0.009255649521946907, + 0.009255570359528065, + 0.009019575081765652, + 0.011394205503165722, + 0.009563674218952656, + 0.007203192915767431, + 0.005261202808469534, + 0.007186692673712969, + 0.010496925562620163, + 0.009887868538498878, + 0.01654849201440811, + 0.014687520451843739, + 0.01334252581000328, + 0.009928693994879723, + 0.008707757107913494, + 0.009785277768969536, + 0.008325628004968166, + 0.009491091594099998, + 0.011153175495564938, + 0.010317562147974968, + 0.01006852462887764, + 0.02007647231221199, + 0.006602548994123936, + 0.01478252187371254, + 0.011584348045289516, + 0.014788833446800709, + 0.011762294918298721, + 0.023201700299978256, + 0.018516957759857178, + 0.012938452884554863, + 0.016507869586348534, + 0.012573666870594025, + 0.006933531258255243, + 0.006390605587512255, + 0.009131019003689289, + 0.010208948515355587, + 0.009851127862930298, + 0.01318720355629921, + 0.007738280110061169, + 0.006749042309820652, + 0.004117689095437527, + 0.0059645879082381725, + 0.0038393097929656506, + 0.006628422532230616, + 0.00622659083455801, + 0.004628823138773441, + 0.005235311575233936, + 0.012200978584587574, + 0.014932026155292988, + 0.013005579821765423, + 0.016704175621271133, + 0.019182303920388222, + 0.010458774864673615, + 0.016992265358567238, + 0.015532666817307472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.002499791095033288 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "static, cinematic shot captures a massive, multi-jointed robotic arm as it rotates majestically and deliberately within the vast hall of an atmospheric reactor. The arm's flawless, polished chrome surface acts like a distorted mirror, reflecting the eerie, pulsating azure glow of the submerged reactor core deep within the circular fissure below. The camera remains perfectly still, emphasizing the machinery's immense scale and contrasting starkly with the savage concrete architecture of the silent chamber. A fine mist hangs in the air, capturing and amplifying the sharp, volumetric, harsh white light that cuts through from unseen fixtures high in the cavernous ceiling. As the arm moves with an unstoppable, silent grace, the blue light twists and slides across its metallic contours, creating a liquid-like effect. The scene, dominated by a cool palette of deep blue, metallic silver, and shadowy black, creates a powerful, awe-inspiring atmosphere of energized and futuristic power, a powerful and awe-inspiring force under control.", + "seed": 19875, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_6/index_9415_seed_19875_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.3159609876107424, + "vlm_logprob": -0.3159609876107424, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.3159609876107424, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive, multi-jointed robotic arm visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robotic arm's surface polished chrome that reflects light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the arm rotating within a vast hall containing a submerged reactor core with an eerie, pulsating azure glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool palette of deep blue, metallic silver, and shadowy black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03962372988462448 + ], + "opposite_logprob_tokens": [ + -3.289623737335205 + ], + "logprob_reward": -0.03962372988462448, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03962372988462448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24951475858688354, + 0.13025839626789093, + 0.05571800842881203, + 0.07170406728982925, + 0.11296505481004715, + 0.10987932235002518, + 0.04342690482735634, + 0.03725225850939751, + 0.14525707066059113, + 0.04402420297265053 + ], + "credit_map": [ + 0.028260018676519394, + 0.04353965446352959, + 0.020555894821882248, + 0.021874986588954926, + 0.03308654949069023, + 0.035286854952573776, + 0.042560968548059464, + 0.02434983104467392, + 0.04867865890264511, + 0.009323074482381344, + 0.005272834096103907, + 0.00442257896065712, + 0.006292416248470545, + 0.04611987620592117, + 0.005312785040587187, + 0.004836163483560085, + 0.014565026387572289, + 0.011135599575936794, + 0.004077852703630924, + 0.0033671280834823847, + 0.006376087665557861, + 0.0063691046088933945, + 0.006025239359587431, + 0.003801973769441247, + 0.009950418956577778, + 0.014323275536298752, + 0.008486783131957054, + 0.0052966768853366375, + 0.011839209124445915, + 0.005399319808930159, + 0.011098743416368961, + 0.005309640895575285, + 0.017016597092151642, + 0.019286449998617172, + 0.015163392759859562, + 0.009959760122001171, + 0.010035098530352116, + 0.014120488427579403, + 0.014350676909089088, + 0.01303259190171957, + 0.007468950003385544, + 0.03324040025472641, + 0.004792476538568735, + 0.004615350626409054, + 0.012769361026585102, + 0.034338273108005524, + 0.00829017162322998, + 0.004364342428743839, + 0.004917986690998077, + 0.006076747551560402, + 0.0039269509725272655, + 0.004520336166024208, + 0.0053209345787763596, + 0.006552455015480518, + 0.008363519795238972, + 0.00374797685071826, + 0.004965746775269508, + 0.005719681270420551, + 0.0033945434261113405, + 0.0028354867827147245, + 0.004387072287499905, + 0.006738551426678896, + 0.005358927417546511, + 0.0038522451650351286, + 0.017929596826434135, + 0.027263212949037552, + 0.014970120042562485, + 0.014243785291910172, + 0.020202437415719032, + 0.020115908235311508, + 0.017741495743393898, + 0.012790501117706299, + 0.005412364844232798, + 0.006164878606796265, + 0.004562333691865206, + 0.004032812546938658, + 0.007726293057203293, + 0.006091057788580656, + 0.006728907581418753, + 0.0033055529929697514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.03962372988462448 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive, multi-jointed robotic arm visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robotic arm's surface polished chrome that reflects light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the arm rotating within a vast hall containing a submerged reactor core with an eerie, pulsating azure glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool palette of deep blue, metallic silver, and shadowy black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029809661209583282 + ], + "opposite_logprob_tokens": [ + -3.5298097133636475 + ], + "logprob_reward": -0.029809661209583282, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029809661209583282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12514226138591766, + 0.2053600400686264, + 0.07438674569129944, + 0.13804976642131805, + 0.07354030758142471, + 0.09057841449975967, + 0.06744704395532608, + 0.06784844398498535, + 0.09411892294883728, + 0.06352804601192474 + ], + "credit_map": [ + 0.022028665989637375, + 0.014488914050161839, + 0.018874898552894592, + 0.01229291781783104, + 0.012295518070459366, + 0.016759244725108147, + 0.013398971408605576, + 0.01500313263386488, + 0.07288149744272232, + 0.015717793256044388, + 0.007842651568353176, + 0.0076390216127038, + 0.010151528753340244, + 0.0724841058254242, + 0.008775472640991211, + 0.009867951273918152, + 0.018056757748126984, + 0.014097011648118496, + 0.006500318646430969, + 0.004406353924423456, + 0.008356370031833649, + 0.011621690355241299, + 0.005374791566282511, + 0.005973446648567915, + 0.015115717425942421, + 0.020618611946702003, + 0.013583838939666748, + 0.013466574251651764, + 0.021151883527636528, + 0.015690159052610397, + 0.024523677304387093, + 0.013899306766688824, + 0.010540148243308067, + 0.010522974655032158, + 0.010207219049334526, + 0.006497712340205908, + 0.006165347062051296, + 0.010388923808932304, + 0.010515918955206871, + 0.008702059276401997, + 0.007270001340657473, + 0.024388520047068596, + 0.005322565324604511, + 0.006263913586735725, + 0.007599045988172293, + 0.023133499547839165, + 0.007540876045823097, + 0.009059988893568516, + 0.0061978562735021114, + 0.010037140920758247, + 0.007407252676784992, + 0.005849875044077635, + 0.007920135743916035, + 0.01047225296497345, + 0.008515587076544762, + 0.011046943254768848, + 0.012670347467064857, + 0.019202670082449913, + 0.00642265984788537, + 0.004815032705664635, + 0.004779411014169455, + 0.009992537088692188, + 0.0056542144156992435, + 0.0043115634471178055, + 0.011184221133589745, + 0.016791371628642082, + 0.010129277594387531, + 0.009534726850688457, + 0.012403570115566254, + 0.013091040775179863, + 0.011933684349060059, + 0.009051021188497543, + 0.00618134206160903, + 0.007575619500130415, + 0.005485796835273504, + 0.006602829787880182, + 0.011317669413983822, + 0.00813143141567707, + 0.012896050699055195, + 0.0053373053669929504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.029809661209583282 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive, multi-jointed robotic arm visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robotic arm's surface polished chrome that reflects light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the arm rotating within a vast hall containing a submerged reactor core with an eerie, pulsating azure glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool palette of deep blue, metallic silver, and shadowy black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00764006981626153 + ], + "opposite_logprob_tokens": [ + -4.8826398849487305 + ], + "logprob_reward": -0.00764006981626153, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00764006981626153, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1966676265001297, + 0.1864551454782486, + 0.05343223363161087, + 0.04266240820288658, + 0.0999457836151123, + 0.1603604406118393, + 0.06936913728713989, + 0.05461082607507706, + 0.057251207530498505, + 0.0792451947927475 + ], + "credit_map": [ + 0.034815724939107895, + 0.0344250351190567, + 0.023815792053937912, + 0.013069598004221916, + 0.02344435267150402, + 0.028116274625062943, + 0.021912548691034317, + 0.01706829108297825, + 0.06402630358934402, + 0.014971047639846802, + 0.009248570539057255, + 0.007395680993795395, + 0.011937848292291164, + 0.06381203234195709, + 0.007527864072471857, + 0.007535789627581835, + 0.01045927032828331, + 0.010011586360633373, + 0.0060022128745913506, + 0.003683490213006735, + 0.005344211123883724, + 0.008632880635559559, + 0.00355965131893754, + 0.005738933105021715, + 0.005292220506817102, + 0.007627126295119524, + 0.0035744989290833473, + 0.004547666292637587, + 0.006135694682598114, + 0.005696858745068312, + 0.004968760069459677, + 0.004819580353796482, + 0.01772099919617176, + 0.019681131467223167, + 0.01325115654617548, + 0.009088337421417236, + 0.008345301263034344, + 0.008970145136117935, + 0.009633178822696209, + 0.013255540281534195, + 0.014427358284592628, + 0.057219576090574265, + 0.006324631627649069, + 0.008795294910669327, + 0.008542695082724094, + 0.04918418452143669, + 0.007823912426829338, + 0.008042789064347744, + 0.008597086183726788, + 0.009379951283335686, + 0.006894748657941818, + 0.007877083495259285, + 0.008662949316203594, + 0.011443039402365685, + 0.010579204186797142, + 0.005935078952461481, + 0.007932417094707489, + 0.007588233333081007, + 0.005484065040946007, + 0.004799637012183666, + 0.005348008591681719, + 0.009753848426043987, + 0.007621732074767351, + 0.006082882639020681, + 0.010982153005897999, + 0.011142428033053875, + 0.00509702367708087, + 0.005253246519714594, + 0.007949317805469036, + 0.007963965646922588, + 0.005084242206066847, + 0.0037788348272442818, + 0.009094325825572014, + 0.012316709384322166, + 0.007830600254237652, + 0.008056466467678547, + 0.014037474058568478, + 0.01107260212302208, + 0.010109656490385532, + 0.006727362051606178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.00764006981626153 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive, multi-jointed robotic arm visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robotic arm's surface polished chrome that reflects light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the arm rotating within a vast hall containing a submerged reactor core with an eerie, pulsating azure glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool palette of deep blue, metallic silver, and shadowy black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.5020931959152222 + ], + "opposite_logprob_tokens": [ + -0.25209322571754456 + ], + "logprob_reward": -1.5020931959152222, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.5020931959152222, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.12688784301280975, + 0.5626512169837952, + 0.10192280262708664, + 0.03060317598283291, + 0.03993374854326248, + 0.026320891454815865, + 0.046413443982601166, + 0.023203745484352112, + 0.02849307283759117, + 0.013570000417530537 + ], + "credit_map": [ + 0.014030582271516323, + 0.024004466831684113, + 0.01213754154741764, + 0.009642474353313446, + 0.021078769117593765, + 0.018220925703644753, + 0.017663316801190376, + 0.010109765455126762, + 0.21893396973609924, + 0.042142193764448166, + 0.01659799925982952, + 0.010866610333323479, + 0.027319874614477158, + 0.21614350378513336, + 0.01150413230061531, + 0.01914302073419094, + 0.01593954861164093, + 0.01774417981505394, + 0.010152969509363174, + 0.007891731336712837, + 0.01394662819802761, + 0.017192885279655457, + 0.008740517310798168, + 0.010314359329640865, + 0.005277362186461687, + 0.007893494330346584, + 0.003145823022350669, + 0.0028793676756322384, + 0.003660946385934949, + 0.0021654251031577587, + 0.0034243441186845303, + 0.002156414557248354, + 0.006603787187486887, + 0.006904084701091051, + 0.00465386314317584, + 0.0034990983549505472, + 0.0030941497534513474, + 0.003353677922859788, + 0.0035026974510401487, + 0.008322392590343952, + 0.0036649706307798624, + 0.0068958173505961895, + 0.0014910638565197587, + 0.002550545847043395, + 0.002633101539686322, + 0.0038491268642246723, + 0.0027721496298909187, + 0.0024641184136271477, + 0.005202566273510456, + 0.00604181969538331, + 0.004739861004054546, + 0.00499449810013175, + 0.005886624101549387, + 0.006518299225717783, + 0.007998984307050705, + 0.00503079779446125, + 0.0029604218434542418, + 0.0033128801733255386, + 0.0022827056236565113, + 0.0019685854203999043, + 0.0029146578162908554, + 0.0044019934721291065, + 0.0027420076075941324, + 0.00262049725279212, + 0.0038503806572407484, + 0.0037334877997636795, + 0.0037886351346969604, + 0.00332177197560668, + 0.003301936900243163, + 0.003651128150522709, + 0.003796522505581379, + 0.0030492101795971394, + 0.0018104615155607462, + 0.002696071518585086, + 0.0011336178286001086, + 0.0012865806929767132, + 0.002287076786160469, + 0.0018925215117633343, + 0.0015589146642014384, + 0.0009047571802511811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -1.5020931959152222 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive, multi-jointed robotic arm visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robotic arm's surface polished chrome that reflects light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the arm rotating within a vast hall containing a submerged reactor core with an eerie, pulsating azure glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool palette of deep blue, metallic silver, and shadowy black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006382812280207872 + ], + "opposite_logprob_tokens": [ + -7.375638484954834 + ], + "logprob_reward": -0.0006382812280207872, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006382812280207872, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1214677095413208, + 0.47624316811561584, + 0.06019424647092819, + 0.055221471935510635, + 0.051315389573574066, + 0.028908470645546913, + 0.054905619472265244, + 0.02723359502851963, + 0.10269681364297867, + 0.021813463419675827 + ], + "credit_map": [ + 0.018276819959282875, + 0.02042992226779461, + 0.014320511370897293, + 0.009043887257575989, + 0.01651189848780632, + 0.016156837344169617, + 0.014962119981646538, + 0.011765721254050732, + 0.18534858524799347, + 0.03581928834319115, + 0.014020734466612339, + 0.009523916058242321, + 0.023031769320368767, + 0.18242701888084412, + 0.010448400862514973, + 0.015623470768332481, + 0.007828853093087673, + 0.01117868535220623, + 0.005723831709474325, + 0.005306954495608807, + 0.008146257139742374, + 0.00888371467590332, + 0.0068098329938948154, + 0.006316120736300945, + 0.009366274811327457, + 0.01354471780359745, + 0.005441238638013601, + 0.004520130343735218, + 0.00698827626183629, + 0.00417599081993103, + 0.006654208991676569, + 0.004530635196715593, + 0.009024957194924355, + 0.007499976083636284, + 0.006809331942349672, + 0.00417869770899415, + 0.004600884858518839, + 0.006870209239423275, + 0.007023762911558151, + 0.005307570565491915, + 0.0025599105283617973, + 0.007749367970973253, + 0.0020906690042465925, + 0.0018225254025310278, + 0.0035201346036046743, + 0.006734628230333328, + 0.002373172203078866, + 0.0020580627024173737, + 0.006478770636022091, + 0.0060799261555075645, + 0.0051163542084395885, + 0.006186190061271191, + 0.007098386529833078, + 0.00833496730774641, + 0.010781525634229183, + 0.004829498939216137, + 0.0038658587727695704, + 0.005005722399801016, + 0.0023742499761283398, + 0.002980967750772834, + 0.002102659549564123, + 0.003970978315919638, + 0.004182824864983559, + 0.0027503324672579765, + 0.013259186409413815, + 0.01880905032157898, + 0.010852527804672718, + 0.010338014923036098, + 0.01413083914667368, + 0.013490360230207443, + 0.01230636890977621, + 0.009510459378361702, + 0.0022982629016041756, + 0.004501506686210632, + 0.0017648653592914343, + 0.002121325582265854, + 0.0032740242313593626, + 0.003069228958338499, + 0.0031368224881589413, + 0.0016474261647090316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.0006382812280207872 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "static, cinematic shot captures a massive, multi-jointed robotic arm as it rotates majestically and deliberately within the vast hall of an atmospheric reactor. The arm's flawless, polished chrome surface acts like a distorted mirror, reflecting the eerie, pulsating azure glow of the submerged reactor core deep within the circular fissure below. The camera remains perfectly still, emphasizing the machinery's immense scale and contrasting starkly with the savage concrete architecture of the silent chamber. A fine mist hangs in the air, capturing and amplifying the sharp, volumetric, harsh white light that cuts through from unseen fixtures high in the cavernous ceiling. As the arm moves with an unstoppable, silent grace, the blue light twists and slides across its metallic contours, creating a liquid-like effect. The scene, dominated by a cool palette of deep blue, metallic silver, and shadowy black, creates a powerful, awe-inspiring atmosphere of energized and futuristic power, a powerful and awe-inspiring force under control.", + "seed": 19876, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_6/index_9415_seed_19876_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.43646154839079826, + "vlm_logprob": -0.43646154839079826, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.43646154839079826, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive, multi-jointed robotic arm visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robotic arm's surface polished chrome that reflects light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the arm rotating within a vast hall containing a submerged reactor core with an eerie, pulsating azure glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool palette of deep blue, metallic silver, and shadowy black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006597285624593496 + ], + "opposite_logprob_tokens": [ + -5.256597518920898 + ], + "logprob_reward": -0.006597285624593496, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.006597285624593496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0879904106259346, + 0.12273105978965759, + 0.03853918984532356, + 0.030667509883642197, + 0.5164336562156677, + 0.06678687781095505, + 0.028125984594225883, + 0.03580363839864731, + 0.0359472781419754, + 0.036974452435970306 + ], + "credit_map": [ + 0.010122926905751228, + 0.005498731043189764, + 0.005623611155897379, + 0.00763432402163744, + 0.015698688104748726, + 0.012394953519105911, + 0.012410469353199005, + 0.01860671304166317, + 0.01626993529498577, + 0.006241364870220423, + 0.007447357289493084, + 0.008332428522408009, + 0.02397739700973034, + 0.028140323236584663, + 0.017157286405563354, + 0.015164965763688087, + 0.004615252837538719, + 0.00481461314484477, + 0.005639438051730394, + 0.006070930976420641, + 0.0055230422876775265, + 0.0033547214698046446, + 0.004828449338674545, + 0.0036927396431565285, + 0.004020127467811108, + 0.003906607162207365, + 0.003398260800167918, + 0.0033612775150686502, + 0.004687292966991663, + 0.0029618788976222277, + 0.004860846325755119, + 0.0034712182823568583, + 0.05412980169057846, + 0.061471112072467804, + 0.06998040527105331, + 0.06182662770152092, + 0.09713917225599289, + 0.06272576749324799, + 0.056160349398851395, + 0.053000371903181076, + 0.007217525504529476, + 0.005513973534107208, + 0.00930183194577694, + 0.007894754409790039, + 0.01062935683876276, + 0.007765266112983227, + 0.008056560531258583, + 0.010407612659037113, + 0.003649236634373665, + 0.0021821726113557816, + 0.0035037954803556204, + 0.0030369029846042395, + 0.004403676372021437, + 0.004197294358164072, + 0.0029751223046332598, + 0.004177783615887165, + 0.00369091285392642, + 0.0042092096991837025, + 0.004184260498732328, + 0.004485274665057659, + 0.005806724540889263, + 0.005219925660640001, + 0.003457203973084688, + 0.004750124178826809, + 0.0028612155001610518, + 0.0029252776876091957, + 0.0059813461266458035, + 0.004265448544174433, + 0.005150876939296722, + 0.0026424259413033724, + 0.006935851648449898, + 0.005184835754334927, + 0.003745599649846554, + 0.003030236344784498, + 0.004661098588258028, + 0.007771131582558155, + 0.004844241309911013, + 0.004571103490889072, + 0.0038477880880236626, + 0.004503252916038036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.006597285624593496 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive, multi-jointed robotic arm visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robotic arm's surface polished chrome that reflects light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the arm rotating within a vast hall containing a submerged reactor core with an eerie, pulsating azure glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool palette of deep blue, metallic silver, and shadowy black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810957819223404 + ], + "opposite_logprob_tokens": [ + -3.288109540939331 + ], + "logprob_reward": -0.03810957819223404, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810957819223404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18970175087451935, + 0.0873769074678421, + 0.07058865576982498, + 0.062476277351379395, + 0.11590836197137833, + 0.1350524127483368, + 0.06941003352403641, + 0.1425817906856537, + 0.0886591374874115, + 0.038244638592004776 + ], + "credit_map": [ + 0.02562762051820755, + 0.01716688647866249, + 0.010777195915579796, + 0.01619306206703186, + 0.0348413921892643, + 0.025529025122523308, + 0.02396875061094761, + 0.03559783473610878, + 0.0073287212289869785, + 0.007028401829302311, + 0.007011469453573227, + 0.007931569591164589, + 0.020312493667006493, + 0.013788223266601562, + 0.011109190993010998, + 0.012866842560470104, + 0.00983503833413124, + 0.009827491827309132, + 0.009307642467319965, + 0.0096732247620821, + 0.010354786179959774, + 0.005210986360907555, + 0.009010979905724525, + 0.0073685068637132645, + 0.009566426277160645, + 0.00553936418145895, + 0.007838500663638115, + 0.005469714291393757, + 0.007975979708135128, + 0.0064932601526379585, + 0.010067526251077652, + 0.009525507688522339, + 0.014453411102294922, + 0.01117157656699419, + 0.01455004420131445, + 0.015575350262224674, + 0.019061924889683723, + 0.013336099684238434, + 0.01159766037017107, + 0.01616230048239231, + 0.015498490072786808, + 0.012738887220621109, + 0.01644054241478443, + 0.022973624989390373, + 0.022777866572141647, + 0.013280835933983326, + 0.011260808445513248, + 0.020081376656889915, + 0.006834924686700106, + 0.0069716330617666245, + 0.009966481477022171, + 0.01172005943953991, + 0.010274486616253853, + 0.005170299671590328, + 0.008415198884904385, + 0.010056954808533192, + 0.014888790436089039, + 0.0171574167907238, + 0.01882380060851574, + 0.01827283762395382, + 0.023500118404626846, + 0.016043124720454216, + 0.014119863510131836, + 0.01977584883570671, + 0.008604638278484344, + 0.007972294464707375, + 0.012494655326008797, + 0.01159907691180706, + 0.018838705494999886, + 0.008616065606474876, + 0.012415993958711624, + 0.008117714896798134, + 0.006452393718063831, + 0.002683438826352358, + 0.004476207308471203, + 0.005878340918570757, + 0.00601832615211606, + 0.004077897872775793, + 0.003820497775450349, + 0.004837538581341505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.03810957819223404 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive, multi-jointed robotic arm visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robotic arm's surface polished chrome that reflects light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the arm rotating within a vast hall containing a submerged reactor core with an eerie, pulsating azure glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool palette of deep blue, metallic silver, and shadowy black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008643713779747486 + ], + "opposite_logprob_tokens": [ + -4.758643627166748 + ], + "logprob_reward": -0.008643713779747486, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008643713779747486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10748563706874847, + 0.04068799316883087, + 0.04657763987779617, + 0.032054923474788666, + 0.23997445404529572, + 0.1405474841594696, + 0.03463085740804672, + 0.0986584946513176, + 0.1462903469800949, + 0.11309215426445007 + ], + "credit_map": [ + 0.016819020733237267, + 0.006919555831700563, + 0.00816427543759346, + 0.008964123204350471, + 0.01393857505172491, + 0.023796377703547478, + 0.011992597952485085, + 0.016891121864318848, + 0.004673017654567957, + 0.002582520479336381, + 0.0029586164746433496, + 0.004166182596236467, + 0.006626976188272238, + 0.007558902725577354, + 0.005669290199875832, + 0.006452489644289017, + 0.005455774255096912, + 0.006816494278609753, + 0.007213160395622253, + 0.0056601837277412415, + 0.006276197265833616, + 0.004495732020586729, + 0.005761456675827503, + 0.004898648709058762, + 0.004103605169802904, + 0.0032938546501100063, + 0.005526942200958729, + 0.0033629874233156443, + 0.00465399120002985, + 0.004098559729754925, + 0.004202932585030794, + 0.0028120537754148245, + 0.023366175591945648, + 0.03324573487043381, + 0.03146304190158844, + 0.02831941470503807, + 0.05169355124235153, + 0.024221213534474373, + 0.026917917653918266, + 0.020747419446706772, + 0.016526872292160988, + 0.011549343355000019, + 0.01751580648124218, + 0.01753127947449684, + 0.02480814792215824, + 0.020858243107795715, + 0.013614963740110397, + 0.01814284175634384, + 0.0032216899562627077, + 0.0029889133293181658, + 0.0036196995060890913, + 0.003803006839007139, + 0.004298865795135498, + 0.004329712595790625, + 0.004770956933498383, + 0.007598016876727343, + 0.010114160366356373, + 0.00802244246006012, + 0.01297753769904375, + 0.007898537442088127, + 0.015127144753932953, + 0.01755598746240139, + 0.01071703527122736, + 0.016245665028691292, + 0.013059440068900585, + 0.012536666356027126, + 0.021835660561919212, + 0.0176071859896183, + 0.024556977674365044, + 0.011824704706668854, + 0.027600068598985672, + 0.01726965606212616, + 0.01948491856455803, + 0.008805590681731701, + 0.014254485256969929, + 0.014287037774920464, + 0.013435912318527699, + 0.01548706367611885, + 0.0167645663022995, + 0.010572590865194798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 5 + ], + "selected_reward": -0.008643713779747486 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive, multi-jointed robotic arm visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robotic arm's surface polished chrome that reflects light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the arm rotating within a vast hall containing a submerged reactor core with an eerie, pulsating azure glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool palette of deep blue, metallic silver, and shadowy black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.1270108222961426 + ], + "opposite_logprob_tokens": [ + -0.127010777592659 + ], + "logprob_reward": -2.1270108222961426, + "yes_no_margin_reward": -2.0, + "token_credit_reward": -2.1270108222961426, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.14035199582576752, + 0.08894287049770355, + 0.07592430710792542, + 0.04441153258085251, + 0.262599915266037, + 0.07185149192810059, + 0.034690167754888535, + 0.08832038938999176, + 0.13968920707702637, + 0.053218159824609756 + ], + "credit_map": [ + 0.0215532835572958, + 0.009933387860655785, + 0.013354546390473843, + 0.012511184439063072, + 0.0200828704982996, + 0.026751047000288963, + 0.017677202820777893, + 0.018488477915525436, + 0.010267547331750393, + 0.006197756621986628, + 0.008596229366958141, + 0.009660407900810242, + 0.011386044323444366, + 0.022010622546076775, + 0.011253210715949535, + 0.009571054950356483, + 0.007824618369340897, + 0.009405137039721012, + 0.010984254069626331, + 0.01088256947696209, + 0.011058731004595757, + 0.007432336453348398, + 0.00882028043270111, + 0.009516378864645958, + 0.005360548850148916, + 0.0050951330922544, + 0.011950364336371422, + 0.004863767419010401, + 0.004828789737075567, + 0.003939363639801741, + 0.00566558912396431, + 0.002707978943362832, + 0.026506885886192322, + 0.0342661589384079, + 0.034311406314373016, + 0.029609359800815582, + 0.055515073239803314, + 0.02941286563873291, + 0.02915629744529724, + 0.023821856826543808, + 0.006147713400423527, + 0.005432573147118092, + 0.00886787660419941, + 0.008266720920801163, + 0.009400504641234875, + 0.009340006858110428, + 0.007351976353675127, + 0.01704411767423153, + 0.0045845345593988895, + 0.0034767494071274996, + 0.004399648401886225, + 0.003787699854001403, + 0.004398967605084181, + 0.006085811648517847, + 0.0031080390326678753, + 0.004848714452236891, + 0.01039769034832716, + 0.009885500185191631, + 0.009900640696287155, + 0.010931387543678284, + 0.013284023851156235, + 0.012214011512696743, + 0.009164277464151382, + 0.012542854063212872, + 0.010302718728780746, + 0.014115625061094761, + 0.01830504648387432, + 0.021454425528645515, + 0.031294532120227814, + 0.007928434759378433, + 0.01961033046245575, + 0.0166780985891819, + 0.009484016336500645, + 0.0036061278078705072, + 0.006475582718849182, + 0.006433793809264898, + 0.005859179422259331, + 0.006918060593307018, + 0.00853628572076559, + 0.005905112251639366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -2.1270108222961426 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive, multi-jointed robotic arm visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robotic arm's surface polished chrome that reflects light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the arm rotating within a vast hall containing a submerged reactor core with an eerie, pulsating azure glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool palette of deep blue, metallic silver, and shadowy black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019463420612737536 + ], + "opposite_logprob_tokens": [ + -6.251946449279785 + ], + "logprob_reward": -0.0019463420612737536, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019463420612737536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2380671501159668, + 0.0856483057141304, + 0.05374859273433685, + 0.042479466646909714, + 0.18959151208400726, + 0.08462513983249664, + 0.045746877789497375, + 0.052335284650325775, + 0.14876927435398102, + 0.05898839980363846 + ], + "credit_map": [ + 0.0314316488802433, + 0.016641369089484215, + 0.017284009605646133, + 0.01889556646347046, + 0.040791209787130356, + 0.04898938536643982, + 0.02658693492412567, + 0.03744702413678169, + 0.008793733082711697, + 0.008823278360068798, + 0.010665012523531914, + 0.009928856045007706, + 0.011483583599328995, + 0.013527275994420052, + 0.01084479596465826, + 0.01158177200704813, + 0.005838208366185427, + 0.005100231617689133, + 0.006158796604722738, + 0.005921937525272369, + 0.008906201459467411, + 0.0064983363263309, + 0.006147970445454121, + 0.009176908060908318, + 0.005314833018928766, + 0.004522404633462429, + 0.007897431962192059, + 0.00436120992526412, + 0.00685530062764883, + 0.004691355396062136, + 0.005695900414139032, + 0.003141032997518778, + 0.018593434244394302, + 0.024388546124100685, + 0.024994315579533577, + 0.021426543593406677, + 0.03933629393577576, + 0.022605804726481438, + 0.021245989948511124, + 0.01700058951973915, + 0.009455565363168716, + 0.007167017553001642, + 0.01076375413686037, + 0.010419485159218311, + 0.015400855801999569, + 0.01072449330240488, + 0.009242258965969086, + 0.011451699770987034, + 0.006669906899333, + 0.005498844664543867, + 0.005596476141363382, + 0.005123564973473549, + 0.006581995636224747, + 0.006534201558679342, + 0.004222603980451822, + 0.0055192806757986546, + 0.0062263538129627705, + 0.005745733622461557, + 0.006546399090439081, + 0.005904426332563162, + 0.00816843006759882, + 0.006054741330444813, + 0.007087694946676493, + 0.006601505912840366, + 0.012126054614782333, + 0.01498868502676487, + 0.01984563283622265, + 0.02392977476119995, + 0.03156311810016632, + 0.00969394575804472, + 0.019149238243699074, + 0.017472833395004272, + 0.010064234957098961, + 0.003707112977281213, + 0.008451069705188274, + 0.006313712801784277, + 0.006127968896180391, + 0.009498662315309048, + 0.00874853041023016, + 0.006077107973396778 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.0019463420612737536 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "static, cinematic shot captures a massive, multi-jointed robotic arm as it rotates majestically and deliberately within the vast hall of an atmospheric reactor. The arm's flawless, polished chrome surface acts like a distorted mirror, reflecting the eerie, pulsating azure glow of the submerged reactor core deep within the circular fissure below. The camera remains perfectly still, emphasizing the machinery's immense scale and contrasting starkly with the savage concrete architecture of the silent chamber. A fine mist hangs in the air, capturing and amplifying the sharp, volumetric, harsh white light that cuts through from unseen fixtures high in the cavernous ceiling. As the arm moves with an unstoppable, silent grace, the blue light twists and slides across its metallic contours, creating a liquid-like effect. The scene, dominated by a cool palette of deep blue, metallic silver, and shadowy black, creates a powerful, awe-inspiring atmosphere of energized and futuristic power, a powerful and awe-inspiring force under control.", + "seed": 19877, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_6/index_9415_seed_19877_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.3667482728138566, + "vlm_logprob": -0.3667482728138566, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.3667482728138566, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive, multi-jointed robotic arm visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robotic arm's surface polished chrome that reflects light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the arm rotating within a vast hall containing a submerged reactor core with an eerie, pulsating azure glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool palette of deep blue, metallic silver, and shadowy black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07171749323606491 + ], + "opposite_logprob_tokens": [ + -2.6967175006866455 + ], + "logprob_reward": -0.07171749323606491, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07171749323606491, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17462091147899628, + 0.07568053156137466, + 0.09898269921541214, + 0.08711422979831696, + 0.04990671947598457, + 0.06546586751937866, + 0.25591227412223816, + 0.12867173552513123, + 0.017515478655695915, + 0.04612957313656807 + ], + "credit_map": [ + 0.02347017079591751, + 0.020386187359690666, + 0.015063208527863026, + 0.023846488445997238, + 0.024627935141324997, + 0.02279946208000183, + 0.02515191026031971, + 0.019275566563010216, + 0.007312068250030279, + 0.012405375950038433, + 0.006016753148287535, + 0.012068063020706177, + 0.011651413515210152, + 0.007245311979204416, + 0.013095241039991379, + 0.00588630884885788, + 0.01187924388796091, + 0.013447158969938755, + 0.011072151362895966, + 0.009686214849352837, + 0.015478057786822319, + 0.015093089081346989, + 0.014998954720795155, + 0.007327822502702475, + 0.009174312464892864, + 0.013982624746859074, + 0.008233848959207535, + 0.010784253478050232, + 0.012649742886424065, + 0.01194656454026699, + 0.012929297983646393, + 0.007413591258227825, + 0.007541323080658913, + 0.008336084894835949, + 0.005830284673720598, + 0.005127560347318649, + 0.007011218462139368, + 0.004508747719228268, + 0.008000388741493225, + 0.0035511150490492582, + 0.012261604890227318, + 0.010654419660568237, + 0.009540379047393799, + 0.004832631908357143, + 0.010024630464613438, + 0.005987989250570536, + 0.005139152519404888, + 0.007025059778243303, + 0.02308659814298153, + 0.040543947368860245, + 0.02929263934493065, + 0.0390625074505806, + 0.05332862213253975, + 0.033293090760707855, + 0.021593127399683, + 0.01571173034608364, + 0.014268014580011368, + 0.020983997732400894, + 0.014784746803343296, + 0.0136198028922081, + 0.020155489444732666, + 0.017924124374985695, + 0.009849516674876213, + 0.01708604209125042, + 0.002297378843650222, + 0.0031693519558757544, + 0.0018313906621187925, + 0.0019967234693467617, + 0.0022960586939007044, + 0.002606435678899288, + 0.0017881059320643544, + 0.0015300348168238997, + 0.005837744567543268, + 0.008061506785452366, + 0.00686684250831604, + 0.004962701816111803, + 0.00621416000649333, + 0.005863599944859743, + 0.005099014844745398, + 0.0032240028958767653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.07171749323606491 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive, multi-jointed robotic arm visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robotic arm's surface polished chrome that reflects light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the arm rotating within a vast hall containing a submerged reactor core with an eerie, pulsating azure glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool palette of deep blue, metallic silver, and shadowy black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307127743959427 + ], + "opposite_logprob_tokens": [ + -3.1680712699890137 + ], + "logprob_reward": -0.04307127743959427, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307127743959427, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4457326829433441, + 0.11249952018260956, + 0.05502960830926895, + 0.06115623563528061, + 0.032538093626499176, + 0.05884614586830139, + 0.10519534349441528, + 0.0635785162448883, + 0.030012816190719604, + 0.03541112691164017 + ], + "credit_map": [ + 0.061451658606529236, + 0.05397489294409752, + 0.044820815324783325, + 0.05364394187927246, + 0.06323088705539703, + 0.05378288775682449, + 0.06249566376209259, + 0.05233193188905716, + 0.013708721846342087, + 0.01209295354783535, + 0.01328494492918253, + 0.015623548999428749, + 0.01625245064496994, + 0.017285630106925964, + 0.015903064981102943, + 0.008348195813596249, + 0.006130583118647337, + 0.007006142754107714, + 0.006328819785267115, + 0.0057605598121881485, + 0.009524168446660042, + 0.008620396256446838, + 0.006863264832645655, + 0.004795667715370655, + 0.007416848558932543, + 0.010014023631811142, + 0.006654633209109306, + 0.005934523418545723, + 0.008875280618667603, + 0.007489665877074003, + 0.008627518080174923, + 0.006143736187368631, + 0.003718703519552946, + 0.003964116331189871, + 0.004554474726319313, + 0.0037454054690897465, + 0.0047390214167535305, + 0.004288150928914547, + 0.004441940225660801, + 0.003086281241849065, + 0.005808608140796423, + 0.008503283374011517, + 0.003930739592760801, + 0.004612750373780727, + 0.011727293953299522, + 0.01101324800401926, + 0.0055205728858709335, + 0.007729643024504185, + 0.009908562526106834, + 0.015932219102978706, + 0.01332163717597723, + 0.015570604242384434, + 0.021435296162962914, + 0.014333313331007957, + 0.008288897573947906, + 0.006404804531484842, + 0.0068665239959955215, + 0.01040139701217413, + 0.007547952234745026, + 0.006789186969399452, + 0.009958875365555286, + 0.007828197441995144, + 0.005330040119588375, + 0.008856340311467648, + 0.004856747575104237, + 0.0037559978663921356, + 0.0033484448213130236, + 0.0037300502881407738, + 0.004089317284524441, + 0.004927820526063442, + 0.0028210151940584183, + 0.002483420306816697, + 0.0038874356541782618, + 0.005280163139104843, + 0.005921423900872469, + 0.0035270657390356064, + 0.005031263455748558, + 0.004677078686654568, + 0.004161547403782606, + 0.0029251468367874622 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.04307127743959427 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive, multi-jointed robotic arm visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robotic arm's surface polished chrome that reflects light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the arm rotating within a vast hall containing a submerged reactor core with an eerie, pulsating azure glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool palette of deep blue, metallic silver, and shadowy black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016077518463134766 + ], + "opposite_logprob_tokens": [ + -4.141077518463135 + ], + "logprob_reward": -0.016077518463134766, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016077518463134766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20988908410072327, + 0.09217952191829681, + 0.06780838221311569, + 0.08601024746894836, + 0.08707426488399506, + 0.14694388210773468, + 0.09882374107837677, + 0.10369382798671722, + 0.04947109520435333, + 0.0581059604883194 + ], + "credit_map": [ + 0.03249623253941536, + 0.02426244504749775, + 0.019217275083065033, + 0.02927866019308567, + 0.03070162422955036, + 0.02625166065990925, + 0.02658410556614399, + 0.021097077056765556, + 0.011424378491938114, + 0.011874007061123848, + 0.01227226760238409, + 0.014536041766405106, + 0.012873460538685322, + 0.012243570759892464, + 0.010335436090826988, + 0.006620357278734446, + 0.007483400870114565, + 0.008400972001254559, + 0.0064491224475204945, + 0.007831385359168053, + 0.011231627315282822, + 0.014215913601219654, + 0.006917242426425219, + 0.005278720986098051, + 0.011326081119477749, + 0.010519220493733883, + 0.007950765080749989, + 0.013000226579606533, + 0.010724222287535667, + 0.009936215355992317, + 0.01284085400402546, + 0.009712664410471916, + 0.01137420441955328, + 0.014569897204637527, + 0.010397076606750488, + 0.00948718748986721, + 0.012074689380824566, + 0.00869250949472189, + 0.01385585032403469, + 0.006622852757573128, + 0.018632622435688972, + 0.016387004405260086, + 0.014207121916115284, + 0.011626104824244976, + 0.022820547223091125, + 0.030797433108091354, + 0.013025643303990364, + 0.019447404891252518, + 0.009848805144429207, + 0.019446562975645065, + 0.007421562448143959, + 0.010867240838706493, + 0.01829293742775917, + 0.01217333972454071, + 0.009287362918257713, + 0.011485934257507324, + 0.01164968777447939, + 0.018358353525400162, + 0.009772468358278275, + 0.010789050720632076, + 0.017704425379633904, + 0.011194376274943352, + 0.008062316104769707, + 0.01616314798593521, + 0.004527064505964518, + 0.006177332252264023, + 0.004978025797754526, + 0.00574725354090333, + 0.007453459780663252, + 0.00798735860735178, + 0.006195288151502609, + 0.0064053102396428585, + 0.006480636540800333, + 0.008848230354487896, + 0.008607257157564163, + 0.005985707510262728, + 0.007846416905522346, + 0.008673849515616894, + 0.006260786205530167, + 0.005403077229857445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.016077518463134766 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive, multi-jointed robotic arm visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robotic arm's surface polished chrome that reflects light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the arm rotating within a vast hall containing a submerged reactor core with an eerie, pulsating azure glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool palette of deep blue, metallic silver, and shadowy black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.7015360593795776 + ], + "opposite_logprob_tokens": [ + -0.20153602957725525 + ], + "logprob_reward": -1.7015360593795776, + "yes_no_margin_reward": -1.5, + "token_credit_reward": -1.7015360593795776, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.36910656094551086, + 0.10758426785469055, + 0.09642909467220306, + 0.04671569913625717, + 0.10621597617864609, + 0.070010244846344, + 0.06736034899950027, + 0.04643720015883446, + 0.03706787899136543, + 0.05307267606258392 + ], + "credit_map": [ + 0.05202984809875488, + 0.04715663194656372, + 0.04215962812304497, + 0.03946002945303917, + 0.05300823971629143, + 0.04072576388716698, + 0.05956351384520531, + 0.035002898424863815, + 0.012732913717627525, + 0.01509827934205532, + 0.009672985412180424, + 0.013109114952385426, + 0.016383223235607147, + 0.012316412292420864, + 0.01784561201930046, + 0.010425728745758533, + 0.009167327545583248, + 0.011457090266048908, + 0.010506147518754005, + 0.013948431238532066, + 0.014399673789739609, + 0.013133207336068153, + 0.015337004326283932, + 0.008480213582515717, + 0.0057499585673213005, + 0.007742965128272772, + 0.00414593331515789, + 0.006421138532459736, + 0.005747983232140541, + 0.007456630934029818, + 0.0053538307547569275, + 0.004097257275134325, + 0.017229346558451653, + 0.017926111817359924, + 0.01296892948448658, + 0.009508085437119007, + 0.015078757889568806, + 0.011686492711305618, + 0.012875729240477085, + 0.008942521177232265, + 0.009180938825011253, + 0.010441985912621021, + 0.007821162231266499, + 0.0074632130563259125, + 0.01267239823937416, + 0.007585173472762108, + 0.00724533898755908, + 0.007600035052746534, + 0.007077101152390242, + 0.015676943585276604, + 0.0063092876225709915, + 0.007474476005882025, + 0.011656586080789566, + 0.005531485192477703, + 0.007767103146761656, + 0.005867370404303074, + 0.004766174592077732, + 0.007494369521737099, + 0.005386697128415108, + 0.005255701020359993, + 0.007111781742423773, + 0.005831219255924225, + 0.004847698844969273, + 0.00574355898424983, + 0.0036419080570340157, + 0.004426444880664349, + 0.005310199689120054, + 0.006745434366166592, + 0.00513701606541872, + 0.003621090902015567, + 0.004667030181735754, + 0.003518757177516818, + 0.007012042682617903, + 0.006844134535640478, + 0.008334130980074406, + 0.004697723314166069, + 0.008593033067882061, + 0.006495364010334015, + 0.004931195173412561, + 0.006165051367133856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -1.7015360593795776 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive, multi-jointed robotic arm visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robotic arm's surface polished chrome that reflects light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the arm rotating within a vast hall containing a submerged reactor core with an eerie, pulsating azure glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool palette of deep blue, metallic silver, and shadowy black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013390155509114265 + ], + "opposite_logprob_tokens": [ + -6.626338958740234 + ], + "logprob_reward": -0.0013390155509114265, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013390155509114265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2039181888103485, + 0.09649647027254105, + 0.0731339156627655, + 0.06648936867713928, + 0.08706147223711014, + 0.13362441956996918, + 0.09164859354496002, + 0.13125452399253845, + 0.05683185160160065, + 0.05954117327928543 + ], + "credit_map": [ + 0.021610043942928314, + 0.02703605219721794, + 0.023083405569195747, + 0.02320699952542782, + 0.028890378773212433, + 0.03226898983120918, + 0.03203563019633293, + 0.015786703675985336, + 0.010056078433990479, + 0.014946172013878822, + 0.0074781510047614574, + 0.011912227608263493, + 0.01586882956326008, + 0.014398043975234032, + 0.015010660514235497, + 0.0068263160064816475, + 0.006670408416539431, + 0.009277748875319958, + 0.008729856461286545, + 0.010721907950937748, + 0.01151771005243063, + 0.010126538574695587, + 0.010502802208065987, + 0.005586947314441204, + 0.009896911680698395, + 0.009697641246020794, + 0.007078921422362328, + 0.007028934545814991, + 0.00936390832066536, + 0.008658595383167267, + 0.007993743754923344, + 0.006770715117454529, + 0.01191625464707613, + 0.01645287312567234, + 0.007643624674528837, + 0.008251728489995003, + 0.012411052361130714, + 0.009337494149804115, + 0.011858112178742886, + 0.009190340526401997, + 0.016542673110961914, + 0.02112605981528759, + 0.017554624006152153, + 0.012104334309697151, + 0.0259938333183527, + 0.013377699069678783, + 0.01300430204719305, + 0.013920902274549007, + 0.008352803997695446, + 0.014336883090436459, + 0.010631127282977104, + 0.012851334176957607, + 0.01919550821185112, + 0.00991361029446125, + 0.009873717091977596, + 0.006493622437119484, + 0.013198607601225376, + 0.021295927464962006, + 0.014684069901704788, + 0.014485739171504974, + 0.020834490656852722, + 0.016701187938451767, + 0.012396583333611488, + 0.017657935619354248, + 0.0044560180976986885, + 0.005962377414107323, + 0.007071454543620348, + 0.01072795782238245, + 0.007724096532911062, + 0.007582915481179953, + 0.007608865853399038, + 0.005698172841221094, + 0.00977035891264677, + 0.008094101212918758, + 0.008486858569085598, + 0.0066496459767222404, + 0.008556128479540348, + 0.007984578609466553, + 0.0046589188277721405, + 0.005340591073036194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.0013390155509114265 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_7/index_tensor([9039, 9039])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_7/index_tensor([9039, 9039])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c966a15f12023c826f5fa486642ec1b70fbbcc5f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_7/index_tensor([9039, 9039])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium close-up captures a poignant scene, rendered in a stark graphic novel style, with thick, bold black ink and dramatic crosshatching shaping every form. The central focus is a Tapanuli orangutan mother, her form both monumental and gentle. Her long, shaggy cinnamon-colored fur is depicted not as limp hair but as a tapestry of richly textured lines, deeply crosshatched and imbued with immense volume. Her intelligent, dark eyes, set on a calm, weathered face, convey a serene, watchful expression. Her tiny infant clings to her chest, its tiny body a compact mass of light-colored fur, its hands and feet disappearing into the mother's thick coat. In a moment of profound tenderness, the mother's large, dark, leathery hand, marked with intricate ink lines, gently and deliberately offers a broad green leaf to the infant. The lighting is high-contrast and directional, casting deep, angular shadows that carve their figures from the minimalist background, creating a powerful, motionless frame that emphasizes the quiet intimacy of the moment and the silent, loving exchange between mother and child.", + "seed": 23557, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_7/index_9039_seed_23557_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.041366803413257006, + "vlm_logprob": -0.041366803413257006, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.041366803413257006, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Tapanuli orangutan mother with a tiny infant clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark graphic novel style with thick, bold black ink and dramatic crosshatching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mother's fur depicted as cinnamon-colored with richly textured lines and immense volume?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mother's hand gently offering a broad green leaf to the infant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framing a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04526781663298607 + ], + "opposite_logprob_tokens": [ + -3.1702678203582764 + ], + "logprob_reward": -0.04526781663298607, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04526781663298607, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2540224492549896, + 0.06818944960832596, + 0.03425367921590805, + 0.21323366463184357, + 0.03823050111532211, + 0.12766329944133759, + 0.08475606888532639, + 0.08616824448108673, + 0.054723598062992096, + 0.03875898942351341 + ], + "credit_map": [ + 0.017821287736296654, + 0.015525740571320057, + 0.01448198314756155, + 0.016317764297127724, + 0.027210120111703873, + 0.03580513224005699, + 0.07815568894147873, + 0.04870476573705673, + 0.006555647123605013, + 0.007495209574699402, + 0.007701207883656025, + 0.005678282584995031, + 0.010774962604045868, + 0.010053995065391064, + 0.011490019969642162, + 0.008440127596259117, + 0.0029030346777290106, + 0.0015512463869526982, + 0.0014758240431547165, + 0.00205156859010458, + 0.0024621798656880856, + 0.0035138018429279327, + 0.009775586426258087, + 0.010520439594984055, + 0.028603356331586838, + 0.011116295121610165, + 0.013909952715039253, + 0.02196759171783924, + 0.015313553623855114, + 0.0438808798789978, + 0.032627690583467484, + 0.04581436887383461, + 0.003697600681334734, + 0.0027522933669388294, + 0.003405017778277397, + 0.003315684152767062, + 0.003943716641515493, + 0.008775303140282631, + 0.004744370002299547, + 0.00759652117267251, + 0.018961239606142044, + 0.008581540547311306, + 0.0061567979864776134, + 0.011616404168307781, + 0.013515077531337738, + 0.018328849226236343, + 0.012295577675104141, + 0.03820781782269478, + 0.007859209552407265, + 0.007299025077372789, + 0.007170910947024822, + 0.010879624634981155, + 0.010112494230270386, + 0.010310347191989422, + 0.010526853613555431, + 0.020597616210579872, + 0.006681850645691156, + 0.007655810564756393, + 0.007923695258796215, + 0.010371957905590534, + 0.009731222875416279, + 0.012646426446735859, + 0.010313112288713455, + 0.020844174548983574, + 0.0052764760330319405, + 0.004447969142347574, + 0.0037446911446750164, + 0.006660636980086565, + 0.005846004001796246, + 0.008149131201207638, + 0.007726604584604502, + 0.012872090563178062, + 0.004254369530826807, + 0.0050032478757202625, + 0.00277310935780406, + 0.00344462669454515, + 0.008447565138339996, + 0.005262382794171572, + 0.003949644975364208, + 0.0056240432895720005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.04526781663298607 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Tapanuli orangutan mother with a tiny infant clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark graphic novel style with thick, bold black ink and dramatic crosshatching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mother's fur depicted as cinnamon-colored with richly textured lines and immense volume?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mother's hand gently offering a broad green leaf to the infant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framing a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032261162996292 + ], + "opposite_logprob_tokens": [ + -2.350322723388672 + ], + "logprob_reward": -0.10032261162996292, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032261162996292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13140621781349182, + 0.23698492348194122, + 0.08411208540201187, + 0.19447489082813263, + 0.08148728311061859, + 0.03605467081069946, + 0.0643530860543251, + 0.031798701733350754, + 0.06802354753017426, + 0.07130462676286697 + ], + "credit_map": [ + 0.007214134093374014, + 0.011442101560533047, + 0.009293876588344574, + 0.0095980828627944, + 0.019246449694037437, + 0.017183315008878708, + 0.03530604764819145, + 0.022122206166386604, + 0.02060707099735737, + 0.024860691279172897, + 0.01970791630446911, + 0.018505962565541267, + 0.038948625326156616, + 0.03696835786104202, + 0.057427991181612015, + 0.019958319142460823, + 0.0060349805280566216, + 0.004352718126028776, + 0.004509745165705681, + 0.003720309119671583, + 0.007492017466574907, + 0.011026682332158089, + 0.024309100583195686, + 0.022666532546281815, + 0.023757802322506905, + 0.012593360617756844, + 0.013821219094097614, + 0.022096989676356316, + 0.02279992774128914, + 0.032114963978528976, + 0.05525065213441849, + 0.012039966881275177, + 0.005290142726153135, + 0.005992973688989878, + 0.004558113403618336, + 0.006603326182812452, + 0.007546715438365936, + 0.007283839862793684, + 0.020780395716428757, + 0.02343177981674671, + 0.004734768997877836, + 0.0047048539854586124, + 0.002894728910177946, + 0.004155528265982866, + 0.0055186268873512745, + 0.0045716771855950356, + 0.00457259314134717, + 0.004901893902570009, + 0.004471937194466591, + 0.006248634308576584, + 0.007868091575801373, + 0.011533667333424091, + 0.008087949827313423, + 0.008617152459919453, + 0.008492942899465561, + 0.009032714180648327, + 0.002541492460295558, + 0.0018866710597649217, + 0.0017052139155566692, + 0.002334397053346038, + 0.0031265178695321083, + 0.004801616072654724, + 0.004847781732678413, + 0.010555011220276356, + 0.0056908102706074715, + 0.007137455977499485, + 0.007799159269779921, + 0.008897163905203342, + 0.012419884093105793, + 0.00909188948571682, + 0.010488010942935944, + 0.006499174050986767, + 0.006456208880990744, + 0.007098787929862738, + 0.004709325265139341, + 0.004081584047526121, + 0.012639911845326424, + 0.011269798502326012, + 0.006281588692218065, + 0.01876741833984852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.10032261162996292 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Tapanuli orangutan mother with a tiny infant clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark graphic novel style with thick, bold black ink and dramatic crosshatching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mother's fur depicted as cinnamon-colored with richly textured lines and immense volume?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mother's hand gently offering a broad green leaf to the infant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framing a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028242485132068396 + ], + "opposite_logprob_tokens": [ + -5.877824306488037 + ], + "logprob_reward": -0.0028242485132068396, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028242485132068396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09343836456537247, + 0.28104084730148315, + 0.12852665781974792, + 0.05579997971653938, + 0.06250631809234619, + 0.029692519456148148, + 0.06255634129047394, + 0.10943475365638733, + 0.1107393130660057, + 0.06626493483781815 + ], + "credit_map": [ + 0.007898173294961452, + 0.009324431419372559, + 0.009277569130063057, + 0.008457733318209648, + 0.013747400604188442, + 0.01615067943930626, + 0.013681533746421337, + 0.014900846406817436, + 0.023437052965164185, + 0.03667458891868591, + 0.03057090938091278, + 0.02434440515935421, + 0.056189924478530884, + 0.03491439297795296, + 0.04513005167245865, + 0.029779504984617233, + 0.016540810465812683, + 0.0062165698036551476, + 0.006596026476472616, + 0.009195573627948761, + 0.011159003712236881, + 0.028257962316274643, + 0.0426342710852623, + 0.00792643427848816, + 0.004480279982089996, + 0.004642178304493427, + 0.004592482000589371, + 0.010443431325256824, + 0.006280200090259314, + 0.006990023888647556, + 0.009462199173867702, + 0.008909187279641628, + 0.0041072433814406395, + 0.004207937978208065, + 0.004898827522993088, + 0.004844884853810072, + 0.006264631170779467, + 0.00937202014029026, + 0.010586457327008247, + 0.01822431944310665, + 0.0019095026655122638, + 0.004443824291229248, + 0.0022720550186932087, + 0.004665903281420469, + 0.006174744106829166, + 0.0033609422389417887, + 0.004256165120750666, + 0.0026093816850334406, + 0.006474326830357313, + 0.005183279048651457, + 0.00440359627828002, + 0.0061722504906356335, + 0.007214225362986326, + 0.00807186122983694, + 0.007433763705193996, + 0.017603036016225815, + 0.008831650950014591, + 0.01136479526758194, + 0.012005976401269436, + 0.01579323038458824, + 0.013457856141030788, + 0.013792723417282104, + 0.013417014852166176, + 0.020771509036421776, + 0.0064355130307376385, + 0.005951746366918087, + 0.025605667382478714, + 0.007637286093086004, + 0.010095251724123955, + 0.01204884797334671, + 0.011451855301856995, + 0.031513143330812454, + 0.004214981570839882, + 0.0054739960469305515, + 0.00484224921092391, + 0.003819542471319437, + 0.007657142821699381, + 0.017018694430589676, + 0.004890508018434048, + 0.01834782212972641 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 8 + ], + "selected_reward": -0.0028242485132068396 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Tapanuli orangutan mother with a tiny infant clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark graphic novel style with thick, bold black ink and dramatic crosshatching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mother's fur depicted as cinnamon-colored with richly textured lines and immense volume?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mother's hand gently offering a broad green leaf to the infant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framing a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009777394123375416 + ], + "opposite_logprob_tokens": [ + -4.634777545928955 + ], + "logprob_reward": -0.009777394123375416, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009777394123375416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2779480218887329, + 0.06643544882535934, + 0.08009453117847443, + 0.24513983726501465, + 0.029497861862182617, + 0.016740962862968445, + 0.029821833595633507, + 0.18187306821346283, + 0.023552287369966507, + 0.04889621585607529 + ], + "credit_map": [ + 0.019683528691530228, + 0.015532735735177994, + 0.017364351078867912, + 0.0174870565533638, + 0.02825314737856388, + 0.04087289422750473, + 0.08711724728345871, + 0.05163702741265297, + 0.005466231610625982, + 0.005974698346108198, + 0.006200669799000025, + 0.005109545774757862, + 0.012815933674573898, + 0.0115215964615345, + 0.013779607601463795, + 0.005567159038037062, + 0.006659364327788353, + 0.006472609471529722, + 0.006265562027692795, + 0.014447389170527458, + 0.009268653579056263, + 0.009961477480828762, + 0.01882830075919628, + 0.008191166445612907, + 0.03008526936173439, + 0.014835665002465248, + 0.017473282292485237, + 0.03416842222213745, + 0.018017686903476715, + 0.04607921466231346, + 0.03374553844332695, + 0.050734743475914, + 0.0025455281138420105, + 0.002596887992694974, + 0.002791525097563863, + 0.0022626693826168776, + 0.0034666010178625584, + 0.005902432836592197, + 0.0030497435946017504, + 0.006882470101118088, + 0.0022828488145023584, + 0.0011612316593527794, + 0.0012748544104397297, + 0.0024627759121358395, + 0.0015784018905833364, + 0.002332546981051564, + 0.0018752196338027716, + 0.0037730818148702383, + 0.002916926983743906, + 0.002156822942197323, + 0.0023815843742340803, + 0.002670266665518284, + 0.0033532658126205206, + 0.003996709361672401, + 0.003601323114708066, + 0.008744930848479271, + 0.014349622651934624, + 0.014800979755818844, + 0.01299494132399559, + 0.02363993227481842, + 0.01739724539220333, + 0.025716260075569153, + 0.02245408296585083, + 0.05051996931433678, + 0.00169843016192317, + 0.0021807574667036533, + 0.0016440397594124079, + 0.0024362073745578527, + 0.0029278004076331854, + 0.0034902843181043863, + 0.0029863102827221155, + 0.006188454572111368, + 0.007183225825428963, + 0.003570766421034932, + 0.0037993632722646, + 0.0026870914734899998, + 0.006445401348173618, + 0.014471005648374557, + 0.005322474520653486, + 0.005416878964751959 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.009777394123375416 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Tapanuli orangutan mother with a tiny infant clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark graphic novel style with thick, bold black ink and dramatic crosshatching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mother's fur depicted as cinnamon-colored with richly textured lines and immense volume?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mother's hand gently offering a broad green leaf to the infant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framing a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864194616675377 + ], + "opposite_logprob_tokens": [ + -3.0486419200897217 + ], + "logprob_reward": -0.04864194616675377, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864194616675377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04407317936420441, + 0.13811174035072327, + 0.11563971638679504, + 0.34393471479415894, + 0.11216694116592407, + 0.10679192841053009, + 0.05367625877261162, + 0.01607130654156208, + 0.03383096680045128, + 0.03570322319865227 + ], + "credit_map": [ + 0.002461126074194908, + 0.00456066383048892, + 0.0027901947032660246, + 0.0038973295595496893, + 0.007192407269030809, + 0.005156687460839748, + 0.010444177314639091, + 0.00757059408351779, + 0.012771778739988804, + 0.016316240653395653, + 0.01263484451919794, + 0.013148637488484383, + 0.025641415268182755, + 0.01923958770930767, + 0.01953805424273014, + 0.01882120594382286, + 0.010096779093146324, + 0.005828456953167915, + 0.006199072115123272, + 0.007394528016448021, + 0.009833297692239285, + 0.019281750544905663, + 0.039378806948661804, + 0.01762702874839306, + 0.047631870955228806, + 0.01801462285220623, + 0.02217322587966919, + 0.0302641112357378, + 0.026214851066470146, + 0.07212972640991211, + 0.05641854181885719, + 0.0710877925157547, + 0.0061271097511053085, + 0.008018734864890575, + 0.006353336852043867, + 0.009755350649356842, + 0.011490695178508759, + 0.029696976765990257, + 0.033668164163827896, + 0.0070565794594585896, + 0.015335765667259693, + 0.009192457422614098, + 0.005710021127015352, + 0.008670439943671227, + 0.01519928127527237, + 0.014632356353104115, + 0.0101645952090621, + 0.027887018397450447, + 0.005560908932238817, + 0.00539438147097826, + 0.004276301246136427, + 0.006537271197885275, + 0.0069100274704396725, + 0.006418049801141024, + 0.00634614797309041, + 0.012233174405992031, + 0.0012206364190205932, + 0.001749458839185536, + 0.00209537404589355, + 0.0027917022816836834, + 0.0018884832970798016, + 0.002215081825852394, + 0.0026297811418771744, + 0.0014807903207838535, + 0.0025531246792525053, + 0.0023341935593634844, + 0.0029121162369847298, + 0.002605888294056058, + 0.0027444143779575825, + 0.005786967929452658, + 0.00376753113232553, + 0.011126731522381306, + 0.0029661713633686304, + 0.0021190079860389233, + 0.0017545046284794807, + 0.0017312305280938745, + 0.00287256040610373, + 0.009901727549731731, + 0.002430358435958624, + 0.011927668005228043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.04864194616675377 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium close-up captures a poignant scene, rendered in a stark graphic novel style, with thick, bold black ink and dramatic crosshatching shaping every form. The central focus is a Tapanuli orangutan mother, her form both monumental and gentle. Her long, shaggy cinnamon-colored fur is depicted not as limp hair but as a tapestry of richly textured lines, deeply crosshatched and imbued with immense volume. Her intelligent, dark eyes, set on a calm, weathered face, convey a serene, watchful expression. Her tiny infant clings to her chest, its tiny body a compact mass of light-colored fur, its hands and feet disappearing into the mother's thick coat. In a moment of profound tenderness, the mother's large, dark, leathery hand, marked with intricate ink lines, gently and deliberately offers a broad green leaf to the infant. The lighting is high-contrast and directional, casting deep, angular shadows that carve their figures from the minimalist background, creating a powerful, motionless frame that emphasizes the quiet intimacy of the moment and the silent, loving exchange between mother and child.", + "seed": 23558, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_7/index_9039_seed_23558_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06516417833045125, + "vlm_logprob": -0.06516417833045125, + "vlm_yes_no_margin": 3.2249999761581423, + "vlm_token_credit": -0.06516417833045125, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Tapanuli orangutan mother with a tiny infant clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark graphic novel style with thick, bold black ink and dramatic crosshatching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mother's fur depicted as cinnamon-colored with richly textured lines and immense volume?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mother's hand gently offering a broad green leaf to the infant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framing a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.045916229486465454 + ], + "opposite_logprob_tokens": [ + -3.1709163188934326 + ], + "logprob_reward": -0.045916229486465454, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.045916229486465454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20571893453598022, + 0.14762338995933533, + 0.12632636725902557, + 0.15582875907421112, + 0.05018151178956032, + 0.030898837372660637, + 0.03561273217201233, + 0.15075325965881348, + 0.06588739901781082, + 0.031168825924396515 + ], + "credit_map": [ + 0.017964942380785942, + 0.02057952806353569, + 0.04585227742791176, + 0.028942350298166275, + 0.023746967315673828, + 0.013738550245761871, + 0.022193053737282753, + 0.03270125761628151, + 0.015330838039517403, + 0.01623508892953396, + 0.02346520684659481, + 0.02232757769525051, + 0.018643245100975037, + 0.013633158057928085, + 0.015001138672232628, + 0.022987131029367447, + 0.010430988855659962, + 0.011746914125978947, + 0.014509853906929493, + 0.019180087372660637, + 0.017246680334210396, + 0.013440025970339775, + 0.016157003119587898, + 0.023614803329110146, + 0.019015930593013763, + 0.008744136430323124, + 0.038571327924728394, + 0.01631859503686428, + 0.018162483349442482, + 0.013603233732283115, + 0.018258176743984222, + 0.023154856637120247, + 0.004275924060493708, + 0.003933559637516737, + 0.0072058020159602165, + 0.006829243153333664, + 0.008243105374276638, + 0.003953324165195227, + 0.00723715266212821, + 0.008503403514623642, + 0.0025524189695715904, + 0.0030036221724003553, + 0.003847314976155758, + 0.005175876896828413, + 0.003306424943730235, + 0.0040632206946611404, + 0.0036173586267977953, + 0.0053325979970395565, + 0.003358231158927083, + 0.0030926004983484745, + 0.004526512231677771, + 0.0035649628844112158, + 0.0045321728102862835, + 0.005220583640038967, + 0.0051893084309995174, + 0.006128362379968166, + 0.012276342138648033, + 0.014482256025075912, + 0.016518205404281616, + 0.015754196792840958, + 0.018337104469537735, + 0.0151429558172822, + 0.03570947051048279, + 0.022532712668180466, + 0.004471919033676386, + 0.006810989696532488, + 0.010798669420182705, + 0.007150172255933285, + 0.009702206589281559, + 0.007032504305243492, + 0.009467702358961105, + 0.010453231632709503, + 0.0027167422231286764, + 0.003567293519154191, + 0.005644786637276411, + 0.0029022947419434786, + 0.0027013772632926702, + 0.0034429472871124744, + 0.003967648837715387, + 0.006225732155144215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.045916229486465454 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Tapanuli orangutan mother with a tiny infant clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark graphic novel style with thick, bold black ink and dramatic crosshatching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mother's fur depicted as cinnamon-colored with richly textured lines and immense volume?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mother's hand gently offering a broad green leaf to the infant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framing a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14281854033470154 + ], + "opposite_logprob_tokens": [ + -2.0178184509277344 + ], + "logprob_reward": -0.14281854033470154, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14281854033470154, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20552368462085724, + 0.23540492355823517, + 0.1581932008266449, + 0.06274031102657318, + 0.07465343177318573, + 0.03781792148947716, + 0.03533853963017464, + 0.04604087397456169, + 0.04916417971253395, + 0.09512291103601456 + ], + "credit_map": [ + 0.015012305229902267, + 0.019076300784945488, + 0.05498580262064934, + 0.01806517131626606, + 0.022896723821759224, + 0.01998710073530674, + 0.0301672276109457, + 0.025333061814308167, + 0.0253616850823164, + 0.020356466993689537, + 0.030222652480006218, + 0.03448980674147606, + 0.029989192262291908, + 0.032853707671165466, + 0.02501651830971241, + 0.03711489215493202, + 0.013007430359721184, + 0.01337567251175642, + 0.017627455294132233, + 0.021198896691203117, + 0.02240075170993805, + 0.021745406091213226, + 0.02132675051689148, + 0.02751084230840206, + 0.005918330512940884, + 0.005323187913745642, + 0.008774920366704464, + 0.009146166034042835, + 0.006561746820807457, + 0.008708421140909195, + 0.007506316527724266, + 0.0108012231066823, + 0.00797104649245739, + 0.005280727054923773, + 0.011187492869794369, + 0.013106616213917732, + 0.006779211573302746, + 0.008982641622424126, + 0.00844390969723463, + 0.012901781126856804, + 0.00437066238373518, + 0.0028945638332515955, + 0.007291312795132399, + 0.0053269462659955025, + 0.003974254243075848, + 0.003190350951626897, + 0.0040319692343473434, + 0.006737858522683382, + 0.0026870928704738617, + 0.003627239726483822, + 0.004975879564881325, + 0.004468080122023821, + 0.00493662990629673, + 0.0029954537749290466, + 0.0067421323619782925, + 0.0049060313031077385, + 0.0029893165919929743, + 0.006122838705778122, + 0.006743211764842272, + 0.0063379500061273575, + 0.004884809721261263, + 0.00743242260068655, + 0.004432260058820248, + 0.007098067086189985, + 0.004871127661317587, + 0.006872713100165129, + 0.005130717996507883, + 0.006706863176077604, + 0.004997631534934044, + 0.0053949542343616486, + 0.008647115901112556, + 0.0065430570393800735, + 0.008403315208852291, + 0.011177328415215015, + 0.0167061947286129, + 0.011697436682879925, + 0.009190907701849937, + 0.01053458359092474, + 0.010617966763675213, + 0.01679518073797226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.14281854033470154 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Tapanuli orangutan mother with a tiny infant clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark graphic novel style with thick, bold black ink and dramatic crosshatching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mother's fur depicted as cinnamon-colored with richly textured lines and immense volume?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mother's hand gently offering a broad green leaf to the infant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framing a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003198272082954645 + ], + "opposite_logprob_tokens": [ + -5.753198146820068 + ], + "logprob_reward": -0.003198272082954645, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003198272082954645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27023810148239136, + 0.06397852301597595, + 0.2193423956632614, + 0.07126622647047043, + 0.047805238515138626, + 0.03484725579619408, + 0.08628268539905548, + 0.1236264780163765, + 0.026834558695554733, + 0.05577851086854935 + ], + "credit_map": [ + 0.034180838614702225, + 0.023047318682074547, + 0.08396552503108978, + 0.02582416869699955, + 0.02620684541761875, + 0.01810149848461151, + 0.018839804455637932, + 0.040072113275527954, + 0.008785351179540157, + 0.005385722033679485, + 0.012645838782191277, + 0.008558901026844978, + 0.006499051582068205, + 0.0071059586480259895, + 0.006298927124589682, + 0.008698769845068455, + 0.018308328464627266, + 0.02763490565121174, + 0.02519058622419834, + 0.0349728986620903, + 0.029217537492513657, + 0.020830543711781502, + 0.02683195285499096, + 0.0363556370139122, + 0.006353689357638359, + 0.005539585370570421, + 0.014828486368060112, + 0.009562164545059204, + 0.007235003635287285, + 0.007616905961185694, + 0.008361609652638435, + 0.011768779717385769, + 0.004017829429358244, + 0.002889574272558093, + 0.006714163813740015, + 0.006535406224429607, + 0.008124791085720062, + 0.005438090302050114, + 0.005960377398878336, + 0.008125008083879948, + 0.004475828260183334, + 0.0032366712111979723, + 0.005697927437722683, + 0.00658479705452919, + 0.0030457735992968082, + 0.003609229577705264, + 0.003067710204049945, + 0.005129320081323385, + 0.004782580770552158, + 0.00843722652643919, + 0.008846507407724857, + 0.013360475189983845, + 0.012478847056627274, + 0.007616941351443529, + 0.01446709968149662, + 0.016293006017804146, + 0.007175751496106386, + 0.00977326463907957, + 0.014985291287302971, + 0.016026170924305916, + 0.010997229255735874, + 0.010622273199260235, + 0.03605615720152855, + 0.017990339547395706, + 0.002329880138859153, + 0.002444556448608637, + 0.004393241833895445, + 0.003409345168620348, + 0.0021023047156631947, + 0.003941005561500788, + 0.0034592736046761274, + 0.004754952155053616, + 0.004024883266538382, + 0.00524780061095953, + 0.011848434805870056, + 0.006339838728308678, + 0.00480756675824523, + 0.005958942696452141, + 0.006765204481780529, + 0.010785840451717377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.003198272082954645 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Tapanuli orangutan mother with a tiny infant clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark graphic novel style with thick, bold black ink and dramatic crosshatching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mother's fur depicted as cinnamon-colored with richly textured lines and immense volume?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mother's hand gently offering a broad green leaf to the infant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framing a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05493883788585663 + ], + "opposite_logprob_tokens": [ + -2.929938793182373 + ], + "logprob_reward": -0.05493883788585663, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493883788585663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16596385836601257, + 0.08658092468976974, + 0.12983380258083344, + 0.05439468100667, + 0.05708545818924904, + 0.04121488332748413, + 0.029650326818227768, + 0.3724376857280731, + 0.029236841946840286, + 0.033601514995098114 + ], + "credit_map": [ + 0.017691442742943764, + 0.018006794154644012, + 0.03257587179541588, + 0.02228059619665146, + 0.024643991142511368, + 0.009590496309101582, + 0.016754552721977234, + 0.024420134723186493, + 0.006826141383498907, + 0.00964593980461359, + 0.01159355603158474, + 0.015394382178783417, + 0.012239539995789528, + 0.009269185364246368, + 0.008102420717477798, + 0.013509765267372131, + 0.01332664955407381, + 0.013319138437509537, + 0.01618845947086811, + 0.019947310909628868, + 0.015310297720134258, + 0.01286244485527277, + 0.016506386920809746, + 0.022373124957084656, + 0.004631453193724155, + 0.0039740195497870445, + 0.00686303898692131, + 0.0073215896263718605, + 0.005778979975730181, + 0.008176110684871674, + 0.005983301438391209, + 0.011666189879179, + 0.0052824425511062145, + 0.003533638082444668, + 0.0061258189380168915, + 0.010370726697146893, + 0.008337792940437794, + 0.007542160339653492, + 0.005554725416004658, + 0.010338153690099716, + 0.0034561518114060163, + 0.0038247883785516024, + 0.005976792890578508, + 0.0068081957288086414, + 0.004747668746858835, + 0.004955071024596691, + 0.004495580680668354, + 0.006950635928660631, + 0.0033277058973908424, + 0.004181401338428259, + 0.003825870342552662, + 0.003441462991759181, + 0.0030720781069248915, + 0.0035716562997549772, + 0.004094329662621021, + 0.004135820548981428, + 0.023578669875860214, + 0.03440665453672409, + 0.03670399636030197, + 0.04542567580938339, + 0.0413997657597065, + 0.033319585025310516, + 0.09992583096027374, + 0.05767751485109329, + 0.0024048639461398125, + 0.002923699328675866, + 0.0036776140332221985, + 0.0029746301006525755, + 0.0042788623832166195, + 0.0032374465372413397, + 0.004653017967939377, + 0.005086709279567003, + 0.0021930611692368984, + 0.003969988785684109, + 0.003258325392380357, + 0.00384234800003469, + 0.0038196807727217674, + 0.002481686184182763, + 0.008891753852367401, + 0.005144672933965921 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.05493883788585663 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Tapanuli orangutan mother with a tiny infant clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark graphic novel style with thick, bold black ink and dramatic crosshatching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mother's fur depicted as cinnamon-colored with richly textured lines and immense volume?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mother's hand gently offering a broad green leaf to the infant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framing a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07894901186227798 + ], + "opposite_logprob_tokens": [ + -2.578948974609375 + ], + "logprob_reward": -0.07894901186227798, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07894901186227798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31675365567207336, + 0.09929532557725906, + 0.28584375977516174, + 0.045706287026405334, + 0.05302415415644646, + 0.025401823222637177, + 0.036073122173547745, + 0.10293393582105637, + 0.020212220028042793, + 0.014755669981241226 + ], + "credit_map": [ + 0.03211182728409767, + 0.018324987962841988, + 0.0998997762799263, + 0.029996581375598907, + 0.02780814841389656, + 0.02995927259325981, + 0.040686365216970444, + 0.03796672448515892, + 0.00810934230685234, + 0.011020446196198463, + 0.015520481392741203, + 0.014178924262523651, + 0.010912997648119926, + 0.011125966906547546, + 0.012637672014534473, + 0.015789499506354332, + 0.02558344416320324, + 0.03759640455245972, + 0.025981254875659943, + 0.04304597154259682, + 0.03942026570439339, + 0.027391783893108368, + 0.03595602139830589, + 0.050868622958660126, + 0.004062135238200426, + 0.005117752589285374, + 0.00728304497897625, + 0.006898924708366394, + 0.00552284624427557, + 0.004976099822670221, + 0.004853004589676857, + 0.006992483977228403, + 0.004732698202133179, + 0.004726127255707979, + 0.0062072183936834335, + 0.006474933121353388, + 0.009350636973977089, + 0.00463598920032382, + 0.00792230200022459, + 0.00897425040602684, + 0.003029267303645611, + 0.0026245603803545237, + 0.003979350905865431, + 0.0034388615749776363, + 0.0036990002263337374, + 0.002282478380948305, + 0.002580131171271205, + 0.0037681744433939457, + 0.0033892602659761906, + 0.0053407889790833, + 0.004179308656603098, + 0.005475960206240416, + 0.004172181244939566, + 0.0033103213645517826, + 0.0036627112422138453, + 0.006542595103383064, + 0.007707909680902958, + 0.015313999727368355, + 0.013551535084843636, + 0.011665153317153454, + 0.010553883388638496, + 0.01183015014976263, + 0.015118609182536602, + 0.017192702740430832, + 0.0018508898792788386, + 0.00224257237277925, + 0.002626068191602826, + 0.0021998833399266005, + 0.0024029300548136234, + 0.0025073655415326357, + 0.003420709865167737, + 0.002961802529171109, + 0.0014171019429340959, + 0.001819137716665864, + 0.0022261403501033783, + 0.002112913178279996, + 0.0015612763818353415, + 0.0016126161208376288, + 0.0019269188633188605, + 0.0020795671734958887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.07894901186227798 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium close-up captures a poignant scene, rendered in a stark graphic novel style, with thick, bold black ink and dramatic crosshatching shaping every form. The central focus is a Tapanuli orangutan mother, her form both monumental and gentle. Her long, shaggy cinnamon-colored fur is depicted not as limp hair but as a tapestry of richly textured lines, deeply crosshatched and imbued with immense volume. Her intelligent, dark eyes, set on a calm, weathered face, convey a serene, watchful expression. Her tiny infant clings to her chest, its tiny body a compact mass of light-colored fur, its hands and feet disappearing into the mother's thick coat. In a moment of profound tenderness, the mother's large, dark, leathery hand, marked with intricate ink lines, gently and deliberately offers a broad green leaf to the infant. The lighting is high-contrast and directional, casting deep, angular shadows that carve their figures from the minimalist background, creating a powerful, motionless frame that emphasizes the quiet intimacy of the moment and the silent, loving exchange between mother and child.", + "seed": 23559, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_7/index_9039_seed_23559_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05150234838947654, + "vlm_logprob": -0.05150234838947654, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.05150234838947654, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Tapanuli orangutan mother with a tiny infant clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark graphic novel style with thick, bold black ink and dramatic crosshatching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mother's fur depicted as cinnamon-colored with richly textured lines and immense volume?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mother's hand gently offering a broad green leaf to the infant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framing a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07256763428449631 + ], + "opposite_logprob_tokens": [ + -2.6975677013397217 + ], + "logprob_reward": -0.07256763428449631, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07256763428449631, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14776715636253357, + 0.05973786488175392, + 0.32272136211395264, + 0.05252944305539131, + 0.10285090655088425, + 0.055778637528419495, + 0.1032341942191124, + 0.026452558115124702, + 0.02581404522061348, + 0.10311383754014969 + ], + "credit_map": [ + 0.012151563540101051, + 0.029513875022530556, + 0.016541510820388794, + 0.013301273807883263, + 0.027567310258746147, + 0.01173121016472578, + 0.017894774675369263, + 0.019065631553530693, + 0.006456192582845688, + 0.013230432756245136, + 0.0072965845465660095, + 0.005474538542330265, + 0.009319484233856201, + 0.006987953092902899, + 0.005320415832102299, + 0.005652262829244137, + 0.020969104021787643, + 0.013343166559934616, + 0.11910872161388397, + 0.018327821046113968, + 0.01864955574274063, + 0.09498851001262665, + 0.011990613304078579, + 0.025343863293528557, + 0.005642076954245567, + 0.008199229836463928, + 0.004132848232984543, + 0.005573865957558155, + 0.006666588131338358, + 0.0049153040163218975, + 0.010383290238678455, + 0.007016239687800407, + 0.007980442605912685, + 0.016891857609152794, + 0.016620945185422897, + 0.010287554934620857, + 0.012053050100803375, + 0.019716007634997368, + 0.009374579414725304, + 0.009926465339958668, + 0.006500832736492157, + 0.0077263182029128075, + 0.009626733139157295, + 0.006291114259511232, + 0.0066965799778699875, + 0.009187422692775726, + 0.004776935558766127, + 0.004972703289240599, + 0.012595321983098984, + 0.014150988310575485, + 0.012414691969752312, + 0.011792545206844807, + 0.013942176476120949, + 0.01725568063557148, + 0.009454567916691303, + 0.011628220789134502, + 0.0024321170058101416, + 0.0029704465996474028, + 0.005920721683651209, + 0.0021856490056961775, + 0.0025385606568306684, + 0.0018065980402752757, + 0.0016967960400506854, + 0.006901669315993786, + 0.0022419351153075695, + 0.004143994767218828, + 0.0020529390312731266, + 0.002557475818321109, + 0.002818251959979534, + 0.0032708863727748394, + 0.0029740980826318264, + 0.005754463840276003, + 0.011668977327644825, + 0.010472471825778484, + 0.009737161919474602, + 0.018476974219083786, + 0.008839011192321777, + 0.009866103529930115, + 0.01858099177479744, + 0.01547214388847351 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.07256763428449631 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Tapanuli orangutan mother with a tiny infant clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark graphic novel style with thick, bold black ink and dramatic crosshatching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mother's fur depicted as cinnamon-colored with richly textured lines and immense volume?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mother's hand gently offering a broad green leaf to the infant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framing a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371596708893776 + ], + "opposite_logprob_tokens": [ + -3.4087159633636475 + ], + "logprob_reward": -0.03371596708893776, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371596708893776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11138395965099335, + 0.09741900116205215, + 0.23693257570266724, + 0.14662505686283112, + 0.05186036229133606, + 0.0862530767917633, + 0.06485211104154587, + 0.05595066770911217, + 0.11584987491369247, + 0.03287334367632866 + ], + "credit_map": [ + 0.012871784158051014, + 0.015709495171904564, + 0.019395392388105392, + 0.010307627730071545, + 0.014572439715266228, + 0.010375695303082466, + 0.009585337713360786, + 0.018566187471151352, + 0.013485969975590706, + 0.014037085697054863, + 0.013918884098529816, + 0.01578349992632866, + 0.010768813081085682, + 0.010502994060516357, + 0.009206273593008518, + 0.009715481661260128, + 0.0190149936825037, + 0.01045900583267212, + 0.08381838351488113, + 0.015696324408054352, + 0.013065110892057419, + 0.05416039749979973, + 0.01216601487249136, + 0.028552349656820297, + 0.015737401321530342, + 0.02029578387737274, + 0.011723421514034271, + 0.01407947763800621, + 0.013379460200667381, + 0.0283915176987648, + 0.021084753796458244, + 0.02193325199186802, + 0.0070678857155144215, + 0.0082778949290514, + 0.00479386979714036, + 0.00666325306519866, + 0.007953056134283543, + 0.0060730730183422565, + 0.005379887297749519, + 0.005651443265378475, + 0.0077511402778327465, + 0.0064430576749145985, + 0.023974671959877014, + 0.00667233020067215, + 0.008048324845731258, + 0.024007346481084824, + 0.004362424369901419, + 0.0049937814474105835, + 0.009324214421212673, + 0.008966655470430851, + 0.007097260095179081, + 0.008015611208975315, + 0.006217907648533583, + 0.008303381502628326, + 0.006621827837079763, + 0.010305246338248253, + 0.004675509873777628, + 0.006113458890467882, + 0.0033031697385013103, + 0.0033337417989969254, + 0.0070995911955833435, + 0.01574186235666275, + 0.003125541377812624, + 0.012557794339954853, + 0.011485898867249489, + 0.0074580879881978035, + 0.006058444734662771, + 0.006398300174623728, + 0.008045069873332977, + 0.03865257278084755, + 0.0053589968010783195, + 0.03239250183105469, + 0.00401969812810421, + 0.0037706077564507723, + 0.0023651495575904846, + 0.005616143345832825, + 0.0031274931970983744, + 0.0038517038337886333, + 0.006126122549176216, + 0.003996421117335558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 8 + ], + "selected_reward": -0.03371596708893776 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Tapanuli orangutan mother with a tiny infant clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark graphic novel style with thick, bold black ink and dramatic crosshatching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mother's fur depicted as cinnamon-colored with richly textured lines and immense volume?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mother's hand gently offering a broad green leaf to the infant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framing a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036227568052709103 + ], + "opposite_logprob_tokens": [ + -5.628622531890869 + ], + "logprob_reward": -0.0036227568052709103, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036227568052709103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05347706004977226, + 0.042396608740091324, + 0.4715660810470581, + 0.13300982117652893, + 0.02572515606880188, + 0.04051128402352333, + 0.024950014427304268, + 0.07549140602350235, + 0.02893027663230896, + 0.10394225269556046 + ], + "credit_map": [ + 0.00580807588994503, + 0.009576425887644291, + 0.006789747159928083, + 0.00726426811888814, + 0.006818647962063551, + 0.0050971293821930885, + 0.005297174211591482, + 0.006825586315244436, + 0.004669880028814077, + 0.004808331839740276, + 0.010916403494775295, + 0.002953999675810337, + 0.004135728348046541, + 0.0030962422024458647, + 0.0033565987832844257, + 0.00845942273736, + 0.03064335510134697, + 0.012006843462586403, + 0.18519246578216553, + 0.028481226414442062, + 0.023791642859578133, + 0.13527518510818481, + 0.011639432981610298, + 0.044535938650369644, + 0.012367249466478825, + 0.018349869176745415, + 0.01104720588773489, + 0.016962936148047447, + 0.014730705879628658, + 0.01996258832514286, + 0.02335691824555397, + 0.01623234897851944, + 0.0021613985300064087, + 0.0030443272553384304, + 0.0021077326964586973, + 0.0027827387675642967, + 0.003827168606221676, + 0.002520417794585228, + 0.005328904837369919, + 0.00395246734842658, + 0.0029499828815460205, + 0.006288433447480202, + 0.006185880862176418, + 0.00449241092428565, + 0.005347513593733311, + 0.006982374005019665, + 0.0039004504214972258, + 0.0043642399832606316, + 0.0025892863050103188, + 0.0038929691072553396, + 0.002722136676311493, + 0.002644937951117754, + 0.003683453891426325, + 0.0025764519814401865, + 0.0030483261216431856, + 0.0037924516946077347, + 0.007019042037427425, + 0.005795078352093697, + 0.005572232883423567, + 0.0037377583794295788, + 0.005657666828483343, + 0.028017202392220497, + 0.003976146690547466, + 0.015716280788183212, + 0.0020348539110273123, + 0.0029106028378009796, + 0.0016335025429725647, + 0.0026247561909258366, + 0.0024298420175909996, + 0.007473175413906574, + 0.0023352312855422497, + 0.007488312665373087, + 0.011654098518192768, + 0.010083266533911228, + 0.009205607697367668, + 0.018598828464746475, + 0.008693059906363487, + 0.010427476838231087, + 0.020434992387890816, + 0.014844926074147224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 9 + ], + "selected_reward": -0.0036227568052709103 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Tapanuli orangutan mother with a tiny infant clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark graphic novel style with thick, bold black ink and dramatic crosshatching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mother's fur depicted as cinnamon-colored with richly textured lines and immense volume?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mother's hand gently offering a broad green leaf to the infant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framing a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703125178813934 + ], + "opposite_logprob_tokens": [ + -2.1270313262939453 + ], + "logprob_reward": -0.12703125178813934, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12703125178813934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15455500781536102, + 0.12095191329717636, + 0.170283704996109, + 0.04455474764108658, + 0.04430869221687317, + 0.07500487565994263, + 0.021382346749305725, + 0.07124993950128555, + 0.0567755289375782, + 0.24093320965766907 + ], + "credit_map": [ + 0.012748989276587963, + 0.031614750623703, + 0.01487373374402523, + 0.015156540088355541, + 0.028493991121649742, + 0.011446374468505383, + 0.022417636588215828, + 0.017802996560931206, + 0.009450101293623447, + 0.024899248033761978, + 0.016102876514196396, + 0.014906511642038822, + 0.019181830808520317, + 0.010763011872768402, + 0.01161851268261671, + 0.014029812067747116, + 0.010527435690164566, + 0.007951788604259491, + 0.06581547111272812, + 0.011230765841901302, + 0.010338525287806988, + 0.03846703842282295, + 0.009822458028793335, + 0.016130218282341957, + 0.0035498656798154116, + 0.007728653959929943, + 0.004309733863919973, + 0.003974585793912411, + 0.006361259147524834, + 0.009339271113276482, + 0.005046059377491474, + 0.004245321732014418, + 0.004143554251641035, + 0.007658854592591524, + 0.005942681338638067, + 0.005160379223525524, + 0.004649638198316097, + 0.008402882143855095, + 0.004039868246763945, + 0.004310833290219307, + 0.01033986546099186, + 0.00764594366773963, + 0.017989935353398323, + 0.005616086069494486, + 0.006800516042858362, + 0.01458678487688303, + 0.005931898485869169, + 0.006093847565352917, + 0.003043537959456444, + 0.0031705645378679037, + 0.002560697728767991, + 0.0023661539889872074, + 0.0028095997404307127, + 0.003213027026504278, + 0.0024579507298767567, + 0.0017608139896765351, + 0.0054947854951024055, + 0.006614937447011471, + 0.0048677208833396435, + 0.003909020684659481, + 0.006930347066372633, + 0.022323423996567726, + 0.004026766866445541, + 0.01708293706178665, + 0.004428215324878693, + 0.0049864971078932285, + 0.0030320168007165194, + 0.005125115625560284, + 0.004653085954487324, + 0.01303684338927269, + 0.006038023624569178, + 0.015475730411708355, + 0.026668980717658997, + 0.02242480404675007, + 0.01915760152041912, + 0.046770427376031876, + 0.01935059018433094, + 0.021101288497447968, + 0.05128539726138115, + 0.0341741219162941 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.12703125178813934 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Tapanuli orangutan mother with a tiny infant clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark graphic novel style with thick, bold black ink and dramatic crosshatching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mother's fur depicted as cinnamon-colored with richly textured lines and immense volume?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mother's hand gently offering a broad green leaf to the infant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framing a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020574131980538368 + ], + "opposite_logprob_tokens": [ + -3.8955740928649902 + ], + "logprob_reward": -0.020574131980538368, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020574131980538368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12890572845935822, + 0.06128226965665817, + 0.10996830463409424, + 0.07123659551143646, + 0.05302121490240097, + 0.21029990911483765, + 0.05989672988653183, + 0.09063968062400818, + 0.0802375078201294, + 0.13451208174228668 + ], + "credit_map": [ + 0.018999116495251656, + 0.017611335963010788, + 0.012913789600133896, + 0.013143081218004227, + 0.022309569641947746, + 0.010474304668605328, + 0.017946859821677208, + 0.015507671050727367, + 0.005489903036504984, + 0.00936609972268343, + 0.013431997038424015, + 0.0059174285270273685, + 0.006738165393471718, + 0.004781445953994989, + 0.006387128960341215, + 0.009170101955533028, + 0.016435477882623672, + 0.02031336911022663, + 0.012376326136291027, + 0.010946177877485752, + 0.014571611769497395, + 0.0106753408908844, + 0.009602430276572704, + 0.015047571621835232, + 0.007324774283915758, + 0.01688302680850029, + 0.007049965672194958, + 0.007793126627802849, + 0.009491230361163616, + 0.008044527843594551, + 0.007198451552540064, + 0.007451492827385664, + 0.008494574576616287, + 0.008420786820352077, + 0.005573058966547251, + 0.005477124359458685, + 0.00690337223932147, + 0.005528035573661327, + 0.0064376844093203545, + 0.006186577957123518, + 0.012404397130012512, + 0.02095843479037285, + 0.06449718773365021, + 0.013428070582449436, + 0.01915435679256916, + 0.055076733231544495, + 0.009218844585120678, + 0.015561878681182861, + 0.005703080911189318, + 0.008972495794296265, + 0.008608166128396988, + 0.005936361849308014, + 0.0075985644944012165, + 0.011577987112104893, + 0.006256552878767252, + 0.005243524443358183, + 0.007790125906467438, + 0.013263801112771034, + 0.017790628597140312, + 0.009676423855125904, + 0.011966945603489876, + 0.012297331355512142, + 0.007165580987930298, + 0.010688843205571175, + 0.005461768247187138, + 0.008451517671346664, + 0.0042524379678070545, + 0.008119300939142704, + 0.007733655162155628, + 0.012703550048172474, + 0.00954272598028183, + 0.02397255226969719, + 0.01644892431795597, + 0.013246678747236729, + 0.01199951209127903, + 0.022809594869613647, + 0.010857945308089256, + 0.015330074355006218, + 0.02672279067337513, + 0.017096564173698425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.020574131980538368 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium close-up captures a poignant scene, rendered in a stark graphic novel style, with thick, bold black ink and dramatic crosshatching shaping every form. The central focus is a Tapanuli orangutan mother, her form both monumental and gentle. Her long, shaggy cinnamon-colored fur is depicted not as limp hair but as a tapestry of richly textured lines, deeply crosshatched and imbued with immense volume. Her intelligent, dark eyes, set on a calm, weathered face, convey a serene, watchful expression. Her tiny infant clings to her chest, its tiny body a compact mass of light-colored fur, its hands and feet disappearing into the mother's thick coat. In a moment of profound tenderness, the mother's large, dark, leathery hand, marked with intricate ink lines, gently and deliberately offers a broad green leaf to the infant. The lighting is high-contrast and directional, casting deep, angular shadows that carve their figures from the minimalist background, creating a powerful, motionless frame that emphasizes the quiet intimacy of the moment and the silent, loving exchange between mother and child.", + "seed": 23560, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_7/index_9039_seed_23560_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05668752761557698, + "vlm_logprob": -0.05668752761557698, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.05668752761557698, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Tapanuli orangutan mother with a tiny infant clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark graphic novel style with thick, bold black ink and dramatic crosshatching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mother's fur depicted as cinnamon-colored with richly textured lines and immense volume?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mother's hand gently offering a broad green leaf to the infant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framing a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.036520037800073624 + ], + "opposite_logprob_tokens": [ + -3.411520004272461 + ], + "logprob_reward": -0.036520037800073624, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.036520037800073624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13000617921352386, + 0.07209530472755432, + 0.03179975971579552, + 0.08497141301631927, + 0.1978975534439087, + 0.03452908247709274, + 0.03701767325401306, + 0.26241546869277954, + 0.07619672268629074, + 0.07307083159685135 + ], + "credit_map": [ + 0.02398349717259407, + 0.014906027354300022, + 0.013270488940179348, + 0.010411153547465801, + 0.008263573981821537, + 0.006998336408287287, + 0.036825601011514664, + 0.015347491949796677, + 0.009069929830729961, + 0.012553004547953606, + 0.008309838362038136, + 0.008785193786025047, + 0.007328384090214968, + 0.004861988592892885, + 0.011750398203730583, + 0.009436561726033688, + 0.005529043264687061, + 0.005291022825986147, + 0.003998795989900827, + 0.0032169176265597343, + 0.0025707141030579805, + 0.002331634284928441, + 0.0050012110732495785, + 0.003860423108562827, + 0.0091279037296772, + 0.015636920928955078, + 0.008243727497756481, + 0.007797784637659788, + 0.009983338415622711, + 0.009120792150497437, + 0.012305821292102337, + 0.012755126692354679, + 0.03014223463833332, + 0.02634027972817421, + 0.016959114000201225, + 0.028467969968914986, + 0.017872445285320282, + 0.019201038405299187, + 0.04077758267521858, + 0.0181368850171566, + 0.004424676299095154, + 0.00544316740706563, + 0.003519682912155986, + 0.005109844729304314, + 0.003719298169016838, + 0.0032987745944410563, + 0.00577171053737402, + 0.003241929691284895, + 0.005050330422818661, + 0.004547944758087397, + 0.004409813322126865, + 0.007559491321444511, + 0.0033330523874610662, + 0.003913374152034521, + 0.004390819929540157, + 0.0038128450978547335, + 0.05361676961183548, + 0.015340090729296207, + 0.01831919699907303, + 0.09109033644199371, + 0.028624562546610832, + 0.010161749087274075, + 0.026810452342033386, + 0.01845230534672737, + 0.0218980573117733, + 0.006259600631892681, + 0.014220556244254112, + 0.009210800752043724, + 0.005729228723794222, + 0.003864775411784649, + 0.008149242028594017, + 0.006864463910460472, + 0.014122608117759228, + 0.00555307837203145, + 0.014801074750721455, + 0.008429007604718208, + 0.006695612333714962, + 0.005484362598508596, + 0.009353176690638065, + 0.00863190833479166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.036520037800073624 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Tapanuli orangutan mother with a tiny infant clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark graphic novel style with thick, bold black ink and dramatic crosshatching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mother's fur depicted as cinnamon-colored with richly textured lines and immense volume?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mother's hand gently offering a broad green leaf to the infant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framing a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20155075192451477 + ], + "opposite_logprob_tokens": [ + -1.7015507221221924 + ], + "logprob_reward": -0.20155075192451477, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20155075192451477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06740628927946091, + 0.04197889193892479, + 0.04865412786602974, + 0.04550475999712944, + 0.06006995588541031, + 0.12495068460702896, + 0.0409541092813015, + 0.185066357254982, + 0.2283182144165039, + 0.15709665417671204 + ], + "credit_map": [ + 0.009293155744671822, + 0.011256101541221142, + 0.006847864482551813, + 0.005409850738942623, + 0.005791943520307541, + 0.005980231333523989, + 0.012893939390778542, + 0.009933200664818287, + 0.004295456223189831, + 0.007245355751365423, + 0.004198248032480478, + 0.004547270014882088, + 0.004615258891135454, + 0.0034213142935186625, + 0.009535465389490128, + 0.004120524041354656, + 0.007881023921072483, + 0.007494120858609676, + 0.0042136418633162975, + 0.005717468447983265, + 0.004218508023768663, + 0.0043181851506233215, + 0.006072539836168289, + 0.00873863697052002, + 0.004813649691641331, + 0.009370073676109314, + 0.005041127558797598, + 0.0038683577440679073, + 0.004086931701749563, + 0.004513064865022898, + 0.008082203567028046, + 0.005729349330067635, + 0.007761394139379263, + 0.010875537991523743, + 0.006424523890018463, + 0.006623866967856884, + 0.007178231608122587, + 0.005047635640949011, + 0.010009353049099445, + 0.00614940794184804, + 0.015629425644874573, + 0.02087372913956642, + 0.013896566815674305, + 0.020409859716892242, + 0.011816688813269138, + 0.010669001378118992, + 0.01900092139840126, + 0.012654485180974007, + 0.007962760515511036, + 0.0051883067935705185, + 0.00442899577319622, + 0.005788828246295452, + 0.005014841910451651, + 0.0030146492645144463, + 0.00427576620131731, + 0.00527995964512229, + 0.03884338214993477, + 0.01034703478217125, + 0.011986392550170422, + 0.06418439000844955, + 0.018650105223059654, + 0.007985423319041729, + 0.01931336149573326, + 0.013756238855421543, + 0.057347752153873444, + 0.010864968411624432, + 0.08215377479791641, + 0.016892965883016586, + 0.015303948894143105, + 0.005866602063179016, + 0.02850688062608242, + 0.011381298303604126, + 0.024487463757395744, + 0.013765119016170502, + 0.04484661668539047, + 0.017203185707330704, + 0.011433304287493229, + 0.010566741228103638, + 0.02149120531976223, + 0.013303007930517197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.20155075192451477 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Tapanuli orangutan mother with a tiny infant clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark graphic novel style with thick, bold black ink and dramatic crosshatching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mother's fur depicted as cinnamon-colored with richly textured lines and immense volume?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mother's hand gently offering a broad green leaf to the infant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framing a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001517812255769968 + ], + "opposite_logprob_tokens": [ + -6.5015177726745605 + ], + "logprob_reward": -0.001517812255769968, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001517812255769968, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0856265053153038, + 0.07606597989797592, + 0.05922669544816017, + 0.09069624543190002, + 0.08064207434654236, + 0.060609713196754456, + 0.05289578437805176, + 0.18632055819034576, + 0.15649762749671936, + 0.1514187902212143 + ], + "credit_map": [ + 0.013511620461940765, + 0.012142335064709187, + 0.009894762188196182, + 0.011082648299634457, + 0.007969164289534092, + 0.006978418678045273, + 0.012151623144745827, + 0.01189594715833664, + 0.007205937057733536, + 0.011284149251878262, + 0.010188776068389416, + 0.010491087101399899, + 0.007254460360854864, + 0.007343731354922056, + 0.010804559104144573, + 0.011493288911879063, + 0.007714554201811552, + 0.010049697011709213, + 0.006851268466562033, + 0.006975266151130199, + 0.004679387900978327, + 0.006501384079456329, + 0.007435904815793037, + 0.009019238874316216, + 0.011297172866761684, + 0.01595870591700077, + 0.007847286760807037, + 0.00822349451482296, + 0.01278330571949482, + 0.009671087376773357, + 0.015017110854387283, + 0.009898084215819836, + 0.009418468922376633, + 0.01674732007086277, + 0.007123011164367199, + 0.010987093672156334, + 0.010625822469592094, + 0.0072624292224645615, + 0.010029856115579605, + 0.008448082953691483, + 0.009383361786603928, + 0.006987445987761021, + 0.005901391617953777, + 0.007763152476400137, + 0.005119150038808584, + 0.005217603873461485, + 0.013664081692695618, + 0.006573531311005354, + 0.009696238674223423, + 0.004967279266566038, + 0.005245017819106579, + 0.011053601279854774, + 0.006225242279469967, + 0.004899399820715189, + 0.006023977883160114, + 0.004785032477229834, + 0.03455725312232971, + 0.017074240371584892, + 0.01579425297677517, + 0.045914795249700546, + 0.0293749887496233, + 0.011295284144580364, + 0.01783212460577488, + 0.014477629214525223, + 0.05203954130411148, + 0.00979557354003191, + 0.024075640365481377, + 0.02821120247244835, + 0.01139924954622984, + 0.007089547347277403, + 0.01410355232656002, + 0.009783328510820866, + 0.02746499702334404, + 0.01097107119858265, + 0.05148885026574135, + 0.01140347309410572, + 0.010056113824248314, + 0.0067503685131669044, + 0.022562125697731972, + 0.010721812956035137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.001517812255769968 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Tapanuli orangutan mother with a tiny infant clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark graphic novel style with thick, bold black ink and dramatic crosshatching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mother's fur depicted as cinnamon-colored with richly textured lines and immense volume?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mother's hand gently offering a broad green leaf to the infant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framing a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020570512861013412 + ], + "opposite_logprob_tokens": [ + -3.8955705165863037 + ], + "logprob_reward": -0.020570512861013412, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020570512861013412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04483714699745178, + 0.08256246149539948, + 0.033014364540576935, + 0.03857039287686348, + 0.08617303520441055, + 0.027464265003800392, + 0.03295290470123291, + 0.42427384853363037, + 0.0800953134894371, + 0.15005621314048767 + ], + "credit_map": [ + 0.005996769294142723, + 0.008309971541166306, + 0.0041287606582045555, + 0.00524377916008234, + 0.004472953267395496, + 0.004009065218269825, + 0.007522967644035816, + 0.005152876954525709, + 0.008669079281389713, + 0.01581892929971218, + 0.009296699427068233, + 0.012387353926897049, + 0.007029013242572546, + 0.007643559947609901, + 0.01186450757086277, + 0.009853317402303219, + 0.004191217478364706, + 0.004230307415127754, + 0.0026971001643687487, + 0.004095605108886957, + 0.005049549974501133, + 0.002919710474088788, + 0.004296392202377319, + 0.005534484051167965, + 0.00461561419069767, + 0.009277251549065113, + 0.0041245585307478905, + 0.004442926496267319, + 0.005197909194976091, + 0.0026091805193573236, + 0.004697370808571577, + 0.0036055820528417826, + 0.011980815790593624, + 0.011106732301414013, + 0.00906949769705534, + 0.0150226429104805, + 0.008954853750765324, + 0.009064665995538235, + 0.010673191398382187, + 0.01030062884092331, + 0.003840507473796606, + 0.0033457365352660418, + 0.0031161063816398382, + 0.004653576295822859, + 0.0030480017885565758, + 0.0026806851383298635, + 0.004070568364113569, + 0.002709081396460533, + 0.004662781022489071, + 0.003759979037567973, + 0.00414703506976366, + 0.004274129867553711, + 0.004298726562410593, + 0.0037752054631710052, + 0.004250067286193371, + 0.0037849803920835257, + 0.08301518112421036, + 0.02766263298690319, + 0.029418375343084335, + 0.1394101083278656, + 0.053010810166597366, + 0.01880171336233616, + 0.04262794554233551, + 0.03032711148262024, + 0.012258732691407204, + 0.004017174243927002, + 0.030841046944260597, + 0.0055785635486245155, + 0.005662547890096903, + 0.002168523147702217, + 0.011678608134388924, + 0.007890121079981327, + 0.022222667932510376, + 0.010781968012452126, + 0.054041605442762375, + 0.01216644886881113, + 0.013100484386086464, + 0.007911247201263905, + 0.019716860726475716, + 0.010114945471286774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 4 + ], + "selected_reward": -0.020570512861013412 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Tapanuli orangutan mother with a tiny infant clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark graphic novel style with thick, bold black ink and dramatic crosshatching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mother's fur depicted as cinnamon-colored with richly textured lines and immense volume?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mother's hand gently offering a broad green leaf to the infant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framing a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023278523236513138 + ], + "opposite_logprob_tokens": [ + -3.7732784748077393 + ], + "logprob_reward": -0.023278523236513138, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023278523236513138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04325736314058304, + 0.018307283520698547, + 0.014854044653475285, + 0.030695701017975807, + 0.039747659116983414, + 0.03617981821298599, + 0.10028156638145447, + 0.03711410239338875, + 0.633289635181427, + 0.04627281799912453 + ], + "credit_map": [ + 0.004416363779455423, + 0.008104663342237473, + 0.005193796474486589, + 0.005478023551404476, + 0.004378551617264748, + 0.0030387400183826685, + 0.008286968804895878, + 0.00436025345697999, + 0.0024457285180687904, + 0.002737766830250621, + 0.0020504454150795937, + 0.0025053962599486113, + 0.0021123290061950684, + 0.001107779098674655, + 0.0036008779425174, + 0.0017469588201493025, + 0.0019936764147132635, + 0.0024431543424725533, + 0.0016609965823590755, + 0.0020247928332537413, + 0.001697859144769609, + 0.001527339336462319, + 0.001999728614464402, + 0.001506497384980321, + 0.003290426218882203, + 0.005201470106840134, + 0.0033227172680199146, + 0.003902958007529378, + 0.0044777411967515945, + 0.002437751041725278, + 0.00442271726205945, + 0.003639919450506568, + 0.00584285007789731, + 0.008039387874305248, + 0.003647188190370798, + 0.0037685928400605917, + 0.0053798360750079155, + 0.003887356258928776, + 0.005344761535525322, + 0.00383768486790359, + 0.004846783820539713, + 0.006119545549154282, + 0.004420291632413864, + 0.005844504572451115, + 0.0037642773240804672, + 0.0027085416950285435, + 0.00497658783569932, + 0.0034992857836186886, + 0.013993711210787296, + 0.00687085697427392, + 0.01011831033974886, + 0.035122133791446686, + 0.00697791762650013, + 0.008260829374194145, + 0.01121491752564907, + 0.007722897455096245, + 0.0044015576131641865, + 0.004262417554855347, + 0.005224066320806742, + 0.005690095014870167, + 0.006178184412419796, + 0.003251222660765052, + 0.005046099424362183, + 0.0030604589264839888, + 0.12569081783294678, + 0.029313987120985985, + 0.25091996788978577, + 0.04195650666952133, + 0.04075862467288971, + 0.01197105273604393, + 0.08581351488828659, + 0.046865157783031464, + 0.009062016382813454, + 0.0037208269350230694, + 0.010378570295870304, + 0.004030156414955854, + 0.005427460186183453, + 0.002668959554284811, + 0.005688016302883625, + 0.005296813324093819 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.023278523236513138 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_8/index_tensor([9116, 9116])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_8/index_tensor([9116, 9116])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3d2d9088bd83f092db51ef05b72834e1e56631f0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_8/index_tensor([9116, 9116])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, high-angle shot allows the camera to track steadily to the right, following a rectangular bar of milky white soap sliding across a damp, dark slate-gray tile surface. The soap bar, worn smooth from use, its surface glistening with water droplets and a thin layer of translucent foam, appears soft and almost plastic. Propelled by an unseen force, it moves steadily and deliberately, leaving behind a thick, milky white trail topped with dense, foamy lather that contrasts sharply against the dark, reflective tile. A single, soft, overhead light casts bright, mirror-like highlights on the water, the glossy soap, and the wet tile, while creating deep shadows in the cement line between them, creating a minimalist, clean, and tranquil atmosphere focused entirely on the satisfying interplay of tactile texture and movement.", + "seed": 31513, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_8/index_9116_seed_31513_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06035535540431738, + "vlm_logprob": -0.06035535540431738, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.06035535540431738, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a rectangular bar of milky white soap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soap bar sliding across a damp, dark slate-gray tile surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soap surface appear glistening with water droplets and a thin layer of translucent foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the soap leave behind a thick, milky white trail topped with dense, foamy lather as it moves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily to the right following the soap?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01198036503046751 + ], + "opposite_logprob_tokens": [ + -6.1369805335998535 + ], + "logprob_reward": -0.01198036503046751, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.01198036503046751, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14764955639839172, + 0.05622969940304756, + 0.02376095950603485, + 0.024708159267902374, + 0.10037542134523392, + 0.05056224763393402, + 0.2863823175430298, + 0.17672199010849, + 0.08505797386169434, + 0.048551734536886215 + ], + "credit_map": [ + 0.010375109501183033, + 0.03227440267801285, + 0.010324221104383469, + 0.018115108832716942, + 0.019770139828324318, + 0.011307135224342346, + 0.034008901566267014, + 0.011474528349936008, + 0.004572185687720776, + 0.0072866519913077354, + 0.006320554297417402, + 0.008600176312029362, + 0.0067789312452077866, + 0.006460738368332386, + 0.007334349676966667, + 0.008876105770468712, + 0.0021300888620316982, + 0.0023355945013463497, + 0.002385544590651989, + 0.0036748407874256372, + 0.0023710832465440035, + 0.003745530266314745, + 0.002254120772704482, + 0.0048641543835401535, + 0.0029188927728682756, + 0.002406410174444318, + 0.0030395707581192255, + 0.003590075299143791, + 0.0023951951880007982, + 0.0024496465921401978, + 0.0029804136138409376, + 0.004927953705191612, + 0.017659248784184456, + 0.011861477978527546, + 0.011528215371072292, + 0.01691613160073757, + 0.00844447873532772, + 0.012946495786309242, + 0.009168172255158424, + 0.011851181276142597, + 0.009901807643473148, + 0.0066743940114974976, + 0.006441274657845497, + 0.007327834609895945, + 0.00454133003950119, + 0.004374466370791197, + 0.0048551252111792564, + 0.00644600810483098, + 0.05158407986164093, + 0.08765014261007309, + 0.013379257172346115, + 0.018082648515701294, + 0.02309504710137844, + 0.010848268866539001, + 0.03227316960692406, + 0.04946966469287872, + 0.035658374428749084, + 0.016142835840582848, + 0.009721217676997185, + 0.00973211508244276, + 0.007424998562783003, + 0.010500670410692692, + 0.036265287548303604, + 0.05127647519111633, + 0.01800437830388546, + 0.023858068510890007, + 0.007371498737484217, + 0.0060386802069842815, + 0.007723354734480381, + 0.005201491992920637, + 0.008616323582828045, + 0.008244171738624573, + 0.005259666126221418, + 0.005035076756030321, + 0.011046218685805798, + 0.005047469399869442, + 0.004471906460821629, + 0.006126930471509695, + 0.004587971139699221, + 0.006976489443331957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.01198036503046751 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a rectangular bar of milky white soap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soap bar sliding across a damp, dark slate-gray tile surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soap surface appear glistening with water droplets and a thin layer of translucent foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the soap leave behind a thick, milky white trail topped with dense, foamy lather as it moves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily to the right following the soap?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009795927442610264 + ], + "opposite_logprob_tokens": [ + -4.634796142578125 + ], + "logprob_reward": -0.009795927442610264, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009795927442610264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18956546485424042, + 0.03553575649857521, + 0.04477928951382637, + 0.028271842747926712, + 0.12782637774944305, + 0.08877858519554138, + 0.05283467844128609, + 0.1077677458524704, + 0.27439987659454346, + 0.050240423530340195 + ], + "credit_map": [ + 0.01182575523853302, + 0.05928073823451996, + 0.01050626952201128, + 0.011813115328550339, + 0.043265532702207565, + 0.011529694311320782, + 0.02797449566423893, + 0.01336986105889082, + 0.004054802469909191, + 0.0041170865297317505, + 0.005788088776171207, + 0.004821191541850567, + 0.0038489880971610546, + 0.004598156549036503, + 0.003607396502047777, + 0.004700046498328447, + 0.006102840881794691, + 0.0052339001558721066, + 0.004216824658215046, + 0.0060393488965928555, + 0.004766142927110195, + 0.009198114275932312, + 0.003774553071707487, + 0.005447564646601677, + 0.0031782686710357666, + 0.002565690316259861, + 0.0028215572237968445, + 0.0035395363811403513, + 0.0022441050969064236, + 0.002435495378449559, + 0.003447561990469694, + 0.008039628155529499, + 0.019799286499619484, + 0.014815703965723515, + 0.01759602315723896, + 0.02587827853858471, + 0.010931123048067093, + 0.01508523989468813, + 0.010664239525794983, + 0.013056487776339054, + 0.017844729125499725, + 0.0124198654666543, + 0.010088232345879078, + 0.013074650429189205, + 0.010416419245302677, + 0.009137227199971676, + 0.00631926953792572, + 0.009478195570409298, + 0.006909425836056471, + 0.007991642691195011, + 0.006231279112398624, + 0.006919038947671652, + 0.005365846212953329, + 0.0054217856377363205, + 0.006664968561381102, + 0.007330691441893578, + 0.017448272556066513, + 0.01352798193693161, + 0.010657182894647121, + 0.010681994259357452, + 0.006161868572235107, + 0.010280699469149113, + 0.01640329696238041, + 0.02260645292699337, + 0.07035139203071594, + 0.08131327480077744, + 0.01956784725189209, + 0.01332815084606409, + 0.016870534047484398, + 0.013440994545817375, + 0.031626902520656586, + 0.027900777757167816, + 0.004432431887835264, + 0.004482720047235489, + 0.008465152233839035, + 0.0055793956853449345, + 0.004247799050062895, + 0.004547028802335262, + 0.007701943162828684, + 0.010783950798213482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.009795927442610264 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a rectangular bar of milky white soap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soap bar sliding across a damp, dark slate-gray tile surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soap surface appear glistening with water droplets and a thin layer of translucent foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the soap leave behind a thick, milky white trail topped with dense, foamy lather as it moves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily to the right following the soap?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006746253930032253 + ], + "opposite_logprob_tokens": [ + -5.006746292114258 + ], + "logprob_reward": -0.006746253930032253, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006746253930032253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39331740140914917, + 0.06100122630596161, + 0.027702752500772476, + 0.016421586275100708, + 0.04644268751144409, + 0.02096150629222393, + 0.07297034561634064, + 0.26893627643585205, + 0.06607560813426971, + 0.026170626282691956 + ], + "credit_map": [ + 0.02209477126598358, + 0.1252286434173584, + 0.02050468884408474, + 0.026961032301187515, + 0.08336645364761353, + 0.019264303147792816, + 0.07237686961889267, + 0.02352064475417137, + 0.006864740047603846, + 0.008423994295299053, + 0.012036638334393501, + 0.009098134934902191, + 0.007464277558028698, + 0.005557808559387922, + 0.0062199183739721775, + 0.005335708614438772, + 0.002906552981585264, + 0.0031470530666410923, + 0.0026440881192684174, + 0.005075098481029272, + 0.004770798142999411, + 0.002105716383084655, + 0.0027016850654035807, + 0.004351761192083359, + 0.0018962047761306167, + 0.0017765838420018554, + 0.0016330841463059187, + 0.0027428800240159035, + 0.002163119148463011, + 0.0015681814402341843, + 0.002059241058304906, + 0.002582291839644313, + 0.007295658811926842, + 0.005438289605081081, + 0.006075515877455473, + 0.008572439663112164, + 0.0038869339041411877, + 0.005415196530520916, + 0.004266068339347839, + 0.005492585711181164, + 0.00280260038562119, + 0.002186480211094022, + 0.0025604001712054014, + 0.005092372186481953, + 0.0020682127214968204, + 0.0014358788030222058, + 0.0026197361294180155, + 0.00219582486897707, + 0.011128083802759647, + 0.018726641312241554, + 0.006413714494556189, + 0.0064103309996426105, + 0.005943714641034603, + 0.0060009704902768135, + 0.00776906730607152, + 0.01057782955467701, + 0.0559033527970314, + 0.028849834576249123, + 0.01262905914336443, + 0.009589538909494877, + 0.012024742551147938, + 0.01198627706617117, + 0.056827135384082794, + 0.08112633973360062, + 0.015787342563271523, + 0.025694437325000763, + 0.004419778939336538, + 0.0039983829483389854, + 0.005924062803387642, + 0.0024787753354758024, + 0.0026236504781991243, + 0.005149178206920624, + 0.0037693921476602554, + 0.002709140768274665, + 0.004458094481378794, + 0.003995400387793779, + 0.0023183536250144243, + 0.0027376243378967047, + 0.0030364906415343285, + 0.003146131755784154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.006746253930032253 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a rectangular bar of milky white soap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soap bar sliding across a damp, dark slate-gray tile surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soap surface appear glistening with water droplets and a thin layer of translucent foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the soap leave behind a thick, milky white trail topped with dense, foamy lather as it moves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily to the right following the soap?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293192207813263 + ], + "opposite_logprob_tokens": [ + -2.237931966781616 + ], + "logprob_reward": -0.11293192207813263, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293192207813263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25543636083602905, + 0.07418608665466309, + 0.07300693541765213, + 0.04301660507917404, + 0.07459180802106857, + 0.038824718445539474, + 0.07614307850599289, + 0.07924078404903412, + 0.17021416127681732, + 0.11533946543931961 + ], + "credit_map": [ + 0.01672830805182457, + 0.07569725066423416, + 0.011188247241079807, + 0.019703557714819908, + 0.051983363926410675, + 0.02230527624487877, + 0.03979107365012169, + 0.018039286136627197, + 0.009302339516580105, + 0.008782222867012024, + 0.009264066815376282, + 0.011942198500037193, + 0.007869862951338291, + 0.006919454783201218, + 0.011409923434257507, + 0.008696022443473339, + 0.007641152944415808, + 0.009852863848209381, + 0.006903715897351503, + 0.00947822816669941, + 0.0063762846402823925, + 0.018644526600837708, + 0.005346376448869705, + 0.008763788267970085, + 0.0038796672597527504, + 0.004597477614879608, + 0.005364701617509127, + 0.00738495634868741, + 0.005647195037454367, + 0.003451254917308688, + 0.005514705553650856, + 0.007176645565778017, + 0.0068253763020038605, + 0.007767798379063606, + 0.010042701847851276, + 0.014047268778085709, + 0.007705294992774725, + 0.009649607352912426, + 0.008226647973060608, + 0.010327118448913097, + 0.00441146083176136, + 0.003980803303420544, + 0.005367442965507507, + 0.008224566467106342, + 0.004735113121569157, + 0.003618387971073389, + 0.0044153365306556225, + 0.004071605391800404, + 0.011297176592051983, + 0.0180986188352108, + 0.010682961903512478, + 0.009778731502592564, + 0.005560815334320068, + 0.0038205033633857965, + 0.0069526140578091145, + 0.009951653890311718, + 0.010136932134628296, + 0.010078792460262775, + 0.011851307936012745, + 0.009969537146389484, + 0.0055527337826788425, + 0.007196086458861828, + 0.010942167602479458, + 0.013513223268091679, + 0.04876333102583885, + 0.053520914167165756, + 0.011901233345270157, + 0.009235715493559837, + 0.008590913377702236, + 0.008032368496060371, + 0.015481853857636452, + 0.014687835238873959, + 0.010733228176832199, + 0.011200441978871822, + 0.015340945683419704, + 0.01731223613023758, + 0.009974665008485317, + 0.009747236967086792, + 0.01796254888176918, + 0.023068159818649292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.11293192207813263 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a rectangular bar of milky white soap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soap bar sliding across a damp, dark slate-gray tile surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soap surface appear glistening with water droplets and a thin layer of translucent foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the soap leave behind a thick, milky white trail topped with dense, foamy lather as it moves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily to the right following the soap?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16032230854034424 + ], + "opposite_logprob_tokens": [ + -1.9103223085403442 + ], + "logprob_reward": -0.16032230854034424, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16032230854034424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10014423727989197, + 0.026195162907242775, + 0.03486385568976402, + 0.026471950113773346, + 0.04832777380943298, + 0.05405587702989578, + 0.14978113770484924, + 0.28104713559150696, + 0.2261306494474411, + 0.05298220366239548 + ], + "credit_map": [ + 0.005170093383640051, + 0.03144305199384689, + 0.005740344058722258, + 0.008652145974338055, + 0.022020990028977394, + 0.006056496873497963, + 0.015498796477913857, + 0.005562318954616785, + 0.002310601994395256, + 0.00256686401553452, + 0.0029515407513827085, + 0.004580691922456026, + 0.0032491430174559355, + 0.003317402908578515, + 0.0035296145360916853, + 0.0036893021315336227, + 0.003919610753655434, + 0.004684587009251118, + 0.0034324515145272017, + 0.0037793158553540707, + 0.0032052823808044195, + 0.009044340811669827, + 0.0026466897688806057, + 0.00415157712996006, + 0.003352686995640397, + 0.0027200127951800823, + 0.0029273766558617353, + 0.004808933939784765, + 0.0026782515924423933, + 0.0023891832679510117, + 0.0029642218723893166, + 0.004631283227354288, + 0.005112864077091217, + 0.005028708837926388, + 0.005356843117624521, + 0.007720747962594032, + 0.004835901781916618, + 0.006645314395427704, + 0.0060872603207826614, + 0.0075401305221021175, + 0.005048925057053566, + 0.004781663883477449, + 0.0073425862938165665, + 0.008465859107673168, + 0.0055151768028736115, + 0.008844856172800064, + 0.006134513765573502, + 0.007922296412289143, + 0.025234660133719444, + 0.040632497519254684, + 0.010793058201670647, + 0.013811505399644375, + 0.011949548497796059, + 0.007771936245262623, + 0.01592935621738434, + 0.023658597841858864, + 0.060569193214178085, + 0.031173869967460632, + 0.012062225490808487, + 0.009931121952831745, + 0.011930684559047222, + 0.011817926540970802, + 0.05852089822292328, + 0.08504121005535126, + 0.06176389753818512, + 0.06425700336694717, + 0.013812776654958725, + 0.008482147008180618, + 0.01130308024585247, + 0.007619621232151985, + 0.029937364161014557, + 0.02895476669073105, + 0.0067560020834207535, + 0.0025452382396906614, + 0.016193244606256485, + 0.005317989736795425, + 0.004288350697606802, + 0.0021837002132087946, + 0.006373182870447636, + 0.009324492886662483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.16032230854034424 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, high-angle shot allows the camera to track steadily to the right, following a rectangular bar of milky white soap sliding across a damp, dark slate-gray tile surface. The soap bar, worn smooth from use, its surface glistening with water droplets and a thin layer of translucent foam, appears soft and almost plastic. Propelled by an unseen force, it moves steadily and deliberately, leaving behind a thick, milky white trail topped with dense, foamy lather that contrasts sharply against the dark, reflective tile. A single, soft, overhead light casts bright, mirror-like highlights on the water, the glossy soap, and the wet tile, while creating deep shadows in the cement line between them, creating a minimalist, clean, and tranquil atmosphere focused entirely on the satisfying interplay of tactile texture and movement.", + "seed": 31514, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_8/index_9116_seed_31514_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07071552220731973, + "vlm_logprob": -0.07071552220731973, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.07071552220731973, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a rectangular bar of milky white soap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soap bar sliding across a damp, dark slate-gray tile surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soap surface appear glistening with water droplets and a thin layer of translucent foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the soap leave behind a thick, milky white trail topped with dense, foamy lather as it moves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily to the right following the soap?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017829736694693565 + ], + "opposite_logprob_tokens": [ + -4.517829895019531 + ], + "logprob_reward": -0.017829736694693565, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.017829736694693565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0906621515750885, + 0.06161316856741905, + 0.04618481546640396, + 0.08385388553142548, + 0.03003176674246788, + 0.16439209878444672, + 0.2803906500339508, + 0.12288249284029007, + 0.0722130537033081, + 0.04777592420578003 + ], + "credit_map": [ + 0.010208102874457836, + 0.006727826315909624, + 0.020997438579797745, + 0.014468574896454811, + 0.005172085016965866, + 0.012537356466054916, + 0.010354330763220787, + 0.010196441784501076, + 0.00497192470356822, + 0.006649367045611143, + 0.011723705567419529, + 0.008368200622498989, + 0.006916402839124203, + 0.007831476628780365, + 0.008067523129284382, + 0.007084574084728956, + 0.00482951058074832, + 0.004431545734405518, + 0.005561710800975561, + 0.007336825132369995, + 0.0052290721796453, + 0.005514107178896666, + 0.004726629238575697, + 0.00855542067438364, + 0.01706545241177082, + 0.009697966277599335, + 0.007881222292780876, + 0.008267517201602459, + 0.010312891565263271, + 0.01587674766778946, + 0.005014812108129263, + 0.009737282060086727, + 0.0036374391056597233, + 0.003636909183114767, + 0.0019290385534986854, + 0.0028574082534760237, + 0.005255811847746372, + 0.003654655534774065, + 0.0043041822500526905, + 0.004756325390189886, + 0.014721166342496872, + 0.01870211958885193, + 0.0090035330504179, + 0.037969015538692474, + 0.015460404567420483, + 0.021849017590284348, + 0.03365625441074371, + 0.013030610047280788, + 0.023127876222133636, + 0.0258471816778183, + 0.014753923751413822, + 0.05989595875144005, + 0.033306993544101715, + 0.03324740752577782, + 0.05425257608294487, + 0.03595873713493347, + 0.010643449611961842, + 0.012855160981416702, + 0.008911480195820332, + 0.027606287971138954, + 0.015250393189489841, + 0.014551518484950066, + 0.019204074516892433, + 0.013860132545232773, + 0.008200828917324543, + 0.010068290866911411, + 0.004475421272218227, + 0.00936154741793871, + 0.014470112510025501, + 0.007202876731753349, + 0.008950063027441502, + 0.009483919478952885, + 0.005506995599716902, + 0.005741650238633156, + 0.0030175282154232264, + 0.006444994360208511, + 0.008832764811813831, + 0.006272585596889257, + 0.005752282217144966, + 0.006207124330103397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.017829736694693565 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a rectangular bar of milky white soap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soap bar sliding across a damp, dark slate-gray tile surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soap surface appear glistening with water droplets and a thin layer of translucent foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the soap leave behind a thick, milky white trail topped with dense, foamy lather as it moves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily to the right following the soap?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026361288502812386 + ], + "opposite_logprob_tokens": [ + -3.6513612270355225 + ], + "logprob_reward": -0.026361288502812386, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026361288502812386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13149699568748474, + 0.09080930799245834, + 0.11097809672355652, + 0.07423597574234009, + 0.04058028385043144, + 0.1248699277639389, + 0.11094515025615692, + 0.061354685574769974, + 0.0700630247592926, + 0.18466652929782867 + ], + "credit_map": [ + 0.015893157571554184, + 0.011458334513008595, + 0.02732664905488491, + 0.01565191149711609, + 0.009212944656610489, + 0.01724224165081978, + 0.01952344737946987, + 0.015188331715762615, + 0.008374583907425404, + 0.008639276959002018, + 0.016256464645266533, + 0.011521942913532257, + 0.011100463569164276, + 0.011594176292419434, + 0.013264152221381664, + 0.010058249346911907, + 0.010974234901368618, + 0.011230195872485638, + 0.016214169561862946, + 0.01790289580821991, + 0.009819180704653263, + 0.014090522192418575, + 0.011636434122920036, + 0.019110459834337234, + 0.011023461818695068, + 0.00950683280825615, + 0.007289811968803406, + 0.010476235300302505, + 0.01056598499417305, + 0.009693492203950882, + 0.0066309161484241486, + 0.009049246087670326, + 0.005979056004434824, + 0.004640332888811827, + 0.00228674104437232, + 0.003991716541349888, + 0.00517078535631299, + 0.005391205660998821, + 0.005741843953728676, + 0.007378606591373682, + 0.015995915979146957, + 0.015230122022330761, + 0.00836495216935873, + 0.02104358747601509, + 0.014110464602708817, + 0.021651102229952812, + 0.015598713420331478, + 0.01287507638335228, + 0.010597064159810543, + 0.011111748404800892, + 0.00628194073215127, + 0.021382132545113564, + 0.015498959459364414, + 0.011081600561738014, + 0.022462481632828712, + 0.01252923347055912, + 0.005975642241537571, + 0.007487404625862837, + 0.004508194047957659, + 0.009078985080122948, + 0.009074440225958824, + 0.007672050967812538, + 0.007635689340531826, + 0.009922286495566368, + 0.008219538256525993, + 0.010676754638552666, + 0.0036560578737407923, + 0.008796957321465015, + 0.01373742613941431, + 0.008033049292862415, + 0.007573117967694998, + 0.009370128624141216, + 0.025692319497466087, + 0.022300098091363907, + 0.014701719395816326, + 0.01817680336534977, + 0.03196703642606735, + 0.028207646682858467, + 0.018084317445755005, + 0.025536594912409782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.026361288502812386 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a rectangular bar of milky white soap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soap bar sliding across a damp, dark slate-gray tile surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soap surface appear glistening with water droplets and a thin layer of translucent foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the soap leave behind a thick, milky white trail topped with dense, foamy lather as it moves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily to the right following the soap?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00865092221647501 + ], + "opposite_logprob_tokens": [ + -4.758650779724121 + ], + "logprob_reward": -0.00865092221647501, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00865092221647501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1466899961233139, + 0.07971727848052979, + 0.06156102940440178, + 0.09333137422800064, + 0.059829775243997574, + 0.08953004330396652, + 0.12274519354104996, + 0.09201159328222275, + 0.21319414675235748, + 0.04138956964015961 + ], + "credit_map": [ + 0.014103507623076439, + 0.015622483566403389, + 0.02421959675848484, + 0.01991288736462593, + 0.01087714172899723, + 0.029139302670955658, + 0.01633770763874054, + 0.01647736132144928, + 0.006943830754607916, + 0.008678354322910309, + 0.011185094714164734, + 0.008206564001739025, + 0.010230065323412418, + 0.0157721359282732, + 0.010576799511909485, + 0.008124428801238537, + 0.006418421398848295, + 0.007133713457733393, + 0.006332475692033768, + 0.008992449380457401, + 0.008484508842229843, + 0.007117392960935831, + 0.0060397181659936905, + 0.011042345315217972, + 0.01630535162985325, + 0.009201745502650738, + 0.007815000601112843, + 0.012876268476247787, + 0.012042098678648472, + 0.014598674140870571, + 0.007312283385545015, + 0.013179946690797806, + 0.0062825302593410015, + 0.007564150728285313, + 0.003332916647195816, + 0.0064877052791416645, + 0.011099415831267834, + 0.006772145163267851, + 0.007491152733564377, + 0.010799759067595005, + 0.007848515175282955, + 0.009366240352392197, + 0.005460857879370451, + 0.018536154180765152, + 0.012048698030412197, + 0.011299800127744675, + 0.01574747823178768, + 0.009222297929227352, + 0.01095463428646326, + 0.01478786300867796, + 0.008355190977454185, + 0.016425466164946556, + 0.020380187779664993, + 0.014952220022678375, + 0.01724972203373909, + 0.019639896228909492, + 0.013234245590865612, + 0.011719753034412861, + 0.008349830284714699, + 0.010733285918831825, + 0.012297446839511395, + 0.013957949355244637, + 0.009245364926755428, + 0.01247372105717659, + 0.01590464822947979, + 0.03079265169799328, + 0.01818440482020378, + 0.03484443202614784, + 0.031539369374513626, + 0.028493309393525124, + 0.025088980793952942, + 0.0283463504165411, + 0.0033982160966843367, + 0.00455192057415843, + 0.0022862451151013374, + 0.007360993418842554, + 0.007950465194880962, + 0.004950239323079586, + 0.006129077170044184, + 0.004762408323585987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.00865092221647501 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a rectangular bar of milky white soap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soap bar sliding across a damp, dark slate-gray tile surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soap surface appear glistening with water droplets and a thin layer of translucent foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the soap leave behind a thick, milky white trail topped with dense, foamy lather as it moves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily to the right following the soap?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866022989153862 + ], + "opposite_logprob_tokens": [ + -3.0486602783203125 + ], + "logprob_reward": -0.04866022989153862, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866022989153862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0917179062962532, + 0.04321978613734245, + 0.1317005604505539, + 0.055509958416223526, + 0.030232250690460205, + 0.10684796422719955, + 0.07916395366191864, + 0.1223529651761055, + 0.2947811484336853, + 0.044473543763160706 + ], + "credit_map": [ + 0.011360319331288338, + 0.0077311755158007145, + 0.019975963979959488, + 0.012312899343669415, + 0.007091963663697243, + 0.01149591151624918, + 0.011220581829547882, + 0.010529096238315105, + 0.004630980081856251, + 0.004224477335810661, + 0.006965144071727991, + 0.00648464635014534, + 0.004868280608206987, + 0.006326222326606512, + 0.005324367433786392, + 0.0043956683948636055, + 0.01391369104385376, + 0.015650033950805664, + 0.015121320262551308, + 0.01747037097811699, + 0.016323408111929893, + 0.01692468486726284, + 0.01126567181199789, + 0.025031371042132378, + 0.0072426749393343925, + 0.008772090077400208, + 0.003220890648663044, + 0.006837594322860241, + 0.012812051922082901, + 0.006033876910805702, + 0.004503380041569471, + 0.006087402813136578, + 0.004333922173827887, + 0.0032580012921243906, + 0.0010776208946481347, + 0.0037999104242771864, + 0.0035411203280091286, + 0.0033622318878769875, + 0.007704754360020161, + 0.00315469061024487, + 0.01061580702662468, + 0.012678489089012146, + 0.006266768090426922, + 0.019204942509531975, + 0.01370762288570404, + 0.016912860795855522, + 0.015828153118491173, + 0.01163332350552082, + 0.007095147855579853, + 0.010750639252364635, + 0.006588962394744158, + 0.006828088779002428, + 0.017331648617982864, + 0.009788374416530132, + 0.005892020184546709, + 0.014889076352119446, + 0.01759134605526924, + 0.015089761465787888, + 0.010508522391319275, + 0.015892265364527702, + 0.015285491943359375, + 0.019054485484957695, + 0.01421280950307846, + 0.014718286693096161, + 0.030051598325371742, + 0.050169773399829865, + 0.018803110346198082, + 0.03870689123868942, + 0.0543128065764904, + 0.034821219742298126, + 0.036773063242435455, + 0.031142709776759148, + 0.004224050790071487, + 0.005543916951864958, + 0.002331619383767247, + 0.007434029597789049, + 0.008643888868391514, + 0.00472263852134347, + 0.006780112162232399, + 0.004793287254869938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 7 + ], + "selected_reward": -0.04866022989153862 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a rectangular bar of milky white soap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soap bar sliding across a damp, dark slate-gray tile surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soap surface appear glistening with water droplets and a thin layer of translucent foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the soap leave behind a thick, milky white trail topped with dense, foamy lather as it moves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily to the right following the soap?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520754337310791 + ], + "opposite_logprob_tokens": [ + -1.502075433731079 + ], + "logprob_reward": -0.2520754337310791, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520754337310791, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11470338702201843, + 0.09843676537275314, + 0.07232353836297989, + 0.08914753794670105, + 0.08901229500770569, + 0.10814071446657181, + 0.1139722466468811, + 0.05312391370534897, + 0.14688190817832947, + 0.11425767838954926 + ], + "credit_map": [ + 0.012516016140580177, + 0.010981744155287743, + 0.02196367271244526, + 0.01216669287532568, + 0.008912677876651287, + 0.01985209994018078, + 0.012041112408041954, + 0.016269372776150703, + 0.00903596356511116, + 0.011500023305416107, + 0.015042999759316444, + 0.010868550278246403, + 0.013046173378825188, + 0.013334106653928757, + 0.013548559509217739, + 0.012060409411787987, + 0.007394281215965748, + 0.007352374494075775, + 0.005703171715140343, + 0.016414443030953407, + 0.008812138810753822, + 0.007789795286953449, + 0.009238338097929955, + 0.009618998505175114, + 0.018036991357803345, + 0.008714446797966957, + 0.00930776633322239, + 0.010254793800413609, + 0.008704998530447483, + 0.017385970801115036, + 0.005771121941506863, + 0.010971457697451115, + 0.012127174064517021, + 0.009881260804831982, + 0.005707876291126013, + 0.008357757702469826, + 0.013709616847336292, + 0.012295395135879517, + 0.009931906126439571, + 0.017001314088702202, + 0.01143792737275362, + 0.013116828165948391, + 0.00894797034561634, + 0.01477664615958929, + 0.013307394459843636, + 0.017772341147065163, + 0.013073353096842766, + 0.015708258375525475, + 0.0091925710439682, + 0.012271348387002945, + 0.008671356365084648, + 0.01849963329732418, + 0.01620037667453289, + 0.014803801663219929, + 0.017287004739046097, + 0.017046159133315086, + 0.006457554176449776, + 0.00616866210475564, + 0.0039850021712481976, + 0.008374891243875027, + 0.006770515348762274, + 0.006179331801831722, + 0.006371769588440657, + 0.008816192857921124, + 0.0174663458019495, + 0.016232168301939964, + 0.01218855008482933, + 0.020451121032238007, + 0.02017034776508808, + 0.016728373244404793, + 0.02326187863945961, + 0.020383134484291077, + 0.014040905050933361, + 0.014951176941394806, + 0.008465156890451908, + 0.012642375193536282, + 0.02396010048687458, + 0.01116072479635477, + 0.013287580572068691, + 0.015749668702483177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.2520754337310791 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, high-angle shot allows the camera to track steadily to the right, following a rectangular bar of milky white soap sliding across a damp, dark slate-gray tile surface. The soap bar, worn smooth from use, its surface glistening with water droplets and a thin layer of translucent foam, appears soft and almost plastic. Propelled by an unseen force, it moves steadily and deliberately, leaving behind a thick, milky white trail topped with dense, foamy lather that contrasts sharply against the dark, reflective tile. A single, soft, overhead light casts bright, mirror-like highlights on the water, the glossy soap, and the wet tile, while creating deep shadows in the cement line between them, creating a minimalist, clean, and tranquil atmosphere focused entirely on the satisfying interplay of tactile texture and movement.", + "seed": 31515, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_8/index_9116_seed_31515_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.035288408678025004, + "vlm_logprob": -0.035288408678025004, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.035288408678025004, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a rectangular bar of milky white soap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soap bar sliding across a damp, dark slate-gray tile surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soap surface appear glistening with water droplets and a thin layer of translucent foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the soap leave behind a thick, milky white trail topped with dense, foamy lather as it moves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily to the right following the soap?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010602100752294064 + ], + "opposite_logprob_tokens": [ + -5.385601997375488 + ], + "logprob_reward": -0.010602100752294064, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.010602100752294064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06762588769197464, + 0.04937845095992088, + 0.05659473314881325, + 0.07527290284633636, + 0.08902167528867722, + 0.04769482836127281, + 0.29260316491127014, + 0.10599852353334427, + 0.1476796716451645, + 0.0681302398443222 + ], + "credit_map": [ + 0.00427576107904315, + 0.013166451826691628, + 0.0059920912608504295, + 0.010410412214696407, + 0.009444464929401875, + 0.0068562026135623455, + 0.008450504392385483, + 0.009029993787407875, + 0.005581789184361696, + 0.005474259611219168, + 0.007171764504164457, + 0.00595935108140111, + 0.007675938308238983, + 0.0040926202200353146, + 0.005904770456254482, + 0.00751795619726181, + 0.01166430301964283, + 0.014646024443209171, + 0.005994407460093498, + 0.0046793813817203045, + 0.005853159818798304, + 0.005117971915751696, + 0.0054190494120121, + 0.0032204347662627697, + 0.007153367158025503, + 0.007320302538573742, + 0.01893409714102745, + 0.01286152470856905, + 0.006970955524593592, + 0.007028729654848576, + 0.010027497075498104, + 0.004976428113877773, + 0.008179397322237492, + 0.015405481681227684, + 0.011585609056055546, + 0.007714956533163786, + 0.023323677480220795, + 0.006625002715736628, + 0.00583390798419714, + 0.010353635996580124, + 0.0050394730642437935, + 0.005417993292212486, + 0.007437212858349085, + 0.006505319382995367, + 0.006982512306421995, + 0.005139457061886787, + 0.004646028392016888, + 0.006526834797114134, + 0.010048958472907543, + 0.015819469466805458, + 0.011841872707009315, + 0.011793901212513447, + 0.10063060373067856, + 0.12438879907131195, + 0.007160256151109934, + 0.010919307358562946, + 0.011474436149001122, + 0.01119889598339796, + 0.018535830080509186, + 0.009372475557029247, + 0.015585649758577347, + 0.01660478487610817, + 0.01384793408215046, + 0.009378521703183651, + 0.014136250130832195, + 0.011845966801047325, + 0.03580203652381897, + 0.020422374829649925, + 0.028117496520280838, + 0.01273408718407154, + 0.016345348209142685, + 0.008276098407804966, + 0.006100240629166365, + 0.007238670717924833, + 0.003938014153391123, + 0.006090488750487566, + 0.011833353899419308, + 0.007426055613905191, + 0.02041439712047577, + 0.005089016165584326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.010602100752294064 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a rectangular bar of milky white soap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soap bar sliding across a damp, dark slate-gray tile surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soap surface appear glistening with water droplets and a thin layer of translucent foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the soap leave behind a thick, milky white trail topped with dense, foamy lather as it moves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily to the right following the soap?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046497792936861515 + ], + "opposite_logprob_tokens": [ + -5.379649639129639 + ], + "logprob_reward": -0.0046497792936861515, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046497792936861515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18354378640651703, + 0.0649634525179863, + 0.05214424803853035, + 0.06691334396600723, + 0.0340791679918766, + 0.05373132601380348, + 0.10680670291185379, + 0.1377338469028473, + 0.2097533792257309, + 0.09033076465129852 + ], + "credit_map": [ + 0.008513271808624268, + 0.043450627475976944, + 0.017204737290740013, + 0.029927637428045273, + 0.020688004791736603, + 0.015355696901679039, + 0.021177228540182114, + 0.027226589620113373, + 0.005434802733361721, + 0.006802024319767952, + 0.007380225230008364, + 0.00776151567697525, + 0.012817000970244408, + 0.007834510877728462, + 0.009501376189291477, + 0.00743199372664094, + 0.01270067784935236, + 0.015189100988209248, + 0.005002025049179792, + 0.0036729963030666113, + 0.006081310100853443, + 0.0024329437874257565, + 0.004603862296789885, + 0.0024613349232822657, + 0.007780517917126417, + 0.0047724125906825066, + 0.013041896745562553, + 0.01391773670911789, + 0.00849733967334032, + 0.005856867879629135, + 0.009268916212022305, + 0.003777657402679324, + 0.004240275826305151, + 0.005710451398044825, + 0.0041737426072359085, + 0.004155855160206556, + 0.0055074081756174564, + 0.003287816420197487, + 0.003224305110052228, + 0.003779312130063772, + 0.004657837562263012, + 0.006742764264345169, + 0.008577116765081882, + 0.008801908232271671, + 0.005881331395357847, + 0.005787968635559082, + 0.006478501949459314, + 0.006803893484175205, + 0.010546458885073662, + 0.008451846428215504, + 0.007581104524433613, + 0.007283901795744896, + 0.026718704029917717, + 0.03239559009671211, + 0.007764813955873251, + 0.006064280401915312, + 0.017025941982865334, + 0.017623765394091606, + 0.019441938027739525, + 0.012768787331879139, + 0.01972479745745659, + 0.01972202956676483, + 0.01940448023378849, + 0.012022114358842373, + 0.008256446570158005, + 0.00880453735589981, + 0.02287784218788147, + 0.015279000625014305, + 0.054042182862758636, + 0.016456475481390953, + 0.0671299546957016, + 0.01690693199634552, + 0.009235839359462261, + 0.012189587578177452, + 0.010596626438200474, + 0.017366329208016396, + 0.012339577078819275, + 0.011284721083939075, + 0.009908112697303295, + 0.007409980054944754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.0046497792936861515 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a rectangular bar of milky white soap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soap bar sliding across a damp, dark slate-gray tile surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soap surface appear glistening with water droplets and a thin layer of translucent foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the soap leave behind a thick, milky white trail topped with dense, foamy lather as it moves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily to the right following the soap?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005956161767244339 + ], + "opposite_logprob_tokens": [ + -5.130956172943115 + ], + "logprob_reward": -0.005956161767244339, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005956161767244339, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23707939684391022, + 0.044190533459186554, + 0.06578675657510757, + 0.09678813070058823, + 0.09628977626562119, + 0.037950947880744934, + 0.038641832768917084, + 0.08622831106185913, + 0.18268492817878723, + 0.11435937881469727 + ], + "credit_map": [ + 0.009546327404677868, + 0.053319741040468216, + 0.028104305267333984, + 0.03721931576728821, + 0.03014475107192993, + 0.02190912887454033, + 0.028449861332774162, + 0.028385953977704048, + 0.006306075025349855, + 0.004932187031954527, + 0.007556319702416658, + 0.005168860778212547, + 0.006091957446187735, + 0.003523878986015916, + 0.005658173467963934, + 0.004953082650899887, + 0.017840556800365448, + 0.020555254071950912, + 0.00553867407143116, + 0.004264913033694029, + 0.007013909984380007, + 0.0031428574584424496, + 0.004492316860705614, + 0.0029382752254605293, + 0.0098658986389637, + 0.0071550761349499226, + 0.019568007439374924, + 0.018301017582416534, + 0.009385732933878899, + 0.010319013148546219, + 0.013601895421743393, + 0.008591491729021072, + 0.00813968200236559, + 0.02094763144850731, + 0.008939757943153381, + 0.005811112932860851, + 0.028122326359152794, + 0.007258351426571608, + 0.006416363641619682, + 0.010654546320438385, + 0.0038510384038090706, + 0.004957843571901321, + 0.006884498987346888, + 0.0045804125256836414, + 0.004149586893618107, + 0.004743683617562056, + 0.003490723902359605, + 0.005293156951665878, + 0.004311221651732922, + 0.004648404195904732, + 0.0038855564780533314, + 0.004282967187464237, + 0.006783683784306049, + 0.00897325947880745, + 0.0026122897397726774, + 0.003144450020045042, + 0.012056651525199413, + 0.00989221315830946, + 0.010477020405232906, + 0.007009304128587246, + 0.01331521850079298, + 0.014410502277314663, + 0.010770641267299652, + 0.008296758867800236, + 0.01263691857457161, + 0.00906460639089346, + 0.03907964751124382, + 0.019337516278028488, + 0.018174175173044205, + 0.052307747304439545, + 0.024170169606804848, + 0.00791415199637413, + 0.01206983346492052, + 0.008822055533528328, + 0.006150615401566029, + 0.008236153982579708, + 0.01752701960504055, + 0.013311040587723255, + 0.0416870191693306, + 0.006555637344717979 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.005956161767244339 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a rectangular bar of milky white soap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soap bar sliding across a damp, dark slate-gray tile surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soap surface appear glistening with water droplets and a thin layer of translucent foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the soap leave behind a thick, milky white trail topped with dense, foamy lather as it moves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily to the right following the soap?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495204031467438 + ], + "opposite_logprob_tokens": [ + -2.9299521446228027 + ], + "logprob_reward": -0.05495204031467438, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495204031467438, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23726101219654083, + 0.05080195888876915, + 0.03629113733768463, + 0.0428992323577404, + 0.044973574578762054, + 0.03566873073577881, + 0.1379593312740326, + 0.036057248711586, + 0.20074987411499023, + 0.1773378700017929 + ], + "credit_map": [ + 0.010131324641406536, + 0.0552673414349556, + 0.029402604326605797, + 0.03662494570016861, + 0.028675932437181473, + 0.021171968430280685, + 0.029760826379060745, + 0.026226088404655457, + 0.005500634200870991, + 0.006641444284468889, + 0.005903976038098335, + 0.006418890319764614, + 0.009731956757605076, + 0.0048431516624987125, + 0.0068159690126776695, + 0.00494594220072031, + 0.009690760634839535, + 0.009753684513270855, + 0.004252532497048378, + 0.002952837385237217, + 0.0032274590339511633, + 0.001677399268373847, + 0.002275299746543169, + 0.0024611675180494785, + 0.004202772863209248, + 0.003399922978132963, + 0.007075217552483082, + 0.006897582672536373, + 0.004423141945153475, + 0.006135448347777128, + 0.006797784939408302, + 0.00396736292168498, + 0.005418488755822182, + 0.006086476147174835, + 0.005654498469084501, + 0.004809490870684385, + 0.006875815335661173, + 0.004585728514939547, + 0.0058522955514490604, + 0.005690784193575382, + 0.004381007514894009, + 0.003973538056015968, + 0.006276067346334457, + 0.0037925331853330135, + 0.003896457841619849, + 0.005114298779517412, + 0.003546485910192132, + 0.004688344895839691, + 0.003475965466350317, + 0.006075595505535603, + 0.0036246650852262974, + 0.004337858874350786, + 0.05072833225131035, + 0.062267474830150604, + 0.003458096645772457, + 0.003991350531578064, + 0.005571061745285988, + 0.00501960888504982, + 0.003959454130381346, + 0.00252034654840827, + 0.005449414718896151, + 0.0052974773570895195, + 0.005190122872591019, + 0.003049766644835472, + 0.00892993900924921, + 0.008149455301463604, + 0.02547614462673664, + 0.014880308881402016, + 0.01884021796286106, + 0.050678614526987076, + 0.060068529099226, + 0.013726677745580673, + 0.013803951442241669, + 0.013697480782866478, + 0.006849354598671198, + 0.011578640900552273, + 0.028980713337659836, + 0.020101627334952354, + 0.07381681352853775, + 0.008509309962391853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.05495204031467438 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a rectangular bar of milky white soap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soap bar sliding across a damp, dark slate-gray tile surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soap surface appear glistening with water droplets and a thin layer of translucent foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the soap leave behind a thick, milky white trail topped with dense, foamy lather as it moves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily to the right following the soap?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1002819612622261 + ], + "opposite_logprob_tokens": [ + -2.3502819538116455 + ], + "logprob_reward": -0.1002819612622261, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1002819612622261, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07085928320884705, + 0.05479585379362106, + 0.02794557251036167, + 0.0567563995718956, + 0.047195471823215485, + 0.07206505537033081, + 0.1807672381401062, + 0.09761106222867966, + 0.3343435823917389, + 0.057660479098558426 + ], + "credit_map": [ + 0.003520095720887184, + 0.016822636127471924, + 0.007911182940006256, + 0.010733814910054207, + 0.008587782271206379, + 0.006581414025276899, + 0.008086164481937885, + 0.008616185747087002, + 0.006923021283000708, + 0.007107852958142757, + 0.006705595646053553, + 0.00779041089117527, + 0.007351868320256472, + 0.005506942514330149, + 0.006468433421105146, + 0.0069417282938957214, + 0.0057688867673277855, + 0.005908014718443155, + 0.002894975943490863, + 0.0027627150993794203, + 0.002949680667370558, + 0.002591072814539075, + 0.0026998608373105526, + 0.0023703649640083313, + 0.006177615374326706, + 0.006781978067010641, + 0.00989474542438984, + 0.008190622553229332, + 0.0071529000997543335, + 0.006296165287494659, + 0.008402818813920021, + 0.0038595523219555616, + 0.003403658280149102, + 0.009246502071619034, + 0.003612372325733304, + 0.004078994505107403, + 0.013040703721344471, + 0.003819830249994993, + 0.0036320535000413656, + 0.0063613541424274445, + 0.00568566145375371, + 0.010372993536293507, + 0.013502180576324463, + 0.009041955694556236, + 0.0078870989382267, + 0.007065104320645332, + 0.00969680305570364, + 0.008813261985778809, + 0.0208743829280138, + 0.014683423563838005, + 0.015606400556862354, + 0.014130116440355778, + 0.04008165001869202, + 0.048090945929288864, + 0.014281214214861393, + 0.013019118458032608, + 0.012435287237167358, + 0.010575855150818825, + 0.009907140396535397, + 0.009211141616106033, + 0.01528854202479124, + 0.016579026356339455, + 0.012351184152066708, + 0.011262878775596619, + 0.014090358279645443, + 0.013555649667978287, + 0.037438031286001205, + 0.0262642540037632, + 0.03028041310608387, + 0.08487901091575623, + 0.10462025552988052, + 0.0232156440615654, + 0.004716143012046814, + 0.004958929028362036, + 0.0028921805787831545, + 0.004963529761880636, + 0.007696997839957476, + 0.009613312780857086, + 0.01934383623301983, + 0.0034755519591271877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.1002819612622261 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, high-angle shot allows the camera to track steadily to the right, following a rectangular bar of milky white soap sliding across a damp, dark slate-gray tile surface. The soap bar, worn smooth from use, its surface glistening with water droplets and a thin layer of translucent foam, appears soft and almost plastic. Propelled by an unseen force, it moves steadily and deliberately, leaving behind a thick, milky white trail topped with dense, foamy lather that contrasts sharply against the dark, reflective tile. A single, soft, overhead light casts bright, mirror-like highlights on the water, the glossy soap, and the wet tile, while creating deep shadows in the cement line between them, creating a minimalist, clean, and tranquil atmosphere focused entirely on the satisfying interplay of tactile texture and movement.", + "seed": 31516, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_8/index_9116_seed_31516_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0453021751716733, + "vlm_logprob": -0.0453021751716733, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.0453021751716733, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a rectangular bar of milky white soap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soap bar sliding across a damp, dark slate-gray tile surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soap surface appear glistening with water droplets and a thin layer of translucent foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the soap leave behind a thick, milky white trail topped with dense, foamy lather as it moves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily to the right following the soap?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012009929865598679 + ], + "opposite_logprob_tokens": [ + -5.262010097503662 + ], + "logprob_reward": -0.012009929865598679, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.012009929865598679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1727694571018219, + 0.08242316544055939, + 0.06302504986524582, + 0.08992209285497665, + 0.29589951038360596, + 0.05586101487278938, + 0.07525216788053513, + 0.03664363920688629, + 0.04861271381378174, + 0.07959121465682983 + ], + "credit_map": [ + 0.012743955478072166, + 0.018191004171967506, + 0.007869490422308445, + 0.011283974163234234, + 0.01825374737381935, + 0.038562625646591187, + 0.017788950353860855, + 0.04807570204138756, + 0.006068552378565073, + 0.008862083777785301, + 0.013124760240316391, + 0.0125584052875638, + 0.010611589066684246, + 0.009267620742321014, + 0.012778679840266705, + 0.009151477366685867, + 0.0061563048511743546, + 0.007116375956684351, + 0.005528578069061041, + 0.005544314160943031, + 0.0077927191741764545, + 0.01033853366971016, + 0.005964543204754591, + 0.014583681710064411, + 0.008996149525046349, + 0.009940549731254578, + 0.00924156978726387, + 0.007084615062922239, + 0.01065138541162014, + 0.017935099080204964, + 0.0074547971598804, + 0.01861792616546154, + 0.015108523890376091, + 0.021970126777887344, + 0.031187407672405243, + 0.031014611944556236, + 0.034149862825870514, + 0.04330676048994064, + 0.06343062967061996, + 0.05573158711194992, + 0.006012182217091322, + 0.005285254213958979, + 0.006682487670332193, + 0.006355554796755314, + 0.006813338492065668, + 0.009086408652365208, + 0.005500718951225281, + 0.010125067085027695, + 0.00828813761472702, + 0.009145795367658138, + 0.009008013643324375, + 0.014148617163300514, + 0.011040621437132359, + 0.008950418792665005, + 0.007287842221558094, + 0.007382721174508333, + 0.0034375032410025597, + 0.005594538990408182, + 0.007034578826278448, + 0.00415960606187582, + 0.0039034432265907526, + 0.0066255745477974415, + 0.0026534677017480135, + 0.003234927775338292, + 0.004791883286088705, + 0.01686634123325348, + 0.004760325886309147, + 0.0023922184482216835, + 0.012167694047093391, + 0.002578067360445857, + 0.002352799056097865, + 0.0027033851947635412, + 0.011662453413009644, + 0.011221790686249733, + 0.01609968952834606, + 0.011329961940646172, + 0.008926981128752232, + 0.0060005770064890385, + 0.006881885230541229, + 0.007467867806553841 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.012009929865598679 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a rectangular bar of milky white soap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soap bar sliding across a damp, dark slate-gray tile surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soap surface appear glistening with water droplets and a thin layer of translucent foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the soap leave behind a thick, milky white trail topped with dense, foamy lather as it moves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily to the right following the soap?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009799941442906857 + ], + "opposite_logprob_tokens": [ + -4.634799957275391 + ], + "logprob_reward": -0.009799941442906857, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009799941442906857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3143601417541504, + 0.0651571974158287, + 0.04386402294039726, + 0.037236664444208145, + 0.13527853786945343, + 0.06911195069551468, + 0.04251250997185707, + 0.0947389304637909, + 0.05648678541183472, + 0.14125323295593262 + ], + "credit_map": [ + 0.021582776680588722, + 0.020504893735051155, + 0.00923984032124281, + 0.007196144666522741, + 0.024563467130064964, + 0.0974501445889473, + 0.010267707519233227, + 0.12355518341064453, + 0.0049390774220228195, + 0.005559540819376707, + 0.007235452067106962, + 0.010664077475667, + 0.012773964554071426, + 0.006505556404590607, + 0.01040254533290863, + 0.007076993118971586, + 0.0037922037299722433, + 0.006044710986316204, + 0.00474070617929101, + 0.0030281280633062124, + 0.008413019590079784, + 0.005939716473221779, + 0.0062551177106797695, + 0.005650423001497984, + 0.003149121766909957, + 0.0036542899906635284, + 0.003691293764859438, + 0.0027143466286361217, + 0.003518631448969245, + 0.007824364118278027, + 0.0031905879732221365, + 0.009494028985500336, + 0.008943225257098675, + 0.011447473429143429, + 0.019007476046681404, + 0.01309987436980009, + 0.009952244348824024, + 0.020167959854006767, + 0.025089198723435402, + 0.027571093291044235, + 0.006647356785833836, + 0.00812196359038353, + 0.00783168151974678, + 0.009398066438734531, + 0.010457183234393597, + 0.009781226515769958, + 0.0073875198140740395, + 0.009486961178481579, + 0.0038426131941378117, + 0.0060892002657055855, + 0.005458053667098284, + 0.007761902175843716, + 0.006057254038751125, + 0.0049622999504208565, + 0.0040873149409890175, + 0.004253878723829985, + 0.01047173235565424, + 0.017225561663508415, + 0.014320342801511288, + 0.009129500947892666, + 0.01187954656779766, + 0.016656825318932533, + 0.006286259274929762, + 0.008769172243773937, + 0.007593156769871712, + 0.020903579890727997, + 0.00622958317399025, + 0.004372913856059313, + 0.008686787448823452, + 0.004028148949146271, + 0.0023286184296011925, + 0.0023440015502274036, + 0.015271029435098171, + 0.038719434291124344, + 0.01919456571340561, + 0.015788599848747253, + 0.026262570172548294, + 0.008625430054962635, + 0.01003559585660696, + 0.00735603179782629 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.009799941442906857 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a rectangular bar of milky white soap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soap bar sliding across a damp, dark slate-gray tile surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soap surface appear glistening with water droplets and a thin layer of translucent foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the soap leave behind a thick, milky white trail topped with dense, foamy lather as it moves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily to the right following the soap?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007642081007361412 + ], + "opposite_logprob_tokens": [ + -4.8826422691345215 + ], + "logprob_reward": -0.007642081007361412, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007642081007361412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32820942997932434, + 0.12179921567440033, + 0.12885987758636475, + 0.06908624619245529, + 0.05156265199184418, + 0.05130864679813385, + 0.027368076145648956, + 0.02942315675318241, + 0.07844056934118271, + 0.11394210904836655 + ], + "credit_map": [ + 0.02113039791584015, + 0.019581763073801994, + 0.011712388135492802, + 0.012936705723404884, + 0.02502010390162468, + 0.10003356635570526, + 0.015869615599513054, + 0.12192489951848984, + 0.0099570257589221, + 0.011830205097794533, + 0.0122571662068367, + 0.01216429378837347, + 0.03271733969449997, + 0.01170004066079855, + 0.01728993095457554, + 0.013883225619792938, + 0.011465883813798428, + 0.019971337169408798, + 0.00723981624469161, + 0.009019757620990276, + 0.026726219803094864, + 0.01949654147028923, + 0.011342908255755901, + 0.023597419261932373, + 0.006275508552789688, + 0.010746990330517292, + 0.00845009833574295, + 0.006314850877970457, + 0.009440078400075436, + 0.008188041858375072, + 0.006469035055488348, + 0.013201639987528324, + 0.003476149868220091, + 0.00649312324821949, + 0.007204308640211821, + 0.006440301891416311, + 0.004565519746392965, + 0.00665406184270978, + 0.010628479532897472, + 0.006100705359131098, + 0.004111090209335089, + 0.00683139031752944, + 0.005489672534167767, + 0.007302975747734308, + 0.006615643389523029, + 0.0074989208951592445, + 0.004828231874853373, + 0.008630718104541302, + 0.0032004299573600292, + 0.003508790861815214, + 0.003700242144986987, + 0.004641249310225248, + 0.003626326099038124, + 0.0031814633402973413, + 0.002663694554939866, + 0.0028458782471716404, + 0.002978784032166004, + 0.005753441713750362, + 0.004501330200582743, + 0.002977883443236351, + 0.004114639014005661, + 0.00416121631860733, + 0.002163713099434972, + 0.002772149397060275, + 0.008850540965795517, + 0.027246741577982903, + 0.0069806440733373165, + 0.0029030530713498592, + 0.02064693346619606, + 0.003825433785095811, + 0.002852685982361436, + 0.00513453409075737, + 0.017904814332723618, + 0.02119595929980278, + 0.027232488617300987, + 0.018664546310901642, + 0.010158364661037922, + 0.007012596819549799, + 0.006317330524325371, + 0.005456000100821257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.007642081007361412 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a rectangular bar of milky white soap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soap bar sliding across a damp, dark slate-gray tile surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soap surface appear glistening with water droplets and a thin layer of translucent foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the soap leave behind a thick, milky white trail topped with dense, foamy lather as it moves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily to the right following the soap?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003460079431534 + ], + "opposite_logprob_tokens": [ + -2.6950345039367676 + ], + "logprob_reward": -0.07003460079431534, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003460079431534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2473616600036621, + 0.11386531591415405, + 0.03986240550875664, + 0.10606199502944946, + 0.05102686956524849, + 0.057744100689888, + 0.09043402969837189, + 0.06490714102983475, + 0.03202546015381813, + 0.19671104848384857 + ], + "credit_map": [ + 0.018533773720264435, + 0.023136617615818977, + 0.010031703859567642, + 0.01181692536920309, + 0.02106030099093914, + 0.06703587621450424, + 0.01662174053490162, + 0.07912472635507584, + 0.008301179856061935, + 0.011599630117416382, + 0.014056112617254257, + 0.015800274908542633, + 0.016444550827145576, + 0.012630670331418514, + 0.019411621615290642, + 0.01562126912176609, + 0.0032467551063746214, + 0.004996758420020342, + 0.005979534704238176, + 0.0032071107998490334, + 0.006116877309978008, + 0.00572304567322135, + 0.0057015628553926945, + 0.00489076180383563, + 0.008373663760721684, + 0.013088996522128582, + 0.00919532310217619, + 0.005427482072263956, + 0.011294706724584103, + 0.02364042028784752, + 0.006286812014877796, + 0.02875458635389805, + 0.0035932783503085375, + 0.006162773817777634, + 0.004513850435614586, + 0.007073361426591873, + 0.005012422800064087, + 0.008019221015274525, + 0.006042448803782463, + 0.010609510354697704, + 0.004630221053957939, + 0.007875126786530018, + 0.005361499264836311, + 0.008732344955205917, + 0.008332887664437294, + 0.007535489276051521, + 0.0065412516705691814, + 0.008735278621315956, + 0.007910942658782005, + 0.012066558003425598, + 0.011557131074368954, + 0.01968163065612316, + 0.011685931123793125, + 0.009101574309170246, + 0.010103720240294933, + 0.008326544426381588, + 0.007729158736765385, + 0.010157412849366665, + 0.00906829908490181, + 0.009113977663218975, + 0.008797496557235718, + 0.008376387879252434, + 0.005075778812170029, + 0.006588628049939871, + 0.0028534322045743465, + 0.007237023673951626, + 0.0033504071179777384, + 0.002670405665412545, + 0.008669045753777027, + 0.0027984012849628925, + 0.0017008199356496334, + 0.002745925448834896, + 0.028264204040169716, + 0.0328742153942585, + 0.044538382440805435, + 0.03280450403690338, + 0.018418196588754654, + 0.014390087686479092, + 0.01559997908771038, + 0.009821467101573944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.07003460079431534 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a rectangular bar of milky white soap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soap bar sliding across a damp, dark slate-gray tile surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the soap surface appear glistening with water droplets and a thin layer of translucent foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the soap leave behind a thick, milky white trail topped with dense, foamy lather as it moves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily to the right following the soap?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270243227481842 + ], + "opposite_logprob_tokens": [ + -2.1270244121551514 + ], + "logprob_reward": -0.1270243227481842, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270243227481842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10969103872776031, + 0.20735381543636322, + 0.0703100860118866, + 0.13536687195301056, + 0.052372049540281296, + 0.10532104223966599, + 0.0682264193892479, + 0.05011945217847824, + 0.06865134835243225, + 0.13258787989616394 + ], + "credit_map": [ + 0.008806144818663597, + 0.014869928359985352, + 0.009227980859577656, + 0.011585250496864319, + 0.022859876975417137, + 0.010808208957314491, + 0.018925504758954048, + 0.012608146294951439, + 0.015778861939907074, + 0.02211916819214821, + 0.030180469155311584, + 0.02993745356798172, + 0.03442917391657829, + 0.019410384818911552, + 0.031154757365584373, + 0.02434353530406952, + 0.005771552678197622, + 0.007114777807146311, + 0.00864130724221468, + 0.006825853604823351, + 0.006161691155284643, + 0.011895610950887203, + 0.006444231141358614, + 0.01745506376028061, + 0.016099346801638603, + 0.018122714012861252, + 0.009212185628712177, + 0.009207434020936489, + 0.0184178426861763, + 0.0258018858730793, + 0.010689174756407738, + 0.0278162844479084, + 0.0033306144177913666, + 0.004172659479081631, + 0.005281933583319187, + 0.0054351151920855045, + 0.006550433114171028, + 0.007132907398045063, + 0.010520866140723228, + 0.009947518818080425, + 0.0068574449978768826, + 0.010773145593702793, + 0.01205763965845108, + 0.01648067496716976, + 0.017280030995607376, + 0.015889856964349747, + 0.011749964207410812, + 0.014232286252081394, + 0.007148963864892721, + 0.008168604224920273, + 0.008881510235369205, + 0.01165306381881237, + 0.00963519886136055, + 0.00795270036906004, + 0.006963411346077919, + 0.007822960615158081, + 0.006372001022100449, + 0.007834934629499912, + 0.007176932413130999, + 0.008660092949867249, + 0.007015599869191647, + 0.005109370220452547, + 0.0038867457769811153, + 0.0040637725032866, + 0.007041746750473976, + 0.024748126044869423, + 0.0075950296595692635, + 0.0038533639162778854, + 0.015597176738083363, + 0.003139280714094639, + 0.0030691740103065968, + 0.0036074514500796795, + 0.015702415257692337, + 0.01667146384716034, + 0.027426235377788544, + 0.021942168474197388, + 0.019592953845858574, + 0.011319302953779697, + 0.01294319611042738, + 0.006990140303969383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 9 + ], + "selected_reward": -0.1270243227481842 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_9/index_tensor([8986, 8986])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_9/index_tensor([8986, 8986])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e83ab31b23d9724571898f0fa4229bffbd7489a3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_9/index_tensor([8986, 8986])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Adopting the desaturated, gritty look of 1990s cinema, a steady medium shot captures the moment of industrial disaster with chilling clarity. From the vast ceiling of an abandoned factory, a complex, greasy mechanism of gears and chains violently breaks loose. The assembly, a tangled web of interlocking cast-iron gears and thick roller chains, coated in a dark, sticky lubricant, shears from its rusted mounting bracket in a puff of metallic dust. Under immense grinding force, the entire rusted assembly slides heavily along the angled surface of the large, grimy skylight panel, its immense weight causing the grime-caked glass to groan and shatter. The camera remains stationary, emphasizing the weight and danger as the jagged edges of the machinery scrape deep grooves into the grime before it finally tumbles over the edge of the steel frame, slowly spinning in the air for a terrifying moment before beginning its inexorable fall, creating a palpable atmosphere of impending impact and mechanical violence.", + "seed": 22742, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_9/index_8986_seed_22742_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.14128770772367716, + "vlm_logprob": -0.14128770772367716, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.14128770772367716, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the vast ceiling of an abandoned factory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a complex assembly of interlocking cast-iron gears and thick roller chains coated in dark lubricant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the machinery shear from a rusted mounting bracket and slide along a large, grimy skylight panel before tumbling over the edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a steady medium shot where the camera remains stationary throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style desaturated and gritty, resembling the look of 1990s cinema?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.23846150934696198 + ], + "opposite_logprob_tokens": [ + -1.6134614944458008 + ], + "logprob_reward": -0.23846150934696198, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.23846150934696198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07184316962957382, + 0.02880876138806343, + 0.045118581503629684, + 0.31816551089286804, + 0.1796666830778122, + 0.06723861396312714, + 0.03799087181687355, + 0.09216400980949402, + 0.07322116196155548, + 0.08578269928693771 + ], + "credit_map": [ + 0.00971938855946064, + 0.00618814816698432, + 0.007102373521775007, + 0.007205060217529535, + 0.006435480900108814, + 0.009961548261344433, + 0.010131921619176865, + 0.015099240466952324, + 0.0033491086214780807, + 0.00378888426348567, + 0.003742692293599248, + 0.004007876385003328, + 0.002364647574722767, + 0.003919616807252169, + 0.003738474566489458, + 0.003897458780556917, + 0.005397286731749773, + 0.004100523889064789, + 0.008112205192446709, + 0.007136342581361532, + 0.004644989501684904, + 0.004258094355463982, + 0.0056703295558691025, + 0.00579880690202117, + 0.03216060250997543, + 0.030955906957387924, + 0.028053291141986847, + 0.034871168434619904, + 0.04656499624252319, + 0.04370710253715515, + 0.04607875272631645, + 0.05577365309000015, + 0.027323752641677856, + 0.028417326509952545, + 0.013599284924566746, + 0.018841303884983063, + 0.024080034345388412, + 0.024806460365653038, + 0.01832093484699726, + 0.02427755482494831, + 0.010953998193144798, + 0.009209433570504189, + 0.005629555322229862, + 0.010028531774878502, + 0.00661035580560565, + 0.007751044351607561, + 0.007213517092168331, + 0.009842174127697945, + 0.004237270448356867, + 0.003879935946315527, + 0.004259821958839893, + 0.005373142659664154, + 0.004114303272217512, + 0.005025469232350588, + 0.005833460949361324, + 0.00526746641844511, + 0.007698903791606426, + 0.015632212162017822, + 0.007244485430419445, + 0.0063543361611664295, + 0.025659669190645218, + 0.014241643249988556, + 0.006736913230270147, + 0.008595839142799377, + 0.011231303215026855, + 0.009939318522810936, + 0.011518994346261024, + 0.005262331105768681, + 0.011748711578547955, + 0.006505220662802458, + 0.005182880442589521, + 0.01183239370584488, + 0.011451547965407372, + 0.01114455796778202, + 0.008703862316906452, + 0.010036070831120014, + 0.010964983142912388, + 0.010713130235671997, + 0.01025164220482111, + 0.01251689251512289 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 7 + ], + "selected_reward": -0.23846150934696198 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the vast ceiling of an abandoned factory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a complex assembly of interlocking cast-iron gears and thick roller chains coated in dark lubricant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the machinery shear from a rusted mounting bracket and slide along a large, grimy skylight panel before tumbling over the edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a steady medium shot where the camera remains stationary throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style desaturated and gritty, resembling the look of 1990s cinema?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009819891303777695 + ], + "opposite_logprob_tokens": [ + -4.634819984436035 + ], + "logprob_reward": -0.009819891303777695, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009819891303777695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16094182431697845, + 0.028719408437609673, + 0.07732295244932175, + 0.3681551516056061, + 0.11264842003583908, + 0.053613122552633286, + 0.06955447793006897, + 0.04544268921017647, + 0.030438948422670364, + 0.053162992000579834 + ], + "credit_map": [ + 0.019568150863051414, + 0.014833115041255951, + 0.0140532860532403, + 0.01952548511326313, + 0.015881890431046486, + 0.01573907770216465, + 0.026967305690050125, + 0.03437351435422897, + 0.003588230349123478, + 0.0028995817992836237, + 0.003408578922972083, + 0.004238588269799948, + 0.0038001835346221924, + 0.0034145181998610497, + 0.003996316809207201, + 0.0033734100870788097, + 0.009271114133298397, + 0.007157839369028807, + 0.008752496913075447, + 0.00906912051141262, + 0.009679993614554405, + 0.00789479911327362, + 0.011727808974683285, + 0.013769780285656452, + 0.03682239353656769, + 0.03277164697647095, + 0.03223318234086037, + 0.0395599901676178, + 0.051742054522037506, + 0.05116324499249458, + 0.05546349287033081, + 0.06839916855096817, + 0.01489671878516674, + 0.018322797492146492, + 0.00970379076898098, + 0.011451187543570995, + 0.015162141062319279, + 0.013825827278196812, + 0.011586534790694714, + 0.017699414864182472, + 0.007795638870447874, + 0.006180556025356054, + 0.0061982241459190845, + 0.005997548345476389, + 0.005557723343372345, + 0.005595392547547817, + 0.009514582343399525, + 0.00677345460280776, + 0.007022274658083916, + 0.008768624626100063, + 0.006287948694080114, + 0.010429364629089832, + 0.008752716705203056, + 0.011951147578656673, + 0.007031301036477089, + 0.009311096742749214, + 0.004549532663077116, + 0.00488034775480628, + 0.005893365945667028, + 0.005365738645195961, + 0.0075499797239899635, + 0.006169190164655447, + 0.00580102764070034, + 0.0052335066720843315, + 0.003694960381835699, + 0.004035842604935169, + 0.004998100455850363, + 0.003658180357888341, + 0.0047604115679860115, + 0.003424972528591752, + 0.002414081245660782, + 0.003452397882938385, + 0.007599995471537113, + 0.007994616404175758, + 0.004756199661642313, + 0.005217391531914473, + 0.00788943376392126, + 0.005882772617042065, + 0.005536411888897419, + 0.008286172524094582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.009819891303777695 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the vast ceiling of an abandoned factory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a complex assembly of interlocking cast-iron gears and thick roller chains coated in dark lubricant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the machinery shear from a rusted mounting bracket and slide along a large, grimy skylight panel before tumbling over the edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a steady medium shot where the camera remains stationary throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style desaturated and gritty, resembling the look of 1990s cinema?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011084645055234432 + ], + "opposite_logprob_tokens": [ + -4.51108455657959 + ], + "logprob_reward": -0.011084645055234432, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011084645055234432, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09087783098220825, + 0.02820352092385292, + 0.09512774646282196, + 0.1383160650730133, + 0.08306208997964859, + 0.07510656863451004, + 0.11358874291181564, + 0.14963768422603607, + 0.048476606607437134, + 0.17760314047336578 + ], + "credit_map": [ + 0.010108046233654022, + 0.007876838557422161, + 0.012761961668729782, + 0.011061357334256172, + 0.007278156001120806, + 0.012668837793171406, + 0.011996897868812084, + 0.017125733196735382, + 0.0029011107981204987, + 0.0030994429253041744, + 0.0037185384426265955, + 0.004140505101531744, + 0.003010559594258666, + 0.0034534209407866, + 0.004119740799069405, + 0.003760203719139099, + 0.010806074365973473, + 0.008758644573390484, + 0.01302773505449295, + 0.018147306516766548, + 0.006811118219047785, + 0.009812096133828163, + 0.014123900793492794, + 0.013640872202813625, + 0.014175395481288433, + 0.015830963850021362, + 0.011939521878957748, + 0.021234851330518723, + 0.02191978693008423, + 0.012601413764059544, + 0.0246589295566082, + 0.015955206006765366, + 0.010448851622641087, + 0.01253323070704937, + 0.007717127911746502, + 0.00988945085555315, + 0.00847087986767292, + 0.011277807876467705, + 0.009722959250211716, + 0.013001779094338417, + 0.011280637234449387, + 0.01176291611045599, + 0.006741255056113005, + 0.010234033688902855, + 0.006854139268398285, + 0.008683260530233383, + 0.008477266877889633, + 0.011073066852986813, + 0.012120703235268593, + 0.009307275526225567, + 0.010623959824442863, + 0.013950767926871777, + 0.015694238245487213, + 0.020277881994843483, + 0.016039974987506866, + 0.015573943965137005, + 0.0197743009775877, + 0.021725287660956383, + 0.015920421108603477, + 0.011297518387436867, + 0.03276930749416351, + 0.018761664628982544, + 0.01513857115060091, + 0.01425060722976923, + 0.007102855481207371, + 0.0061259143985807896, + 0.005998630076646805, + 0.007743320893496275, + 0.006271078251302242, + 0.005024330224841833, + 0.005315013229846954, + 0.004895460791885853, + 0.03983045369386673, + 0.028809448704123497, + 0.01535513810813427, + 0.016642656177282333, + 0.018650008365511894, + 0.02242126874625683, + 0.014453424140810966, + 0.021440744400024414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 3 + ], + "selected_reward": -0.011084645055234432 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the vast ceiling of an abandoned factory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a complex assembly of interlocking cast-iron gears and thick roller chains coated in dark lubricant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the machinery shear from a rusted mounting bracket and slide along a large, grimy skylight panel before tumbling over the edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a steady medium shot where the camera remains stationary throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style desaturated and gritty, resembling the look of 1990s cinema?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42888343334198 + ], + "opposite_logprob_tokens": [ + -1.05388343334198 + ], + "logprob_reward": -0.42888343334198, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42888343334198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06252356618642807, + 0.023415114730596542, + 0.03712894394993782, + 0.3380005359649658, + 0.11322224885225296, + 0.06614255905151367, + 0.11435945332050323, + 0.05964028462767601, + 0.11446042358875275, + 0.07110689580440521 + ], + "credit_map": [ + 0.007169574964791536, + 0.006197781767696142, + 0.006094068288803101, + 0.009138718247413635, + 0.00556447496637702, + 0.005683773662894964, + 0.010621565394103527, + 0.012053602375090122, + 0.002629199530929327, + 0.0023120276164263487, + 0.0027031833305954933, + 0.002356651471927762, + 0.001978115178644657, + 0.003134530270472169, + 0.003551319008693099, + 0.004750086460262537, + 0.00428871251642704, + 0.0036237812601029873, + 0.0030669427942484617, + 0.005113613326102495, + 0.005563018843531609, + 0.003395162522792816, + 0.006425702944397926, + 0.005652004387229681, + 0.03363995626568794, + 0.032518744468688965, + 0.028546886518597603, + 0.03750082105398178, + 0.051234837621450424, + 0.04449204355478287, + 0.05166257172822952, + 0.058404695242643356, + 0.017012612894177437, + 0.00934643391519785, + 0.008703511208295822, + 0.008669338189065456, + 0.012692664749920368, + 0.017367973923683167, + 0.01008650939911604, + 0.029343195259571075, + 0.00833787303417921, + 0.009504874236881733, + 0.008275160565972328, + 0.008983825333416462, + 0.005713198333978653, + 0.006971112918108702, + 0.006618153769522905, + 0.01173835527151823, + 0.015549098141491413, + 0.01318043377250433, + 0.010664744302630424, + 0.01592315174639225, + 0.015546459704637527, + 0.01664484478533268, + 0.013226552866399288, + 0.013624165207147598, + 0.009892306290566921, + 0.006162792444229126, + 0.011520256288349628, + 0.006523825693875551, + 0.008803733624517918, + 0.0072284601628780365, + 0.004672996699810028, + 0.004835908766835928, + 0.01651284657418728, + 0.016674218699336052, + 0.021541466936469078, + 0.010625600814819336, + 0.013241932727396488, + 0.011094920337200165, + 0.00924819428473711, + 0.01552122924476862, + 0.018995897844433784, + 0.013052263297140598, + 0.005124525632709265, + 0.0043706041760742664, + 0.007477407809346914, + 0.004836937412619591, + 0.006182503886520863, + 0.011066748760640621 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 6 + ], + "selected_reward": -0.42888343334198 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the vast ceiling of an abandoned factory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a complex assembly of interlocking cast-iron gears and thick roller chains coated in dark lubricant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the machinery shear from a rusted mounting bracket and slide along a large, grimy skylight panel before tumbling over the edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a steady medium shot where the camera remains stationary throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style desaturated and gritty, resembling the look of 1990s cinema?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818905957043171 + ], + "opposite_logprob_tokens": [ + -4.018188953399658 + ], + "logprob_reward": -0.01818905957043171, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818905957043171, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1737837940454483, + 0.048383790999650955, + 0.06678876280784607, + 0.08320234715938568, + 0.15360502898693085, + 0.04762100428342819, + 0.0824192687869072, + 0.11469082534313202, + 0.07362529635429382, + 0.1558798998594284 + ], + "credit_map": [ + 0.017605489119887352, + 0.012729381211102009, + 0.018754437565803528, + 0.030735986307263374, + 0.016109494492411613, + 0.019435789436101913, + 0.031252771615982056, + 0.027160445228219032, + 0.005794069729745388, + 0.004764482844620943, + 0.004829525016248226, + 0.007605529855936766, + 0.004557740408927202, + 0.005224787164479494, + 0.007920365780591965, + 0.007687287870794535, + 0.0075204093009233475, + 0.007158438675105572, + 0.008877615444362164, + 0.011104132048785686, + 0.007248673588037491, + 0.0065697492100298405, + 0.010005452670156956, + 0.008304291404783726, + 0.007932877168059349, + 0.006232443731278181, + 0.007655731867998838, + 0.013551944866776466, + 0.008944074623286724, + 0.009033150970935822, + 0.013863599859178066, + 0.01598852500319481, + 0.023628905415534973, + 0.016674408689141273, + 0.014516095630824566, + 0.013233140110969543, + 0.020074494183063507, + 0.020511524751782417, + 0.015296724624931812, + 0.029669731855392456, + 0.007355305831879377, + 0.006173033267259598, + 0.006258565466850996, + 0.005036190152168274, + 0.00659289350733161, + 0.005312157794833183, + 0.005470696836709976, + 0.0054221609607338905, + 0.009549249894917011, + 0.008999908342957497, + 0.007715288549661636, + 0.011541178449988365, + 0.010837328620254993, + 0.012732346542179585, + 0.008085235022008419, + 0.012958738952875137, + 0.011071916669607162, + 0.012750997208058834, + 0.015040731057524681, + 0.011586896143853664, + 0.021065879613161087, + 0.017252562567591667, + 0.011875806376338005, + 0.01404604222625494, + 0.011939561925828457, + 0.011297191493213177, + 0.014128840528428555, + 0.005842631682753563, + 0.012358210980892181, + 0.005404955707490444, + 0.00512526324018836, + 0.007528642192482948, + 0.035752423107624054, + 0.02690728008747101, + 0.013200357556343079, + 0.012560787610709667, + 0.017998212948441505, + 0.013423779048025608, + 0.01043655350804329, + 0.025600502267479897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.01818905957043171 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Adopting the desaturated, gritty look of 1990s cinema, a steady medium shot captures the moment of industrial disaster with chilling clarity. From the vast ceiling of an abandoned factory, a complex, greasy mechanism of gears and chains violently breaks loose. The assembly, a tangled web of interlocking cast-iron gears and thick roller chains, coated in a dark, sticky lubricant, shears from its rusted mounting bracket in a puff of metallic dust. Under immense grinding force, the entire rusted assembly slides heavily along the angled surface of the large, grimy skylight panel, its immense weight causing the grime-caked glass to groan and shatter. The camera remains stationary, emphasizing the weight and danger as the jagged edges of the machinery scrape deep grooves into the grime before it finally tumbles over the edge of the steel frame, slowly spinning in the air for a terrifying moment before beginning its inexorable fall, creating a palpable atmosphere of impending impact and mechanical violence.", + "seed": 22743, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_9/index_8986_seed_22743_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.10035703349858523, + "vlm_logprob": -0.10035703349858523, + "vlm_yes_no_margin": 3.275000011920929, + "vlm_token_credit": -0.10035703349858523, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the vast ceiling of an abandoned factory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a complex assembly of interlocking cast-iron gears and thick roller chains coated in dark lubricant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the machinery shear from a rusted mounting bracket and slide along a large, grimy skylight panel before tumbling over the edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a steady medium shot where the camera remains stationary throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style desaturated and gritty, resembling the look of 1990s cinema?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11142901331186295 + ], + "opposite_logprob_tokens": [ + -2.48642897605896 + ], + "logprob_reward": -0.11142901331186295, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.11142901331186295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2395276576280594, + 0.03614986687898636, + 0.03644797205924988, + 0.06748086214065552, + 0.2408062070608139, + 0.11846398562192917, + 0.1156480461359024, + 0.08375074714422226, + 0.0423520989716053, + 0.019372625276446342 + ], + "credit_map": [ + 0.02229049801826477, + 0.016405871137976646, + 0.014593886211514473, + 0.023904459550976753, + 0.015959614887833595, + 0.020096806809306145, + 0.0726931020617485, + 0.05358339473605156, + 0.003998212516307831, + 0.002363467589020729, + 0.005607535596936941, + 0.0057077263481915, + 0.005258904304355383, + 0.0032296113204210997, + 0.0051905312575399876, + 0.004793878644704819, + 0.005695217289030552, + 0.0038332806434482336, + 0.0036066018510609865, + 0.004364972934126854, + 0.004603077191859484, + 0.003709719516336918, + 0.006571110803633928, + 0.004063992295414209, + 0.010776517912745476, + 0.007469636853784323, + 0.007311106193810701, + 0.006228823680430651, + 0.008309407159686089, + 0.006517530418932438, + 0.008182300254702568, + 0.012685539200901985, + 0.025128284469246864, + 0.037611398845911026, + 0.03392413631081581, + 0.02871430665254593, + 0.03885844722390175, + 0.019806690514087677, + 0.022961478680372238, + 0.03380145877599716, + 0.011335496790707111, + 0.01192692294716835, + 0.009605731815099716, + 0.009306429885327816, + 0.022169211879372597, + 0.014400728046894073, + 0.022071607410907745, + 0.017647862434387207, + 0.014938822947442532, + 0.008600758388638496, + 0.011262397281825542, + 0.006793718319386244, + 0.007175305858254433, + 0.010799111798405647, + 0.04219583421945572, + 0.013882092200219631, + 0.008877167478203773, + 0.007098210509866476, + 0.015989534556865692, + 0.013016575016081333, + 0.0087870042771101, + 0.0060715931467711926, + 0.011817152611911297, + 0.012093511410057545, + 0.004944028332829475, + 0.006273144856095314, + 0.006890599150210619, + 0.004374858923256397, + 0.005532998591661453, + 0.005826916545629501, + 0.003927056211978197, + 0.004582492634654045, + 0.0017907421570271254, + 0.0019094324670732021, + 0.003674635896459222, + 0.002671385183930397, + 0.0016918993787840009, + 0.002282720059156418, + 0.0032776272855699062, + 0.0020741831976920366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.11142901331186295 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the vast ceiling of an abandoned factory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a complex assembly of interlocking cast-iron gears and thick roller chains coated in dark lubricant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the machinery shear from a rusted mounting bracket and slide along a large, grimy skylight panel before tumbling over the edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a steady medium shot where the camera remains stationary throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style desaturated and gritty, resembling the look of 1990s cinema?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008674322627484798 + ], + "opposite_logprob_tokens": [ + -4.758674144744873 + ], + "logprob_reward": -0.008674322627484798, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008674322627484798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34169650077819824, + 0.049539078027009964, + 0.053680919110774994, + 0.11234596371650696, + 0.08993299305438995, + 0.11519312858581543, + 0.10861213505268097, + 0.08506659418344498, + 0.024052055552601814, + 0.01988064870238304 + ], + "credit_map": [ + 0.05057777091860771, + 0.016819681972265244, + 0.018786432221531868, + 0.02828470803797245, + 0.022931993007659912, + 0.026200734078884125, + 0.11393788456916809, + 0.06415730714797974, + 0.004885285627096891, + 0.0036108065396547318, + 0.007407360710203648, + 0.006389293819665909, + 0.006343022920191288, + 0.0045727151446044445, + 0.00854512769728899, + 0.007785463705658913, + 0.007270475849509239, + 0.002936924109235406, + 0.0042037200182676315, + 0.004049299750477076, + 0.004103910643607378, + 0.005362179130315781, + 0.015738509595394135, + 0.010015900246798992, + 0.014723862521350384, + 0.013232049532234669, + 0.015168999321758747, + 0.011689811013638973, + 0.01484775822609663, + 0.01100605446845293, + 0.011557154357433319, + 0.020120279863476753, + 0.010146277956664562, + 0.013508057221770287, + 0.012669370509684086, + 0.009970132261514664, + 0.010513288900256157, + 0.009734129533171654, + 0.009441901929676533, + 0.013949836604297161, + 0.010212205350399017, + 0.011859889142215252, + 0.009101483970880508, + 0.010173073969781399, + 0.02534235082566738, + 0.012682237662374973, + 0.01950584538280964, + 0.016316037625074387, + 0.01827811636030674, + 0.008333221077919006, + 0.009104065597057343, + 0.006586930714547634, + 0.007577604614198208, + 0.010016304440796375, + 0.03659067675471306, + 0.012125218287110329, + 0.008051755838096142, + 0.008450636640191078, + 0.01205823477357626, + 0.006934402510523796, + 0.00994444452226162, + 0.00868119765073061, + 0.017145680263638496, + 0.0138002410531044, + 0.002429173095151782, + 0.0026884949766099453, + 0.0027646636590361595, + 0.0038239911664277315, + 0.0029627156909555197, + 0.0024844766594469547, + 0.0034172467421740294, + 0.003481294959783554, + 0.001964143943041563, + 0.0025776957627385855, + 0.0026630915235728025, + 0.0023911476600915194, + 0.0027422942221164703, + 0.0023597143590450287, + 0.0023085824213922024, + 0.0028739783447235823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.008674322627484798 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the vast ceiling of an abandoned factory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a complex assembly of interlocking cast-iron gears and thick roller chains coated in dark lubricant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the machinery shear from a rusted mounting bracket and slide along a large, grimy skylight panel before tumbling over the edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a steady medium shot where the camera remains stationary throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style desaturated and gritty, resembling the look of 1990s cinema?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023299487307667732 + ], + "opposite_logprob_tokens": [ + -3.7732994556427 + ], + "logprob_reward": -0.023299487307667732, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023299487307667732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2994065582752228, + 0.026918552815914154, + 0.03684800863265991, + 0.05632905662059784, + 0.08204110711812973, + 0.25715702772140503, + 0.0584305115044117, + 0.09013666957616806, + 0.06857046484947205, + 0.02416202798485756 + ], + "credit_map": [ + 0.05305697023868561, + 0.013974172994494438, + 0.01967540942132473, + 0.02520524337887764, + 0.02022891491651535, + 0.021203499287366867, + 0.10185588896274567, + 0.04420647770166397, + 0.004374715033918619, + 0.002146301791071892, + 0.004544188268482685, + 0.0029333257116377354, + 0.0022285713348537683, + 0.002988636028021574, + 0.0030839177779853344, + 0.004618895705789328, + 0.006026448216289282, + 0.002520247595384717, + 0.004185139201581478, + 0.0030988333746790886, + 0.003156225197017193, + 0.0032143602147698402, + 0.005936030298471451, + 0.008710727095603943, + 0.007332182954996824, + 0.00674973102286458, + 0.007103880401700735, + 0.006876317784190178, + 0.00837085023522377, + 0.006805637385696173, + 0.005715584848076105, + 0.007374870125204325, + 0.006901104003190994, + 0.009387610480189323, + 0.015025191940367222, + 0.00908940751105547, + 0.010473526082932949, + 0.0061574215069413185, + 0.0165096465498209, + 0.008497197180986404, + 0.02423563413321972, + 0.024941949173808098, + 0.01911637745797634, + 0.022115642204880714, + 0.05127028748393059, + 0.029698628932237625, + 0.047139886766672134, + 0.0386386401951313, + 0.006277892738580704, + 0.006842930801212788, + 0.00592981930822134, + 0.0034781789872795343, + 0.005277639254927635, + 0.007367510348558426, + 0.017505090683698654, + 0.0057514444924890995, + 0.005137616768479347, + 0.010080023668706417, + 0.012281002476811409, + 0.010004017502069473, + 0.015213937498629093, + 0.012956901453435421, + 0.013998395763337612, + 0.01046477910131216, + 0.007762722205370665, + 0.009756427258253098, + 0.010442942380905151, + 0.007234977092593908, + 0.009467259980738163, + 0.010540507733821869, + 0.006320984102785587, + 0.007044646423310041, + 0.0012988627422600985, + 0.0033378645312041044, + 0.001758280210196972, + 0.0022564867977052927, + 0.004750133957713842, + 0.004369562026113272, + 0.0029163279104977846, + 0.0034745100419968367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.023299487307667732 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the vast ceiling of an abandoned factory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a complex assembly of interlocking cast-iron gears and thick roller chains coated in dark lubricant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the machinery shear from a rusted mounting bracket and slide along a large, grimy skylight panel before tumbling over the edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a steady medium shot where the camera remains stationary throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style desaturated and gritty, resembling the look of 1990s cinema?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486022353172302 + ], + "opposite_logprob_tokens": [ + -1.223602294921875 + ], + "logprob_reward": -0.3486022353172302, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.3486022353172302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3091038167476654, + 0.034842539578676224, + 0.08582060784101486, + 0.11252228915691376, + 0.037722449749708176, + 0.07178019732236862, + 0.2170393019914627, + 0.07127077877521515, + 0.04254956170916557, + 0.017348479479551315 + ], + "credit_map": [ + 0.045837871730327606, + 0.01464530173689127, + 0.016969500109553337, + 0.02503245137631893, + 0.020167449489235878, + 0.022606559097766876, + 0.10726070404052734, + 0.05658397078514099, + 0.00378264463506639, + 0.0024424022994935513, + 0.006474250927567482, + 0.004182292148470879, + 0.0028495974838733673, + 0.004818949848413467, + 0.004974749870598316, + 0.005317653529345989, + 0.008676337078213692, + 0.005397813394665718, + 0.005034556612372398, + 0.006051446311175823, + 0.005875976290553808, + 0.008656401187181473, + 0.02303077094256878, + 0.023097306489944458, + 0.016427911818027496, + 0.012990213930606842, + 0.015490245074033737, + 0.01144565362483263, + 0.016121968626976013, + 0.013320401310920715, + 0.011251872405409813, + 0.015474018640816212, + 0.0032386956736445427, + 0.006077330559492111, + 0.006099744699895382, + 0.0038797983434051275, + 0.00828922912478447, + 0.003087664255872369, + 0.003319748677313328, + 0.003730237251147628, + 0.006499933544546366, + 0.007047498598694801, + 0.009074137546122074, + 0.009585591964423656, + 0.008873272687196732, + 0.011232811026275158, + 0.010370039381086826, + 0.009096913039684296, + 0.021659651771187782, + 0.02039886824786663, + 0.023190636187791824, + 0.01311343815177679, + 0.01639186218380928, + 0.025230104103684425, + 0.07328929752111435, + 0.0237654447555542, + 0.0060034580528736115, + 0.006489213090389967, + 0.007755033206194639, + 0.006932631600648165, + 0.008599827066063881, + 0.007630371954292059, + 0.01656240224838257, + 0.011297840625047684, + 0.0036105618346482515, + 0.005725296214222908, + 0.008696412667632103, + 0.005127153359353542, + 0.005406034644693136, + 0.0057756067253649235, + 0.00392159866169095, + 0.004286895040422678, + 0.001780663849785924, + 0.0023781622294336557, + 0.002437603659927845, + 0.002254095859825611, + 0.002473606029525399, + 0.002524075098335743, + 0.0017057855147868395, + 0.001794487820006907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.3486022353172302 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the vast ceiling of an abandoned factory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a complex assembly of interlocking cast-iron gears and thick roller chains coated in dark lubricant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the machinery shear from a rusted mounting bracket and slide along a large, grimy skylight panel before tumbling over the edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a steady medium shot where the camera remains stationary throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style desaturated and gritty, resembling the look of 1990s cinema?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978010892868042 + ], + "opposite_logprob_tokens": [ + -4.634779930114746 + ], + "logprob_reward": -0.00978010892868042, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978010892868042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13813932240009308, + 0.032154303044080734, + 0.02665870450437069, + 0.035219814628362656, + 0.04937770590186119, + 0.0685945525765419, + 0.27809134125709534, + 0.17817248404026031, + 0.11013306677341461, + 0.08345865458250046 + ], + "credit_map": [ + 0.030278243124485016, + 0.007285310886800289, + 0.012841897085309029, + 0.020054621621966362, + 0.01122637465596199, + 0.015018019825220108, + 0.021673288196325302, + 0.01976155862212181, + 0.004254708997905254, + 0.0029395606834441423, + 0.005337453912943602, + 0.0030086080078035593, + 0.003816501470282674, + 0.003283890662714839, + 0.004133270122110844, + 0.005380307324230671, + 0.0024771224707365036, + 0.0023985847365111113, + 0.002547904849052429, + 0.0035921994131058455, + 0.003862033598124981, + 0.0034691502805799246, + 0.004005607683211565, + 0.004306101705878973, + 0.004769686609506607, + 0.003755758050829172, + 0.004472403321415186, + 0.004672465845942497, + 0.004658627789467573, + 0.004193825647234917, + 0.004028869792819023, + 0.004668178502470255, + 0.004334093537181616, + 0.0074852751567959785, + 0.005723833106458187, + 0.006009915843605995, + 0.0073339082300662994, + 0.005309341941028833, + 0.0066618467681109905, + 0.006519494578242302, + 0.006179343443363905, + 0.005135578569024801, + 0.006210354622453451, + 0.012930458411574364, + 0.007411569822579622, + 0.008603238500654697, + 0.01371600478887558, + 0.008408005349338055, + 0.03160044550895691, + 0.02163529396057129, + 0.029374567791819572, + 0.016957418993115425, + 0.019077451899647713, + 0.027838172391057014, + 0.09872303158044815, + 0.03288496658205986, + 0.013166023418307304, + 0.020029010251164436, + 0.021342825144529343, + 0.018118588253855705, + 0.027603911235928535, + 0.025191832333803177, + 0.026545973494648933, + 0.02617432177066803, + 0.009245472960174084, + 0.014484471641480923, + 0.014484144747257233, + 0.012311572209000587, + 0.017812101170420647, + 0.019988209009170532, + 0.010021289810538292, + 0.011785803362727165, + 0.005861677695065737, + 0.009831245057284832, + 0.00902376975864172, + 0.010933219455182552, + 0.014434462413191795, + 0.01565035805106163, + 0.0079615144059062, + 0.00976240448653698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.00978010892868042 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Adopting the desaturated, gritty look of 1990s cinema, a steady medium shot captures the moment of industrial disaster with chilling clarity. From the vast ceiling of an abandoned factory, a complex, greasy mechanism of gears and chains violently breaks loose. The assembly, a tangled web of interlocking cast-iron gears and thick roller chains, coated in a dark, sticky lubricant, shears from its rusted mounting bracket in a puff of metallic dust. Under immense grinding force, the entire rusted assembly slides heavily along the angled surface of the large, grimy skylight panel, its immense weight causing the grime-caked glass to groan and shatter. The camera remains stationary, emphasizing the weight and danger as the jagged edges of the machinery scrape deep grooves into the grime before it finally tumbles over the edge of the steel frame, slowly spinning in the air for a terrifying moment before beginning its inexorable fall, creating a palpable atmosphere of impending impact and mechanical violence.", + "seed": 22744, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_9/index_8986_seed_22744_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.051006811857223514, + "vlm_logprob": -0.051006811857223514, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.051006811857223514, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the vast ceiling of an abandoned factory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a complex assembly of interlocking cast-iron gears and thick roller chains coated in dark lubricant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the machinery shear from a rusted mounting bracket and slide along a large, grimy skylight panel before tumbling over the edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a steady medium shot where the camera remains stationary throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style desaturated and gritty, resembling the look of 1990s cinema?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09267295897006989 + ], + "opposite_logprob_tokens": [ + -2.5926730632781982 + ], + "logprob_reward": -0.09267295897006989, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.09267295897006989, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1364186853170395, + 0.052958421409130096, + 0.12169227004051208, + 0.12432236224412918, + 0.02867838181555271, + 0.08385831862688065, + 0.12996426224708557, + 0.21251164376735687, + 0.07695570588111877, + 0.032639987766742706 + ], + "credit_map": [ + 0.012022779323160648, + 0.017022786661982536, + 0.022877013310790062, + 0.025603653863072395, + 0.014678404666483402, + 0.013282576575875282, + 0.016442179679870605, + 0.014489292167127132, + 0.0049420795403420925, + 0.005947910249233246, + 0.007415281143039465, + 0.009438847191631794, + 0.005561358295381069, + 0.00467294454574585, + 0.006688366644084454, + 0.008291631005704403, + 0.009540800005197525, + 0.006483397912234068, + 0.009629134088754654, + 0.024615049362182617, + 0.006003980524837971, + 0.02269865944981575, + 0.008862024173140526, + 0.033859219402074814, + 0.011391331441700459, + 0.01454020757228136, + 0.01133601926267147, + 0.020756276324391365, + 0.02126527577638626, + 0.014500364661216736, + 0.017460815608501434, + 0.013072081841528416, + 0.003231715876609087, + 0.0020670995581895113, + 0.005353747401386499, + 0.00705370306968689, + 0.0018921929877251387, + 0.0016109785065054893, + 0.005094653461128473, + 0.0023742883931845427, + 0.007688905578106642, + 0.006801756098866463, + 0.010991297662258148, + 0.011881492100656033, + 0.011647295206785202, + 0.006228944752365351, + 0.015495575033128262, + 0.013123057782649994, + 0.011997648514807224, + 0.010888637974858284, + 0.018888264894485474, + 0.026655320078134537, + 0.009880444034934044, + 0.009314077906310558, + 0.024528639391064644, + 0.01781122386455536, + 0.01309049129486084, + 0.007945814169943333, + 0.020585065707564354, + 0.030751293525099754, + 0.008051521144807339, + 0.013215774670243263, + 0.06923254579305649, + 0.04963913559913635, + 0.009111897088587284, + 0.005728005897253752, + 0.013993993401527405, + 0.016792623326182365, + 0.009221819229424, + 0.005331367254257202, + 0.008989227935671806, + 0.007786779198795557, + 0.0016067437827587128, + 0.0013595462078228593, + 0.006594538688659668, + 0.008805911988019943, + 0.0022943627554923296, + 0.001831072848290205, + 0.006541110575199127, + 0.003606697777286172 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.09267295897006989 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the vast ceiling of an abandoned factory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a complex assembly of interlocking cast-iron gears and thick roller chains coated in dark lubricant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the machinery shear from a rusted mounting bracket and slide along a large, grimy skylight panel before tumbling over the edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a steady medium shot where the camera remains stationary throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style desaturated and gritty, resembling the look of 1990s cinema?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03374950587749481 + ], + "opposite_logprob_tokens": [ + -3.408749580383301 + ], + "logprob_reward": -0.03374950587749481, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03374950587749481, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2607058584690094, + 0.09601166844367981, + 0.09013387560844421, + 0.07806391268968582, + 0.10086353868246078, + 0.0379251129925251, + 0.10738617181777954, + 0.11180675029754639, + 0.061595212668180466, + 0.05550789833068848 + ], + "credit_map": [ + 0.02190260961651802, + 0.0383150689303875, + 0.04585275053977966, + 0.06005680933594704, + 0.01690441370010376, + 0.023237518966197968, + 0.03007400780916214, + 0.02436269447207451, + 0.008411606773734093, + 0.007023858837783337, + 0.018648164346814156, + 0.022684922441840172, + 0.00622344808652997, + 0.009439044632017612, + 0.009682326577603817, + 0.013898296281695366, + 0.006887937895953655, + 0.0060045053251087666, + 0.012692813761532307, + 0.017172526568174362, + 0.017127197235822678, + 0.007705546449869871, + 0.013670837506651878, + 0.008872516453266144, + 0.00689860712736845, + 0.005394032225012779, + 0.014681266620755196, + 0.02180921658873558, + 0.0052701798267662525, + 0.004327920265495777, + 0.012071525678038597, + 0.007611165754497051, + 0.009657317772507668, + 0.006707296706736088, + 0.01775374636054039, + 0.02424636483192444, + 0.007602106314152479, + 0.0066735208965837955, + 0.02117788977921009, + 0.00704528484493494, + 0.002718234434723854, + 0.004125650506466627, + 0.006294145714491606, + 0.00638204999268055, + 0.003827138105407357, + 0.0034397547133266926, + 0.006913480814546347, + 0.004224658012390137, + 0.010255323722958565, + 0.008091764524579048, + 0.014190555550158024, + 0.024082263931632042, + 0.00815720483660698, + 0.006931177340447903, + 0.02357262372970581, + 0.012105255387723446, + 0.011734501458704472, + 0.00743078114464879, + 0.023323388770222664, + 0.03290313482284546, + 0.00874103419482708, + 0.00674174539744854, + 0.012327461503446102, + 0.008604710921645164, + 0.004109530709683895, + 0.003687717020511627, + 0.00883588008582592, + 0.018896328285336494, + 0.00418767798691988, + 0.003992654383182526, + 0.010198901407420635, + 0.007686526048928499, + 0.004720448050647974, + 0.003945726901292801, + 0.012336207553744316, + 0.012848901562392712, + 0.0031839811708778143, + 0.003877563402056694, + 0.006380260456353426, + 0.008214809000492096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.03374950587749481 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the vast ceiling of an abandoned factory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a complex assembly of interlocking cast-iron gears and thick roller chains coated in dark lubricant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the machinery shear from a rusted mounting bracket and slide along a large, grimy skylight panel before tumbling over the edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a steady medium shot where the camera remains stationary throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style desaturated and gritty, resembling the look of 1990s cinema?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636094018816948 + ], + "opposite_logprob_tokens": [ + -3.6513609886169434 + ], + "logprob_reward": -0.02636094018816948, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636094018816948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10655193030834198, + 0.06925071775913239, + 0.1298968344926834, + 0.04312090948224068, + 0.09192556887865067, + 0.03961770236492157, + 0.14850588142871857, + 0.21453534066677094, + 0.08362226188182831, + 0.0729728639125824 + ], + "credit_map": [ + 0.00917167030274868, + 0.013383699581027031, + 0.020208386704325676, + 0.01721205748617649, + 0.007743591908365488, + 0.011155025102198124, + 0.016735071316361427, + 0.010942420922219753, + 0.006565747782588005, + 0.00423816591501236, + 0.014076480641961098, + 0.012419654056429863, + 0.005023440811783075, + 0.0072352951392531395, + 0.011065619066357613, + 0.008626312017440796, + 0.009814065881073475, + 0.005922878626734018, + 0.010049372911453247, + 0.024985766038298607, + 0.005395613145083189, + 0.025089632719755173, + 0.010488354600965977, + 0.038151148706674576, + 0.0031572673469781876, + 0.003087272634729743, + 0.0057726772502064705, + 0.0071938359178602695, + 0.004811247810721397, + 0.005290615372359753, + 0.007465361151844263, + 0.006342633161693811, + 0.010053794831037521, + 0.007526956964284182, + 0.013760370202362537, + 0.0172142181545496, + 0.008963556960225105, + 0.010248892940580845, + 0.015533088706433773, + 0.00862469244748354, + 0.00478648254647851, + 0.0028050376567989588, + 0.006694200448691845, + 0.006361121311783791, + 0.003730371128767729, + 0.004401858896017075, + 0.005903959274291992, + 0.0049346694722771645, + 0.01268320344388485, + 0.007978568784892559, + 0.02178049460053444, + 0.032796189188957214, + 0.008832424879074097, + 0.006249565631151199, + 0.035951025784015656, + 0.02223440632224083, + 0.01577765867114067, + 0.007979296147823334, + 0.04773305729031563, + 0.04267231002449989, + 0.01002360600978136, + 0.009907164610922337, + 0.05643187090754509, + 0.024010371416807175, + 0.00821428932249546, + 0.0038749058730900288, + 0.017568493261933327, + 0.020908452570438385, + 0.0055447896011173725, + 0.004535182379186153, + 0.015091661363840103, + 0.007884488441050053, + 0.003935257438570261, + 0.0032703527249395847, + 0.008371714502573013, + 0.00825544074177742, + 0.0035511243622750044, + 0.0043967911042273045, + 0.024487795308232307, + 0.016704386100172997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 2 + ], + "selected_reward": -0.02636094018816948 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the vast ceiling of an abandoned factory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a complex assembly of interlocking cast-iron gears and thick roller chains coated in dark lubricant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the machinery shear from a rusted mounting bracket and slide along a large, grimy skylight panel before tumbling over the edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a steady medium shot where the camera remains stationary throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style desaturated and gritty, resembling the look of 1990s cinema?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.078966423869133 + ], + "opposite_logprob_tokens": [ + -2.5789663791656494 + ], + "logprob_reward": -0.078966423869133, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.078966423869133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07551253587007523, + 0.06268572062253952, + 0.04038480669260025, + 0.10412467271089554, + 0.07629211246967316, + 0.13231995701789856, + 0.2807750403881073, + 0.15042123198509216, + 0.04122510552406311, + 0.036258865147829056 + ], + "credit_map": [ + 0.005749737843871117, + 0.009642107412219048, + 0.012643895111978054, + 0.012813560664653778, + 0.009562946856021881, + 0.005493102129548788, + 0.012641197070479393, + 0.006965989712625742, + 0.0054168300703167915, + 0.005014899652451277, + 0.012196392752230167, + 0.013322456739842892, + 0.009518512524664402, + 0.0029178825207054615, + 0.008846822194755077, + 0.005451921373605728, + 0.005158443469554186, + 0.004088550806045532, + 0.005455056671053171, + 0.00611122278496623, + 0.007193322759121656, + 0.004586236085742712, + 0.004096324555575848, + 0.0036956474650651217, + 0.01307910680770874, + 0.0052780709229409695, + 0.012169155292212963, + 0.014772373251616955, + 0.01290963962674141, + 0.01577986776828766, + 0.013752161525189877, + 0.016384292393922806, + 0.005209889728575945, + 0.0038158053066581488, + 0.014256957918405533, + 0.022079870104789734, + 0.005501095205545425, + 0.005989781115204096, + 0.012031948193907738, + 0.007406755816191435, + 0.010469087399542332, + 0.009462689980864525, + 0.022975653409957886, + 0.02031605876982212, + 0.012754506431519985, + 0.01123872771859169, + 0.02410874329507351, + 0.02099447511136532, + 0.028485476970672607, + 0.019479693844914436, + 0.06207401305437088, + 0.06087064743041992, + 0.017964519560337067, + 0.01686454378068447, + 0.0468684658408165, + 0.028167668730020523, + 0.011001445353031158, + 0.007057195529341698, + 0.02016463316977024, + 0.02677331678569317, + 0.008811737410724163, + 0.008615458384156227, + 0.041344206780195236, + 0.026653237640857697, + 0.003672239137813449, + 0.0022653068881481886, + 0.008107057772576809, + 0.010716342367231846, + 0.0025663264095783234, + 0.002936525968834758, + 0.00537415174767375, + 0.005587148945778608, + 0.002487967722117901, + 0.001900080475024879, + 0.004003706853836775, + 0.004556442145258188, + 0.001957847736775875, + 0.002181641524657607, + 0.01024810504168272, + 0.00892307236790657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.078966423869133 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the vast ceiling of an abandoned factory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a complex assembly of interlocking cast-iron gears and thick roller chains coated in dark lubricant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the machinery shear from a rusted mounting bracket and slide along a large, grimy skylight panel before tumbling over the edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a steady medium shot where the camera remains stationary throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style desaturated and gritty, resembling the look of 1990s cinema?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328423038125038 + ], + "opposite_logprob_tokens": [ + -3.7732841968536377 + ], + "logprob_reward": -0.02328423038125038, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328423038125038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19388747215270996, + 0.050306785851716995, + 0.03437548503279686, + 0.0398729145526886, + 0.031948041170835495, + 0.053536899387836456, + 0.1735602468252182, + 0.2598852515220642, + 0.06417949497699738, + 0.09844743460416794 + ], + "credit_map": [ + 0.019597506150603294, + 0.021075116470456123, + 0.037609271705150604, + 0.030035145580768585, + 0.03592713922262192, + 0.012778029777109623, + 0.016791606321930885, + 0.020073650404810905, + 0.005672682076692581, + 0.004312737379223108, + 0.010057596489787102, + 0.00677477428689599, + 0.006981946527957916, + 0.00369204836897552, + 0.006940937135368586, + 0.005874063353985548, + 0.003640953917056322, + 0.0018346054712310433, + 0.004851873032748699, + 0.0063875587657094, + 0.0029591398779302835, + 0.003962232731282711, + 0.0036613685078918934, + 0.00707775354385376, + 0.0032995669171214104, + 0.0036739876959472895, + 0.004561415407806635, + 0.010112485848367214, + 0.006071959622204304, + 0.004203790333122015, + 0.004919445142149925, + 0.003030266147106886, + 0.0030754003673791885, + 0.0020872517488896847, + 0.006249912548810244, + 0.006389175541698933, + 0.0022433593403548002, + 0.004021286498755217, + 0.004600866232067347, + 0.003280787728726864, + 0.0035778931342065334, + 0.00473566772416234, + 0.0056951274164021015, + 0.012342344038188457, + 0.007247793022543192, + 0.004552963189780712, + 0.009522982873022556, + 0.005862129386514425, + 0.016895323991775513, + 0.009988207370042801, + 0.024876996874809265, + 0.03897416964173317, + 0.01136299129575491, + 0.008657827973365784, + 0.039821017533540726, + 0.022983694449067116, + 0.020468175411224365, + 0.009823323227465153, + 0.04569282382726669, + 0.04571780189871788, + 0.011544008739292622, + 0.012888121418654919, + 0.07449246942996979, + 0.039258528500795364, + 0.007169097196310759, + 0.0038376986049115658, + 0.013099145144224167, + 0.015366451814770699, + 0.004123196471482515, + 0.003871372202411294, + 0.008131110109388828, + 0.008581424131989479, + 0.006315240170806646, + 0.004201587755233049, + 0.015593662858009338, + 0.01512511819601059, + 0.005275482311844826, + 0.005521885119378567, + 0.027760837227106094, + 0.018653612583875656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.02328423038125038 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Adopting the desaturated, gritty look of 1990s cinema, a steady medium shot captures the moment of industrial disaster with chilling clarity. From the vast ceiling of an abandoned factory, a complex, greasy mechanism of gears and chains violently breaks loose. The assembly, a tangled web of interlocking cast-iron gears and thick roller chains, coated in a dark, sticky lubricant, shears from its rusted mounting bracket in a puff of metallic dust. Under immense grinding force, the entire rusted assembly slides heavily along the angled surface of the large, grimy skylight panel, its immense weight causing the grime-caked glass to groan and shatter. The camera remains stationary, emphasizing the weight and danger as the jagged edges of the machinery scrape deep grooves into the grime before it finally tumbles over the edge of the steel frame, slowly spinning in the air for a terrifying moment before beginning its inexorable fall, creating a palpable atmosphere of impending impact and mechanical violence.", + "seed": 22745, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_9/index_8986_seed_22745_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06298492974601685, + "vlm_logprob": -0.06298492974601685, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.06298492974601685, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the vast ceiling of an abandoned factory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a complex assembly of interlocking cast-iron gears and thick roller chains coated in dark lubricant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the machinery shear from a rusted mounting bracket and slide along a large, grimy skylight panel before tumbling over the edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a steady medium shot where the camera remains stationary throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style desaturated and gritty, resembling the look of 1990s cinema?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.051755163818597794 + ], + "opposite_logprob_tokens": [ + -3.42675518989563 + ], + "logprob_reward": -0.051755163818597794, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.051755163818597794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1062307059764862, + 0.043902695178985596, + 0.15756407380104065, + 0.11601081490516663, + 0.05034669116139412, + 0.04463901370763779, + 0.039982955902814865, + 0.051435161381959915, + 0.050639498978853226, + 0.3392484188079834 + ], + "credit_map": [ + 0.006233463529497385, + 0.010275919921696186, + 0.021667158231139183, + 0.01910693198442459, + 0.013738786801695824, + 0.006731972098350525, + 0.016573265194892883, + 0.011903203092515469, + 0.004489972721785307, + 0.0031838989816606045, + 0.008614717982709408, + 0.006576736457645893, + 0.004736016504466534, + 0.0038967051077634096, + 0.007574028801172972, + 0.004830621182918549, + 0.010686592198908329, + 0.012952005490660667, + 0.024023305624723434, + 0.023067399859428406, + 0.01930422894656658, + 0.01777416095137596, + 0.02706499584019184, + 0.022691374644637108, + 0.010160301811993122, + 0.007750791031867266, + 0.02678784541785717, + 0.024978959932923317, + 0.007366296835243702, + 0.008454685099422932, + 0.016635751351714134, + 0.013876179233193398, + 0.004495859611779451, + 0.006290880963206291, + 0.009713600389659405, + 0.00960333552211523, + 0.003969929181039333, + 0.0042098527774214745, + 0.007013618014752865, + 0.005049614235758781, + 0.0033923648297786713, + 0.0032043226528912783, + 0.010066196322441101, + 0.010441451333463192, + 0.0023618778213858604, + 0.0038551699835807085, + 0.006879979744553566, + 0.004437650553882122, + 0.001940038986504078, + 0.004634491167962551, + 0.006930523552000523, + 0.007060233037918806, + 0.005700042471289635, + 0.002881578868255019, + 0.005936058238148689, + 0.004899988882243633, + 0.0038845750968903303, + 0.004050903487950563, + 0.012003691866993904, + 0.00673328060656786, + 0.004980480298399925, + 0.0032175746746361256, + 0.007439221255481243, + 0.009125432930886745, + 0.0038089260924607515, + 0.003590049920603633, + 0.007648508995771408, + 0.007200370077043772, + 0.009320487268269062, + 0.004270999692380428, + 0.007852627895772457, + 0.0069475253112614155, + 0.02378111705183983, + 0.020599042996764183, + 0.05289081484079361, + 0.056229326874017715, + 0.06424444913864136, + 0.022336546331644058, + 0.045458365231752396, + 0.053708743304014206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 3 + ], + "selected_reward": -0.051755163818597794 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the vast ceiling of an abandoned factory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a complex assembly of interlocking cast-iron gears and thick roller chains coated in dark lubricant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the machinery shear from a rusted mounting bracket and slide along a large, grimy skylight panel before tumbling over the edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a steady medium shot where the camera remains stationary throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style desaturated and gritty, resembling the look of 1990s cinema?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028406528290361166 + ], + "opposite_logprob_tokens": [ + -5.877840518951416 + ], + "logprob_reward": -0.0028406528290361166, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028406528290361166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14041344821453094, + 0.11301280558109283, + 0.03939878195524216, + 0.16246092319488525, + 0.08245490491390228, + 0.07968427985906601, + 0.09021221846342087, + 0.03356422483921051, + 0.09537369757890701, + 0.16342468559741974 + ], + "credit_map": [ + 0.011313943192362785, + 0.012272815220057964, + 0.02696133218705654, + 0.02813025191426277, + 0.013530587777495384, + 0.012572742998600006, + 0.021053127944469452, + 0.01457864511758089, + 0.0107556888833642, + 0.009685139171779156, + 0.019526613876223564, + 0.03594426438212395, + 0.008097206242382526, + 0.004575763363391161, + 0.009995193220674992, + 0.014432936906814575, + 0.002942498540505767, + 0.004213096108287573, + 0.00705077825114131, + 0.006916950922459364, + 0.0043460167944431305, + 0.002562472829595208, + 0.006261363159865141, + 0.005105608142912388, + 0.011338606476783752, + 0.014837951399385929, + 0.037561479955911636, + 0.029586318880319595, + 0.013075530529022217, + 0.014248299412429333, + 0.02078394405543804, + 0.0210287943482399, + 0.005245642736554146, + 0.007333079818636179, + 0.018331686034798622, + 0.016509385779500008, + 0.008795643225312233, + 0.005462388042360544, + 0.012574470601975918, + 0.008202611468732357, + 0.005797534249722958, + 0.00793399941176176, + 0.017370468005537987, + 0.016664225608110428, + 0.004816614557057619, + 0.004970434587448835, + 0.01353735476732254, + 0.00859364215284586, + 0.006880355067551136, + 0.011306718923151493, + 0.016215382143855095, + 0.017322391271591187, + 0.011161703616380692, + 0.0051994589157402515, + 0.01141304150223732, + 0.010713172145187855, + 0.002835706574842334, + 0.003125182818621397, + 0.006534055806696415, + 0.006418407894670963, + 0.003615214256569743, + 0.0021064819302409887, + 0.00461128493770957, + 0.004317891784012318, + 0.007686129771173, + 0.005342514254152775, + 0.017333047464489937, + 0.016459226608276367, + 0.015325915068387985, + 0.007425693329423666, + 0.015029076486825943, + 0.010772096924483776, + 0.008509544655680656, + 0.014005322940647602, + 0.02286573499441147, + 0.01850101351737976, + 0.045809056609869, + 0.008843746036291122, + 0.013493852689862251, + 0.03139641880989075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.0028406528290361166 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the vast ceiling of an abandoned factory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a complex assembly of interlocking cast-iron gears and thick roller chains coated in dark lubricant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the machinery shear from a rusted mounting bracket and slide along a large, grimy skylight panel before tumbling over the edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a steady medium shot where the camera remains stationary throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style desaturated and gritty, resembling the look of 1990s cinema?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014203409664332867 + ], + "opposite_logprob_tokens": [ + -4.2642035484313965 + ], + "logprob_reward": -0.014203409664332867, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014203409664332867, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14365524053573608, + 0.0938454419374466, + 0.11097695678472519, + 0.10902819782495499, + 0.03597713261842728, + 0.0956469401717186, + 0.02752538211643696, + 0.0655415952205658, + 0.19044074416160583, + 0.12736237049102783 + ], + "credit_map": [ + 0.009680653922259808, + 0.011506585404276848, + 0.022418105974793434, + 0.023197073489427567, + 0.027872469276189804, + 0.009566865861415863, + 0.023037012666463852, + 0.016376463696360588, + 0.008907217532396317, + 0.006040314212441444, + 0.01722881756722927, + 0.030014080926775932, + 0.007022653706371784, + 0.004792051389813423, + 0.008669035509228706, + 0.011171267367899418, + 0.008039116859436035, + 0.010181833989918232, + 0.013177762739360332, + 0.01373912114650011, + 0.023416828364133835, + 0.013013768009841442, + 0.01465595792979002, + 0.014752564951777458, + 0.009244218468666077, + 0.007122549694031477, + 0.024953028187155724, + 0.02319946326315403, + 0.009481054730713367, + 0.007651846390217543, + 0.01409038808196783, + 0.013285651803016663, + 0.002644899068400264, + 0.0031496852170675993, + 0.008418659679591656, + 0.0055719828233122826, + 0.003970008343458176, + 0.0024071927182376385, + 0.006002085283398628, + 0.0038126243744045496, + 0.005771332420408726, + 0.007316648494452238, + 0.020912598818540573, + 0.016132259741425514, + 0.007213082164525986, + 0.008796093054115772, + 0.014519087970256805, + 0.014985834248363972, + 0.0023947868030518293, + 0.003039218718186021, + 0.004734866786748171, + 0.00420118635520339, + 0.002654652576893568, + 0.0019397918367758393, + 0.004872946068644524, + 0.0036879314575344324, + 0.004539179615676403, + 0.005633344873785973, + 0.014597541652619839, + 0.011022980324923992, + 0.0071192169561982155, + 0.004033293575048447, + 0.009303811937570572, + 0.009292220696806908, + 0.014209943823516369, + 0.012068706564605236, + 0.02912139520049095, + 0.037484344094991684, + 0.02755817025899887, + 0.01222851313650608, + 0.029617635533213615, + 0.028152024373412132, + 0.0074424115009605885, + 0.008404508233070374, + 0.025436697527766228, + 0.0265886839479208, + 0.020891224965453148, + 0.007760594133287668, + 0.014339170418679714, + 0.016499077901244164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.014203409664332867 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the vast ceiling of an abandoned factory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a complex assembly of interlocking cast-iron gears and thick roller chains coated in dark lubricant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the machinery shear from a rusted mounting bracket and slide along a large, grimy skylight panel before tumbling over the edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a steady medium shot where the camera remains stationary throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style desaturated and gritty, resembling the look of 1990s cinema?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22554545104503632 + ], + "opposite_logprob_tokens": [ + -1.6005454063415527 + ], + "logprob_reward": -0.22554545104503632, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22554545104503632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08619622141122818, + 0.05871910601854324, + 0.049128614366054535, + 0.13517868518829346, + 0.10844359546899796, + 0.06997598707675934, + 0.03299196437001228, + 0.14023487269878387, + 0.1457545906305313, + 0.17337633669376373 + ], + "credit_map": [ + 0.007194581441581249, + 0.007730263285338879, + 0.016800707206130028, + 0.02145647257566452, + 0.0055810753256082535, + 0.006453228648751974, + 0.011541367508471012, + 0.009438530541956425, + 0.00560547411441803, + 0.006896358914673328, + 0.009254850447177887, + 0.007619411218911409, + 0.01040716003626585, + 0.004292311146855354, + 0.008809235878288746, + 0.005834304261952639, + 0.0031481306068599224, + 0.005757339298725128, + 0.008818401023745537, + 0.010259570553898811, + 0.005050900392234325, + 0.004197451751679182, + 0.006671357899904251, + 0.005225461442023516, + 0.00988768134266138, + 0.014510172419250011, + 0.027641674503684044, + 0.023906003683805466, + 0.015434644185006618, + 0.01167075801640749, + 0.016927054151892662, + 0.01520069595426321, + 0.007629786152392626, + 0.010276580229401588, + 0.0205739326775074, + 0.022286538034677505, + 0.009408907033503056, + 0.006958228535950184, + 0.019215648993849754, + 0.012093978002667427, + 0.005580043885856867, + 0.007468903437256813, + 0.012882840819656849, + 0.011791756376624107, + 0.005924860015511513, + 0.006872440222650766, + 0.012395945377647877, + 0.007059198804199696, + 0.002376012969762087, + 0.003026906866580248, + 0.006070333998650312, + 0.005613698158413172, + 0.003711969591677189, + 0.0020145857706665993, + 0.005904793739318848, + 0.004273661877959967, + 0.008985036984086037, + 0.009156554937362671, + 0.041185278445482254, + 0.02141866460442543, + 0.009517000056803226, + 0.006758136209100485, + 0.01262133102864027, + 0.030592873692512512, + 0.007867584004998207, + 0.00996439903974533, + 0.030257077887654305, + 0.02524968795478344, + 0.013590444810688496, + 0.011669319123029709, + 0.02468053065240383, + 0.02247556298971176, + 0.013477618806064129, + 0.009530640207231045, + 0.02359415404498577, + 0.03190023452043533, + 0.030321428552269936, + 0.010422870516777039, + 0.023706922307610512, + 0.030422482639551163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.22554545104503632 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the vast ceiling of an abandoned factory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a complex assembly of interlocking cast-iron gears and thick roller chains coated in dark lubricant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the machinery shear from a rusted mounting bracket and slide along a large, grimy skylight panel before tumbling over the edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a steady medium shot where the camera remains stationary throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style desaturated and gritty, resembling the look of 1990s cinema?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020579971373081207 + ], + "opposite_logprob_tokens": [ + -3.8955800533294678 + ], + "logprob_reward": -0.020579971373081207, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020579971373081207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1150214821100235, + 0.07852815836668015, + 0.0732702687382698, + 0.11862974613904953, + 0.051333144307136536, + 0.028023652732372284, + 0.026126395910978317, + 0.0332907997071743, + 0.045302726328372955, + 0.4304737150669098 + ], + "credit_map": [ + 0.006414012983441353, + 0.012570164166390896, + 0.024756206199526787, + 0.025693466886878014, + 0.008742446079850197, + 0.0078108906745910645, + 0.01844891905784607, + 0.010585373267531395, + 0.007008228451013565, + 0.005877231247723103, + 0.014087141491472721, + 0.02079417183995247, + 0.00566788949072361, + 0.005393123719841242, + 0.01098680030554533, + 0.008713562972843647, + 0.005578354001045227, + 0.006744409911334515, + 0.01174665242433548, + 0.012664759531617165, + 0.006512452382594347, + 0.008121169172227383, + 0.01084333099424839, + 0.01105913519859314, + 0.009095815010368824, + 0.009866850450634956, + 0.020025696605443954, + 0.022542333230376244, + 0.010214647278189659, + 0.01066578645259142, + 0.01627560704946518, + 0.019943002611398697, + 0.0033467374742031097, + 0.007491541560739279, + 0.009346740320324898, + 0.00913831777870655, + 0.005555539391934872, + 0.002787928329780698, + 0.007737679407000542, + 0.005928657948970795, + 0.0024543232284486294, + 0.0027889804914593697, + 0.005167464260011911, + 0.0051193865947425365, + 0.002192530082538724, + 0.0032782412599772215, + 0.00416620122268796, + 0.0028565216343849897, + 0.0024724670220166445, + 0.002465379424393177, + 0.004361366853117943, + 0.003385726595297456, + 0.0035070525482296944, + 0.002578569343313575, + 0.004156820941716433, + 0.003199012018740177, + 0.0030669812113046646, + 0.0027690085116773844, + 0.007242103572934866, + 0.005999122746288776, + 0.003932279068976641, + 0.002414085902273655, + 0.003986399620771408, + 0.0038808146491646767, + 0.0028912904672324657, + 0.004711411893367767, + 0.00769441295415163, + 0.007675801869481802, + 0.0049965353682637215, + 0.0035465341061353683, + 0.006185055710375309, + 0.0076016769744455814, + 0.02921172045171261, + 0.033486656844615936, + 0.04731205850839615, + 0.04200154170393944, + 0.1254911869764328, + 0.022273821756243706, + 0.03994264081120491, + 0.09075405448675156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.020579971373081207 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_0/index_tensor([4474, 4474])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_0/index_tensor([4474, 4474])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6d4cec00a401f80f1876445acca6d45d2d67d198 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_0/index_tensor([4474, 4474])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "tight, slightly high-angled, static shot captures the delicate movements of a searing black cast iron grill, focused by the kitchen lights. A chef's hands, clean and expertly executed, firmly grasp a wide stainless steel spatula with a black composite handle. With swift, decisive precision, the hand slides the spatula beneath the thick, glistening beef patty, revealing dark, crisscrossed charcoal. Then, with a deft flick of the wrist, it flips over, releasing a puff of fragrant steam. The hand then gently places the spatula on the adjacent stainless steel counter, making a soft clinking sound. Without hesitation, the same hand turns to the soft, golden brioche bun, methodically placing the top half directly onto the cooking patty to heat and absorb the juices. Immediately afterwards, the same hand deftly removes several thin, round, flour-dusted flatbreads from a nearby pile and carefully places them, one by one, on the open, oiled section of the grill, where they begin to puff slightly from the heat, creating an atmosphere of focus, expertise, and deeply appetizing flavor.", + "seed": 30408, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_0/index_4474_seed_30408_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05057453345507383, + "vlm_logprob": -0.05057453345507383, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.05057453345507383, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, slightly high-angled, static view of a searing black cast iron grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a chef's hand use a wide stainless steel spatula with a black composite handle to flip a thick beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden brioche bun placed on top of the cooking patty to heat and absorb juices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, round, flour-dusted flatbreads placed one by one on the open, oiled section of the grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flatbreads on the grill begin to puff slightly from the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01050631981343031 + ], + "opposite_logprob_tokens": [ + -5.2605061531066895 + ], + "logprob_reward": -0.01050631981343031, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.01050631981343031, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28423526883125305, + 0.048716843128204346, + 0.08135481923818588, + 0.06772370636463165, + 0.1029210016131401, + 0.06783407181501389, + 0.05951514467597008, + 0.05887468159198761, + 0.16288262605667114, + 0.06594187021255493 + ], + "credit_map": [ + 0.02230166457593441, + 0.015932681038975716, + 0.010385420173406601, + 0.0075553059577941895, + 0.09755086153745651, + 0.09549185633659363, + 0.012707242742180824, + 0.022310223430395126, + 0.0045637949369847775, + 0.003349201986566186, + 0.0033626577351242304, + 0.006875452119857073, + 0.00852909218519926, + 0.006028680130839348, + 0.01002511940896511, + 0.0059828441590070724, + 0.008795308880507946, + 0.00638945447281003, + 0.004476578440517187, + 0.010636205784976482, + 0.008821855299174786, + 0.010099113918840885, + 0.019702160730957985, + 0.012434137985110283, + 0.005557992961257696, + 0.0084099555388093, + 0.005489223171025515, + 0.01337299682199955, + 0.0075790733098983765, + 0.007249838672578335, + 0.011477576568722725, + 0.008587050251662731, + 0.011957191862165928, + 0.008224098943173885, + 0.014117367565631866, + 0.014326287433505058, + 0.015310132876038551, + 0.01351427286863327, + 0.012030480429530144, + 0.013441167771816254, + 0.007977352477610111, + 0.009170196950435638, + 0.006985192187130451, + 0.010140045545995235, + 0.012739871628582478, + 0.008837314322590828, + 0.006503014825284481, + 0.005481088068336248, + 0.004509760998189449, + 0.009615842252969742, + 0.004796571563929319, + 0.008505916222929955, + 0.00743667408823967, + 0.006838313303887844, + 0.010067936964333057, + 0.007744130678474903, + 0.007524078246206045, + 0.006143264938145876, + 0.004083854611963034, + 0.010741469450294971, + 0.005304238758981228, + 0.005641473922878504, + 0.008055449463427067, + 0.011380854062736034, + 0.02592959813773632, + 0.019710412248969078, + 0.04076240956783295, + 0.014617805369198322, + 0.01714969612658024, + 0.02194531075656414, + 0.015842629596590996, + 0.006924772635102272, + 0.005497025325894356, + 0.0061242301017045975, + 0.004552697297185659, + 0.011082500219345093, + 0.006092076655477285, + 0.021307650953531265, + 0.0053217667154967785, + 0.005963916890323162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.01050631981343031 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, slightly high-angled, static view of a searing black cast iron grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a chef's hand use a wide stainless steel spatula with a black composite handle to flip a thick beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden brioche bun placed on top of the cooking patty to heat and absorb juices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, round, flour-dusted flatbreads placed one by one on the open, oiled section of the grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flatbreads on the grill begin to puff slightly from the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023317422717809677 + ], + "opposite_logprob_tokens": [ + -3.773317337036133 + ], + "logprob_reward": -0.023317422717809677, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023317422717809677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24390247464179993, + 0.046904221177101135, + 0.09371962398290634, + 0.061956170946359634, + 0.07481897622346878, + 0.03827454522252083, + 0.039720453321933746, + 0.08285271376371384, + 0.08085651695728302, + 0.23699423670768738 + ], + "credit_map": [ + 0.02383372001349926, + 0.01832733489573002, + 0.014418931677937508, + 0.013449492864310741, + 0.07328778505325317, + 0.06302299350500107, + 0.0186154842376709, + 0.01894674263894558, + 0.005415864288806915, + 0.004733438603579998, + 0.0031250794418156147, + 0.004622105974704027, + 0.009754636324942112, + 0.005019168835133314, + 0.00946146622300148, + 0.0047724684700369835, + 0.009567122906446457, + 0.00783987995237112, + 0.003992942161858082, + 0.011568492278456688, + 0.010018538683652878, + 0.010919350199401379, + 0.024989917874336243, + 0.014823392033576965, + 0.0036797316279262304, + 0.006333858706057072, + 0.004784741904586554, + 0.011111155152320862, + 0.0058953650295734406, + 0.01063233520835638, + 0.011012411676347256, + 0.00850658118724823, + 0.005669326521456242, + 0.007743454538285732, + 0.007515881676226854, + 0.009831136092543602, + 0.006942262407392263, + 0.007811347022652626, + 0.017911093309521675, + 0.011394479312002659, + 0.003786121029406786, + 0.005474457051604986, + 0.004238257184624672, + 0.006095303688198328, + 0.006021582055836916, + 0.0052675059996545315, + 0.0038189785555005074, + 0.003572340589016676, + 0.0037723046261817217, + 0.005090833641588688, + 0.0036193763371556997, + 0.004587216302752495, + 0.00968183670192957, + 0.0058211577124893665, + 0.003870125161483884, + 0.0032776077277958393, + 0.010064676403999329, + 0.007626428734511137, + 0.005736437626183033, + 0.011737325228750706, + 0.008651875890791416, + 0.009040990844368935, + 0.013055478222668171, + 0.016939517110586166, + 0.019402408972382545, + 0.0065640779212117195, + 0.028635475784540176, + 0.005775290075689554, + 0.005513589829206467, + 0.007725446484982967, + 0.004230097867548466, + 0.0030101400334388018, + 0.01679398864507675, + 0.006054193712770939, + 0.04326711595058441, + 0.03466442972421646, + 0.025303784757852554, + 0.092958465218544, + 0.007850635796785355, + 0.01010166760534048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.023317422717809677 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, slightly high-angled, static view of a searing black cast iron grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a chef's hand use a wide stainless steel spatula with a black composite handle to flip a thick beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden brioche bun placed on top of the cooking patty to heat and absorb juices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, round, flour-dusted flatbreads placed one by one on the open, oiled section of the grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flatbreads on the grill begin to puff slightly from the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038120366632938385 + ], + "opposite_logprob_tokens": [ + -3.2881202697753906 + ], + "logprob_reward": -0.038120366632938385, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038120366632938385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03917896747589111, + 0.04740144684910774, + 0.02311556041240692, + 0.03670436888933182, + 0.04157729819417, + 0.02711331844329834, + 0.017154298722743988, + 0.03625519201159477, + 0.4636569023132324, + 0.2678426206111908 + ], + "credit_map": [ + 0.00333423656411469, + 0.002862347289919853, + 0.003078147303313017, + 0.0029792520217597485, + 0.010029172524809837, + 0.009836000390350819, + 0.003449093084782362, + 0.0036107173655182123, + 0.0046192193403840065, + 0.0033405902795493603, + 0.002422383986413479, + 0.005722333211451769, + 0.01055966317653656, + 0.005493330303579569, + 0.010024181567132473, + 0.005219743587076664, + 0.0033276770263910294, + 0.002104307059198618, + 0.0013320520520210266, + 0.0023369721602648497, + 0.0025204552803188562, + 0.0029044323600828648, + 0.0050292229279875755, + 0.0035604401491582394, + 0.0020670131780207157, + 0.00290037808008492, + 0.0025321939028799534, + 0.006462988443672657, + 0.003875414840877056, + 0.005659511312842369, + 0.007874475792050362, + 0.005332393571734428, + 0.004623827058821917, + 0.0040030935779213905, + 0.004159021656960249, + 0.005794530268758535, + 0.004705358762294054, + 0.004313738085329533, + 0.008761288598179817, + 0.0052164397202432156, + 0.003236540826037526, + 0.0022483165375888348, + 0.0026640130672603846, + 0.003164112102240324, + 0.005096386652439833, + 0.004468380473554134, + 0.0033564739860594273, + 0.002879094798117876, + 0.0013116954360157251, + 0.0021357203368097544, + 0.0015406616730615497, + 0.0023594629019498825, + 0.0028372325468808413, + 0.002235901076346636, + 0.0026959707029163837, + 0.0020376534666866064, + 0.004027518909424543, + 0.0034105994272977114, + 0.0034246519207954407, + 0.003980816341936588, + 0.00409476924687624, + 0.005009930580854416, + 0.004648829810321331, + 0.007658075541257858, + 0.13445162773132324, + 0.020993700250983238, + 0.18525001406669617, + 0.028586456552147865, + 0.023460628464818, + 0.04121607542037964, + 0.015450702048838139, + 0.014247706159949303, + 0.018550872802734375, + 0.0050026848912239075, + 0.05207538604736328, + 0.04016667231917381, + 0.024718597531318665, + 0.10788658261299133, + 0.0074433074332773685, + 0.011998521164059639 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 1 + ], + "selected_reward": -0.038120366632938385 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, slightly high-angled, static view of a searing black cast iron grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a chef's hand use a wide stainless steel spatula with a black composite handle to flip a thick beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden brioche bun placed on top of the cooking patty to heat and absorb juices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, round, flour-dusted flatbreads placed one by one on the open, oiled section of the grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flatbreads on the grill begin to puff slightly from the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381036102771759 + ], + "opposite_logprob_tokens": [ + -3.2881035804748535 + ], + "logprob_reward": -0.0381036102771759, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381036102771759, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10833179950714111, + 0.0674126073718071, + 0.05317096784710884, + 0.041311703622341156, + 0.04363726079463959, + 0.030760306864976883, + 0.034452520310878754, + 0.04683603718876839, + 0.5004274845123291, + 0.07365933060646057 + ], + "credit_map": [ + 0.010931660421192646, + 0.008471516892313957, + 0.011250379495322704, + 0.008225569501519203, + 0.02490587718784809, + 0.02475920505821705, + 0.010242991149425507, + 0.009544599801301956, + 0.007093128282576799, + 0.0059491791762411594, + 0.00556652806699276, + 0.010053752921521664, + 0.011314861476421356, + 0.006005879491567612, + 0.013171021826565266, + 0.008258254267275333, + 0.00935570988804102, + 0.0055451784282922745, + 0.003403695300221443, + 0.005471528973430395, + 0.006102428305894136, + 0.005264149513095617, + 0.009911399334669113, + 0.008116877637803555, + 0.0031004019547253847, + 0.0038307178765535355, + 0.0040580518543720245, + 0.006729601416736841, + 0.005204825196415186, + 0.004707254469394684, + 0.008479801937937737, + 0.005201051477342844, + 0.004582850728183985, + 0.004936020355671644, + 0.0071012042462825775, + 0.006830053869634867, + 0.005186138208955526, + 0.004909317474812269, + 0.005385534837841988, + 0.00470614293590188, + 0.00348380533978343, + 0.003564892802387476, + 0.0029731050599366426, + 0.004422353580594063, + 0.005554830189794302, + 0.003698094515129924, + 0.00408389326184988, + 0.0029793321155011654, + 0.002267546020448208, + 0.0023905278649181128, + 0.0033195160795003176, + 0.005583573132753372, + 0.004967792890965939, + 0.0047005545347929, + 0.006689961533993483, + 0.004533044528216124, + 0.006057457532733679, + 0.005639664828777313, + 0.003348604077473283, + 0.00458613270893693, + 0.004975393880158663, + 0.0033084205351769924, + 0.007327169179916382, + 0.011593191884458065, + 0.1444425731897354, + 0.024118155241012573, + 0.19730184972286224, + 0.03115801140666008, + 0.026016104966402054, + 0.04302661120891571, + 0.018785376101732254, + 0.015578816644847393, + 0.003899236908182502, + 0.006202344316989183, + 0.0026868553832173347, + 0.010616774670779705, + 0.014935142360627651, + 0.023846585303544998, + 0.005984221585094929, + 0.005488176830112934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.0381036102771759 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, slightly high-angled, static view of a searing black cast iron grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a chef's hand use a wide stainless steel spatula with a black composite handle to flip a thick beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden brioche bun placed on top of the cooking patty to heat and absorb juices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, round, flour-dusted flatbreads placed one by one on the open, oiled section of the grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flatbreads on the grill begin to puff slightly from the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428249478340149 + ], + "opposite_logprob_tokens": [ + -2.01782488822937 + ], + "logprob_reward": -0.1428249478340149, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1428249478340149, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07897940278053284, + 0.048727259039878845, + 0.038248833268880844, + 0.09154435247182846, + 0.03030954860150814, + 0.024816516786813736, + 0.09403910487890244, + 0.03310784697532654, + 0.47503843903541565, + 0.08518867194652557 + ], + "credit_map": [ + 0.009945021010935307, + 0.008392869494855404, + 0.007830225862562656, + 0.004794032778590918, + 0.018712077289819717, + 0.015220863744616508, + 0.00784089881926775, + 0.006243419833481312, + 0.0060457494109869, + 0.00447824876755476, + 0.0030385267455130816, + 0.005556994117796421, + 0.008423920720815659, + 0.005755789112299681, + 0.009182993322610855, + 0.006245041731745005, + 0.004325373563915491, + 0.004851324949413538, + 0.0018554527778178453, + 0.0037368834018707275, + 0.004668132867664099, + 0.0045004370622336864, + 0.008387397974729538, + 0.005923834163695574, + 0.004898408893495798, + 0.008487014099955559, + 0.007447982206940651, + 0.01466826256364584, + 0.01174677349627018, + 0.01204114779829979, + 0.017364956438541412, + 0.014889816753566265, + 0.0025840834714472294, + 0.003061640076339245, + 0.0025444955099374056, + 0.003461351851001382, + 0.003100003581494093, + 0.005242779850959778, + 0.005363766103982925, + 0.004951432812958956, + 0.003395093372091651, + 0.0022410876117646694, + 0.0034952983260154724, + 0.0034572959411889315, + 0.003652446437627077, + 0.00303508248180151, + 0.0024582359474152327, + 0.003081980627030134, + 0.006063718814402819, + 0.010736237280070782, + 0.006127370987087488, + 0.01453781221061945, + 0.01605098880827427, + 0.00973060354590416, + 0.0179404579102993, + 0.012851925566792488, + 0.004198036622256041, + 0.00500519247725606, + 0.0023167049512267113, + 0.0036466310266405344, + 0.003577080089598894, + 0.003044143784791231, + 0.004596156533807516, + 0.006723903119564056, + 0.13295240700244904, + 0.02471771650016308, + 0.19018954038619995, + 0.03033457323908806, + 0.024772480130195618, + 0.04701052978634834, + 0.010666271671652794, + 0.014394951052963734, + 0.004782446194440126, + 0.003019318450242281, + 0.015469119884073734, + 0.01355315838009119, + 0.006732035428285599, + 0.03373534977436066, + 0.00334744481369853, + 0.004549808334559202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 3 + ], + "selected_reward": -0.1428249478340149 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, slightly high-angled, static shot captures the delicate movements of a searing black cast iron grill, focused by the kitchen lights. A chef's hands, clean and expertly executed, firmly grasp a wide stainless steel spatula with a black composite handle. With swift, decisive precision, the hand slides the spatula beneath the thick, glistening beef patty, revealing dark, crisscrossed charcoal. Then, with a deft flick of the wrist, it flips over, releasing a puff of fragrant steam. The hand then gently places the spatula on the adjacent stainless steel counter, making a soft clinking sound. Without hesitation, the same hand turns to the soft, golden brioche bun, methodically placing the top half directly onto the cooking patty to heat and absorb the juices. Immediately afterwards, the same hand deftly removes several thin, round, flour-dusted flatbreads from a nearby pile and carefully places them, one by one, on the open, oiled section of the grill, where they begin to puff slightly from the heat, creating an atmosphere of focus, expertise, and deeply appetizing flavor.", + "seed": 30409, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_0/index_4474_seed_30409_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04557411186397076, + "vlm_logprob": -0.04557411186397076, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.04557411186397076, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, slightly high-angled, static view of a searing black cast iron grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a chef's hand use a wide stainless steel spatula with a black composite handle to flip a thick beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden brioche bun placed on top of the cooking patty to heat and absorb juices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, round, flour-dusted flatbreads placed one by one on the open, oiled section of the grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flatbreads on the grill begin to puff slightly from the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017396362498402596 + ], + "opposite_logprob_tokens": [ + -4.892396450042725 + ], + "logprob_reward": -0.017396362498402596, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.017396362498402596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07569339871406555, + 0.023075753822922707, + 0.022538399323821068, + 0.059000223875045776, + 0.16135606169700623, + 0.19270296394824982, + 0.04986514151096344, + 0.051750823855400085, + 0.21411718428134918, + 0.14989997446537018 + ], + "credit_map": [ + 0.007379884831607342, + 0.008503938093781471, + 0.007967162877321243, + 0.014178906567394733, + 0.008129068650305271, + 0.006427150219678879, + 0.014526952989399433, + 0.008580341003835201, + 0.002335703931748867, + 0.0013897698372602463, + 0.0025251416955143213, + 0.005534932017326355, + 0.0024404104333370924, + 0.00232623890042305, + 0.003068811958655715, + 0.003454746911302209, + 0.0016960055800154805, + 0.0021414891816675663, + 0.00301473424769938, + 0.0042886994779109955, + 0.0019585092086344957, + 0.0027229797560721636, + 0.0036965373437851667, + 0.003019446274265647, + 0.006194190587848425, + 0.006049421615898609, + 0.0044836318120360374, + 0.009390873834490776, + 0.007343290373682976, + 0.007888562977313995, + 0.009518575854599476, + 0.008131683804094791, + 0.013688665814697742, + 0.012692674994468689, + 0.017963523045182228, + 0.04008716344833374, + 0.01315967459231615, + 0.015493347309529781, + 0.02347526326775551, + 0.024795766919851303, + 0.021619416773319244, + 0.021371891722083092, + 0.01867629773914814, + 0.03377119079232216, + 0.02470092661678791, + 0.013302544131875038, + 0.031666241586208344, + 0.027594471350312233, + 0.005837644916027784, + 0.007310837972909212, + 0.004607789684087038, + 0.006470485590398312, + 0.00825535412877798, + 0.004515761975198984, + 0.006294404622167349, + 0.006572866812348366, + 0.003250471083447337, + 0.0029113320633769035, + 0.0049154385924339294, + 0.00599330011755228, + 0.005230027716606855, + 0.0030147761572152376, + 0.011604245752096176, + 0.014831236563622952, + 0.008295658975839615, + 0.00928550772368908, + 0.03569343686103821, + 0.018179098144173622, + 0.011038067750632763, + 0.011617670767009258, + 0.04240158572793007, + 0.07760617136955261, + 0.007389000616967678, + 0.005257381591945887, + 0.032357219606637955, + 0.008677594363689423, + 0.005654609762132168, + 0.013848895207047462, + 0.02243192121386528, + 0.05428337678313255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 4 + ], + "selected_reward": -0.017396362498402596 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, slightly high-angled, static view of a searing black cast iron grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a chef's hand use a wide stainless steel spatula with a black composite handle to flip a thick beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden brioche bun placed on top of the cooking patty to heat and absorb juices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, round, flour-dusted flatbreads placed one by one on the open, oiled section of the grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flatbreads on the grill begin to puff slightly from the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023315908387303352 + ], + "opposite_logprob_tokens": [ + -3.773315906524658 + ], + "logprob_reward": -0.023315908387303352, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023315908387303352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.113816037774086, + 0.23398365080356598, + 0.045030947774648666, + 0.10697972029447556, + 0.03805827721953392, + 0.10420359671115875, + 0.026513991877436638, + 0.09433688968420029, + 0.09660068154335022, + 0.14047618210315704 + ], + "credit_map": [ + 0.009517627768218517, + 0.015669992193579674, + 0.00995232816785574, + 0.016136104241013527, + 0.012238300405442715, + 0.023372424766421318, + 0.0132567984983325, + 0.013672477565705776, + 0.017700666561722755, + 0.017668461427092552, + 0.03398154675960541, + 0.03403209149837494, + 0.028272582218050957, + 0.024849241599440575, + 0.05021227151155472, + 0.027266817167401314, + 0.00388609291985631, + 0.0034195531625300646, + 0.005851227790117264, + 0.006885725073516369, + 0.0047430736012756824, + 0.007402781862765551, + 0.006677172612398863, + 0.006165326572954655, + 0.005316538270562887, + 0.006471897475421429, + 0.010900039225816727, + 0.022552402690052986, + 0.009986168704926968, + 0.011327648535370827, + 0.022632448002696037, + 0.01779259368777275, + 0.002592529868707061, + 0.002092642243951559, + 0.004383221734315157, + 0.008940239436924458, + 0.0025773856323212385, + 0.0030115642584860325, + 0.007487583439797163, + 0.0069731115363538265, + 0.011342884972691536, + 0.01820010133087635, + 0.00670106615871191, + 0.018072644248604774, + 0.012853537686169147, + 0.006730654276907444, + 0.017417486757040024, + 0.012885231524705887, + 0.0035601966083049774, + 0.0025395264383405447, + 0.0024384590797126293, + 0.004190228879451752, + 0.004028496332466602, + 0.001917947200126946, + 0.003565739141777158, + 0.004273402038961649, + 0.007910137996077538, + 0.009737644344568253, + 0.010191786102950573, + 0.01375175453722477, + 0.01586005836725235, + 0.0076286974363029, + 0.01453151274472475, + 0.014725304208695889, + 0.006270493380725384, + 0.004255396779626608, + 0.02820543572306633, + 0.00864191260188818, + 0.005355390720069408, + 0.0066030798479914665, + 0.018488852307200432, + 0.01878013089299202, + 0.007762244902551174, + 0.005463459063321352, + 0.047457173466682434, + 0.009386474266648293, + 0.004365730099380016, + 0.010496987029910088, + 0.022746769711375237, + 0.032797351479530334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.023315908387303352 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, slightly high-angled, static view of a searing black cast iron grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a chef's hand use a wide stainless steel spatula with a black composite handle to flip a thick beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden brioche bun placed on top of the cooking patty to heat and absorb juices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, round, flour-dusted flatbreads placed one by one on the open, oiled section of the grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flatbreads on the grill begin to puff slightly from the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033741552382707596 + ], + "opposite_logprob_tokens": [ + -3.4087414741516113 + ], + "logprob_reward": -0.033741552382707596, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033741552382707596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08159022778272629, + 0.05748576670885086, + 0.051126446574926376, + 0.1328846514225006, + 0.06412311643362045, + 0.06761544197797775, + 0.026950620114803314, + 0.14415891468524933, + 0.154827281832695, + 0.21923759579658508 + ], + "credit_map": [ + 0.005598521325737238, + 0.005883585661649704, + 0.007199161686003208, + 0.017826996743679047, + 0.009736474603414536, + 0.007478957064449787, + 0.01627248339354992, + 0.011594039388000965, + 0.00496315024793148, + 0.00338463531807065, + 0.005417979322373867, + 0.013119830749928951, + 0.00441017234697938, + 0.005689454264938831, + 0.011506793089210987, + 0.008993744850158691, + 0.0035508298315107822, + 0.0023982864804565907, + 0.008382979780435562, + 0.01241410244256258, + 0.003336372785270214, + 0.004437008872628212, + 0.008693702518939972, + 0.007913157343864441, + 0.008294038474559784, + 0.01091211847960949, + 0.012454847805202007, + 0.03186708316206932, + 0.01260733138769865, + 0.013595045544207096, + 0.02217610366642475, + 0.020978081971406937, + 0.007001660298556089, + 0.0070112613029778, + 0.008946394547820091, + 0.011024697683751583, + 0.005902118515223265, + 0.0046735648065805435, + 0.010496185161173344, + 0.009067224338650703, + 0.00845204759389162, + 0.006955687887966633, + 0.0067974659614264965, + 0.00871983915567398, + 0.007041838485747576, + 0.006629780866205692, + 0.012176021002233028, + 0.010842760093510151, + 0.0027644624933600426, + 0.001963692717254162, + 0.0026251489762216806, + 0.0043809860944747925, + 0.003796377219259739, + 0.0037144517991691828, + 0.0037227850407361984, + 0.003982711583375931, + 0.005975741893053055, + 0.005606008227914572, + 0.009408775717020035, + 0.01091714110225439, + 0.011289632879197598, + 0.008704990148544312, + 0.038306400179862976, + 0.05395021289587021, + 0.00995210837572813, + 0.006857537664473057, + 0.040816906839609146, + 0.012696187943220139, + 0.00875599030405283, + 0.012210996821522713, + 0.013284621760249138, + 0.05025291070342064, + 0.0115269236266613, + 0.007945554330945015, + 0.06577688455581665, + 0.017275283113121986, + 0.0056639001704752445, + 0.01948416791856289, + 0.01568194292485714, + 0.0758829265832901 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.033741552382707596 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, slightly high-angled, static view of a searing black cast iron grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a chef's hand use a wide stainless steel spatula with a black composite handle to flip a thick beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden brioche bun placed on top of the cooking patty to heat and absorb juices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, round, flour-dusted flatbreads placed one by one on the open, oiled section of the grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flatbreads on the grill begin to puff slightly from the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634863369166851 + ], + "opposite_logprob_tokens": [ + -3.65134859085083 + ], + "logprob_reward": -0.02634863369166851, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634863369166851, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1252630054950714, + 0.04476693645119667, + 0.08015967905521393, + 0.12152963876724243, + 0.050460513681173325, + 0.034835897386074066, + 0.03260711953043938, + 0.13959835469722748, + 0.1310768574476242, + 0.23970195651054382 + ], + "credit_map": [ + 0.0076620085164904594, + 0.014153708703815937, + 0.010240118019282818, + 0.023944836109876633, + 0.011499622836709023, + 0.02266349457204342, + 0.018962282687425613, + 0.016136925667524338, + 0.0034448967780917883, + 0.003122104099020362, + 0.006318265106528997, + 0.006776418536901474, + 0.005167173221707344, + 0.00468523520976305, + 0.008697835728526115, + 0.006555006839334965, + 0.008854594081640244, + 0.006408554036170244, + 0.009099485352635384, + 0.014406655915081501, + 0.006856986321508884, + 0.011368303559720516, + 0.011618802323937416, + 0.01154630072414875, + 0.007353648077696562, + 0.011647682636976242, + 0.012202507816255093, + 0.02961714006960392, + 0.013732335530221462, + 0.014065273106098175, + 0.016108965501189232, + 0.01680208556354046, + 0.004619606304913759, + 0.00346467480994761, + 0.00739941094070673, + 0.007981278002262115, + 0.004914197605103254, + 0.004179481416940689, + 0.009584137238562107, + 0.008317725732922554, + 0.003073126310482621, + 0.002282708417624235, + 0.002937684301286936, + 0.007008356973528862, + 0.004625069443136454, + 0.003248099237680435, + 0.005850259680300951, + 0.0058105927892029285, + 0.0028727538883686066, + 0.0019891969859600067, + 0.0025578809436410666, + 0.0060731531120836735, + 0.006221226882189512, + 0.0037990135606378317, + 0.004643524996936321, + 0.00445036543533206, + 0.007926895283162594, + 0.005086130928248167, + 0.012284665368497372, + 0.009881818667054176, + 0.009266231209039688, + 0.009289784356951714, + 0.03446672484278679, + 0.05139610171318054, + 0.007477934006601572, + 0.005568896420300007, + 0.02909761667251587, + 0.01266537792980671, + 0.009185965172946453, + 0.008078179322183132, + 0.04040217399597168, + 0.018600698560476303, + 0.01299919281154871, + 0.008306674659252167, + 0.08202607929706573, + 0.016957592219114304, + 0.006301585119217634, + 0.019080698490142822, + 0.030331438407301903, + 0.0636986792087555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.02634863369166851 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, slightly high-angled, static view of a searing black cast iron grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a chef's hand use a wide stainless steel spatula with a black composite handle to flip a thick beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden brioche bun placed on top of the cooking patty to heat and absorb juices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, round, flour-dusted flatbreads placed one by one on the open, oiled section of the grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flatbreads on the grill begin to puff slightly from the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706810235977173 + ], + "opposite_logprob_tokens": [ + -2.127068042755127 + ], + "logprob_reward": -0.12706810235977173, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12706810235977173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11653237044811249, + 0.08314739912748337, + 0.15458476543426514, + 0.16479016840457916, + 0.07005557417869568, + 0.11962399631738663, + 0.03244096785783768, + 0.08034177869558334, + 0.1278872936964035, + 0.050595711916685104 + ], + "credit_map": [ + 0.010478248819708824, + 0.014110271818935871, + 0.013283153995871544, + 0.017686758190393448, + 0.016389675438404083, + 0.012710881419479847, + 0.017482679337263107, + 0.014390692114830017, + 0.00752723915502429, + 0.005359988659620285, + 0.009380744770169258, + 0.021898308768868446, + 0.007302781566977501, + 0.007510540075600147, + 0.011376832611858845, + 0.012790950946509838, + 0.014257586561143398, + 0.011725381016731262, + 0.019747084006667137, + 0.03110363334417343, + 0.016238350421190262, + 0.017282143235206604, + 0.023426365107297897, + 0.02080421894788742, + 0.009347186423838139, + 0.013695231638848782, + 0.017220603302121162, + 0.03862842917442322, + 0.015820007771253586, + 0.019917473196983337, + 0.024905193597078323, + 0.02525603212416172, + 0.007087578997015953, + 0.00604100339114666, + 0.008571109734475613, + 0.01113609503954649, + 0.005575450602918863, + 0.006237882189452648, + 0.013587920926511288, + 0.011818530969321728, + 0.01607217825949192, + 0.016076724976301193, + 0.010356154292821884, + 0.01402227021753788, + 0.01323788147419691, + 0.007776268757879734, + 0.022917060181498528, + 0.019165456295013428, + 0.003973538987338543, + 0.0031048855744302273, + 0.0030214067082852125, + 0.004339029546827078, + 0.006473687011748552, + 0.0033022412098944187, + 0.003800438717007637, + 0.0044257366098463535, + 0.0075505380518734455, + 0.006730691529810429, + 0.008237877860665321, + 0.011147947050631046, + 0.010588932782411575, + 0.0065185329876840115, + 0.014039970934391022, + 0.015527279116213322, + 0.008612497709691525, + 0.0054845307022333145, + 0.03737352415919304, + 0.008873493410646915, + 0.006698494777083397, + 0.009259848855435848, + 0.013810044154524803, + 0.03777484968304634, + 0.004505071323364973, + 0.0025232562329620123, + 0.0102932658046484, + 0.003919691313058138, + 0.0026385451201349497, + 0.0065598394721746445, + 0.004990709945559502, + 0.015165325254201889 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 8 + ], + "selected_reward": -0.12706810235977173 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, slightly high-angled, static shot captures the delicate movements of a searing black cast iron grill, focused by the kitchen lights. A chef's hands, clean and expertly executed, firmly grasp a wide stainless steel spatula with a black composite handle. With swift, decisive precision, the hand slides the spatula beneath the thick, glistening beef patty, revealing dark, crisscrossed charcoal. Then, with a deft flick of the wrist, it flips over, releasing a puff of fragrant steam. The hand then gently places the spatula on the adjacent stainless steel counter, making a soft clinking sound. Without hesitation, the same hand turns to the soft, golden brioche bun, methodically placing the top half directly onto the cooking patty to heat and absorb the juices. Immediately afterwards, the same hand deftly removes several thin, round, flour-dusted flatbreads from a nearby pile and carefully places them, one by one, on the open, oiled section of the grill, where they begin to puff slightly from the heat, creating an atmosphere of focus, expertise, and deeply appetizing flavor.", + "seed": 30410, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_0/index_4474_seed_30410_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.034829413332045076, + "vlm_logprob": -0.034829413332045076, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.034829413332045076, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, slightly high-angled, static view of a searing black cast iron grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a chef's hand use a wide stainless steel spatula with a black composite handle to flip a thick beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden brioche bun placed on top of the cooking patty to heat and absorb juices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, round, flour-dusted flatbreads placed one by one on the open, oiled section of the grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flatbreads on the grill begin to puff slightly from the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009346184320747852 + ], + "opposite_logprob_tokens": [ + -5.259346008300781 + ], + "logprob_reward": -0.009346184320747852, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.009346184320747852, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20065850019454956, + 0.07743220776319504, + 0.1706245243549347, + 0.015993379056453705, + 0.04854576289653778, + 0.04984099045395851, + 0.09397279471158981, + 0.11235249042510986, + 0.1415264755487442, + 0.0890529528260231 + ], + "credit_map": [ + 0.02568022347986698, + 0.019413063302636147, + 0.020408155396580696, + 0.022370178252458572, + 0.02968144230544567, + 0.027635624632239342, + 0.029428333044052124, + 0.026041468605399132, + 0.007358571980148554, + 0.007341071963310242, + 0.012367116287350655, + 0.011750251986086369, + 0.00631055748090148, + 0.0085260896012187, + 0.01473364606499672, + 0.009044898673892021, + 0.00814269483089447, + 0.007081655319780111, + 0.012962560169398785, + 0.00769395288079977, + 0.019034987315535545, + 0.0507957749068737, + 0.01004138495773077, + 0.05487152189016342, + 0.0012847606558352709, + 0.001628198428079486, + 0.0017712541157379746, + 0.0015421647112816572, + 0.0024811404291540384, + 0.003456238890066743, + 0.0016170768067240715, + 0.002212545834481716, + 0.008391273207962513, + 0.00567439803853631, + 0.005336274392902851, + 0.003523807041347027, + 0.005221155006438494, + 0.006279290188103914, + 0.006449723616242409, + 0.0076698423363268375, + 0.006681674625724554, + 0.005712599493563175, + 0.005731706507503986, + 0.007263172417879105, + 0.0052579897455871105, + 0.00528599601238966, + 0.0064401086419820786, + 0.007467744406312704, + 0.016176460310816765, + 0.01437514927238226, + 0.006348806899040937, + 0.005391137208789587, + 0.012586022727191448, + 0.007749490439891815, + 0.016839632764458656, + 0.014506091363728046, + 0.012341638095676899, + 0.015743233263492584, + 0.01288188062608242, + 0.013058159500360489, + 0.014853518456220627, + 0.008182518184185028, + 0.014869660139083862, + 0.02042187750339508, + 0.022247740998864174, + 0.026850562542676926, + 0.01097511500120163, + 0.015206019394099712, + 0.01852669194340706, + 0.011066916398704052, + 0.025058511644601822, + 0.011594907380640507, + 0.008769873529672623, + 0.00890157651156187, + 0.013758865185081959, + 0.011322504840791225, + 0.009466474875807762, + 0.008123340085148811, + 0.008013335056602955, + 0.020696980878710747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.009346184320747852 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, slightly high-angled, static view of a searing black cast iron grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a chef's hand use a wide stainless steel spatula with a black composite handle to flip a thick beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden brioche bun placed on top of the cooking patty to heat and absorb juices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, round, flour-dusted flatbreads placed one by one on the open, oiled section of the grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flatbreads on the grill begin to puff slightly from the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020601458847522736 + ], + "opposite_logprob_tokens": [ + -3.895601511001587 + ], + "logprob_reward": -0.020601458847522736, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020601458847522736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12406200170516968, + 0.051389794796705246, + 0.16640938818454742, + 0.013952855952084064, + 0.041452426463365555, + 0.10238233208656311, + 0.10182935744524002, + 0.12508727610111237, + 0.07593300938606262, + 0.19750161468982697 + ], + "credit_map": [ + 0.013733576983213425, + 0.016842886805534363, + 0.013491908088326454, + 0.016852542757987976, + 0.017338689416646957, + 0.016680575907230377, + 0.013107527047395706, + 0.01601429283618927, + 0.005308233667165041, + 0.004277386702597141, + 0.00753250764682889, + 0.0049838777631521225, + 0.004550542216747999, + 0.008937201462686062, + 0.006103979889303446, + 0.009696059860289097, + 0.009736467152833939, + 0.0072371396236121655, + 0.017057647928595543, + 0.009511918760836124, + 0.017385607585310936, + 0.045606836676597595, + 0.010509057901799679, + 0.049364712089300156, + 0.0012293545296415687, + 0.0019010689575225115, + 0.00192217156291008, + 0.0019057344179600477, + 0.0014018721412867308, + 0.001761251362040639, + 0.001935748034156859, + 0.0018956540152430534, + 0.005453227553516626, + 0.005022012162953615, + 0.004435060545802116, + 0.002633411902934313, + 0.004927793517708778, + 0.006376834120601416, + 0.005849079228937626, + 0.006755006033927202, + 0.013817084021866322, + 0.012755731120705605, + 0.010786066763103008, + 0.01572383940219879, + 0.011162619106471539, + 0.011117168702185154, + 0.012316995300352573, + 0.014702818356454372, + 0.01541267428547144, + 0.01101610530167818, + 0.00888858549296856, + 0.009134654887020588, + 0.013027725741267204, + 0.010758580639958382, + 0.020941028371453285, + 0.01264999620616436, + 0.008844709023833275, + 0.012933827005326748, + 0.007901694625616074, + 0.018306508660316467, + 0.016672855243086815, + 0.01031744945794344, + 0.03266666829586029, + 0.017443548887968063, + 0.012054942548274994, + 0.011073979549109936, + 0.007272695656865835, + 0.009510164149105549, + 0.009054753929376602, + 0.004746236372739077, + 0.013666721992194653, + 0.008553511463105679, + 0.013400164432823658, + 0.018495354801416397, + 0.05880056321620941, + 0.014673049561679363, + 0.015731841325759888, + 0.007376900874078274, + 0.017299961298704147, + 0.05172376334667206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 7 + ], + "selected_reward": -0.020601458847522736 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, slightly high-angled, static view of a searing black cast iron grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a chef's hand use a wide stainless steel spatula with a black composite handle to flip a thick beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden brioche bun placed on top of the cooking patty to heat and absorb juices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, round, flour-dusted flatbreads placed one by one on the open, oiled section of the grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flatbreads on the grill begin to puff slightly from the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330857142806053 + ], + "opposite_logprob_tokens": [ + -3.773308515548706 + ], + "logprob_reward": -0.02330857142806053, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330857142806053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060745712369680405, + 0.12094877660274506, + 0.039919834583997726, + 0.016999626532197, + 0.0413239449262619, + 0.04566024988889694, + 0.2009197175502777, + 0.2488158494234085, + 0.09059169143438339, + 0.1340745985507965 + ], + "credit_map": [ + 0.008195297792553902, + 0.0061643775552511215, + 0.004834218882024288, + 0.006509180646389723, + 0.00986315868794918, + 0.01048752386122942, + 0.0061212461441755295, + 0.008570706471800804, + 0.01334386970847845, + 0.014676855877041817, + 0.020086372271180153, + 0.019727444276213646, + 0.009744030423462391, + 0.014341884292662144, + 0.01298595778644085, + 0.01604236103594303, + 0.002518892288208008, + 0.0020990094635635614, + 0.005072786472737789, + 0.0030952992383390665, + 0.0045337677001953125, + 0.009555323980748653, + 0.003167280461639166, + 0.009877479635179043, + 0.0014174494426697493, + 0.0014792904257774353, + 0.00211062072776258, + 0.0019691477064043283, + 0.0020254994742572308, + 0.003078465349972248, + 0.0019776476547122, + 0.002941505517810583, + 0.005732659250497818, + 0.003953823819756508, + 0.00411986093968153, + 0.0030828791204839945, + 0.006249377503991127, + 0.005739042069762945, + 0.004905762616544962, + 0.007540540769696236, + 0.005571568384766579, + 0.006546859163790941, + 0.00412947591394186, + 0.007039432879537344, + 0.0045188418589532375, + 0.005436399020254612, + 0.006756778806447983, + 0.005660889204591513, + 0.03236905112862587, + 0.025112036615610123, + 0.009240429848432541, + 0.010589401237666607, + 0.025532394647598267, + 0.020897064357995987, + 0.04334092512726784, + 0.0338384211063385, + 0.014430513605475426, + 0.018649429082870483, + 0.025371814146637917, + 0.03873351588845253, + 0.022555120289325714, + 0.018402691930532455, + 0.06476827710866928, + 0.045904502272605896, + 0.015449924394488335, + 0.015428248792886734, + 0.0055795577354729176, + 0.009632567875087261, + 0.014101608656346798, + 0.00824673194438219, + 0.01398716401308775, + 0.008165891282260418, + 0.010025604628026485, + 0.013225467875599861, + 0.03490153327584267, + 0.009350868873298168, + 0.012097693048417568, + 0.006051914766430855, + 0.015474013052880764, + 0.03294750675559044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.02330857142806053 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, slightly high-angled, static view of a searing black cast iron grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a chef's hand use a wide stainless steel spatula with a black composite handle to flip a thick beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden brioche bun placed on top of the cooking patty to heat and absorb juices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, round, flour-dusted flatbreads placed one by one on the open, oiled section of the grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flatbreads on the grill begin to puff slightly from the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020583007484674454 + ], + "opposite_logprob_tokens": [ + -3.895582914352417 + ], + "logprob_reward": -0.020583007484674454, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020583007484674454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0746716856956482, + 0.2265729159116745, + 0.1123352199792862, + 0.014419753104448318, + 0.034357473254203796, + 0.028832975775003433, + 0.14886675775051117, + 0.1878672093153, + 0.07618243247270584, + 0.09589353948831558 + ], + "credit_map": [ + 0.009016976691782475, + 0.008072561584413052, + 0.008921392261981964, + 0.008375855162739754, + 0.011334248818457127, + 0.011733724735677242, + 0.006968339905142784, + 0.010248586535453796, + 0.02821965701878071, + 0.03682209551334381, + 0.034019555896520615, + 0.0330238901078701, + 0.018096476793289185, + 0.02398073300719261, + 0.023040127009153366, + 0.0293703842908144, + 0.006001359783113003, + 0.006351938936859369, + 0.016543487086892128, + 0.006983079947531223, + 0.012118670158088207, + 0.02843281999230385, + 0.006464317906647921, + 0.029439544305205345, + 0.0012943227775394917, + 0.0013667340390384197, + 0.0019486439414322376, + 0.0018327205907553434, + 0.0013978270580992103, + 0.0027988606598228216, + 0.0011594833340495825, + 0.002621160354465246, + 0.0037846523337066174, + 0.004221559502184391, + 0.0034519669134169817, + 0.002813725732266903, + 0.004662191029638052, + 0.0048748524859547615, + 0.005242759361863136, + 0.005305766593664885, + 0.003020412987098098, + 0.002710978500545025, + 0.0027401496190577745, + 0.004891149699687958, + 0.002745340345427394, + 0.0039908261969685555, + 0.005203812848776579, + 0.0035303079057484865, + 0.029444411396980286, + 0.016702422872185707, + 0.010724262334406376, + 0.009882569313049316, + 0.027765488252043724, + 0.012733837589621544, + 0.02360588125884533, + 0.018007885664701462, + 0.01360529288649559, + 0.01976579986512661, + 0.026100503280758858, + 0.03155108168721199, + 0.01725958287715912, + 0.014935052953660488, + 0.02876337245106697, + 0.03588651865720749, + 0.015372366644442081, + 0.011840424500405788, + 0.006310592871159315, + 0.008071593940258026, + 0.008296428248286247, + 0.005809763912111521, + 0.013014940544962883, + 0.007466321345418692, + 0.0080430768430233, + 0.012058665044605732, + 0.01874740608036518, + 0.006898684427142143, + 0.010711734183132648, + 0.0044292001985013485, + 0.014026920311152935, + 0.020977849140763283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 6 + ], + "selected_reward": -0.020583007484674454 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, slightly high-angled, static view of a searing black cast iron grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a chef's hand use a wide stainless steel spatula with a black composite handle to flip a thick beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden brioche bun placed on top of the cooking patty to heat and absorb juices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, round, flour-dusted flatbreads placed one by one on the open, oiled section of the grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flatbreads on the grill begin to puff slightly from the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030784457921982 + ], + "opposite_logprob_tokens": [ + -2.3503079414367676 + ], + "logprob_reward": -0.10030784457921982, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030784457921982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06427771598100662, + 0.2120451182126999, + 0.04051393270492554, + 0.013680292293429375, + 0.026798861101269722, + 0.03660866618156433, + 0.0788845419883728, + 0.2854703664779663, + 0.06551457196474075, + 0.17620594799518585 + ], + "credit_map": [ + 0.010520067065954208, + 0.007806332316249609, + 0.004923112690448761, + 0.004579001571983099, + 0.01288224384188652, + 0.008860616944730282, + 0.005541760940104723, + 0.009164584800601006, + 0.0266653373837471, + 0.03152354806661606, + 0.029897967353463173, + 0.03251831978559494, + 0.017601624131202698, + 0.024447232484817505, + 0.020425546914339066, + 0.028965555131435394, + 0.00280548632144928, + 0.0026279210578650236, + 0.008756381459534168, + 0.004852084908634424, + 0.0028191336896270514, + 0.006725340615957975, + 0.0033006395678967237, + 0.00862694438546896, + 0.0013986013364046812, + 0.0011443918338045478, + 0.0016687267925590277, + 0.0011621725279837847, + 0.0017804568633437157, + 0.0031272040214389563, + 0.0013458726461976767, + 0.002052866853773594, + 0.0033087965566664934, + 0.002677141223102808, + 0.0029780494514852762, + 0.0015558251179754734, + 0.0024049216881394386, + 0.005277525167912245, + 0.003357000881806016, + 0.005239601247012615, + 0.004028534982353449, + 0.0039919354021549225, + 0.0034216211643069983, + 0.005451251287013292, + 0.004319943021982908, + 0.005319241899996996, + 0.004848949611186981, + 0.005227190908044577, + 0.008456268347799778, + 0.007859837263822556, + 0.0063529247418046, + 0.005450247786939144, + 0.006442193407565355, + 0.009757647290825844, + 0.021075746044516563, + 0.013489682227373123, + 0.022389965131878853, + 0.029871877282857895, + 0.034173596650362015, + 0.04386836290359497, + 0.033421676605939865, + 0.024703292176127434, + 0.05143623426556587, + 0.0456053726375103, + 0.00685633160173893, + 0.007353873457759619, + 0.0046183522790670395, + 0.005772983655333519, + 0.01004528533667326, + 0.005854226183146238, + 0.01636122353374958, + 0.008652296848595142, + 0.010717766359448433, + 0.010405952110886574, + 0.05831865221261978, + 0.012708950787782669, + 0.016960352659225464, + 0.006901544984430075, + 0.01583280600607395, + 0.044359926134347916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 9 + ], + "selected_reward": -0.10030784457921982 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, slightly high-angled, static shot captures the delicate movements of a searing black cast iron grill, focused by the kitchen lights. A chef's hands, clean and expertly executed, firmly grasp a wide stainless steel spatula with a black composite handle. With swift, decisive precision, the hand slides the spatula beneath the thick, glistening beef patty, revealing dark, crisscrossed charcoal. Then, with a deft flick of the wrist, it flips over, releasing a puff of fragrant steam. The hand then gently places the spatula on the adjacent stainless steel counter, making a soft clinking sound. Without hesitation, the same hand turns to the soft, golden brioche bun, methodically placing the top half directly onto the cooking patty to heat and absorb the juices. Immediately afterwards, the same hand deftly removes several thin, round, flour-dusted flatbreads from a nearby pile and carefully places them, one by one, on the open, oiled section of the grill, where they begin to puff slightly from the heat, creating an atmosphere of focus, expertise, and deeply appetizing flavor.", + "seed": 30411, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_0/index_4474_seed_30411_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04798878077417612, + "vlm_logprob": -0.04798878077417612, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.04798878077417612, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, slightly high-angled, static view of a searing black cast iron grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a chef's hand use a wide stainless steel spatula with a black composite handle to flip a thick beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden brioche bun placed on top of the cooking patty to heat and absorb juices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, round, flour-dusted flatbreads placed one by one on the open, oiled section of the grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flatbreads on the grill begin to puff slightly from the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010587592609226704 + ], + "opposite_logprob_tokens": [ + -5.385587692260742 + ], + "logprob_reward": -0.010587592609226704, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.010587592609226704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18585360050201416, + 0.05433651804924011, + 0.37941041588783264, + 0.12330308556556702, + 0.03168958052992821, + 0.03651728108525276, + 0.024620892480015755, + 0.03219638764858246, + 0.06229249760508537, + 0.06977973878383636 + ], + "credit_map": [ + 0.00995287299156189, + 0.02567915804684162, + 0.030209681019186974, + 0.023871861398220062, + 0.023284178227186203, + 0.029115594923496246, + 0.019884128123521805, + 0.02385612763464451, + 0.004463057965040207, + 0.0064189983531832695, + 0.011144597083330154, + 0.005067400634288788, + 0.008563566952943802, + 0.005387941841036081, + 0.003885810263454914, + 0.009405144490301609, + 0.01664857007563114, + 0.030378643423318863, + 0.039403434842824936, + 0.03975921496748924, + 0.06382685899734497, + 0.07670017331838608, + 0.05802371725440025, + 0.054669808596372604, + 0.005961192771792412, + 0.012213686481118202, + 0.013576335273683071, + 0.014501940459012985, + 0.01746959425508976, + 0.018014878034591675, + 0.01726663112640381, + 0.024298829957842827, + 0.0017692090477794409, + 0.0021584148053079844, + 0.005887728650122881, + 0.0041127330623567104, + 0.0037358200643211603, + 0.004957288969308138, + 0.003289497923105955, + 0.005778886843472719, + 0.0021902681328356266, + 0.0025366011541336775, + 0.0074300337582826614, + 0.0040115900337696075, + 0.0037784494925290346, + 0.005577703006565571, + 0.0037576393224298954, + 0.007234996650367975, + 0.0016885811928659678, + 0.002208071295171976, + 0.006778113078325987, + 0.002610674826428294, + 0.002621662337332964, + 0.002125250408425927, + 0.0017717217560857534, + 0.004816817585378885, + 0.0011966363526880741, + 0.0019287419272586703, + 0.005538663826882839, + 0.004045091103762388, + 0.0034715370275080204, + 0.0036766191478818655, + 0.006739573087543249, + 0.005599526222795248, + 0.0021609009709209204, + 0.009370838291943073, + 0.010068235918879509, + 0.006300424225628376, + 0.009211055003106594, + 0.009525460191071033, + 0.005568850319832563, + 0.010086729191243649, + 0.002829608740285039, + 0.005258549470454454, + 0.014815208502113819, + 0.010693425312638283, + 0.007668328005820513, + 0.0075934105552732944, + 0.008172951638698578, + 0.012748251669108868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.010587592609226704 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, slightly high-angled, static view of a searing black cast iron grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a chef's hand use a wide stainless steel spatula with a black composite handle to flip a thick beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden brioche bun placed on top of the cooking patty to heat and absorb juices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, round, flour-dusted flatbreads placed one by one on the open, oiled section of the grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flatbreads on the grill begin to puff slightly from the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373682498931885 + ], + "opposite_logprob_tokens": [ + -3.4087367057800293 + ], + "logprob_reward": -0.03373682498931885, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373682498931885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12798576056957245, + 0.06130804121494293, + 0.25465697050094604, + 0.024794353172183037, + 0.032102178782224655, + 0.02366511896252632, + 0.040226440876722336, + 0.02807416208088398, + 0.0979454517364502, + 0.30924156308174133 + ], + "credit_map": [ + 0.005817235913127661, + 0.01600709557533264, + 0.018336042761802673, + 0.015778692439198494, + 0.016127821058034897, + 0.028838658705353737, + 0.010552226565778255, + 0.016527989879250526, + 0.003214494325220585, + 0.0032136435620486736, + 0.01168131735175848, + 0.010625418275594711, + 0.005909656174480915, + 0.005275568459182978, + 0.011065575294196606, + 0.010322369635105133, + 0.011304412968456745, + 0.020286966115236282, + 0.029332803562283516, + 0.02477724850177765, + 0.043147820979356766, + 0.051441680639982224, + 0.03695507347583771, + 0.03741094470024109, + 0.0015539521118625998, + 0.0028170838486403227, + 0.004789240192621946, + 0.0029546543955802917, + 0.0026709609664976597, + 0.00352474395185709, + 0.00239346781745553, + 0.0040902504697442055, + 0.0016546028200536966, + 0.0017354132141917944, + 0.004223833326250315, + 0.007354408036917448, + 0.0026364633813500404, + 0.0034713868517428637, + 0.005588631611317396, + 0.005437443032860756, + 0.0012785975122824311, + 0.001805263222195208, + 0.003318755654618144, + 0.0042090024799108505, + 0.0026221773587167263, + 0.0025379774160683155, + 0.003822009777650237, + 0.004071335773915052, + 0.0019509216072037816, + 0.0024339326191693544, + 0.008235259912908077, + 0.005142873153090477, + 0.005385075230151415, + 0.0053306762129068375, + 0.0050662606954574585, + 0.006681440863758326, + 0.001575850648805499, + 0.001688862219452858, + 0.009210589341819286, + 0.002122160978615284, + 0.0015361414989456534, + 0.003214349737390876, + 0.0024661244824528694, + 0.006260082591325045, + 0.0033747730776667595, + 0.013612328097224236, + 0.021593444049358368, + 0.008604902774095535, + 0.01296208519488573, + 0.013782218098640442, + 0.008120608516037464, + 0.015895098447799683, + 0.011428005993366241, + 0.03477856144309044, + 0.0829259604215622, + 0.027657410129904747, + 0.033326100558042526, + 0.037487804889678955, + 0.03021707944571972, + 0.05142064392566681 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.03373682498931885 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, slightly high-angled, static view of a searing black cast iron grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a chef's hand use a wide stainless steel spatula with a black composite handle to flip a thick beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden brioche bun placed on top of the cooking patty to heat and absorb juices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, round, flour-dusted flatbreads placed one by one on the open, oiled section of the grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flatbreads on the grill begin to puff slightly from the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060507982969284 + ], + "opposite_logprob_tokens": [ + -3.8956050872802734 + ], + "logprob_reward": -0.02060507982969284, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060507982969284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09389795362949371, + 0.04160566255450249, + 0.2219572365283966, + 0.10500673949718475, + 0.06104189157485962, + 0.032907549291849136, + 0.07600908726453781, + 0.07276930660009384, + 0.12896142899990082, + 0.1658431738615036 + ], + "credit_map": [ + 0.005074300803244114, + 0.011347630992531776, + 0.014251327142119408, + 0.013920940458774567, + 0.010874561034142971, + 0.01381231751292944, + 0.011927991174161434, + 0.012688891030848026, + 0.0029517384245991707, + 0.004485391080379486, + 0.00852535106241703, + 0.005172139964997768, + 0.004777559079229832, + 0.004043778870254755, + 0.004062580410391092, + 0.007587124593555927, + 0.010712848976254463, + 0.01897955872118473, + 0.025615131482481956, + 0.021763043478131294, + 0.03494028002023697, + 0.04334786534309387, + 0.03365504741668701, + 0.03294345363974571, + 0.0056066494435071945, + 0.013752804137766361, + 0.011884169653058052, + 0.012444625608623028, + 0.008490457199513912, + 0.015598809346556664, + 0.010346137918531895, + 0.026883086189627647, + 0.0028796601109206676, + 0.004595245700329542, + 0.008518718183040619, + 0.010884666815400124, + 0.008387412875890732, + 0.007603035774081945, + 0.009151021018624306, + 0.009022128768265247, + 0.0026745584327727556, + 0.0026199216954410076, + 0.006253550294786692, + 0.0033464781008660793, + 0.004248538054525852, + 0.005378650035709143, + 0.002638362580910325, + 0.005747491028159857, + 0.004060747101902962, + 0.005592676345258951, + 0.011527332477271557, + 0.012802764773368835, + 0.009198701940476894, + 0.00960447546094656, + 0.010992438532412052, + 0.012229946441948414, + 0.002666375134140253, + 0.006863473914563656, + 0.007321774028241634, + 0.009672220796346664, + 0.016018353402614594, + 0.009370476938784122, + 0.009083740413188934, + 0.011772887781262398, + 0.004428898449987173, + 0.022850699722766876, + 0.025619763880968094, + 0.013334158807992935, + 0.016078941524028778, + 0.01729908585548401, + 0.008608403615653515, + 0.02074149250984192, + 0.007316387724131346, + 0.015779243782162666, + 0.0320199579000473, + 0.02065039612352848, + 0.021622639149427414, + 0.023461027070879936, + 0.0168777983635664, + 0.02811571955680847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 8 + ], + "selected_reward": -0.02060507982969284 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, slightly high-angled, static view of a searing black cast iron grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a chef's hand use a wide stainless steel spatula with a black composite handle to flip a thick beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden brioche bun placed on top of the cooking patty to heat and absorb juices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, round, flour-dusted flatbreads placed one by one on the open, oiled section of the grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flatbreads on the grill begin to puff slightly from the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206374615430832 + ], + "opposite_logprob_tokens": [ + -2.812063694000244 + ], + "logprob_reward": -0.06206374615430832, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206374615430832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07463295757770538, + 0.05196596682071686, + 0.37137073278427124, + 0.10436823964118958, + 0.08333741873502731, + 0.03882336989045143, + 0.030903875827789307, + 0.11442418396472931, + 0.058916885405778885, + 0.07125639915466309 + ], + "credit_map": [ + 0.0034403640311211348, + 0.004765210673213005, + 0.01653217524290085, + 0.013399451039731503, + 0.006400889717042446, + 0.006967176683247089, + 0.010990659706294537, + 0.012137027457356453, + 0.00271576177328825, + 0.0035606336314231157, + 0.0074228099547326565, + 0.009246718138456345, + 0.0066709271632134914, + 0.006094636395573616, + 0.0083952397108078, + 0.007859238423407078, + 0.015166227705776691, + 0.030337847769260406, + 0.03960033878684044, + 0.0403977632522583, + 0.06313911825418472, + 0.07653523236513138, + 0.05364168435335159, + 0.052552517503499985, + 0.0052425977773964405, + 0.005172512494027615, + 0.015080127865076065, + 0.016748113557696342, + 0.01391485147178173, + 0.01649562641978264, + 0.013703052885830402, + 0.018011363223195076, + 0.005503923166543245, + 0.007717365864664316, + 0.014345125295221806, + 0.01502638403326273, + 0.009891826659440994, + 0.00829674955457449, + 0.00949296448379755, + 0.013063077814877033, + 0.0019129273714497685, + 0.0029574616346508265, + 0.006142118480056524, + 0.004658492282032967, + 0.003662524279206991, + 0.007534030359238386, + 0.003436486469581723, + 0.008519330993294716, + 0.0016876125009730458, + 0.002352343173697591, + 0.004427789244800806, + 0.005097184795886278, + 0.00432214792817831, + 0.003953078296035528, + 0.0038127524312585592, + 0.005250968039035797, + 0.004506927914917469, + 0.009041077457368374, + 0.01268770918250084, + 0.016394905745983124, + 0.02117058075964451, + 0.015495785512030125, + 0.014984460547566414, + 0.02014273963868618, + 0.0030490290373563766, + 0.008774468675255775, + 0.008235852234065533, + 0.00632941210642457, + 0.010715348646044731, + 0.007622716017067432, + 0.005889270454645157, + 0.008300790563225746, + 0.003200843930244446, + 0.00391271710395813, + 0.019136326387524605, + 0.009290083311498165, + 0.005923187360167503, + 0.007926997728645802, + 0.010212224908173084, + 0.011654016561806202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 3 + ], + "selected_reward": -0.06206374615430832 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, slightly high-angled, static view of a searing black cast iron grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a chef's hand use a wide stainless steel spatula with a black composite handle to flip a thick beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden brioche bun placed on top of the cooking patty to heat and absorb juices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, round, flour-dusted flatbreads placed one by one on the open, oiled section of the grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flatbreads on the grill begin to puff slightly from the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295066028833389 + ], + "opposite_logprob_tokens": [ + -2.237950563430786 + ], + "logprob_reward": -0.11295066028833389, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295066028833389, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11060703545808792, + 0.058521050959825516, + 0.5117409825325012, + 0.02891356311738491, + 0.048930998891592026, + 0.030088545754551888, + 0.03685309737920761, + 0.04448371380567551, + 0.05603056773543358, + 0.07383046299219131 + ], + "credit_map": [ + 0.006231472361832857, + 0.010208982974290848, + 0.021483460441231728, + 0.016767842695116997, + 0.011323370039463043, + 0.011143079027533531, + 0.016589131206274033, + 0.016859695315361023, + 0.0031793222296983004, + 0.004711475223302841, + 0.011950844898819923, + 0.008489123545587063, + 0.0061430903151631355, + 0.0056759933941066265, + 0.00781330931931734, + 0.010557892732322216, + 0.02135874517261982, + 0.04054844379425049, + 0.05345498397946358, + 0.05235348269343376, + 0.08942562341690063, + 0.1071852296590805, + 0.07346987724304199, + 0.0739445686340332, + 0.0017294756835326552, + 0.0019807033240795135, + 0.005712566897273064, + 0.003934322390705347, + 0.002172526903450489, + 0.004642745479941368, + 0.003136520739644766, + 0.005604701582342386, + 0.0034652138128876686, + 0.0036118896678090096, + 0.007993748411536217, + 0.009242629632353783, + 0.00491808308288455, + 0.004967421293258667, + 0.007241059094667435, + 0.007490954827517271, + 0.0019965292885899544, + 0.0020190102513879538, + 0.0078106350265443325, + 0.0018778154626488686, + 0.0035150055773556232, + 0.0056260148994624615, + 0.001807735301554203, + 0.005435799714177847, + 0.001967241521924734, + 0.002431496512144804, + 0.00869312696158886, + 0.004903711844235659, + 0.0035715403500944376, + 0.003949692007154226, + 0.004641232546418905, + 0.006695054937154055, + 0.0015171434497460723, + 0.003371923929080367, + 0.006297567393630743, + 0.007234128657728434, + 0.007529853843152523, + 0.005055328365415335, + 0.006136194337159395, + 0.007341573014855385, + 0.0024913009256124496, + 0.005684640724211931, + 0.006704785395413637, + 0.005989712197333574, + 0.010560490190982819, + 0.011056015267968178, + 0.004087265580892563, + 0.00945635698735714, + 0.003630864666774869, + 0.006587360519915819, + 0.012226426973938942, + 0.012609959580004215, + 0.007552525494247675, + 0.009192409925162792, + 0.009449878707528114, + 0.012581042014062405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.11295066028833389 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_1/index_tensor([4980, 4980])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_1/index_tensor([4980, 4980])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..bb78f77bccc09cc0f57db28a1d1a49d239d9c94f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_1/index_tensor([4980, 4980])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, wide-angle shot, the camera slowly pans right along the riverbank, capturing the majestic grandeur of the canyon at a high frame rate. The entire scene is rendered with the style of a delicate picture book. The towering canyon walls are a cascade of colors painted with soft brushstrokes, transitioning from warm terracotta to deep ochre. The outlines are simplified and expressive, and under the even, bright, diffuse light, there is not a single harsh shadow. At the bottom of the canyon, a colorful river rushes. The water is not a single blue, but a fantastical mixture of turquoise, sapphire, and lavender, shimmering with a warm luster under the soft light. The powerful current surges forward with unstoppable force, violently impacting the rounded boulders in the riverbed and stirring up massive whitewater waves. Thanks to the high frame rate's clear capture, each splash is broken down into countless crystalline individual droplets, which trace clear, smooth, and graceful arcs in the air before gracefully falling back onto the churning waters. The entire process is imbued with an exquisite sense of dynamic beauty. This extreme dynamic clarity contrasts sharply with the static aesthetic of the picture book-like background, creating a unique visual experience that is both imbued with powerful natural movement and imbued with the tranquility and beauty of a fairytale.", + "seed": 26508, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_1/index_4980_seed_26508_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.19278621478006244, + "vlm_logprob": -0.19278621478006244, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.19278621478006244, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right along the riverbank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the canyon walls rendered with simplified outlines and soft brushstrokes in a picture book style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river water a mixture of turquoise, sapphire, and lavender colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the water splashes captured as individual crystalline droplets tracing smooth arcs in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07455834001302719 + ], + "opposite_logprob_tokens": [ + -2.6995582580566406 + ], + "logprob_reward": -0.07455834001302719, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07455834001302719, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7674217224121094, + 0.03633666783571243, + 0.06503914296627045, + 0.016432352364063263, + 0.030817991122603416, + 0.009086959064006805, + 0.015809940174221992, + 0.019446657970547676, + 0.020362213253974915, + 0.019246336072683334 + ], + "credit_map": [ + 0.02344268187880516, + 0.02189633622765541, + 0.024242714047431946, + 0.32053741812705994, + 0.023153871297836304, + 0.017165319994091988, + 0.31281062960624695, + 0.02417278289794922, + 0.0034230751916766167, + 0.0031794572714716196, + 0.0038158632814884186, + 0.005433852318674326, + 0.0036589899100363255, + 0.0031167324632406235, + 0.01034451276063919, + 0.003364181611686945, + 0.010192341171205044, + 0.006539091933518648, + 0.009237540885806084, + 0.007973422296345234, + 0.0068407137878239155, + 0.005614034365862608, + 0.011021449230611324, + 0.007620546966791153, + 0.0022471463307738304, + 0.002332025673240423, + 0.002181223127990961, + 0.0019212191691622138, + 0.0022543384693562984, + 0.001581557677127421, + 0.002288978546857834, + 0.001625863485969603, + 0.003574630944058299, + 0.0049198828637599945, + 0.004020129796117544, + 0.0038524519186466932, + 0.003992512356489897, + 0.0034346245229244232, + 0.0029276625718921423, + 0.004096096847206354, + 0.0010336500126868486, + 0.0009871938964352012, + 0.0013807936338707805, + 0.0013113727327436209, + 0.0009857512777671218, + 0.0008715435978956521, + 0.0012825684389099479, + 0.0012340854154899716, + 0.002322453074157238, + 0.002098767552524805, + 0.0024463715963065624, + 0.002348249312490225, + 0.0017518256790935993, + 0.0014698085142299533, + 0.0015032149385660887, + 0.001869248109869659, + 0.003209738526493311, + 0.002719441894441843, + 0.0025806783232837915, + 0.0029317724984139204, + 0.0018352364422753453, + 0.00199881661683321, + 0.0016137846978381276, + 0.002557189902290702, + 0.002765534445643425, + 0.002609855029731989, + 0.0031259057577699423, + 0.0034914666321128607, + 0.0018798564560711384, + 0.0019848153460770845, + 0.0018788583111017942, + 0.0026259201113134623, + 0.0023757005110383034, + 0.002217258792370558, + 0.0035516167990863323, + 0.0033289645798504353, + 0.0016085656825453043, + 0.0017930628964677453, + 0.002202532719820738, + 0.0021686346735805273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.07455834001302719 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right along the riverbank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the canyon walls rendered with simplified outlines and soft brushstrokes in a picture book style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river water a mixture of turquoise, sapphire, and lavender colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the water splashes captured as individual crystalline droplets tracing smooth arcs in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261444568634033 + ], + "opposite_logprob_tokens": [ + -0.5761444568634033 + ], + "logprob_reward": -0.8261444568634033, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261444568634033, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.6538513898849487, + 0.03257681056857109, + 0.15029524266719818, + 0.032450173050165176, + 0.031219543889164925, + 0.014010848477482796, + 0.015830999240279198, + 0.02723909541964531, + 0.023373184725642204, + 0.01915266178548336 + ], + "credit_map": [ + 0.02067154087126255, + 0.01826327107846737, + 0.02039160765707493, + 0.2678409516811371, + 0.020477669313549995, + 0.01627976819872856, + 0.2683744430541992, + 0.021552182734012604, + 0.003282880876213312, + 0.0035012357402592897, + 0.004131026100367308, + 0.004940820392221212, + 0.003348275786265731, + 0.0026718578301370144, + 0.006414072122424841, + 0.004286644048988819, + 0.023366758599877357, + 0.015950243920087814, + 0.020271675661206245, + 0.017796535044908524, + 0.017217176035046577, + 0.013633428141474724, + 0.024473031982779503, + 0.017586402595043182, + 0.003947286866605282, + 0.003395380452275276, + 0.00611156364902854, + 0.004463558550924063, + 0.0036461118143051863, + 0.002768164034932852, + 0.004445819649845362, + 0.0036722859367728233, + 0.004034508019685745, + 0.004181692376732826, + 0.005123374052345753, + 0.004154087509959936, + 0.003836997551843524, + 0.0033241675700992346, + 0.0031094825826585293, + 0.0034552353899925947, + 0.0019927348475903273, + 0.0016654932405799627, + 0.002067707944661379, + 0.0015498026041314006, + 0.001981689129024744, + 0.001686831470578909, + 0.0011675730347633362, + 0.0018990172538906336, + 0.0019746895413845778, + 0.0014845600817352533, + 0.003012946341186762, + 0.0021772701293230057, + 0.001456995029002428, + 0.0013972400920465589, + 0.0021966290660202503, + 0.0021306700073182583, + 0.003934614825993776, + 0.0033997504506260157, + 0.003865625476464629, + 0.003753514727577567, + 0.002628133399412036, + 0.003185475943610072, + 0.002795831998810172, + 0.0036761490628123283, + 0.002500622533261776, + 0.002390838460996747, + 0.0041612968780100346, + 0.003334317822009325, + 0.002291747136041522, + 0.0019474312430247664, + 0.003726484952494502, + 0.0030204446520656347, + 0.002264593495056033, + 0.0021039629355072975, + 0.003346291370689869, + 0.0034775796812027693, + 0.0016844384372234344, + 0.001840015291236341, + 0.001954183215275407, + 0.0024815970100462437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.8261444568634033 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right along the riverbank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the canyon walls rendered with simplified outlines and soft brushstrokes in a picture book style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river water a mixture of turquoise, sapphire, and lavender colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the water splashes captured as individual crystalline droplets tracing smooth arcs in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005951777566224337 + ], + "opposite_logprob_tokens": [ + -5.130951881408691 + ], + "logprob_reward": -0.005951777566224337, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005951777566224337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3069998621940613, + 0.04656675085425377, + 0.37505394220352173, + 0.02850584127008915, + 0.059985700994729996, + 0.05107855051755905, + 0.037563908845186234, + 0.036244239658117294, + 0.03189805895090103, + 0.026103146374225616 + ], + "credit_map": [ + 0.013734305277466774, + 0.019173260778188705, + 0.015410320833325386, + 0.11391223967075348, + 0.01804174855351448, + 0.012209203094244003, + 0.1005634069442749, + 0.01395536307245493, + 0.005239979363977909, + 0.005481064319610596, + 0.0064893863163888454, + 0.008423777297139168, + 0.0044112312607467175, + 0.0036108666099607944, + 0.008079119957983494, + 0.004831324331462383, + 0.0534830205142498, + 0.034811995923519135, + 0.05554855614900589, + 0.04761684685945511, + 0.038465820252895355, + 0.032067980617284775, + 0.06778134405612946, + 0.045278400182724, + 0.0033518089912831783, + 0.003750069532543421, + 0.005062831100076437, + 0.00353144109249115, + 0.003177145728841424, + 0.003372570965439081, + 0.003352596191689372, + 0.0029073769692331553, + 0.0069898092187941074, + 0.008630001917481422, + 0.00825542863458395, + 0.0075442492961883545, + 0.007592020556330681, + 0.0067358301021158695, + 0.006390869617462158, + 0.007847492583096027, + 0.006350329145789146, + 0.006669966038316488, + 0.0067188856191933155, + 0.005508510395884514, + 0.007770789321511984, + 0.006345239467918873, + 0.004976117517799139, + 0.00673871161416173, + 0.003884436795488, + 0.0035892280284315348, + 0.006624982226639986, + 0.00599895603954792, + 0.0028823092579841614, + 0.003877823008224368, + 0.00516984099522233, + 0.005536334589123726, + 0.004654917400330305, + 0.004808246623724699, + 0.004570468328893185, + 0.004837004467844963, + 0.00435229204595089, + 0.004454748705029488, + 0.003684684168547392, + 0.004881875589489937, + 0.003284421982243657, + 0.0032090311869978905, + 0.006017843261361122, + 0.004737597890198231, + 0.003316112793982029, + 0.0032779185567051172, + 0.005350410472601652, + 0.0027047230396419764, + 0.0031219571828842163, + 0.0027677922043949366, + 0.005256724078208208, + 0.003829980967566371, + 0.0025632143951952457, + 0.0026000274810940027, + 0.003272686153650284, + 0.002690763445571065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.005951777566224337 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right along the riverbank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the canyon walls rendered with simplified outlines and soft brushstrokes in a picture book style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river water a mixture of turquoise, sapphire, and lavender colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the water splashes captured as individual crystalline droplets tracing smooth arcs in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008635913021862507 + ], + "opposite_logprob_tokens": [ + -4.758635997772217 + ], + "logprob_reward": -0.008635913021862507, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008635913021862507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8244784474372864, + 0.018147608265280724, + 0.05580909922719002, + 0.011638187803328037, + 0.015304910950362682, + 0.010647459886968136, + 0.008543620817363262, + 0.014037087559700012, + 0.02662118710577488, + 0.014772357419133186 + ], + "credit_map": [ + 0.024312403053045273, + 0.021547049283981323, + 0.025472138077020645, + 0.3497159481048584, + 0.022504979744553566, + 0.017379155382514, + 0.33726924657821655, + 0.026277631521224976, + 0.0016502696089446545, + 0.0011672942200675607, + 0.0022492047864943743, + 0.003558251541107893, + 0.0012305061100050807, + 0.0011531503405421972, + 0.004874122329056263, + 0.002264810958877206, + 0.008351169526576996, + 0.0064190155826509, + 0.007861916907131672, + 0.006444432772696018, + 0.006193674169480801, + 0.005507520399987698, + 0.008812571875751019, + 0.006218805443495512, + 0.001223846571519971, + 0.001498165656812489, + 0.0018691420555114746, + 0.0013339215656742454, + 0.0015619341284036636, + 0.00155678391456604, + 0.0015369573375210166, + 0.0010574378538876772, + 0.002139010466635227, + 0.002160395961254835, + 0.002215401502326131, + 0.0017804185627028346, + 0.0018258214695379138, + 0.001611568033695221, + 0.0016584689728915691, + 0.0019138269126415253, + 0.0009971302933990955, + 0.0008480677497573197, + 0.0019547247793525457, + 0.0014702330809086561, + 0.0008354337187483907, + 0.0008868054719641805, + 0.0018776659853756428, + 0.0017773999134078622, + 0.0009699696092866361, + 0.0009016286348924041, + 0.0016707411268725991, + 0.00127055321354419, + 0.00060801807558164, + 0.0007430087425746024, + 0.0010950438445433974, + 0.0012846591416746378, + 0.001751862233504653, + 0.0014941886765882373, + 0.0025873288977891207, + 0.0019714345689862967, + 0.0011139706475660205, + 0.0014099956024438143, + 0.0021602988708764315, + 0.0015480086440220475, + 0.003571368521079421, + 0.0036449364852160215, + 0.0035187501925975084, + 0.0035978583618998528, + 0.002727008191868663, + 0.0027063421439379454, + 0.0026516795624047518, + 0.004203249234706163, + 0.0015372460475191474, + 0.0014504477148875594, + 0.002722079399973154, + 0.002193994354456663, + 0.0011654543923214078, + 0.0013492362340912223, + 0.0018689270364120603, + 0.002484974218532443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.008635913021862507 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right along the riverbank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the canyon walls rendered with simplified outlines and soft brushstrokes in a picture book style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river water a mixture of turquoise, sapphire, and lavender colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the water splashes captured as individual crystalline droplets tracing smooth arcs in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864058643579483 + ], + "opposite_logprob_tokens": [ + -3.048640489578247 + ], + "logprob_reward": -0.04864058643579483, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864058643579483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6840658187866211, + 0.04220564290881157, + 0.08303254097700119, + 0.036071840673685074, + 0.022439805790781975, + 0.035121478140354156, + 0.025430873036384583, + 0.02447129413485527, + 0.019521983340382576, + 0.027638724073767662 + ], + "credit_map": [ + 0.022694436833262444, + 0.02325156144797802, + 0.023763660341501236, + 0.28442373871803284, + 0.022763026878237724, + 0.015289434231817722, + 0.2685092091560364, + 0.023370826616883278, + 0.005675192456692457, + 0.004233663436025381, + 0.006000000983476639, + 0.005988925229758024, + 0.004118151031434536, + 0.003166691865772009, + 0.008490299805998802, + 0.0045327176339924335, + 0.0117027023807168, + 0.010939396917819977, + 0.010619143955409527, + 0.010447260923683643, + 0.012115213088691235, + 0.009732568636536598, + 0.007453411817550659, + 0.010022848844528198, + 0.002933057490736246, + 0.0037809035275131464, + 0.006729631684720516, + 0.004486828111112118, + 0.003675836604088545, + 0.004004510119557381, + 0.005465625319629908, + 0.004995445255190134, + 0.0019480212358757854, + 0.0021710398141294718, + 0.004161575343459845, + 0.003269697306677699, + 0.0016297181136906147, + 0.002234524581581354, + 0.0032513048499822617, + 0.003773924196138978, + 0.0038001947104930878, + 0.0037132869474589825, + 0.006077922414988279, + 0.0055045452900230885, + 0.003511898685246706, + 0.0037481221370399, + 0.004633989185094833, + 0.004131520166993141, + 0.003376429434865713, + 0.0034662901889532804, + 0.003780214348807931, + 0.003966562915593386, + 0.0025849135126918554, + 0.0021458491683006287, + 0.0026924503035843372, + 0.0034181629307568073, + 0.002933705225586891, + 0.0033499717246741056, + 0.0035025596152991056, + 0.003929187078028917, + 0.0032062663231045008, + 0.0026156397070735693, + 0.0028381706215441227, + 0.002095794305205345, + 0.002152313943952322, + 0.0018920989241451025, + 0.003765630768612027, + 0.0032016304321587086, + 0.0016160501400008798, + 0.0016890853876248002, + 0.003282866207882762, + 0.001922308700159192, + 0.002643908141180873, + 0.004088715184479952, + 0.004571360070258379, + 0.00484311580657959, + 0.0028550056740641594, + 0.002542209578678012, + 0.00354405608959496, + 0.0025503523647785187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.04864058643579483 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide-angle shot, the camera slowly pans right along the riverbank, capturing the majestic grandeur of the canyon at a high frame rate. The entire scene is rendered with the style of a delicate picture book. The towering canyon walls are a cascade of colors painted with soft brushstrokes, transitioning from warm terracotta to deep ochre. The outlines are simplified and expressive, and under the even, bright, diffuse light, there is not a single harsh shadow. At the bottom of the canyon, a colorful river rushes. The water is not a single blue, but a fantastical mixture of turquoise, sapphire, and lavender, shimmering with a warm luster under the soft light. The powerful current surges forward with unstoppable force, violently impacting the rounded boulders in the riverbed and stirring up massive whitewater waves. Thanks to the high frame rate's clear capture, each splash is broken down into countless crystalline individual droplets, which trace clear, smooth, and graceful arcs in the air before gracefully falling back onto the churning waters. The entire process is imbued with an exquisite sense of dynamic beauty. This extreme dynamic clarity contrasts sharply with the static aesthetic of the picture book-like background, creating a unique visual experience that is both imbued with powerful natural movement and imbued with the tranquility and beauty of a fairytale.", + "seed": 26509, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_1/index_4980_seed_26509_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.10752648627385497, + "vlm_logprob": -0.10752648627385497, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.10752648627385497, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right along the riverbank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the canyon walls rendered with simplified outlines and soft brushstrokes in a picture book style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river water a mixture of turquoise, sapphire, and lavender colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the water splashes captured as individual crystalline droplets tracing smooth arcs in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02094496227800846 + ], + "opposite_logprob_tokens": [ + -4.270945072174072 + ], + "logprob_reward": -0.02094496227800846, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.02094496227800846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06510542333126068, + 0.12486181408166885, + 0.0236187856644392, + 0.06091730669140816, + 0.17790254950523376, + 0.023760778829455376, + 0.08752816170454025, + 0.1344427466392517, + 0.21650904417037964, + 0.08535332977771759 + ], + "credit_map": [ + 0.005116312764585018, + 0.008726483210921288, + 0.006252377759665251, + 0.006351414602249861, + 0.008962607011198997, + 0.004496951587498188, + 0.008746224455535412, + 0.016453057527542114, + 0.012875411659479141, + 0.014978830702602863, + 0.007998152635991573, + 0.011262577958405018, + 0.007778842467814684, + 0.006281564477831125, + 0.03049445152282715, + 0.033191997557878494, + 0.004303723108023405, + 0.0018991580000147223, + 0.0018873418448492885, + 0.0030662789940834045, + 0.002261882182210684, + 0.0018614826258271933, + 0.005674807354807854, + 0.002664112951606512, + 0.004272822290658951, + 0.005055089946836233, + 0.012400027364492416, + 0.008826162666082382, + 0.0052757179364562035, + 0.0043212734162807465, + 0.006602388806641102, + 0.014163830317556858, + 0.010939601808786392, + 0.03773968294262886, + 0.007513060700148344, + 0.06817197799682617, + 0.0076256198808550835, + 0.003623270895332098, + 0.031286220997571945, + 0.011003151535987854, + 0.0021018676925450563, + 0.003266523825004697, + 0.001628945698030293, + 0.0054405685514211655, + 0.0018207681132480502, + 0.0013908690307289362, + 0.005777792539447546, + 0.0023334468714892864, + 0.009863425977528095, + 0.012732842937111855, + 0.007649545092135668, + 0.01000169850885868, + 0.013922113925218582, + 0.01053616963326931, + 0.011069666594266891, + 0.011752712540328503, + 0.008746599778532982, + 0.03638788312673569, + 0.004490670748054981, + 0.01180937048047781, + 0.018796367570757866, + 0.00564485602080822, + 0.040826696902513504, + 0.007740316446870565, + 0.008411293849349022, + 0.030092928558588028, + 0.010866613127291203, + 0.0826488807797432, + 0.01367093250155449, + 0.0067626433447003365, + 0.05725794658064842, + 0.006797830108553171, + 0.024346280843019485, + 0.01508763525635004, + 0.004623951856046915, + 0.018385274335741997, + 0.007941183634102345, + 0.0030671083368360996, + 0.00728892395272851, + 0.0046129850670695305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.02094496227800846 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right along the riverbank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the canyon walls rendered with simplified outlines and soft brushstrokes in a picture book style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river water a mixture of turquoise, sapphire, and lavender colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the water splashes captured as individual crystalline droplets tracing smooth arcs in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134394884109497 + ], + "opposite_logprob_tokens": [ + -1.3134394884109497 + ], + "logprob_reward": -0.3134394884109497, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134394884109497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05725334212183952, + 0.044246818870306015, + 0.08899487555027008, + 0.036996643990278244, + 0.5978187918663025, + 0.01993459276854992, + 0.024112863466143608, + 0.07173427939414978, + 0.015392831526696682, + 0.04351497441530228 + ], + "credit_map": [ + 0.00577236944809556, + 0.007376395631581545, + 0.003935358952730894, + 0.0074431197717785835, + 0.009827310219407082, + 0.004580192733556032, + 0.007461003493517637, + 0.010857593268156052, + 0.006633571349084377, + 0.004901939537376165, + 0.004680958576500416, + 0.0042474158108234406, + 0.004300955682992935, + 0.002404123079031706, + 0.011073216795921326, + 0.006004637107253075, + 0.018996257334947586, + 0.005447698291391134, + 0.0039802538231015205, + 0.009218174032866955, + 0.006429840810596943, + 0.00447527039796114, + 0.029027916491031647, + 0.011419467628002167, + 0.004003363195806742, + 0.0033793034963309765, + 0.004140816628932953, + 0.005233990494161844, + 0.0034883376210927963, + 0.0024319684598594904, + 0.004976837895810604, + 0.00934202503412962, + 0.05172983184456825, + 0.15209317207336426, + 0.023256244137883186, + 0.20568343997001648, + 0.030828697606921196, + 0.01487753540277481, + 0.08652884513139725, + 0.03282102197408676, + 0.004694145172834396, + 0.0077406540513038635, + 0.0007746462360955775, + 0.0010865969816222787, + 0.0020116076339036226, + 0.0009379875264130533, + 0.0018110598903149366, + 0.000877894985023886, + 0.002622873755171895, + 0.0032116654329001904, + 0.0017127060564234853, + 0.0030887287575751543, + 0.003451436758041382, + 0.0031389438081532717, + 0.002609046408906579, + 0.004277462139725685, + 0.004977488424628973, + 0.016379643231630325, + 0.0029466599225997925, + 0.006863428745418787, + 0.010564073920249939, + 0.003687971970066428, + 0.02178088203072548, + 0.0045341383665800095, + 0.00178140786010772, + 0.003143428359180689, + 0.0014297724701464176, + 0.0023396755568683147, + 0.001828245003707707, + 0.0016129720024764538, + 0.001685830415226519, + 0.001571498578414321, + 0.006311493460088968, + 0.013154620304703712, + 0.004176351707428694, + 0.0043027400970458984, + 0.007257804740220308, + 0.0019688827451318502, + 0.004131132271140814, + 0.002211946062743664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 7 + ], + "selected_reward": -0.3134394884109497 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right along the riverbank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the canyon walls rendered with simplified outlines and soft brushstrokes in a picture book style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river water a mixture of turquoise, sapphire, and lavender colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the water splashes captured as individual crystalline droplets tracing smooth arcs in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028203255496919155 + ], + "opposite_logprob_tokens": [ + -5.8778204917907715 + ], + "logprob_reward": -0.0028203255496919155, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028203255496919155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13141275942325592, + 0.03234829008579254, + 0.19920501112937927, + 0.09102939814329147, + 0.08560259640216827, + 0.03068026900291443, + 0.07242260128259659, + 0.022653359919786453, + 0.1992873102426529, + 0.13535840809345245 + ], + "credit_map": [ + 0.01084880344569683, + 0.01570393517613411, + 0.010511520318686962, + 0.021323779597878456, + 0.02073802798986435, + 0.010167386382818222, + 0.020169071853160858, + 0.021950237452983856, + 0.003895119996741414, + 0.00491007324308157, + 0.003194127231836319, + 0.004400656558573246, + 0.00405321316793561, + 0.002039753831923008, + 0.004651293624192476, + 0.0052040498703718185, + 0.04717661067843437, + 0.01056691538542509, + 0.007474821526557207, + 0.018793093040585518, + 0.01457203272730112, + 0.008359121158719063, + 0.06730562448501587, + 0.02495678700506687, + 0.006818393245339394, + 0.009303548373281956, + 0.01034380029886961, + 0.01606118120253086, + 0.01063512172549963, + 0.007619510404765606, + 0.014806213788688183, + 0.015441632829606533, + 0.009661060757935047, + 0.021825842559337616, + 0.004396194592118263, + 0.02368687093257904, + 0.006067825946956873, + 0.0026915480848401785, + 0.011138959787786007, + 0.006134302821010351, + 0.005755553022027016, + 0.009647074155509472, + 0.0018395904917269945, + 0.002263509202748537, + 0.003445184091106057, + 0.0026201317086815834, + 0.003083419520407915, + 0.00202580657787621, + 0.008307495154440403, + 0.010725230909883976, + 0.007410983555018902, + 0.008410642854869366, + 0.011466323398053646, + 0.008053217083215714, + 0.009484346956014633, + 0.008564361371099949, + 0.0024792603217065334, + 0.0046065496280789375, + 0.0013807734940201044, + 0.0017253855476155877, + 0.0036613813135772943, + 0.0034845396876335144, + 0.003001469885930419, + 0.00231400178745389, + 0.0069901105016469955, + 0.03052959404885769, + 0.00935746356844902, + 0.07397140562534332, + 0.01309872791171074, + 0.006414113100618124, + 0.05224798247218132, + 0.006677928846329451, + 0.02697942405939102, + 0.014700336381793022, + 0.008859912864863873, + 0.02500731311738491, + 0.01326531358063221, + 0.005686192773282528, + 0.03138086944818497, + 0.009479052387177944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 9 + ], + "selected_reward": -0.0028203255496919155 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right along the riverbank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the canyon walls rendered with simplified outlines and soft brushstrokes in a picture book style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river water a mixture of turquoise, sapphire, and lavender colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the water splashes captured as individual crystalline droplets tracing smooth arcs in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057156339287758 + ], + "opposite_logprob_tokens": [ + -3.89557147026062 + ], + "logprob_reward": -0.02057156339287758, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057156339287758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08761008828878403, + 0.04659882187843323, + 0.054123491048812866, + 0.04100026935338974, + 0.03027157299220562, + 0.05016596242785454, + 0.04320095106959343, + 0.08752527832984924, + 0.2294165939092636, + 0.3300870358943939 + ], + "credit_map": [ + 0.007750650867819786, + 0.010915365070104599, + 0.007361702620983124, + 0.009319336153566837, + 0.011237971484661102, + 0.008065561763942242, + 0.01404409110546112, + 0.018915394321084023, + 0.007057976443320513, + 0.004506359342485666, + 0.004003547132015228, + 0.006031604018062353, + 0.004490061197429895, + 0.004638059996068478, + 0.005136318504810333, + 0.010734887793660164, + 0.009748812764883041, + 0.005036130081862211, + 0.003049680031836033, + 0.006608270574361086, + 0.00706273689866066, + 0.003944584634155035, + 0.013092921115458012, + 0.005580350290983915, + 0.002942607505246997, + 0.004672388080507517, + 0.004283555783331394, + 0.005513378884643316, + 0.007182813249528408, + 0.004249973222613335, + 0.005328739993274212, + 0.0068268077448010445, + 0.003019683761522174, + 0.008594883605837822, + 0.001677224994637072, + 0.0077800084836781025, + 0.0029252765234559774, + 0.0013150464510545135, + 0.002687857486307621, + 0.00227158865891397, + 0.0036296448670327663, + 0.009115181863307953, + 0.0038069456350058317, + 0.009222868829965591, + 0.004202560056000948, + 0.0028134516905993223, + 0.0141353914514184, + 0.003239911515265703, + 0.004631110467016697, + 0.00563840800896287, + 0.004249142482876778, + 0.006984614301472902, + 0.0055197481997311115, + 0.005021141842007637, + 0.004490670748054981, + 0.006666108034551144, + 0.006593657657504082, + 0.02263503149151802, + 0.0036781448870897293, + 0.006695335730910301, + 0.012024447321891785, + 0.006356069818139076, + 0.021312760189175606, + 0.008229823783040047, + 0.010194783098995686, + 0.03478945419192314, + 0.012086371891200542, + 0.0828302875161171, + 0.01622881181538105, + 0.007944985292851925, + 0.05644572153687477, + 0.008896159939467907, + 0.106193408370018, + 0.07954443991184235, + 0.014295253902673721, + 0.016412822529673576, + 0.03616057336330414, + 0.00942651554942131, + 0.05671314522624016, + 0.011340854689478874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.02057156339287758 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right along the riverbank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the canyon walls rendered with simplified outlines and soft brushstrokes in a picture book style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river water a mixture of turquoise, sapphire, and lavender colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the water splashes captured as individual crystalline droplets tracing smooth arcs in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798560917377472 + ], + "opposite_logprob_tokens": [ + -1.8048560619354248 + ], + "logprob_reward": -0.1798560917377472, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798560917377472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0803925171494484, + 0.05236439406871796, + 0.13556048274040222, + 0.03533579036593437, + 0.041378989815711975, + 0.06487033516168594, + 0.04951076954603195, + 0.28091952204704285, + 0.1501755565404892, + 0.10949167609214783 + ], + "credit_map": [ + 0.006513145286589861, + 0.009201470762491226, + 0.004896522033959627, + 0.009742663241922855, + 0.008010153658688068, + 0.006361787207424641, + 0.013875185512006283, + 0.02179158292710781, + 0.012419778853654861, + 0.007001393008977175, + 0.004559350665658712, + 0.004789809230715036, + 0.006070986855775118, + 0.004903491120785475, + 0.006842587608844042, + 0.005776994861662388, + 0.0322083979845047, + 0.006654917262494564, + 0.005618195980787277, + 0.012631328776478767, + 0.009434562176465988, + 0.0055189961567521095, + 0.04594220593571663, + 0.01755187101662159, + 0.003735697129741311, + 0.003941055852919817, + 0.005612686276435852, + 0.005118417087942362, + 0.004613328259438276, + 0.0025084668304771185, + 0.0041890195570886135, + 0.005617116112262011, + 0.002050461946055293, + 0.0021060053259134293, + 0.001988430740311742, + 0.015281083062291145, + 0.0025561319198459387, + 0.0018089365912601352, + 0.012540377676486969, + 0.003047561040148139, + 0.012266561388969421, + 0.025010917335748672, + 0.002184404293075204, + 0.003108769888058305, + 0.0061898562125861645, + 0.002883733483031392, + 0.010269307531416416, + 0.0029567803721874952, + 0.005373694933950901, + 0.007765381596982479, + 0.0044395518489181995, + 0.006558144465088844, + 0.0066340528428554535, + 0.004795207176357508, + 0.005992670077830553, + 0.007952064275741577, + 0.01723588816821575, + 0.07430597394704819, + 0.006996560841798782, + 0.025028161704540253, + 0.03998285531997681, + 0.012156588025391102, + 0.08818788826465607, + 0.01702558435499668, + 0.005719217471778393, + 0.024925118312239647, + 0.008485431782901287, + 0.05301966518163681, + 0.01086566410958767, + 0.0051942546851933, + 0.036689601838588715, + 0.005276591517031193, + 0.03316981717944145, + 0.032211676239967346, + 0.005398258101195097, + 0.006787408608943224, + 0.01502935029566288, + 0.0038390306290239096, + 0.010411899536848068, + 0.002644235035404563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 2 + ], + "selected_reward": -0.1798560917377472 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide-angle shot, the camera slowly pans right along the riverbank, capturing the majestic grandeur of the canyon at a high frame rate. The entire scene is rendered with the style of a delicate picture book. The towering canyon walls are a cascade of colors painted with soft brushstrokes, transitioning from warm terracotta to deep ochre. The outlines are simplified and expressive, and under the even, bright, diffuse light, there is not a single harsh shadow. At the bottom of the canyon, a colorful river rushes. The water is not a single blue, but a fantastical mixture of turquoise, sapphire, and lavender, shimmering with a warm luster under the soft light. The powerful current surges forward with unstoppable force, violently impacting the rounded boulders in the riverbed and stirring up massive whitewater waves. Thanks to the high frame rate's clear capture, each splash is broken down into countless crystalline individual droplets, which trace clear, smooth, and graceful arcs in the air before gracefully falling back onto the churning waters. The entire process is imbued with an exquisite sense of dynamic beauty. This extreme dynamic clarity contrasts sharply with the static aesthetic of the picture book-like background, creating a unique visual experience that is both imbued with powerful natural movement and imbued with the tranquility and beauty of a fairytale.", + "seed": 26510, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_1/index_4980_seed_26510_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07552619606722147, + "vlm_logprob": -0.07552619606722147, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.07552619606722147, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right along the riverbank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the canyon walls rendered with simplified outlines and soft brushstrokes in a picture book style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river water a mixture of turquoise, sapphire, and lavender colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the water splashes captured as individual crystalline droplets tracing smooth arcs in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018736518919467926 + ], + "opposite_logprob_tokens": [ + -4.518736362457275 + ], + "logprob_reward": -0.018736518919467926, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.018736518919467926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44149306416511536, + 0.03544111177325249, + 0.04044552519917488, + 0.2223922461271286, + 0.04968603700399399, + 0.0449485145509243, + 0.03326456993818283, + 0.07341176271438599, + 0.031324513256549835, + 0.02759268507361412 + ], + "credit_map": [ + 0.05681143328547478, + 0.06797707825899124, + 0.05244547128677368, + 0.04387957602739334, + 0.04456253722310066, + 0.04776611551642418, + 0.0837286114692688, + 0.044322237372398376, + 0.00546607468277216, + 0.004861267749220133, + 0.0051166401244699955, + 0.0034612715244293213, + 0.003677290165796876, + 0.003994578029960394, + 0.003861508797854185, + 0.0050024790689349174, + 0.005574275739490986, + 0.003752616001293063, + 0.005980797111988068, + 0.006183347664773464, + 0.003156191436573863, + 0.0039365594275295734, + 0.0055985404178500175, + 0.006263194140046835, + 0.017682582139968872, + 0.018205255270004272, + 0.024729259312152863, + 0.036744046956300735, + 0.016236480325460434, + 0.024104520678520203, + 0.05790604650974274, + 0.02678404375910759, + 0.004692633170634508, + 0.004428742919117212, + 0.005807588808238506, + 0.003989944700151682, + 0.010564901866018772, + 0.010008729994297028, + 0.0049548023380339146, + 0.00523869413882494, + 0.003527985420078039, + 0.0040371594950556755, + 0.012567833997309208, + 0.004154416266828775, + 0.011278090998530388, + 0.00405624695122242, + 0.0025706421583890915, + 0.0027561390306800604, + 0.0028575428295880556, + 0.003606364829465747, + 0.004168159794062376, + 0.0032489581499248743, + 0.00425553135573864, + 0.007532583549618721, + 0.003976935986429453, + 0.0036184911150485277, + 0.007933266460895538, + 0.01718314178287983, + 0.006460977718234062, + 0.018362276256084442, + 0.005353735759854317, + 0.007013349328190088, + 0.0068925232626497746, + 0.004212495870888233, + 0.003489559516310692, + 0.005158430431038141, + 0.0029391131829470396, + 0.00488756550475955, + 0.005140206310898066, + 0.005782422609627247, + 0.0020975221414119005, + 0.001829691813327372, + 0.0021226138342171907, + 0.0021538170985877514, + 0.0025678446982055902, + 0.006293770857155323, + 0.006198767572641373, + 0.0025653087068349123, + 0.0030853368807584047, + 0.002605224959552288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.018736518919467926 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right along the riverbank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the canyon walls rendered with simplified outlines and soft brushstrokes in a picture book style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river water a mixture of turquoise, sapphire, and lavender colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the water splashes captured as individual crystalline droplets tracing smooth arcs in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25209951400756836 + ], + "opposite_logprob_tokens": [ + -1.5020995140075684 + ], + "logprob_reward": -0.25209951400756836, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25209951400756836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3034292161464691, + 0.04484200105071068, + 0.046208009123802185, + 0.21739257872104645, + 0.052317045629024506, + 0.039235975593328476, + 0.04237450659275055, + 0.12362469732761383, + 0.10728597640991211, + 0.023289939388632774 + ], + "credit_map": [ + 0.04660268500447273, + 0.041027892380952835, + 0.03529014065861702, + 0.03230234235525131, + 0.0266598928719759, + 0.029104266315698624, + 0.06724972277879715, + 0.02519230544567108, + 0.008598304353654385, + 0.006071906071156263, + 0.005154786165803671, + 0.004859919659793377, + 0.0036304001696407795, + 0.004841960500925779, + 0.006700723432004452, + 0.004984007216989994, + 0.0035380234476178885, + 0.005203539505600929, + 0.004516783636063337, + 0.008268974721431732, + 0.005057745147496462, + 0.005350059363991022, + 0.009504683315753937, + 0.004768204875290394, + 0.01657995767891407, + 0.021344315260648727, + 0.022148311138153076, + 0.032437827438116074, + 0.019171487540006638, + 0.023785002529621124, + 0.04757419973611832, + 0.03435150906443596, + 0.004185905214399099, + 0.00716479541733861, + 0.0075760832987725735, + 0.004001276101917028, + 0.011764861643314362, + 0.009661898016929626, + 0.004111503250896931, + 0.0038507278077304363, + 0.0034575413446873426, + 0.0034441850148141384, + 0.0039113848470151424, + 0.003227618522942066, + 0.010495942085981369, + 0.007677158340811729, + 0.0035936080384999514, + 0.0034285408910363913, + 0.0034732529893517494, + 0.003727829083800316, + 0.005997173488140106, + 0.0037428135983645916, + 0.00792913231998682, + 0.008924991823732853, + 0.004467901773750782, + 0.0041114152409136295, + 0.011464724317193031, + 0.03419126570224762, + 0.010451676324009895, + 0.031520724296569824, + 0.007770840544253588, + 0.012511140666902065, + 0.00943477638065815, + 0.006279557477682829, + 0.010320799425244331, + 0.014302420429885387, + 0.009906052611768246, + 0.020364897325634956, + 0.015117517672479153, + 0.014658215455710888, + 0.008849969133734703, + 0.013766119256615639, + 0.003011114429682493, + 0.0027022056747227907, + 0.002780923619866371, + 0.0026529417373239994, + 0.002619201550260186, + 0.004376180935651064, + 0.0029657497070729733, + 0.002181625459343195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.25209951400756836 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right along the riverbank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the canyon walls rendered with simplified outlines and soft brushstrokes in a picture book style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river water a mixture of turquoise, sapphire, and lavender colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the water splashes captured as individual crystalline droplets tracing smooth arcs in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017174031818285584 + ], + "opposite_logprob_tokens": [ + -6.376717567443848 + ], + "logprob_reward": -0.0017174031818285584, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017174031818285584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22526907920837402, + 0.1023402214050293, + 0.20954959094524384, + 0.06810726970434189, + 0.13751524686813354, + 0.08524223417043686, + 0.05149071663618088, + 0.027869008481502533, + 0.05161658301949501, + 0.041000012308359146 + ], + "credit_map": [ + 0.04726380482316017, + 0.020956605672836304, + 0.02822575531899929, + 0.025832796469330788, + 0.013116943649947643, + 0.02264428697526455, + 0.0435626246035099, + 0.023666253313422203, + 0.01065992470830679, + 0.01201537437736988, + 0.01935514807701111, + 0.008678581565618515, + 0.009743645787239075, + 0.009784531779587269, + 0.013624478131532669, + 0.01847853884100914, + 0.0213566143065691, + 0.019646165892481804, + 0.023382611572742462, + 0.04849477857351303, + 0.015556764788925648, + 0.01842845417559147, + 0.045136965811252594, + 0.017547236755490303, + 0.00515188230201602, + 0.005338794551789761, + 0.0069815972819924355, + 0.011343983933329582, + 0.005738206673413515, + 0.009055853821337223, + 0.016733506694436073, + 0.0077634453773498535, + 0.011119834147393703, + 0.015960924327373505, + 0.018327398225665092, + 0.010609475895762444, + 0.0331682451069355, + 0.025441551581025124, + 0.0115627720952034, + 0.011325046420097351, + 0.007280571386218071, + 0.007951623760163784, + 0.014949572272598743, + 0.005713712889701128, + 0.02424183487892151, + 0.01256531197577715, + 0.0057169003412127495, + 0.006822704337537289, + 0.0027558559086173773, + 0.002525984076783061, + 0.014595426619052887, + 0.005054164677858353, + 0.015942519530653954, + 0.005715565290302038, + 0.0028248473536223173, + 0.0020763538777828217, + 0.002598382765427232, + 0.007297073025256395, + 0.0023895581252872944, + 0.006584784481674433, + 0.0021205267403274775, + 0.002657725242897868, + 0.0024498256389051676, + 0.0017711329273879528, + 0.004172960761934519, + 0.00451150257140398, + 0.009656769223511219, + 0.01502828299999237, + 0.006518120411783457, + 0.0046415794640779495, + 0.003886936930939555, + 0.003200432052835822, + 0.0033129912335425615, + 0.003387197619304061, + 0.003381229704245925, + 0.008401108905673027, + 0.010375197045505047, + 0.004961926024407148, + 0.003751272801309824, + 0.0034290875773876905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.0017174031818285584 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right along the riverbank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the canyon walls rendered with simplified outlines and soft brushstrokes in a picture book style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river water a mixture of turquoise, sapphire, and lavender colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the water splashes captured as individual crystalline droplets tracing smooth arcs in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016062619164586067 + ], + "opposite_logprob_tokens": [ + -4.1410627365112305 + ], + "logprob_reward": -0.016062619164586067, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016062619164586067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08997739851474762, + 0.049722641706466675, + 0.07966980338096619, + 0.30531105399131775, + 0.054264675825834274, + 0.037919823080301285, + 0.2113047093153, + 0.06479751318693161, + 0.027019616216421127, + 0.08001280575990677 + ], + "credit_map": [ + 0.012604750692844391, + 0.015299713239073753, + 0.008659936487674713, + 0.006987844128161669, + 0.01060574408620596, + 0.011042175814509392, + 0.01666060835123062, + 0.008116629905998707, + 0.005505639594048262, + 0.004786203615367413, + 0.0058548180386424065, + 0.005221524741500616, + 0.00526822404935956, + 0.0064964876510202885, + 0.005848349072039127, + 0.010741393081843853, + 0.003976879641413689, + 0.005930281709879637, + 0.004857373423874378, + 0.023060396313667297, + 0.006460641976445913, + 0.007701286114752293, + 0.0205797478556633, + 0.0071031986735761166, + 0.026077212765812874, + 0.027968913316726685, + 0.030750827863812447, + 0.0468786284327507, + 0.02343614399433136, + 0.030856691300868988, + 0.075102299451828, + 0.04424033686518669, + 0.004362115170806646, + 0.005868073552846909, + 0.00669869314879179, + 0.004592054523527622, + 0.012632360681891441, + 0.010828164406120777, + 0.0044581047259271145, + 0.00482510682195425, + 0.0027354450430721045, + 0.0026387555990368128, + 0.007476369384676218, + 0.0024473608937114477, + 0.01208458561450243, + 0.00593913160264492, + 0.0022059946786612272, + 0.00239218189381063, + 0.011014713905751705, + 0.009081719443202019, + 0.062474366277456284, + 0.020984061062335968, + 0.0659748986363411, + 0.02436637692153454, + 0.010315542109310627, + 0.007093027234077454, + 0.005910075735300779, + 0.017070330679416656, + 0.005809526890516281, + 0.01765422336757183, + 0.0036621461622416973, + 0.006299164146184921, + 0.005129355471581221, + 0.0032626900356262922, + 0.002456053625792265, + 0.0031120055355131626, + 0.004798055626451969, + 0.00669266190379858, + 0.0032757523003965616, + 0.0024172193370759487, + 0.00229810643941164, + 0.0019697619136422873, + 0.004963885992765427, + 0.003518576268106699, + 0.004952103830873966, + 0.023592641577124596, + 0.02792828157544136, + 0.007127705495804548, + 0.004533111583441496, + 0.003396496642380953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.016062619164586067 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right along the riverbank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the canyon walls rendered with simplified outlines and soft brushstrokes in a picture book style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river water a mixture of turquoise, sapphire, and lavender colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the water splashes captured as individual crystalline droplets tracing smooth arcs in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890149250626564 + ], + "opposite_logprob_tokens": [ + -2.464015007019043 + ], + "logprob_reward": -0.0890149250626564, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890149250626564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1734832227230072, + 0.06736025214195251, + 0.0522962287068367, + 0.4113740622997284, + 0.021279368549585342, + 0.04684877768158913, + 0.08880381286144257, + 0.04421555623412132, + 0.06095091253519058, + 0.03338777273893356 + ], + "credit_map": [ + 0.023778120055794716, + 0.03012848272919655, + 0.015845241025090218, + 0.01560342125594616, + 0.017353089526295662, + 0.01992722600698471, + 0.03486288711428642, + 0.015984762459993362, + 0.006567119620740414, + 0.009655158966779709, + 0.007879745215177536, + 0.006807253696024418, + 0.007003359962254763, + 0.007344114128500223, + 0.009212160483002663, + 0.012891346588730812, + 0.0034367649350315332, + 0.0039336285553872585, + 0.004005942493677139, + 0.012610040605068207, + 0.004390930291265249, + 0.005913360510021448, + 0.012506334111094475, + 0.005499228835105896, + 0.033694129437208176, + 0.037251900881528854, + 0.044558290392160416, + 0.0649857148528099, + 0.03080020844936371, + 0.04145946726202965, + 0.09990084171295166, + 0.05872352421283722, + 0.002130686305463314, + 0.002220458583906293, + 0.003976214211434126, + 0.0025974696036428213, + 0.00316700991243124, + 0.0023405891843140125, + 0.002557130763307214, + 0.0022898116149008274, + 0.003836310002952814, + 0.0038314014673233032, + 0.010464171878993511, + 0.004209601785987616, + 0.011001262813806534, + 0.006358888000249863, + 0.002815070329234004, + 0.004332077223807573, + 0.004783469717949629, + 0.004628833383321762, + 0.025740167126059532, + 0.008765595965087414, + 0.02718666009604931, + 0.00975156482309103, + 0.004249724559485912, + 0.0036978102289140224, + 0.004747874103486538, + 0.011280792765319347, + 0.003474325640127063, + 0.009646320715546608, + 0.003587437095120549, + 0.004673422779887915, + 0.003787922440096736, + 0.003017466515302658, + 0.005554940085858107, + 0.007467247545719147, + 0.006878713145852089, + 0.013386772945523262, + 0.00843819510191679, + 0.007538821082562208, + 0.00472209881991148, + 0.006964129861444235, + 0.0019984568934887648, + 0.001764315296895802, + 0.003827617270871997, + 0.007952908053994179, + 0.006816530600190163, + 0.0029914970509707928, + 0.004616045393049717, + 0.003420403692871332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.0890149250626564 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide-angle shot, the camera slowly pans right along the riverbank, capturing the majestic grandeur of the canyon at a high frame rate. The entire scene is rendered with the style of a delicate picture book. The towering canyon walls are a cascade of colors painted with soft brushstrokes, transitioning from warm terracotta to deep ochre. The outlines are simplified and expressive, and under the even, bright, diffuse light, there is not a single harsh shadow. At the bottom of the canyon, a colorful river rushes. The water is not a single blue, but a fantastical mixture of turquoise, sapphire, and lavender, shimmering with a warm luster under the soft light. The powerful current surges forward with unstoppable force, violently impacting the rounded boulders in the riverbed and stirring up massive whitewater waves. Thanks to the high frame rate's clear capture, each splash is broken down into countless crystalline individual droplets, which trace clear, smooth, and graceful arcs in the air before gracefully falling back onto the churning waters. The entire process is imbued with an exquisite sense of dynamic beauty. This extreme dynamic clarity contrasts sharply with the static aesthetic of the picture book-like background, creating a unique visual experience that is both imbued with powerful natural movement and imbued with the tranquility and beauty of a fairytale.", + "seed": 26511, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_1/index_4980_seed_26511_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10354684730991721, + "vlm_logprob": -0.10354684730991721, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.10354684730991721, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right along the riverbank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the canyon walls rendered with simplified outlines and soft brushstrokes in a picture book style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river water a mixture of turquoise, sapphire, and lavender colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the water splashes captured as individual crystalline droplets tracing smooth arcs in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029200272634625435 + ], + "opposite_logprob_tokens": [ + -3.904200315475464 + ], + "logprob_reward": -0.029200272634625435, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.029200272634625435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1840362250804901, + 0.11696877330541611, + 0.3054872155189514, + 0.0330667644739151, + 0.1475038230419159, + 0.03702908754348755, + 0.06720144301652908, + 0.04111760854721069, + 0.030314244329929352, + 0.03727484866976738 + ], + "credit_map": [ + 0.02032117359340191, + 0.027882305905222893, + 0.021147122606635094, + 0.019035153090953827, + 0.02939884550869465, + 0.03371667116880417, + 0.016628580167889595, + 0.01590634323656559, + 0.014156333170831203, + 0.02400452457368374, + 0.010189587250351906, + 0.014483819715678692, + 0.0191146619617939, + 0.014161852188408375, + 0.010522649623453617, + 0.010335328057408333, + 0.022098565474152565, + 0.03801245987415314, + 0.02322416938841343, + 0.07302626967430115, + 0.08642444014549255, + 0.020327825099229813, + 0.027415314689278603, + 0.014958134852349758, + 0.0039660013280808926, + 0.004018251784145832, + 0.005048729944974184, + 0.0032700994051992893, + 0.006694283336400986, + 0.0034370727371424437, + 0.002992961322888732, + 0.0036393599584698677, + 0.01738925091922283, + 0.018305784091353416, + 0.016400886699557304, + 0.020443877205252647, + 0.017295261844992638, + 0.017369352281093597, + 0.020354896783828735, + 0.01994450017809868, + 0.004275673534721136, + 0.003974223975092173, + 0.0053446595557034016, + 0.003578097792342305, + 0.005942082032561302, + 0.007020529825240374, + 0.0028192526660859585, + 0.00407456373795867, + 0.007078087888658047, + 0.009213373996317387, + 0.008192410692572594, + 0.007185814436525106, + 0.009208230301737785, + 0.008403606712818146, + 0.008164373226463795, + 0.009755538776516914, + 0.006440389901399612, + 0.005650503095239401, + 0.004708047490566969, + 0.004224604461342096, + 0.004470495972782373, + 0.0037368552293628454, + 0.006001182831823826, + 0.00588552700355649, + 0.004943152889609337, + 0.004549928940832615, + 0.0025637103244662285, + 0.0031864983029663563, + 0.004552361089736223, + 0.004663048777729273, + 0.0028749953489750624, + 0.0029805446974933147, + 0.005037609487771988, + 0.005329091101884842, + 0.0029200860299170017, + 0.004461533389985561, + 0.005623058881610632, + 0.0064579276368021965, + 0.003525987034663558, + 0.003919551149010658 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.029200272634625435 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right along the riverbank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the canyon walls rendered with simplified outlines and soft brushstrokes in a picture book style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river water a mixture of turquoise, sapphire, and lavender colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the water splashes captured as individual crystalline droplets tracing smooth arcs in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34861552715301514 + ], + "opposite_logprob_tokens": [ + -1.2236155271530151 + ], + "logprob_reward": -0.34861552715301514, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34861552715301514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13267064094543457, + 0.10936237126588821, + 0.09860393404960632, + 0.04082394018769264, + 0.36396360397338867, + 0.028354670852422714, + 0.05837678909301758, + 0.02590424381196499, + 0.03655729815363884, + 0.10538249462842941 + ], + "credit_map": [ + 0.02237159200012684, + 0.021713484078645706, + 0.011888722889125347, + 0.020736822858452797, + 0.01645774208009243, + 0.019486214965581894, + 0.010728047229349613, + 0.009288025088608265, + 0.01367980893701315, + 0.023187534883618355, + 0.008444998413324356, + 0.011077308095991611, + 0.020479312166571617, + 0.014265166595578194, + 0.00872771069407463, + 0.009500529617071152, + 0.007327445782721043, + 0.015133700333535671, + 0.009837035089731216, + 0.01985902152955532, + 0.023062363266944885, + 0.007525735534727573, + 0.008945044130086899, + 0.006913576275110245, + 0.00671797851100564, + 0.007125621661543846, + 0.004167620558291674, + 0.0035372613929212093, + 0.005241574253886938, + 0.006785821169614792, + 0.0034280072432011366, + 0.0038200540002435446, + 0.04127958416938782, + 0.0376451201736927, + 0.04125961661338806, + 0.058195531368255615, + 0.04219580814242363, + 0.04146689176559448, + 0.05014347285032272, + 0.05177755281329155, + 0.002846950199455023, + 0.003624010132625699, + 0.005557720549404621, + 0.0038431992288678885, + 0.0044523137621581554, + 0.002265437040477991, + 0.0028897018637508154, + 0.00287533993832767, + 0.005670452956110239, + 0.005556935910135508, + 0.008273246698081493, + 0.007026453036814928, + 0.007495030295103788, + 0.007276176009327173, + 0.007638746406883001, + 0.009439744986593723, + 0.0038235336542129517, + 0.0027945328038185835, + 0.002558933338150382, + 0.004430806264281273, + 0.0031637840438634157, + 0.003478211583569646, + 0.002583124442026019, + 0.0030713158193975687, + 0.006890950258821249, + 0.006108072120696306, + 0.003417454194277525, + 0.003713022219017148, + 0.005005331244319677, + 0.00440414622426033, + 0.0036945389583706856, + 0.0033237801399081945, + 0.014030488207936287, + 0.017401380464434624, + 0.008253931067883968, + 0.009155801497399807, + 0.020354406908154488, + 0.0165146142244339, + 0.00920277088880539, + 0.010469104163348675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.34861552715301514 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right along the riverbank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the canyon walls rendered with simplified outlines and soft brushstrokes in a picture book style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river water a mixture of turquoise, sapphire, and lavender colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the water splashes captured as individual crystalline droplets tracing smooth arcs in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005949051585048437 + ], + "opposite_logprob_tokens": [ + -5.130949020385742 + ], + "logprob_reward": -0.005949051585048437, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005949051585048437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15886923670768738, + 0.04180360212922096, + 0.4766623377799988, + 0.017317572608590126, + 0.06264307349920273, + 0.10574077814817429, + 0.021066579967737198, + 0.041003093123435974, + 0.03358517214655876, + 0.041308626532554626 + ], + "credit_map": [ + 0.02002010866999626, + 0.028799299150705338, + 0.015361146070063114, + 0.019098173826932907, + 0.024214740842580795, + 0.02013586275279522, + 0.01697208359837532, + 0.014267802238464355, + 0.006407040171325207, + 0.006605509202927351, + 0.005124472547322512, + 0.004836666863411665, + 0.006528795231133699, + 0.004873950034379959, + 0.003975543659180403, + 0.00345161953009665, + 0.03570608049631119, + 0.05881219357252121, + 0.03354482352733612, + 0.11212863028049469, + 0.1418105959892273, + 0.030582696199417114, + 0.0414227694272995, + 0.022654499858617783, + 0.0020981081761419773, + 0.002262673806399107, + 0.0019250746117904782, + 0.0024632010608911514, + 0.0028678125236183405, + 0.002217750996351242, + 0.0017211433732882142, + 0.0017618071287870407, + 0.007308317348361015, + 0.0076107457280159, + 0.010373950935900211, + 0.015153477899730206, + 0.00631649000570178, + 0.005348140839487314, + 0.005849150009453297, + 0.0046827904880046844, + 0.01004620362073183, + 0.007243230007588863, + 0.024844272062182426, + 0.018165607005357742, + 0.011681354604661465, + 0.008358050137758255, + 0.011850474402308464, + 0.013551575131714344, + 0.0024476463440805674, + 0.0021605913061648607, + 0.003023129189386964, + 0.003377708373591304, + 0.002087875036522746, + 0.0017530708573758602, + 0.0031712432391941547, + 0.003045313525944948, + 0.003911213483661413, + 0.004000306595116854, + 0.004495896399021149, + 0.0056303199380636215, + 0.00432737497612834, + 0.004063994158059359, + 0.007160318549722433, + 0.0074136643670499325, + 0.005041064694523811, + 0.004870167002081871, + 0.003067322075366974, + 0.0030970685184001923, + 0.005384438205510378, + 0.004728434607386589, + 0.003352214116603136, + 0.004044459201395512, + 0.007409487385302782, + 0.006633587181568146, + 0.003150716656818986, + 0.004708106629550457, + 0.006107947323471308, + 0.005002617835998535, + 0.0035808023530989885, + 0.004715355578809977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.005949051585048437 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right along the riverbank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the canyon walls rendered with simplified outlines and soft brushstrokes in a picture book style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river water a mixture of turquoise, sapphire, and lavender colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the water splashes captured as individual crystalline droplets tracing smooth arcs in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368634358048439 + ], + "opposite_logprob_tokens": [ + -3.408686399459839 + ], + "logprob_reward": -0.03368634358048439, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368634358048439, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.059833966195583344, + 0.03556784242391586, + 0.3162982165813446, + 0.05009333789348602, + 0.06126327067613602, + 0.058268994092941284, + 0.04013271629810333, + 0.15403467416763306, + 0.1008586436510086, + 0.12364830076694489 + ], + "credit_map": [ + 0.00798000954091549, + 0.010692167095839977, + 0.0065961191430687904, + 0.006910465192049742, + 0.006773272063583136, + 0.008483042009174824, + 0.00564314192160964, + 0.006755751091986895, + 0.004180082585662603, + 0.007692627143114805, + 0.0031257353257387877, + 0.004681859165430069, + 0.004856904037296772, + 0.004174354020506144, + 0.0036890392657369375, + 0.0031672411132603884, + 0.02230978012084961, + 0.03742823749780655, + 0.02283787727355957, + 0.07783184200525284, + 0.09169795364141464, + 0.019706428050994873, + 0.029061494395136833, + 0.015424617566168308, + 0.006963287480175495, + 0.006461271084845066, + 0.005756719969213009, + 0.008907734416425228, + 0.006421301048249006, + 0.006970707792788744, + 0.004577511455863714, + 0.004034810699522495, + 0.005798493977636099, + 0.006826701108366251, + 0.009104817174375057, + 0.010665316134691238, + 0.006925572175532579, + 0.006976045202463865, + 0.007560855709016323, + 0.007405473385006189, + 0.003737064078450203, + 0.0062682535499334335, + 0.011200681328773499, + 0.007401017472147942, + 0.008402911946177483, + 0.004429046995937824, + 0.007875710725784302, + 0.008954311721026897, + 0.0054023670963943005, + 0.005693132523447275, + 0.0036984372418373823, + 0.004814364016056061, + 0.004401658661663532, + 0.004751327447593212, + 0.005183347035199404, + 0.0061880857683718204, + 0.017423288896679878, + 0.015978535637259483, + 0.01598784327507019, + 0.01949908770620823, + 0.01818266697227955, + 0.017336620017886162, + 0.02248750440776348, + 0.02713914029300213, + 0.01692701317369938, + 0.013308470137417316, + 0.008901053108274937, + 0.011203082278370857, + 0.01506522111594677, + 0.012302334420382977, + 0.010660805739462376, + 0.012490681372582912, + 0.0168767087161541, + 0.020379552617669106, + 0.009923655539751053, + 0.01338577177375555, + 0.02265661582350731, + 0.018136434257030487, + 0.010452463291585445, + 0.011837108060717583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 9 + ], + "selected_reward": -0.03368634358048439 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right along the riverbank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the canyon walls rendered with simplified outlines and soft brushstrokes in a picture book style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river water a mixture of turquoise, sapphire, and lavender colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the water splashes captured as individual crystalline droplets tracing smooth arcs in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10028304159641266 + ], + "opposite_logprob_tokens": [ + -2.350283145904541 + ], + "logprob_reward": -0.10028304159641266, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10028304159641266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1246643140912056, + 0.040181297808885574, + 0.39990922808647156, + 0.03477027639746666, + 0.19684892892837524, + 0.03170251101255417, + 0.026113631203770638, + 0.07574386149644852, + 0.03585316985845566, + 0.03421274200081825 + ], + "credit_map": [ + 0.018235914409160614, + 0.028323445469141006, + 0.009909983724355698, + 0.014636877924203873, + 0.017591208219528198, + 0.01861739158630371, + 0.00890977494418621, + 0.008439731784164906, + 0.004270209465175867, + 0.005275053903460503, + 0.005060995928943157, + 0.007068878039717674, + 0.006819194182753563, + 0.0034241625107824802, + 0.003707072464749217, + 0.004555735737085342, + 0.02706793323159218, + 0.04526427760720253, + 0.029694758355617523, + 0.10048648715019226, + 0.11516787111759186, + 0.025428403168916702, + 0.03750796243548393, + 0.019291555508971214, + 0.003765829373151064, + 0.004451758693903685, + 0.008507225662469864, + 0.0050399573519825935, + 0.0033959385473281145, + 0.0025943827349692583, + 0.00407791743054986, + 0.002937269164249301, + 0.02646060846745968, + 0.01834910549223423, + 0.02263326570391655, + 0.029828457161784172, + 0.02242053672671318, + 0.020061468705534935, + 0.02739889919757843, + 0.029696594923734665, + 0.0035603472497314215, + 0.0041540851816535, + 0.0037142857909202576, + 0.004341559950262308, + 0.003944781608879566, + 0.0029590255580842495, + 0.004623458255082369, + 0.004404969047755003, + 0.0029732645489275455, + 0.0024136644788086414, + 0.0056742834858596325, + 0.0031334813684225082, + 0.003605139208957553, + 0.0026870733126997948, + 0.002860012697055936, + 0.002766713732853532, + 0.008729026652872562, + 0.007905412465333939, + 0.00757596967741847, + 0.011218714527785778, + 0.007229018025100231, + 0.008164171129465103, + 0.013041676953434944, + 0.011879878118634224, + 0.005586072336882353, + 0.00383792188949883, + 0.0040430533699691296, + 0.004600184503942728, + 0.0034285138826817274, + 0.0030909141059964895, + 0.006189175881445408, + 0.0050773355178534985, + 0.005609969142824411, + 0.00538531644269824, + 0.0037352677900344133, + 0.003810405032709241, + 0.003598492592573166, + 0.005234928335994482, + 0.003752427641302347, + 0.00308593618683517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.10028304159641266 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_10/index_tensor([4930, 4930])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_10/index_tensor([4930, 4930])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a369943b6cbc4519aeb1288eb2992754362b9583 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_10/index_tensor([4930, 4930])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium, static shot, taken from a slightly upward angle, captures a commanding ancient general, his complexion crimson like a red date and his presence imposing. His thick, slender black beard falls to his chest, his eyes piercing with lightning. He clad in a finely crafted suit of bronze armor, featuring thick shoulder pads and intricately carved animal-face patterns. With his leather-guarded right hand, he grips the hilt of his massive crescent-shaped blade with a force that threatens to crush it. The veins in his massive forearm, tangled like tangled tree roots, are visible from the sheer force. Every detail reveals the immense power contained within. The background is a battlefield filled with yellow sand and smoke. The sky is gloomy, and the cold sunlight penetrates the thin clouds, casting sharp metallic reflections on his armor and cold blade. The main colors of the picture tend to be gloomy ochre and cold metallic colors. The whole scene is filled with the tension of impending storm and a powerful aura that is ready to explode.", + "seed": 30720, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_10/index_4930_seed_30720_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.025050134724006058, + "vlm_logprob": -0.025050134724006058, + "vlm_yes_no_margin": 4.349999952316284, + "vlm_token_credit": -0.025050134724006058, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient general with a crimson complexion and a thick black beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general wearing bronze armor with thick shoulder pads and carved animal-face patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general gripping a massive crescent-shaped blade with his leather-guarded right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a battlefield filled with yellow sand and smoke under a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01891445554792881 + ], + "opposite_logprob_tokens": [ + -4.018914222717285 + ], + "logprob_reward": -0.01891445554792881, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01891445554792881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4197288155555725, + 0.08357936143875122, + 0.030404193326830864, + 0.05133502557873726, + 0.05073828250169754, + 0.08159321546554565, + 0.05372662469744682, + 0.08878910541534424, + 0.08429720997810364, + 0.055808186531066895 + ], + "credit_map": [ + 0.042469169944524765, + 0.0878896415233612, + 0.05262056738138199, + 0.041646674275398254, + 0.057814907282590866, + 0.036750759929418564, + 0.05172141268849373, + 0.04881569370627403, + 0.011871964670717716, + 0.010777967981994152, + 0.009182138368487358, + 0.010453636758029461, + 0.01179756224155426, + 0.008012722246348858, + 0.01025531068444252, + 0.011228060349822044, + 0.004845347721129656, + 0.003037572605535388, + 0.0031706371810287237, + 0.0033723972737789154, + 0.00461138179525733, + 0.0036499795969575644, + 0.0040820688009262085, + 0.003634805791079998, + 0.008533371612429619, + 0.0069858478382229805, + 0.00571878906339407, + 0.005522418301552534, + 0.00769699364900589, + 0.006294236984103918, + 0.005104159004986286, + 0.005479206796735525, + 0.008380221202969551, + 0.009151835925877094, + 0.00500407675281167, + 0.005659265909343958, + 0.007832654751837254, + 0.006671964190900326, + 0.004010406788438559, + 0.004027856048196554, + 0.013983247801661491, + 0.010550879873335361, + 0.009863942861557007, + 0.012150882743299007, + 0.009909164160490036, + 0.008696235716342926, + 0.007540558464825153, + 0.008898304775357246, + 0.006833394058048725, + 0.007240316830575466, + 0.006397467106580734, + 0.007165207527577877, + 0.00639504985883832, + 0.005300905555486679, + 0.0072854491882026196, + 0.007108835503458977, + 0.012650015763938427, + 0.013880360871553421, + 0.007169103249907494, + 0.009007926099002361, + 0.014639104716479778, + 0.010258236899971962, + 0.007535542827099562, + 0.013648810796439648, + 0.021188993006944656, + 0.011719792149960995, + 0.006903160363435745, + 0.00837708916515112, + 0.008090391755104065, + 0.013918319717049599, + 0.00740086380392313, + 0.006698599085211754, + 0.008091503754258156, + 0.007011951878666878, + 0.005429532378911972, + 0.0073865074664354324, + 0.008200363256037235, + 0.0061849369667470455, + 0.008478120900690556, + 0.005025269463658333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.01891445554792881 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient general with a crimson complexion and a thick black beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general wearing bronze armor with thick shoulder pads and carved animal-face patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general gripping a massive crescent-shaped blade with his leather-guarded right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a battlefield filled with yellow sand and smoke under a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00362691399641335 + ], + "opposite_logprob_tokens": [ + -5.628626823425293 + ], + "logprob_reward": -0.00362691399641335, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00362691399641335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33078479766845703, + 0.08649349957704544, + 0.04053664207458496, + 0.05198463052511215, + 0.07250670343637466, + 0.055072762072086334, + 0.053135402500629425, + 0.09699368476867676, + 0.11783673614263535, + 0.09465516358613968 + ], + "credit_map": [ + 0.042936936020851135, + 0.04581087455153465, + 0.03578324243426323, + 0.03894277289509773, + 0.04055916145443916, + 0.03533494472503662, + 0.0370272658765316, + 0.05438954383134842, + 0.0105856042355299, + 0.010984504595398903, + 0.009042371064424515, + 0.008946687914431095, + 0.011997085064649582, + 0.009440159425139427, + 0.013686362653970718, + 0.011810719035565853, + 0.005038468167185783, + 0.004981457721441984, + 0.004849147982895374, + 0.005953795276582241, + 0.005968907848000526, + 0.0036450542975217104, + 0.0061153024435043335, + 0.00398450531065464, + 0.005381591152399778, + 0.007301828823983669, + 0.005272149574011564, + 0.005856924690306187, + 0.007998045533895493, + 0.004520989954471588, + 0.007110584527254105, + 0.008542519062757492, + 0.008446495980024338, + 0.01089767087250948, + 0.010061771608889103, + 0.010694285854697227, + 0.009823865257203579, + 0.006225086748600006, + 0.007207210175693035, + 0.009150311350822449, + 0.008476722054183483, + 0.009275594726204872, + 0.004546678625047207, + 0.006665064487606287, + 0.009722011163830757, + 0.006028143223375082, + 0.00498985406011343, + 0.00536868954077363, + 0.007480916567146778, + 0.005601306911557913, + 0.005277068819850683, + 0.005378250032663345, + 0.008434254676103592, + 0.005716186016798019, + 0.006085558794438839, + 0.009161852300167084, + 0.011072319932281971, + 0.013425104320049286, + 0.009405598044395447, + 0.010016760788857937, + 0.018354419618844986, + 0.009881253354251385, + 0.010565240867435932, + 0.014272981323301792, + 0.018168384209275246, + 0.018505247309803963, + 0.009781252592802048, + 0.011142667382955551, + 0.016430426388978958, + 0.014780684374272823, + 0.01303800381720066, + 0.01599005237221718, + 0.007994404062628746, + 0.013063515536487103, + 0.010027590207755566, + 0.012997782789170742, + 0.013841048814356327, + 0.009303657338023186, + 0.013598565943539143, + 0.01382859144359827 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.00362691399641335 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient general with a crimson complexion and a thick black beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general wearing bronze armor with thick shoulder pads and carved animal-face patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general gripping a massive crescent-shaped blade with his leather-guarded right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a battlefield filled with yellow sand and smoke under a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020572146400809288 + ], + "opposite_logprob_tokens": [ + -3.8955721855163574 + ], + "logprob_reward": -0.020572146400809288, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020572146400809288, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28175705671310425, + 0.06758349388837814, + 0.05491645261645317, + 0.0709397941827774, + 0.08652948588132858, + 0.10296299308538437, + 0.1439458727836609, + 0.07056427747011185, + 0.0619424507021904, + 0.058858081698417664 + ], + "credit_map": [ + 0.03523989021778107, + 0.04839450493454933, + 0.03095654956996441, + 0.02461359277367592, + 0.043884847313165665, + 0.026548270136117935, + 0.02849951572716236, + 0.04361989349126816, + 0.009812257252633572, + 0.008349108509719372, + 0.007170126773416996, + 0.00758429616689682, + 0.01022628415375948, + 0.008724498562514782, + 0.008447980508208275, + 0.007268947549164295, + 0.008361946791410446, + 0.006257764995098114, + 0.005346032325178385, + 0.007719376124441624, + 0.005900098942220211, + 0.005429379642009735, + 0.005978872068226337, + 0.009922989644110203, + 0.014021678827702999, + 0.007046624552458525, + 0.005759100429713726, + 0.006459786556661129, + 0.011163754388689995, + 0.010011817328631878, + 0.007832109928131104, + 0.008644928224384785, + 0.01373363472521305, + 0.009147326461970806, + 0.011722502298653126, + 0.013448822312057018, + 0.009644357487559319, + 0.00863020122051239, + 0.008670120500028133, + 0.011532525531947613, + 0.023055778816342354, + 0.012981678359210491, + 0.00928818341344595, + 0.010045577771961689, + 0.011798224411904812, + 0.01670023612678051, + 0.008413484320044518, + 0.010679841972887516, + 0.03181104362010956, + 0.01887405663728714, + 0.014588038437068462, + 0.014211129397153854, + 0.016406772658228874, + 0.019485685974359512, + 0.014231705106794834, + 0.014337453991174698, + 0.008243164047598839, + 0.010838955640792847, + 0.006482359487563372, + 0.006512217223644257, + 0.012683730572462082, + 0.007130327634513378, + 0.007873859256505966, + 0.010799667797982693, + 0.01562427170574665, + 0.008280964568257332, + 0.005264054052531719, + 0.004440769087523222, + 0.006341895554214716, + 0.01065844390541315, + 0.005471937824040651, + 0.005860117729753256, + 0.008190950378775597, + 0.008776001632213593, + 0.006196130532771349, + 0.007325687911361456, + 0.00875565130263567, + 0.006380137987434864, + 0.007676103617995977, + 0.005557428114116192 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.020572146400809288 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient general with a crimson complexion and a thick black beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general wearing bronze armor with thick shoulder pads and carved animal-face patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general gripping a massive crescent-shaped blade with his leather-guarded right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a battlefield filled with yellow sand and smoke under a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003189241047948599 + ], + "opposite_logprob_tokens": [ + -5.7531890869140625 + ], + "logprob_reward": -0.003189241047948599, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003189241047948599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21473854780197144, + 0.13063696026802063, + 0.16454319655895233, + 0.06568243354558945, + 0.0556020550429821, + 0.08127664029598236, + 0.14120523631572723, + 0.05648654326796532, + 0.04647251218557358, + 0.043355900794267654 + ], + "credit_map": [ + 0.023943772539496422, + 0.0404839813709259, + 0.022899385541677475, + 0.019551243633031845, + 0.041818927973508835, + 0.019758298993110657, + 0.021836889907717705, + 0.024446064606308937, + 0.021024908870458603, + 0.013194583356380463, + 0.015838533639907837, + 0.01459657121449709, + 0.01354013942182064, + 0.015035453252494335, + 0.012915998697280884, + 0.024490773677825928, + 0.022951960563659668, + 0.019697923213243484, + 0.016981003805994987, + 0.015078801661729813, + 0.02178204618394375, + 0.017687898129224777, + 0.01900990679860115, + 0.0313536673784256, + 0.012324206531047821, + 0.00922046322375536, + 0.008417502976953983, + 0.007181030698120594, + 0.006941474974155426, + 0.009039249271154404, + 0.006292259786278009, + 0.006266247481107712, + 0.010640475898981094, + 0.006030566059052944, + 0.0063848006539046764, + 0.00791032426059246, + 0.006509057711809874, + 0.0076470826752483845, + 0.00409145001322031, + 0.006388295907527208, + 0.011822186410427094, + 0.010906807146966457, + 0.005814809817820787, + 0.008558601140975952, + 0.014809873886406422, + 0.009156235493719578, + 0.008524514734745026, + 0.011683614924550056, + 0.037023499608039856, + 0.015351357869803905, + 0.011984491720795631, + 0.012571783736348152, + 0.01538405753672123, + 0.0227765254676342, + 0.011917470954358578, + 0.014196042902767658, + 0.011885719373822212, + 0.006699790246784687, + 0.005430776625871658, + 0.0050473264418542385, + 0.008938956074416637, + 0.007851884700357914, + 0.005024169106036425, + 0.005607923027127981, + 0.011975551955401897, + 0.005633765831589699, + 0.0038678362034261227, + 0.004171617794781923, + 0.005700198467820883, + 0.007546734530478716, + 0.0036438507959246635, + 0.003932960797101259, + 0.006679827813059092, + 0.0053435759618878365, + 0.0038206574972718954, + 0.004541232716292143, + 0.006565420422703028, + 0.004631944932043552, + 0.005537544377148151, + 0.006235698238015175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.003189241047948599 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient general with a crimson complexion and a thick black beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general wearing bronze armor with thick shoulder pads and carved animal-face patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general gripping a massive crescent-shaped blade with his leather-guarded right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a battlefield filled with yellow sand and smoke under a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07894791662693024 + ], + "opposite_logprob_tokens": [ + -2.5789480209350586 + ], + "logprob_reward": -0.07894791662693024, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07894791662693024, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25266650319099426, + 0.07974014431238174, + 0.06729195266962051, + 0.0953705683350563, + 0.18619555234909058, + 0.09351421892642975, + 0.10741626471281052, + 0.043010346591472626, + 0.035602446645498276, + 0.03919195383787155 + ], + "credit_map": [ + 0.041296180337667465, + 0.0324840173125267, + 0.01762726902961731, + 0.02458253689110279, + 0.033465150743722916, + 0.030392078682780266, + 0.035272106528282166, + 0.03754717856645584, + 0.012606756761670113, + 0.010366384871304035, + 0.0074519566260278225, + 0.008661113679409027, + 0.01186059508472681, + 0.009795110672712326, + 0.01043026614934206, + 0.008567960932850838, + 0.009278546087443829, + 0.01040291041135788, + 0.00656613614410162, + 0.007444208953529596, + 0.012378234416246414, + 0.006733781658113003, + 0.007057954091578722, + 0.007430184166878462, + 0.019757766276597977, + 0.01129066664725542, + 0.008510222658514977, + 0.012426861561834812, + 0.01144022773951292, + 0.013703946955502033, + 0.010332213714718819, + 0.00790866557508707, + 0.03242884576320648, + 0.021790064871311188, + 0.025533165782690048, + 0.032936688512563705, + 0.017699578776955605, + 0.02034028246998787, + 0.015083038248121738, + 0.020383894443511963, + 0.021683204919099808, + 0.011825925670564175, + 0.007209172006696463, + 0.008217273280024529, + 0.010091195814311504, + 0.01627330668270588, + 0.00863308273255825, + 0.009581061080098152, + 0.01790551468729973, + 0.01059882715344429, + 0.011974195018410683, + 0.012399137020111084, + 0.013703996315598488, + 0.014002205803990364, + 0.010486756451427937, + 0.016345638781785965, + 0.011965492740273476, + 0.004113000817596912, + 0.0037040153983980417, + 0.0028095843736082315, + 0.00508795166388154, + 0.007412404287606478, + 0.003569873748347163, + 0.0043480233289301395, + 0.007298422511667013, + 0.004068935755640268, + 0.0025807530619204044, + 0.0043307640589773655, + 0.004848766606301069, + 0.00513893598690629, + 0.0034538402687758207, + 0.003882026532664895, + 0.003282813122496009, + 0.004717647098004818, + 0.0041639553382992744, + 0.0060900612734258175, + 0.008046132512390614, + 0.0036132431123405695, + 0.005977985914796591, + 0.00330011616460979 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.07894791662693024 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium, static shot, taken from a slightly upward angle, captures a commanding ancient general, his complexion crimson like a red date and his presence imposing. His thick, slender black beard falls to his chest, his eyes piercing with lightning. He clad in a finely crafted suit of bronze armor, featuring thick shoulder pads and intricately carved animal-face patterns. With his leather-guarded right hand, he grips the hilt of his massive crescent-shaped blade with a force that threatens to crush it. The veins in his massive forearm, tangled like tangled tree roots, are visible from the sheer force. Every detail reveals the immense power contained within. The background is a battlefield filled with yellow sand and smoke. The sky is gloomy, and the cold sunlight penetrates the thin clouds, casting sharp metallic reflections on his armor and cold blade. The main colors of the picture tend to be gloomy ochre and cold metallic colors. The whole scene is filled with the tension of impending storm and a powerful aura that is ready to explode.", + "seed": 30721, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_10/index_4930_seed_30721_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08042259383946657, + "vlm_logprob": -0.08042259383946657, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.08042259383946657, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient general with a crimson complexion and a thick black beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general wearing bronze armor with thick shoulder pads and carved animal-face patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general gripping a massive crescent-shaped blade with his leather-guarded right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a battlefield filled with yellow sand and smoke under a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03860185667872429 + ], + "opposite_logprob_tokens": [ + -3.288601875305176 + ], + "logprob_reward": -0.03860185667872429, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03860185667872429, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16141235828399658, + 0.07137661427259445, + 0.37303927540779114, + 0.04139658063650131, + 0.08597370982170105, + 0.017438989132642746, + 0.023301882669329643, + 0.09487231820821762, + 0.10287167876958847, + 0.02831657975912094 + ], + "credit_map": [ + 0.018110405653715134, + 0.018087150529026985, + 0.02957700565457344, + 0.04789797589182854, + 0.009850353933870792, + 0.014302772469818592, + 0.011059313081204891, + 0.012527380138635635, + 0.008983018808066845, + 0.012418125756084919, + 0.009258449077606201, + 0.016915995627641678, + 0.006959831807762384, + 0.004497839137911797, + 0.005843706429004669, + 0.006499639246612787, + 0.018581002950668335, + 0.0276755653321743, + 0.010926740244030952, + 0.02698373608291149, + 0.015767047181725502, + 0.13339436054229736, + 0.01319984532892704, + 0.12651097774505615, + 0.003688319120556116, + 0.008761849254369736, + 0.003199787111952901, + 0.009311451576650143, + 0.005139012821018696, + 0.0039021209813654423, + 0.0035954329650849104, + 0.0037986028473824263, + 0.00835946761071682, + 0.019762765616178513, + 0.007008562795817852, + 0.010253836400806904, + 0.011552420444786549, + 0.01119830273091793, + 0.008126896806061268, + 0.009711448103189468, + 0.001689837547019124, + 0.0028839909937232733, + 0.0019093649461865425, + 0.0024108781944960356, + 0.00254500238224864, + 0.002249118871986866, + 0.0020720872562378645, + 0.00167870766017586, + 0.0024952010717242956, + 0.0035295349080115557, + 0.0022859356831759214, + 0.002499575726687908, + 0.0041036768816411495, + 0.004016575403511524, + 0.0022927385289222, + 0.0020786430686712265, + 0.010832173749804497, + 0.018703851848840714, + 0.006898609455674887, + 0.00902183074504137, + 0.020656781271100044, + 0.012969045899808407, + 0.008811824955046177, + 0.006978197488933802, + 0.014805838465690613, + 0.014983137138187885, + 0.004821606911718845, + 0.007025141268968582, + 0.012545146048069, + 0.03408045694231987, + 0.008525163866579533, + 0.006085173226892948, + 0.0030746604315936565, + 0.005228135734796524, + 0.0020136532839387655, + 0.002846968127414584, + 0.004665724467486143, + 0.0034339213743805885, + 0.003999214619398117, + 0.0030543014872819185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.03860185667872429 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient general with a crimson complexion and a thick black beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general wearing bronze armor with thick shoulder pads and carved animal-face patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general gripping a massive crescent-shaped blade with his leather-guarded right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a battlefield filled with yellow sand and smoke under a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008650331757962704 + ], + "opposite_logprob_tokens": [ + -4.758650302886963 + ], + "logprob_reward": -0.008650331757962704, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008650331757962704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13404136896133423, + 0.05558821186423302, + 0.5518864989280701, + 0.04518251121044159, + 0.03404001146554947, + 0.017575865611433983, + 0.021342914551496506, + 0.04072745516896248, + 0.08098169416189194, + 0.018633466213941574 + ], + "credit_map": [ + 0.017141709104180336, + 0.012701169587671757, + 0.029800962656736374, + 0.03694246709346771, + 0.007002291735261679, + 0.008747455663979053, + 0.00899860542267561, + 0.012706708163022995, + 0.006722956895828247, + 0.0067888228222727776, + 0.007527250796556473, + 0.009924850426614285, + 0.006779064889997244, + 0.006091895513236523, + 0.005423456430435181, + 0.0063299150206148624, + 0.022716248407959938, + 0.024796582758426666, + 0.011612948030233383, + 0.03136647492647171, + 0.017058363184332848, + 0.22807829082012177, + 0.018085848540067673, + 0.1981717348098755, + 0.003698303597047925, + 0.0071415649726986885, + 0.005028979852795601, + 0.006356472615152597, + 0.0043987249955534935, + 0.010962065309286118, + 0.004358247853815556, + 0.0032381543423980474, + 0.004027698654681444, + 0.006050719413906336, + 0.0038147384766489267, + 0.004330464173108339, + 0.0050748963840305805, + 0.005241761449724436, + 0.002711717039346695, + 0.002788018900901079, + 0.0019306314643472433, + 0.0029175893869251013, + 0.0016414360143244267, + 0.0020086783915758133, + 0.0029289391823112965, + 0.002890663919970393, + 0.0017821984365582466, + 0.001475726836360991, + 0.002323601394891739, + 0.0033342784736305475, + 0.0021805923897773027, + 0.0020763296633958817, + 0.00333566521294415, + 0.0039742509834468365, + 0.001998127670958638, + 0.002120070159435272, + 0.004189463332295418, + 0.005820234306156635, + 0.004805686417967081, + 0.004143794532865286, + 0.005831994116306305, + 0.007531153503805399, + 0.004169889260083437, + 0.004235239699482918, + 0.012396767735481262, + 0.008856577798724174, + 0.0033506995532661676, + 0.0051667108200490475, + 0.009725931100547314, + 0.029101945459842682, + 0.007225753739476204, + 0.005157300271093845, + 0.0016897623427212238, + 0.0030332410242408514, + 0.0016318736597895622, + 0.0018601586343720555, + 0.0036123828031122684, + 0.0024812682531774044, + 0.0023615886457264423, + 0.001963190268725157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.008650331757962704 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient general with a crimson complexion and a thick black beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general wearing bronze armor with thick shoulder pads and carved animal-face patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general gripping a massive crescent-shaped blade with his leather-guarded right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a battlefield filled with yellow sand and smoke under a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023280387744307518 + ], + "opposite_logprob_tokens": [ + -3.773280382156372 + ], + "logprob_reward": -0.023280387744307518, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023280387744307518, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06150529161095619, + 0.040620237588882446, + 0.22754749655723572, + 0.16834717988967896, + 0.0844990536570549, + 0.03485731780529022, + 0.030014224350452423, + 0.12860475480556488, + 0.1761760264635086, + 0.04782850295305252 + ], + "credit_map": [ + 0.006819036789238453, + 0.008125649765133858, + 0.006528007332235575, + 0.010839181020855904, + 0.009141342714428902, + 0.005941209848970175, + 0.004902564454823732, + 0.00920829176902771, + 0.004168620333075523, + 0.00744854100048542, + 0.004684247076511383, + 0.007101237308233976, + 0.005891272332519293, + 0.0022653047926723957, + 0.0042164684273302555, + 0.004844543058425188, + 0.009625914506614208, + 0.014788334257900715, + 0.005227481015026569, + 0.014383809641003609, + 0.010576310567557812, + 0.09470148384571075, + 0.007177605759352446, + 0.07106653600931168, + 0.012034735642373562, + 0.02738366089761257, + 0.025242028757929802, + 0.02917327545583248, + 0.01551978848874569, + 0.02432062476873398, + 0.01956118643283844, + 0.01511186920106411, + 0.008388602174818516, + 0.012992276810109615, + 0.008162004873156548, + 0.00998335238546133, + 0.014813330955803394, + 0.015410587191581726, + 0.008648121729493141, + 0.006100768223404884, + 0.0035597477108240128, + 0.003376720007508993, + 0.002288710791617632, + 0.005025382153689861, + 0.0032722933683544397, + 0.0061167520470917225, + 0.007999461144208908, + 0.003218247089534998, + 0.003404720686376095, + 0.003933629486709833, + 0.0021775036584585905, + 0.0026611967477947474, + 0.007059919182211161, + 0.005722532048821449, + 0.002509047044441104, + 0.0025456715375185013, + 0.013888943940401077, + 0.014797838404774666, + 0.007847810164093971, + 0.011849486269056797, + 0.027586013078689575, + 0.025047406554222107, + 0.01856295019388199, + 0.009024295955896378, + 0.026761947199702263, + 0.017299890518188477, + 0.007593635469675064, + 0.011494191363453865, + 0.020405244082212448, + 0.06329819560050964, + 0.016619935631752014, + 0.012702974490821362, + 0.0059826853685081005, + 0.005050171632319689, + 0.003013771027326584, + 0.00533357635140419, + 0.00553177623078227, + 0.010988060384988785, + 0.0062943147495388985, + 0.0056341467425227165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 3 + ], + "selected_reward": -0.023280387744307518 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient general with a crimson complexion and a thick black beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general wearing bronze armor with thick shoulder pads and carved animal-face patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general gripping a massive crescent-shaped blade with his leather-guarded right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a battlefield filled with yellow sand and smoke under a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018182270228862762 + ], + "opposite_logprob_tokens": [ + -4.018182277679443 + ], + "logprob_reward": -0.018182270228862762, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018182270228862762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5941111445426941, + 0.04036007076501846, + 0.2125079184770584, + 0.06026897579431534, + 0.0114347068592906, + 0.008446775376796722, + 0.028691094368696213, + 0.010828457772731781, + 0.02637474611401558, + 0.006976109463721514 + ], + "credit_map": [ + 0.05960281193256378, + 0.04597049579024315, + 0.1363030970096588, + 0.19434654712677002, + 0.026520859450101852, + 0.047575440257787704, + 0.04496224224567413, + 0.03882969543337822, + 0.006632830947637558, + 0.006566844880580902, + 0.004741957411170006, + 0.007120734546333551, + 0.004244431387633085, + 0.002893861848860979, + 0.0040396228432655334, + 0.004119787830859423, + 0.009195784106850624, + 0.01017400249838829, + 0.004121553152799606, + 0.011550367809832096, + 0.007880273275077343, + 0.0917186439037323, + 0.006910661235451698, + 0.07095663994550705, + 0.0046281046234071255, + 0.009225469082593918, + 0.007941672578454018, + 0.01033965777605772, + 0.00584589596837759, + 0.009947474114596844, + 0.00701707461848855, + 0.005323633551597595, + 0.0010630751494318247, + 0.0017167668556794524, + 0.0012770190369337797, + 0.001414958736859262, + 0.002035405719652772, + 0.0017475294880568981, + 0.0013506787363439798, + 0.0008292733691632748, + 0.0007186605362221599, + 0.001265662140212953, + 0.0008468999294564128, + 0.0012237587943673134, + 0.0012533188564702868, + 0.000988306594081223, + 0.0011648231884464622, + 0.0009853458032011986, + 0.0032567400485277176, + 0.003959570080041885, + 0.0015869956696406007, + 0.0023620727006345987, + 0.007719978224486113, + 0.005446649622172117, + 0.002954899100586772, + 0.001404190668836236, + 0.0012384018627926707, + 0.0020253192633390427, + 0.0007030472042970359, + 0.0009293318726122379, + 0.0015200036577880383, + 0.0020168088376522064, + 0.00116444262675941, + 0.001231102622114122, + 0.0035064658150076866, + 0.0035860794596374035, + 0.0014676115242764354, + 0.0017386659746989608, + 0.003070227336138487, + 0.007985907606780529, + 0.0025546918623149395, + 0.002465098164975643, + 0.001027375110425055, + 0.000982254627160728, + 0.00043721028487198055, + 0.0005027095321565866, + 0.000814855913631618, + 0.002015928737819195, + 0.0007434063591063023, + 0.00045236884034238756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.018182270228862762 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient general with a crimson complexion and a thick black beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general wearing bronze armor with thick shoulder pads and carved animal-face patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general gripping a massive crescent-shaped blade with his leather-guarded right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a battlefield filled with yellow sand and smoke under a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3133981227874756 + ], + "opposite_logprob_tokens": [ + -1.3133981227874756 + ], + "logprob_reward": -0.3133981227874756, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3133981227874756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20225803554058075, + 0.02602921426296234, + 0.38867437839508057, + 0.169953390955925, + 0.015351945534348488, + 0.018764108419418335, + 0.031533382833004, + 0.031761594116687775, + 0.11052349954843521, + 0.005150441080331802 + ], + "credit_map": [ + 0.024080660194158554, + 0.01446377020329237, + 0.056308213621377945, + 0.06286311149597168, + 0.010089903138577938, + 0.010537896305322647, + 0.010680741630494595, + 0.013233746401965618, + 0.003538043238222599, + 0.005512708332389593, + 0.002995919669046998, + 0.004183168523013592, + 0.0028634879272431135, + 0.001672869548201561, + 0.00250584096647799, + 0.0027571774553507566, + 0.015276811085641384, + 0.017182426527142525, + 0.007639499381184578, + 0.021869011223316193, + 0.012967973947525024, + 0.16363166272640228, + 0.011892857030034065, + 0.138214111328125, + 0.012178130447864532, + 0.029537271708250046, + 0.02249055728316307, + 0.03027362748980522, + 0.016634056344628334, + 0.025648333132267, + 0.019618704915046692, + 0.013572710566222668, + 0.0015096249990165234, + 0.0034059486351907253, + 0.0016684202710166574, + 0.0026093011256307364, + 0.0017771178390830755, + 0.001356193097308278, + 0.0012516150018200278, + 0.0017737246816977859, + 0.002133323810994625, + 0.002594210673123598, + 0.001547243562527001, + 0.0019791575614362955, + 0.002395091112703085, + 0.003934134729206562, + 0.0022523063234984875, + 0.0019286409951746464, + 0.004125823266804218, + 0.0034096038434654474, + 0.0013551785377785563, + 0.0023257930297404528, + 0.010222584940493107, + 0.0062044113874435425, + 0.0025074465665966272, + 0.001382541609928012, + 0.003422398352995515, + 0.004979559686034918, + 0.002526708412915468, + 0.0034986601676791906, + 0.0055627888068556786, + 0.005515629425644875, + 0.003602535929530859, + 0.00265331263653934, + 0.016922978684306145, + 0.012237003073096275, + 0.004886702634394169, + 0.007321407552808523, + 0.013248695991933346, + 0.039500750601291656, + 0.009693785570561886, + 0.006712178699672222, + 0.0005649623344652355, + 0.0006553258863277733, + 0.00043785697198472917, + 0.0004962519742548466, + 0.000896821788046509, + 0.0010209084721282125, + 0.000627651228569448, + 0.0004506621335167438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.3133981227874756 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium, static shot, taken from a slightly upward angle, captures a commanding ancient general, his complexion crimson like a red date and his presence imposing. His thick, slender black beard falls to his chest, his eyes piercing with lightning. He clad in a finely crafted suit of bronze armor, featuring thick shoulder pads and intricately carved animal-face patterns. With his leather-guarded right hand, he grips the hilt of his massive crescent-shaped blade with a force that threatens to crush it. The veins in his massive forearm, tangled like tangled tree roots, are visible from the sheer force. Every detail reveals the immense power contained within. The background is a battlefield filled with yellow sand and smoke. The sky is gloomy, and the cold sunlight penetrates the thin clouds, casting sharp metallic reflections on his armor and cold blade. The main colors of the picture tend to be gloomy ochre and cold metallic colors. The whole scene is filled with the tension of impending storm and a powerful aura that is ready to explode.", + "seed": 30722, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_10/index_4930_seed_30722_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09004267323762179, + "vlm_logprob": -0.09004267323762179, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.09004267323762179, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient general with a crimson complexion and a thick black beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general wearing bronze armor with thick shoulder pads and carved animal-face patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general gripping a massive crescent-shaped blade with his leather-guarded right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a battlefield filled with yellow sand and smoke under a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016792723909020424 + ], + "opposite_logprob_tokens": [ + -4.1417927742004395 + ], + "logprob_reward": -0.016792723909020424, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016792723909020424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07621239870786667, + 0.015886934474110603, + 0.041404154151678085, + 0.04980612173676491, + 0.29782673716545105, + 0.031222550198435783, + 0.26621484756469727, + 0.17571745812892914, + 0.0266416072845459, + 0.01906716264784336 + ], + "credit_map": [ + 0.009869267232716084, + 0.007899045012891293, + 0.015256459824740887, + 0.00964977964758873, + 0.009131445549428463, + 0.010747282765805721, + 0.004754982423037291, + 0.008904136717319489, + 0.001643161871470511, + 0.0017756713787093759, + 0.0024897682014852762, + 0.0020553043577820063, + 0.0030577972065657377, + 0.0011225538328289986, + 0.0015621412312611938, + 0.002180536277592182, + 0.0038402096834033728, + 0.0039776102639734745, + 0.004628643859177828, + 0.007236181292682886, + 0.009291324764490128, + 0.0038297621067613363, + 0.003341298084706068, + 0.005259125493466854, + 0.003921744879335165, + 0.004115127958357334, + 0.00608646497130394, + 0.008268341422080994, + 0.00863988883793354, + 0.005868564359843731, + 0.00689274538308382, + 0.006013248115777969, + 0.020085984840989113, + 0.010016784071922302, + 0.030430639162659645, + 0.11315274983644485, + 0.055249836295843124, + 0.00995650328695774, + 0.020197968930006027, + 0.03873627632856369, + 0.00244503072462976, + 0.0030479654669761658, + 0.002870231168344617, + 0.003932399209588766, + 0.008457248099148273, + 0.002364552114158869, + 0.003251150716096163, + 0.004853971302509308, + 0.037308692932128906, + 0.014245455153286457, + 0.013899815268814564, + 0.04261980205774307, + 0.040102098137140274, + 0.011277446523308754, + 0.0162044744938612, + 0.0905570387840271, + 0.012958915904164314, + 0.012490734457969666, + 0.02032199688255787, + 0.047391634434461594, + 0.022915123030543327, + 0.009406983852386475, + 0.009878222830593586, + 0.040353864431381226, + 0.002741232980042696, + 0.001865701051428914, + 0.002687447937205434, + 0.003624029690399766, + 0.004509051330387592, + 0.002966940402984619, + 0.003464918117970228, + 0.00478228647261858, + 0.0015960120363160968, + 0.0020511772017925978, + 0.0029639648273587227, + 0.002380723599344492, + 0.00313794263638556, + 0.002886540722101927, + 0.001892355619929731, + 0.0021584464702755213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 7 + ], + "selected_reward": -0.016792723909020424 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient general with a crimson complexion and a thick black beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general wearing bronze armor with thick shoulder pads and carved animal-face patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general gripping a massive crescent-shaped blade with his leather-guarded right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a battlefield filled with yellow sand and smoke under a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004101318307220936 + ], + "opposite_logprob_tokens": [ + -5.504101276397705 + ], + "logprob_reward": -0.004101318307220936, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004101318307220936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1301337629556656, + 0.09888055920600891, + 0.06517976522445679, + 0.03893231227993965, + 0.3623637557029724, + 0.053389932960271835, + 0.09189930558204651, + 0.07955785095691681, + 0.02945871837437153, + 0.05020405724644661 + ], + "credit_map": [ + 0.015062776394188404, + 0.018980899825692177, + 0.023614337667822838, + 0.019047338515520096, + 0.018600281327962875, + 0.010201891884207726, + 0.01129913330078125, + 0.013327107764780521, + 0.016154028475284576, + 0.007475328631699085, + 0.0071061099879443645, + 0.01221421267837286, + 0.033835750073194504, + 0.0064281136728823185, + 0.006589305121451616, + 0.009077709168195724, + 0.006339751649647951, + 0.006005614530295134, + 0.0074395351111888885, + 0.008603001944720745, + 0.0125748161226511, + 0.00926463957875967, + 0.0049931881949305534, + 0.009959212504327297, + 0.0033496383111923933, + 0.0035469122231006622, + 0.004487296566367149, + 0.005192902870476246, + 0.007827657274901867, + 0.006074086297303438, + 0.004312152974307537, + 0.004141664132475853, + 0.025187965482473373, + 0.010823669843375683, + 0.03684529289603233, + 0.13981562852859497, + 0.0676230639219284, + 0.010588194243609905, + 0.02438284456729889, + 0.04709707573056221, + 0.006717396434396505, + 0.005596895236521959, + 0.0063534090295434, + 0.006245859432965517, + 0.008206530474126339, + 0.0034274118952453136, + 0.004422337748110294, + 0.012420091778039932, + 0.01225508563220501, + 0.005210492759943008, + 0.005662926472723484, + 0.015460778959095478, + 0.012237491086125374, + 0.005154269747436047, + 0.00533882575109601, + 0.030579442158341408, + 0.005881527904421091, + 0.006768784951418638, + 0.011940281838178635, + 0.01372339017689228, + 0.015661969780921936, + 0.005882124416530132, + 0.007530596107244492, + 0.012169175781309605, + 0.0027462763246148825, + 0.0022320207208395004, + 0.003911309875547886, + 0.004708535969257355, + 0.004248171113431454, + 0.002758253365755081, + 0.004146581050008535, + 0.004707569256424904, + 0.00470479903742671, + 0.005326048005372286, + 0.006806134711951017, + 0.007479021791368723, + 0.0077101062051951885, + 0.00363847310654819, + 0.007994911633431911, + 0.006544569041579962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.004101318307220936 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient general with a crimson complexion and a thick black beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general wearing bronze armor with thick shoulder pads and carved animal-face patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general gripping a massive crescent-shaped blade with his leather-guarded right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a battlefield filled with yellow sand and smoke under a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012532057240605354 + ], + "opposite_logprob_tokens": [ + -4.3875322341918945 + ], + "logprob_reward": -0.012532057240605354, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012532057240605354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12782742083072662, + 0.06350978463888168, + 0.10601910203695297, + 0.04598262906074524, + 0.2711842656135559, + 0.08418183773756027, + 0.10192783921957016, + 0.08639397472143173, + 0.051654741168022156, + 0.06131839007139206 + ], + "credit_map": [ + 0.009287342429161072, + 0.017637629061937332, + 0.02393222600221634, + 0.016365399584174156, + 0.01817101053893566, + 0.012535905465483665, + 0.013162828981876373, + 0.01673506759107113, + 0.009936455637216568, + 0.004185039084404707, + 0.005018888507038355, + 0.007327223662286997, + 0.015614802949130535, + 0.004583100788295269, + 0.006235241889953613, + 0.010609031654894352, + 0.009555996395647526, + 0.010520043782889843, + 0.011485227383673191, + 0.02013571187853813, + 0.02011597715318203, + 0.011868248693645, + 0.009457293897867203, + 0.012880600057542324, + 0.003733177436515689, + 0.00382772134616971, + 0.0065392679534852505, + 0.008524148724973202, + 0.007848642766475677, + 0.0053590284660458565, + 0.004580025561153889, + 0.005570615641772747, + 0.01725548878312111, + 0.01190789882093668, + 0.025416571646928787, + 0.09859716892242432, + 0.05063123628497124, + 0.013937427662312984, + 0.01680946536362171, + 0.03662902116775513, + 0.007863178849220276, + 0.010736426338553429, + 0.009283722378313541, + 0.009688761085271835, + 0.016512859612703323, + 0.00937726255506277, + 0.005754930432885885, + 0.014964696951210499, + 0.012844828888773918, + 0.007024636026471853, + 0.010450190864503384, + 0.014697813428938389, + 0.017440546303987503, + 0.007467861287295818, + 0.007006899453699589, + 0.02499506063759327, + 0.006926416885107756, + 0.00744116585701704, + 0.008292723447084427, + 0.01812310330569744, + 0.015900690108537674, + 0.005418207962065935, + 0.005488371476531029, + 0.018803300336003304, + 0.0045871613547205925, + 0.0054078418761491776, + 0.00719190388917923, + 0.008796891197562218, + 0.009176412597298622, + 0.0041588107123970985, + 0.006838710512965918, + 0.005497011821717024, + 0.0051154629327356815, + 0.006081657018512487, + 0.009407063946127892, + 0.010259022004902363, + 0.010095072910189629, + 0.0048125190660357475, + 0.007902738638222218, + 0.007644850295037031 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.012532057240605354 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient general with a crimson complexion and a thick black beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general wearing bronze armor with thick shoulder pads and carved animal-face patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general gripping a massive crescent-shaped blade with his leather-guarded right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a battlefield filled with yellow sand and smoke under a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978825569152832 + ], + "opposite_logprob_tokens": [ + -3.5297882556915283 + ], + "logprob_reward": -0.02978825569152832, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978825569152832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2648463249206543, + 0.039153799414634705, + 0.030675683170557022, + 0.07911860942840576, + 0.1403905153274536, + 0.0491192489862442, + 0.06373688578605652, + 0.14385424554347992, + 0.05966990068554878, + 0.1294347196817398 + ], + "credit_map": [ + 0.022617993876338005, + 0.022978445515036583, + 0.028043441474437714, + 0.037745401263237, + 0.03434981778264046, + 0.0350579209625721, + 0.028518706560134888, + 0.05553458258509636, + 0.005015999544411898, + 0.003879878669977188, + 0.003919561859220266, + 0.005684328265488148, + 0.01016614306718111, + 0.002496131230145693, + 0.002956679556518793, + 0.00503507861867547, + 0.002361955586820841, + 0.0025858236476778984, + 0.0032382239587605, + 0.007646968122571707, + 0.005790745373815298, + 0.002687244676053524, + 0.002878520404919982, + 0.003486199537292123, + 0.005617646500468254, + 0.010222546756267548, + 0.007729141041636467, + 0.013727286830544472, + 0.01740925759077072, + 0.00866222195327282, + 0.006990351714193821, + 0.00876015517860651, + 0.010095621459186077, + 0.007211036514490843, + 0.015544641762971878, + 0.047481928020715714, + 0.02384207211434841, + 0.009285656735301018, + 0.010142097249627113, + 0.016787471249699593, + 0.004576089791953564, + 0.004360978491604328, + 0.00655911210924387, + 0.008435268886387348, + 0.009396976791322231, + 0.004083344712853432, + 0.0037298768293112516, + 0.007977601140737534, + 0.005268760956823826, + 0.00599240604788065, + 0.006305586080998182, + 0.010572196915745735, + 0.012273779138922691, + 0.006050428841263056, + 0.0060633947141468525, + 0.011210331693291664, + 0.012035133317112923, + 0.012180555611848831, + 0.025758450850844383, + 0.024004386737942696, + 0.020444227382540703, + 0.013392863795161247, + 0.013900450430810452, + 0.022138185799121857, + 0.007225040812045336, + 0.005984001327306032, + 0.007443008944392204, + 0.009586403146386147, + 0.008152681402862072, + 0.004002727102488279, + 0.007410567253828049, + 0.00986547116190195, + 0.010784192010760307, + 0.014681730419397354, + 0.022882401943206787, + 0.02188590168952942, + 0.017167404294013977, + 0.011783624067902565, + 0.015818364918231964, + 0.014431100338697433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.02978825569152832 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient general with a crimson complexion and a thick black beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general wearing bronze armor with thick shoulder pads and carved animal-face patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general gripping a massive crescent-shaped blade with his leather-guarded right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a battlefield filled with yellow sand and smoke under a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3869990110397339 + ], + "opposite_logprob_tokens": [ + -1.1369990110397339 + ], + "logprob_reward": -0.3869990110397339, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3869990110397339, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07298275828361511, + 0.16247442364692688, + 0.09315175563097, + 0.09319346398115158, + 0.07382523268461227, + 0.03971574455499649, + 0.05713857710361481, + 0.2945724427700043, + 0.05556328967213631, + 0.057382356375455856 + ], + "credit_map": [ + 0.007637463975697756, + 0.015021445229649544, + 0.010671150870621204, + 0.010694815777242184, + 0.010987897403538227, + 0.005259137600660324, + 0.005888979882001877, + 0.006821867544203997, + 0.030315900221467018, + 0.009774458594620228, + 0.010201106779277325, + 0.02008882910013199, + 0.059015508741140366, + 0.009902416728436947, + 0.008361292071640491, + 0.014814892783761024, + 0.006638228893280029, + 0.0062094456516206264, + 0.007864225655794144, + 0.014264159835875034, + 0.01911994256079197, + 0.01467693317681551, + 0.007256036624312401, + 0.017122777178883553, + 0.008084533736109734, + 0.01079977210611105, + 0.012529919855296612, + 0.01512644812464714, + 0.017253750935196877, + 0.011560902930796146, + 0.007731813006103039, + 0.01010632049292326, + 0.005791449919342995, + 0.005268172826617956, + 0.006383486092090607, + 0.01573755592107773, + 0.013029396533966064, + 0.013952834531664848, + 0.004941808059811592, + 0.008720524609088898, + 0.005445484071969986, + 0.0037993849255144596, + 0.006510044913738966, + 0.005708681419491768, + 0.007106798235327005, + 0.003134471131488681, + 0.0027056620456278324, + 0.005305218975991011, + 0.006227548699826002, + 0.004519638139754534, + 0.005981743801385164, + 0.008641551248729229, + 0.010370248928666115, + 0.004038546234369278, + 0.004096664022654295, + 0.013262633234262466, + 0.02148991823196411, + 0.023694658651947975, + 0.03802376985549927, + 0.07380547374486923, + 0.03589492291212082, + 0.018199067562818527, + 0.023244667798280716, + 0.06021996587514877, + 0.007960249669849873, + 0.00576361920684576, + 0.0065718889236450195, + 0.0070235771127045155, + 0.006641930900514126, + 0.0035725319758057594, + 0.003989716060459614, + 0.014039775356650352, + 0.00584699772298336, + 0.006235155742615461, + 0.007365486118942499, + 0.010163557715713978, + 0.006823723670095205, + 0.004390212241560221, + 0.009712561964988708, + 0.006844659801572561 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 3 + ], + "selected_reward": -0.3869990110397339 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium, static shot, taken from a slightly upward angle, captures a commanding ancient general, his complexion crimson like a red date and his presence imposing. His thick, slender black beard falls to his chest, his eyes piercing with lightning. He clad in a finely crafted suit of bronze armor, featuring thick shoulder pads and intricately carved animal-face patterns. With his leather-guarded right hand, he grips the hilt of his massive crescent-shaped blade with a force that threatens to crush it. The veins in his massive forearm, tangled like tangled tree roots, are visible from the sheer force. Every detail reveals the immense power contained within. The background is a battlefield filled with yellow sand and smoke. The sky is gloomy, and the cold sunlight penetrates the thin clouds, casting sharp metallic reflections on his armor and cold blade. The main colors of the picture tend to be gloomy ochre and cold metallic colors. The whole scene is filled with the tension of impending storm and a powerful aura that is ready to explode.", + "seed": 30723, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_10/index_4930_seed_30723_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.020641452632844448, + "vlm_logprob": -0.020641452632844448, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.020641452632844448, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient general with a crimson complexion and a thick black beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general wearing bronze armor with thick shoulder pads and carved animal-face patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general gripping a massive crescent-shaped blade with his leather-guarded right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a battlefield filled with yellow sand and smoke under a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02152346819639206 + ], + "opposite_logprob_tokens": [ + -3.8965234756469727 + ], + "logprob_reward": -0.02152346819639206, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02152346819639206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2017161250114441, + 0.08923199772834778, + 0.09204110503196716, + 0.16296488046646118, + 0.0658053308725357, + 0.05453933775424957, + 0.07715332508087158, + 0.057400237768888474, + 0.13138140738010406, + 0.0677662342786789 + ], + "credit_map": [ + 0.02030724473297596, + 0.022659584879875183, + 0.038862887769937515, + 0.01770637556910515, + 0.02754977159202099, + 0.03165523335337639, + 0.020607849583029747, + 0.02236720360815525, + 0.00964432768523693, + 0.008249001577496529, + 0.01422718446701765, + 0.009542137384414673, + 0.011944492347538471, + 0.01268377248197794, + 0.010578375309705734, + 0.012362713925540447, + 0.011617629788815975, + 0.0078009446151554585, + 0.018864421173930168, + 0.008116086944937706, + 0.01283747423440218, + 0.010829842649400234, + 0.009859784506261349, + 0.012114926241338253, + 0.018291322514414787, + 0.020036833360791206, + 0.019164010882377625, + 0.02022816427052021, + 0.030185235664248466, + 0.013694167137145996, + 0.017360905185341835, + 0.02400425635278225, + 0.008079985156655312, + 0.006911515723913908, + 0.010547145269811153, + 0.005767639260739088, + 0.009940841235220432, + 0.010725014843046665, + 0.0071755689568817616, + 0.006657624617218971, + 0.006487753242254257, + 0.006009874865412712, + 0.009459749795496464, + 0.0051206364296376705, + 0.00801598746329546, + 0.006552813109010458, + 0.005787610542029142, + 0.00710491556674242, + 0.00994007010012865, + 0.008060150779783726, + 0.01674853079020977, + 0.0077633000910282135, + 0.009549160487949848, + 0.009233739227056503, + 0.006932681892067194, + 0.00892570149153471, + 0.0076228859834373, + 0.00725591741502285, + 0.010553582571446896, + 0.005561613943427801, + 0.008841280825436115, + 0.005599752999842167, + 0.006030884571373463, + 0.005934324115514755, + 0.01601998880505562, + 0.015280433930456638, + 0.025611860677599907, + 0.012620371766388416, + 0.02113170549273491, + 0.01700470596551895, + 0.01214243471622467, + 0.011569907888770103, + 0.007375903427600861, + 0.008328710682690144, + 0.008232025429606438, + 0.00735773891210556, + 0.01460763718932867, + 0.008223157376050949, + 0.007997037842869759, + 0.005644027143716812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.02152346819639206 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient general with a crimson complexion and a thick black beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general wearing bronze armor with thick shoulder pads and carved animal-face patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general gripping a massive crescent-shaped blade with his leather-guarded right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a battlefield filled with yellow sand and smoke under a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01421845331788063 + ], + "opposite_logprob_tokens": [ + -4.264218330383301 + ], + "logprob_reward": -0.01421845331788063, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01421845331788063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2079509049654007, + 0.07603263109922409, + 0.06845651566982269, + 0.07561018317937851, + 0.05111461505293846, + 0.047916561365127563, + 0.05309252440929413, + 0.12832987308502197, + 0.16328983008861542, + 0.12820638716220856 + ], + "credit_map": [ + 0.02082671783864498, + 0.023268720135092735, + 0.03855488821864128, + 0.020946672186255455, + 0.029656510800123215, + 0.028962641954421997, + 0.024393389001488686, + 0.021341366693377495, + 0.008797180838882923, + 0.008980846032500267, + 0.01636154018342495, + 0.008109958842396736, + 0.012190015986561775, + 0.006881686393171549, + 0.009098613634705544, + 0.0056127943098545074, + 0.007609378546476364, + 0.007810460403561592, + 0.013798600994050503, + 0.005747142713516951, + 0.011220520362257957, + 0.008180616423487663, + 0.008790389634668827, + 0.005299407057464123, + 0.008653930388391018, + 0.010647289454936981, + 0.011963200755417347, + 0.009059783071279526, + 0.01173257827758789, + 0.00897122174501419, + 0.007625901140272617, + 0.006956279277801514, + 0.005710107274353504, + 0.008170515298843384, + 0.008539709262549877, + 0.005130356643348932, + 0.0067992135882377625, + 0.006006515119224787, + 0.005357826594263315, + 0.005400368943810463, + 0.0054877451620996, + 0.0061185299418866634, + 0.008229240775108337, + 0.00432920316234231, + 0.006802414543926716, + 0.006547244731336832, + 0.004865250550210476, + 0.005536934360861778, + 0.006980931386351585, + 0.00570527371019125, + 0.009354914538562298, + 0.004078903701156378, + 0.006751688662916422, + 0.007582725491374731, + 0.005169063340872526, + 0.007469020783901215, + 0.017824305221438408, + 0.019285382702946663, + 0.024527644738554955, + 0.011479660868644714, + 0.012643842026591301, + 0.020756827667355537, + 0.010576597414910793, + 0.011235610581934452, + 0.02176891453564167, + 0.016882551833987236, + 0.02411336451768875, + 0.01656053401529789, + 0.02835955284535885, + 0.024518199265003204, + 0.015021705068647861, + 0.016065018251538277, + 0.014185942709445953, + 0.017019668594002724, + 0.011301177553832531, + 0.015328365378081799, + 0.03218180686235428, + 0.015586074441671371, + 0.014659583568572998, + 0.0079437755048275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.01421845331788063 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient general with a crimson complexion and a thick black beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general wearing bronze armor with thick shoulder pads and carved animal-face patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general gripping a massive crescent-shaped blade with his leather-guarded right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a battlefield filled with yellow sand and smoke under a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014194007031619549 + ], + "opposite_logprob_tokens": [ + -4.264194011688232 + ], + "logprob_reward": -0.014194007031619549, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014194007031619549, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14071044325828552, + 0.08013391494750977, + 0.06796149164438248, + 0.05281967297196388, + 0.074373759329319, + 0.0790754035115242, + 0.0678730458021164, + 0.07990631461143494, + 0.2717630863189697, + 0.08538290113210678 + ], + "credit_map": [ + 0.01747129112482071, + 0.014727088622748852, + 0.03224167600274086, + 0.010129662230610847, + 0.01980932056903839, + 0.021969400346279144, + 0.012458092533051968, + 0.011903917416930199, + 0.008317356929183006, + 0.00860578566789627, + 0.013758216984570026, + 0.007509274408221245, + 0.01641797088086605, + 0.008811564184725285, + 0.009516432881355286, + 0.007197315338999033, + 0.009502498432993889, + 0.0074218278750777245, + 0.007766639348119497, + 0.007962030358612537, + 0.00918719731271267, + 0.007534285541623831, + 0.010267170146107674, + 0.008319839835166931, + 0.007273517549037933, + 0.004351979587227106, + 0.014169512316584587, + 0.005116602871567011, + 0.007676581386476755, + 0.006755211390554905, + 0.0035603572614490986, + 0.003915911540389061, + 0.008873025886714458, + 0.006646002642810345, + 0.011749411933124065, + 0.006957660894840956, + 0.010806970298290253, + 0.011151475831866264, + 0.007616318296641111, + 0.010572889819741249, + 0.010599233210086823, + 0.006855654995888472, + 0.016475405544042587, + 0.007217475678771734, + 0.007477212697267532, + 0.01245169248431921, + 0.006791119463741779, + 0.011207614094018936, + 0.009264626540243626, + 0.009947432205080986, + 0.01406280230730772, + 0.006727483589202166, + 0.008546443656086922, + 0.008238180540502071, + 0.004384965170174837, + 0.0067011164501309395, + 0.009238185361027718, + 0.007714126259088516, + 0.011878721415996552, + 0.007423378061503172, + 0.011346745304763317, + 0.013898590579628944, + 0.006690075155347586, + 0.011716497130692005, + 0.03213032707571983, + 0.02661880850791931, + 0.0529434010386467, + 0.026662085205316544, + 0.03256687521934509, + 0.037762802094221115, + 0.03074220009148121, + 0.032336581498384476, + 0.009959324263036251, + 0.0095552122220397, + 0.0074401069432497025, + 0.010614958591759205, + 0.019364533945918083, + 0.01017430517822504, + 0.010429448448121548, + 0.007845012471079826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.014194007031619549 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient general with a crimson complexion and a thick black beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general wearing bronze armor with thick shoulder pads and carved animal-face patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general gripping a massive crescent-shaped blade with his leather-guarded right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a battlefield filled with yellow sand and smoke under a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004635065793991089 + ], + "opposite_logprob_tokens": [ + -5.379634857177734 + ], + "logprob_reward": -0.004635065793991089, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004635065793991089, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10482929646968842, + 0.08802026510238647, + 0.13026681542396545, + 0.049856998026371, + 0.14542940258979797, + 0.0826144590973854, + 0.06781301647424698, + 0.019977690652012825, + 0.07278037071228027, + 0.23841169476509094 + ], + "credit_map": [ + 0.011626987718045712, + 0.009424659423530102, + 0.020166289061307907, + 0.009825291112065315, + 0.012181619182229042, + 0.01577656902372837, + 0.012513007037341595, + 0.013314872980117798, + 0.010947613045573235, + 0.006637279875576496, + 0.020311500877141953, + 0.009500809013843536, + 0.009380546398460865, + 0.014031503349542618, + 0.009261267259716988, + 0.007949749939143658, + 0.02134530059993267, + 0.01023386511951685, + 0.02821740135550499, + 0.011761660687625408, + 0.012546314857900143, + 0.025134440511465073, + 0.010912993922829628, + 0.010114850476384163, + 0.006509582512080669, + 0.005702187307178974, + 0.008752971887588501, + 0.004574484657496214, + 0.007506809663027525, + 0.008025364018976688, + 0.0041470834985375404, + 0.004638512618839741, + 0.01909923553466797, + 0.011793388985097408, + 0.023777863010764122, + 0.011370532214641571, + 0.017423639073967934, + 0.028844540938735008, + 0.012370810844004154, + 0.02074938826262951, + 0.01122442726045847, + 0.007946285419166088, + 0.015497591346502304, + 0.008679702877998352, + 0.009507684037089348, + 0.015534739010035992, + 0.008146852254867554, + 0.006077174562960863, + 0.007846665568649769, + 0.008457791991531849, + 0.010914618149399757, + 0.007565440144389868, + 0.010307584889233112, + 0.008891462348401546, + 0.007498502731323242, + 0.00633094971999526, + 0.0026622682344168425, + 0.0020800279453396797, + 0.005421672947704792, + 0.0015781905967742205, + 0.002421316923573613, + 0.002824394730851054, + 0.0014599493006244302, + 0.0015298681100830436, + 0.010170364752411842, + 0.009712021797895432, + 0.012556856498122215, + 0.006526811979711056, + 0.010205212980508804, + 0.011259689927101135, + 0.006739649921655655, + 0.005609763320535421, + 0.0321311391890049, + 0.02681748941540718, + 0.037604980170726776, + 0.02180447056889534, + 0.032630544155836105, + 0.03929906710982323, + 0.019614754244685173, + 0.02850925736129284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 2 + ], + "selected_reward": -0.004635065793991089 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient general with a crimson complexion and a thick black beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general wearing bronze armor with thick shoulder pads and carved animal-face patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general gripping a massive crescent-shaped blade with his leather-guarded right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a battlefield filled with yellow sand and smoke under a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863626882433891 + ], + "opposite_logprob_tokens": [ + -3.0486361980438232 + ], + "logprob_reward": -0.04863626882433891, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863626882433891, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.260132759809494, + 0.07760697603225708, + 0.07646206021308899, + 0.09274633973836899, + 0.06851588189601898, + 0.05731582269072533, + 0.046303533017635345, + 0.06826086342334747, + 0.14375165104866028, + 0.10890413820743561 + ], + "credit_map": [ + 0.02789793349802494, + 0.02327951043844223, + 0.031995389610528946, + 0.030897488817572594, + 0.03721877932548523, + 0.03380342200398445, + 0.03606396168470383, + 0.03897624462842941, + 0.008073735050857067, + 0.00948434416204691, + 0.012347224168479443, + 0.009635575115680695, + 0.009426301345229149, + 0.010723525658249855, + 0.008945346809923649, + 0.008970926515758038, + 0.009901386685669422, + 0.007363113574683666, + 0.016679057851433754, + 0.008801247924566269, + 0.009645624086260796, + 0.010198215954005718, + 0.008314203470945358, + 0.005559206940233707, + 0.013035907410085201, + 0.009995547123253345, + 0.01926671713590622, + 0.009953703731298447, + 0.011957776732742786, + 0.010728946886956692, + 0.007571021560579538, + 0.010236716829240322, + 0.00824613869190216, + 0.007281175348907709, + 0.010322687216103077, + 0.006738077383488417, + 0.013706563971936703, + 0.009163562208414078, + 0.007078905124217272, + 0.005978775676339865, + 0.007942958734929562, + 0.0048662894405424595, + 0.011868048459291458, + 0.005319719202816486, + 0.008150083012878895, + 0.009241752326488495, + 0.004610189236700535, + 0.0053167869336903095, + 0.005848912987858057, + 0.0056573087349534035, + 0.0073867300525307655, + 0.003628309816122055, + 0.006319430656731129, + 0.007283071521669626, + 0.004832641687244177, + 0.005347128491848707, + 0.007683749310672283, + 0.007086710073053837, + 0.014160319231450558, + 0.006791756488382816, + 0.007812224328517914, + 0.006973194889724255, + 0.0071173072792589664, + 0.010635597631335258, + 0.018282484263181686, + 0.012478819116950035, + 0.026354704052209854, + 0.018128838390111923, + 0.018627069890499115, + 0.023816466331481934, + 0.014386861585080624, + 0.011676410213112831, + 0.013316182419657707, + 0.014025251381099224, + 0.01554306410253048, + 0.009413782507181168, + 0.019903196021914482, + 0.016144197434186935, + 0.009093300439417362, + 0.011465166695415974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.04863626882433891 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_11/index_tensor([4046, 4046])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_11/index_tensor([4046, 4046])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e6842d8a88a2339b088dd00eec6549c4d6d6567c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_11/index_tensor([4046, 4046])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, rotating tracking shot, set in a Stardew Valley-esque pixel world, captures a \"Yangtianxuelu\" tea spirit, composed of emerald green pixels. Composed of varying shades of green, bordered by darker dark green pixels and dotted with two large, gleaming black pixel eyes, it gracefully leaps from the top of the screen in a graceful arc. The camera rotates gracefully around it, following its slow tumble through the air, ultimately gracing a massive, pixelated ceramic teacup resting atop a warm, wood-grained pixel tabletop. The off-white cup, massive and shimmering with a soft, pixelated sheen, is filled with amber tea, wisps of steam rising from it composed of white and light gray pixels. The tea spirit descends gracefully, its tip delicately touching the surface of the steaming tea, instantly sending ripples of tiny cubes radiating outward. The entire scene is bathed in soft, warm indoor light, creating an atmosphere both magical and deeply warm and tranquil.", + "seed": 681, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_11/index_4046_seed_681_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06894646817818284, + "vlm_logprob": -0.06894646817818284, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.06894646817818284, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tea spirit composed of emerald green pixels with two large black pixel eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tea spirit leap from the top of the screen in a graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera rotating around the tea spirit while it tumbles through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tea spirit land in a massive off-white pixelated ceramic teacup filled with amber tea and white/light gray steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny cube ripples radiate outward from the tea surface when the spirit touches it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011322175152599812 + ], + "opposite_logprob_tokens": [ + -4.636322021484375 + ], + "logprob_reward": -0.011322175152599812, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.011322175152599812, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.510333776473999, + 0.19839966297149658, + 0.04824420064687729, + 0.033388666808605194, + 0.03245072066783905, + 0.026403460651636124, + 0.052026934921741486, + 0.04742947965860367, + 0.030662162229418755, + 0.02066092938184738 + ], + "credit_map": [ + 0.0374680794775486, + 0.022171538323163986, + 0.035343535244464874, + 0.09265770018100739, + 0.033224500715732574, + 0.06534842401742935, + 0.17717373371124268, + 0.04694626107811928, + 0.02669944427907467, + 0.04394160211086273, + 0.02616778388619423, + 0.021489733830094337, + 0.028632409870624542, + 0.01364878099411726, + 0.01999114453792572, + 0.01782876066863537, + 0.004809962119907141, + 0.005065129138529301, + 0.00434254202991724, + 0.006848755292594433, + 0.005614466965198517, + 0.0063935937359929085, + 0.00680225295946002, + 0.00836749654263258, + 0.0020687486976385117, + 0.0035129657480865717, + 0.0030828046146780252, + 0.00489385612308979, + 0.00708965677767992, + 0.0036779914516955614, + 0.004469500854611397, + 0.0045931399799883366, + 0.004087081644684076, + 0.004008710850030184, + 0.004169732332229614, + 0.003437525825574994, + 0.004798773676156998, + 0.004757091403007507, + 0.003209197660908103, + 0.003982606343924999, + 0.0035534962080419064, + 0.0017151060746982694, + 0.003420336404815316, + 0.003914349712431431, + 0.005095259286463261, + 0.0026771987322717905, + 0.00284113478846848, + 0.003186580026522279, + 0.008956985548138618, + 0.0061266678385436535, + 0.008659451268613338, + 0.0049835736863315105, + 0.004679420031607151, + 0.006603439804166555, + 0.003983830567449331, + 0.008033565245568752, + 0.004584731999784708, + 0.008182263001799583, + 0.006209301296621561, + 0.0038037991616874933, + 0.008379418402910233, + 0.004415726754814386, + 0.004467819817364216, + 0.007386420853435993, + 0.0029555996879935265, + 0.002846579998731613, + 0.007231688592582941, + 0.002298240549862385, + 0.001908553997054696, + 0.004076283425092697, + 0.002008981304243207, + 0.007336236536502838, + 0.0021184408105909824, + 0.0035241765435785055, + 0.0019461401971057057, + 0.003770848037675023, + 0.0022500467021018267, + 0.0023258677683770657, + 0.001874878304079175, + 0.0028505295049399137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.011322175152599812 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tea spirit composed of emerald green pixels with two large black pixel eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tea spirit leap from the top of the screen in a graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera rotating around the tea spirit while it tumbles through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tea spirit land in a massive off-white pixelated ceramic teacup filled with amber tea and white/light gray steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny cube ripples radiate outward from the tea surface when the spirit touches it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048674989491701126 + ], + "opposite_logprob_tokens": [ + -3.048675060272217 + ], + "logprob_reward": -0.048674989491701126, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048674989491701126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3137628436088562, + 0.2451462596654892, + 0.08384987711906433, + 0.13000532984733582, + 0.03695430979132652, + 0.04521432891488075, + 0.06592827290296555, + 0.026038769632577896, + 0.027023445814847946, + 0.026076601818203926 + ], + "credit_map": [ + 0.023947669193148613, + 0.03885393962264061, + 0.02426626719534397, + 0.04915599152445793, + 0.03310712054371834, + 0.035618461668491364, + 0.08625161647796631, + 0.022561734542250633, + 0.027658458799123764, + 0.0296963881701231, + 0.03920326381921768, + 0.034467052668333054, + 0.02647016942501068, + 0.01974937878549099, + 0.033911410719156265, + 0.03399012237787247, + 0.006169257685542107, + 0.008086987771093845, + 0.007251253817230463, + 0.012222236953675747, + 0.010903933085501194, + 0.009312550537288189, + 0.014566999860107899, + 0.015336643904447556, + 0.008412002585828304, + 0.011731600388884544, + 0.009688855148851871, + 0.0221724733710289, + 0.02972140535712242, + 0.013127975165843964, + 0.021815817803144455, + 0.013335189782083035, + 0.0030520118307322264, + 0.004077923018485308, + 0.003041779389604926, + 0.003628814360126853, + 0.010265757329761982, + 0.0037862418685108423, + 0.005205762106925249, + 0.003896016161888838, + 0.007487637922167778, + 0.0035491837188601494, + 0.006443789694458246, + 0.005618513096123934, + 0.008695981465280056, + 0.0031592778395861387, + 0.0034336699172854424, + 0.006826272699981928, + 0.010093007236719131, + 0.005731284152716398, + 0.008850052952766418, + 0.0072700646705925465, + 0.0068238601088523865, + 0.00894572027027607, + 0.005379105918109417, + 0.012835171073675156, + 0.002017027698457241, + 0.0046717096120119095, + 0.002789735095575452, + 0.002379660727456212, + 0.004281878471374512, + 0.004132831934839487, + 0.0019488577963784337, + 0.0038170653861016035, + 0.0019002911867573857, + 0.004468183033168316, + 0.0029923676047474146, + 0.004579984117299318, + 0.003497126279398799, + 0.0028306450694799423, + 0.0027577371802181005, + 0.003997110761702061, + 0.0033348475117236376, + 0.0030097365379333496, + 0.0041993348859250546, + 0.003461717162281275, + 0.002369108609855175, + 0.0035378062166273594, + 0.002822266425937414, + 0.003341781673952937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.048674989491701126 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tea spirit composed of emerald green pixels with two large black pixel eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tea spirit leap from the top of the screen in a graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera rotating around the tea spirit while it tumbles through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tea spirit land in a massive off-white pixelated ceramic teacup filled with amber tea and white/light gray steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny cube ripples radiate outward from the tea surface when the spirit touches it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.252111554145813 + ], + "opposite_logprob_tokens": [ + -1.502111554145813 + ], + "logprob_reward": -0.252111554145813, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.252111554145813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2796159088611603, + 0.18080604076385498, + 0.10242471098899841, + 0.06997086107730865, + 0.16308268904685974, + 0.03101147897541523, + 0.024693971499800682, + 0.08384113758802414, + 0.02978229708969593, + 0.03477087989449501 + ], + "credit_map": [ + 0.0347144789993763, + 0.03470718115568161, + 0.033631518483161926, + 0.04048696532845497, + 0.03248237818479538, + 0.03784772381186485, + 0.03284277021884918, + 0.03290291130542755, + 0.02577199786901474, + 0.025450080633163452, + 0.03385930880904198, + 0.01803499273955822, + 0.027126694098114967, + 0.014546645805239677, + 0.01822713017463684, + 0.01778917945921421, + 0.009113345295190811, + 0.009624924510717392, + 0.006483256351202726, + 0.020117560401558876, + 0.0129162622615695, + 0.007009417284280062, + 0.025991715490818024, + 0.011168226599693298, + 0.005791211500763893, + 0.0079122856259346, + 0.00582157215103507, + 0.008626340888440609, + 0.011890342459082603, + 0.011859580874443054, + 0.009009081870317459, + 0.009060446172952652, + 0.009814703837037086, + 0.015938978642225266, + 0.017107676714658737, + 0.023606877774000168, + 0.04169263318181038, + 0.016347941011190414, + 0.01895519718527794, + 0.019618671387434006, + 0.004454885143786669, + 0.0044498383067548275, + 0.0038504397962242365, + 0.004181317985057831, + 0.0035921065136790276, + 0.0035267327912151814, + 0.002946781227365136, + 0.004009376745671034, + 0.002240987028926611, + 0.004796133376657963, + 0.0032056637573987246, + 0.0034932377748191357, + 0.0029137160163372755, + 0.0029974656645208597, + 0.0023879234213382006, + 0.002658843295648694, + 0.004364322405308485, + 0.018059637397527695, + 0.010418177582323551, + 0.005554779898375273, + 0.017659245058894157, + 0.00834573619067669, + 0.005795313511043787, + 0.013643925078213215, + 0.0035777008160948753, + 0.004377009812742472, + 0.005201995838433504, + 0.0033592109102755785, + 0.00280493451282382, + 0.0035480062942951918, + 0.00277935853227973, + 0.0041340794414281845, + 0.004041730426251888, + 0.0058323475532233715, + 0.004903582856059074, + 0.004083919804543257, + 0.003836806397885084, + 0.0036837784573435783, + 0.004031741060316563, + 0.004356972873210907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.252111554145813 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tea spirit composed of emerald green pixels with two large black pixel eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tea spirit leap from the top of the screen in a graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera rotating around the tea spirit while it tumbles through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tea spirit land in a massive off-white pixelated ceramic teacup filled with amber tea and white/light gray steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny cube ripples radiate outward from the tea surface when the spirit touches it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028203255496919155 + ], + "opposite_logprob_tokens": [ + -5.8778204917907715 + ], + "logprob_reward": -0.0028203255496919155, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028203255496919155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2108638882637024, + 0.07813576608896255, + 0.04869456589221954, + 0.07395534962415695, + 0.05415381118655205, + 0.033781498670578, + 0.093259796500206, + 0.15809878706932068, + 0.21239008009433746, + 0.03666643798351288 + ], + "credit_map": [ + 0.02172836847603321, + 0.01764659956097603, + 0.015731653198599815, + 0.041414808481931686, + 0.01749223843216896, + 0.018055442720651627, + 0.062390074133872986, + 0.01640469953417778, + 0.009371927008032799, + 0.009465797804296017, + 0.014190218411386013, + 0.009536907076835632, + 0.009650825522840023, + 0.009483827278017998, + 0.008320341818034649, + 0.008115926757454872, + 0.003442660439759493, + 0.002814584644511342, + 0.003333279164507985, + 0.010684767737984657, + 0.0044015501625835896, + 0.006592185236513615, + 0.01171999890357256, + 0.005705541465431452, + 0.005474398843944073, + 0.007052777335047722, + 0.006010159850120544, + 0.010194383561611176, + 0.020975571125745773, + 0.00617580208927393, + 0.010481061413884163, + 0.0075911893509328365, + 0.003547555999830365, + 0.008493009954690933, + 0.008205137215554714, + 0.006513835862278938, + 0.005992426536977291, + 0.008538829162716866, + 0.005792784038931131, + 0.007070230785757303, + 0.004107465501874685, + 0.0029281715396791697, + 0.0036551612429320812, + 0.004490850493311882, + 0.006785963661968708, + 0.00362334749661386, + 0.004430051892995834, + 0.0037604891695082188, + 0.009499027393758297, + 0.011371170170605183, + 0.013008063659071922, + 0.008522887714207172, + 0.014427325688302517, + 0.01132462453097105, + 0.009117885492742062, + 0.015988819301128387, + 0.00933827180415392, + 0.03530922159552574, + 0.015936974436044693, + 0.012891756370663643, + 0.03958071395754814, + 0.01346774771809578, + 0.011846994049847126, + 0.019727109000086784, + 0.020996637642383575, + 0.019937613978981972, + 0.05322897806763649, + 0.020474372431635857, + 0.012044449336826801, + 0.023829571902751923, + 0.01368004735559225, + 0.04819841682910919, + 0.005601359531283379, + 0.004623496904969215, + 0.00561346672475338, + 0.0033600821625441313, + 0.002993364352732897, + 0.005797309335321188, + 0.0030538130085915327, + 0.005623546428978443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.0028203255496919155 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tea spirit composed of emerald green pixels with two large black pixel eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tea spirit leap from the top of the screen in a graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera rotating around the tea spirit while it tumbles through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tea spirit land in a massive off-white pixelated ceramic teacup filled with amber tea and white/light gray steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny cube ripples radiate outward from the tea surface when the spirit touches it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980329655110836 + ], + "opposite_logprob_tokens": [ + -3.5298032760620117 + ], + "logprob_reward": -0.02980329655110836, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980329655110836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5133693814277649, + 0.06664497405290604, + 0.03665740787982941, + 0.03845831751823425, + 0.034077588468790054, + 0.03236807882785797, + 0.12583404779434204, + 0.09727583080530167, + 0.027225084602832794, + 0.028089167550206184 + ], + "credit_map": [ + 0.0383443608880043, + 0.027213802561163902, + 0.030415980145335197, + 0.08012531697750092, + 0.03161351755261421, + 0.08897783607244492, + 0.17470252513885498, + 0.04197611287236214, + 0.00789773091673851, + 0.008307796902954578, + 0.008792220614850521, + 0.010930784977972507, + 0.006543922238051891, + 0.004783011041581631, + 0.010734389536082745, + 0.008655126206576824, + 0.0031362660229206085, + 0.0031054194550961256, + 0.002371900249272585, + 0.006405854597687721, + 0.0046026622876524925, + 0.004003254231065512, + 0.007924525067210197, + 0.00510752759873867, + 0.002716274932026863, + 0.003586036153137684, + 0.003346377518028021, + 0.00581758888438344, + 0.008001704700291157, + 0.004694515373557806, + 0.0061749862506985664, + 0.004120841156691313, + 0.002941110171377659, + 0.004660081584006548, + 0.0047652944922447205, + 0.004165101330727339, + 0.0052899629808962345, + 0.004700968042016029, + 0.0037535850424319506, + 0.003801488783210516, + 0.0025692381896078587, + 0.002038366161286831, + 0.006198281422257423, + 0.0030387502629309893, + 0.005449902731925249, + 0.0030516174156218767, + 0.0034715402871370316, + 0.0065503837540745735, + 0.012815126217901707, + 0.014150083996355534, + 0.019922824576497078, + 0.012782823294401169, + 0.021145053207874298, + 0.011541626416146755, + 0.010838931426405907, + 0.022637588903307915, + 0.00677857268601656, + 0.018854819238185883, + 0.014943067915737629, + 0.006433134898543358, + 0.02066657319664955, + 0.009004073217511177, + 0.006046051159501076, + 0.014549545012414455, + 0.0021503239404410124, + 0.004193121567368507, + 0.0046342588029801846, + 0.0025440678000450134, + 0.003128413576632738, + 0.002968364395201206, + 0.0024782069958746433, + 0.005128331482410431, + 0.0027222619391977787, + 0.004513255786150694, + 0.0035269437357783318, + 0.006216639187186956, + 0.0024649137631058693, + 0.002814836800098419, + 0.0022759989369660616, + 0.003554319031536579 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.02980329655110836 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, rotating tracking shot, set in a Stardew Valley-esque pixel world, captures a \"Yangtianxuelu\" tea spirit, composed of emerald green pixels. Composed of varying shades of green, bordered by darker dark green pixels and dotted with two large, gleaming black pixel eyes, it gracefully leaps from the top of the screen in a graceful arc. The camera rotates gracefully around it, following its slow tumble through the air, ultimately gracing a massive, pixelated ceramic teacup resting atop a warm, wood-grained pixel tabletop. The off-white cup, massive and shimmering with a soft, pixelated sheen, is filled with amber tea, wisps of steam rising from it composed of white and light gray pixels. The tea spirit descends gracefully, its tip delicately touching the surface of the steaming tea, instantly sending ripples of tiny cubes radiating outward. The entire scene is bathed in soft, warm indoor light, creating an atmosphere both magical and deeply warm and tranquil.", + "seed": 682, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_11/index_4046_seed_682_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04027398448670283, + "vlm_logprob": -0.04027398448670283, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.04027398448670283, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tea spirit composed of emerald green pixels with two large black pixel eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tea spirit leap from the top of the screen in a graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera rotating around the tea spirit while it tumbles through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tea spirit land in a massive off-white pixelated ceramic teacup filled with amber tea and white/light gray steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny cube ripples radiate outward from the tea surface when the spirit touches it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011843248270452023 + ], + "opposite_logprob_tokens": [ + -4.761843204498291 + ], + "logprob_reward": -0.011843248270452023, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.011843248270452023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34323662519454956, + 0.11128640174865723, + 0.0922703817486763, + 0.04726635292172432, + 0.08052139729261398, + 0.13580287992954254, + 0.038099393248558044, + 0.03357852250337601, + 0.06973310559988022, + 0.04820491746068001 + ], + "credit_map": [ + 0.07141022384166718, + 0.02673456445336342, + 0.07339455187320709, + 0.03414012864232063, + 0.03602292016148567, + 0.025455111637711525, + 0.034731995314359665, + 0.041347142308950424, + 0.01963304728269577, + 0.011817374266684055, + 0.014871552586555481, + 0.007842855527997017, + 0.02027156949043274, + 0.014091094955801964, + 0.006867311429232359, + 0.015891587361693382, + 0.011975696310400963, + 0.008849511854350567, + 0.010627619922161102, + 0.00892258808016777, + 0.014694460667669773, + 0.01338986586779356, + 0.008776519447565079, + 0.015034112147986889, + 0.006794337183237076, + 0.005585463251918554, + 0.006580648478120565, + 0.005836102645844221, + 0.005612022243440151, + 0.005780121311545372, + 0.0073744384571909904, + 0.0037032198160886765, + 0.011340618133544922, + 0.00697522796690464, + 0.006962754298001528, + 0.005425708368420601, + 0.019487518817186356, + 0.019258709624409676, + 0.005325960461050272, + 0.005744899623095989, + 0.022003615275025368, + 0.030375199392437935, + 0.012541152536869049, + 0.016856810078024864, + 0.021820558235049248, + 0.012911196798086166, + 0.011563196778297424, + 0.00773114338517189, + 0.005385239142924547, + 0.0033963562455028296, + 0.006205595564097166, + 0.0028980267234146595, + 0.005783936008810997, + 0.005083558615297079, + 0.0032314255367964506, + 0.0061152540147304535, + 0.0028808622155338526, + 0.004051369149237871, + 0.004471681080758572, + 0.0034826945047825575, + 0.005138017702847719, + 0.004144546575844288, + 0.005908253137022257, + 0.003501100931316614, + 0.006830394268035889, + 0.006147406995296478, + 0.018443020060658455, + 0.004252713639289141, + 0.00807657465338707, + 0.007527289912104607, + 0.0060729533433914185, + 0.012382753193378448, + 0.004697948694229126, + 0.004421066492795944, + 0.005919511895626783, + 0.007577516604214907, + 0.005681007634848356, + 0.005178277380764484, + 0.007988334633409977, + 0.006741256918758154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.011843248270452023 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tea spirit composed of emerald green pixels with two large black pixel eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tea spirit leap from the top of the screen in a graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera rotating around the tea spirit while it tumbles through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tea spirit land in a massive off-white pixelated ceramic teacup filled with amber tea and white/light gray steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny cube ripples radiate outward from the tea surface when the spirit touches it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048669882118701935 + ], + "opposite_logprob_tokens": [ + -3.0486698150634766 + ], + "logprob_reward": -0.048669882118701935, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048669882118701935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1529436707496643, + 0.09221038967370987, + 0.0525614470243454, + 0.06001219525933266, + 0.07711940258741379, + 0.35678645968437195, + 0.05268583446741104, + 0.048709504306316376, + 0.058383289724588394, + 0.04858775436878204 + ], + "credit_map": [ + 0.02375686541199684, + 0.020632481202483177, + 0.023544812574982643, + 0.02660721354186535, + 0.015367207117378712, + 0.011603167280554771, + 0.019377857446670532, + 0.012054074555635452, + 0.027660591527819633, + 0.009566076099872589, + 0.00659489631652832, + 0.00756427738815546, + 0.006357226986438036, + 0.0067350054159760475, + 0.0052916742861270905, + 0.02244063839316368, + 0.01031995378434658, + 0.006793328560888767, + 0.005804360378533602, + 0.004242127761244774, + 0.0070901475846767426, + 0.0073767900466918945, + 0.006255218759179115, + 0.00467951875180006, + 0.011577572673559189, + 0.009783138521015644, + 0.005980204790830612, + 0.004549236502498388, + 0.009440180845558643, + 0.008272446691989899, + 0.005624939221888781, + 0.004784476477652788, + 0.007781616412103176, + 0.011732245795428753, + 0.009430327452719212, + 0.006189366336911917, + 0.013448390178382397, + 0.01744760572910309, + 0.006251615937799215, + 0.004838236141949892, + 0.04841994866728783, + 0.0828780084848404, + 0.02937743440270424, + 0.03830147162079811, + 0.06750976294279099, + 0.039029549807310104, + 0.03076919913291931, + 0.020501071587204933, + 0.004574161022901535, + 0.006167348939925432, + 0.00755687290802598, + 0.005201362539082766, + 0.007763022556900978, + 0.006505965255200863, + 0.008574639447033405, + 0.006342459470033646, + 0.005269972607493401, + 0.005560336168855429, + 0.005849575623869896, + 0.004091387614607811, + 0.005979728884994984, + 0.007840852253139019, + 0.007149938028305769, + 0.006967713125050068, + 0.005648934748023748, + 0.00536585645750165, + 0.013506442308425903, + 0.004037196282297373, + 0.007474208250641823, + 0.007384161464869976, + 0.0042815678752958775, + 0.010684923268854618, + 0.004161196295171976, + 0.005526482127606869, + 0.00644240528345108, + 0.0062892623245716095, + 0.00825011357665062, + 0.005112206097692251, + 0.006515202112495899, + 0.006290885619819164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.048669882118701935 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tea spirit composed of emerald green pixels with two large black pixel eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tea spirit leap from the top of the screen in a graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera rotating around the tea spirit while it tumbles through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tea spirit land in a massive off-white pixelated ceramic teacup filled with amber tea and white/light gray steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny cube ripples radiate outward from the tea surface when the spirit touches it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002559304237366 + ], + "opposite_logprob_tokens": [ + -2.695025682449341 + ], + "logprob_reward": -0.07002559304237366, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002559304237366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3256716728210449, + 0.148564875125885, + 0.04559771344065666, + 0.025318413972854614, + 0.04183650761842728, + 0.12332319468259811, + 0.05208641290664673, + 0.03323749825358391, + 0.1378438025712967, + 0.06652002036571503 + ], + "credit_map": [ + 0.052508141845464706, + 0.02982950210571289, + 0.04761068895459175, + 0.06081276014447212, + 0.03502408415079117, + 0.02535298839211464, + 0.04302114620804787, + 0.03151232376694679, + 0.04287990182638168, + 0.015890218317508698, + 0.019219398498535156, + 0.00854021217674017, + 0.028956234455108643, + 0.014022930525243282, + 0.008482986129820347, + 0.010572974570095539, + 0.0075970799662172794, + 0.007976561784744263, + 0.003803168423473835, + 0.003221606370061636, + 0.008763205260038376, + 0.0076820156536996365, + 0.003515382995828986, + 0.003038686001673341, + 0.0034189678262919188, + 0.003474007360637188, + 0.0032570494804531336, + 0.0035461271181702614, + 0.002269759774208069, + 0.0027536891866475344, + 0.004134142305701971, + 0.002464667893946171, + 0.005655813962221146, + 0.006742210127413273, + 0.004670285154134035, + 0.004170573782175779, + 0.005049380473792553, + 0.006497393827885389, + 0.00514057045802474, + 0.003910275641828775, + 0.014513787813484669, + 0.02285708859562874, + 0.009418620727956295, + 0.010332204401493073, + 0.03038892336189747, + 0.014423222281038761, + 0.012005617842078209, + 0.00938370916992426, + 0.004498147405683994, + 0.003980168607085943, + 0.00962767843157053, + 0.005370152182877064, + 0.005461075343191624, + 0.00885091070085764, + 0.006910671014338732, + 0.007387601770460606, + 0.002899090526625514, + 0.005481702275574207, + 0.005056027788668871, + 0.002648904686793685, + 0.005192078649997711, + 0.004303731955587864, + 0.003355389228090644, + 0.004300571046769619, + 0.014098911546170712, + 0.0093567855656147, + 0.04532681033015251, + 0.007550064008682966, + 0.007372771389782429, + 0.012390718795359135, + 0.007836912758648396, + 0.033910807222127914, + 0.01030693855136633, + 0.007838382385671139, + 0.00799297634512186, + 0.010485269129276276, + 0.006656194105744362, + 0.007385619916021824, + 0.008704887703061104, + 0.0071497526951134205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.07002559304237366 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tea spirit composed of emerald green pixels with two large black pixel eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tea spirit leap from the top of the screen in a graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera rotating around the tea spirit while it tumbles through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tea spirit land in a massive off-white pixelated ceramic teacup filled with amber tea and white/light gray steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny cube ripples radiate outward from the tea surface when the spirit touches it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008132726070471108 + ], + "opposite_logprob_tokens": [ + -7.1258134841918945 + ], + "logprob_reward": -0.0008132726070471108, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008132726070471108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16070257127285004, + 0.15646208822727203, + 0.0990677922964096, + 0.05255613848567009, + 0.06800174713134766, + 0.07110374420881271, + 0.08770262449979782, + 0.09439827501773834, + 0.14215689897537231, + 0.06784818321466446 + ], + "credit_map": [ + 0.026994844898581505, + 0.016777850687503815, + 0.029692385345697403, + 0.014861592091619968, + 0.026106085628271103, + 0.0143212815746665, + 0.015757642686367035, + 0.016190888360142708, + 0.05190390720963478, + 0.016253894194960594, + 0.012952656485140324, + 0.008386548608541489, + 0.018065445125102997, + 0.013768211007118225, + 0.007098372094333172, + 0.028033047914505005, + 0.012660392560064793, + 0.01749699003994465, + 0.009397992864251137, + 0.01187218353152275, + 0.015322498045861721, + 0.012272399850189686, + 0.008957242593169212, + 0.011088082566857338, + 0.007463247049599886, + 0.007531050592660904, + 0.00640846649184823, + 0.006815237458795309, + 0.004996064119040966, + 0.006178570911288261, + 0.0066747176460921764, + 0.006488784682005644, + 0.010933446697890759, + 0.008510752581059933, + 0.006162962876260281, + 0.005479181185364723, + 0.013999829068779945, + 0.009833336807787418, + 0.00800949614495039, + 0.00507274130359292, + 0.010514344088733196, + 0.009674410335719585, + 0.007786714006215334, + 0.006320077460259199, + 0.011240142397582531, + 0.011379447765648365, + 0.007232992444187403, + 0.0069556157104671, + 0.00892801396548748, + 0.008638757281005383, + 0.010905769653618336, + 0.009203142486512661, + 0.011388980783522129, + 0.014461041428148746, + 0.014386851340532303, + 0.009790066629648209, + 0.012200113385915756, + 0.011300227604806423, + 0.009425009600818157, + 0.008845791220664978, + 0.012620501220226288, + 0.011622299440205097, + 0.01668793521821499, + 0.011696389876306057, + 0.011652231216430664, + 0.0084098931401968, + 0.04409383237361908, + 0.007277338299900293, + 0.009571068920195103, + 0.011745311319828033, + 0.007627298589795828, + 0.04177991300821304, + 0.008128155954182148, + 0.006740497425198555, + 0.009847569279372692, + 0.006532918196171522, + 0.009183399379253387, + 0.00939879659563303, + 0.007555473130196333, + 0.010461374185979366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.0008132726070471108 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tea spirit composed of emerald green pixels with two large black pixel eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tea spirit leap from the top of the screen in a graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera rotating around the tea spirit while it tumbles through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tea spirit land in a massive off-white pixelated ceramic teacup filled with amber tea and white/light gray steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny cube ripples radiate outward from the tea surface when the spirit touches it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001792639493942 + ], + "opposite_logprob_tokens": [ + -2.6950178146362305 + ], + "logprob_reward": -0.07001792639493942, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001792639493942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20589758455753326, + 0.14462162554264069, + 0.04828687012195587, + 0.04491420462727547, + 0.04732850193977356, + 0.12536253035068512, + 0.05895266309380531, + 0.08136659860610962, + 0.17903906106948853, + 0.06423034518957138 + ], + "credit_map": [ + 0.03094811923801899, + 0.016484739258885384, + 0.027348103001713753, + 0.02444741688668728, + 0.031289104372262955, + 0.02200380153954029, + 0.023314159363508224, + 0.030062133446335793, + 0.026649463921785355, + 0.016304805874824524, + 0.01491867657750845, + 0.007851244881749153, + 0.02876361273229122, + 0.022583642974495888, + 0.006908176466822624, + 0.020642001181840897, + 0.00890129990875721, + 0.004937923979014158, + 0.005043467972427607, + 0.005634902976453304, + 0.0053449347615242004, + 0.008629144169390202, + 0.0047996207140386105, + 0.00499557564035058, + 0.008750173263251781, + 0.004797447007149458, + 0.004151832312345505, + 0.005100405775010586, + 0.00579817732796073, + 0.005168565548956394, + 0.006584199145436287, + 0.004563408438116312, + 0.005711189471185207, + 0.0067294384352862835, + 0.0036227479577064514, + 0.0031194111797958612, + 0.009336581453680992, + 0.01120669860392809, + 0.004466806538403034, + 0.0031356278341263533, + 0.015220049768686295, + 0.017480894923210144, + 0.014882650226354599, + 0.01502624899148941, + 0.021792998537421227, + 0.01444566622376442, + 0.01298688817769289, + 0.01352713443338871, + 0.0063009788282215595, + 0.007607594598084688, + 0.007863381877541542, + 0.0049814791418612, + 0.00877680815756321, + 0.010454727336764336, + 0.007095989305526018, + 0.0058717019855976105, + 0.010136831551790237, + 0.009881793521344662, + 0.010628728196024895, + 0.006541659124195576, + 0.011795482598245144, + 0.009604597464203835, + 0.011638453230261803, + 0.011139054782688618, + 0.016835713759064674, + 0.009868763387203217, + 0.057218872010707855, + 0.010646095499396324, + 0.011499764397740364, + 0.018267378211021423, + 0.01147421170026064, + 0.04322824999690056, + 0.006110733840614557, + 0.005998993758112192, + 0.00817048642784357, + 0.008789719082415104, + 0.009272165596485138, + 0.007506173104047775, + 0.00783291831612587, + 0.010549156926572323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.07001792639493942 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, rotating tracking shot, set in a Stardew Valley-esque pixel world, captures a \"Yangtianxuelu\" tea spirit, composed of emerald green pixels. Composed of varying shades of green, bordered by darker dark green pixels and dotted with two large, gleaming black pixel eyes, it gracefully leaps from the top of the screen in a graceful arc. The camera rotates gracefully around it, following its slow tumble through the air, ultimately gracing a massive, pixelated ceramic teacup resting atop a warm, wood-grained pixel tabletop. The off-white cup, massive and shimmering with a soft, pixelated sheen, is filled with amber tea, wisps of steam rising from it composed of white and light gray pixels. The tea spirit descends gracefully, its tip delicately touching the surface of the steaming tea, instantly sending ripples of tiny cubes radiating outward. The entire scene is bathed in soft, warm indoor light, creating an atmosphere both magical and deeply warm and tranquil.", + "seed": 683, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_11/index_4046_seed_683_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05570533315185457, + "vlm_logprob": -0.05570533315185457, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.05570533315185457, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tea spirit composed of emerald green pixels with two large black pixel eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tea spirit leap from the top of the screen in a graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera rotating around the tea spirit while it tumbles through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tea spirit land in a massive off-white pixelated ceramic teacup filled with amber tea and white/light gray steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny cube ripples radiate outward from the tea surface when the spirit touches it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00846394058316946 + ], + "opposite_logprob_tokens": [ + -5.0084638595581055 + ], + "logprob_reward": -0.00846394058316946, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00846394058316946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12663589417934418, + 0.2446192502975464, + 0.23815584182739258, + 0.10293379426002502, + 0.02975207194685936, + 0.030153606086969376, + 0.031615372747182846, + 0.02742449752986431, + 0.017315980046987534, + 0.15139375627040863 + ], + "credit_map": [ + 0.012507102452218533, + 0.012873155064880848, + 0.016324926167726517, + 0.016019625589251518, + 0.01908874697983265, + 0.020355205982923508, + 0.010889081284403801, + 0.018578045070171356, + 0.022341342642903328, + 0.02891744300723076, + 0.019117604941129684, + 0.04290249943733215, + 0.02794704958796501, + 0.039880070835351944, + 0.034979939460754395, + 0.02853326126933098, + 0.06454277038574219, + 0.04807563126087189, + 0.018793383613228798, + 0.01254065427929163, + 0.030488120391964912, + 0.025132566690444946, + 0.014400019310414791, + 0.02418268471956253, + 0.013771414756774902, + 0.010388215072453022, + 0.012516781687736511, + 0.00674401456490159, + 0.02246006950736046, + 0.017682312056422234, + 0.0059981499798595905, + 0.013372819870710373, + 0.003037132788449526, + 0.004622488748282194, + 0.0030433228239417076, + 0.0028836342971771955, + 0.0029853235464543104, + 0.006766119040548801, + 0.0018222355283796787, + 0.0045918128453195095, + 0.0057459441013634205, + 0.0030359867960214615, + 0.0038633672520518303, + 0.003707375843077898, + 0.003069024533033371, + 0.004476562608033419, + 0.0023129298351705074, + 0.003942415118217468, + 0.005567857995629311, + 0.0022249671164900064, + 0.0063828942365944386, + 0.0018922348972409964, + 0.003159986110404134, + 0.007271856535226107, + 0.001439990708604455, + 0.0036755793262273073, + 0.004032469820231199, + 0.004568873904645443, + 0.002745215781033039, + 0.00184432917740196, + 0.0035141881089657545, + 0.004952459596097469, + 0.00220831879414618, + 0.0035586394369602203, + 0.0024896354880183935, + 0.0016647729789838195, + 0.002803824609145522, + 0.001164061133749783, + 0.0022874323185533285, + 0.002356430049985647, + 0.0016680972184985876, + 0.0028817239217460155, + 0.009167182259261608, + 0.0030473065562546253, + 0.004144023172557354, + 0.0026403756346553564, + 0.002822592156007886, + 0.06384578347206116, + 0.003601050702854991, + 0.062125422060489655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 9 + ], + "selected_reward": -0.00846394058316946 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tea spirit composed of emerald green pixels with two large black pixel eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tea spirit leap from the top of the screen in a graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera rotating around the tea spirit while it tumbles through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tea spirit land in a massive off-white pixelated ceramic teacup filled with amber tea and white/light gray steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny cube ripples radiate outward from the tea surface when the spirit touches it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371354565024376 + ], + "opposite_logprob_tokens": [ + -3.4087135791778564 + ], + "logprob_reward": -0.03371354565024376, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371354565024376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13193835318088531, + 0.13310761749744415, + 0.27345600724220276, + 0.05724221467971802, + 0.05632529407739639, + 0.1283983290195465, + 0.05558864772319794, + 0.026562489569187164, + 0.02546936646103859, + 0.11191174387931824 + ], + "credit_map": [ + 0.014108375646173954, + 0.014903400093317032, + 0.008470260538160801, + 0.009561371989548206, + 0.024098338559269905, + 0.019577503204345703, + 0.013730902224779129, + 0.027488185092806816, + 0.01837977208197117, + 0.018806755542755127, + 0.012778472155332565, + 0.016500674188137054, + 0.014432819560170174, + 0.024400772526860237, + 0.01620837301015854, + 0.01159996259957552, + 0.07618669420480728, + 0.03111257404088974, + 0.06498038023710251, + 0.021974297240376472, + 0.02485976554453373, + 0.019518837332725525, + 0.015249805524945259, + 0.019573621451854706, + 0.007176416926085949, + 0.007104677613824606, + 0.007951819337904453, + 0.005711916834115982, + 0.00906428974121809, + 0.010091838426887989, + 0.004377681761980057, + 0.005763568449765444, + 0.005210352595895529, + 0.01086513139307499, + 0.005593820475041866, + 0.004580640699714422, + 0.0049508679658174515, + 0.0129935247823596, + 0.00417008250951767, + 0.007960868999361992, + 0.01878778636455536, + 0.015284661203622818, + 0.014868839643895626, + 0.01116463914513588, + 0.020349372178316116, + 0.021237755194306374, + 0.0091691380366683, + 0.017536120489239693, + 0.01004757173359394, + 0.0038440153002738953, + 0.012238920666277409, + 0.0033068787306547165, + 0.003993350081145763, + 0.01265690103173256, + 0.00317557854577899, + 0.006325429305434227, + 0.004435393027961254, + 0.0027474709786474705, + 0.0042380052618682384, + 0.0017533143982291222, + 0.0016071193385869265, + 0.005353159736841917, + 0.0019571399316191673, + 0.004470886196941137, + 0.0034892838448286057, + 0.0016291762003675103, + 0.004406393971294165, + 0.0013217584928497672, + 0.0023721489124000072, + 0.00392460310831666, + 0.0018383330898359418, + 0.006487667094916105, + 0.00703484145924449, + 0.0036739911884069443, + 0.0028263351414352655, + 0.002408571308478713, + 0.003211502218618989, + 0.044617507606744766, + 0.003181933658197522, + 0.04495704546570778 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.03371354565024376 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tea spirit composed of emerald green pixels with two large black pixel eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tea spirit leap from the top of the screen in a graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera rotating around the tea spirit while it tumbles through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tea spirit land in a massive off-white pixelated ceramic teacup filled with amber tea and white/light gray steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny cube ripples radiate outward from the tea surface when the spirit touches it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798863708972931 + ], + "opposite_logprob_tokens": [ + -1.8048863410949707 + ], + "logprob_reward": -0.1798863708972931, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798863708972931, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06331899762153625, + 0.09721330553293228, + 0.43031951785087585, + 0.04758813977241516, + 0.030605437234044075, + 0.14316624402999878, + 0.08476407825946808, + 0.012794368900358677, + 0.03877602145075798, + 0.05145391821861267 + ], + "credit_map": [ + 0.007144173141568899, + 0.005647086538374424, + 0.00715513713657856, + 0.006709787994623184, + 0.008687392808496952, + 0.008086217567324638, + 0.008022507652640343, + 0.01186670083552599, + 0.00872442964464426, + 0.013639450073242188, + 0.007165386341512203, + 0.014979111962020397, + 0.011387504637241364, + 0.013028762303292751, + 0.015731144696474075, + 0.012557515874505043, + 0.11582916975021362, + 0.0367126539349556, + 0.11326508969068527, + 0.03617618605494499, + 0.033719006925821304, + 0.03644965961575508, + 0.02403152734041214, + 0.034136202186346054, + 0.006045530084520578, + 0.00716162147000432, + 0.0054122027941048145, + 0.006216936744749546, + 0.007128051016479731, + 0.006242988631129265, + 0.004029971081763506, + 0.005350836087018251, + 0.002321571810171008, + 0.00717565743252635, + 0.003774025710299611, + 0.004014299716800451, + 0.0030400087125599384, + 0.004392215982079506, + 0.002392764901742339, + 0.0034948934335261583, + 0.021967696025967598, + 0.020297948271036148, + 0.01559726893901825, + 0.011633271351456642, + 0.026256779208779335, + 0.019407348707318306, + 0.009971496649086475, + 0.0180344358086586, + 0.010786878876388073, + 0.00633359607309103, + 0.01900235377252102, + 0.0036021580453962088, + 0.007759586907923222, + 0.023176096379756927, + 0.003941087052226067, + 0.010162314400076866, + 0.002614879049360752, + 0.0011904343264177442, + 0.0014898637309670448, + 0.0006305953138507903, + 0.0011853411560878158, + 0.002919056685641408, + 0.0008831123122945428, + 0.0018810862675309181, + 0.005489469040185213, + 0.0016427435912191868, + 0.004690604284405708, + 0.0010939595522359014, + 0.0018835877999663353, + 0.00707505876198411, + 0.002551998710259795, + 0.014348596334457397, + 0.0032816894818097353, + 0.0024051449727267027, + 0.0027628070674836636, + 0.0013839952880516648, + 0.001513318857178092, + 0.019370486959815025, + 0.0015964702470228076, + 0.019140003249049187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 1 + ], + "selected_reward": -0.1798863708972931 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tea spirit composed of emerald green pixels with two large black pixel eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tea spirit leap from the top of the screen in a graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera rotating around the tea spirit while it tumbles through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tea spirit land in a massive off-white pixelated ceramic teacup filled with amber tea and white/light gray steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny cube ripples radiate outward from the tea surface when the spirit touches it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015125750796869397 + ], + "opposite_logprob_tokens": [ + -6.50151252746582 + ], + "logprob_reward": -0.0015125750796869397, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015125750796869397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16336165368556976, + 0.09803535789251328, + 0.32702702283859253, + 0.046848416328430176, + 0.05223853886127472, + 0.09059032052755356, + 0.02834058552980423, + 0.045070234686136246, + 0.03411813825368881, + 0.11436974257230759 + ], + "credit_map": [ + 0.017401909455657005, + 0.018107324838638306, + 0.022109363228082657, + 0.02080352231860161, + 0.022474108263850212, + 0.02672475576400757, + 0.015450620092451572, + 0.020290059968829155, + 0.011036364361643791, + 0.01314181461930275, + 0.009693389758467674, + 0.014338064007461071, + 0.009863737970590591, + 0.0124705471098423, + 0.016196036711335182, + 0.011295408941805363, + 0.08136674016714096, + 0.028443768620491028, + 0.09303794801235199, + 0.027194630354642868, + 0.02660464122891426, + 0.02710588648915291, + 0.018868062645196915, + 0.02440537139773369, + 0.006527879275381565, + 0.004820412490516901, + 0.007767383474856615, + 0.004825827665627003, + 0.007559524849057198, + 0.0049571567215025425, + 0.0035400239285081625, + 0.006850213278084993, + 0.005245388951152563, + 0.012173369526863098, + 0.0064870077185332775, + 0.006919562350958586, + 0.006223280448466539, + 0.005928029306232929, + 0.0042753336019814014, + 0.004986574873328209, + 0.013615796342492104, + 0.014928845688700676, + 0.010364902205765247, + 0.0061313989572227, + 0.019066940993070602, + 0.011491450481116772, + 0.008000082336366177, + 0.006990910042077303, + 0.004794117994606495, + 0.004178576171398163, + 0.003944569732993841, + 0.001657273038290441, + 0.004173024091869593, + 0.004445319529622793, + 0.0015825775917619467, + 0.0035651277285069227, + 0.008920758962631226, + 0.005077731795608997, + 0.006082703359425068, + 0.0023341048508882523, + 0.0033975117839872837, + 0.009459331631660461, + 0.003365034470334649, + 0.0064330617897212505, + 0.005384810268878937, + 0.001950794248841703, + 0.006320746149867773, + 0.0017739966278895736, + 0.001698954962193966, + 0.005280747078359127, + 0.0022301599383354187, + 0.009477932937443256, + 0.007972034625709057, + 0.004249834921211004, + 0.008382579311728477, + 0.0039583491161465645, + 0.004039622377604246, + 0.04034164920449257, + 0.005596877075731754, + 0.039828818291425705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.0015125750796869397 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tea spirit composed of emerald green pixels with two large black pixel eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tea spirit leap from the top of the screen in a graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera rotating around the tea spirit while it tumbles through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tea spirit land in a massive off-white pixelated ceramic teacup filled with amber tea and white/light gray steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny cube ripples radiate outward from the tea surface when the spirit touches it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495023354887962 + ], + "opposite_logprob_tokens": [ + -2.92995023727417 + ], + "logprob_reward": -0.05495023354887962, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495023354887962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16944198310375214, + 0.10098763555288315, + 0.2361103594303131, + 0.07815875113010406, + 0.046247754245996475, + 0.09514153748750687, + 0.04540715366601944, + 0.08564332127571106, + 0.06582798808813095, + 0.07703351974487305 + ], + "credit_map": [ + 0.01780763268470764, + 0.017914559692144394, + 0.024325266480445862, + 0.02101912535727024, + 0.026154104620218277, + 0.022019505500793457, + 0.015641160309314728, + 0.02456061914563179, + 0.012057331390678883, + 0.012381206266582012, + 0.01018837746232748, + 0.01490831095725298, + 0.011820538900792599, + 0.015556013211607933, + 0.014155326411128044, + 0.00992052536457777, + 0.07339164614677429, + 0.022728372365236282, + 0.04726817458868027, + 0.01683885231614113, + 0.014176120050251484, + 0.020124375820159912, + 0.013122372329235077, + 0.028460459783673286, + 0.010518479160964489, + 0.0119176609441638, + 0.009006458334624767, + 0.009172534570097923, + 0.010497432202100754, + 0.011378697119653225, + 0.005212565418332815, + 0.010454918257892132, + 0.004791182465851307, + 0.008250338956713676, + 0.005079449154436588, + 0.007726583629846573, + 0.0043787769973278046, + 0.006394130177795887, + 0.0036339564248919487, + 0.00599333643913269, + 0.012712319381535053, + 0.01009087823331356, + 0.011653712019324303, + 0.011536981910467148, + 0.017068497836589813, + 0.014385310001671314, + 0.006772096734493971, + 0.010921746492385864, + 0.006533043459057808, + 0.004477089270949364, + 0.007619759067893028, + 0.0036987115163356066, + 0.005198169965296984, + 0.008529683575034142, + 0.0032714048866182566, + 0.006079291459172964, + 0.009648664854466915, + 0.011203286238014698, + 0.010958565399050713, + 0.006929001770913601, + 0.007606357801705599, + 0.01315650437027216, + 0.011140002869069576, + 0.015000938437879086, + 0.01160485576838255, + 0.0033135428093373775, + 0.00584548432379961, + 0.002467074664309621, + 0.003693060949444771, + 0.012091133743524551, + 0.0036613098345696926, + 0.023151526227593422, + 0.006693825125694275, + 0.006533757317811251, + 0.00863735657185316, + 0.003967631608247757, + 0.0038964555133134127, + 0.020917387679219246, + 0.004075855016708374, + 0.02231125719845295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.05495023354887962 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, rotating tracking shot, set in a Stardew Valley-esque pixel world, captures a \"Yangtianxuelu\" tea spirit, composed of emerald green pixels. Composed of varying shades of green, bordered by darker dark green pixels and dotted with two large, gleaming black pixel eyes, it gracefully leaps from the top of the screen in a graceful arc. The camera rotates gracefully around it, following its slow tumble through the air, ultimately gracing a massive, pixelated ceramic teacup resting atop a warm, wood-grained pixel tabletop. The off-white cup, massive and shimmering with a soft, pixelated sheen, is filled with amber tea, wisps of steam rising from it composed of white and light gray pixels. The tea spirit descends gracefully, its tip delicately touching the surface of the steaming tea, instantly sending ripples of tiny cubes radiating outward. The entire scene is bathed in soft, warm indoor light, creating an atmosphere both magical and deeply warm and tranquil.", + "seed": 684, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_11/index_4046_seed_684_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05405100556090474, + "vlm_logprob": -0.05405100556090474, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.05405100556090474, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tea spirit composed of emerald green pixels with two large black pixel eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tea spirit leap from the top of the screen in a graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera rotating around the tea spirit while it tumbles through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tea spirit land in a massive off-white pixelated ceramic teacup filled with amber tea and white/light gray steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny cube ripples radiate outward from the tea surface when the spirit touches it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022312667220830917 + ], + "opposite_logprob_tokens": [ + -3.897312641143799 + ], + "logprob_reward": -0.022312667220830917, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.022312667220830917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19343511760234833, + 0.16439762711524963, + 0.07060492038726807, + 0.06844739615917206, + 0.047661274671554565, + 0.03989233821630478, + 0.1980651468038559, + 0.10778743028640747, + 0.05769519880414009, + 0.05201348289847374 + ], + "credit_map": [ + 0.018999554216861725, + 0.023314373567700386, + 0.023073138669133186, + 0.028769930824637413, + 0.02407268062233925, + 0.022330373525619507, + 0.01725710742175579, + 0.035617973655462265, + 0.03491315245628357, + 0.023896999657154083, + 0.030729178339242935, + 0.015139476396143436, + 0.011006714776158333, + 0.01143037062138319, + 0.015310313552618027, + 0.021971451118588448, + 0.01277334988117218, + 0.006800931412726641, + 0.007166728377342224, + 0.0077734338119626045, + 0.011996949091553688, + 0.005081217736005783, + 0.011009203270077705, + 0.008003114722669125, + 0.007195481564849615, + 0.005748630501329899, + 0.008730686269700527, + 0.013210965320467949, + 0.00414610980078578, + 0.007601861841976643, + 0.011464189738035202, + 0.01034947857260704, + 0.006281209643930197, + 0.004939067177474499, + 0.003991341684013605, + 0.006706765852868557, + 0.006888426840305328, + 0.0043771108612418175, + 0.009489504620432854, + 0.004987851716578007, + 0.003946431912481785, + 0.004060954321175814, + 0.003923216834664345, + 0.00569278746843338, + 0.004194870125502348, + 0.003932437859475613, + 0.010552303865551949, + 0.0035893411841243505, + 0.042923443019390106, + 0.013205486349761486, + 0.03608736768364906, + 0.022948816418647766, + 0.011950582265853882, + 0.01881401799619198, + 0.03611106052994728, + 0.0160243958234787, + 0.013411113061010838, + 0.014704008586704731, + 0.013766476884484291, + 0.00842368509620428, + 0.012825338169932365, + 0.017669053748250008, + 0.016674384474754333, + 0.010313384234905243, + 0.005268377251923084, + 0.0052354722283780575, + 0.006007193587720394, + 0.00505494698882103, + 0.009787572547793388, + 0.0064252265729010105, + 0.014057625085115433, + 0.0058587947860360146, + 0.006446987856179476, + 0.004403427243232727, + 0.00567464018240571, + 0.007659994065761566, + 0.007070259656757116, + 0.005307525862008333, + 0.00830591470003128, + 0.007144736126065254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.022312667220830917 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tea spirit composed of emerald green pixels with two large black pixel eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tea spirit leap from the top of the screen in a graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera rotating around the tea spirit while it tumbles through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tea spirit land in a massive off-white pixelated ceramic teacup filled with amber tea and white/light gray steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny cube ripples radiate outward from the tea surface when the spirit touches it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0486721508204937 + ], + "opposite_logprob_tokens": [ + -3.0486721992492676 + ], + "logprob_reward": -0.0486721508204937, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0486721508204937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2143358439207077, + 0.15281586349010468, + 0.11088702082633972, + 0.04558112099766731, + 0.08755944669246674, + 0.08187627047300339, + 0.07146292924880981, + 0.07373325526714325, + 0.08372925966978073, + 0.07801902294158936 + ], + "credit_map": [ + 0.02106219343841076, + 0.026441581547260284, + 0.017596829682588577, + 0.05256957560777664, + 0.030625540763139725, + 0.023089412599802017, + 0.017421633005142212, + 0.02552907168865204, + 0.031379640102386475, + 0.028564320877194405, + 0.015640700235962868, + 0.015796754509210587, + 0.01326694991439581, + 0.01194432470947504, + 0.02089628390967846, + 0.015326892025768757, + 0.009325106628239155, + 0.013579292222857475, + 0.00913830567151308, + 0.014609375968575478, + 0.02360960841178894, + 0.008361018262803555, + 0.02156483195722103, + 0.010699485428631306, + 0.006056827027350664, + 0.003529514418914914, + 0.004578599706292152, + 0.0071088336408138275, + 0.0035403261426836252, + 0.0050621237605810165, + 0.011825181543827057, + 0.0038797163870185614, + 0.006657194811850786, + 0.014424562454223633, + 0.007971550337970257, + 0.009892492555081844, + 0.016540659591555595, + 0.008179014548659325, + 0.01684226095676422, + 0.007051713299006224, + 0.009997598826885223, + 0.01057775691151619, + 0.008847839199006557, + 0.010062805376946926, + 0.012692291289567947, + 0.009224825538694859, + 0.013936336152255535, + 0.006536813452839851, + 0.010326397605240345, + 0.0064930785447359085, + 0.008405216038227081, + 0.0066175684332847595, + 0.007152770180255175, + 0.012808061204850674, + 0.011208578944206238, + 0.008451255038380623, + 0.008650841191411018, + 0.005988831631839275, + 0.0142081743106246, + 0.006740138866007328, + 0.007685445714741945, + 0.013550960458815098, + 0.009406563825905323, + 0.007502302993088961, + 0.006896780338138342, + 0.011602913029491901, + 0.0058731879107654095, + 0.013066370971500874, + 0.014628015458583832, + 0.009713038802146912, + 0.014169370755553246, + 0.007779583800584078, + 0.008264393545687199, + 0.008286112919449806, + 0.007630712818354368, + 0.013146249577403069, + 0.011230682954192162, + 0.006035769358277321, + 0.01314797718077898, + 0.010277117602527142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0486721508204937 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tea spirit composed of emerald green pixels with two large black pixel eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tea spirit leap from the top of the screen in a graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera rotating around the tea spirit while it tumbles through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tea spirit land in a massive off-white pixelated ceramic teacup filled with amber tea and white/light gray steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny cube ripples radiate outward from the tea surface when the spirit touches it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704816460609436 + ], + "opposite_logprob_tokens": [ + -2.1270482540130615 + ], + "logprob_reward": -0.12704816460609436, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12704816460609436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10684839636087418, + 0.20919571816921234, + 0.05436168238520622, + 0.05837613716721535, + 0.06810351461172104, + 0.053680773824453354, + 0.17940890789031982, + 0.1341870278120041, + 0.09249559044837952, + 0.043342214077711105 + ], + "credit_map": [ + 0.009795584715902805, + 0.00994498748332262, + 0.011721063405275345, + 0.01635286770761013, + 0.01716293767094612, + 0.019336143508553505, + 0.00955653004348278, + 0.01297828834503889, + 0.04357928782701492, + 0.03717206045985222, + 0.043548911809921265, + 0.017238954082131386, + 0.017188267782330513, + 0.011228892020881176, + 0.02288014069199562, + 0.016359198838472366, + 0.008450889028608799, + 0.005304260179400444, + 0.005222199950367212, + 0.009419851005077362, + 0.007102214731276035, + 0.0038964208215475082, + 0.01047650445252657, + 0.00448934081941843, + 0.00613769656047225, + 0.003573056310415268, + 0.00558779900893569, + 0.013600298203527927, + 0.00428033759817481, + 0.0048105851747095585, + 0.01491722371429205, + 0.005469139199703932, + 0.00394139951094985, + 0.006514651235193014, + 0.006747598294168711, + 0.007121389731764793, + 0.015042017214000225, + 0.009088631719350815, + 0.012209642678499222, + 0.007438190747052431, + 0.0053022271022200584, + 0.005914881359785795, + 0.006295608822256327, + 0.00764478137716651, + 0.005702517461031675, + 0.004577069077640772, + 0.012410838156938553, + 0.005832853261381388, + 0.019356846809387207, + 0.01596914418041706, + 0.013782159425318241, + 0.014189698733389378, + 0.02814723365008831, + 0.04468384385108948, + 0.029550809413194656, + 0.013729175552725792, + 0.018765900284051895, + 0.016490114852786064, + 0.025185788050293922, + 0.010141902603209019, + 0.017756879329681396, + 0.017620472237467766, + 0.01743250899016857, + 0.010793466120958328, + 0.007549378555268049, + 0.00851949118077755, + 0.005575506016612053, + 0.012013443745672703, + 0.018916040658950806, + 0.005754518788307905, + 0.026074934750795364, + 0.008092268370091915, + 0.003485350636765361, + 0.003781316801905632, + 0.004091454669833183, + 0.012509740889072418, + 0.004830294754356146, + 0.0034533250145614147, + 0.006485561840236187, + 0.004705169703811407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 7 + ], + "selected_reward": -0.12704816460609436 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tea spirit composed of emerald green pixels with two large black pixel eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tea spirit leap from the top of the screen in a graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera rotating around the tea spirit while it tumbles through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tea spirit land in a massive off-white pixelated ceramic teacup filled with amber tea and white/light gray steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny cube ripples radiate outward from the tea surface when the spirit touches it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021971152164041996 + ], + "opposite_logprob_tokens": [ + -6.127197265625 + ], + "logprob_reward": -0.0021971152164041996, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021971152164041996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12282214313745499, + 0.1438896656036377, + 0.06408441811800003, + 0.10735853016376495, + 0.10754035413265228, + 0.10998642444610596, + 0.06531823426485062, + 0.10381066054105759, + 0.08364313095808029, + 0.0915464237332344 + ], + "credit_map": [ + 0.010936494916677475, + 0.010808076709508896, + 0.01430812943726778, + 0.019301602616906166, + 0.01513693667948246, + 0.014815596863627434, + 0.012231392785906792, + 0.02528391219675541, + 0.028081320226192474, + 0.017488274723291397, + 0.028520064428448677, + 0.011291014030575752, + 0.012214242480695248, + 0.01191416010260582, + 0.01521383598446846, + 0.01916675642132759, + 0.006335723679512739, + 0.0040000141598284245, + 0.007455367594957352, + 0.014112901873886585, + 0.00736161507666111, + 0.0033357113134115934, + 0.013689404353499413, + 0.007793678902089596, + 0.012053701095283031, + 0.0066300430335104465, + 0.010680838488042355, + 0.015980621799826622, + 0.007009513210505247, + 0.009552761912345886, + 0.03296707570552826, + 0.012483987025916576, + 0.007543903775513172, + 0.007071434985846281, + 0.010454811155796051, + 0.011467954143881798, + 0.012419931590557098, + 0.010042747482657433, + 0.038254719227552414, + 0.010284855030477047, + 0.007491590455174446, + 0.015833042562007904, + 0.010830209590494633, + 0.012892643921077251, + 0.0240471251308918, + 0.013337231241166592, + 0.013998749665915966, + 0.011555842123925686, + 0.008080169558525085, + 0.0063786618411540985, + 0.008554022759199142, + 0.007573877461254597, + 0.009011080488562584, + 0.008383884094655514, + 0.009622765704989433, + 0.007713770493865013, + 0.009584710001945496, + 0.014390697702765465, + 0.012102962471544743, + 0.00896413717418909, + 0.01609090529382229, + 0.016272738575935364, + 0.019490035250782967, + 0.006914468947798014, + 0.004413566552102566, + 0.005660408176481724, + 0.004625286906957626, + 0.006767881102859974, + 0.022128857672214508, + 0.006405415944755077, + 0.027952630072832108, + 0.005689084995537996, + 0.007575992960482836, + 0.011963185854256153, + 0.0072007691487669945, + 0.018524546176195145, + 0.018187550827860832, + 0.007422262337058783, + 0.013937613926827908, + 0.0067344969138503075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.0021971152164041996 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tea spirit composed of emerald green pixels with two large black pixel eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tea spirit leap from the top of the screen in a graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera rotating around the tea spirit while it tumbles through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tea spirit land in a massive off-white pixelated ceramic teacup filled with amber tea and white/light gray steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny cube ripples radiate outward from the tea surface when the spirit touches it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002492994070053 + ], + "opposite_logprob_tokens": [ + -2.6950249671936035 + ], + "logprob_reward": -0.07002492994070053, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002492994070053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12211007624864578, + 0.1637321412563324, + 0.055723633617162704, + 0.06311995536088943, + 0.08349461853504181, + 0.04777903854846954, + 0.09063751250505447, + 0.16176550090312958, + 0.05632904917001724, + 0.1553085446357727 + ], + "credit_map": [ + 0.011279459111392498, + 0.012278462760150433, + 0.01571817323565483, + 0.021342234686017036, + 0.01742277666926384, + 0.013660234399139881, + 0.010005882009863853, + 0.020402856171131134, + 0.04353202134370804, + 0.030033031478524208, + 0.02408747375011444, + 0.010864989832043648, + 0.01276860199868679, + 0.010434266179800034, + 0.016804613173007965, + 0.015207134187221527, + 0.006363484542816877, + 0.005675457417964935, + 0.005859760567545891, + 0.008803583681583405, + 0.008289121091365814, + 0.00514728669077158, + 0.01050576288253069, + 0.005079175345599651, + 0.005254407878965139, + 0.005486726760864258, + 0.0068964408710598946, + 0.01436618622392416, + 0.00467697624117136, + 0.00832650437951088, + 0.01140139251947403, + 0.006711322348564863, + 0.008378492668271065, + 0.011711121536791325, + 0.007074405904859304, + 0.009880617260932922, + 0.012194670736789703, + 0.009916470386087894, + 0.014789840206503868, + 0.009549000300467014, + 0.005813430529087782, + 0.006332752760499716, + 0.0049603963270783424, + 0.005763720255345106, + 0.007685198448598385, + 0.004292414989322424, + 0.008199701085686684, + 0.00473141809925437, + 0.011482044123113155, + 0.00691034272313118, + 0.017163027077913284, + 0.008996902033686638, + 0.011724529787898064, + 0.01606576330959797, + 0.010738893412053585, + 0.007556009106338024, + 0.020074917003512383, + 0.016996966674923897, + 0.02606121636927128, + 0.014568229205906391, + 0.023254375904798508, + 0.02724911831319332, + 0.017906105145812035, + 0.015654565766453743, + 0.004523485898971558, + 0.004615806508809328, + 0.004447294864803553, + 0.013503816910088062, + 0.008024298585951328, + 0.005256104748696089, + 0.009287072345614433, + 0.006671170238405466, + 0.01456299889832735, + 0.017158815637230873, + 0.013050718232989311, + 0.03460923582315445, + 0.027351567521691322, + 0.013802817091345787, + 0.02232762612402439, + 0.012444762513041496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 9 + ], + "selected_reward": -0.07002492994070053 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_12/index_tensor([9705, 9705])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_12/index_tensor([9705, 9705])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b29e9650272a76eeb5692aa45e01ef254d614e8a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_12/index_tensor([9705, 9705])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, low-angle, wide-angle shot captures the cataclysmic moment as a towering bronze statue of an ancient king, situated in the center of a vast, stone-paved plaza, begins its demise. The king, clad in dark, weathered bronze streaked with green patina, stands in an imperial pose, one hand resting on a greatsword, his bearded face set in a permanent, stern expression. Beneath a dark, turbulent sky, dramatic, low-angle sunlight penetrates the clouds, casting long, stark shadows that dimly shimmer across the metal surface. The entire scene is captured by a violent, low-frequency vibration that courses through the statue's massive form, causing the camera to tremble slightly from the shockwave. This is immediately followed by high-frequency dusting, removing oxidized metal flakes and small stone fragments from its massive base. With the muffled groan of shearing metal, a visible crack spreads from the statue's base, and the colossal figure begins to slowly and inexorably tilt forward. The statue's immense weight becomes the instrument of its own destruction, its once-proud posture twisting into a grotesque tilt as it begins to collapse, creating an atmosphere of terrifying, apocalyptic grandeur amidst the ruins.", + "seed": 16675, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_12/index_9705_seed_16675_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06654242114163936, + "vlm_logprob": -0.06654242114163936, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.06654242114163936, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering bronze statue of an ancient king situated in the center of a vast, stone-paved plaza?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue clad in dark, weathered bronze streaked with green patina, with one hand resting on a greatsword and a bearded face in a stern expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dark, turbulent sky with low-angle sunlight penetrating the clouds to cast long, stark shadows on the metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue exhibit a visible crack spreading from its base and begin to slowly tilt forward as it collapses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle view that trembles slightly due to a violent, low-frequency vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10179319977760315 + ], + "opposite_logprob_tokens": [ + -2.3517932891845703 + ], + "logprob_reward": -0.10179319977760315, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10179319977760315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15577150881290436, + 0.05152561143040657, + 0.13677388429641724, + 0.12000321596860886, + 0.08278302848339081, + 0.06677258014678955, + 0.10736291855573654, + 0.07594748586416245, + 0.15368466079235077, + 0.04937510937452316 + ], + "credit_map": [ + 0.022925078868865967, + 0.019582077860832214, + 0.020324021577835083, + 0.012003224343061447, + 0.023965487256646156, + 0.02293725498020649, + 0.017007650807499886, + 0.017026718705892563, + 0.007622201927006245, + 0.0040653422474861145, + 0.006866298615932465, + 0.003389943391084671, + 0.007211478427052498, + 0.00799411628395319, + 0.006543382070958614, + 0.007832847535610199, + 0.025979869067668915, + 0.019195962697267532, + 0.011547228321433067, + 0.009161886759102345, + 0.013114847242832184, + 0.020543400198221207, + 0.020289968699216843, + 0.01694071665406227, + 0.023350996896624565, + 0.010964155197143555, + 0.008809156715869904, + 0.0052936929278075695, + 0.014210280030965805, + 0.021319719031453133, + 0.0102507583796978, + 0.02580445632338524, + 0.0204482302069664, + 0.007868604734539986, + 0.005470700562000275, + 0.003993770107626915, + 0.008541460148990154, + 0.011798471212387085, + 0.011078343726694584, + 0.01358344592154026, + 0.010288300924003124, + 0.011992412619292736, + 0.005667939316481352, + 0.006198346149176359, + 0.00891539640724659, + 0.007369528524577618, + 0.010778410360217094, + 0.005562249105423689, + 0.013513194397091866, + 0.012320122681558132, + 0.01391498651355505, + 0.009420224465429783, + 0.01297848578542471, + 0.01582355424761772, + 0.013313697651028633, + 0.016078650951385498, + 0.009997228160500526, + 0.005248145665973425, + 0.0173319224268198, + 0.006370361428707838, + 0.007413615006953478, + 0.00910390168428421, + 0.011189213022589684, + 0.009293094277381897, + 0.016072561964392662, + 0.008625250309705734, + 0.013733470812439919, + 0.0178094282746315, + 0.015158820897340775, + 0.04620895907282829, + 0.025944264605641365, + 0.010131907649338245, + 0.0037440797314047813, + 0.00514534767717123, + 0.00797748751938343, + 0.00833117589354515, + 0.005305595230311155, + 0.007551020011305809, + 0.006393509916961193, + 0.004926894325762987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.10179319977760315 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering bronze statue of an ancient king situated in the center of a vast, stone-paved plaza?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue clad in dark, weathered bronze streaked with green patina, with one hand resting on a greatsword and a bearded face in a stern expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dark, turbulent sky with low-angle sunlight penetrating the clouds to cast long, stark shadows on the metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue exhibit a visible crack spreading from its base and begin to slowly tilt forward as it collapses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle view that trembles slightly due to a violent, low-frequency vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028325694147497416 + ], + "opposite_logprob_tokens": [ + -5.877832412719727 + ], + "logprob_reward": -0.0028325694147497416, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028325694147497416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09687177836894989, + 0.042215920984745026, + 0.0347038209438324, + 0.060237444937229156, + 0.11704445630311966, + 0.1204850897192955, + 0.04140384495258331, + 0.039348967373371124, + 0.3667515814304352, + 0.08093702793121338 + ], + "credit_map": [ + 0.01238791923969984, + 0.0165843665599823, + 0.011068418622016907, + 0.008348386734724045, + 0.01526589971035719, + 0.013565023429691792, + 0.009876304306089878, + 0.009775460697710514, + 0.007410542573779821, + 0.0038339486345648766, + 0.006005721632391214, + 0.0032777038868516684, + 0.004824142903089523, + 0.006091736257076263, + 0.00485310610383749, + 0.0059190187603235245, + 0.00661956612020731, + 0.003972808830440044, + 0.003036285052075982, + 0.002158602699637413, + 0.0046875206753611565, + 0.005382295232266188, + 0.005349616054445505, + 0.0034971225541085005, + 0.013207131996750832, + 0.005181676242500544, + 0.004080496728420258, + 0.0025838096626102924, + 0.007287504151463509, + 0.01033227052539587, + 0.004882874432951212, + 0.012681678868830204, + 0.011934120208024979, + 0.02032610960304737, + 0.006553187966346741, + 0.0061777327209711075, + 0.011728870682418346, + 0.014190193265676498, + 0.025148507207632065, + 0.020985733717679977, + 0.017896348610520363, + 0.018398793414235115, + 0.008571941405534744, + 0.010682234540581703, + 0.017083460465073586, + 0.014198016375303268, + 0.0236325915902853, + 0.010021708905696869, + 0.006696242373436689, + 0.0031957996543496847, + 0.00695491349324584, + 0.003089119214564562, + 0.0038707852363586426, + 0.006231218576431274, + 0.004550461191684008, + 0.006815303582698107, + 0.003146774834021926, + 0.002526306314393878, + 0.007304811384528875, + 0.003486384404823184, + 0.0035304699558764696, + 0.008106501772999763, + 0.007912051863968372, + 0.0033356656786054373, + 0.031177224591374397, + 0.019367048516869545, + 0.020802272483706474, + 0.01790371723473072, + 0.023265810683369637, + 0.13562768697738647, + 0.09430310875177383, + 0.024304715916514397, + 0.007301222067326307, + 0.009820764884352684, + 0.011743275448679924, + 0.01149703562259674, + 0.012727892957627773, + 0.009166243486106396, + 0.012054555118083954, + 0.006626040209084749 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 4 + ], + "selected_reward": -0.0028325694147497416 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering bronze statue of an ancient king situated in the center of a vast, stone-paved plaza?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue clad in dark, weathered bronze streaked with green patina, with one hand resting on a greatsword and a bearded face in a stern expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dark, turbulent sky with low-angle sunlight penetrating the clouds to cast long, stark shadows on the metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue exhibit a visible crack spreading from its base and begin to slowly tilt forward as it collapses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle view that trembles slightly due to a violent, low-frequency vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005949170328676701 + ], + "opposite_logprob_tokens": [ + -5.130949020385742 + ], + "logprob_reward": -0.005949170328676701, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005949170328676701, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04159031808376312, + 0.031453438103199005, + 0.14132417738437653, + 0.11638016253709793, + 0.07235587388277054, + 0.09980815649032593, + 0.043819550424814224, + 0.05038628727197647, + 0.32197022438049316, + 0.08091174811124802 + ], + "credit_map": [ + 0.0039045291487127542, + 0.0030478057451546192, + 0.005893701687455177, + 0.0035593281500041485, + 0.00802739430218935, + 0.006158028729259968, + 0.005888018757104874, + 0.005111516918987036, + 0.005479669198393822, + 0.002832406433299184, + 0.0036741760559380054, + 0.0027567485813051462, + 0.003264708211645484, + 0.004791524726897478, + 0.003965423442423344, + 0.004688781686127186, + 0.025806700810790062, + 0.020290106534957886, + 0.012089183554053307, + 0.008773588575422764, + 0.013854347169399261, + 0.02198490872979164, + 0.01955370232462883, + 0.018971655517816544, + 0.025024132803082466, + 0.00814107432961464, + 0.009135177358984947, + 0.004832696169614792, + 0.013953601941466331, + 0.0204600989818573, + 0.007745612412691116, + 0.027087779715657234, + 0.00788519624620676, + 0.012910207733511925, + 0.004621748346835375, + 0.004951823968440294, + 0.008593408390879631, + 0.008641485124826431, + 0.016236644238233566, + 0.008515365421772003, + 0.014282691292464733, + 0.017310718074440956, + 0.007406236138194799, + 0.008212015964090824, + 0.013530121184885502, + 0.011762043461203575, + 0.018306180834770203, + 0.008998152799904346, + 0.0068064783699810505, + 0.003480395767837763, + 0.007512363139539957, + 0.002912576077505946, + 0.0043631368316709995, + 0.006228912156075239, + 0.0046800607815384865, + 0.007835633121430874, + 0.004816778469830751, + 0.0026451575104147196, + 0.011767134070396423, + 0.004307333845645189, + 0.004259753972291946, + 0.008948693051934242, + 0.008895273320376873, + 0.0047461665235459805, + 0.02713211439549923, + 0.01817922107875347, + 0.017190346494317055, + 0.015531730838119984, + 0.0209356676787138, + 0.11868180334568024, + 0.08404110372066498, + 0.020278258249163628, + 0.007751100696623325, + 0.010623766109347343, + 0.010908644646406174, + 0.011785218492150307, + 0.013655356131494045, + 0.007763293571770191, + 0.012295597232878208, + 0.006128779146820307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 3 + ], + "selected_reward": -0.005949170328676701 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering bronze statue of an ancient king situated in the center of a vast, stone-paved plaza?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue clad in dark, weathered bronze streaked with green patina, with one hand resting on a greatsword and a bearded face in a stern expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dark, turbulent sky with low-angle sunlight penetrating the clouds to cast long, stark shadows on the metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue exhibit a visible crack spreading from its base and begin to slowly tilt forward as it collapses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle view that trembles slightly due to a violent, low-frequency vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20155532658100128 + ], + "opposite_logprob_tokens": [ + -1.7015553712844849 + ], + "logprob_reward": -0.20155532658100128, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20155532658100128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03764534741640091, + 0.024825049564242363, + 0.06018751859664917, + 0.0353827066719532, + 0.021746324375271797, + 0.037911511957645416, + 0.07312659919261932, + 0.04685400053858757, + 0.5481343269348145, + 0.1141865998506546 + ], + "credit_map": [ + 0.005768630653619766, + 0.0024264685343950987, + 0.0055350386537611485, + 0.0031832780223339796, + 0.004751121625304222, + 0.0051221405155956745, + 0.0052791680209338665, + 0.005579502787441015, + 0.0032109511084854603, + 0.002984333084896207, + 0.002931149210780859, + 0.0024562301114201546, + 0.003618248039856553, + 0.0035327766090631485, + 0.0034367952030152082, + 0.002654565265402198, + 0.011022075079381466, + 0.008660736493766308, + 0.00489019462838769, + 0.004020775202661753, + 0.006093932781368494, + 0.0091328676789999, + 0.008047444745898247, + 0.008319489657878876, + 0.005185512825846672, + 0.0034926333464682102, + 0.003319003153592348, + 0.0036507542245090008, + 0.004192310851067305, + 0.004452227149158716, + 0.0057665579952299595, + 0.00532370712608099, + 0.0024733857717365026, + 0.0027465440798550844, + 0.0016809514490887523, + 0.002031497424468398, + 0.0024769536685198545, + 0.0030601462349295616, + 0.0036054528318345547, + 0.003671392099931836, + 0.0063111064955592155, + 0.005376208573579788, + 0.0031987265683710575, + 0.003343811258673668, + 0.004942215047776699, + 0.004109901841729879, + 0.006594513542950153, + 0.004035026300698519, + 0.008500452153384686, + 0.012846534140408039, + 0.007621455937623978, + 0.006879731081426144, + 0.009730050340294838, + 0.00879470445215702, + 0.010223429650068283, + 0.008530237711966038, + 0.0050134421326220036, + 0.003596722148358822, + 0.006219259463250637, + 0.004952090326696634, + 0.0061193229630589485, + 0.006760443560779095, + 0.009044278413057327, + 0.005148442927747965, + 0.04149162024259567, + 0.032822445034980774, + 0.02303401753306389, + 0.026579800993204117, + 0.03788500279188156, + 0.19725172221660614, + 0.1566401720046997, + 0.032429564744234085, + 0.011006047017872334, + 0.014776729978621006, + 0.01595756597816944, + 0.014962885528802872, + 0.01941383257508278, + 0.012443248182535172, + 0.01572989486157894, + 0.009896400384604931 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.20155532658100128 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering bronze statue of an ancient king situated in the center of a vast, stone-paved plaza?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue clad in dark, weathered bronze streaked with green patina, with one hand resting on a greatsword and a bearded face in a stern expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dark, turbulent sky with low-angle sunlight penetrating the clouds to cast long, stark shadows on the metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue exhibit a visible crack spreading from its base and begin to slowly tilt forward as it collapses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle view that trembles slightly due to a violent, low-frequency vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581839606165886 + ], + "opposite_logprob_tokens": [ + -3.8955817222595215 + ], + "logprob_reward": -0.020581839606165886, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581839606165886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1284945160150528, + 0.03800145536661148, + 0.07788956165313721, + 0.07768506556749344, + 0.047239214181900024, + 0.06763947010040283, + 0.09895305335521698, + 0.05729184299707413, + 0.36582159996032715, + 0.04098426550626755 + ], + "credit_map": [ + 0.016142021864652634, + 0.011640516109764576, + 0.0185588076710701, + 0.010857225395739079, + 0.019824953749775887, + 0.016774991527199745, + 0.018239960074424744, + 0.016456034034490585, + 0.006309489253908396, + 0.004450538195669651, + 0.004695953335613012, + 0.0027478523552417755, + 0.0043728528544306755, + 0.005478049628436565, + 0.0053224097937345505, + 0.0046243066899478436, + 0.013108020648360252, + 0.005267859902232885, + 0.007528283633291721, + 0.00509792473167181, + 0.012509158812463284, + 0.011999581009149551, + 0.013796558603644371, + 0.008582181297242641, + 0.010332778096199036, + 0.00908153597265482, + 0.005385372322052717, + 0.005437878426164389, + 0.010815339162945747, + 0.009739461354911327, + 0.013803121633827686, + 0.01308958139270544, + 0.00543151143938303, + 0.006016845349222422, + 0.0037673928309231997, + 0.003296640934422612, + 0.005156580824404955, + 0.006962854880839586, + 0.007055283989757299, + 0.009552101604640484, + 0.00947641208767891, + 0.011338828131556511, + 0.0059311361983418465, + 0.005506516899913549, + 0.008628140203654766, + 0.008203733712434769, + 0.01121748797595501, + 0.007337215356528759, + 0.010549711994826794, + 0.01588454097509384, + 0.008810468949377537, + 0.011120661161839962, + 0.015981096774339676, + 0.011810320429503918, + 0.01431768573820591, + 0.010478567332029343, + 0.0059331804513931274, + 0.0034244381822645664, + 0.010125864297151566, + 0.0036818210501223803, + 0.005434161517769098, + 0.011674937792122364, + 0.010632294230163097, + 0.0063851457089185715, + 0.0248105525970459, + 0.02047557756304741, + 0.013014265336096287, + 0.02351783961057663, + 0.03150784969329834, + 0.1193285658955574, + 0.11278358101844788, + 0.020383354276418686, + 0.0041140662506222725, + 0.004861766006797552, + 0.006300259847193956, + 0.006638599559664726, + 0.004088628571480513, + 0.006042877212166786, + 0.006086086854338646, + 0.002851977711543441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.020581839606165886 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide-angle shot captures the cataclysmic moment as a towering bronze statue of an ancient king, situated in the center of a vast, stone-paved plaza, begins its demise. The king, clad in dark, weathered bronze streaked with green patina, stands in an imperial pose, one hand resting on a greatsword, his bearded face set in a permanent, stern expression. Beneath a dark, turbulent sky, dramatic, low-angle sunlight penetrates the clouds, casting long, stark shadows that dimly shimmer across the metal surface. The entire scene is captured by a violent, low-frequency vibration that courses through the statue's massive form, causing the camera to tremble slightly from the shockwave. This is immediately followed by high-frequency dusting, removing oxidized metal flakes and small stone fragments from its massive base. With the muffled groan of shearing metal, a visible crack spreads from the statue's base, and the colossal figure begins to slowly and inexorably tilt forward. The statue's immense weight becomes the instrument of its own destruction, its once-proud posture twisting into a grotesque tilt as it begins to collapse, creating an atmosphere of terrifying, apocalyptic grandeur amidst the ruins.", + "seed": 16676, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_12/index_9705_seed_16676_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0609543533064425, + "vlm_logprob": -0.0609543533064425, + "vlm_yes_no_margin": 3.6500000238418577, + "vlm_token_credit": -0.0609543533064425, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering bronze statue of an ancient king situated in the center of a vast, stone-paved plaza?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue clad in dark, weathered bronze streaked with green patina, with one hand resting on a greatsword and a bearded face in a stern expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dark, turbulent sky with low-angle sunlight penetrating the clouds to cast long, stark shadows on the metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue exhibit a visible crack spreading from its base and begin to slowly tilt forward as it collapses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle view that trembles slightly due to a violent, low-frequency vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12835197150707245 + ], + "opposite_logprob_tokens": [ + -2.128351926803589 + ], + "logprob_reward": -0.12835197150707245, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12835197150707245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5280965566635132, + 0.03511859476566315, + 0.09520465135574341, + 0.05910240858793259, + 0.06939317286014557, + 0.05022028088569641, + 0.05201973021030426, + 0.02914067544043064, + 0.0638357475399971, + 0.017868205904960632 + ], + "credit_map": [ + 0.06584977358579636, + 0.044458989053964615, + 0.09613311290740967, + 0.055062539875507355, + 0.037463996559381485, + 0.07951945811510086, + 0.024607859551906586, + 0.12500084936618805, + 0.005320689175277948, + 0.004299014341086149, + 0.005207900423556566, + 0.0033739430364221334, + 0.00377137353643775, + 0.0037772043142467737, + 0.004352692514657974, + 0.0050157769583165646, + 0.01232190802693367, + 0.010806376114487648, + 0.013054762966930866, + 0.006882032845169306, + 0.011027169413864613, + 0.013668403960764408, + 0.006480121985077858, + 0.02096387930214405, + 0.007932409644126892, + 0.0070626805536448956, + 0.006936423014849424, + 0.0066696214489638805, + 0.009400758892297745, + 0.005410460755228996, + 0.009734715335071087, + 0.00595533661544323, + 0.009762188419699669, + 0.008287959732115269, + 0.015225442126393318, + 0.00634062709286809, + 0.009155893698334694, + 0.0049329837784171104, + 0.008554721251130104, + 0.007133360020816326, + 0.010324640199542046, + 0.005357654765248299, + 0.0050404928624629974, + 0.0042093670926988125, + 0.007754950318485498, + 0.007484655827283859, + 0.006142577156424522, + 0.0039059435948729515, + 0.00696879718452692, + 0.005080451723188162, + 0.0043859477154910564, + 0.005174028687179089, + 0.006240606773644686, + 0.005937236361205578, + 0.007472049444913864, + 0.010760613717138767, + 0.003767576301470399, + 0.004250253550708294, + 0.0033554958645254374, + 0.0043054791167378426, + 0.0035143340937793255, + 0.0018989413511008024, + 0.004434525500983, + 0.003614068264141679, + 0.005357303656637669, + 0.005663966294378042, + 0.0038860742934048176, + 0.015943637117743492, + 0.005135750398039818, + 0.0041083381511271, + 0.013842283748090267, + 0.009898393414914608, + 0.0017115904483944178, + 0.0018003300065174699, + 0.0020163669250905514, + 0.002327440772205591, + 0.002562256995588541, + 0.0015397057868540287, + 0.0034686920698732138, + 0.0024418230168521404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.12835197150707245 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering bronze statue of an ancient king situated in the center of a vast, stone-paved plaza?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue clad in dark, weathered bronze streaked with green patina, with one hand resting on a greatsword and a bearded face in a stern expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dark, turbulent sky with low-angle sunlight penetrating the clouds to cast long, stark shadows on the metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue exhibit a visible crack spreading from its base and begin to slowly tilt forward as it collapses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle view that trembles slightly due to a violent, low-frequency vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003626320045441389 + ], + "opposite_logprob_tokens": [ + -5.628626346588135 + ], + "logprob_reward": -0.003626320045441389, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003626320045441389, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.478667289018631, + 0.034660812467336655, + 0.045647356659173965, + 0.06267528980970383, + 0.034772466868162155, + 0.19026440382003784, + 0.06507705897092819, + 0.035803500562906265, + 0.03668053820729256, + 0.01575128734111786 + ], + "credit_map": [ + 0.05923590064048767, + 0.039514876902103424, + 0.08697054535150528, + 0.05070552974939346, + 0.03361119329929352, + 0.07232002913951874, + 0.02242053858935833, + 0.1138887107372284, + 0.0055047632195055485, + 0.003789092181250453, + 0.00490266177803278, + 0.0030462276190519333, + 0.0039076716639101505, + 0.003908136859536171, + 0.004131672438234091, + 0.005470586940646172, + 0.0053759170696139336, + 0.004757388960570097, + 0.005059409886598587, + 0.0033600840251892805, + 0.006839660461992025, + 0.006947449874132872, + 0.0033987427596002817, + 0.00990870501846075, + 0.008197975344955921, + 0.009084965102374554, + 0.006089094094932079, + 0.007198281120508909, + 0.009696191176772118, + 0.005491578485816717, + 0.010984218679368496, + 0.005932987201958895, + 0.006557559594511986, + 0.005508859176188707, + 0.004896028898656368, + 0.0027760688681155443, + 0.004463098477572203, + 0.002391334157437086, + 0.003967176657170057, + 0.004212342202663422, + 0.039930760860443115, + 0.021844202652573586, + 0.016234450042247772, + 0.016407020390033722, + 0.028294693678617477, + 0.03446514159440994, + 0.020168893039226532, + 0.012919247150421143, + 0.009618079289793968, + 0.005097422283142805, + 0.008163507096469402, + 0.004959208890795708, + 0.007131492253392935, + 0.007336496375501156, + 0.009231925010681152, + 0.013538933359086514, + 0.005443252623081207, + 0.0049897897988557816, + 0.004945538472384214, + 0.0033831971231848, + 0.0042788018472492695, + 0.00224540987983346, + 0.00754742044955492, + 0.002970094559714198, + 0.0032672923989593983, + 0.0029014183674007654, + 0.0028727457392960787, + 0.007385379169136286, + 0.006088509690016508, + 0.00171898293774575, + 0.00866895355284214, + 0.0037772583309561014, + 0.0016318653943017125, + 0.0019395961426198483, + 0.001838567783124745, + 0.002239715773612261, + 0.0018316665664315224, + 0.001655464293435216, + 0.0022679350804537535, + 0.0023464772384613752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.003626320045441389 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering bronze statue of an ancient king situated in the center of a vast, stone-paved plaza?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue clad in dark, weathered bronze streaked with green patina, with one hand resting on a greatsword and a bearded face in a stern expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dark, turbulent sky with low-angle sunlight penetrating the clouds to cast long, stark shadows on the metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue exhibit a visible crack spreading from its base and begin to slowly tilt forward as it collapses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle view that trembles slightly due to a violent, low-frequency vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006733820773661137 + ], + "opposite_logprob_tokens": [ + -5.0067338943481445 + ], + "logprob_reward": -0.006733820773661137, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006733820773661137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3318093717098236, + 0.04161577299237251, + 0.10330178588628769, + 0.085805244743824, + 0.10982872545719147, + 0.03864576667547226, + 0.07670489698648453, + 0.0661647617816925, + 0.11736519634723663, + 0.028758468106389046 + ], + "credit_map": [ + 0.04189585521817207, + 0.026622312143445015, + 0.06786265969276428, + 0.0320398174226284, + 0.019717229530215263, + 0.04896775260567665, + 0.014940119348466396, + 0.07976364344358444, + 0.005793705116957426, + 0.0057052792981266975, + 0.008082093670964241, + 0.003057965775951743, + 0.004444669466465712, + 0.0045386869460344315, + 0.0035310983657836914, + 0.006462275516241789, + 0.0150818582624197, + 0.013188834302127361, + 0.014002140611410141, + 0.00682045379653573, + 0.0094191525131464, + 0.014932294376194477, + 0.0066413250751793385, + 0.023215731605887413, + 0.012426400557160378, + 0.012256238609552383, + 0.011604051105678082, + 0.009101590141654015, + 0.011725407093763351, + 0.007253152783960104, + 0.013927195221185684, + 0.007511208299547434, + 0.017527759075164795, + 0.017616214230656624, + 0.020754368975758553, + 0.011018750257790089, + 0.014198875986039639, + 0.006969812326133251, + 0.013056144118309021, + 0.008686797693371773, + 0.006897491402924061, + 0.004249608609825373, + 0.00521206296980381, + 0.003538643941283226, + 0.00425547081977129, + 0.006639981642365456, + 0.005606423132121563, + 0.0022460848558694124, + 0.011471078731119633, + 0.006134429015219212, + 0.006923970300704241, + 0.006475267931818962, + 0.007681398186832666, + 0.01080949604511261, + 0.010309803299605846, + 0.01689945161342621, + 0.009869816713035107, + 0.008871778845787048, + 0.011344283819198608, + 0.007284226361662149, + 0.004688773769885302, + 0.004126921761780977, + 0.013153607957065105, + 0.006825350224971771, + 0.012131509371101856, + 0.011740781366825104, + 0.009542139247059822, + 0.027338191866874695, + 0.011319045908749104, + 0.007742729038000107, + 0.016485771164298058, + 0.021065030246973038, + 0.004393039271235466, + 0.0034433051478117704, + 0.002995803952217102, + 0.003012578235939145, + 0.004736057948321104, + 0.0026479328516870737, + 0.0037266789004206657, + 0.003803073428571224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.006733820773661137 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering bronze statue of an ancient king situated in the center of a vast, stone-paved plaza?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue clad in dark, weathered bronze streaked with green patina, with one hand resting on a greatsword and a bearded face in a stern expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dark, turbulent sky with low-angle sunlight penetrating the clouds to cast long, stark shadows on the metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue exhibit a visible crack spreading from its base and begin to slowly tilt forward as it collapses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle view that trembles slightly due to a violent, low-frequency vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14278019964694977 + ], + "opposite_logprob_tokens": [ + -2.017780303955078 + ], + "logprob_reward": -0.14278019964694977, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14278019964694977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22484774887561798, + 0.037159696221351624, + 0.06900419294834137, + 0.031271450221538544, + 0.06095445156097412, + 0.13158681988716125, + 0.06865730881690979, + 0.04899762570858002, + 0.2177184373140335, + 0.10980232059955597 + ], + "credit_map": [ + 0.02903757430613041, + 0.01889827661216259, + 0.04210631921887398, + 0.022577593103051186, + 0.017259225249290466, + 0.03361199423670769, + 0.009702197276055813, + 0.05165453255176544, + 0.006926773115992546, + 0.0051668002270162106, + 0.004637064412236214, + 0.004364640451967716, + 0.003987883683294058, + 0.0038332052063196898, + 0.0036951808724552393, + 0.004548140801489353, + 0.008779538795351982, + 0.008723358623683453, + 0.008236739784479141, + 0.005253949202597141, + 0.009259889833629131, + 0.008147808723151684, + 0.008739843033254147, + 0.011863055638968945, + 0.005297370720654726, + 0.004452654626220465, + 0.0041211447678506374, + 0.0031794654205441475, + 0.004414882976561785, + 0.0029786007944494486, + 0.0033452212810516357, + 0.0034821107983589172, + 0.010886156000196934, + 0.008732595480978489, + 0.009052366018295288, + 0.006718517746776342, + 0.0075576468370854855, + 0.004431072156876326, + 0.007026259321719408, + 0.0065498314797878265, + 0.027693524956703186, + 0.015255672857165337, + 0.013495517894625664, + 0.011411051265895367, + 0.015348726883530617, + 0.029685935005545616, + 0.009898402728140354, + 0.008797981776297092, + 0.00893661193549633, + 0.006608629133552313, + 0.017184922471642494, + 0.008041957393288612, + 0.005919239018112421, + 0.007758778985589743, + 0.007827579975128174, + 0.006379582453519106, + 0.0073647573590278625, + 0.00816641841083765, + 0.008587640710175037, + 0.007379851769655943, + 0.0038022086955606937, + 0.003794983960688114, + 0.00635779881849885, + 0.003543958766385913, + 0.017732489854097366, + 0.02793039381504059, + 0.01244764868170023, + 0.04106347635388374, + 0.02055879309773445, + 0.015167051926255226, + 0.044533904641866684, + 0.03828464820981026, + 0.012974449433386326, + 0.015064412727952003, + 0.011803108267486095, + 0.016833234578371048, + 0.0136452941223979, + 0.010898275300860405, + 0.012975966557860374, + 0.015607568435370922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.14278019964694977 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering bronze statue of an ancient king situated in the center of a vast, stone-paved plaza?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue clad in dark, weathered bronze streaked with green patina, with one hand resting on a greatsword and a bearded face in a stern expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dark, turbulent sky with low-angle sunlight penetrating the clouds to cast long, stark shadows on the metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue exhibit a visible crack spreading from its base and begin to slowly tilt forward as it collapses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle view that trembles slightly due to a violent, low-frequency vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023279454559087753 + ], + "opposite_logprob_tokens": [ + -3.7732794284820557 + ], + "logprob_reward": -0.023279454559087753, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023279454559087753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26132410764694214, + 0.06341742724180222, + 0.09587164223194122, + 0.13793455064296722, + 0.08980414271354675, + 0.09751188009977341, + 0.05686350539326668, + 0.043551813811063766, + 0.10223210602998734, + 0.05148880183696747 + ], + "credit_map": [ + 0.032439835369586945, + 0.01946345344185829, + 0.0508638434112072, + 0.03293328359723091, + 0.021314991638064384, + 0.03436616435647011, + 0.016568291932344437, + 0.05337424948811531, + 0.008748534135520458, + 0.007347778882831335, + 0.008234700188040733, + 0.006461234297603369, + 0.007939803414046764, + 0.007080381270498037, + 0.00794276874512434, + 0.009662224911153316, + 0.012591768056154251, + 0.011046350933611393, + 0.011598414741456509, + 0.007850940339267254, + 0.008620581589639187, + 0.013950234279036522, + 0.007852054201066494, + 0.02236129716038704, + 0.018999703228473663, + 0.01901962421834469, + 0.016818907111883163, + 0.014011785387992859, + 0.02082357183098793, + 0.01184488832950592, + 0.023793671280145645, + 0.012622412294149399, + 0.016041016206145287, + 0.015574385412037373, + 0.009256409481167793, + 0.008650209754705429, + 0.013564024120569229, + 0.005672926548868418, + 0.012646452523767948, + 0.008398719131946564, + 0.01235990971326828, + 0.015303028747439384, + 0.011763567104935646, + 0.015121354721486568, + 0.015397391282022, + 0.008985035121440887, + 0.010591761209070683, + 0.007989831268787384, + 0.008521960116922855, + 0.004810762591660023, + 0.005542742554098368, + 0.004935564938932657, + 0.005758768878877163, + 0.007876891642808914, + 0.007207667920738459, + 0.012209143489599228, + 0.005927297752350569, + 0.007552572526037693, + 0.005611644126474857, + 0.006655986420810223, + 0.004925113171339035, + 0.002546781674027443, + 0.006086550187319517, + 0.004245870281010866, + 0.008975381031632423, + 0.011510319076478481, + 0.007545083295553923, + 0.019839677959680557, + 0.010026740841567516, + 0.006493730470538139, + 0.018313845619559288, + 0.019527319818735123, + 0.00597400264814496, + 0.007583350874483585, + 0.0058072456158697605, + 0.007887075655162334, + 0.004942098166793585, + 0.0056031872518360615, + 0.005903994664549828, + 0.007787847425788641 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.023279454559087753 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide-angle shot captures the cataclysmic moment as a towering bronze statue of an ancient king, situated in the center of a vast, stone-paved plaza, begins its demise. The king, clad in dark, weathered bronze streaked with green patina, stands in an imperial pose, one hand resting on a greatsword, his bearded face set in a permanent, stern expression. Beneath a dark, turbulent sky, dramatic, low-angle sunlight penetrates the clouds, casting long, stark shadows that dimly shimmer across the metal surface. The entire scene is captured by a violent, low-frequency vibration that courses through the statue's massive form, causing the camera to tremble slightly from the shockwave. This is immediately followed by high-frequency dusting, removing oxidized metal flakes and small stone fragments from its massive base. With the muffled groan of shearing metal, a visible crack spreads from the statue's base, and the colossal figure begins to slowly and inexorably tilt forward. The statue's immense weight becomes the instrument of its own destruction, its once-proud posture twisting into a grotesque tilt as it begins to collapse, creating an atmosphere of terrifying, apocalyptic grandeur amidst the ruins.", + "seed": 16677, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_12/index_9705_seed_16677_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02284500761888921, + "vlm_logprob": -0.02284500761888921, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.02284500761888921, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering bronze statue of an ancient king situated in the center of a vast, stone-paved plaza?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue clad in dark, weathered bronze streaked with green patina, with one hand resting on a greatsword and a bearded face in a stern expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dark, turbulent sky with low-angle sunlight penetrating the clouds to cast long, stark shadows on the metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue exhibit a visible crack spreading from its base and begin to slowly tilt forward as it collapses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle view that trembles slightly due to a violent, low-frequency vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024484386667609215 + ], + "opposite_logprob_tokens": [ + -3.774484395980835 + ], + "logprob_reward": -0.024484386667609215, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.024484386667609215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20008975267410278, + 0.09309308975934982, + 0.020633339881896973, + 0.02735566534101963, + 0.06725791841745377, + 0.3609238862991333, + 0.10201244801282883, + 0.03268192335963249, + 0.041766367852687836, + 0.05418558046221733 + ], + "credit_map": [ + 0.020104559138417244, + 0.011866588145494461, + 0.021710319444537163, + 0.013486613519489765, + 0.02365400455892086, + 0.030284350737929344, + 0.03648211807012558, + 0.04250122979283333, + 0.009757410734891891, + 0.010112753137946129, + 0.006436425726860762, + 0.006363763939589262, + 0.015784641727805138, + 0.014433738775551319, + 0.014641981571912766, + 0.015562387183308601, + 0.001838300726376474, + 0.0018354335334151983, + 0.001478957710787654, + 0.002363962819799781, + 0.0032792009878903627, + 0.003396472195163369, + 0.004258452914655209, + 0.002182560507208109, + 0.002651435090228915, + 0.0022209621965885162, + 0.001724736881442368, + 0.002302490873262286, + 0.006170779466629028, + 0.0042258575558662415, + 0.0043289558961987495, + 0.0037304486613720655, + 0.006952466443181038, + 0.004675667267292738, + 0.005081782117486, + 0.006836917717009783, + 0.011093486100435257, + 0.014600205235183239, + 0.011122709140181541, + 0.006894692778587341, + 0.02384829893708229, + 0.01824001967906952, + 0.018367167562246323, + 0.009998174384236336, + 0.10925782471895218, + 0.1374214142560959, + 0.028935471549630165, + 0.014855539426207542, + 0.010256164707243443, + 0.008630294352769852, + 0.003863120684400201, + 0.004702643025666475, + 0.03482276573777199, + 0.02423935942351818, + 0.009209743700921535, + 0.0062883696518838406, + 0.003517103847116232, + 0.003602539421990514, + 0.0032696560956537724, + 0.006658249534666538, + 0.004954532720148563, + 0.003982531372457743, + 0.004180547781288624, + 0.0025167649146169424, + 0.0023821289651095867, + 0.004272341728210449, + 0.0053676278330385685, + 0.007329843007028103, + 0.004868411924690008, + 0.0045091332867741585, + 0.008515037596225739, + 0.004521851893514395, + 0.002972974209114909, + 0.006262174807488918, + 0.006330333184450865, + 0.010187630541622639, + 0.0074506537057459354, + 0.006644715089350939, + 0.008253333158791065, + 0.0060837711207568645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.024484386667609215 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering bronze statue of an ancient king situated in the center of a vast, stone-paved plaza?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue clad in dark, weathered bronze streaked with green patina, with one hand resting on a greatsword and a bearded face in a stern expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dark, turbulent sky with low-angle sunlight penetrating the clouds to cast long, stark shadows on the metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue exhibit a visible crack spreading from its base and begin to slowly tilt forward as it collapses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle view that trembles slightly due to a violent, low-frequency vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003203738247975707 + ], + "opposite_logprob_tokens": [ + -5.753203868865967 + ], + "logprob_reward": -0.003203738247975707, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003203738247975707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3224446773529053, + 0.03191839158535004, + 0.01869240216910839, + 0.018826769664883614, + 0.05582296848297119, + 0.4157257378101349, + 0.07621698826551437, + 0.026805441826581955, + 0.02088983729481697, + 0.012656754814088345 + ], + "credit_map": [ + 0.0335964560508728, + 0.02482573129236698, + 0.028347620740532875, + 0.023001590743660927, + 0.04251541942358017, + 0.046717289835214615, + 0.05416429415345192, + 0.06927628815174103, + 0.0025035508442670107, + 0.0022584721446037292, + 0.0037161922082304955, + 0.002732600085437298, + 0.0056370217353105545, + 0.0044799852184951305, + 0.006488779094070196, + 0.004101788625121117, + 0.001939122099429369, + 0.0014225252671167254, + 0.00137263722717762, + 0.0016125820111483335, + 0.0032090325839817524, + 0.002607951872050762, + 0.004100052639842033, + 0.002428496954962611, + 0.0020852177403867245, + 0.001318081165663898, + 0.0012253086315467954, + 0.0016273449873551726, + 0.004088003188371658, + 0.003602334763854742, + 0.002385045401751995, + 0.002495433669537306, + 0.004812508821487427, + 0.0029060414526611567, + 0.004977212753146887, + 0.00437541026622057, + 0.01052102167159319, + 0.014029907062649727, + 0.009936743415892124, + 0.004264122806489468, + 0.026466364040970802, + 0.02129299007356167, + 0.019011981785297394, + 0.008718599565327168, + 0.12833033502101898, + 0.16164197027683258, + 0.03271174803376198, + 0.01755174808204174, + 0.007378210313618183, + 0.0074506583623588085, + 0.0036588183138519526, + 0.004906468093395233, + 0.022916218265891075, + 0.017764220014214516, + 0.006797912996262312, + 0.00534448679536581, + 0.002427532570436597, + 0.005282096564769745, + 0.003385573858395219, + 0.003256093477830291, + 0.0027998469304293394, + 0.004039421211928129, + 0.0031737436074763536, + 0.002441133139654994, + 0.0015758122317492962, + 0.003541388548910618, + 0.002883713925257325, + 0.0023465834092348814, + 0.0019644536077976227, + 0.002597084501758218, + 0.0026132813654839993, + 0.0033675204031169415, + 0.0006910245283506811, + 0.0015326131833717227, + 0.0013858700403943658, + 0.0017083178972825408, + 0.0017856825143098831, + 0.0015259000938385725, + 0.0026285098865628242, + 0.0013988364953547716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.003203738247975707 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering bronze statue of an ancient king situated in the center of a vast, stone-paved plaza?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue clad in dark, weathered bronze streaked with green patina, with one hand resting on a greatsword and a bearded face in a stern expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dark, turbulent sky with low-angle sunlight penetrating the clouds to cast long, stark shadows on the metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue exhibit a visible crack spreading from its base and begin to slowly tilt forward as it collapses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle view that trembles slightly due to a violent, low-frequency vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252371542155743 + ], + "opposite_logprob_tokens": [ + -5.255252361297607 + ], + "logprob_reward": -0.005252371542155743, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252371542155743, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34920230507850647, + 0.07796822488307953, + 0.0971786230802536, + 0.04153268784284592, + 0.05081799998879433, + 0.07003358751535416, + 0.17606550455093384, + 0.051815565675497055, + 0.0499873086810112, + 0.03539818525314331 + ], + "credit_map": [ + 0.03651708737015724, + 0.020453521981835365, + 0.03632637858390808, + 0.021852843463420868, + 0.03419345244765282, + 0.052524033933877945, + 0.05741072818636894, + 0.08992428332567215, + 0.0075382087379693985, + 0.007806703913956881, + 0.007624875288456678, + 0.005606197752058506, + 0.014814366586506367, + 0.012265539728105068, + 0.012191414833068848, + 0.010120918974280357, + 0.009578808210790157, + 0.008515977300703526, + 0.009882834739983082, + 0.009601673111319542, + 0.014794771559536457, + 0.01343044824898243, + 0.017922598868608475, + 0.013451519422233105, + 0.005594203714281321, + 0.0028374555986374617, + 0.0029565237928181887, + 0.0036876769736409187, + 0.006356281228363514, + 0.006309536285698414, + 0.006934175733476877, + 0.00685683311894536, + 0.005585066974163055, + 0.002546419855207205, + 0.004699359182268381, + 0.003610235871747136, + 0.008344342932105064, + 0.013584069907665253, + 0.008371377363801003, + 0.004077129065990448, + 0.006018682848662138, + 0.003525784704834223, + 0.0044782585464417934, + 0.0033280414063483477, + 0.018946973606944084, + 0.02413507178425789, + 0.005967113189399242, + 0.00363366212695837, + 0.014039658941328526, + 0.02739475853741169, + 0.012060881592333317, + 0.013686522841453552, + 0.04095468297600746, + 0.03594856709241867, + 0.01857246458530426, + 0.013407960534095764, + 0.005006431136280298, + 0.010187606327235699, + 0.004809753503650427, + 0.004941870458424091, + 0.007655999157577753, + 0.01021616905927658, + 0.004815998952835798, + 0.004181738942861557, + 0.003051692619919777, + 0.006215399596840143, + 0.005935085471719503, + 0.008632883429527283, + 0.0059207165613770485, + 0.005229760892689228, + 0.007996665313839912, + 0.007005103398114443, + 0.0024205874651670456, + 0.004896134603768587, + 0.0038062231615185738, + 0.0037801482249051332, + 0.0038305455818772316, + 0.00645704660564661, + 0.005795121192932129, + 0.0044123767875134945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.005252371542155743 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering bronze statue of an ancient king situated in the center of a vast, stone-paved plaza?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue clad in dark, weathered bronze streaked with green patina, with one hand resting on a greatsword and a bearded face in a stern expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dark, turbulent sky with low-angle sunlight penetrating the clouds to cast long, stark shadows on the metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue exhibit a visible crack spreading from its base and begin to slowly tilt forward as it collapses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle view that trembles slightly due to a violent, low-frequency vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494380369782448 + ], + "opposite_logprob_tokens": [ + -2.929943799972534 + ], + "logprob_reward": -0.05494380369782448, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494380369782448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13263538479804993, + 0.04231851175427437, + 0.052401404827833176, + 0.03209272772073746, + 0.09383314102888107, + 0.21169975399971008, + 0.3852525055408478, + 0.018295062705874443, + 0.019511526450514793, + 0.011959971860051155 + ], + "credit_map": [ + 0.014128814451396465, + 0.010003378614783287, + 0.011359553784132004, + 0.009221874177455902, + 0.01712035946547985, + 0.01969929039478302, + 0.020602215081453323, + 0.030499903485178947, + 0.004095590673387051, + 0.0029151616618037224, + 0.004628606606274843, + 0.003392627928406, + 0.006387067027390003, + 0.005928988568484783, + 0.007700522430241108, + 0.0072699496522545815, + 0.005915847606956959, + 0.0028824240434914827, + 0.003430331125855446, + 0.0038337847217917442, + 0.009807164780795574, + 0.008371284231543541, + 0.010587346740067005, + 0.007573223672807217, + 0.005023718345910311, + 0.0014121225103735924, + 0.0020669475197792053, + 0.0019410740351304412, + 0.0037013408727943897, + 0.004908862058073282, + 0.004705466330051422, + 0.008333194069564342, + 0.009660869836807251, + 0.004441550001502037, + 0.008558251895010471, + 0.006492608692497015, + 0.01508120633661747, + 0.02637644298374653, + 0.01597519777715206, + 0.007247008848935366, + 0.015419424511492252, + 0.01294296607375145, + 0.010584448464214802, + 0.011587056331336498, + 0.05716817080974579, + 0.07336050271987915, + 0.019372427836060524, + 0.011264762841165066, + 0.03741082921624184, + 0.04132581874728203, + 0.017703527584671974, + 0.022602420300245285, + 0.11858831346035004, + 0.08908295631408691, + 0.03354461491107941, + 0.024994025006890297, + 0.0018733252072706819, + 0.0035739063750952482, + 0.0022002155892550945, + 0.0030826309230178595, + 0.002258103806525469, + 0.0018856850219890475, + 0.0018505186308175325, + 0.00157067715190351, + 0.0015895285177975893, + 0.0031129205599427223, + 0.002231884514912963, + 0.002426733262836933, + 0.002516543958336115, + 0.001967779127880931, + 0.003343658521771431, + 0.0023224777542054653, + 0.0007693881634622812, + 0.001478541293181479, + 0.001653827028349042, + 0.0012994944117963314, + 0.0017779739573597908, + 0.001866431557573378, + 0.0021301917731761932, + 0.0009841242572292686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.05494380369782448 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering bronze statue of an ancient king situated in the center of a vast, stone-paved plaza?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue clad in dark, weathered bronze streaked with green patina, with one hand resting on a greatsword and a bearded face in a stern expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dark, turbulent sky with low-angle sunlight penetrating the clouds to cast long, stark shadows on the metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue exhibit a visible crack spreading from its base and begin to slowly tilt forward as it collapses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle view that trembles slightly due to a violent, low-frequency vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634073793888092 + ], + "opposite_logprob_tokens": [ + -3.6513407230377197 + ], + "logprob_reward": -0.02634073793888092, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634073793888092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19313021004199982, + 0.06715881824493408, + 0.028568046167492867, + 0.07131941616535187, + 0.07629141956567764, + 0.23016130924224854, + 0.1074962466955185, + 0.09441573917865753, + 0.09615278989076614, + 0.035306014120578766 + ], + "credit_map": [ + 0.021224960684776306, + 0.010773123241961002, + 0.019607655704021454, + 0.012829374521970749, + 0.020671484991908073, + 0.029150279238820076, + 0.03103153593838215, + 0.04784180596470833, + 0.007926774211227894, + 0.006718733813613653, + 0.004490104038268328, + 0.00478758430108428, + 0.01304133702069521, + 0.010366244241595268, + 0.01167334709316492, + 0.008154687471687794, + 0.0028788854833692312, + 0.002414771355688572, + 0.001961621455848217, + 0.002743793884292245, + 0.005055845249444246, + 0.004195426125079393, + 0.005958567839115858, + 0.0033591329120099545, + 0.008256400935351849, + 0.004343012347817421, + 0.007149743847548962, + 0.007539017591625452, + 0.013119603507220745, + 0.009819889441132545, + 0.01202133297920227, + 0.009070409461855888, + 0.00799925159662962, + 0.002763443859294057, + 0.007789086550474167, + 0.0048043071292340755, + 0.008818771690130234, + 0.025230035185813904, + 0.013022995553910732, + 0.005863532889634371, + 0.016482191160321236, + 0.012195724062621593, + 0.011322058737277985, + 0.010067625902593136, + 0.06620204448699951, + 0.0838710218667984, + 0.018835967406630516, + 0.011184664443135262, + 0.008692240342497826, + 0.021010981872677803, + 0.007861160673201084, + 0.008787407539784908, + 0.021518833935260773, + 0.018631476908922195, + 0.011824888177216053, + 0.009169254451990128, + 0.00852440856397152, + 0.020368466153740883, + 0.009555185213685036, + 0.010861762799322605, + 0.012465476989746094, + 0.015999624505639076, + 0.009674743749201298, + 0.006966080982238054, + 0.006457297131419182, + 0.012300373055040836, + 0.01155600231140852, + 0.016643982380628586, + 0.011098995804786682, + 0.010489009320735931, + 0.01423417218029499, + 0.013372957706451416, + 0.0021664465311914682, + 0.0036106782499700785, + 0.004843275528401136, + 0.003996584564447403, + 0.004005115944892168, + 0.006133695598691702, + 0.005677377805113792, + 0.004872840829193592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.02634073793888092 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide-angle shot captures the cataclysmic moment as a towering bronze statue of an ancient king, situated in the center of a vast, stone-paved plaza, begins its demise. The king, clad in dark, weathered bronze streaked with green patina, stands in an imperial pose, one hand resting on a greatsword, his bearded face set in a permanent, stern expression. Beneath a dark, turbulent sky, dramatic, low-angle sunlight penetrates the clouds, casting long, stark shadows that dimly shimmer across the metal surface. The entire scene is captured by a violent, low-frequency vibration that courses through the statue's massive form, causing the camera to tremble slightly from the shockwave. This is immediately followed by high-frequency dusting, removing oxidized metal flakes and small stone fragments from its massive base. With the muffled groan of shearing metal, a visible crack spreads from the statue's base, and the colossal figure begins to slowly and inexorably tilt forward. The statue's immense weight becomes the instrument of its own destruction, its once-proud posture twisting into a grotesque tilt as it begins to collapse, creating an atmosphere of terrifying, apocalyptic grandeur amidst the ruins.", + "seed": 16678, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_12/index_9705_seed_16678_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0422086332924664, + "vlm_logprob": -0.0422086332924664, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.0422086332924664, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering bronze statue of an ancient king situated in the center of a vast, stone-paved plaza?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue clad in dark, weathered bronze streaked with green patina, with one hand resting on a greatsword and a bearded face in a stern expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dark, turbulent sky with low-angle sunlight penetrating the clouds to cast long, stark shadows on the metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue exhibit a visible crack spreading from its base and begin to slowly tilt forward as it collapses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle view that trembles slightly due to a violent, low-frequency vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16337250173091888 + ], + "opposite_logprob_tokens": [ + -1.91337251663208 + ], + "logprob_reward": -0.16337250173091888, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16337250173091888, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0753922313451767, + 0.1487503945827484, + 0.0942089781165123, + 0.14022338390350342, + 0.07693032920360565, + 0.10182751715183258, + 0.08355749398469925, + 0.12838299572467804, + 0.10936141014099121, + 0.04136529564857483 + ], + "credit_map": [ + 0.0131594343110919, + 0.014854224398732185, + 0.01036565937101841, + 0.007477087434381247, + 0.007648513186722994, + 0.008273831568658352, + 0.00975753366947174, + 0.003855952061712742, + 0.03006221354007721, + 0.03200786933302879, + 0.024750152602791786, + 0.01866295374929905, + 0.012827569618821144, + 0.012813310138881207, + 0.009787971153855324, + 0.007838346995413303, + 0.017246024683117867, + 0.010614980943500996, + 0.012483352795243263, + 0.013469982892274857, + 0.010302804410457611, + 0.013525268994271755, + 0.010084660723805428, + 0.006481901276856661, + 0.022343028336763382, + 0.01482962816953659, + 0.028303813189268112, + 0.01703670062124729, + 0.01331004872918129, + 0.01582402177155018, + 0.013867633417248726, + 0.0147085040807724, + 0.011078448966145515, + 0.009242186322808266, + 0.010274199768900871, + 0.008964729495346546, + 0.0098278122022748, + 0.010930903255939484, + 0.008745940402150154, + 0.007866106927394867, + 0.017955554649233818, + 0.014072616584599018, + 0.02209951914846897, + 0.01176067441701889, + 0.007675099186599255, + 0.012258358299732208, + 0.007237029727548361, + 0.008768659085035324, + 0.010469729080796242, + 0.008291666395962238, + 0.017761357128620148, + 0.008423790335655212, + 0.013294607400894165, + 0.009158202446997166, + 0.008497371338307858, + 0.007660773582756519, + 0.017257967963814735, + 0.016316896304488182, + 0.021159199997782707, + 0.013842414133250713, + 0.010803584940731525, + 0.02444375306367874, + 0.011628646403551102, + 0.012930531986057758, + 0.012192307971417904, + 0.018741972744464874, + 0.012823222205042839, + 0.018330978229641914, + 0.013596471399068832, + 0.007288769353181124, + 0.019839255139231682, + 0.006548442877829075, + 0.00404725968837738, + 0.005122975446283817, + 0.007130156271159649, + 0.00835739728063345, + 0.004255115054547787, + 0.002495196647942066, + 0.004636021796613932, + 0.005321172531694174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 7 + ], + "selected_reward": -0.16337250173091888 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering bronze statue of an ancient king situated in the center of a vast, stone-paved plaza?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue clad in dark, weathered bronze streaked with green patina, with one hand resting on a greatsword and a bearded face in a stern expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dark, turbulent sky with low-angle sunlight penetrating the clouds to cast long, stark shadows on the metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue exhibit a visible crack spreading from its base and begin to slowly tilt forward as it collapses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle view that trembles slightly due to a violent, low-frequency vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028348281048238277 + ], + "opposite_logprob_tokens": [ + -5.877834796905518 + ], + "logprob_reward": -0.0028348281048238277, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028348281048238277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06836817413568497, + 0.0609254352748394, + 0.054070595651865005, + 0.1788727343082428, + 0.2596161961555481, + 0.12520231306552887, + 0.047163818031549454, + 0.04073466360569, + 0.12653866410255432, + 0.03850739449262619 + ], + "credit_map": [ + 0.010524223558604717, + 0.007681282237172127, + 0.008022611029446125, + 0.006725254934281111, + 0.01113258395344019, + 0.006596673745661974, + 0.012453625909984112, + 0.005231916438788176, + 0.010636839084327221, + 0.011014255695044994, + 0.009804022498428822, + 0.008410192094743252, + 0.005221118219196796, + 0.006184672005474567, + 0.005503489635884762, + 0.004150847904384136, + 0.007823582738637924, + 0.004562908783555031, + 0.007349397987127304, + 0.006657297722995281, + 0.006292376667261124, + 0.008248385041952133, + 0.008472996763885021, + 0.004663650877773762, + 0.01836373284459114, + 0.019220663234591484, + 0.033616021275520325, + 0.02429213933646679, + 0.02003544010221958, + 0.021486559882760048, + 0.02130725234746933, + 0.020550936460494995, + 0.042051687836647034, + 0.05271950364112854, + 0.03226342797279358, + 0.03422215208411217, + 0.024523306638002396, + 0.02727349102497101, + 0.02400985360145569, + 0.022552762180566788, + 0.018163148313760757, + 0.021275119855999947, + 0.02682391367852688, + 0.012150770984590054, + 0.011213407851755619, + 0.015172039158642292, + 0.0071690864861011505, + 0.013234833255410194, + 0.005537330638617277, + 0.0036789507139474154, + 0.011631407774984837, + 0.0037799435667693615, + 0.007475357968360186, + 0.005324868485331535, + 0.0038236381951719522, + 0.0059123230166733265, + 0.006017317995429039, + 0.0031735182274132967, + 0.006544376723468304, + 0.004563448019325733, + 0.0035778535529971123, + 0.008731868118047714, + 0.0029621291905641556, + 0.0051641520112752914, + 0.013052167370915413, + 0.020136233419179916, + 0.012294207699596882, + 0.017253557220101357, + 0.01882743090391159, + 0.008334596641361713, + 0.02944863773882389, + 0.007191830314695835, + 0.003541381098330021, + 0.00546179199591279, + 0.005498321261256933, + 0.005103965289890766, + 0.0054163397289812565, + 0.002454242203384638, + 0.006758169736713171, + 0.0042731850408017635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 8 + ], + "selected_reward": -0.0028348281048238277 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering bronze statue of an ancient king situated in the center of a vast, stone-paved plaza?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue clad in dark, weathered bronze streaked with green patina, with one hand resting on a greatsword and a bearded face in a stern expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dark, turbulent sky with low-angle sunlight penetrating the clouds to cast long, stark shadows on the metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue exhibit a visible crack spreading from its base and begin to slowly tilt forward as it collapses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle view that trembles slightly due to a violent, low-frequency vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005951540544629097 + ], + "opposite_logprob_tokens": [ + -5.130951404571533 + ], + "logprob_reward": -0.005951540544629097, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005951540544629097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08186992257833481, + 0.05116916075348854, + 0.11721073091030121, + 0.16100063920021057, + 0.16651345789432526, + 0.142014279961586, + 0.08182651549577713, + 0.08481616526842117, + 0.061969317495822906, + 0.05160985141992569 + ], + "credit_map": [ + 0.011909843422472477, + 0.008467474021017551, + 0.01245969533920288, + 0.008845478296279907, + 0.013438266701996326, + 0.010227060876786709, + 0.01055445708334446, + 0.005967646837234497, + 0.009035624563694, + 0.009499775245785713, + 0.006552518345415592, + 0.006472098175436258, + 0.0044374060817062855, + 0.005192461423575878, + 0.005929021630436182, + 0.004050255287438631, + 0.02376614697277546, + 0.020544273778796196, + 0.01513922493904829, + 0.01929803006350994, + 0.00798533484339714, + 0.010352221317589283, + 0.01141190156340599, + 0.008713598363101482, + 0.026139963418245316, + 0.009455251507461071, + 0.06129109486937523, + 0.023952340707182884, + 0.010000042617321014, + 0.013956863433122635, + 0.006290228106081486, + 0.009914846159517765, + 0.024416737258434296, + 0.03541528806090355, + 0.020573187619447708, + 0.01988133043050766, + 0.01564823091030121, + 0.01983189769089222, + 0.013818278908729553, + 0.016928499564528465, + 0.02472788281738758, + 0.01759358122944832, + 0.03659125417470932, + 0.013380654156208038, + 0.010540194809436798, + 0.017278701066970825, + 0.009268627502024174, + 0.012633390724658966, + 0.011229372583329678, + 0.004596528131514788, + 0.027739031240344048, + 0.006726451683789492, + 0.007801525294780731, + 0.010541540570557117, + 0.0045195454731583595, + 0.008672520518302917, + 0.009214423596858978, + 0.009236167185008526, + 0.014480989426374435, + 0.008551720529794693, + 0.00578294089064002, + 0.01544512715190649, + 0.006316787097603083, + 0.015788007527589798, + 0.006598807405680418, + 0.009609876200556755, + 0.004856716375797987, + 0.008094330318272114, + 0.00880524329841137, + 0.004337093327194452, + 0.0160263329744339, + 0.0036409106105566025, + 0.006289750803261995, + 0.007388368248939514, + 0.007826012559235096, + 0.006446720100939274, + 0.007716614753007889, + 0.0034265059512108564, + 0.0069692786782979965, + 0.005546598695218563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.005951540544629097 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering bronze statue of an ancient king situated in the center of a vast, stone-paved plaza?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue clad in dark, weathered bronze streaked with green patina, with one hand resting on a greatsword and a bearded face in a stern expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dark, turbulent sky with low-angle sunlight penetrating the clouds to cast long, stark shadows on the metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue exhibit a visible crack spreading from its base and begin to slowly tilt forward as it collapses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle view that trembles slightly due to a violent, low-frequency vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541004456579685 + ], + "opposite_logprob_tokens": [ + -4.387540817260742 + ], + "logprob_reward": -0.012541004456579685, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541004456579685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.059015609323978424, + 0.04838845878839493, + 0.05186604708433151, + 0.09889468550682068, + 0.09279683232307434, + 0.13620176911354065, + 0.06263195723295212, + 0.14184461534023285, + 0.23098455369472504, + 0.07737544924020767 + ], + "credit_map": [ + 0.007180826272815466, + 0.0068678949028253555, + 0.008096137084066868, + 0.004975644405931234, + 0.007120717316865921, + 0.009390578605234623, + 0.011095643974840641, + 0.004288170021027327, + 0.007763360161334276, + 0.007104567252099514, + 0.006595185492187738, + 0.00474583962932229, + 0.004760390147566795, + 0.004845699295401573, + 0.009328698739409447, + 0.0032447187695652246, + 0.007196015678346157, + 0.008807233534753323, + 0.0054065166041255, + 0.00614819023758173, + 0.006325867492705584, + 0.005334187299013138, + 0.008415176533162594, + 0.004232858773320913, + 0.008535114116966724, + 0.007929213345050812, + 0.030803794041275978, + 0.016003983095288277, + 0.009002394042909145, + 0.014997080899775028, + 0.0060854158364236355, + 0.005537690129131079, + 0.017089879140257835, + 0.012880704365670681, + 0.010095101781189442, + 0.012630407698452473, + 0.010066375136375427, + 0.010252543725073338, + 0.009196423925459385, + 0.010585394687950611, + 0.021899258717894554, + 0.015214473009109497, + 0.03536675497889519, + 0.013557575643062592, + 0.009680929593741894, + 0.01549841370433569, + 0.011303027160465717, + 0.013681343756616116, + 0.008193441666662693, + 0.00683562271296978, + 0.014016090892255306, + 0.005589762702584267, + 0.008273879066109657, + 0.006554021965712309, + 0.006043312139809132, + 0.007125824689865112, + 0.025614166632294655, + 0.023763475939631462, + 0.017242394387722015, + 0.02033112570643425, + 0.010031811892986298, + 0.013646436855196953, + 0.01109755877405405, + 0.020117662847042084, + 0.032274674624204636, + 0.030836736783385277, + 0.03223889693617821, + 0.026848653331398964, + 0.026663653552532196, + 0.018015993759036064, + 0.034110598266124725, + 0.02999533712863922, + 0.007409951649606228, + 0.008529591374099255, + 0.012916822917759418, + 0.010145394131541252, + 0.010497920215129852, + 0.004671281669288874, + 0.012436087243258953, + 0.010768403299152851 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.012541004456579685 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering bronze statue of an ancient king situated in the center of a vast, stone-paved plaza?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue clad in dark, weathered bronze streaked with green patina, with one hand resting on a greatsword and a bearded face in a stern expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a dark, turbulent sky with low-angle sunlight penetrating the clouds to cast long, stark shadows on the metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue exhibit a visible crack spreading from its base and begin to slowly tilt forward as it collapses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle view that trembles slightly due to a violent, low-frequency vibration?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026343291625380516 + ], + "opposite_logprob_tokens": [ + -3.65134334564209 + ], + "logprob_reward": -0.026343291625380516, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026343291625380516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09958545863628387, + 0.06707355380058289, + 0.11124477535486221, + 0.30854731798171997, + 0.06184070557355881, + 0.11765551567077637, + 0.08120544999837875, + 0.07881470024585724, + 0.03809081390500069, + 0.03594169765710831 + ], + "credit_map": [ + 0.017787713557481766, + 0.01644832082092762, + 0.014697105623781681, + 0.009430106729269028, + 0.012520932592451572, + 0.008425327017903328, + 0.01458069309592247, + 0.005695255007594824, + 0.010186223313212395, + 0.013493579812347889, + 0.009473064914345741, + 0.007298095151782036, + 0.008820225484669209, + 0.005639825947582722, + 0.007208473980426788, + 0.0049540637992322445, + 0.015163721516728401, + 0.02282227762043476, + 0.012966918759047985, + 0.017982574179768562, + 0.010129895061254501, + 0.011737481690943241, + 0.010352794080972672, + 0.010089118964970112, + 0.04617227613925934, + 0.014791585505008698, + 0.12397854030132294, + 0.04698700085282326, + 0.017268890514969826, + 0.030143046751618385, + 0.012192386202514172, + 0.017013603821396828, + 0.009332520887255669, + 0.009663816541433334, + 0.008288772776722908, + 0.008543742820620537, + 0.006346115376800299, + 0.008487271144986153, + 0.00462359981611371, + 0.0065548657439649105, + 0.019176475703716278, + 0.009377460926771164, + 0.028245598077774048, + 0.011675997637212276, + 0.01151641272008419, + 0.013686876744031906, + 0.012793335132300854, + 0.011183357797563076, + 0.012082740664482117, + 0.0069985599257051945, + 0.012407993897795677, + 0.01195753924548626, + 0.011628007516264915, + 0.007624584250152111, + 0.007533562369644642, + 0.010972458869218826, + 0.012733226642012596, + 0.01108461432158947, + 0.01371617428958416, + 0.0075227562338113785, + 0.005732391495257616, + 0.010328870266675949, + 0.005836649797856808, + 0.011860013008117676, + 0.005971834994852543, + 0.004109065979719162, + 0.004466896411031485, + 0.005958653520792723, + 0.004456672351807356, + 0.0031420111190527678, + 0.005466872360557318, + 0.004518804140388966, + 0.004405323415994644, + 0.0052401176653802395, + 0.005259157624095678, + 0.004968029912561178, + 0.0038427847903221846, + 0.002446422353386879, + 0.006447742227464914, + 0.003332118270918727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.026343291625380516 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_13/index_tensor([2501, 2501])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_13/index_tensor([2501, 2501])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c25d8df38f13f435e41aa7363243b20297881e0b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_13/index_tensor([2501, 2501])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, wide shot slowly pans right, following a narrow, wooden boat as it glides silently and steadily across a mirror-like river. A thick, milky morning mist shrouds the scene, completely swallowing up the distant riverbank and any possible scenery, creating a sense of secluded tranquility. The first rays of morning sunlight struggle to penetrate the thick fog, casting a wide swath of warm, saturated golden light on the water, contrasting sharply with the soft blue tones permeating the mist, and imbuing the image with Steve McCurry's signature rich color style. The boat is made of weathered, dark wood, the scratches and grain of its hull clearly visible in the sidelight. The bow of the boat steadily breaks the surface of the water, leaving behind soft, diffuse ripples. On the boat stands a solitary figure in simple clothes and a conical hat. His back appears both small and powerful in the interweaving of morning light and mist. He paddles the oars rhythmically, each entry into the water and each lift off with precision and composure, as if he has become one with this ancient river. The whole scene is filled with a tranquil, mysterious and almost dreamy atmosphere, like a vivid painting frozen in time.", + "seed": 2843, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_13/index_2501_seed_2843_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.141940488293767, + "vlm_logprob": -0.141940488293767, + "vlm_yes_no_margin": 3.074999988079071, + "vlm_token_credit": -0.141940488293767, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a narrow wooden boat made of weathered dark wood with visible scratches and grain gliding across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure wearing simple clothes and a conical hat standing in the boat and paddling rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a river shrouded in thick milky morning mist with soft blue tones and warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the boat's bow break the water surface creating soft diffuse ripples while the distant riverbank is completely swallowed by mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3624691963195801 + ], + "opposite_logprob_tokens": [ + -1.23746919631958 + ], + "logprob_reward": -0.3624691963195801, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3624691963195801, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24254603683948517, + 0.06286907941102982, + 0.05178489163517952, + 0.12553377449512482, + 0.04612191766500473, + 0.043962664902210236, + 0.05556920915842056, + 0.06003003194928169, + 0.20707488059997559, + 0.1045074537396431 + ], + "credit_map": [ + 0.016065526753664017, + 0.031092777848243713, + 0.016341887414455414, + 0.039465680718421936, + 0.04389316216111183, + 0.04758969321846962, + 0.020738204941153526, + 0.02735910750925541, + 0.004279209766536951, + 0.008610693737864494, + 0.004681665450334549, + 0.007450425997376442, + 0.009984237141907215, + 0.010144568048417568, + 0.005432827863842249, + 0.01228545606136322, + 0.005060585215687752, + 0.008207369595766068, + 0.0042767999693751335, + 0.0048837922513484955, + 0.007956727407872677, + 0.01008824072778225, + 0.004498012363910675, + 0.00681336922571063, + 0.0124605568125844, + 0.01587470807135105, + 0.01011930126696825, + 0.021941915154457092, + 0.015352625399827957, + 0.020068209618330002, + 0.0104704350233078, + 0.019246038049459457, + 0.00405195401981473, + 0.007081711199134588, + 0.00396593427285552, + 0.008219151757657528, + 0.005362126976251602, + 0.0057306792587041855, + 0.0040450249798595905, + 0.007665339857339859, + 0.0033043627627193928, + 0.005000697448849678, + 0.005139512941241264, + 0.009038503281772137, + 0.004706878215074539, + 0.006744561716914177, + 0.004332138691097498, + 0.0056960126385092735, + 0.005751103162765503, + 0.008037630468606949, + 0.004031147807836533, + 0.006406697910279036, + 0.007770522031933069, + 0.01192270964384079, + 0.004594484809786081, + 0.007054918445646763, + 0.005690265446901321, + 0.008588937111198902, + 0.0053641884587705135, + 0.006592911668121815, + 0.005974495783448219, + 0.011744383722543716, + 0.005706860218197107, + 0.01036799419671297, + 0.014632265083491802, + 0.03308414667844772, + 0.02741970121860504, + 0.01649543270468712, + 0.023204293102025986, + 0.029094379395246506, + 0.021859081462025642, + 0.04128560051321983, + 0.008085881359875202, + 0.012492561712861061, + 0.01353757455945015, + 0.012189826928079128, + 0.011474776081740856, + 0.0170623566955328, + 0.010190010070800781, + 0.01947447843849659 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.3624691963195801 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a narrow wooden boat made of weathered dark wood with visible scratches and grain gliding across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure wearing simple clothes and a conical hat standing in the boat and paddling rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a river shrouded in thick milky morning mist with soft blue tones and warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the boat's bow break the water surface creating soft diffuse ripples while the distant riverbank is completely swallowed by mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31350022554397583 + ], + "opposite_logprob_tokens": [ + -1.313500165939331 + ], + "logprob_reward": -0.31350022554397583, + "yes_no_margin_reward": 0.9999999403953552, + "token_credit_reward": -0.31350022554397583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.164756178855896, + 0.16180483996868134, + 0.06722874939441681, + 0.0507475845515728, + 0.04694758355617523, + 0.17831580340862274, + 0.06530767679214478, + 0.0428108349442482, + 0.20328983664512634, + 0.01879090629518032 + ], + "credit_map": [ + 0.010681720450520515, + 0.0202940721064806, + 0.011487259529531002, + 0.015039480291306973, + 0.021376797929406166, + 0.026697542518377304, + 0.014500710181891918, + 0.044678591191768646, + 0.013439169153571129, + 0.01873316429555416, + 0.013554207980632782, + 0.01817493699491024, + 0.020813167095184326, + 0.028098084032535553, + 0.02109968289732933, + 0.02789243496954441, + 0.006519025191664696, + 0.008673107251524925, + 0.0056763868778944016, + 0.010218160226941109, + 0.008684673346579075, + 0.011900559067726135, + 0.0063162026926875114, + 0.00924063939601183, + 0.004225780721753836, + 0.007781710475683212, + 0.003558297408744693, + 0.005648930091410875, + 0.009132379665970802, + 0.01088304165750742, + 0.0033145921770483255, + 0.006202852353453636, + 0.004380380269140005, + 0.006661084946244955, + 0.004404814448207617, + 0.010781409218907356, + 0.004739857744425535, + 0.005573956295847893, + 0.003888932755216956, + 0.0065171485766768456, + 0.019116481766104698, + 0.02062023989856243, + 0.024071237072348595, + 0.0271627027541399, + 0.016586625948548317, + 0.020476702600717545, + 0.02380804903805256, + 0.02647375874221325, + 0.0046522649936378, + 0.008412977680563927, + 0.005813992116600275, + 0.011077886447310448, + 0.0076348367147147655, + 0.013723993673920631, + 0.007419004570692778, + 0.00657272431999445, + 0.0031573502346873283, + 0.0065844678319990635, + 0.0055170441046357155, + 0.005413161125034094, + 0.006732727400958538, + 0.005856426898390055, + 0.004935065284371376, + 0.004614592995494604, + 0.014034726656973362, + 0.033459946513175964, + 0.024174075573682785, + 0.023412901908159256, + 0.03172554448246956, + 0.023177271708846092, + 0.02581317536532879, + 0.027492178604006767, + 0.0019525215029716492, + 0.002119701122865081, + 0.0023734073620289564, + 0.0030400357209146023, + 0.002042942913249135, + 0.0025527209509164095, + 0.0022981807123869658, + 0.0024113955441862345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.31350022554397583 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a narrow wooden boat made of weathered dark wood with visible scratches and grain gliding across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure wearing simple clothes and a conical hat standing in the boat and paddling rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a river shrouded in thick milky morning mist with soft blue tones and warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the boat's bow break the water surface creating soft diffuse ripples while the distant riverbank is completely swallowed by mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01255148183554411 + ], + "opposite_logprob_tokens": [ + -4.387551307678223 + ], + "logprob_reward": -0.01255148183554411, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01255148183554411, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1824582815170288, + 0.10569749772548676, + 0.13600465655326843, + 0.20837752521038055, + 0.06290063261985779, + 0.04669083282351494, + 0.09520389139652252, + 0.09228131175041199, + 0.04473056271672249, + 0.025654833763837814 + ], + "credit_map": [ + 0.014971334487199783, + 0.021755993366241455, + 0.022219501435756683, + 0.032469604164361954, + 0.02457846887409687, + 0.03275278955698013, + 0.01800929382443428, + 0.01570129580795765, + 0.009551746770739555, + 0.012973209843039513, + 0.00956195779144764, + 0.01777341403067112, + 0.01264481246471405, + 0.017108192667365074, + 0.01016340870410204, + 0.01592075638473034, + 0.009773651137948036, + 0.0198200736194849, + 0.008606073446571827, + 0.012806297279894352, + 0.02132406085729599, + 0.03745711222290993, + 0.011289575137197971, + 0.014927811920642853, + 0.01695389300584793, + 0.02629115805029869, + 0.023241041228175163, + 0.05118579789996147, + 0.02203594520688057, + 0.02346399426460266, + 0.017962809652090073, + 0.02724290080368519, + 0.005544349551200867, + 0.008476341143250465, + 0.005969678983092308, + 0.014533781446516514, + 0.005506240762770176, + 0.007481842301785946, + 0.004153725691139698, + 0.011234677396714687, + 0.004768901038914919, + 0.005799591541290283, + 0.005594010930508375, + 0.00679710553959012, + 0.0059256465174257755, + 0.0068121436052024364, + 0.005980647634714842, + 0.005012787878513336, + 0.00943638663738966, + 0.012552471831440926, + 0.0070966994389891624, + 0.014259954914450645, + 0.010944431647658348, + 0.02300839126110077, + 0.008865341544151306, + 0.009040207602083683, + 0.007680182810872793, + 0.01869182102382183, + 0.008645712397992611, + 0.012813800945878029, + 0.01459641195833683, + 0.012327794916927814, + 0.00789946224540472, + 0.009626124054193497, + 0.003037126502022147, + 0.007478056009858847, + 0.006506609730422497, + 0.003858159761875868, + 0.006225335877388716, + 0.004660269245505333, + 0.004530185367912054, + 0.008434821851551533, + 0.0020315158180892467, + 0.0034115933813154697, + 0.003337280824780464, + 0.0031148537527769804, + 0.002410935703665018, + 0.005324752070009708, + 0.0019995758775621653, + 0.0040243263356387615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.01255148183554411 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a narrow wooden boat made of weathered dark wood with visible scratches and grain gliding across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure wearing simple clothes and a conical hat standing in the boat and paddling rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a river shrouded in thick milky morning mist with soft blue tones and warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the boat's bow break the water surface creating soft diffuse ripples while the distant riverbank is completely swallowed by mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012538178823888302 + ], + "opposite_logprob_tokens": [ + -4.387537956237793 + ], + "logprob_reward": -0.012538178823888302, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012538178823888302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25036078691482544, + 0.1317545771598816, + 0.11383794248104095, + 0.15750256180763245, + 0.06016208976507187, + 0.06157825514674187, + 0.08606979250907898, + 0.09205321967601776, + 0.022839291021227837, + 0.02384147047996521 + ], + "credit_map": [ + 0.020849721506237984, + 0.031031465157866478, + 0.013417703099548817, + 0.017156392335891724, + 0.03498692438006401, + 0.04995345324277878, + 0.024496303871273994, + 0.058468833565711975, + 0.010776122100651264, + 0.018068427219986916, + 0.006827231030911207, + 0.0090264193713665, + 0.023771777749061584, + 0.03605307266116142, + 0.007637643720954657, + 0.019593896344304085, + 0.009642143733799458, + 0.02127087116241455, + 0.004510986618697643, + 0.005814770236611366, + 0.02096700295805931, + 0.03322213888168335, + 0.004104395397007465, + 0.01430563349276781, + 0.01198112964630127, + 0.026217486709356308, + 0.007104854565113783, + 0.01012889388948679, + 0.027638349682092667, + 0.036990538239479065, + 0.00973592884838581, + 0.02770538069307804, + 0.004715512972325087, + 0.009654957801103592, + 0.003118969267234206, + 0.005349024664610624, + 0.011172543279826641, + 0.01449295412749052, + 0.004107286222279072, + 0.0075508421286940575, + 0.00472797267138958, + 0.008391836658120155, + 0.004302049521356821, + 0.005306660197675228, + 0.009406124241650105, + 0.015372587367892265, + 0.004020968917757273, + 0.010050058364868164, + 0.007966121658682823, + 0.014192966744303703, + 0.0028426628559827805, + 0.00345259509049356, + 0.013252455741167068, + 0.026635712012648582, + 0.003914317116141319, + 0.013812960125505924, + 0.006465972401201725, + 0.021802620962262154, + 0.003893931396305561, + 0.002900801133364439, + 0.017690135166049004, + 0.018930166959762573, + 0.004730017390102148, + 0.015639569610357285, + 0.001884938683360815, + 0.002866618800908327, + 0.0026230458170175552, + 0.002228704048320651, + 0.0030903478618711233, + 0.0043929992243647575, + 0.0026919147931039333, + 0.003060721792280674, + 0.0015909923240542412, + 0.0046676285564899445, + 0.0011820952640846372, + 0.0015183317009359598, + 0.003253482747823, + 0.004769444465637207, + 0.0016143653774634004, + 0.00524513004347682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.012538178823888302 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a narrow wooden boat made of weathered dark wood with visible scratches and grain gliding across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure wearing simple clothes and a conical hat standing in the boat and paddling rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a river shrouded in thick milky morning mist with soft blue tones and warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the boat's bow break the water surface creating soft diffuse ripples while the distant riverbank is completely swallowed by mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008643358945846558 + ], + "opposite_logprob_tokens": [ + -4.75864315032959 + ], + "logprob_reward": -0.008643358945846558, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008643358945846558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2680833637714386, + 0.10276775807142258, + 0.13529998064041138, + 0.1033342033624649, + 0.0660359114408493, + 0.09276600927114487, + 0.04568704590201378, + 0.07827514410018921, + 0.06071330979466438, + 0.0470372699201107 + ], + "credit_map": [ + 0.019996697083115578, + 0.03939395770430565, + 0.015368404798209667, + 0.03178388997912407, + 0.05776406452059746, + 0.0545649491250515, + 0.02183164656162262, + 0.027379751205444336, + 0.008307591080665588, + 0.01438593864440918, + 0.00765899196267128, + 0.013012424111366272, + 0.01556738093495369, + 0.021575164049863815, + 0.008760524913668633, + 0.013499735854566097, + 0.009436527267098427, + 0.022853899747133255, + 0.005717214196920395, + 0.007741071283817291, + 0.026414962485432625, + 0.04663431644439697, + 0.005586585495620966, + 0.010915402323007584, + 0.008823014795780182, + 0.012519522570073605, + 0.007317656651139259, + 0.01322442851960659, + 0.013379906304180622, + 0.019382797181606293, + 0.007988705299794674, + 0.020698171108961105, + 0.0062361168675124645, + 0.010401614010334015, + 0.0044748554937541485, + 0.006260393187403679, + 0.011608066968619823, + 0.014470741152763367, + 0.0050422693602740765, + 0.007541859522461891, + 0.006197093985974789, + 0.012227492406964302, + 0.007245982531458139, + 0.008509792387485504, + 0.015735840424895287, + 0.02194872871041298, + 0.008270680904388428, + 0.012630399316549301, + 0.004633911419659853, + 0.007362423464655876, + 0.0024646406527608633, + 0.0045565953478217125, + 0.0066136266104876995, + 0.010494164191186428, + 0.0043899789452552795, + 0.005171709693968296, + 0.006527898367494345, + 0.014435862191021442, + 0.0032648183405399323, + 0.004709117580205202, + 0.01199332345277071, + 0.017457937821745872, + 0.004891423042863607, + 0.014994769357144833, + 0.004304183647036552, + 0.010017124004662037, + 0.0045889453031122684, + 0.005664394702762365, + 0.010651243850588799, + 0.010603351518511772, + 0.008079293183982372, + 0.006804770790040493, + 0.004277095664292574, + 0.0068887751549482346, + 0.0032929712906479836, + 0.003176593454554677, + 0.007096076384186745, + 0.01072463858872652, + 0.0036301296204328537, + 0.007950989529490471 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.008643358945846558 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide shot slowly pans right, following a narrow, wooden boat as it glides silently and steadily across a mirror-like river. A thick, milky morning mist shrouds the scene, completely swallowing up the distant riverbank and any possible scenery, creating a sense of secluded tranquility. The first rays of morning sunlight struggle to penetrate the thick fog, casting a wide swath of warm, saturated golden light on the water, contrasting sharply with the soft blue tones permeating the mist, and imbuing the image with Steve McCurry's signature rich color style. The boat is made of weathered, dark wood, the scratches and grain of its hull clearly visible in the sidelight. The bow of the boat steadily breaks the surface of the water, leaving behind soft, diffuse ripples. On the boat stands a solitary figure in simple clothes and a conical hat. His back appears both small and powerful in the interweaving of morning light and mist. He paddles the oars rhythmically, each entry into the water and each lift off with precision and composure, as if he has become one with this ancient river. The whole scene is filled with a tranquil, mysterious and almost dreamy atmosphere, like a vivid painting frozen in time.", + "seed": 2844, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_13/index_2501_seed_2844_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02493084855377674, + "vlm_logprob": -0.02493084855377674, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.02493084855377674, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a narrow wooden boat made of weathered dark wood with visible scratches and grain gliding across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure wearing simple clothes and a conical hat standing in the boat and paddling rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a river shrouded in thick milky morning mist with soft blue tones and warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the boat's bow break the water surface creating soft diffuse ripples while the distant riverbank is completely swallowed by mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05892226845026016 + ], + "opposite_logprob_tokens": [ + -3.183922290802002 + ], + "logprob_reward": -0.05892226845026016, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.05892226845026016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07886341959238052, + 0.07591111958026886, + 0.039728276431560516, + 0.04431579262018204, + 0.11294729262590408, + 0.10990619659423828, + 0.28681620955467224, + 0.029848111793398857, + 0.007461908273398876, + 0.21420170366764069 + ], + "credit_map": [ + 0.009925371035933495, + 0.009466270916163921, + 0.007143969181925058, + 0.008216076530516148, + 0.01159784197807312, + 0.013440124690532684, + 0.010873675346374512, + 0.00820008572191, + 0.004292537923902273, + 0.003424137830734253, + 0.002274945843964815, + 0.003088521771132946, + 0.028724122792482376, + 0.02471052296459675, + 0.0046648201532661915, + 0.0047315070405602455, + 0.004355229903012514, + 0.0033184783533215523, + 0.0019921923521906137, + 0.002763614524155855, + 0.008529501967132092, + 0.007665340788662434, + 0.004248549696058035, + 0.0068553658202290535, + 0.004482700023800135, + 0.005055964924395084, + 0.003686954267323017, + 0.0032476806081831455, + 0.00913156382739544, + 0.0061235129833221436, + 0.007619502954185009, + 0.004967914428561926, + 0.013610309921205044, + 0.01826612278819084, + 0.004649307578802109, + 0.0069693997502326965, + 0.018850818276405334, + 0.02187647856771946, + 0.013642939738929272, + 0.015081914141774178, + 0.017008963972330093, + 0.016449222341179848, + 0.00622920598834753, + 0.011119633913040161, + 0.014779865741729736, + 0.015648409724235535, + 0.015372131951153278, + 0.013298767618834972, + 0.04397526755928993, + 0.03710193559527397, + 0.02241029217839241, + 0.024173839017748833, + 0.022165587171912193, + 0.036340322345495224, + 0.04291070997714996, + 0.057738278061151505, + 0.004058739636093378, + 0.004898937419056892, + 0.0018297045025974512, + 0.0035790828987956047, + 0.004180937074124813, + 0.003081109607592225, + 0.004352619405835867, + 0.0038669821806252003, + 0.0006980359903536737, + 0.0007331743254326284, + 0.0005904415738768876, + 0.0007926099351607263, + 0.0008001932874321938, + 0.001119834603741765, + 0.0010572539176791906, + 0.0016703642904758453, + 0.014323392882943153, + 0.08461599797010422, + 0.0025753523223102093, + 0.005121039692312479, + 0.009939148090779781, + 0.08328137546777725, + 0.0039124870672822, + 0.01043290551751852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 4 + ], + "selected_reward": -0.05892226845026016 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a narrow wooden boat made of weathered dark wood with visible scratches and grain gliding across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure wearing simple clothes and a conical hat standing in the boat and paddling rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a river shrouded in thick milky morning mist with soft blue tones and warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the boat's bow break the water surface creating soft diffuse ripples while the distant riverbank is completely swallowed by mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813057765364647 + ], + "opposite_logprob_tokens": [ + -3.288130521774292 + ], + "logprob_reward": -0.03813057765364647, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813057765364647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15865541994571686, + 0.1508191078901291, + 0.050917401909828186, + 0.03636382892727852, + 0.2283186912536621, + 0.04924984276294708, + 0.07796414941549301, + 0.040247842669487, + 0.021503575146198273, + 0.18596011400222778 + ], + "credit_map": [ + 0.016264120116829872, + 0.01841791905462742, + 0.015452499501407146, + 0.011563354171812534, + 0.01742532290518284, + 0.03344564512372017, + 0.02274906635284424, + 0.02333749830722809, + 0.008177435025572777, + 0.009226260706782341, + 0.0034663008991628885, + 0.004759901203215122, + 0.05950796976685524, + 0.05261315032839775, + 0.004748737905174494, + 0.008319356478750706, + 0.003847370622679591, + 0.005605801474303007, + 0.002388815162703395, + 0.0032533868215978146, + 0.013209972530603409, + 0.013343426398932934, + 0.0038529345765709877, + 0.005415694322437048, + 0.0023530111648142338, + 0.0022946111857891083, + 0.0018472768133506179, + 0.0021724284160882235, + 0.008623562753200531, + 0.00846804864704609, + 0.005403666291385889, + 0.005201227497309446, + 0.027939055114984512, + 0.03791433945298195, + 0.009869914501905441, + 0.014920325949788094, + 0.03820732608437538, + 0.04043415188789368, + 0.028228670358657837, + 0.030804922804236412, + 0.005379753652960062, + 0.007138730958104134, + 0.0035901167429983616, + 0.005803532898426056, + 0.0071526928804814816, + 0.005351748317480087, + 0.007535771932452917, + 0.007297496777027845, + 0.011507497169077396, + 0.009823628701269627, + 0.00588202802464366, + 0.0057260519824922085, + 0.006362174637615681, + 0.009504228830337524, + 0.012324086390435696, + 0.016834450885653496, + 0.00438019959256053, + 0.006882672663778067, + 0.003336170455440879, + 0.004659147933125496, + 0.004890010692179203, + 0.00723078241571784, + 0.005211460869759321, + 0.003657398046925664, + 0.003109412267804146, + 0.0035623377189040184, + 0.0016871315892785788, + 0.00354609708301723, + 0.002476551802828908, + 0.0028462500777095556, + 0.0022682349663227797, + 0.0020075596403330564, + 0.012443343177437782, + 0.06880156695842743, + 0.0038363677449524403, + 0.008820499293506145, + 0.008967743255198002, + 0.06826040148735046, + 0.00485267722979188, + 0.009977529756724834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.03813057765364647 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a narrow wooden boat made of weathered dark wood with visible scratches and grain gliding across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure wearing simple clothes and a conical hat standing in the boat and paddling rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a river shrouded in thick milky morning mist with soft blue tones and warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the boat's bow break the water surface creating soft diffuse ripples while the distant riverbank is completely swallowed by mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009788372553884983 + ], + "opposite_logprob_tokens": [ + -4.634788513183594 + ], + "logprob_reward": -0.009788372553884983, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009788372553884983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1395438015460968, + 0.1836574524641037, + 0.053444258868694305, + 0.030730707570910454, + 0.39327919483184814, + 0.05095570906996727, + 0.07406654208898544, + 0.011672035790979862, + 0.008627236820757389, + 0.05402301996946335 + ], + "credit_map": [ + 0.012907199561595917, + 0.018948663026094437, + 0.0149690555408597, + 0.012185831554234028, + 0.015401261858642101, + 0.024781472980976105, + 0.021677115932106972, + 0.018673209473490715, + 0.0104560237377882, + 0.006786135025322437, + 0.003969298675656319, + 0.005699132569134235, + 0.07576024532318115, + 0.06559987366199493, + 0.0058246394619345665, + 0.009562111459672451, + 0.002411609748378396, + 0.0035333912819623947, + 0.0016786702908575535, + 0.002101022284477949, + 0.019518684595823288, + 0.017980830743908882, + 0.002501375274732709, + 0.0037186797708272934, + 0.002929462818428874, + 0.0035316902212798595, + 0.001745206885971129, + 0.0019160680240020156, + 0.007111590355634689, + 0.003949295729398727, + 0.0056230006739497185, + 0.003924395889043808, + 0.04303031414747238, + 0.06166538968682289, + 0.017011716961860657, + 0.023255543783307076, + 0.05808244273066521, + 0.08313703536987305, + 0.04830380156636238, + 0.05879294127225876, + 0.005234161391854286, + 0.007058980409055948, + 0.003215656615793705, + 0.004909330978989601, + 0.00923733040690422, + 0.006224960088729858, + 0.008554467000067234, + 0.0065208254382014275, + 0.012359940446913242, + 0.010140606202185154, + 0.005924195982515812, + 0.006234609987586737, + 0.0058814287185668945, + 0.00941440835595131, + 0.009850643575191498, + 0.014260712079703808, + 0.0011209325166419148, + 0.0013630561297759414, + 0.0013617529766634107, + 0.0012905398616567254, + 0.0013874705182388425, + 0.0020224598702043295, + 0.0017103329300880432, + 0.0014154912205412984, + 0.001236101146787405, + 0.0012192833237349987, + 0.000686799583490938, + 0.0011730521218851209, + 0.0011061918921768665, + 0.0010789124062284827, + 0.0010560363298282027, + 0.0010708605404943228, + 0.0035052441526204348, + 0.020032690837979317, + 0.0010090504074469209, + 0.0025552227161824703, + 0.002720074960961938, + 0.019860820844769478, + 0.0015977310249581933, + 0.0027421913109719753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.009788372553884983 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a narrow wooden boat made of weathered dark wood with visible scratches and grain gliding across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure wearing simple clothes and a conical hat standing in the boat and paddling rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a river shrouded in thick milky morning mist with soft blue tones and warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the boat's bow break the water surface creating soft diffuse ripples while the distant riverbank is completely swallowed by mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003613610751926899 + ], + "opposite_logprob_tokens": [ + -5.628613471984863 + ], + "logprob_reward": -0.003613610751926899, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003613610751926899, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1099429577589035, + 0.07280094921588898, + 0.16696590185165405, + 0.12223467230796814, + 0.031773824244737625, + 0.03974949195981026, + 0.07615616917610168, + 0.013249116949737072, + 0.008778374642133713, + 0.358348548412323 + ], + "credit_map": [ + 0.009849372319877148, + 0.008819052018225193, + 0.0074378252029418945, + 0.007131331134587526, + 0.014307349920272827, + 0.03213537111878395, + 0.013764731585979462, + 0.01649792678654194, + 0.009713806211948395, + 0.005316098220646381, + 0.0028796044643968344, + 0.0033607042860239744, + 0.02600187249481678, + 0.014508545398712158, + 0.004740195814520121, + 0.0062801288440823555, + 0.007379425223916769, + 0.004601623397320509, + 0.002608558163046837, + 0.0029042158275842667, + 0.06638836860656738, + 0.06492633372545242, + 0.00579244177788496, + 0.012364937923848629, + 0.01029204111546278, + 0.008667593821883202, + 0.005030185449868441, + 0.006722332444041967, + 0.02853579819202423, + 0.0276829581707716, + 0.013790179044008255, + 0.021513577550649643, + 0.0034469745587557554, + 0.0050911325961351395, + 0.0013604139676317573, + 0.0022041539195924997, + 0.004559493623673916, + 0.007021317724138498, + 0.0035364909563213587, + 0.004553849343210459, + 0.005705362651497126, + 0.006201046518981457, + 0.002036582212895155, + 0.0034427354112267494, + 0.004364744294434786, + 0.004793817177414894, + 0.005508678499609232, + 0.00769652659073472, + 0.012498418800532818, + 0.010813991539180279, + 0.00607660785317421, + 0.006456590257585049, + 0.005713053047657013, + 0.009765740483999252, + 0.009851258248090744, + 0.014980507083237171, + 0.0019080195343121886, + 0.002258880762383342, + 0.0007606035796925426, + 0.0010754705872386694, + 0.001111306599341333, + 0.001844750833697617, + 0.0014009129954501987, + 0.002889171941205859, + 0.0009250353323295712, + 0.00127518514636904, + 0.0007731678779236972, + 0.0012397919781506062, + 0.0010023724753409624, + 0.001159865758381784, + 0.0011076059890910983, + 0.0012953504920005798, + 0.024562889710068703, + 0.1435462385416031, + 0.0031311041675508022, + 0.006846828386187553, + 0.016562258824706078, + 0.14019346237182617, + 0.005537046119570732, + 0.01796867698431015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 3 + ], + "selected_reward": -0.003613610751926899 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a narrow wooden boat made of weathered dark wood with visible scratches and grain gliding across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure wearing simple clothes and a conical hat standing in the boat and paddling rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a river shrouded in thick milky morning mist with soft blue tones and warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the boat's bow break the water surface creating soft diffuse ripples while the distant riverbank is completely swallowed by mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014199413359165192 + ], + "opposite_logprob_tokens": [ + -4.264199256896973 + ], + "logprob_reward": -0.014199413359165192, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014199413359165192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07016127556562424, + 0.11832265555858612, + 0.13919034600257874, + 0.03995060175657272, + 0.01874282769858837, + 0.03754706680774689, + 0.10930968821048737, + 0.026981843635439873, + 0.008216099813580513, + 0.43157759308815 + ], + "credit_map": [ + 0.006831951439380646, + 0.009019231423735619, + 0.0061033545061945915, + 0.005265219137072563, + 0.010424908250570297, + 0.014631743542850018, + 0.008928831666707993, + 0.008956038393080235, + 0.011515467427670956, + 0.007641736883670092, + 0.0037809223867952824, + 0.005983199458569288, + 0.04026774317026138, + 0.02809981070458889, + 0.01106241811066866, + 0.009971361607313156, + 0.006700740195810795, + 0.0036495875101536512, + 0.0019395713461562991, + 0.0019056655000895262, + 0.05397011339664459, + 0.053670063614845276, + 0.005700996145606041, + 0.011653607711195946, + 0.004935057368129492, + 0.00501555809751153, + 0.003608459373936057, + 0.0033928982447832823, + 0.005430137738585472, + 0.003903340781107545, + 0.006666925735771656, + 0.006998224649578333, + 0.002510837744921446, + 0.003014948684722185, + 0.0010794466361403465, + 0.0021779825910925865, + 0.002540145069360733, + 0.002194153144955635, + 0.0028168032877147198, + 0.002408510074019432, + 0.004855339881032705, + 0.005006942432373762, + 0.0021496256813406944, + 0.004911323543637991, + 0.004554969258606434, + 0.0048247757367789745, + 0.00562428729608655, + 0.005619805306196213, + 0.017034374177455902, + 0.014462808147072792, + 0.008555003441870213, + 0.009653989225625992, + 0.007963606156408787, + 0.013204392045736313, + 0.015589085407555103, + 0.022846434265375137, + 0.003675460582599044, + 0.004449056927114725, + 0.0014507065061479807, + 0.002600830513983965, + 0.003572369460016489, + 0.004558464977890253, + 0.00306174554862082, + 0.0036132093518972397, + 0.0008653191616758704, + 0.0009469446958974004, + 0.0007283056038431823, + 0.0009084264165721834, + 0.0011266024084761739, + 0.0010722976876422763, + 0.0010791015811264515, + 0.0014891024911776185, + 0.029698777943849564, + 0.17331813275814056, + 0.0038602317217737436, + 0.00779613945633173, + 0.019608525559306145, + 0.1695268154144287, + 0.006241132039576769, + 0.021527862176299095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 1 + ], + "selected_reward": -0.014199413359165192 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide shot slowly pans right, following a narrow, wooden boat as it glides silently and steadily across a mirror-like river. A thick, milky morning mist shrouds the scene, completely swallowing up the distant riverbank and any possible scenery, creating a sense of secluded tranquility. The first rays of morning sunlight struggle to penetrate the thick fog, casting a wide swath of warm, saturated golden light on the water, contrasting sharply with the soft blue tones permeating the mist, and imbuing the image with Steve McCurry's signature rich color style. The boat is made of weathered, dark wood, the scratches and grain of its hull clearly visible in the sidelight. The bow of the boat steadily breaks the surface of the water, leaving behind soft, diffuse ripples. On the boat stands a solitary figure in simple clothes and a conical hat. His back appears both small and powerful in the interweaving of morning light and mist. He paddles the oars rhythmically, each entry into the water and each lift off with precision and composure, as if he has become one with this ancient river. The whole scene is filled with a tranquil, mysterious and almost dreamy atmosphere, like a vivid painting frozen in time.", + "seed": 2845, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_13/index_2501_seed_2845_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02995173865929246, + "vlm_logprob": -0.02995173865929246, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.02995173865929246, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a narrow wooden boat made of weathered dark wood with visible scratches and grain gliding across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure wearing simple clothes and a conical hat standing in the boat and paddling rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a river shrouded in thick milky morning mist with soft blue tones and warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the boat's bow break the water surface creating soft diffuse ripples while the distant riverbank is completely swallowed by mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07596596330404282 + ], + "opposite_logprob_tokens": [ + -2.8259658813476562 + ], + "logprob_reward": -0.07596596330404282, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07596596330404282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05424577370285988, + 0.023218289017677307, + 0.04830343648791313, + 0.04547909274697304, + 0.3254377543926239, + 0.24440887570381165, + 0.03904150426387787, + 0.13359816372394562, + 0.06904962658882141, + 0.017217522487044334 + ], + "credit_map": [ + 0.0062813470140099525, + 0.011319051496684551, + 0.0055300770327448845, + 0.005958642344921827, + 0.005857436452060938, + 0.004573230165988207, + 0.008426510728895664, + 0.006299478001892567, + 0.0019500203197821975, + 0.004239629488438368, + 0.0016924738883972168, + 0.0024620587937533855, + 0.00332300434820354, + 0.004108697175979614, + 0.0029903023969382048, + 0.0024521027226001024, + 0.002127649262547493, + 0.0035527157597243786, + 0.0016716712852939963, + 0.0024959915317595005, + 0.017334388568997383, + 0.016030915081501007, + 0.0019819890148937702, + 0.0031081177294254303, + 0.003437167964875698, + 0.007625017315149307, + 0.00284966710023582, + 0.004953733179718256, + 0.009934298694133759, + 0.007757438812404871, + 0.005209301598370075, + 0.0037124650552868843, + 0.013199752196669579, + 0.011239679530262947, + 0.00431046774610877, + 0.006319625303149223, + 0.14047649502754211, + 0.12982751429080963, + 0.006154326256364584, + 0.01390990149229765, + 0.005808638408780098, + 0.003321701427921653, + 0.004410092253237963, + 0.005811919458210468, + 0.05723923444747925, + 0.10653471201658249, + 0.007465867325663567, + 0.05381670594215393, + 0.0017407214036211371, + 0.003071494633331895, + 0.0016908752731978893, + 0.002854461781680584, + 0.0038095370400696993, + 0.012276966124773026, + 0.0024335230700671673, + 0.011163927614688873, + 0.016879430040717125, + 0.023505037650465965, + 0.009020380675792694, + 0.016439195722341537, + 0.020902907475829124, + 0.02066154032945633, + 0.011455642990767956, + 0.014734040014445782, + 0.010554089210927486, + 0.011823657900094986, + 0.004983643535524607, + 0.007658892776817083, + 0.010291439481079578, + 0.00886316504329443, + 0.007938257418572903, + 0.0069364807568490505, + 0.002075847703963518, + 0.002381908241659403, + 0.0014792378060519695, + 0.0018320555100217462, + 0.003326586214825511, + 0.0019980149809271097, + 0.0024765946436673403, + 0.0016472777351737022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -0.07596596330404282 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a narrow wooden boat made of weathered dark wood with visible scratches and grain gliding across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure wearing simple clothes and a conical hat standing in the boat and paddling rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a river shrouded in thick milky morning mist with soft blue tones and warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the boat's bow break the water surface creating soft diffuse ripples while the distant riverbank is completely swallowed by mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02982296794652939 + ], + "opposite_logprob_tokens": [ + -3.529823064804077 + ], + "logprob_reward": -0.02982296794652939, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982296794652939, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06732383370399475, + 0.03081907145678997, + 0.5196582078933716, + 0.014774727635085583, + 0.29294049739837646, + 0.017398551106452942, + 0.010298443026840687, + 0.02720179595053196, + 0.00919651985168457, + 0.010388338938355446 + ], + "credit_map": [ + 0.0059033045545220375, + 0.01321307010948658, + 0.007110223174095154, + 0.006849967408925295, + 0.008804629556834698, + 0.006027558818459511, + 0.012046393007040024, + 0.0073686749674379826, + 0.002003737725317478, + 0.0033087425399571657, + 0.001450098818168044, + 0.002043595537543297, + 0.004375610500574112, + 0.009402371011674404, + 0.0028135862667113543, + 0.005421322770416737, + 0.0224534310400486, + 0.013306207954883575, + 0.006122265011072159, + 0.011412439867854118, + 0.23670931160449982, + 0.20785194635391235, + 0.006931855343282223, + 0.014870679005980492, + 0.0015007067704573274, + 0.0021644202060997486, + 0.0013320784782990813, + 0.0015029434580355883, + 0.002811738522723317, + 0.0024750845041126013, + 0.0019012921256944537, + 0.0010864614741876721, + 0.010766865685582161, + 0.008689099922776222, + 0.004149150103330612, + 0.005242018960416317, + 0.12470908463001251, + 0.11937924474477768, + 0.006949722766876221, + 0.01305531244724989, + 0.001288463594391942, + 0.0012153979623690248, + 0.0010805964702740312, + 0.0013063587248325348, + 0.005952438805252314, + 0.002046021167188883, + 0.0014681307366117835, + 0.003041140967980027, + 0.000550776778254658, + 0.000915228039957583, + 0.0006395447417162359, + 0.0010357233695685863, + 0.001033745356835425, + 0.0025326618924736977, + 0.0010059941560029984, + 0.002584767760708928, + 0.002968768123537302, + 0.004147535189986229, + 0.0015240763314068317, + 0.002473278669640422, + 0.0051872618496418, + 0.005222965031862259, + 0.0021646658424288034, + 0.003513241419568658, + 0.0013008646201342344, + 0.0013358184369280934, + 0.0011684171622619033, + 0.0012937237042933702, + 0.0010228876490145922, + 0.0012903790920972824, + 0.0009461236768402159, + 0.0008383046369999647, + 0.001778460224159062, + 0.0018307664431631565, + 0.0008060886757448316, + 0.0009892114903777838, + 0.0014588331105187535, + 0.0013412816915661097, + 0.001388226286508143, + 0.0007954707252793014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.02982296794652939 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a narrow wooden boat made of weathered dark wood with visible scratches and grain gliding across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure wearing simple clothes and a conical hat standing in the boat and paddling rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a river shrouded in thick milky morning mist with soft blue tones and warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the boat's bow break the water surface creating soft diffuse ripples while the distant riverbank is completely swallowed by mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00249301316216588 + ], + "opposite_logprob_tokens": [ + -6.002492904663086 + ], + "logprob_reward": -0.00249301316216588, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.00249301316216588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05798738822340965, + 0.020513145253062248, + 0.07750041782855988, + 0.03011455200612545, + 0.3232938349246979, + 0.27015480399131775, + 0.03442346677184105, + 0.14249573647975922, + 0.024163464084267616, + 0.01935325376689434 + ], + "credit_map": [ + 0.00745440274477005, + 0.011814024299383163, + 0.004516137298196554, + 0.0055860551074147224, + 0.007467553950846195, + 0.0073594930581748486, + 0.008052029646933079, + 0.005737693049013615, + 0.0021580997854471207, + 0.002899029292166233, + 0.0015001272549852729, + 0.001794883399270475, + 0.003301484975963831, + 0.0031024226918816566, + 0.003730200231075287, + 0.0020268980879336596, + 0.004056423902511597, + 0.004272095859050751, + 0.0020798577461391687, + 0.0029849912971258163, + 0.031531307846307755, + 0.02758430317044258, + 0.0020673542749136686, + 0.0029240914154797792, + 0.002376151503995061, + 0.0029979227110743523, + 0.0018374179489910603, + 0.0023493950720876455, + 0.008806364610791206, + 0.007327449508011341, + 0.0026822141371667385, + 0.0017376369796693325, + 0.011662681587040424, + 0.006857984699308872, + 0.0039806487038731575, + 0.0048708198592066765, + 0.14355015754699707, + 0.13199272751808167, + 0.006705717649310827, + 0.013673104345798492, + 0.010715652257204056, + 0.006845436058938503, + 0.003166748443618417, + 0.005265298765152693, + 0.12113653123378754, + 0.09431927651166916, + 0.006454253103584051, + 0.02225162833929062, + 0.0012812354834750295, + 0.0019149149302393198, + 0.0012581561459228396, + 0.001834385795518756, + 0.003412386402487755, + 0.012006308883428574, + 0.002217730972915888, + 0.010498348623514175, + 0.016433479264378548, + 0.024180637672543526, + 0.010326996445655823, + 0.018130654469132423, + 0.02056535892188549, + 0.022324243560433388, + 0.013278909958899021, + 0.017255457118153572, + 0.0033975671976804733, + 0.003977363463491201, + 0.0017957087839022279, + 0.0027051211800426245, + 0.0037509857211261988, + 0.003359610680490732, + 0.003045470453798771, + 0.00213163741864264, + 0.002824717201292515, + 0.002932050498202443, + 0.0015861738938838243, + 0.002385130152106285, + 0.0024952543899416924, + 0.0023748367093503475, + 0.0028844792395830154, + 0.0018706105183809996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -0.00249301316216588 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a narrow wooden boat made of weathered dark wood with visible scratches and grain gliding across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure wearing simple clothes and a conical hat standing in the boat and paddling rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a river shrouded in thick milky morning mist with soft blue tones and warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the boat's bow break the water surface creating soft diffuse ripples while the distant riverbank is completely swallowed by mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018181920051574707 + ], + "opposite_logprob_tokens": [ + -4.018181800842285 + ], + "logprob_reward": -0.018181920051574707, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018181920051574707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05313762649893761, + 0.07059124857187271, + 0.12998715043067932, + 0.05922233313322067, + 0.037622176110744476, + 0.5468111634254456, + 0.013763642869889736, + 0.04670466110110283, + 0.014342625625431538, + 0.027817314490675926 + ], + "credit_map": [ + 0.007393339183181524, + 0.007130438927561045, + 0.004709064960479736, + 0.0045633334666490555, + 0.008793147280812263, + 0.008468851447105408, + 0.006440113298594952, + 0.00563934026286006, + 0.0028630392625927925, + 0.003415348706766963, + 0.003142808098345995, + 0.0036311601288616657, + 0.011234475299715996, + 0.02520887553691864, + 0.005025086458772421, + 0.01607045903801918, + 0.005074618384242058, + 0.00305360765196383, + 0.0017891281750053167, + 0.003423125948756933, + 0.05888434499502182, + 0.0526144877076149, + 0.0018582467455416918, + 0.0032895966432988644, + 0.0050332131795585155, + 0.004053197335451841, + 0.003493148135021329, + 0.003716805251315236, + 0.018912354484200478, + 0.016171835362911224, + 0.004580265376716852, + 0.003261517034843564, + 0.001703709247522056, + 0.0019108218839392066, + 0.00123310973867774, + 0.0015726737910881639, + 0.014609684236347675, + 0.012448063120245934, + 0.0018838291289284825, + 0.002260291250422597, + 0.01473455410450697, + 0.010577156208455563, + 0.007280748803168535, + 0.011331971734762192, + 0.21339593827724457, + 0.23112080991268158, + 0.012530801817774773, + 0.04583924263715744, + 0.0008016173378564417, + 0.0020894119516015053, + 0.0008261218899860978, + 0.0016426887596026063, + 0.0021725615952163935, + 0.0024366818834096193, + 0.0021804370917379856, + 0.0016141236992552876, + 0.005900170188397169, + 0.008176417089998722, + 0.0027569907251745462, + 0.004528541583567858, + 0.007659721653908491, + 0.0088299959897995, + 0.003595174988731742, + 0.005257651675492525, + 0.0022173267789185047, + 0.0023289958480745554, + 0.0012648662086576223, + 0.0017497902736067772, + 0.0020931395702064037, + 0.0018160839099436998, + 0.0015932099195197225, + 0.001279214397072792, + 0.004945863503962755, + 0.004349097143858671, + 0.0017297384329140186, + 0.002505996497347951, + 0.002716885646805167, + 0.004294849466532469, + 0.004122511483728886, + 0.003152374643832445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.018181920051574707 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a narrow wooden boat made of weathered dark wood with visible scratches and grain gliding across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure wearing simple clothes and a conical hat standing in the boat and paddling rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a river shrouded in thick milky morning mist with soft blue tones and warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the boat's bow break the water surface creating soft diffuse ripples while the distant riverbank is completely swallowed by mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023294828832149506 + ], + "opposite_logprob_tokens": [ + -3.7732949256896973 + ], + "logprob_reward": -0.023294828832149506, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023294828832149506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037945546209812164, + 0.06800640374422073, + 0.24604909121990204, + 0.043850675225257874, + 0.31289201974868774, + 0.18814095854759216, + 0.013947742991149426, + 0.020981352776288986, + 0.035430263727903366, + 0.03275595232844353 + ], + "credit_map": [ + 0.006586096249520779, + 0.004996879957616329, + 0.0040880474261939526, + 0.003964411560446024, + 0.004892898723483086, + 0.003978325519710779, + 0.006248334888368845, + 0.003190551185980439, + 0.002912437077611685, + 0.0041170804761350155, + 0.0027112006209790707, + 0.0034020196180790663, + 0.009617808274924755, + 0.023141633719205856, + 0.005793207790702581, + 0.01631101965904236, + 0.011379480361938477, + 0.007373997010290623, + 0.003160023596137762, + 0.005408468656241894, + 0.11191336810588837, + 0.09656428545713425, + 0.0036884108558297157, + 0.0065610697492957115, + 0.004714669659733772, + 0.0044787111692130566, + 0.0037211524322628975, + 0.004527909215539694, + 0.007921688258647919, + 0.007253698538988829, + 0.007089134305715561, + 0.004143708385527134, + 0.01189141720533371, + 0.006492449901998043, + 0.00378628633916378, + 0.004577856045216322, + 0.1393287032842636, + 0.12698586285114288, + 0.00612842570990324, + 0.013701020739972591, + 0.009772295132279396, + 0.0060265460051596165, + 0.0029120224062353373, + 0.00394637743011117, + 0.082818903028965, + 0.038172684609889984, + 0.005977666936814785, + 0.03851446509361267, + 0.0006347923190332949, + 0.001472372212447226, + 0.0006972529226914048, + 0.0013310880167409778, + 0.0012757362565025687, + 0.0036299382336437702, + 0.0011106693418696523, + 0.003795894328504801, + 0.0014737085439264774, + 0.002538044238463044, + 0.0016277770046144724, + 0.0027770872693508863, + 0.0023544293362647295, + 0.004091052338480949, + 0.002419077791273594, + 0.0037001764867454767, + 0.004523419309407473, + 0.005282009486109018, + 0.002517808461561799, + 0.003780757775530219, + 0.006935417652130127, + 0.00506161991506815, + 0.004145254846662283, + 0.003183978144079447, + 0.005101208109408617, + 0.00449707917869091, + 0.0029275959823280573, + 0.00356281572021544, + 0.004690251313149929, + 0.004555368330329657, + 0.004080071114003658, + 0.0033415614161640406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 5 + ], + "selected_reward": -0.023294828832149506 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide shot slowly pans right, following a narrow, wooden boat as it glides silently and steadily across a mirror-like river. A thick, milky morning mist shrouds the scene, completely swallowing up the distant riverbank and any possible scenery, creating a sense of secluded tranquility. The first rays of morning sunlight struggle to penetrate the thick fog, casting a wide swath of warm, saturated golden light on the water, contrasting sharply with the soft blue tones permeating the mist, and imbuing the image with Steve McCurry's signature rich color style. The boat is made of weathered, dark wood, the scratches and grain of its hull clearly visible in the sidelight. The bow of the boat steadily breaks the surface of the water, leaving behind soft, diffuse ripples. On the boat stands a solitary figure in simple clothes and a conical hat. His back appears both small and powerful in the interweaving of morning light and mist. He paddles the oars rhythmically, each entry into the water and each lift off with precision and composure, as if he has become one with this ancient river. The whole scene is filled with a tranquil, mysterious and almost dreamy atmosphere, like a vivid painting frozen in time.", + "seed": 2846, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_13/index_2501_seed_2846_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.019017823832109572, + "vlm_logprob": -0.019017823832109572, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.019017823832109572, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a narrow wooden boat made of weathered dark wood with visible scratches and grain gliding across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure wearing simple clothes and a conical hat standing in the boat and paddling rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a river shrouded in thick milky morning mist with soft blue tones and warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the boat's bow break the water surface creating soft diffuse ripples while the distant riverbank is completely swallowed by mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03737007454037666 + ], + "opposite_logprob_tokens": [ + -3.537369966506958 + ], + "logprob_reward": -0.03737007454037666, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03737007454037666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.012728163972496986, + 0.025322329252958298, + 0.04138806834816933, + 0.5128796696662903, + 0.01125310268253088, + 0.007402530871331692, + 0.0076563009060919285, + 0.31784239411354065, + 0.02555275149643421, + 0.03797464817762375 + ], + "credit_map": [ + 0.0011822215747088194, + 0.0015705181285738945, + 0.0012022466398775578, + 0.0015394710935652256, + 0.0022618749644607306, + 0.0020914385095238686, + 0.0014513444621115923, + 0.0014290499966591597, + 0.0015343830455094576, + 0.0018328260630369186, + 0.004429628141224384, + 0.004147971048951149, + 0.0034450837410986423, + 0.00466106366366148, + 0.0033463151194155216, + 0.0019250609911978245, + 0.0024099370930343866, + 0.0015845589805394411, + 0.013029905967414379, + 0.0016760174185037613, + 0.004465110599994659, + 0.015489989891648293, + 0.0012737958459183574, + 0.0014587555779144168, + 0.027300382032990456, + 0.020929669961333275, + 0.1723947823047638, + 0.015774328261613846, + 0.04555042088031769, + 0.20196284353733063, + 0.010447445325553417, + 0.018519865348935127, + 0.000921902246773243, + 0.0015889600617811084, + 0.001441210275515914, + 0.0013986743288114667, + 0.0016558999195694923, + 0.002091564703732729, + 0.0010328555945307016, + 0.0011220356682315469, + 0.0008409239235334098, + 0.0016610486200079322, + 0.0007327796192839742, + 0.0006217649206519127, + 0.0013077195035293698, + 0.001109735807403922, + 0.0006809886544942856, + 0.0004475699970498681, + 0.0009237039485014975, + 0.001505750697106123, + 0.0007831506081856787, + 0.0008596716215834022, + 0.001325764344073832, + 0.0010469892295077443, + 0.0005967905162833631, + 0.0006144812214188278, + 0.02850516326725483, + 0.13062138855457306, + 0.006512416992336512, + 0.0034409326035529375, + 0.015273713506758213, + 0.11861361563205719, + 0.004054712597280741, + 0.010820473544299603, + 0.00282464106567204, + 0.010145461186766624, + 0.0008964992011897266, + 0.000811417936347425, + 0.002915418241173029, + 0.006069837603718042, + 0.0008382158703170717, + 0.0010512620210647583, + 0.005026878789067268, + 0.009535351768136024, + 0.0035864782985299826, + 0.0034188306890428066, + 0.005014085676521063, + 0.005187473259866238, + 0.0038044494576752186, + 0.0024011065252125263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 2 + ], + "selected_reward": -0.03737007454037666 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a narrow wooden boat made of weathered dark wood with visible scratches and grain gliding across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure wearing simple clothes and a conical hat standing in the boat and paddling rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a river shrouded in thick milky morning mist with soft blue tones and warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the boat's bow break the water surface creating soft diffuse ripples while the distant riverbank is completely swallowed by mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033716656267642975 + ], + "opposite_logprob_tokens": [ + -3.4087166786193848 + ], + "logprob_reward": -0.033716656267642975, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033716656267642975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09614403545856476, + 0.2223483771085739, + 0.19321106374263763, + 0.26737457513809204, + 0.04487358406186104, + 0.021178165450692177, + 0.020536448806524277, + 0.06024743244051933, + 0.01906309463083744, + 0.05502323433756828 + ], + "credit_map": [ + 0.006891796365380287, + 0.008987348526716232, + 0.008987868204712868, + 0.012427866458892822, + 0.016656894236803055, + 0.021036989986896515, + 0.01219703909009695, + 0.008958229795098305, + 0.013180882669985294, + 0.015333453193306923, + 0.041271813213825226, + 0.034473832696676254, + 0.02970588020980358, + 0.04617559537291527, + 0.025615699589252472, + 0.01659121736884117, + 0.009792923927307129, + 0.007357156369835138, + 0.05558677017688751, + 0.009869592264294624, + 0.01985245943069458, + 0.07405075430870056, + 0.008439745754003525, + 0.008261664770543575, + 0.01482281181961298, + 0.010544808581471443, + 0.09116408973932266, + 0.008004669100046158, + 0.02235424518585205, + 0.10521329194307327, + 0.0050936536863446236, + 0.01017700880765915, + 0.004097470082342625, + 0.005061521660536528, + 0.005227543413639069, + 0.005081124138087034, + 0.00633979681879282, + 0.006266143172979355, + 0.007763642352074385, + 0.0050363396294415, + 0.002201635157689452, + 0.0029698056168854237, + 0.0022415455896407366, + 0.0026795079465955496, + 0.003526432439684868, + 0.002591150114312768, + 0.002969245193526149, + 0.001998842926695943, + 0.002526117954403162, + 0.0034136404283344746, + 0.0019635106436908245, + 0.002788343233987689, + 0.002990084234625101, + 0.002646834123879671, + 0.0023355488665401936, + 0.0018723701359704137, + 0.00631395447999239, + 0.02255255915224552, + 0.0021421078126877546, + 0.002047592541202903, + 0.006575249135494232, + 0.016068316996097565, + 0.002358237747102976, + 0.0021894166711717844, + 0.0017889041919261217, + 0.0042770905420184135, + 0.0015649895649403334, + 0.0013996497727930546, + 0.004411350004374981, + 0.0019765570759773254, + 0.0023535562213510275, + 0.0012909964425489306, + 0.007846934720873833, + 0.013170421123504639, + 0.0047209616750478745, + 0.005231368355453014, + 0.0065582869574427605, + 0.006968105211853981, + 0.006802922114729881, + 0.0037242264952510595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.033716656267642975 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a narrow wooden boat made of weathered dark wood with visible scratches and grain gliding across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure wearing simple clothes and a conical hat standing in the boat and paddling rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a river shrouded in thick milky morning mist with soft blue tones and warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the boat's bow break the water surface creating soft diffuse ripples while the distant riverbank is completely swallowed by mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012546890415251255 + ], + "opposite_logprob_tokens": [ + -4.387547016143799 + ], + "logprob_reward": -0.012546890415251255, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012546890415251255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12560218572616577, + 0.04311946779489517, + 0.16037802398204803, + 0.13343489170074463, + 0.01899428479373455, + 0.020650546997785568, + 0.022359348833560944, + 0.43238404393196106, + 0.028530383482575417, + 0.014546819031238556 + ], + "credit_map": [ + 0.008285153657197952, + 0.010100890882313251, + 0.011828665621578693, + 0.014881031587719917, + 0.03408876433968544, + 0.020484626293182373, + 0.015324704349040985, + 0.010608353652060032, + 0.0025023561902344227, + 0.0033707250840961933, + 0.006014181300997734, + 0.008497490547597408, + 0.005271789152175188, + 0.007026254199445248, + 0.006247865501791239, + 0.004188804887235165, + 0.008151405490934849, + 0.006244512740522623, + 0.048702072352170944, + 0.006589188240468502, + 0.01572326570749283, + 0.06287642568349838, + 0.0054790908470749855, + 0.006612056866288185, + 0.007399995811283588, + 0.006071317009627819, + 0.043489743024110794, + 0.004705161787569523, + 0.01183257158845663, + 0.051786262542009354, + 0.0029953254852443933, + 0.0051545132882893085, + 0.001077247317880392, + 0.0018706078408285975, + 0.0021604965440928936, + 0.0022457020822912455, + 0.002903706394135952, + 0.003562774509191513, + 0.0026680512819439173, + 0.0025056980084627867, + 0.0023271734826266766, + 0.004699100274592638, + 0.0025091723073273897, + 0.001693658996373415, + 0.0034799922723323107, + 0.0029218823183327913, + 0.0017010747687891126, + 0.0013184914132580161, + 0.0016560995718464255, + 0.0034366154577583075, + 0.002152826404199004, + 0.0027948685456067324, + 0.004696955904364586, + 0.0031491334084421396, + 0.002095116302371025, + 0.0023777317255735397, + 0.030852902680635452, + 0.14672355353832245, + 0.010951920412480831, + 0.012122535146772861, + 0.03524775803089142, + 0.16456839442253113, + 0.012370782904326916, + 0.019546182826161385, + 0.0037715639919042587, + 0.004235355649143457, + 0.0023678995203226805, + 0.002442366210743785, + 0.005224539432674646, + 0.0045811086893081665, + 0.0031511744018644094, + 0.0027563723269850016, + 0.0016494536539539695, + 0.002141582081094384, + 0.001587194507010281, + 0.0014269797829911113, + 0.0017478218069300056, + 0.0026030056178569794, + 0.0018913508392870426, + 0.001499428879469633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 3 + ], + "selected_reward": -0.012546890415251255 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a narrow wooden boat made of weathered dark wood with visible scratches and grain gliding across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure wearing simple clothes and a conical hat standing in the boat and paddling rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a river shrouded in thick milky morning mist with soft blue tones and warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the boat's bow break the water surface creating soft diffuse ripples while the distant riverbank is completely swallowed by mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028148575220257044 + ], + "opposite_logprob_tokens": [ + -5.877814769744873 + ], + "logprob_reward": -0.0028148575220257044, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028148575220257044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14346173405647278, + 0.05168278515338898, + 0.20041248202323914, + 0.16343814134597778, + 0.03239059075713158, + 0.047555457800626755, + 0.05281681567430496, + 0.23235906660556793, + 0.04970645532011986, + 0.026176543906331062 + ], + "credit_map": [ + 0.010691655799746513, + 0.012286313809454441, + 0.017394497990608215, + 0.015905063599348068, + 0.016225691884756088, + 0.02667991630733013, + 0.01401741523295641, + 0.030261171981692314, + 0.003252902301028371, + 0.004723441321402788, + 0.0075521282851696014, + 0.006689183413982391, + 0.006592971738427877, + 0.01437427382916212, + 0.005323324352502823, + 0.0031745596788823605, + 0.01038342621177435, + 0.007333760615438223, + 0.06057129055261612, + 0.010350670665502548, + 0.02034931257367134, + 0.07578098028898239, + 0.007175978738814592, + 0.00846706610172987, + 0.008625084534287453, + 0.009267617017030716, + 0.045278485864400864, + 0.009759264066815376, + 0.01918594352900982, + 0.056055549532175064, + 0.007411210332065821, + 0.007854990661144257, + 0.003674904815852642, + 0.00535217160359025, + 0.005768698174506426, + 0.002452368149533868, + 0.004299550782889128, + 0.004674229305237532, + 0.003810509108006954, + 0.0023581604473292828, + 0.005033180117607117, + 0.005731768906116486, + 0.010025721043348312, + 0.004166286904364824, + 0.0083793755620718, + 0.006744229234755039, + 0.004458756186068058, + 0.0030161382164806128, + 0.005202631000429392, + 0.010394509881734848, + 0.0068277777172625065, + 0.004527711775153875, + 0.011589528061449528, + 0.0074439384043216705, + 0.0034917471930384636, + 0.0033389702439308167, + 0.017213137820363045, + 0.08617228269577026, + 0.006446899846196175, + 0.0042228372767567635, + 0.01558079943060875, + 0.08957230299711227, + 0.004226146265864372, + 0.008924675174057484, + 0.004673410207033157, + 0.015720469877123833, + 0.0023544845171272755, + 0.002361687831580639, + 0.006423288024961948, + 0.011743878945708275, + 0.0035635060630738735, + 0.0028657319489866495, + 0.0026390627026557922, + 0.0035173522774130106, + 0.0029602793511003256, + 0.0025751786306500435, + 0.003592838067561388, + 0.0047250776551663876, + 0.00384388561360538, + 0.002322869375348091 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 3 + ], + "selected_reward": -0.0028148575220257044 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a narrow wooden boat made of weathered dark wood with visible scratches and grain gliding across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure wearing simple clothes and a conical hat standing in the boat and paddling rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a river shrouded in thick milky morning mist with soft blue tones and warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the boat's bow break the water surface creating soft diffuse ripples while the distant riverbank is completely swallowed by mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640640415251255 + ], + "opposite_logprob_tokens": [ + -4.758640766143799 + ], + "logprob_reward": -0.008640640415251255, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640640415251255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06752185523509979, + 0.042999058961868286, + 0.3027336001396179, + 0.11187245696783066, + 0.014681397005915642, + 0.018312225118279457, + 0.0076189530082046986, + 0.3740207254886627, + 0.03150556981563568, + 0.028734132647514343 + ], + "credit_map": [ + 0.00394652783870697, + 0.005893700290471315, + 0.007556449621915817, + 0.010105260647833347, + 0.012176735326647758, + 0.011539943516254425, + 0.008712289854884148, + 0.007590945810079575, + 0.002686835592612624, + 0.0027887336909770966, + 0.007313121110200882, + 0.006210134364664555, + 0.006191359832882881, + 0.010017592459917068, + 0.0043585095554590225, + 0.0034327704925090075, + 0.015648014843463898, + 0.011536496691405773, + 0.09234903752803802, + 0.011778791435062885, + 0.030202919617295265, + 0.1197667121887207, + 0.009816787205636501, + 0.011634829454123974, + 0.005772358272224665, + 0.00495501421391964, + 0.03610506281256676, + 0.004228705540299416, + 0.010077360086143017, + 0.04190458357334137, + 0.004199949558824301, + 0.0046294135972857475, + 0.0014430530136451125, + 0.0020489608868956566, + 0.001959498506039381, + 0.0016719497507438064, + 0.0023699456360191107, + 0.0021993147674947977, + 0.0018521070014685392, + 0.001136565930210054, + 0.0020841334480792284, + 0.003908527549356222, + 0.002769643208011985, + 0.0015773775521665812, + 0.0028415978886187077, + 0.0025625487323850393, + 0.0014469630550593138, + 0.001121432171203196, + 0.0006442927406169474, + 0.0009818302933126688, + 0.0009032537927851081, + 0.0009291591704823077, + 0.0017490405589342117, + 0.0010004978394135833, + 0.00060425722040236, + 0.0008066209265962243, + 0.03116677515208721, + 0.14911064505577087, + 0.007516966201364994, + 0.0038391754496842623, + 0.01940571703016758, + 0.14474692940711975, + 0.005161339417099953, + 0.013073155656456947, + 0.003322546137496829, + 0.01263287290930748, + 0.0009074722765944898, + 0.0009350114269182086, + 0.0031708304304629564, + 0.008247745223343372, + 0.0010401587933301926, + 0.001248931628651917, + 0.004318561404943466, + 0.008388325572013855, + 0.0018816388910636306, + 0.001771086361259222, + 0.003708155360072851, + 0.004687262699007988, + 0.0022717283573001623, + 0.0017073751660063863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 3 + ], + "selected_reward": -0.008640640415251255 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_14/index_tensor([7371, 7371])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_14/index_tensor([7371, 7371])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1619778e1a83b02b11fb1d066f86a6063bc74f73 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_14/index_tensor([7371, 7371])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "series of rapid, sharp cuts captures the moment of intense bodily reaction from multiple angles, beginning with a dramatic low-angle shot that emphasizes the intense, unconscious trembling throughout the subject's frame. Then, the perspective abruptly shifts to extreme close-up, focusing entirely on a magnificent, perfectly spherical Afro, its thick, tightly coiled black curls with a soft, velvety texture filling the screen. Against a stark, dark charcoal-gray backdrop, the trembling triggers a dramatic tremor in the hair; a series of micro-vibrations ripples through the mass, making each individual curl appear to dance and vibrate independently. The scene is illuminated by dramatic, high-contrast black and white lighting from a single, harsh, off-camera key light, which sculpts the subject's unseen features in light and shadow while capturing the edges of the trembling curls, creating a shimmering, dynamic halo effect against the profound darkness. The overall atmosphere is intense, raw, and emotionally charged, completely focused on the pure, dynamic expression of powerful, unseen excitement.", + "seed": 5220, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_14/index_7371_seed_5220_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1820066686719656, + "vlm_logprob": -0.1820066686719656, + "vlm_yes_no_margin": 2.25, + "vlm_token_credit": -0.1820066686719656, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video consist of a series of rapid, sharp cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's hair a perfectly spherical Afro with thick, tightly coiled black curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a stark, dark charcoal-gray color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by high-contrast black and white lighting from a single, harsh off-camera key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include an extreme close-up shot focusing entirely on the hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.579658031463623 + ], + "opposite_logprob_tokens": [ + -0.829658031463623 + ], + "logprob_reward": -0.579658031463623, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.579658031463623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048522334545850754, + 0.08911047875881195, + 0.05557041987776756, + 0.11457470804452896, + 0.10543370246887207, + 0.09357825666666031, + 0.03971008583903313, + 0.022898197174072266, + 0.36484962701797485, + 0.06575223058462143 + ], + "credit_map": [ + 0.006009189877659082, + 0.006294772494584322, + 0.007583738770335913, + 0.006060122046619654, + 0.005170953925698996, + 0.005478185135871172, + 0.006120803765952587, + 0.005804568994790316, + 0.01879606395959854, + 0.01731458678841591, + 0.01223925780504942, + 0.01258418895304203, + 0.005888368468731642, + 0.007574062794446945, + 0.0070935459807515144, + 0.007620407268404961, + 0.0064915865659713745, + 0.007128723431378603, + 0.008603563532233238, + 0.006027352064847946, + 0.00796720664948225, + 0.0052765184082090855, + 0.0059118373319506645, + 0.0081636318936944, + 0.008347220718860626, + 0.0173631738871336, + 0.012688687071204185, + 0.010313744656741619, + 0.01997506245970726, + 0.022785626351833344, + 0.015373828820884228, + 0.007727370131760836, + 0.007587856613099575, + 0.0065467096865177155, + 0.007316840346902609, + 0.02185247093439102, + 0.007407048717141151, + 0.008439507335424423, + 0.028095008805394173, + 0.01818825863301754, + 0.014657218009233475, + 0.009312698617577553, + 0.010252779349684715, + 0.016774337738752365, + 0.006409989204257727, + 0.008837462402880192, + 0.015032738447189331, + 0.012301028706133366, + 0.00486854650080204, + 0.003976079635322094, + 0.005723199341446161, + 0.007463250309228897, + 0.003362649818882346, + 0.002461893716827035, + 0.008100833743810654, + 0.0037536306772381067, + 0.0029534040950238705, + 0.002827464137226343, + 0.0020557367242872715, + 0.003911591600626707, + 0.0019400474848225713, + 0.0022738028783351183, + 0.0044045718386769295, + 0.002531575970351696, + 0.007384041789919138, + 0.012814599089324474, + 0.008479331620037556, + 0.14703865349292755, + 0.009961938485503197, + 0.008655473589897156, + 0.15231473743915558, + 0.018200848251581192, + 0.004500329028815031, + 0.0034306966699659824, + 0.006180126685649157, + 0.022814497351646423, + 0.0034651034511625767, + 0.0032645631581544876, + 0.016543051227927208, + 0.005553859286010265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 4 + ], + "selected_reward": -0.579658031463623 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video consist of a series of rapid, sharp cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's hair a perfectly spherical Afro with thick, tightly coiled black curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a stark, dark charcoal-gray color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by high-contrast black and white lighting from a single, harsh off-camera key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include an extreme close-up shot focusing entirely on the hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17994143068790436 + ], + "opposite_logprob_tokens": [ + -1.8049414157867432 + ], + "logprob_reward": -0.17994143068790436, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17994143068790436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19840370118618011, + 0.052402496337890625, + 0.050037872046232224, + 0.05457025393843651, + 0.031034529209136963, + 0.18674838542938232, + 0.03587920218706131, + 0.020549241453409195, + 0.30696240067481995, + 0.06341194361448288 + ], + "credit_map": [ + 0.02452169358730316, + 0.024878736585378647, + 0.030113762244582176, + 0.031206484884023666, + 0.018882904201745987, + 0.022509705275297165, + 0.021159594878554344, + 0.02513081580400467, + 0.0073876213282346725, + 0.006854541599750519, + 0.009221995249390602, + 0.005490757990628481, + 0.0040518841706216335, + 0.004817449953407049, + 0.005921238102018833, + 0.008657006546854973, + 0.004692900460213423, + 0.005630018655210733, + 0.011340107768774033, + 0.006095046643167734, + 0.005245547275990248, + 0.004906638991087675, + 0.006513035390526056, + 0.0056145754642784595, + 0.0041609532199800014, + 0.00827599223703146, + 0.005714100785553455, + 0.005502704996615648, + 0.009989676997065544, + 0.010050932876765728, + 0.006576991640031338, + 0.004298901651054621, + 0.002886870177462697, + 0.0026606214232742786, + 0.001956637715920806, + 0.0062414733693003654, + 0.00266859820112586, + 0.0031171024311333895, + 0.0074558863416314125, + 0.004047338385134935, + 0.026910066604614258, + 0.0143049955368042, + 0.01935729756951332, + 0.03259909525513649, + 0.014658467844128609, + 0.021400712430477142, + 0.028947943821549416, + 0.02856980264186859, + 0.003907159902155399, + 0.003338370006531477, + 0.00468425964936614, + 0.006685857195407152, + 0.003156043589115143, + 0.0021419536788016558, + 0.008064702153205872, + 0.0039008574094623327, + 0.003174816956743598, + 0.0017059362726286054, + 0.002098566386848688, + 0.0035354364663362503, + 0.002077832818031311, + 0.0019435635767877102, + 0.0033604721538722515, + 0.0026526174042373896, + 0.006193225271999836, + 0.009763146750628948, + 0.006299044005572796, + 0.12719930708408356, + 0.009956078603863716, + 0.006337620783597231, + 0.1294931173324585, + 0.011720874346792698, + 0.003655376611277461, + 0.003536020405590534, + 0.005810835864394903, + 0.019306378439068794, + 0.00429012905806303, + 0.003329456551000476, + 0.01853572018444538, + 0.004948025103658438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.17994143068790436 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video consist of a series of rapid, sharp cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's hair a perfectly spherical Afro with thick, tightly coiled black curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a stark, dark charcoal-gray color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by high-contrast black and white lighting from a single, harsh off-camera key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include an extreme close-up shot focusing entirely on the hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809695690870285 + ], + "opposite_logprob_tokens": [ + -3.2880969047546387 + ], + "logprob_reward": -0.03809695690870285, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809695690870285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049479804933071136, + 0.27712351083755493, + 0.05380435660481453, + 0.11013796925544739, + 0.04071923345327377, + 0.33606401085853577, + 0.010154533199965954, + 0.031260933727025986, + 0.03340831398963928, + 0.05784733220934868 + ], + "credit_map": [ + 0.006810393184423447, + 0.006193201988935471, + 0.007265028543770313, + 0.006241959519684315, + 0.004175947979092598, + 0.00598130002617836, + 0.006501092575490475, + 0.006310883443802595, + 0.05510662868618965, + 0.04887670278549194, + 0.04423636570572853, + 0.03692537918686867, + 0.01883210428059101, + 0.022054733708500862, + 0.027493569999933243, + 0.02359805628657341, + 0.004649595823138952, + 0.006881038658320904, + 0.007256563287228346, + 0.006889017764478922, + 0.009899567812681198, + 0.00444754445925355, + 0.009878152050077915, + 0.0039028795436024666, + 0.008771166205406189, + 0.018031124025583267, + 0.012040827423334122, + 0.013091696426272392, + 0.016644984483718872, + 0.019808508455753326, + 0.012890765443444252, + 0.00885890331119299, + 0.003701248439028859, + 0.0021828964818269014, + 0.0023710059467703104, + 0.009036466479301453, + 0.003305901540443301, + 0.003710383316501975, + 0.010625761933624744, + 0.005785573273897171, + 0.05165882781147957, + 0.031122062355279922, + 0.03324766084551811, + 0.06276954710483551, + 0.02208414115011692, + 0.035845790058374405, + 0.05448059365153313, + 0.04485539346933365, + 0.0019174149492755532, + 0.0009098835871554911, + 0.0012116784928366542, + 0.001620988012291491, + 0.0008372912416234612, + 0.0008513693464919925, + 0.0017173818778246641, + 0.0010885267984122038, + 0.005542423576116562, + 0.002859641332179308, + 0.002182022901251912, + 0.004127707798033953, + 0.003366223769262433, + 0.002823334652930498, + 0.005302392411977053, + 0.0050571877509355545, + 0.003242869395762682, + 0.00465867854654789, + 0.0027311393059790134, + 0.00826252531260252, + 0.0025542154908180237, + 0.0016650038305670023, + 0.007824025116860867, + 0.00246985862031579, + 0.003063131356611848, + 0.0032493257895112038, + 0.00518683111295104, + 0.02136075869202614, + 0.0027412953786551952, + 0.0027941118460148573, + 0.01526323240250349, + 0.0041886488907039165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 3 + ], + "selected_reward": -0.03809695690870285 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video consist of a series of rapid, sharp cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's hair a perfectly spherical Afro with thick, tightly coiled black curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a stark, dark charcoal-gray color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by high-contrast black and white lighting from a single, harsh off-camera key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include an extreme close-up shot focusing entirely on the hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905211091041565 + ], + "opposite_logprob_tokens": [ + -2.464052200317383 + ], + "logprob_reward": -0.08905211091041565, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905211091041565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08575131744146347, + 0.14106491208076477, + 0.042858414351940155, + 0.08068642765283585, + 0.031161606311798096, + 0.33058100938796997, + 0.02125893160700798, + 0.023245634511113167, + 0.21940407156944275, + 0.023987704887986183 + ], + "credit_map": [ + 0.008875084109604359, + 0.010992737486958504, + 0.010657500475645065, + 0.01801462285220623, + 0.007983204908668995, + 0.009476550854742527, + 0.008432982489466667, + 0.011318634264171124, + 0.032237425446510315, + 0.029960468411445618, + 0.018313761800527573, + 0.019091816619038582, + 0.007907220162451267, + 0.010553872212767601, + 0.011863334104418755, + 0.011137007735669613, + 0.003745235037058592, + 0.00518527626991272, + 0.0061722127720713615, + 0.003921661525964737, + 0.007045510690659285, + 0.00488534290343523, + 0.0043420810252428055, + 0.007561092264950275, + 0.007714404724538326, + 0.012419195845723152, + 0.00886585097759962, + 0.009565739892423153, + 0.011537074111402035, + 0.012849518097937107, + 0.010729445144534111, + 0.00700519559904933, + 0.003243713639676571, + 0.002584073692560196, + 0.0020313397981226444, + 0.0062284753657877445, + 0.0020821127109229565, + 0.00331287644803524, + 0.007142179179936647, + 0.004536835942417383, + 0.04976929724216461, + 0.030102908611297607, + 0.03315461054444313, + 0.06189711019396782, + 0.021337289363145828, + 0.0345425046980381, + 0.05405235290527344, + 0.04572492092847824, + 0.002799968235194683, + 0.0023672915995121002, + 0.002838298911228776, + 0.003587722312659025, + 0.0015318236546590924, + 0.0013089061249047518, + 0.004477440379559994, + 0.002347478875890374, + 0.003934467677026987, + 0.0025527672842144966, + 0.0023219597060233355, + 0.003559265984222293, + 0.0021592210978269577, + 0.002224615076556802, + 0.004482358694076538, + 0.002010980388149619, + 0.005125812254846096, + 0.005053506698459387, + 0.004982392303645611, + 0.09470328688621521, + 0.007215321063995361, + 0.005755566991865635, + 0.0899980440735817, + 0.0065701426938176155, + 0.0013998787617310882, + 0.0008256718865595758, + 0.0021955498959869146, + 0.009054256603121758, + 0.0008638456929475069, + 0.0013590097660198808, + 0.0067480322904884815, + 0.0015414608642458916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 1 + ], + "selected_reward": -0.08905211091041565 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video consist of a series of rapid, sharp cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's hair a perfectly spherical Afro with thick, tightly coiled black curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a stark, dark charcoal-gray color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by high-contrast black and white lighting from a single, harsh off-camera key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include an extreme close-up shot focusing entirely on the hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328481338918209 + ], + "opposite_logprob_tokens": [ + -3.773284912109375 + ], + "logprob_reward": -0.02328481338918209, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328481338918209, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1690569519996643, + 0.0503888875246048, + 0.08648335188627243, + 0.07241769134998322, + 0.07330242544412613, + 0.18017105758190155, + 0.06387457251548767, + 0.01865880936384201, + 0.1614198088645935, + 0.12422646582126617 + ], + "credit_map": [ + 0.021196510642766953, + 0.022457530722022057, + 0.02552061155438423, + 0.023156853392720222, + 0.015454167500138283, + 0.02261519618332386, + 0.01634112372994423, + 0.022314950823783875, + 0.009018613956868649, + 0.008611928671598434, + 0.006650279741734266, + 0.006059094797819853, + 0.005896587856113911, + 0.003988164942711592, + 0.005232653114944696, + 0.004931563511490822, + 0.00968183483928442, + 0.011065972037613392, + 0.008831828832626343, + 0.007474524434655905, + 0.011217818595468998, + 0.009702229872345924, + 0.01697174459695816, + 0.011537407524883747, + 0.008112326264381409, + 0.013151385821402073, + 0.008717864751815796, + 0.006814020685851574, + 0.01049436442553997, + 0.009418661706149578, + 0.008738118223845959, + 0.006970943417400122, + 0.005688535049557686, + 0.005043032579123974, + 0.0067030442878603935, + 0.01461751852184534, + 0.00596863217651844, + 0.007095266133546829, + 0.016044126823544502, + 0.01214226707816124, + 0.029743777588009834, + 0.015622257255017757, + 0.020895376801490784, + 0.03159768506884575, + 0.01306214090436697, + 0.020114367827773094, + 0.028267379850149155, + 0.020868076011538506, + 0.011934911832213402, + 0.006713497918099165, + 0.008958258666098118, + 0.010828262194991112, + 0.005204785615205765, + 0.004799655172973871, + 0.009243197739124298, + 0.006192005705088377, + 0.0026474783662706614, + 0.001526951091364026, + 0.001887535210698843, + 0.004285145550966263, + 0.0012379945255815983, + 0.0013979815412312746, + 0.003830001223832369, + 0.0018457217374816537, + 0.008689414709806442, + 0.010044360533356667, + 0.007832574658095837, + 0.062071654945611954, + 0.005370303522795439, + 0.004997516982257366, + 0.05225026234984398, + 0.01016373373568058, + 0.011971277184784412, + 0.013261394575238228, + 0.0132406922057271, + 0.0237994734197855, + 0.008488142862915993, + 0.011584537103772163, + 0.02807091549038887, + 0.013810031116008759 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.02328481338918209 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "series of rapid, sharp cuts captures the moment of intense bodily reaction from multiple angles, beginning with a dramatic low-angle shot that emphasizes the intense, unconscious trembling throughout the subject's frame. Then, the perspective abruptly shifts to extreme close-up, focusing entirely on a magnificent, perfectly spherical Afro, its thick, tightly coiled black curls with a soft, velvety texture filling the screen. Against a stark, dark charcoal-gray backdrop, the trembling triggers a dramatic tremor in the hair; a series of micro-vibrations ripples through the mass, making each individual curl appear to dance and vibrate independently. The scene is illuminated by dramatic, high-contrast black and white lighting from a single, harsh, off-camera key light, which sculpts the subject's unseen features in light and shadow while capturing the edges of the trembling curls, creating a shimmering, dynamic halo effect against the profound darkness. The overall atmosphere is intense, raw, and emotionally charged, completely focused on the pure, dynamic expression of powerful, unseen excitement.", + "seed": 5221, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_14/index_7371_seed_5221_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.17210556855425238, + "vlm_logprob": -0.17210556855425238, + "vlm_yes_no_margin": 2.325, + "vlm_token_credit": -0.17210556855425238, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video consist of a series of rapid, sharp cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's hair a perfectly spherical Afro with thick, tightly coiled black curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a stark, dark charcoal-gray color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by high-contrast black and white lighting from a single, harsh off-camera key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include an extreme close-up shot focusing entirely on the hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4352434575557709 + ], + "opposite_logprob_tokens": [ + -1.0602434873580933 + ], + "logprob_reward": -0.4352434575557709, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4352434575557709, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03837239369750023, + 0.06235808506608009, + 0.049230460077524185, + 0.10884977877140045, + 0.05964846909046173, + 0.04726667329668999, + 0.25564682483673096, + 0.12494705617427826, + 0.1965915709733963, + 0.05708871781826019 + ], + "credit_map": [ + 0.004505814053118229, + 0.004075460601598024, + 0.004388106055557728, + 0.0038568300660699606, + 0.0024428118485957384, + 0.006078448612242937, + 0.009322580881416798, + 0.003702339483425021, + 0.007402551360428333, + 0.008592135272920132, + 0.00785236619412899, + 0.011849411763250828, + 0.009960413910448551, + 0.005290644243359566, + 0.0036142454482614994, + 0.00779631407931447, + 0.003198090009391308, + 0.012148304842412472, + 0.0032570522744208574, + 0.016444340348243713, + 0.008328369818627834, + 0.0015502430032938719, + 0.001633108826354146, + 0.0026709504891186953, + 0.01112573966383934, + 0.01263391599059105, + 0.012649727053940296, + 0.028477715328335762, + 0.016395313665270805, + 0.008577450178563595, + 0.010313201695680618, + 0.008676726371049881, + 0.005499343387782574, + 0.005167308729141951, + 0.01113624032586813, + 0.0072853537276387215, + 0.007361325901001692, + 0.0077254995703697205, + 0.010220597498118877, + 0.0052528041414916515, + 0.003924441523849964, + 0.004748344887048006, + 0.003237214870750904, + 0.007462774403393269, + 0.008759497664868832, + 0.008547276258468628, + 0.006708033848553896, + 0.0038790882099419832, + 0.006093160714954138, + 0.007757366634905338, + 0.005547825712710619, + 0.013353497721254826, + 0.0930054560303688, + 0.09482835233211517, + 0.019216854125261307, + 0.015844304114580154, + 0.005102975759655237, + 0.005594782996922731, + 0.006027515511959791, + 0.009874613955616951, + 0.012505846098065376, + 0.029536014422774315, + 0.040259845554828644, + 0.016045456752181053, + 0.007294866722077131, + 0.005301425699144602, + 0.0041608079336583614, + 0.008056769147515297, + 0.0413680374622345, + 0.08265749365091324, + 0.03741777315735817, + 0.010334409773349762, + 0.002534636529162526, + 0.004258378874510527, + 0.0037804432213306427, + 0.008346044458448887, + 0.010547660291194916, + 0.007106286473572254, + 0.008479495532810688, + 0.012035774998366833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.4352434575557709 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video consist of a series of rapid, sharp cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's hair a perfectly spherical Afro with thick, tightly coiled black curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a stark, dark charcoal-gray color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by high-contrast black and white lighting from a single, harsh off-camera key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include an extreme close-up shot focusing entirely on the hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17996403574943542 + ], + "opposite_logprob_tokens": [ + -1.8049640655517578 + ], + "logprob_reward": -0.17996403574943542, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17996403574943542, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12710317969322205, + 0.03699331358075142, + 0.02179255336523056, + 0.06498043239116669, + 0.03557780385017395, + 0.06553737074136734, + 0.156167671084404, + 0.18043817579746246, + 0.14325523376464844, + 0.16815431416034698 + ], + "credit_map": [ + 0.027170738205313683, + 0.01400259230285883, + 0.006531301885843277, + 0.006419946439564228, + 0.006012164521962404, + 0.03749760612845421, + 0.01004072092473507, + 0.019428113475441933, + 0.0034022594336420298, + 0.005531285423785448, + 0.005161973647773266, + 0.004536287859082222, + 0.005124104209244251, + 0.004676312208175659, + 0.0029485917184501886, + 0.005612497683614492, + 0.001443362096324563, + 0.0034147428814321756, + 0.002718138974159956, + 0.006443290505558252, + 0.004448176361620426, + 0.0011262696934863925, + 0.000884189794305712, + 0.0013143840478733182, + 0.006354689598083496, + 0.009404994547367096, + 0.007729376200586557, + 0.016426408663392067, + 0.009736858308315277, + 0.004624335560947657, + 0.00587055878713727, + 0.0048332130536437035, + 0.002868156647309661, + 0.005027633160352707, + 0.004807831719517708, + 0.0059619322419166565, + 0.004587944131344557, + 0.005087742116302252, + 0.004215905908495188, + 0.003020659787580371, + 0.006711896508932114, + 0.004993331152945757, + 0.004865258000791073, + 0.01037635374814272, + 0.010952943935990334, + 0.008783902041614056, + 0.011461355723440647, + 0.007392331026494503, + 0.010021154768764973, + 0.006271077785640955, + 0.019219648092985153, + 0.010368400253355503, + 0.031066348776221275, + 0.05227275192737579, + 0.015035148710012436, + 0.01191314309835434, + 0.010871070437133312, + 0.011128168553113937, + 0.01123962365090847, + 0.013847720809280872, + 0.01886874809861183, + 0.047877732664346695, + 0.032067783176898956, + 0.034537315368652344, + 0.004241261165589094, + 0.004112155642360449, + 0.004143927246332169, + 0.008261719718575478, + 0.0432245209813118, + 0.055261142551898956, + 0.014863110147416592, + 0.009147381410002708, + 0.006666593253612518, + 0.012111928313970566, + 0.006143699400126934, + 0.02223905734717846, + 0.03193444013595581, + 0.02332233265042305, + 0.02548973448574543, + 0.04024652764201164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.17996403574943542 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video consist of a series of rapid, sharp cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's hair a perfectly spherical Afro with thick, tightly coiled black curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a stark, dark charcoal-gray color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by high-contrast black and white lighting from a single, harsh off-camera key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include an extreme close-up shot focusing entirely on the hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897821068763733 + ], + "opposite_logprob_tokens": [ + -2.5789783000946045 + ], + "logprob_reward": -0.07897821068763733, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897821068763733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16896803677082062, + 0.033093083649873734, + 0.16119982302188873, + 0.09552191942930222, + 0.052735090255737305, + 0.07976178824901581, + 0.02420482225716114, + 0.1146094873547554, + 0.18348577618598938, + 0.08642014116048813 + ], + "credit_map": [ + 0.025842634961009026, + 0.014196272008121014, + 0.01128863450139761, + 0.010111107490956783, + 0.007070078980177641, + 0.04832964763045311, + 0.039818987250328064, + 0.01231068093329668, + 0.004828129429370165, + 0.005430220160633326, + 0.006105193868279457, + 0.003364172065630555, + 0.003027109196409583, + 0.0034679279197007418, + 0.0025853835977613926, + 0.004284948110580444, + 0.016287298873066902, + 0.02305164746940136, + 0.03320359066128731, + 0.02614590711891651, + 0.03503883630037308, + 0.008297245018184185, + 0.0069231609813869, + 0.012252137996256351, + 0.010229409672319889, + 0.01144863199442625, + 0.011811877600848675, + 0.02522384747862816, + 0.014720311388373375, + 0.006330200470983982, + 0.009220409207046032, + 0.006537236738950014, + 0.0037914172280579805, + 0.005912963300943375, + 0.01152009516954422, + 0.007055447902530432, + 0.00657618185505271, + 0.005916532129049301, + 0.0069611556828022, + 0.005001297686249018, + 0.004613589029759169, + 0.007153655868023634, + 0.004699409008026123, + 0.00952939409762621, + 0.011649668216705322, + 0.013159996829926968, + 0.012279876507818699, + 0.01667618937790394, + 0.0022736547980457544, + 0.0023110585752874613, + 0.0016923978691920638, + 0.002703295787796378, + 0.0054886192083358765, + 0.0035944313276559114, + 0.0031829108484089375, + 0.0029584525618702173, + 0.004362634848803282, + 0.004480249248445034, + 0.00672003673389554, + 0.007728023454546928, + 0.010463018901646137, + 0.031058942899107933, + 0.038310691714286804, + 0.011485890485346317, + 0.004472002852708101, + 0.0042404904961586, + 0.0023022862151265144, + 0.0068996320478618145, + 0.05533149838447571, + 0.07817021757364273, + 0.024264108389616013, + 0.00780552439391613, + 0.006035184487700462, + 0.007007786072790623, + 0.006818912923336029, + 0.012022278271615505, + 0.01342764776200056, + 0.011445515789091587, + 0.013047367334365845, + 0.016615457832813263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.07897821068763733 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video consist of a series of rapid, sharp cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's hair a perfectly spherical Afro with thick, tightly coiled black curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a stark, dark charcoal-gray color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by high-contrast black and white lighting from a single, harsh off-camera key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include an extreme close-up shot focusing entirely on the hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16039521992206573 + ], + "opposite_logprob_tokens": [ + -1.9103952646255493 + ], + "logprob_reward": -0.16039521992206573, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16039521992206573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09764169156551361, + 0.06866665184497833, + 0.07894327491521835, + 0.05038943886756897, + 0.03258775919675827, + 0.09452684968709946, + 0.22038203477859497, + 0.20791980624198914, + 0.05015473812818527, + 0.098787821829319 + ], + "credit_map": [ + 0.019201848655939102, + 0.025524549186229706, + 0.006441195961087942, + 0.007810099981725216, + 0.006531968712806702, + 0.009011072106659412, + 0.01588832214474678, + 0.007232622243463993, + 0.005991218611598015, + 0.011335181072354317, + 0.01202048547565937, + 0.008064347319304943, + 0.007666310295462608, + 0.00710385711863637, + 0.005734554026275873, + 0.010750685818493366, + 0.0057358345948159695, + 0.016206035390496254, + 0.007024135906249285, + 0.02233152650296688, + 0.01474015973508358, + 0.0046005891636013985, + 0.0028328620828688145, + 0.005472124088555574, + 0.005237575154751539, + 0.010187973268330097, + 0.005179800093173981, + 0.007417968008667231, + 0.0075583006255328655, + 0.005404347088187933, + 0.0055532874539494514, + 0.0038501820527017117, + 0.002575542079284787, + 0.004353594034910202, + 0.005309368949383497, + 0.0044792937114834785, + 0.0036390505265444517, + 0.0036724575329571962, + 0.005063278134912252, + 0.003495172131806612, + 0.010482347570359707, + 0.007592891808599234, + 0.008165187202394009, + 0.011191331781446934, + 0.015132655389606953, + 0.01202742476016283, + 0.012819560244679451, + 0.017115451395511627, + 0.004487278405576944, + 0.0059532346203923225, + 0.0031292459461838007, + 0.010264789685606956, + 0.08023543655872345, + 0.09115075320005417, + 0.015343944542109966, + 0.009817329235374928, + 0.007138072978705168, + 0.007847907021641731, + 0.011135962791740894, + 0.01552286371588707, + 0.02001417987048626, + 0.05436057597398758, + 0.06803420931100845, + 0.023866020143032074, + 0.001475962228141725, + 0.001679642009548843, + 0.0014096847735345364, + 0.002616496291011572, + 0.012178586795926094, + 0.020212706178426743, + 0.00777127081528306, + 0.002810385078191757, + 0.006654195953160524, + 0.008227752521634102, + 0.006318371742963791, + 0.013539189472794533, + 0.0168323777616024, + 0.013230727054178715, + 0.01860920339822769, + 0.015375998802483082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.16039521992206573 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video consist of a series of rapid, sharp cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's hair a perfectly spherical Afro with thick, tightly coiled black curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a stark, dark charcoal-gray color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by high-contrast black and white lighting from a single, harsh off-camera key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include an extreme close-up shot focusing entirely on the hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005946918856352568 + ], + "opposite_logprob_tokens": [ + -5.130947113037109 + ], + "logprob_reward": -0.005946918856352568, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005946918856352568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07877421379089355, + 0.05023621395230293, + 0.041228048503398895, + 0.05859137326478958, + 0.10782526433467865, + 0.028680387884378433, + 0.12068349123001099, + 0.09884873777627945, + 0.39172130823135376, + 0.023410888388752937 + ], + "credit_map": [ + 0.010501162149012089, + 0.011033167131245136, + 0.010617608204483986, + 0.015388778410851955, + 0.004442884586751461, + 0.008680105209350586, + 0.011595659889280796, + 0.006514848675578833, + 0.003778167301788926, + 0.0075974236242473125, + 0.007753192912787199, + 0.006677933968603611, + 0.007752077654004097, + 0.004497428424656391, + 0.004003416281193495, + 0.008176572620868683, + 0.0032868797425180674, + 0.006694411858916283, + 0.007013621274381876, + 0.010980434715747833, + 0.006668821442872286, + 0.0022959087509661913, + 0.001983229536563158, + 0.0023047379218041897, + 0.006183679681271315, + 0.007028013002127409, + 0.007309193257242441, + 0.012553143315017223, + 0.009605019353330135, + 0.0049758669920265675, + 0.005931058432906866, + 0.0050054024904966354, + 0.00883865263313055, + 0.008783922530710697, + 0.019184619188308716, + 0.01218433864414692, + 0.013681266456842422, + 0.012955225072801113, + 0.019491633400321007, + 0.012705613858997822, + 0.0017878840444609523, + 0.0024403531569987535, + 0.0018422635039314628, + 0.0033289510756731033, + 0.0057816761545836926, + 0.0052355206571519375, + 0.003351503284648061, + 0.004912237636744976, + 0.005043315701186657, + 0.010555247776210308, + 0.008867849595844746, + 0.008939584717154503, + 0.034819357097148895, + 0.02895289473235607, + 0.009994164109230042, + 0.013511073775589466, + 0.004278294276446104, + 0.004119714722037315, + 0.006240880116820335, + 0.006685960106551647, + 0.008506438694894314, + 0.02661731280386448, + 0.02910490520298481, + 0.01329523790627718, + 0.008673547767102718, + 0.010642466135323048, + 0.006733827292919159, + 0.01805092580616474, + 0.1216910183429718, + 0.16071724891662598, + 0.04403413459658623, + 0.021178146824240685, + 0.0009692314197309315, + 0.0019672370981425047, + 0.0012125278590247035, + 0.003959443420171738, + 0.004430236294865608, + 0.002773524960502982, + 0.0031495268922299147, + 0.004949161317199469 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 4 + ], + "selected_reward": -0.005946918856352568 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "series of rapid, sharp cuts captures the moment of intense bodily reaction from multiple angles, beginning with a dramatic low-angle shot that emphasizes the intense, unconscious trembling throughout the subject's frame. Then, the perspective abruptly shifts to extreme close-up, focusing entirely on a magnificent, perfectly spherical Afro, its thick, tightly coiled black curls with a soft, velvety texture filling the screen. Against a stark, dark charcoal-gray backdrop, the trembling triggers a dramatic tremor in the hair; a series of micro-vibrations ripples through the mass, making each individual curl appear to dance and vibrate independently. The scene is illuminated by dramatic, high-contrast black and white lighting from a single, harsh, off-camera key light, which sculpts the subject's unseen features in light and shadow while capturing the edges of the trembling curls, creating a shimmering, dynamic halo effect against the profound darkness. The overall atmosphere is intense, raw, and emotionally charged, completely focused on the pure, dynamic expression of powerful, unseen excitement.", + "seed": 5222, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_14/index_7371_seed_5222_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.28621991821564735, + "vlm_logprob": -0.28621991821564735, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.28621991821564735, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video consist of a series of rapid, sharp cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's hair a perfectly spherical Afro with thick, tightly coiled black curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a stark, dark charcoal-gray color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by high-contrast black and white lighting from a single, harsh off-camera key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include an extreme close-up shot focusing entirely on the hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.3168528079986572 + ], + "opposite_logprob_tokens": [ + -0.31685277819633484 + ], + "logprob_reward": -1.3168528079986572, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.3168528079986572, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09630929678678513, + 0.08332697302103043, + 0.07114963978528976, + 0.08114685863256454, + 0.18120616674423218, + 0.19831672310829163, + 0.0833083763718605, + 0.1465485394001007, + 0.041346579790115356, + 0.017340851947665215 + ], + "credit_map": [ + 0.008704271167516708, + 0.007117032539099455, + 0.009651055559515953, + 0.007258411031216383, + 0.019909298047423363, + 0.012528108432888985, + 0.014651489444077015, + 0.016489626839756966, + 0.011644032783806324, + 0.009941895492374897, + 0.007952522486448288, + 0.0062699769623577595, + 0.012773511931300163, + 0.011217832565307617, + 0.01414088997989893, + 0.009386307559907436, + 0.007844031788408756, + 0.009794896468520164, + 0.009425396099686623, + 0.007095961831510067, + 0.010139713063836098, + 0.010935474187135696, + 0.009143677540123463, + 0.006770487409085035, + 0.007881011813879013, + 0.01147710345685482, + 0.011102230288088322, + 0.013082685880362988, + 0.011948533356189728, + 0.008965115062892437, + 0.01177233550697565, + 0.004917837679386139, + 0.011039551347494125, + 0.03806226700544357, + 0.028260614722967148, + 0.045976247638463974, + 0.01482070330530405, + 0.013918058015406132, + 0.020634301006793976, + 0.008494434878230095, + 0.024493763223290443, + 0.025482095777988434, + 0.040060706436634064, + 0.02994621731340885, + 0.014570711180567741, + 0.030712131410837173, + 0.02073131501674652, + 0.012319771572947502, + 0.007943464443087578, + 0.009106164798140526, + 0.013109119608998299, + 0.012394984252750874, + 0.008555951528251171, + 0.009375755675137043, + 0.01215511653572321, + 0.010667815804481506, + 0.013029742054641247, + 0.012116264551877975, + 0.030347663909196854, + 0.017775125801563263, + 0.01601012796163559, + 0.019886169582605362, + 0.022913187742233276, + 0.01447024755179882, + 0.005916429217904806, + 0.004839985631406307, + 0.00644396198913455, + 0.005198655184358358, + 0.004977002739906311, + 0.00410874979570508, + 0.005142693407833576, + 0.004719096701592207, + 0.0033209945540875196, + 0.0018619904294610023, + 0.0020423694513738155, + 0.001617831876501441, + 0.0015837445389479399, + 0.00201661279425025, + 0.0024308571591973305, + 0.0024664506781846285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 7 + ], + "selected_reward": -1.3168528079986572 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video consist of a series of rapid, sharp cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's hair a perfectly spherical Afro with thick, tightly coiled black curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a stark, dark charcoal-gray color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by high-contrast black and white lighting from a single, harsh off-camera key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include an extreme close-up shot focusing entirely on the hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06210934370756149 + ], + "opposite_logprob_tokens": [ + -2.8121092319488525 + ], + "logprob_reward": -0.06210934370756149, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06210934370756149, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07527390122413635, + 0.03351664915680885, + 0.042287882417440414, + 0.06067202240228653, + 0.293743371963501, + 0.03740629181265831, + 0.15379171073436737, + 0.07700452208518982, + 0.07607521861791611, + 0.15022844076156616 + ], + "credit_map": [ + 0.007332341279834509, + 0.006437852047383785, + 0.008354268968105316, + 0.005182460881769657, + 0.012245193123817444, + 0.00883531104773283, + 0.010352854616940022, + 0.01653362065553665, + 0.00475354865193367, + 0.004888430703431368, + 0.002721847267821431, + 0.0034151191357523203, + 0.004661112558096647, + 0.004481900949031115, + 0.005615083500742912, + 0.0029796080198138952, + 0.004982363898307085, + 0.005437998101115227, + 0.004166917409747839, + 0.004206015262752771, + 0.007388418540358543, + 0.005243655294179916, + 0.005570165812969208, + 0.005292345769703388, + 0.0055354381911456585, + 0.006112406961619854, + 0.012663683854043484, + 0.013522105291485786, + 0.0040470901876688, + 0.004405824467539787, + 0.01090503390878439, + 0.003480437910184264, + 0.019847357645630836, + 0.0639176145195961, + 0.04124991223216057, + 0.06344863772392273, + 0.028402751311659813, + 0.027064066380262375, + 0.03414857015013695, + 0.015664448961615562, + 0.003337453817948699, + 0.005326403304934502, + 0.005508133675903082, + 0.005299062468111515, + 0.0035458605270832777, + 0.005471441429108381, + 0.006005201023072004, + 0.0029127360321581364, + 0.013714700937271118, + 0.016191335394978523, + 0.028574606403708458, + 0.02929367870092392, + 0.012014760635793209, + 0.017033753916621208, + 0.019711343571543694, + 0.01725754141807556, + 0.010102979838848114, + 0.005400648340582848, + 0.01795060932636261, + 0.00977590773254633, + 0.006477629765868187, + 0.00937630608677864, + 0.011695845052599907, + 0.006224596872925758, + 0.012062355875968933, + 0.007317985408008099, + 0.013452683575451374, + 0.00790336262434721, + 0.009306790307164192, + 0.0075072553008794785, + 0.009546354413032532, + 0.008978430181741714, + 0.02457631193101406, + 0.015628188848495483, + 0.011520364321768284, + 0.011430610902607441, + 0.016641490161418915, + 0.024008676409721375, + 0.024854904040694237, + 0.021567905321717262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 9 + ], + "selected_reward": -0.06210934370756149 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video consist of a series of rapid, sharp cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's hair a perfectly spherical Afro with thick, tightly coiled black curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a stark, dark charcoal-gray color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by high-contrast black and white lighting from a single, harsh off-camera key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include an extreme close-up shot focusing entirely on the hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026348982006311417 + ], + "opposite_logprob_tokens": [ + -3.6513490676879883 + ], + "logprob_reward": -0.026348982006311417, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026348982006311417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0438997708261013, + 0.09829363971948624, + 0.04838832840323448, + 0.031297799199819565, + 0.5671488046646118, + 0.03763924166560173, + 0.06594792753458023, + 0.060185860842466354, + 0.03255822882056236, + 0.01464033592492342 + ], + "credit_map": [ + 0.0036176193971186876, + 0.0034410913940519094, + 0.004598479252308607, + 0.0027065854519605637, + 0.009155130945146084, + 0.0062608965672552586, + 0.006315135862678289, + 0.007804835215210915, + 0.021951254457235336, + 0.009921655058860779, + 0.007067639380693436, + 0.009466966614127159, + 0.009599510580301285, + 0.01459011435508728, + 0.013011501170694828, + 0.012685002759099007, + 0.0059792655520141125, + 0.006366122514009476, + 0.0041104028932750225, + 0.002993175992742181, + 0.008206568658351898, + 0.01062911655753851, + 0.005411037243902683, + 0.004692647140473127, + 0.0032838094048202038, + 0.003477538237348199, + 0.006486604455858469, + 0.004501803312450647, + 0.0027161056641489267, + 0.002757029375061393, + 0.005463798530399799, + 0.002611114177852869, + 0.036357201635837555, + 0.114839568734169, + 0.0809415876865387, + 0.13589271903038025, + 0.05753111466765404, + 0.046796925365924835, + 0.06351444125175476, + 0.03127530217170715, + 0.004702987149357796, + 0.003883499652147293, + 0.004954924341291189, + 0.0056062061339616776, + 0.004241724964231253, + 0.00414961576461792, + 0.007005844730883837, + 0.003094444051384926, + 0.007378241512924433, + 0.008062129840254784, + 0.010054302401840687, + 0.009055635891854763, + 0.00664510065689683, + 0.008143792860209942, + 0.010154877789318562, + 0.00645385542884469, + 0.010092000477015972, + 0.004342123866081238, + 0.011517888866364956, + 0.006652466487139463, + 0.005990521050989628, + 0.007549998350441456, + 0.008879655040800571, + 0.0051612150855362415, + 0.007017932832241058, + 0.00342002185061574, + 0.00450329203158617, + 0.003153464524075389, + 0.0032136577647179365, + 0.003276559291407466, + 0.004310048185288906, + 0.0036632558330893517, + 0.002429249230772257, + 0.0011229722294956446, + 0.0018643330549821258, + 0.0013471362181007862, + 0.0010793569963425398, + 0.001716300961561501, + 0.002171382075175643, + 0.0029096058569848537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 6 + ], + "selected_reward": -0.026348982006311417 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video consist of a series of rapid, sharp cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's hair a perfectly spherical Afro with thick, tightly coiled black curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a stark, dark charcoal-gray color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by high-contrast black and white lighting from a single, harsh off-camera key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include an extreme close-up shot focusing entirely on the hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023299487307667732 + ], + "opposite_logprob_tokens": [ + -3.7732994556427 + ], + "logprob_reward": -0.023299487307667732, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023299487307667732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07627224177122116, + 0.13298290967941284, + 0.027206746861338615, + 0.0612356923520565, + 0.3438199460506439, + 0.05404308810830116, + 0.15722617506980896, + 0.1110718697309494, + 0.02384999766945839, + 0.012291296385228634 + ], + "credit_map": [ + 0.0065462300553917885, + 0.005106091499328613, + 0.005481521133333445, + 0.00506577966734767, + 0.01457344088703394, + 0.011331203393638134, + 0.013430669903755188, + 0.014737309888005257, + 0.03722319379448891, + 0.01446022093296051, + 0.00864436011761427, + 0.013254334218800068, + 0.011762399226427078, + 0.018509192392230034, + 0.012898006476461887, + 0.01623121090233326, + 0.003222590545192361, + 0.0033412007614970207, + 0.0035150269977748394, + 0.0027103442698717117, + 0.003642016090452671, + 0.0037532218266278505, + 0.0041115316562354565, + 0.0029108168091624975, + 0.005377774126827717, + 0.007089835591614246, + 0.01441824622452259, + 0.009185781702399254, + 0.006348559632897377, + 0.0043961782939732075, + 0.010365237481892109, + 0.0040540797635912895, + 0.022210899740457535, + 0.06573134660720825, + 0.05072524771094322, + 0.08158470690250397, + 0.03613561764359474, + 0.032373979687690735, + 0.039211299270391464, + 0.015846850350499153, + 0.004659721162170172, + 0.005096410401165485, + 0.01212761178612709, + 0.009145205840468407, + 0.0037703777197748423, + 0.00907929427921772, + 0.00626023905351758, + 0.003904225304722786, + 0.0156239690259099, + 0.01622653193771839, + 0.022892789915204048, + 0.023891393095254898, + 0.01302881445735693, + 0.022450925782322884, + 0.022594593465328217, + 0.020517155528068542, + 0.021439358592033386, + 0.007401293143630028, + 0.01776442863047123, + 0.015058309771120548, + 0.00908817071467638, + 0.009976224973797798, + 0.018951328471302986, + 0.01139275636523962, + 0.005321416072547436, + 0.001750521594658494, + 0.004403769504278898, + 0.0018011496867984533, + 0.001846759463660419, + 0.0023120115511119366, + 0.0036091485526412725, + 0.0028052206616848707, + 0.001922108232975006, + 0.0010384299093857408, + 0.0019298671977594495, + 0.0015781433321535587, + 0.0009048691717907786, + 0.001273737521842122, + 0.0020733890123665333, + 0.001570752589032054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 1 + ], + "selected_reward": -0.023299487307667732 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video consist of a series of rapid, sharp cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's hair a perfectly spherical Afro with thick, tightly coiled black curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a stark, dark charcoal-gray color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by high-contrast black and white lighting from a single, harsh off-camera key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include an extreme close-up shot focusing entirely on the hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024889700580388308 + ], + "opposite_logprob_tokens": [ + -6.00248908996582 + ], + "logprob_reward": -0.0024889700580388308, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024889700580388308, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13094301521778107, + 0.07728739827871323, + 0.07248251885175705, + 0.10696586221456528, + 0.18923576176166534, + 0.16889742016792297, + 0.11072287708520889, + 0.054205697029829025, + 0.06529217958450317, + 0.02396726980805397 + ], + "credit_map": [ + 0.010033573023974895, + 0.010655209422111511, + 0.013036690652370453, + 0.01113270502537489, + 0.022464176639914513, + 0.015077419579029083, + 0.022272352129220963, + 0.026270877569913864, + 0.01701478660106659, + 0.011921584606170654, + 0.004412206821143627, + 0.005579509772360325, + 0.011189677752554417, + 0.010542958043515682, + 0.009498256258666515, + 0.007128414232283831, + 0.008431356400251389, + 0.009043265134096146, + 0.008853311650454998, + 0.007075231056660414, + 0.01069062203168869, + 0.01121255848556757, + 0.00929546169936657, + 0.007880711928009987, + 0.010307993739843369, + 0.01153677236288786, + 0.021108360961079597, + 0.023973478004336357, + 0.0066130575723946095, + 0.008043118752539158, + 0.01956111378967762, + 0.0058219656348228455, + 0.014908064156770706, + 0.04025040939450264, + 0.023132646456360817, + 0.031504690647125244, + 0.023183520883321762, + 0.01856749691069126, + 0.023343171924352646, + 0.01434575766324997, + 0.01836208999156952, + 0.019399404525756836, + 0.03820878267288208, + 0.027361180633306503, + 0.009994563646614552, + 0.026980459690093994, + 0.018527984619140625, + 0.010062954388558865, + 0.007937627844512463, + 0.008657815866172314, + 0.02256513573229313, + 0.03117445670068264, + 0.0046940031461417675, + 0.008561071939766407, + 0.013811009004712105, + 0.013321750797331333, + 0.007221360225230455, + 0.004327222239226103, + 0.011402061209082603, + 0.007656771223992109, + 0.00469551095739007, + 0.006068026646971703, + 0.008232609368860722, + 0.004602131899446249, + 0.012044273316860199, + 0.007081856485456228, + 0.009894181042909622, + 0.0067821647971868515, + 0.007516552694141865, + 0.006797322537750006, + 0.008039352484047413, + 0.007136466912925243, + 0.004529546946287155, + 0.0018285649130120873, + 0.0023941821418702602, + 0.0016388104995712638, + 0.001984453061595559, + 0.0025646882131695747, + 0.003964659757912159, + 0.005062363110482693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.0024889700580388308 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "series of rapid, sharp cuts captures the moment of intense bodily reaction from multiple angles, beginning with a dramatic low-angle shot that emphasizes the intense, unconscious trembling throughout the subject's frame. Then, the perspective abruptly shifts to extreme close-up, focusing entirely on a magnificent, perfectly spherical Afro, its thick, tightly coiled black curls with a soft, velvety texture filling the screen. Against a stark, dark charcoal-gray backdrop, the trembling triggers a dramatic tremor in the hair; a series of micro-vibrations ripples through the mass, making each individual curl appear to dance and vibrate independently. The scene is illuminated by dramatic, high-contrast black and white lighting from a single, harsh, off-camera key light, which sculpts the subject's unseen features in light and shadow while capturing the edges of the trembling curls, creating a shimmering, dynamic halo effect against the profound darkness. The overall atmosphere is intense, raw, and emotionally charged, completely focused on the pure, dynamic expression of powerful, unseen excitement.", + "seed": 5223, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_14/index_7371_seed_5223_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.3755967730656266, + "vlm_logprob": -0.3755967730656266, + "vlm_yes_no_margin": 2.375, + "vlm_token_credit": -0.3755967730656266, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video consist of a series of rapid, sharp cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's hair a perfectly spherical Afro with thick, tightly coiled black curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a stark, dark charcoal-gray color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by high-contrast black and white lighting from a single, harsh off-camera key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include an extreme close-up shot focusing entirely on the hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.70521080493927 + ], + "opposite_logprob_tokens": [ + -0.2052108496427536 + ], + "logprob_reward": -1.70521080493927, + "yes_no_margin_reward": -1.5, + "token_credit_reward": -1.70521080493927, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05009147897362709, + 0.13567309081554413, + 0.03496868535876274, + 0.020600972697138786, + 0.010222170501947403, + 0.06418382376432419, + 0.015597624704241753, + 0.025640495121479034, + 0.6266258955001831, + 0.016395816579461098 + ], + "credit_map": [ + 0.006698963697999716, + 0.007760743144899607, + 0.0026149216573685408, + 0.004780338145792484, + 0.00934443436563015, + 0.006281259004026651, + 0.007656134199351072, + 0.004954684991389513, + 0.015604452230036259, + 0.007627821993082762, + 0.010992802679538727, + 0.01055967342108488, + 0.01570841297507286, + 0.020614977926015854, + 0.038479290902614594, + 0.01608564890921116, + 0.0033963655587285757, + 0.00270936475135386, + 0.00488069374114275, + 0.0020570585038512945, + 0.0034961977507919073, + 0.0033916581887751818, + 0.007439505774527788, + 0.00759784085676074, + 0.0030704925302416086, + 0.003229559864848852, + 0.002905286382883787, + 0.0018187706591561437, + 0.0021898006089031696, + 0.0024937582202255726, + 0.002246944233775139, + 0.0026463607791811228, + 0.0015292145544663072, + 0.0011215467238798738, + 0.0014720680192112923, + 0.0013294932432472706, + 0.0011900649406015873, + 0.00122783612459898, + 0.0011267117224633694, + 0.0012252359883859754, + 0.005979527253657579, + 0.003383455565199256, + 0.019030606374144554, + 0.005400215275585651, + 0.0038788416422903538, + 0.0034449794329702854, + 0.0189201757311821, + 0.004146021790802479, + 0.0018631702987477183, + 0.0013590471353381872, + 0.003098392626270652, + 0.0014998263213783503, + 0.0009197433246299624, + 0.0015682341763749719, + 0.003509405767545104, + 0.0017798058688640594, + 0.0024584836792200804, + 0.0020412292797118425, + 0.004997745156288147, + 0.004029692616313696, + 0.002658197423443198, + 0.0019287981558591127, + 0.003957787062972784, + 0.003568560816347599, + 0.038579173386096954, + 0.014156123623251915, + 0.27047622203826904, + 0.25638484954833984, + 0.0037476627621799707, + 0.011778699234127998, + 0.02625853568315506, + 0.005244569852948189, + 0.0035671237856149673, + 0.0016360458685085177, + 0.002270107390359044, + 0.002603195607662201, + 0.0013600344536826015, + 0.0016661904519423842, + 0.001686745323240757, + 0.0016063740476965904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 5 + ], + "selected_reward": -1.70521080493927 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video consist of a series of rapid, sharp cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's hair a perfectly spherical Afro with thick, tightly coiled black curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a stark, dark charcoal-gray color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by high-contrast black and white lighting from a single, harsh off-camera key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include an extreme close-up shot focusing entirely on the hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08913411945104599 + ], + "opposite_logprob_tokens": [ + -2.4641342163085938 + ], + "logprob_reward": -0.08913411945104599, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08913411945104599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03215418756008148, + 0.44945377111434937, + 0.018421407788991928, + 0.011143402196466923, + 0.02159971557557583, + 0.09035920351743698, + 0.01703009381890297, + 0.015323391184210777, + 0.32578712701797485, + 0.01872776821255684 + ], + "credit_map": [ + 0.004445002414286137, + 0.004594274330884218, + 0.0016783762257546186, + 0.0028489134274423122, + 0.00664774002507329, + 0.00432744063436985, + 0.004017821047455072, + 0.003594619920477271, + 0.045144274830818176, + 0.018793564289808273, + 0.07279611378908157, + 0.03622058406472206, + 0.040442679077386856, + 0.05916149169206619, + 0.12059088051319122, + 0.056304171681404114, + 0.001679171691648662, + 0.0010814089328050613, + 0.0026972282212227583, + 0.000976508716121316, + 0.0016628592275083065, + 0.001405495684593916, + 0.00447400938719511, + 0.004444727208465338, + 0.0016734586097300053, + 0.0013478384353220463, + 0.001696533290669322, + 0.0013117892667651176, + 0.0010729876812547445, + 0.0012671101139858365, + 0.0014249756932258606, + 0.0013487078249454498, + 0.0029118009842932224, + 0.0021328264847397804, + 0.003792357398197055, + 0.0026269780937582254, + 0.002228084020316601, + 0.0022059904877096415, + 0.003085004398599267, + 0.0026166748721152544, + 0.007694496773183346, + 0.002191035309806466, + 0.0299761313945055, + 0.004634479060769081, + 0.0034158274065703154, + 0.0034114595036953688, + 0.035151030868291855, + 0.003884746693074703, + 0.002361749531701207, + 0.001189078320749104, + 0.0042739128693938255, + 0.001458971295505762, + 0.0011533439392223954, + 0.000987518928013742, + 0.0038467012345790863, + 0.001758816302753985, + 0.00152741523925215, + 0.0015267165144905448, + 0.002452187705785036, + 0.00246896012686193, + 0.0015463788295164704, + 0.0013147464487701654, + 0.0023748751264065504, + 0.002112112008035183, + 0.01989167183637619, + 0.007567608263343573, + 0.14081411063671112, + 0.13373738527297974, + 0.001649199053645134, + 0.00613826047629118, + 0.013325986452400684, + 0.0026629064232110977, + 0.0028869921807199717, + 0.0017280292231589556, + 0.0026532358024269342, + 0.0028744216542690992, + 0.0017530455952510238, + 0.0021559197921305895, + 0.0026318589225411415, + 0.002044264692813158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 5 + ], + "selected_reward": -0.08913411945104599 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video consist of a series of rapid, sharp cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's hair a perfectly spherical Afro with thick, tightly coiled black curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a stark, dark charcoal-gray color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by high-contrast black and white lighting from a single, harsh off-camera key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include an extreme close-up shot focusing entirely on the hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307219386100769 + ], + "opposite_logprob_tokens": [ + -3.16807222366333 + ], + "logprob_reward": -0.04307219386100769, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307219386100769, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09105952084064484, + 0.28465330600738525, + 0.03350589796900749, + 0.07372932136058807, + 0.0715956836938858, + 0.05750894919037819, + 0.01281727384775877, + 0.007157850079238415, + 0.3484443426132202, + 0.019527852535247803 + ], + "credit_map": [ + 0.0215462539345026, + 0.007007728796452284, + 0.006287621334195137, + 0.007469151634722948, + 0.010197864845395088, + 0.013068346306681633, + 0.01333654671907425, + 0.012146000750362873, + 0.028827721253037453, + 0.01231247466057539, + 0.0494326651096344, + 0.02495512366294861, + 0.025478990748524666, + 0.03665792942047119, + 0.07221122086048126, + 0.03477716073393822, + 0.003977484069764614, + 0.0026527426671236753, + 0.0066725448705255985, + 0.0026050179731100798, + 0.0031916836742311716, + 0.004014085978269577, + 0.00534866051748395, + 0.005043680779635906, + 0.012696933001279831, + 0.009908471256494522, + 0.013449943624436855, + 0.006024185102432966, + 0.006776570808142424, + 0.007990117184817791, + 0.008727745153009892, + 0.008155355229973793, + 0.015851447358727455, + 0.004762616939842701, + 0.017988964915275574, + 0.005616127047687769, + 0.005524955224245787, + 0.00451682647690177, + 0.010246185585856438, + 0.007088559679687023, + 0.004460690077394247, + 0.0020386262331157923, + 0.021308619529008865, + 0.003509078174829483, + 0.0019441654440015554, + 0.0020580654963850975, + 0.019063042476773262, + 0.003126661293208599, + 0.0013804740738123655, + 0.0010542505187913775, + 0.002995368791744113, + 0.001416974700987339, + 0.0008043596753850579, + 0.0008701322949491441, + 0.0028532356955111027, + 0.0014424783876165748, + 0.0012706882553175092, + 0.0006024186150170863, + 0.001280299504287541, + 0.0007846302469260991, + 0.0007891571149230003, + 0.0005806325352750719, + 0.0008425304549746215, + 0.0010074930032715201, + 0.021279385313391685, + 0.008143489249050617, + 0.15047580003738403, + 0.14339521527290344, + 0.0016294867964461446, + 0.00641871802508831, + 0.014365521259605885, + 0.0027367188595235348, + 0.0034808891359716654, + 0.0024617877788841724, + 0.002313226694241166, + 0.002926111686974764, + 0.0024288594722747803, + 0.0023235182743519545, + 0.0016639980021864176, + 0.001929461839608848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.04307219386100769 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video consist of a series of rapid, sharp cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's hair a perfectly spherical Afro with thick, tightly coiled black curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a stark, dark charcoal-gray color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by high-contrast black and white lighting from a single, harsh off-camera key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include an extreme close-up shot focusing entirely on the hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026366746053099632 + ], + "opposite_logprob_tokens": [ + -3.651366710662842 + ], + "logprob_reward": -0.026366746053099632, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026366746053099632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09463632106781006, + 0.11543276160955429, + 0.10720837116241455, + 0.06496089696884155, + 0.026767672970891, + 0.1561439037322998, + 0.032419998198747635, + 0.036181461066007614, + 0.3367179036140442, + 0.029530653730034828 + ], + "credit_map": [ + 0.02021363005042076, + 0.007710132747888565, + 0.007083408534526825, + 0.00785996112972498, + 0.013838869519531727, + 0.012633571401238441, + 0.013592165894806385, + 0.011704584583640099, + 0.00973450019955635, + 0.006243521347641945, + 0.01081874594092369, + 0.007717328146100044, + 0.011612134985625744, + 0.017846647650003433, + 0.03484850749373436, + 0.016611380502581596, + 0.013400577940046787, + 0.009289075620472431, + 0.02048317901790142, + 0.007419806905090809, + 0.01081083808094263, + 0.011051727458834648, + 0.015355310402810574, + 0.01939784735441208, + 0.011216615326702595, + 0.00992453470826149, + 0.010475048795342445, + 0.0061092255637049675, + 0.004548192024230957, + 0.007723896764218807, + 0.007067668251693249, + 0.007895712740719318, + 0.003559222910553217, + 0.002067163586616516, + 0.004571550525724888, + 0.0036422780249267817, + 0.0023116739466786385, + 0.0026578453835099936, + 0.0035463108215481043, + 0.004411627538502216, + 0.01086394488811493, + 0.003729953896254301, + 0.059221409261226654, + 0.008376061916351318, + 0.006329900119453669, + 0.005160864442586899, + 0.05595136061310768, + 0.006510395556688309, + 0.0036408966407179832, + 0.0032300357706844807, + 0.006307070143520832, + 0.0039329142309725285, + 0.003003128804266453, + 0.0033803025726228952, + 0.005920408293604851, + 0.0030052403453737497, + 0.004026926588267088, + 0.0040399134159088135, + 0.0063909986056387424, + 0.004673432093113661, + 0.0044899568893015385, + 0.0034703067503869534, + 0.0042243958450853825, + 0.004865532275289297, + 0.020776458084583282, + 0.007897417061030865, + 0.1450093686580658, + 0.13785424828529358, + 0.001814087387174368, + 0.006238860078155994, + 0.014173711650073528, + 0.0029537633527070284, + 0.006568941753357649, + 0.003067593090236187, + 0.0038092052564024925, + 0.004203000105917454, + 0.002411194611340761, + 0.0036525370087474585, + 0.003057737834751606, + 0.002760444302111864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 1 + ], + "selected_reward": -0.026366746053099632 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video consist of a series of rapid, sharp cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject's hair a perfectly spherical Afro with thick, tightly coiled black curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a stark, dark charcoal-gray color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by high-contrast black and white lighting from a single, harsh off-camera key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include an extreme close-up shot focusing entirely on the hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014200001023709774 + ], + "opposite_logprob_tokens": [ + -4.264200210571289 + ], + "logprob_reward": -0.014200001023709774, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014200001023709774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.032330866903066635, + 0.04052036255598068, + 0.021963732317090034, + 0.006619763560593128, + 0.007225078530609608, + 0.13594549894332886, + 0.018266385421156883, + 0.007300992030650377, + 0.7256636619567871, + 0.004163619130849838 + ], + "credit_map": [ + 0.0042382278479635715, + 0.003601144766435027, + 0.0028679119423031807, + 0.0036095790565013885, + 0.003733311779797077, + 0.0041888742707669735, + 0.005822461098432541, + 0.004269357305020094, + 0.0045627872459590435, + 0.0034309474285691977, + 0.005713779013603926, + 0.004048947244882584, + 0.005246089771389961, + 0.007136116269975901, + 0.0065488810651004314, + 0.0038328166119754314, + 0.0023168607149273157, + 0.0020293269772082567, + 0.0028554790187627077, + 0.0014237455325201154, + 0.002784227952361107, + 0.002366322325542569, + 0.003878391347825527, + 0.004309378564357758, + 0.0008327479590661824, + 0.0007821128237992525, + 0.0010140301892533898, + 0.0008026061696000397, + 0.0007061212090775371, + 0.0007746961200609803, + 0.0007981344824656844, + 0.0009093140251934528, + 0.0008571018697693944, + 0.000680824217852205, + 0.0011528381146490574, + 0.001095141633413732, + 0.0006663841777481139, + 0.0007890393026173115, + 0.0009652091539464891, + 0.001018540933728218, + 0.0074787926860153675, + 0.001704719034023583, + 0.05639627203345299, + 0.004562474321573973, + 0.002563486574217677, + 0.00275132292881608, + 0.05590542405843735, + 0.004583008587360382, + 0.002159354044124484, + 0.0015349378809332848, + 0.0038874198216944933, + 0.0022664146963506937, + 0.001553823472931981, + 0.0016239439137279987, + 0.0034685309510678053, + 0.0017719595925882459, + 0.0007799528539180756, + 0.0005885295104235411, + 0.001275021000765264, + 0.00117878383025527, + 0.0007399435271508992, + 0.0005450125318020582, + 0.001087336684577167, + 0.001106412149965763, + 0.04423871263861656, + 0.017698826268315315, + 0.31276118755340576, + 0.2984359562397003, + 0.003684875089675188, + 0.013635530136525631, + 0.02954057976603508, + 0.005668031983077526, + 0.0006648934213444591, + 0.00043796372483484447, + 0.0006197108305059373, + 0.0005889901658520103, + 0.00035091873724013567, + 0.0004427885578479618, + 0.0005447586881928146, + 0.0005135949468240142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 1 + ], + "selected_reward": -0.014200001023709774 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_15/index_tensor([9888, 9888])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_15/index_tensor([9888, 9888])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..672e8560a20252bde32476d4948818c3d78baebd --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_15/index_tensor([9888, 9888])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, high-angle, wide shot establishes a panoramic view of a deeply chaotic room, steeped in a profound sense of quiet solitude. The primary light source is a bare-bulb floor lamp, casting a warm yellow-orange glow that creates long, dramatic shadows and accentuates the chaotic textures of the environment. Piles of academic books with frayed spines, stacks of old magazines, and discarded clothing create a chaotic mess on the floor and a nearby unmade bed. At the center of this chaos, a solitary figure hunches over a low wooden stool, their back arched in a pose of deep exhaustion or concentration. Their identity is completely obscured, their face obscured by a mass of long, dark, tangled hair and the deep hood of an oversized, faded black hoodie. Before them, on a small patch of clean floor, sits a simple white Styrofoam cup of instant noodles, illuminated by thick white steam rising steadily from it. The figure's pale hand slowly and deliberately lifts a portion of noodles with a pair of disposable chopsticks. This micro-movement suggests a weary, almost mechanical act of maintenance, creating a poignant and atmospheric scene of late-night isolation.", + "seed": 14160, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_15/index_9888_seed_14160_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1632922811433673, + "vlm_logprob": -0.1632922811433673, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.1632922811433673, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a chaotic room viewed from a static, high-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room filled with piles of academic books, old magazines, and discarded clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure hunched over a low wooden stool wearing a faded black hoodie with long dark hair obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a simple white Styrofoam cup of instant noodles with thick white steam rising from it visible on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure use disposable chopsticks to slowly and deliberately lift a portion of noodles from the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.732585072517395 + ], + "opposite_logprob_tokens": [ + -0.732585072517395 + ], + "logprob_reward": -0.732585072517395, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.732585072517395, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.18232783675193787, + 0.393419474363327, + 0.03044317103922367, + 0.05208653211593628, + 0.04713757708668709, + 0.14811868965625763, + 0.043156854808330536, + 0.01754802092909813, + 0.05751850828528404, + 0.028243329375982285 + ], + "credit_map": [ + 0.02311965450644493, + 0.018908709287643433, + 0.02619093470275402, + 0.024688612669706345, + 0.023846358060836792, + 0.022402869537472725, + 0.01871620863676071, + 0.02445448376238346, + 0.0349545031785965, + 0.03627324849367142, + 0.05117711424827576, + 0.07653521746397018, + 0.05070225149393082, + 0.06582000106573105, + 0.04429028928279877, + 0.03366686403751373, + 0.003756890771910548, + 0.003515431424602866, + 0.0056260572746396065, + 0.003702807240188122, + 0.00268688821233809, + 0.003443759633228183, + 0.004243180621415377, + 0.0034681539982557297, + 0.005378579255193472, + 0.004730408079922199, + 0.006365037057548761, + 0.009217897430062294, + 0.0067506334744393826, + 0.007725537288933992, + 0.006194256246089935, + 0.005724182818084955, + 0.005409107077866793, + 0.005728248506784439, + 0.0051868753507733345, + 0.006355300545692444, + 0.004903017543256283, + 0.008536312729120255, + 0.004753641784191132, + 0.006265073083341122, + 0.014386442489922047, + 0.02244790457189083, + 0.02174868807196617, + 0.026505351066589355, + 0.014838448725640774, + 0.021511949598789215, + 0.013738451525568962, + 0.012941454537212849, + 0.00491567375138402, + 0.004013063386082649, + 0.006225238088518381, + 0.0057879481464624405, + 0.004495106637477875, + 0.004963735584169626, + 0.007869093678891659, + 0.004886995535343885, + 0.0019277656683698297, + 0.0017379792407155037, + 0.0028055557049810886, + 0.0028225192800164223, + 0.0015164996730163693, + 0.0031375379767268896, + 0.001414271304383874, + 0.0021858925465494394, + 0.008116328157484531, + 0.006832659710198641, + 0.00928466860204935, + 0.008318365551531315, + 0.004784219432622194, + 0.00787376333028078, + 0.006954062730073929, + 0.005354441236704588, + 0.0035896149929612875, + 0.0028122200164943933, + 0.00579437892884016, + 0.003709950018674135, + 0.0027246885001659393, + 0.002656044904142618, + 0.003642493858933449, + 0.003313938155770302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.732585072517395 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a chaotic room viewed from a static, high-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room filled with piles of academic books, old magazines, and discarded clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure hunched over a low wooden stool wearing a faded black hoodie with long dark hair obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a simple white Styrofoam cup of instant noodles with thick white steam rising from it visible on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure use disposable chopsticks to slowly and deliberately lift a portion of noodles from the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029849810525774956 + ], + "opposite_logprob_tokens": [ + -3.5298497676849365 + ], + "logprob_reward": -0.029849810525774956, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029849810525774956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22360925376415253, + 0.10995257645845413, + 0.09808940440416336, + 0.048537079244852066, + 0.11488095670938492, + 0.10876136273145676, + 0.21886953711509705, + 0.02376491017639637, + 0.027876408770680428, + 0.025658460333943367 + ], + "credit_map": [ + 0.027323633432388306, + 0.018812259659171104, + 0.034665171056985855, + 0.03497821465134621, + 0.01823277398943901, + 0.03788461163640022, + 0.020258044824004173, + 0.03145454451441765, + 0.016160188242793083, + 0.0110364044085145, + 0.016526535153388977, + 0.01943124085664749, + 0.013219705782830715, + 0.01167174894362688, + 0.011184689588844776, + 0.01072207186371088, + 0.013024011626839638, + 0.011411591432988644, + 0.019131934270262718, + 0.011318160220980644, + 0.008630514144897461, + 0.010954288765788078, + 0.008512023836374283, + 0.01510688941925764, + 0.004373341798782349, + 0.0038913008756935596, + 0.006200841628015041, + 0.008130829781293869, + 0.007056578062474728, + 0.007104275748133659, + 0.00682616513222456, + 0.004953749943524599, + 0.02227897383272648, + 0.01138638611882925, + 0.014913205057382584, + 0.016837187111377716, + 0.010748741216957569, + 0.011262328363955021, + 0.011900372803211212, + 0.01555376872420311, + 0.012175868265330791, + 0.014811383560299873, + 0.016678841784596443, + 0.01773804984986782, + 0.012963129207491875, + 0.011475167237222195, + 0.011129703372716904, + 0.011789225973188877, + 0.02752712368965149, + 0.018157193437218666, + 0.04142102226614952, + 0.027535324916243553, + 0.01875992864370346, + 0.024038251489400864, + 0.029969101771712303, + 0.03146158903837204, + 0.003372678766027093, + 0.00194498710334301, + 0.005720089189708233, + 0.0030560351442545652, + 0.0018736382480710745, + 0.002628742251545191, + 0.001619480550289154, + 0.0035492614842951298, + 0.003682199865579605, + 0.0024555560667067766, + 0.005086482036858797, + 0.004159289877861738, + 0.002246837830170989, + 0.004173719324171543, + 0.002891832496970892, + 0.003180492902174592, + 0.00298280268907547, + 0.00205416651442647, + 0.005762980319559574, + 0.003244170919060707, + 0.0018065215554088354, + 0.003451777156442404, + 0.0024017351679503918, + 0.003954305779188871 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.029849810525774956 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a chaotic room viewed from a static, high-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room filled with piles of academic books, old magazines, and discarded clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure hunched over a low wooden stool wearing a faded black hoodie with long dark hair obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a simple white Styrofoam cup of instant noodles with thick white steam rising from it visible on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure use disposable chopsticks to slowly and deliberately lift a portion of noodles from the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003627982921898365 + ], + "opposite_logprob_tokens": [ + -5.628627777099609 + ], + "logprob_reward": -0.003627982921898365, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003627982921898365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31519562005996704, + 0.17135192453861237, + 0.08843319118022919, + 0.16351695358753204, + 0.028848422691226006, + 0.08963331580162048, + 0.05343877524137497, + 0.019825313240289688, + 0.02818339131772518, + 0.041573043912649155 + ], + "credit_map": [ + 0.03776869550347328, + 0.03014315478503704, + 0.034441571682691574, + 0.05845660716295242, + 0.04355233535170555, + 0.04823128134012222, + 0.029866456985473633, + 0.03273552656173706, + 0.018536774441599846, + 0.015078604221343994, + 0.02297970838844776, + 0.03216313198208809, + 0.021419493481516838, + 0.026105690747499466, + 0.01854259893298149, + 0.016525909304618835, + 0.01537824235856533, + 0.007009163964539766, + 0.01029425859451294, + 0.01669141836464405, + 0.010329426266252995, + 0.01069569680839777, + 0.008934710174798965, + 0.009100274182856083, + 0.015223626978695393, + 0.01353753637522459, + 0.018822183832526207, + 0.03311816230416298, + 0.02489851973950863, + 0.02715235948562622, + 0.0168319009244442, + 0.013932676054537296, + 0.004062417894601822, + 0.0029001522343605757, + 0.003506777575239539, + 0.0042123571038246155, + 0.002767637837678194, + 0.004671960603445768, + 0.002440483309328556, + 0.004286637529730797, + 0.013166525401175022, + 0.011522693559527397, + 0.011281875893473625, + 0.014303818345069885, + 0.009339912794530392, + 0.013427816331386566, + 0.007033885456621647, + 0.009556788019835949, + 0.006210595369338989, + 0.004366753622889519, + 0.006668268237262964, + 0.010223941877484322, + 0.005821433383971453, + 0.009852197021245956, + 0.004526661243289709, + 0.005768920760601759, + 0.002806396223604679, + 0.0017778360052034259, + 0.0026210914365947247, + 0.003646804951131344, + 0.002209418686106801, + 0.002779280999675393, + 0.0017028695438057184, + 0.002281616907566786, + 0.003090793499723077, + 0.0021931538358330727, + 0.00424985634163022, + 0.004336474929004908, + 0.002302910201251507, + 0.005201538093388081, + 0.0030615476425737143, + 0.0037471165414899588, + 0.00452874880284071, + 0.0036534760147333145, + 0.005182571709156036, + 0.007879028096795082, + 0.0033671921119093895, + 0.008562094531953335, + 0.004138344898819923, + 0.004261586349457502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.003627982921898365 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a chaotic room viewed from a static, high-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room filled with piles of academic books, old magazines, and discarded clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure hunched over a low wooden stool wearing a faded black hoodie with long dark hair obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a simple white Styrofoam cup of instant noodles with thick white steam rising from it visible on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure use disposable chopsticks to slowly and deliberately lift a portion of noodles from the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020579738542437553 + ], + "opposite_logprob_tokens": [ + -3.8955798149108887 + ], + "logprob_reward": -0.020579738542437553, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020579738542437553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21743443608283997, + 0.24667617678642273, + 0.09409207850694656, + 0.048215605318546295, + 0.06321512162685394, + 0.1835654377937317, + 0.026304781436920166, + 0.03291966766119003, + 0.05593674257397652, + 0.031639911234378815 + ], + "credit_map": [ + 0.020926810801029205, + 0.01955394819378853, + 0.027714258059859276, + 0.034809984266757965, + 0.03319418057799339, + 0.025440029799938202, + 0.026181992143392563, + 0.029613228514790535, + 0.0250293780118227, + 0.02212151698768139, + 0.030436450615525246, + 0.04925588518381119, + 0.031386665999889374, + 0.04174867644906044, + 0.026684464886784554, + 0.020013131201267242, + 0.01385935116559267, + 0.006397162564098835, + 0.016891559585928917, + 0.012632223777472973, + 0.006508974824100733, + 0.010820157825946808, + 0.008004060946404934, + 0.018978584557771683, + 0.010340261273086071, + 0.005155758000910282, + 0.006932614371180534, + 0.006881191395223141, + 0.0033726918045431376, + 0.0041396706365048885, + 0.005264968145638704, + 0.0061284517869353294, + 0.008055943064391613, + 0.00570432934910059, + 0.008186367340385914, + 0.009675118140876293, + 0.009746764786541462, + 0.007754948455840349, + 0.005732937250286341, + 0.008358718827366829, + 0.01828606426715851, + 0.019518662244081497, + 0.022459834814071655, + 0.037351690232753754, + 0.022206192836165428, + 0.030782479792833328, + 0.01764686033129692, + 0.015313653275370598, + 0.0024741438683122396, + 0.002553590340539813, + 0.003916850313544273, + 0.003379468573257327, + 0.0034805440809577703, + 0.003000810742378235, + 0.003212613519281149, + 0.004286758601665497, + 0.005683480761945248, + 0.0030465207528322935, + 0.00431235134601593, + 0.0048859198577702045, + 0.0019915217999368906, + 0.005767017137259245, + 0.0022400845773518085, + 0.004992774222046137, + 0.00668675173074007, + 0.004529063124209642, + 0.009937576949596405, + 0.010176664218306541, + 0.004220324102789164, + 0.010536141693592072, + 0.0038702746387571096, + 0.00597994402050972, + 0.004456287249922752, + 0.0030692405998706818, + 0.006692483555525541, + 0.00445746723562479, + 0.0018658875487744808, + 0.0034241825342178345, + 0.003289822954684496, + 0.0043845404870808125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.020579738542437553 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a chaotic room viewed from a static, high-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room filled with piles of academic books, old magazines, and discarded clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure hunched over a low wooden stool wearing a faded black hoodie with long dark hair obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a simple white Styrofoam cup of instant noodles with thick white steam rising from it visible on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure use disposable chopsticks to slowly and deliberately lift a portion of noodles from the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981880120933056 + ], + "opposite_logprob_tokens": [ + -3.5298187732696533 + ], + "logprob_reward": -0.02981880120933056, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981880120933056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3404220938682556, + 0.23803097009658813, + 0.026687579229474068, + 0.09228131175041199, + 0.026365453377366066, + 0.04482010751962662, + 0.051355812698602676, + 0.038350727409124374, + 0.06311523914337158, + 0.07857068628072739 + ], + "credit_map": [ + 0.04198969155550003, + 0.03307606279850006, + 0.0361294262111187, + 0.06133115664124489, + 0.04325753077864647, + 0.05537218227982521, + 0.03345448151230812, + 0.035811569541692734, + 0.018706506118178368, + 0.022676298394799232, + 0.030141053721308708, + 0.046189941465854645, + 0.03477779030799866, + 0.03565423935651779, + 0.02962208166718483, + 0.02026306465268135, + 0.003966154996305704, + 0.0024795588105916977, + 0.0029401942156255245, + 0.0035840128548443317, + 0.002688287291675806, + 0.003391544334590435, + 0.002846800722181797, + 0.0047910260036587715, + 0.009433525614440441, + 0.009131278842687607, + 0.012335531413555145, + 0.018682939931750298, + 0.014816351234912872, + 0.010465089231729507, + 0.009904724545776844, + 0.007511873263865709, + 0.003910215571522713, + 0.00261366693302989, + 0.003713107667863369, + 0.0038261795416474342, + 0.003674151375889778, + 0.0036276739556342363, + 0.002166610909625888, + 0.0028338467236608267, + 0.009014623239636421, + 0.004730985499918461, + 0.005036832299083471, + 0.007937107235193253, + 0.0037382806185632944, + 0.0064101070165634155, + 0.0044948142021894455, + 0.0034573599696159363, + 0.005667747929692268, + 0.00619744136929512, + 0.008698591031134129, + 0.0068034278228878975, + 0.004469224251806736, + 0.006070986390113831, + 0.00631896173581481, + 0.007129428442567587, + 0.0033815784845501184, + 0.002909855218604207, + 0.00584886409342289, + 0.006764959078282118, + 0.0035053982865065336, + 0.008312517777085304, + 0.0030338543001562357, + 0.004593701101839542, + 0.00972033105790615, + 0.005764209199696779, + 0.009258292615413666, + 0.008816254325211048, + 0.00661398796364665, + 0.005804977845400572, + 0.00879331212490797, + 0.008343870751559734, + 0.00998101569712162, + 0.006853100378066301, + 0.012467483058571815, + 0.011429695412516594, + 0.006812096107751131, + 0.00971993338316679, + 0.011205250397324562, + 0.010102116502821445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.02981880120933056 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle, wide shot establishes a panoramic view of a deeply chaotic room, steeped in a profound sense of quiet solitude. The primary light source is a bare-bulb floor lamp, casting a warm yellow-orange glow that creates long, dramatic shadows and accentuates the chaotic textures of the environment. Piles of academic books with frayed spines, stacks of old magazines, and discarded clothing create a chaotic mess on the floor and a nearby unmade bed. At the center of this chaos, a solitary figure hunches over a low wooden stool, their back arched in a pose of deep exhaustion or concentration. Their identity is completely obscured, their face obscured by a mass of long, dark, tangled hair and the deep hood of an oversized, faded black hoodie. Before them, on a small patch of clean floor, sits a simple white Styrofoam cup of instant noodles, illuminated by thick white steam rising steadily from it. The figure's pale hand slowly and deliberately lifts a portion of noodles with a pair of disposable chopsticks. This micro-movement suggests a weary, almost mechanical act of maintenance, creating a poignant and atmospheric scene of late-night isolation.", + "seed": 14161, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_15/index_9888_seed_14161_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2077091638930142, + "vlm_logprob": -0.2077091638930142, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.2077091638930142, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a chaotic room viewed from a static, high-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room filled with piles of academic books, old magazines, and discarded clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure hunched over a low wooden stool wearing a faded black hoodie with long dark hair obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a simple white Styrofoam cup of instant noodles with thick white steam rising from it visible on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure use disposable chopsticks to slowly and deliberately lift a portion of noodles from the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9174960851669312 + ], + "opposite_logprob_tokens": [ + -0.5424960851669312 + ], + "logprob_reward": -0.9174960851669312, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.9174960851669312, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.0972079336643219, + 0.0399043969810009, + 0.03847777843475342, + 0.023066304624080658, + 0.020713476464152336, + 0.016099827364087105, + 0.5503907799720764, + 0.09591139853000641, + 0.07031308114528656, + 0.04791510850191116 + ], + "credit_map": [ + 0.013456808403134346, + 0.015010847710072994, + 0.006808388512581587, + 0.01352212019264698, + 0.010294007137417793, + 0.013025493361055851, + 0.007650007493793964, + 0.01744026690721512, + 0.00425342284142971, + 0.004764151759445667, + 0.0013593508629128337, + 0.004397423006594181, + 0.0035416539758443832, + 0.009800284169614315, + 0.0018050221260637045, + 0.009983088821172714, + 0.0052977874875068665, + 0.008047341369092464, + 0.002768326783552766, + 0.003995800856500864, + 0.005375127773731947, + 0.005392914637923241, + 0.003049955703318119, + 0.004550527315586805, + 0.0026549287140369415, + 0.0031728078611195087, + 0.0011262226616963744, + 0.003968254663050175, + 0.0023908945731818676, + 0.004692844580858946, + 0.002079590456560254, + 0.002980761229991913, + 0.002551329554989934, + 0.0031439282465726137, + 0.001161866937763989, + 0.003113290760666132, + 0.002865426242351532, + 0.002929306123405695, + 0.002241370500996709, + 0.002706958679482341, + 0.002503017894923687, + 0.0024940255098044872, + 0.0011786228278651834, + 0.001868337974883616, + 0.0018629776313900948, + 0.0024805678986012936, + 0.0017120882403105497, + 0.002000188920646906, + 0.16842134296894073, + 0.020275702700018883, + 0.011666766367852688, + 0.020228318870067596, + 0.014665879309177399, + 0.2255801111459732, + 0.008851198479533195, + 0.08070139586925507, + 0.028137480840086937, + 0.005066162906587124, + 0.0027491687797009945, + 0.004698180593550205, + 0.004179876763373613, + 0.03549515828490257, + 0.002904231660068035, + 0.01268113311380148, + 0.019530408084392548, + 0.004129643552005291, + 0.0022302046418190002, + 0.0061911544762551785, + 0.0034218751825392246, + 0.006100601051002741, + 0.0027318140491843224, + 0.025977376848459244, + 0.0067849005572497845, + 0.006962071638554335, + 0.004806789103895426, + 0.0037024642806500196, + 0.00345643050968647, + 0.011965017765760422, + 0.003560240613296628, + 0.006677193101495504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.9174960851669312 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a chaotic room viewed from a static, high-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room filled with piles of academic books, old magazines, and discarded clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure hunched over a low wooden stool wearing a faded black hoodie with long dark hair obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a simple white Styrofoam cup of instant noodles with thick white steam rising from it visible on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure use disposable chopsticks to slowly and deliberately lift a portion of noodles from the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05500856786966324 + ], + "opposite_logprob_tokens": [ + -2.93000864982605 + ], + "logprob_reward": -0.05500856786966324, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05500856786966324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08806632459163666, + 0.08774080127477646, + 0.05072938650846481, + 0.02898481674492359, + 0.02046525850892067, + 0.015123131684958935, + 0.4947943687438965, + 0.11830239742994308, + 0.07238245755434036, + 0.023411083966493607 + ], + "credit_map": [ + 0.017298530787229538, + 0.010886781848967075, + 0.010038956999778748, + 0.008545396849513054, + 0.00550930155441165, + 0.01701323315501213, + 0.003452288918197155, + 0.015321837738156319, + 0.01765897311270237, + 0.005543520674109459, + 0.0035896559711545706, + 0.0037605860270559788, + 0.002946635242551565, + 0.034297648817300797, + 0.0016762008890509605, + 0.01826758310198784, + 0.01159836258739233, + 0.0065460470505058765, + 0.002139318734407425, + 0.004460020922124386, + 0.004118897952139378, + 0.013764682225883007, + 0.002210038248449564, + 0.005892016459256411, + 0.005803391803056002, + 0.003985174931585789, + 0.0011471896432340145, + 0.0030132525134831667, + 0.002209926024079323, + 0.00788514781743288, + 0.0014872000319883227, + 0.0034535343293100595, + 0.0041495515033602715, + 0.0025030856486409903, + 0.0009955400601029396, + 0.0022626537829637527, + 0.0019923518411815166, + 0.004418873693794012, + 0.0015191402053460479, + 0.0026240607257932425, + 0.002063294406980276, + 0.0022575980983674526, + 0.001278280164115131, + 0.002055671764537692, + 0.001452191499993205, + 0.0022141295485198498, + 0.0015790988691151142, + 0.002222867915406823, + 0.15426281094551086, + 0.018651477992534637, + 0.010421785525977612, + 0.018338358029723167, + 0.013175290077924728, + 0.2018374800682068, + 0.008081300184130669, + 0.07002583146095276, + 0.03084191307425499, + 0.0058089811354875565, + 0.002938717370852828, + 0.006144041195511818, + 0.004127142485231161, + 0.045874614268541336, + 0.002789237769320607, + 0.019777750596404076, + 0.017727112397551537, + 0.0041293613612651825, + 0.0016488671535626054, + 0.005107496865093708, + 0.002693548332899809, + 0.010221472941339016, + 0.00238507566973567, + 0.028469517827033997, + 0.002664304804056883, + 0.004939100705087185, + 0.0016725803725421429, + 0.002298711333423853, + 0.0023322906345129013, + 0.0034187091514468193, + 0.002043061424046755, + 0.004042324144393206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.05500856786966324 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a chaotic room viewed from a static, high-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room filled with piles of academic books, old magazines, and discarded clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure hunched over a low wooden stool wearing a faded black hoodie with long dark hair obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a simple white Styrofoam cup of instant noodles with thick white steam rising from it visible on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure use disposable chopsticks to slowly and deliberately lift a portion of noodles from the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004645389039069414 + ], + "opposite_logprob_tokens": [ + -5.379645347595215 + ], + "logprob_reward": -0.004645389039069414, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004645389039069414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11292271316051483, + 0.06731736660003662, + 0.03141610324382782, + 0.009700313210487366, + 0.009552531875669956, + 0.011364137753844261, + 0.5930801033973694, + 0.10647300630807877, + 0.026693247258663177, + 0.03148047253489494 + ], + "credit_map": [ + 0.033041857182979584, + 0.010548189282417297, + 0.0047119115479290485, + 0.013984529301524162, + 0.007186906412243843, + 0.019736414775252342, + 0.004743929952383041, + 0.018968969583511353, + 0.020472027361392975, + 0.00400986522436142, + 0.0019972866866737604, + 0.0029498557560145855, + 0.0020223883911967278, + 0.026478705927729607, + 0.0012374004581943154, + 0.008149837143719196, + 0.0037960342597216368, + 0.004171581473201513, + 0.001485068816691637, + 0.0021626099478453398, + 0.002097843447700143, + 0.008962652646005154, + 0.001572705339640379, + 0.007167605683207512, + 0.001116483355872333, + 0.0015662249643355608, + 0.00048233367851935327, + 0.0011406937846913934, + 0.0007169035379774868, + 0.0020164449233561754, + 0.00044758475269190967, + 0.002213644329458475, + 0.0017846791306510568, + 0.0011241075117141008, + 0.0005777414189651608, + 0.000968955981079489, + 0.000773830630350858, + 0.001112916972488165, + 0.0008058748790062964, + 0.002404425060376525, + 0.0025835703127086163, + 0.0011324409861117601, + 0.0007367008365690708, + 0.0011278730817139149, + 0.0010168756125494838, + 0.0026028763968497515, + 0.0006336077931337059, + 0.001530192792415619, + 0.1801632046699524, + 0.021602565422654152, + 0.012954497709870338, + 0.0224703811109066, + 0.016363194212317467, + 0.24315588176250458, + 0.009861388243734837, + 0.08650901913642883, + 0.026584794744849205, + 0.004650976043194532, + 0.0034703200217336416, + 0.004917758982628584, + 0.003609428182244301, + 0.042269203811883926, + 0.0024830037727952003, + 0.01848752424120903, + 0.009406737983226776, + 0.0014409919967874885, + 0.0010939216008409858, + 0.002158229472115636, + 0.0010943567613139749, + 0.0017265811329707503, + 0.0012778312666341662, + 0.008494596928358078, + 0.003787544323131442, + 0.005162606481462717, + 0.003537425072863698, + 0.004285992123186588, + 0.0037499030586332083, + 0.003114697290584445, + 0.0030393667984753847, + 0.0048029376193881035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.004645389039069414 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a chaotic room viewed from a static, high-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room filled with piles of academic books, old magazines, and discarded clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure hunched over a low wooden stool wearing a faded black hoodie with long dark hair obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a simple white Styrofoam cup of instant noodles with thick white steam rising from it visible on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure use disposable chopsticks to slowly and deliberately lift a portion of noodles from the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328539453446865 + ], + "opposite_logprob_tokens": [ + -3.773285388946533 + ], + "logprob_reward": -0.02328539453446865, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328539453446865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13729311525821686, + 0.104680635035038, + 0.04535694047808647, + 0.01706823706626892, + 0.029738038778305054, + 0.030345385894179344, + 0.4347441792488098, + 0.10809274762868881, + 0.043785449117422104, + 0.04889519140124321 + ], + "credit_map": [ + 0.03883301466703415, + 0.006973470561206341, + 0.007153213489800692, + 0.006127257831394672, + 0.005423348862677813, + 0.051568031311035156, + 0.0032179250847548246, + 0.01799687184393406, + 0.012321251444518566, + 0.003660133806988597, + 0.003604741068556905, + 0.004723574500530958, + 0.003224398009479046, + 0.04192301258444786, + 0.0025055641308426857, + 0.03271796926856041, + 0.008484945632517338, + 0.006112078670412302, + 0.0021221742499619722, + 0.00363913387991488, + 0.0037878542207181454, + 0.012116368860006332, + 0.0029326188378036022, + 0.006161775439977646, + 0.004788403864949942, + 0.0016514342278242111, + 0.000826587958727032, + 0.001348631689324975, + 0.0012074491241946816, + 0.0026511724572628736, + 0.0012830101186409593, + 0.003311550710350275, + 0.005812203045934439, + 0.00552787771448493, + 0.0017268940573558211, + 0.00282306969165802, + 0.0023738578893244267, + 0.005867264233529568, + 0.001863192766904831, + 0.0037436841521412134, + 0.00666500348597765, + 0.0028932548593729734, + 0.0025295126251876354, + 0.002873486140742898, + 0.0017148394836112857, + 0.008735226467251778, + 0.0012396008241921663, + 0.003694467945024371, + 0.13432414829730988, + 0.01658141426742077, + 0.009864815510809422, + 0.016302982345223427, + 0.011611425317823887, + 0.17692694067955017, + 0.0071912421844899654, + 0.061941295862197876, + 0.032579243183135986, + 0.005831033922731876, + 0.0052053662948310375, + 0.006552064325660467, + 0.0042563434690237045, + 0.03688226640224457, + 0.003191281808540225, + 0.013595160096883774, + 0.01297981571406126, + 0.0026684042531996965, + 0.0027606734074652195, + 0.003356614848598838, + 0.0032199155539274216, + 0.0031629635486751795, + 0.0021929359063506126, + 0.01344413310289383, + 0.009547093883156776, + 0.0032019384671002626, + 0.0033521531149744987, + 0.004362956620752811, + 0.0034354948438704014, + 0.016109509393572807, + 0.0021811279002577066, + 0.006704926025122404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.02328539453446865 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a chaotic room viewed from a static, high-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room filled with piles of academic books, old magazines, and discarded clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure hunched over a low wooden stool wearing a faded black hoodie with long dark hair obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a simple white Styrofoam cup of instant noodles with thick white steam rising from it visible on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure use disposable chopsticks to slowly and deliberately lift a portion of noodles from the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811038285493851 + ], + "opposite_logprob_tokens": [ + -3.2881104946136475 + ], + "logprob_reward": -0.03811038285493851, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811038285493851, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13039277493953705, + 0.09705964475870132, + 0.01472460851073265, + 0.0129545908421278, + 0.009493129327893257, + 0.023468194529414177, + 0.15486685931682587, + 0.4309483766555786, + 0.09338466823101044, + 0.03270718827843666 + ], + "credit_map": [ + 0.03250116854906082, + 0.011365008540451527, + 0.006216621026396751, + 0.011052356101572514, + 0.007633860222995281, + 0.041247978806495667, + 0.0044815088622272015, + 0.015894273295998573, + 0.02134905382990837, + 0.004430979490280151, + 0.002373775700107217, + 0.0044629513286054134, + 0.0028099410701543093, + 0.04047653079032898, + 0.0020091375336050987, + 0.019147276878356934, + 0.0019730261992663145, + 0.0020427536219358444, + 0.001125277834944427, + 0.001913312473334372, + 0.0016570331063121557, + 0.002462009433656931, + 0.0010054195299744606, + 0.0025457763113081455, + 0.0022090154234319925, + 0.0016931003192439675, + 0.000745364639442414, + 0.0016759077552706003, + 0.0008972134673967957, + 0.001833783695474267, + 0.0012216820614412427, + 0.002678523538634181, + 0.0013112041633576155, + 0.00203447206877172, + 0.000929424655623734, + 0.0012358528329059482, + 0.0009934019763022661, + 0.0010099251521751285, + 0.0006068432121537626, + 0.001372005557641387, + 0.003203079802915454, + 0.003037245012819767, + 0.0021931270603090525, + 0.003337879665195942, + 0.0029946411959826946, + 0.0023131328634917736, + 0.002629147144034505, + 0.003759941319003701, + 0.044207487255334854, + 0.007596811279654503, + 0.005094071850180626, + 0.006158705800771713, + 0.00460291001945734, + 0.06054292991757393, + 0.0037147223483771086, + 0.022949213162064552, + 0.12582513689994812, + 0.02127343788743019, + 0.011669794097542763, + 0.019650494679808617, + 0.015319897793233395, + 0.16543066501617432, + 0.010759645141661167, + 0.06101932376623154, + 0.027366159483790398, + 0.004645608365535736, + 0.002892449963837862, + 0.006562885362654924, + 0.004277004394680262, + 0.008480122312903404, + 0.0032873216550797224, + 0.035873107612133026, + 0.004870404489338398, + 0.002457455964758992, + 0.001993236131966114, + 0.005006154999136925, + 0.0022131698206067085, + 0.008248824626207352, + 0.003220594022423029, + 0.004697350785136223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.03811038285493851 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle, wide shot establishes a panoramic view of a deeply chaotic room, steeped in a profound sense of quiet solitude. The primary light source is a bare-bulb floor lamp, casting a warm yellow-orange glow that creates long, dramatic shadows and accentuates the chaotic textures of the environment. Piles of academic books with frayed spines, stacks of old magazines, and discarded clothing create a chaotic mess on the floor and a nearby unmade bed. At the center of this chaos, a solitary figure hunches over a low wooden stool, their back arched in a pose of deep exhaustion or concentration. Their identity is completely obscured, their face obscured by a mass of long, dark, tangled hair and the deep hood of an oversized, faded black hoodie. Before them, on a small patch of clean floor, sits a simple white Styrofoam cup of instant noodles, illuminated by thick white steam rising steadily from it. The figure's pale hand slowly and deliberately lifts a portion of noodles with a pair of disposable chopsticks. This micro-movement suggests a weary, almost mechanical act of maintenance, creating a poignant and atmospheric scene of late-night isolation.", + "seed": 14162, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_15/index_9888_seed_14162_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.33851809818297623, + "vlm_logprob": -0.33851809818297623, + "vlm_yes_no_margin": 2.775, + "vlm_token_credit": -0.33851809818297623, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a chaotic room viewed from a static, high-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room filled with piles of academic books, old magazines, and discarded clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure hunched over a low wooden stool wearing a faded black hoodie with long dark hair obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a simple white Styrofoam cup of instant noodles with thick white steam rising from it visible on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure use disposable chopsticks to slowly and deliberately lift a portion of noodles from the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.5120556354522705 + ], + "opposite_logprob_tokens": [ + -0.2620556950569153 + ], + "logprob_reward": -1.5120556354522705, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.5120556354522705, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3515303134918213, + 0.11280784010887146, + 0.03300580382347107, + 0.11890105903148651, + 0.1315746009349823, + 0.08084037154912949, + 0.053031448274850845, + 0.027433572337031364, + 0.044663313776254654, + 0.04621165990829468 + ], + "credit_map": [ + 0.03795167803764343, + 0.05535523593425751, + 0.04257907345890999, + 0.0492851622402668, + 0.03851168230175972, + 0.05093705654144287, + 0.03303124010562897, + 0.043879225850105286, + 0.01077762246131897, + 0.018454590812325478, + 0.012603678740561008, + 0.015983043238520622, + 0.014135777950286865, + 0.018510326743125916, + 0.00883596483618021, + 0.013506841845810413, + 0.003275760682299733, + 0.004214664921164513, + 0.0038905232213437557, + 0.005421678069978952, + 0.003434533253312111, + 0.005747408140450716, + 0.0028281232807785273, + 0.004193114582449198, + 0.010825345292687416, + 0.013332730159163475, + 0.009115231223404408, + 0.01880098134279251, + 0.011102158576250076, + 0.02015513926744461, + 0.013023522682487965, + 0.022545956075191498, + 0.014173814095556736, + 0.02363421581685543, + 0.012560037896037102, + 0.0181538425385952, + 0.015535574406385422, + 0.012205066159367561, + 0.019318727776408195, + 0.015993338078260422, + 0.006542677525430918, + 0.016165314242243767, + 0.010870196856558323, + 0.011444061994552612, + 0.011305801570415497, + 0.00932814460247755, + 0.005047038663178682, + 0.010137136094272137, + 0.0036854243371635675, + 0.008133267052471638, + 0.0058256108313798904, + 0.009037625044584274, + 0.005239829421043396, + 0.009540100581943989, + 0.004358361475169659, + 0.007211230229586363, + 0.0015904627507552505, + 0.005052065011113882, + 0.0028208657167851925, + 0.004031042102724314, + 0.0027728003915399313, + 0.004606356844305992, + 0.002818040084093809, + 0.003741941647604108, + 0.0034084594808518887, + 0.007396022789180279, + 0.004699985962361097, + 0.007454669568687677, + 0.005939852446317673, + 0.006000191438943148, + 0.003907755482941866, + 0.005856380332261324, + 0.005025269463658333, + 0.005144163034856319, + 0.005029570311307907, + 0.007118069566786289, + 0.005740421824157238, + 0.008497877046465874, + 0.0037368808407336473, + 0.005919411778450012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -1.5120556354522705 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a chaotic room viewed from a static, high-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room filled with piles of academic books, old magazines, and discarded clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure hunched over a low wooden stool wearing a faded black hoodie with long dark hair obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a simple white Styrofoam cup of instant noodles with thick white steam rising from it visible on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure use disposable chopsticks to slowly and deliberately lift a portion of noodles from the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12712238729000092 + ], + "opposite_logprob_tokens": [ + -2.127122402191162 + ], + "logprob_reward": -0.12712238729000092, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12712238729000092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32153061032295227, + 0.11004524677991867, + 0.039008572697639465, + 0.08861392736434937, + 0.16411824524402618, + 0.03407379612326622, + 0.09309939295053482, + 0.04181399941444397, + 0.062075935304164886, + 0.04562024027109146 + ], + "credit_map": [ + 0.02122277393937111, + 0.040682755410671234, + 0.037843357771635056, + 0.05335218086838722, + 0.035338833928108215, + 0.0679406076669693, + 0.02238425239920616, + 0.04276587441563606, + 0.00769202783703804, + 0.015377892181277275, + 0.00997228268533945, + 0.018983019515872, + 0.012119000777602196, + 0.02492750994861126, + 0.007906259968876839, + 0.01306726224720478, + 0.0026104331482201815, + 0.006314436439424753, + 0.0052407607436180115, + 0.005828238558024168, + 0.0037966708187013865, + 0.0067170532420277596, + 0.0028896883595734835, + 0.005611298605799675, + 0.006575564853847027, + 0.009938487783074379, + 0.006296765990555286, + 0.013645726256072521, + 0.008832253515720367, + 0.01815790683031082, + 0.010014310479164124, + 0.015152914449572563, + 0.021229524165391922, + 0.030217178165912628, + 0.015350212343037128, + 0.021998794749379158, + 0.01727631501853466, + 0.02157166600227356, + 0.01657005213201046, + 0.019904518499970436, + 0.0027901979628950357, + 0.005307940766215324, + 0.0033625555224716663, + 0.005739037878811359, + 0.0031428635120391846, + 0.005823976825922728, + 0.003319359850138426, + 0.004587867762893438, + 0.008105838671326637, + 0.016438834369182587, + 0.009702732786536217, + 0.016016749665141106, + 0.008972665295004845, + 0.010504444129765034, + 0.011038871482014656, + 0.012319264002144337, + 0.0024300331715494394, + 0.005341392941772938, + 0.004867109470069408, + 0.006169380620121956, + 0.00544095691293478, + 0.006563300732523203, + 0.004234577063471079, + 0.0067672510631382465, + 0.006194666959345341, + 0.010313245467841625, + 0.009001411497592926, + 0.009730079211294651, + 0.006274511106312275, + 0.007965912111103535, + 0.005374618340283632, + 0.007221495267003775, + 0.004474116023629904, + 0.007479521445930004, + 0.0053540789522230625, + 0.007325237151235342, + 0.005975904408842325, + 0.0064163911156356335, + 0.0030841061379760504, + 0.005510885734111071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.12712238729000092 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a chaotic room viewed from a static, high-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room filled with piles of academic books, old magazines, and discarded clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure hunched over a low wooden stool wearing a faded black hoodie with long dark hair obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a simple white Styrofoam cup of instant noodles with thick white steam rising from it visible on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure use disposable chopsticks to slowly and deliberately lift a portion of noodles from the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036259638145565987 + ], + "opposite_logprob_tokens": [ + -5.628625869750977 + ], + "logprob_reward": -0.0036259638145565987, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036259638145565987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22201915085315704, + 0.10584919154644012, + 0.0404822863638401, + 0.1627599149942398, + 0.11292609572410583, + 0.13463540375232697, + 0.04591476172208786, + 0.06238938868045807, + 0.0706401839852333, + 0.04238360747694969 + ], + "credit_map": [ + 0.013339495286345482, + 0.03385891765356064, + 0.025468863546848297, + 0.03455754742026329, + 0.024333810433745384, + 0.0459979884326458, + 0.016740435734391212, + 0.027722107246518135, + 0.007400908973067999, + 0.012399895116686821, + 0.01233252976089716, + 0.02208033949136734, + 0.010315310209989548, + 0.020769890397787094, + 0.007664963603019714, + 0.01288535911589861, + 0.004130937624722719, + 0.0065366593189537525, + 0.004430677276104689, + 0.0057612219825387, + 0.00439229654148221, + 0.0058985017240047455, + 0.004045907873660326, + 0.005286089144647121, + 0.014178252778947353, + 0.030947836115956306, + 0.018880844116210938, + 0.022448355332016945, + 0.01670222356915474, + 0.020033948123455048, + 0.013670295476913452, + 0.025898177176713943, + 0.011245760135352612, + 0.015926141291856766, + 0.011554259806871414, + 0.015639539808034897, + 0.011157963424921036, + 0.017167197540402412, + 0.013817353174090385, + 0.016417887061834335, + 0.010150658898055553, + 0.02612455189228058, + 0.01830195263028145, + 0.018884574994444847, + 0.018418364226818085, + 0.015425601042807102, + 0.009008951485157013, + 0.01832076720893383, + 0.0032655629329383373, + 0.005612365435808897, + 0.0047677503898739815, + 0.006316297221928835, + 0.006416668649762869, + 0.009037487208843231, + 0.0038768851663917303, + 0.0066217463463544846, + 0.005442233290523291, + 0.010837244801223278, + 0.007761951070278883, + 0.009120747447013855, + 0.006363740190863609, + 0.007550379727035761, + 0.006700112484395504, + 0.0086129829287529, + 0.007047683000564575, + 0.012636872939765453, + 0.009176882915198803, + 0.011865273118019104, + 0.006931751035153866, + 0.008697970770299435, + 0.005940691102296114, + 0.008343066088855267, + 0.004729892127215862, + 0.005939265247434378, + 0.00578963290899992, + 0.006370407063513994, + 0.004375476855784655, + 0.007032365538179874, + 0.0029532278422266245, + 0.0051933457143604755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.0036259638145565987 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a chaotic room viewed from a static, high-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room filled with piles of academic books, old magazines, and discarded clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure hunched over a low wooden stool wearing a faded black hoodie with long dark hair obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a simple white Styrofoam cup of instant noodles with thick white steam rising from it visible on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure use disposable chopsticks to slowly and deliberately lift a portion of noodles from the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606731116771698 + ], + "opposite_logprob_tokens": [ + -4.1410675048828125 + ], + "logprob_reward": -0.01606731116771698, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606731116771698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16857242584228516, + 0.12250158935785294, + 0.05353376269340515, + 0.18228571116924286, + 0.23490984737873077, + 0.08551057428121567, + 0.05990603566169739, + 0.043675292283296585, + 0.027542972937226295, + 0.021561872214078903 + ], + "credit_map": [ + 0.015273689292371273, + 0.03196874260902405, + 0.025625010952353477, + 0.023457448929548264, + 0.01637325994670391, + 0.025596467778086662, + 0.015043153427541256, + 0.015234663151204586, + 0.004859501961618662, + 0.013695395551621914, + 0.008292337879538536, + 0.029153965413570404, + 0.010698786936700344, + 0.033978160470724106, + 0.00788733921945095, + 0.013936104252934456, + 0.005474328529089689, + 0.007316266186535358, + 0.006903565488755703, + 0.0074852569960057735, + 0.005345827899873257, + 0.009031499736011028, + 0.005682035349309444, + 0.006294984370470047, + 0.017765503376722336, + 0.01239048596471548, + 0.015434722416102886, + 0.029819205403327942, + 0.013238444924354553, + 0.04067371040582657, + 0.01504463516175747, + 0.037918996065855026, + 0.022656505927443504, + 0.035301532596349716, + 0.0222709309309721, + 0.03245155140757561, + 0.01981973648071289, + 0.039104610681533813, + 0.02782556414604187, + 0.03547941893339157, + 0.010971052572131157, + 0.009376968257129192, + 0.007884816266596317, + 0.013142668642103672, + 0.006424186751246452, + 0.012182720936834812, + 0.013486713171005249, + 0.012041445821523666, + 0.004999796859920025, + 0.009499636478722095, + 0.006920131854712963, + 0.007896683178842068, + 0.008386239409446716, + 0.008531630039215088, + 0.005856354255229235, + 0.007815558463335037, + 0.003828538116067648, + 0.004972672555595636, + 0.0049046543426811695, + 0.006732763256877661, + 0.0041219270788133144, + 0.006383076310157776, + 0.00557356234639883, + 0.007158098742365837, + 0.0027029921766370535, + 0.004454281646758318, + 0.0035507145803421736, + 0.004385726992040873, + 0.0026386850513517857, + 0.003576514543965459, + 0.002548566786572337, + 0.0036854883655905724, + 0.0016311912331730127, + 0.0031644508708268404, + 0.0028620215598493814, + 0.0033507670741528273, + 0.0018850571941584349, + 0.0036091357469558716, + 0.001602762844413519, + 0.0034564868547022343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.01606731116771698 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a chaotic room viewed from a static, high-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room filled with piles of academic books, old magazines, and discarded clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure hunched over a low wooden stool wearing a faded black hoodie with long dark hair obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a simple white Styrofoam cup of instant noodles with thick white steam rising from it visible on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure use disposable chopsticks to slowly and deliberately lift a portion of noodles from the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371919319033623 + ], + "opposite_logprob_tokens": [ + -3.408719301223755 + ], + "logprob_reward": -0.03371919319033623, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371919319033623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14871488511562347, + 0.07845032960176468, + 0.060637589544057846, + 0.12046132236719131, + 0.34753188490867615, + 0.060189034789800644, + 0.0710129365324974, + 0.060145795345306396, + 0.020843159407377243, + 0.03201300650835037 + ], + "credit_map": [ + 0.012895412743091583, + 0.03207184001803398, + 0.019356489181518555, + 0.02092621847987175, + 0.02086191065609455, + 0.016638515517115593, + 0.01127060316503048, + 0.01469389908015728, + 0.0066915638744831085, + 0.008210218511521816, + 0.007997408509254456, + 0.014709397219121456, + 0.006594737991690636, + 0.0162404403090477, + 0.006846561562269926, + 0.011160008609294891, + 0.006447272840887308, + 0.0092732859775424, + 0.006468529347330332, + 0.00885444413870573, + 0.0046462807804346085, + 0.009665211662650108, + 0.007102207746356726, + 0.008180360309779644, + 0.011259002611041069, + 0.015352043323218822, + 0.009954173117876053, + 0.01936313509941101, + 0.011116414330899715, + 0.019168520346283913, + 0.011186100542545319, + 0.02306193858385086, + 0.04940420761704445, + 0.06415761262178421, + 0.03330054506659508, + 0.04542848467826843, + 0.03648431599140167, + 0.03726520761847496, + 0.04443439096212387, + 0.03705712407827377, + 0.005220352206379175, + 0.008342171087861061, + 0.00559230288490653, + 0.010500052943825722, + 0.005971875973045826, + 0.008733348920941353, + 0.008324194699525833, + 0.007504728157073259, + 0.00651667220517993, + 0.013834472745656967, + 0.007571191992610693, + 0.011041799560189247, + 0.009036256931722164, + 0.006604154594242573, + 0.0072173853404819965, + 0.009191006422042847, + 0.005830835551023483, + 0.011997401714324951, + 0.006511372048407793, + 0.009037686511874199, + 0.006205385085195303, + 0.0058944085612893105, + 0.007672823499888182, + 0.006995882373303175, + 0.0014741759514436126, + 0.0037361946888267994, + 0.0022108026314526796, + 0.003569735214114189, + 0.0024381240364164114, + 0.00257642543874681, + 0.0017490522004663944, + 0.003088648896664381, + 0.002401441801339388, + 0.006060109008103609, + 0.004152845591306686, + 0.004852760583162308, + 0.0038977335207164288, + 0.004510771948844194, + 0.002123223152011633, + 0.004014119505882263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.03371919319033623 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle, wide shot establishes a panoramic view of a deeply chaotic room, steeped in a profound sense of quiet solitude. The primary light source is a bare-bulb floor lamp, casting a warm yellow-orange glow that creates long, dramatic shadows and accentuates the chaotic textures of the environment. Piles of academic books with frayed spines, stacks of old magazines, and discarded clothing create a chaotic mess on the floor and a nearby unmade bed. At the center of this chaos, a solitary figure hunches over a low wooden stool, their back arched in a pose of deep exhaustion or concentration. Their identity is completely obscured, their face obscured by a mass of long, dark, tangled hair and the deep hood of an oversized, faded black hoodie. Before them, on a small patch of clean floor, sits a simple white Styrofoam cup of instant noodles, illuminated by thick white steam rising steadily from it. The figure's pale hand slowly and deliberately lifts a portion of noodles with a pair of disposable chopsticks. This micro-movement suggests a weary, almost mechanical act of maintenance, creating a poignant and atmospheric scene of late-night isolation.", + "seed": 14163, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_15/index_9888_seed_14163_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.24294398464262484, + "vlm_logprob": -0.24294398464262484, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.24294398464262484, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a chaotic room viewed from a static, high-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room filled with piles of academic books, old magazines, and discarded clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure hunched over a low wooden stool wearing a faded black hoodie with long dark hair obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a simple white Styrofoam cup of instant noodles with thick white steam rising from it visible on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure use disposable chopsticks to slowly and deliberately lift a portion of noodles from the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0701359510421753 + ], + "opposite_logprob_tokens": [ + -0.4451359510421753 + ], + "logprob_reward": -1.0701359510421753, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0701359510421753, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.24355821311473846, + 0.113806813955307, + 0.06890444457530975, + 0.20018021762371063, + 0.041972946375608444, + 0.0469447560608387, + 0.05570533126592636, + 0.03531033545732498, + 0.1541052907705307, + 0.039511680603027344 + ], + "credit_map": [ + 0.035170406103134155, + 0.031413257122039795, + 0.03273170441389084, + 0.02356821484863758, + 0.02072400413453579, + 0.015666116029024124, + 0.034316401928663254, + 0.04996810480952263, + 0.023328877985477448, + 0.022910909727215767, + 0.015320367179811, + 0.009952728636562824, + 0.0102124297991395, + 0.00931000616401434, + 0.008959632366895676, + 0.013811858370900154, + 0.018495310097932816, + 0.020219990983605385, + 0.006468994542956352, + 0.0057029202580451965, + 0.005295448936522007, + 0.003155481768772006, + 0.004456385970115662, + 0.005109909921884537, + 0.08059671521186829, + 0.0817093476653099, + 0.006456345785409212, + 0.010107260197401047, + 0.006743514910340309, + 0.00642739562317729, + 0.004113867878913879, + 0.004025761038064957, + 0.008099867030978203, + 0.006402608472853899, + 0.005249123089015484, + 0.0035528549924492836, + 0.004534026607871056, + 0.00301998108625412, + 0.00456305081024766, + 0.006551430094987154, + 0.00850001908838749, + 0.007021244615316391, + 0.006424514576792717, + 0.005175712984055281, + 0.006309280637651682, + 0.003022589022293687, + 0.0043569584377110004, + 0.006134436000138521, + 0.008779126219451427, + 0.006504932884126902, + 0.007590532302856445, + 0.006080314517021179, + 0.008765553124248981, + 0.004604082088917494, + 0.006344985216856003, + 0.007035804912447929, + 0.0044800760224461555, + 0.004522046539932489, + 0.003956791944801807, + 0.003813557792454958, + 0.0036948244087398052, + 0.0029893950559198856, + 0.004694818519055843, + 0.007158822845667601, + 0.025044647976756096, + 0.024171555414795876, + 0.02122390829026699, + 0.024131717160344124, + 0.014682533219456673, + 0.012417130172252655, + 0.015785612165927887, + 0.0166481826454401, + 0.010783769190311432, + 0.010784301906824112, + 0.0025137807242572308, + 0.0028050232212990522, + 0.0028511113487184048, + 0.0032088514417409897, + 0.0024195825681090355, + 0.0041452571749687195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -1.0701359510421753 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a chaotic room viewed from a static, high-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room filled with piles of academic books, old magazines, and discarded clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure hunched over a low wooden stool wearing a faded black hoodie with long dark hair obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a simple white Styrofoam cup of instant noodles with thick white steam rising from it visible on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure use disposable chopsticks to slowly and deliberately lift a portion of noodles from the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0790286585688591 + ], + "opposite_logprob_tokens": [ + -2.579028606414795 + ], + "logprob_reward": -0.0790286585688591, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0790286585688591, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2887203097343445, + 0.10550590604543686, + 0.1952235847711563, + 0.14926281571388245, + 0.055149152874946594, + 0.0511111281812191, + 0.05184754356741905, + 0.02360198087990284, + 0.03106018714606762, + 0.048517484217882156 + ], + "credit_map": [ + 0.037336915731430054, + 0.0391586609184742, + 0.042272504419088364, + 0.03184503689408302, + 0.02780768647789955, + 0.025590764358639717, + 0.05252694711089134, + 0.03218178078532219, + 0.017732569947838783, + 0.018869886174798012, + 0.01493640337139368, + 0.01045627985149622, + 0.009478011168539524, + 0.009992106817662716, + 0.012022322975099087, + 0.012018308974802494, + 0.054407693445682526, + 0.05774673447012901, + 0.018533535301685333, + 0.017677810043096542, + 0.013150438666343689, + 0.009006476029753685, + 0.01259826309978962, + 0.012102616019546986, + 0.05718604475259781, + 0.05982619524002075, + 0.007015962153673172, + 0.007606377359479666, + 0.005502070300281048, + 0.005269235000014305, + 0.0034295646473765373, + 0.0034273371566087008, + 0.007623554207384586, + 0.006577969063073397, + 0.007538301404565573, + 0.006536399479955435, + 0.00560433603823185, + 0.007789463736116886, + 0.006231813691556454, + 0.0072473036125302315, + 0.007544570602476597, + 0.007065577432513237, + 0.00577806681394577, + 0.00514963548630476, + 0.0059499433264136314, + 0.003954208921641111, + 0.0064987363293766975, + 0.009170384146273136, + 0.006787081714719534, + 0.007189098279923201, + 0.006186343263834715, + 0.005244787782430649, + 0.004596197046339512, + 0.003733634715899825, + 0.008267989382147789, + 0.00984240509569645, + 0.003595965914428234, + 0.003086887300014496, + 0.0023298112209886312, + 0.0032712311949580908, + 0.002237798413261771, + 0.002397417090833187, + 0.0022483398206532, + 0.004434526432305574, + 0.005389836151152849, + 0.004270576871931553, + 0.004434788599610329, + 0.004344182554632425, + 0.002874996978789568, + 0.00396755151450634, + 0.0031979340128600597, + 0.002580317435786128, + 0.011856318451464176, + 0.011786958202719688, + 0.004139726050198078, + 0.004183746874332428, + 0.0038690492510795593, + 0.004186489153653383, + 0.004487412981688976, + 0.004007779527455568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.0790286585688591 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a chaotic room viewed from a static, high-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room filled with piles of academic books, old magazines, and discarded clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure hunched over a low wooden stool wearing a faded black hoodie with long dark hair obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a simple white Styrofoam cup of instant noodles with thick white steam rising from it visible on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure use disposable chopsticks to slowly and deliberately lift a portion of noodles from the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005954621359705925 + ], + "opposite_logprob_tokens": [ + -5.130954742431641 + ], + "logprob_reward": -0.005954621359705925, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005954621359705925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2393517941236496, + 0.12541328370571136, + 0.12041190266609192, + 0.09162313491106033, + 0.06466273218393326, + 0.030111711472272873, + 0.05870658531785011, + 0.03997338190674782, + 0.18722011148929596, + 0.042525339871644974 + ], + "credit_map": [ + 0.03433002904057503, + 0.028614401817321777, + 0.033121995627880096, + 0.026870137080550194, + 0.03278995677828789, + 0.028648968786001205, + 0.03151559457182884, + 0.023460689932107925, + 0.02324799634516239, + 0.02308262512087822, + 0.012501776218414307, + 0.010478166863322258, + 0.01567881926894188, + 0.016418293118476868, + 0.011126996017992496, + 0.012878603301942348, + 0.03622020035982132, + 0.037071313709020615, + 0.011706968769431114, + 0.010315566323697567, + 0.008280574344098568, + 0.005461180116981268, + 0.00577512988820672, + 0.005580979865044355, + 0.018377596512436867, + 0.016766779124736786, + 0.008874964900314808, + 0.00808841921389103, + 0.009809233248233795, + 0.007397409528493881, + 0.008302050642669201, + 0.014006681740283966, + 0.00934742484241724, + 0.009598546661436558, + 0.006081821396946907, + 0.005845693405717611, + 0.0066498382948338985, + 0.005112217739224434, + 0.008658227510750294, + 0.013368958607316017, + 0.004350034054368734, + 0.00412522628903389, + 0.003893290413543582, + 0.0031299663241952658, + 0.0028096467722207308, + 0.0025080558843910694, + 0.004812853876501322, + 0.004482637159526348, + 0.0076461005955934525, + 0.007247635163366795, + 0.005742088425904512, + 0.006033888552337885, + 0.006967822555452585, + 0.00676942290738225, + 0.0063282400369644165, + 0.01197138987481594, + 0.00466811191290617, + 0.004958665929734707, + 0.00485531659796834, + 0.004648657515645027, + 0.004361004568636417, + 0.005385252647101879, + 0.003946284297853708, + 0.007150084245949984, + 0.026379313319921494, + 0.027510488405823708, + 0.017594298347830772, + 0.019022108986973763, + 0.015758216381072998, + 0.014706314541399479, + 0.02179243601858616, + 0.044456932693719864, + 0.010189010761678219, + 0.010526472702622414, + 0.0035543395206332207, + 0.0033006498124450445, + 0.0038896326441317797, + 0.0028144819661974907, + 0.003264316590502858, + 0.004986434709280729 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.005954621359705925 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a chaotic room viewed from a static, high-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room filled with piles of academic books, old magazines, and discarded clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure hunched over a low wooden stool wearing a faded black hoodie with long dark hair obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a simple white Styrofoam cup of instant noodles with thick white steam rising from it visible on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure use disposable chopsticks to slowly and deliberately lift a portion of noodles from the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029790569096803665 + ], + "opposite_logprob_tokens": [ + -3.5297906398773193 + ], + "logprob_reward": -0.029790569096803665, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029790569096803665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22961647808551788, + 0.10486900061368942, + 0.2215937077999115, + 0.11911576986312866, + 0.07052755355834961, + 0.04937107115983963, + 0.03678181394934654, + 0.02464461512863636, + 0.06836892664432526, + 0.07511106133460999 + ], + "credit_map": [ + 0.03496271371841431, + 0.029811803251504898, + 0.034760817885398865, + 0.0306728295981884, + 0.018822599202394485, + 0.017281658947467804, + 0.030946554616093636, + 0.03235751390457153, + 0.020794784650206566, + 0.013713531196117401, + 0.01635342836380005, + 0.012661143206059933, + 0.010070273652672768, + 0.00725423963740468, + 0.01257294975221157, + 0.011448647826910019, + 0.03747916594147682, + 0.03602319210767746, + 0.03662063553929329, + 0.02930941991508007, + 0.017408203333616257, + 0.016505658626556396, + 0.027459759265184402, + 0.0207876805216074, + 0.016875676810741425, + 0.015197641216218472, + 0.01770034432411194, + 0.01778332144021988, + 0.012566585093736649, + 0.012517743743956089, + 0.010957533493638039, + 0.01551692746579647, + 0.012017177417874336, + 0.008880059234797955, + 0.010208459571003914, + 0.00825054943561554, + 0.006960605271160603, + 0.009511949494481087, + 0.00850099977105856, + 0.006197753828018904, + 0.008798354305326939, + 0.007772861048579216, + 0.007353618741035461, + 0.005781044717878103, + 0.004096589982509613, + 0.004384136758744717, + 0.006188558880239725, + 0.004995906725525856, + 0.007056730333715677, + 0.00568361347541213, + 0.0048910570330917835, + 0.0045943367294967175, + 0.00307493656873703, + 0.003531461814418435, + 0.004446629900485277, + 0.003503047628328204, + 0.003545515239238739, + 0.0027863322757184505, + 0.0036787926219403744, + 0.0031843483448028564, + 0.0020805643871426582, + 0.0028150370344519615, + 0.003150586737319827, + 0.003403438488021493, + 0.01148828025907278, + 0.009570777416229248, + 0.00832717027515173, + 0.010495233349502087, + 0.006902707275003195, + 0.008513228967785835, + 0.007436475250869989, + 0.005635054782032967, + 0.019658956676721573, + 0.020450275391340256, + 0.005138574633747339, + 0.00548687344416976, + 0.0055716210044920444, + 0.004518568515777588, + 0.005928623024374247, + 0.008357571437954903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.029790569096803665 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a chaotic room viewed from a static, high-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room filled with piles of academic books, old magazines, and discarded clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a solitary figure hunched over a low wooden stool wearing a faded black hoodie with long dark hair obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a simple white Styrofoam cup of instant noodles with thick white steam rising from it visible on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure use disposable chopsticks to slowly and deliberately lift a portion of noodles from the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029810123145580292 + ], + "opposite_logprob_tokens": [ + -3.5298101902008057 + ], + "logprob_reward": -0.029810123145580292, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029810123145580292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.208469420671463, + 0.1534600704908371, + 0.09422417730093002, + 0.16838668286800385, + 0.10036180913448334, + 0.06190602481365204, + 0.04598525911569595, + 0.055747129023075104, + 0.0600438117980957, + 0.0514155775308609 + ], + "credit_map": [ + 0.03528812527656555, + 0.027676403522491455, + 0.029597898945212364, + 0.02803056128323078, + 0.01984759233891964, + 0.01573859713971615, + 0.02027853950858116, + 0.03201170265674591, + 0.02425159327685833, + 0.02083718217909336, + 0.022066863253712654, + 0.02273845113813877, + 0.015080817975103855, + 0.016550099477171898, + 0.01780879497528076, + 0.014126269146800041, + 0.015163831412792206, + 0.015631409361958504, + 0.015990909188985825, + 0.01005618181079626, + 0.013777261599898338, + 0.008356978185474873, + 0.007571598049253225, + 0.007676006760448217, + 0.05982113257050514, + 0.05994858592748642, + 0.00892835482954979, + 0.008439642377197742, + 0.007069599349051714, + 0.006256879307329655, + 0.008950202725827694, + 0.008972282521426678, + 0.014509906992316246, + 0.012242208234965801, + 0.012165345251560211, + 0.010115250945091248, + 0.010429900139570236, + 0.011432835832238197, + 0.012139707803726196, + 0.017326654866337776, + 0.009575868956744671, + 0.00918665062636137, + 0.007768100127577782, + 0.00692502036690712, + 0.0062854173593223095, + 0.004439081996679306, + 0.007368951570242643, + 0.010356933809816837, + 0.006785646080970764, + 0.006444635335355997, + 0.005317375063896179, + 0.0047054593451321125, + 0.005200085695832968, + 0.0042245821096003056, + 0.006397908087819815, + 0.006909569259732962, + 0.008834311738610268, + 0.006986968219280243, + 0.005985150579363108, + 0.00448992196470499, + 0.005219514947384596, + 0.004851727746427059, + 0.006584161426872015, + 0.012795372866094112, + 0.011031189002096653, + 0.008290905505418777, + 0.009806279093027115, + 0.008846867829561234, + 0.004832570906728506, + 0.006353592034429312, + 0.005869862157851458, + 0.005012545268982649, + 0.012082519941031933, + 0.011335296556353569, + 0.007232262287288904, + 0.004998920951038599, + 0.004172675311565399, + 0.0031105801463127136, + 0.0036536389961838722, + 0.004829684738069773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.029810123145580292 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_2/index_tensor([4113, 4113])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_2/index_tensor([4113, 4113])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..687e47f812b8b20c240a5668fdf4ac6c05c02878 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_2/index_tensor([4113, 4113])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "strictly static, bird's-eye view, the shot is focused vertically on a set of weathered bluestone steps. The surface is etched with age, its edges smoothed by age, and wisps of tenacious dark green moss dot the crevices. Suddenly, a thick, dark brown terracotta jar plummets from above, its rough surface grazing the air as it falls. Upon contact with the solid stone steps, the jar disintegrates with a dull crack, shattering into countless terracotta fragments of varying sizes and sharp edges. Centered on the point of impact, these fragments, mixed with a small cloud of rising dust, fly in all directions at high speed, leaving a fleeting and chaotic scene of destruction on the cold stone surface. A harsh, sidelight slices through the scene, highlighting the ancient texture of the stone steps and capturing every fleeting reflection as the fragments fly, creating an atmosphere of tension, suddenness, and irreversibility.", + "seed": 11490, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_2/index_4113_seed_11490_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.11838417164981366, + "vlm_logprob": -0.11838417164981366, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.11838417164981366, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a bird's-eye view focused vertically on weathered bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, dark brown terracotta jar visible falling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jar disintegrate upon impact, shattering into countless fragments of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments fly in all directions at high speed, mixed with a small cloud of rising dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a harsh sidelight that highlights the stone texture and captures reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030887827277183533 + ], + "opposite_logprob_tokens": [ + -3.7808878421783447 + ], + "logprob_reward": -0.030887827277183533, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.030887827277183533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12989690899848938, + 0.05966642498970032, + 0.21090522408485413, + 0.03954562917351723, + 0.16480109095573425, + 0.05836614593863487, + 0.0879996120929718, + 0.06589045375585556, + 0.10861915349960327, + 0.07430937886238098 + ], + "credit_map": [ + 0.015937915071845055, + 0.016083741560578346, + 0.020068829879164696, + 0.018661588430404663, + 0.022137824445962906, + 0.012258653528988361, + 0.013987038284540176, + 0.010761312209069729, + 0.007484664209187031, + 0.006181453820317984, + 0.0056107016280293465, + 0.006833112798631191, + 0.012256362475454807, + 0.008456976152956486, + 0.008591643534600735, + 0.004251502454280853, + 0.021835435181856155, + 0.017830373719334602, + 0.025227287784218788, + 0.022490277886390686, + 0.03358156979084015, + 0.025844799354672432, + 0.03350753337144852, + 0.0305879395455122, + 0.004294832237064838, + 0.0036144121550023556, + 0.004848105367273092, + 0.0036526289768517017, + 0.005882736761122942, + 0.0058333734050393105, + 0.004551818128675222, + 0.006867721676826477, + 0.019930046051740646, + 0.01893921010196209, + 0.012667142786085606, + 0.014815445989370346, + 0.02188137359917164, + 0.026358185335993767, + 0.02180001139640808, + 0.02840965986251831, + 0.008490749634802341, + 0.005876138806343079, + 0.0053943004459142685, + 0.006641466170549393, + 0.005698349792510271, + 0.011860530823469162, + 0.006317985709756613, + 0.008086618036031723, + 0.01252696942538023, + 0.013051491230726242, + 0.007851769216358662, + 0.010196713730692863, + 0.010032105259597301, + 0.011202831752598286, + 0.013171154074370861, + 0.009966567158699036, + 0.006879271473735571, + 0.011284807696938515, + 0.005485005676746368, + 0.006236515007913113, + 0.007981506176292896, + 0.010310495272278786, + 0.00974917970597744, + 0.007963663898408413, + 0.01741250790655613, + 0.021127907559275627, + 0.00994656328111887, + 0.008825470693409443, + 0.010269592516124249, + 0.014891930855810642, + 0.011428832076489925, + 0.014716342091560364, + 0.011885174550116062, + 0.012793675996363163, + 0.007417142391204834, + 0.006782542914152145, + 0.007650298997759819, + 0.011661631055176258, + 0.00824255496263504, + 0.007876347750425339 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.030887827277183533 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a bird's-eye view focused vertically on weathered bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, dark brown terracotta jar visible falling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jar disintegrate upon impact, shattering into countless fragments of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments fly in all directions at high speed, mixed with a small cloud of rising dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a harsh sidelight that highlights the stone texture and captures reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026370810344815254 + ], + "opposite_logprob_tokens": [ + -3.6513707637786865 + ], + "logprob_reward": -0.026370810344815254, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026370810344815254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19407200813293457, + 0.15302453935146332, + 0.14534030854701996, + 0.06707042455673218, + 0.07032638788223267, + 0.06423650681972504, + 0.05590708553791046, + 0.06917287409305573, + 0.13577251136302948, + 0.04507732763886452 + ], + "credit_map": [ + 0.023634277284145355, + 0.02882835641503334, + 0.02295473963022232, + 0.016542134806513786, + 0.03647592291235924, + 0.024057108908891678, + 0.02664409764111042, + 0.014935365878045559, + 0.018418466672301292, + 0.011955959722399712, + 0.017230844125151634, + 0.02393290586769581, + 0.028198176994919777, + 0.01618719846010208, + 0.024205852299928665, + 0.012895136140286922, + 0.012914365157485008, + 0.01606939360499382, + 0.017902294173836708, + 0.017078489065170288, + 0.021474285051226616, + 0.0182440597563982, + 0.02093392051756382, + 0.020723508670926094, + 0.009683972224593163, + 0.007796321529895067, + 0.006640048231929541, + 0.009927280247211456, + 0.010143976658582687, + 0.0066872634924948215, + 0.008315927349030972, + 0.007875637151300907, + 0.0068075028248131275, + 0.006886585149914026, + 0.006575305946171284, + 0.010853271000087261, + 0.00777349341660738, + 0.0106995003297925, + 0.009414652362465858, + 0.01131607685238123, + 0.005480968859046698, + 0.006781858392059803, + 0.006905104964971542, + 0.009360137395560741, + 0.009321178309619427, + 0.009807847440242767, + 0.008980057202279568, + 0.007599347736686468, + 0.008641782216727734, + 0.006495651323348284, + 0.005147915333509445, + 0.00668921647593379, + 0.0074399556033313274, + 0.007677797693759203, + 0.007590086665004492, + 0.006224675569683313, + 0.007533298339694738, + 0.01022165734320879, + 0.006391889415681362, + 0.008199277333915234, + 0.008676713332533836, + 0.00941118411719799, + 0.011956421658396721, + 0.0067824325524270535, + 0.02146111987531185, + 0.021018266677856445, + 0.015578949823975563, + 0.012226455844938755, + 0.014919927343726158, + 0.01910606399178505, + 0.01576979085803032, + 0.015691934153437614, + 0.007222679443657398, + 0.007064148783683777, + 0.003933272324502468, + 0.003773526521399617, + 0.005455528851598501, + 0.0062883454374969006, + 0.006429194938391447, + 0.004910632502287626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.026370810344815254 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a bird's-eye view focused vertically on weathered bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, dark brown terracotta jar visible falling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jar disintegrate upon impact, shattering into countless fragments of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments fly in all directions at high speed, mixed with a small cloud of rising dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a harsh sidelight that highlights the stone texture and captures reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870271146297455 + ], + "opposite_logprob_tokens": [ + -1.1370271444320679 + ], + "logprob_reward": -0.3870271146297455, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870271146297455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18191944062709808, + 0.07973869144916534, + 0.05494224652647972, + 0.0830041840672493, + 0.12073703110218048, + 0.04879343882203102, + 0.08751492947340012, + 0.09467267990112305, + 0.1633426398038864, + 0.08533473312854767 + ], + "credit_map": [ + 0.027128219604492188, + 0.025631586089730263, + 0.025390837341547012, + 0.017266033217310905, + 0.030373288318514824, + 0.018718594685196877, + 0.02653639204800129, + 0.0108744902536273, + 0.01032201573252678, + 0.006643845699727535, + 0.00622660294175148, + 0.011352002620697021, + 0.014367490075528622, + 0.010030312463641167, + 0.013835536316037178, + 0.006960881408303976, + 0.008620286360383034, + 0.004293158650398254, + 0.007888630963861942, + 0.005933927837759256, + 0.006489892490208149, + 0.006067465990781784, + 0.0072830370627343655, + 0.008365843445062637, + 0.009707526303827763, + 0.008111005648970604, + 0.00716700404882431, + 0.013473398052155972, + 0.014280091971158981, + 0.007667872589081526, + 0.012980402447283268, + 0.009616885334253311, + 0.011835646815598011, + 0.010670078918337822, + 0.009514298290014267, + 0.012637626379728317, + 0.011651281267404556, + 0.024375319480895996, + 0.017218897119164467, + 0.02283388189971447, + 0.006624248344451189, + 0.00595804862678051, + 0.0034543436486274004, + 0.004899974912405014, + 0.006191145163029432, + 0.01029682345688343, + 0.006625540554523468, + 0.0047433143481612206, + 0.016580691561102867, + 0.01630263961851597, + 0.007590125780552626, + 0.009558484889566898, + 0.007973589934408665, + 0.009719910100102425, + 0.012699421495199203, + 0.007090069353580475, + 0.012876004911959171, + 0.014166162349283695, + 0.007651980500668287, + 0.009820421226322651, + 0.010124347172677517, + 0.014383327215909958, + 0.016218574717640877, + 0.009431862272322178, + 0.02357669733464718, + 0.03184986487030983, + 0.017847584560513496, + 0.013302089646458626, + 0.01641799882054329, + 0.020249929279088974, + 0.01659277267754078, + 0.023505698889493942, + 0.01285821944475174, + 0.00988998543471098, + 0.009158593602478504, + 0.007642828859388828, + 0.009577502496540546, + 0.015778034925460815, + 0.01001458428800106, + 0.010414978489279747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.3870271146297455 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a bird's-eye view focused vertically on weathered bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, dark brown terracotta jar visible falling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jar disintegrate upon impact, shattering into countless fragments of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments fly in all directions at high speed, mixed with a small cloud of rising dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a harsh sidelight that highlights the stone texture and captures reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058662846684456 + ], + "opposite_logprob_tokens": [ + -3.8955867290496826 + ], + "logprob_reward": -0.02058662846684456, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058662846684456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1313697099685669, + 0.07943976670503616, + 0.0795135572552681, + 0.1322535127401352, + 0.12442375719547272, + 0.0641876757144928, + 0.10842888802289963, + 0.061820246279239655, + 0.12670369446277618, + 0.09185919165611267 + ], + "credit_map": [ + 0.017550315707921982, + 0.016178037971258163, + 0.015753479674458504, + 0.009927622973918915, + 0.020859476178884506, + 0.015694964677095413, + 0.02340739592909813, + 0.011998416855931282, + 0.007662358693778515, + 0.007507754489779472, + 0.008196587674319744, + 0.009862931445240974, + 0.014634200371801853, + 0.010823464021086693, + 0.01344703696668148, + 0.007305440027266741, + 0.010004904121160507, + 0.006144159939140081, + 0.009792526252567768, + 0.008031733334064484, + 0.012746507301926613, + 0.009987812489271164, + 0.011662947945296764, + 0.011142969131469727, + 0.01899689808487892, + 0.014026612974703312, + 0.012262415140867233, + 0.017093487083911896, + 0.023142419755458832, + 0.012642209418118, + 0.018237566575407982, + 0.015851903706789017, + 0.013322646729648113, + 0.010462041944265366, + 0.010495222173631191, + 0.013066777028143406, + 0.015595203265547752, + 0.024472638964653015, + 0.01802212931215763, + 0.01898709498345852, + 0.007889066822826862, + 0.008196376264095306, + 0.0050733028911054134, + 0.005929756443947554, + 0.0074786534532904625, + 0.01271588820964098, + 0.009678889065980911, + 0.007225749082863331, + 0.021325336769223213, + 0.016309892758727074, + 0.011145311407744884, + 0.010703634470701218, + 0.011954917572438717, + 0.014034396968781948, + 0.012790863402187824, + 0.010164537467062473, + 0.009339308366179466, + 0.00904590543359518, + 0.004984494298696518, + 0.006391598377376795, + 0.007814235985279083, + 0.008430395275354385, + 0.009866626001894474, + 0.005947685334831476, + 0.016109174117445946, + 0.018582874909043312, + 0.019861124455928802, + 0.012958252802491188, + 0.016917681321501732, + 0.015006066299974918, + 0.013693657703697681, + 0.013574869371950626, + 0.010754345916211605, + 0.01543452125042677, + 0.008525766432285309, + 0.007688986137509346, + 0.011026313528418541, + 0.018600257113575935, + 0.010056338272988796, + 0.009772670455276966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.02058662846684456 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a bird's-eye view focused vertically on weathered bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, dark brown terracotta jar visible falling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jar disintegrate upon impact, shattering into countless fragments of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments fly in all directions at high speed, mixed with a small cloud of rising dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a harsh sidelight that highlights the stone texture and captures reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704847753047943 + ], + "opposite_logprob_tokens": [ + -2.1270484924316406 + ], + "logprob_reward": -0.12704847753047943, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12704847753047943, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1907126009464264, + 0.10527189075946808, + 0.059346504509449005, + 0.10618490725755692, + 0.07805133610963821, + 0.08202547580003738, + 0.11000752449035645, + 0.08207555115222931, + 0.12914977967739105, + 0.05717439576983452 + ], + "credit_map": [ + 0.017030691727995872, + 0.022001026198267937, + 0.022931993007659912, + 0.027435261756181717, + 0.024386797100305557, + 0.03925877809524536, + 0.021867364645004272, + 0.01580069772899151, + 0.008548753336071968, + 0.009437198750674725, + 0.01455220952630043, + 0.013707530684769154, + 0.01623542606830597, + 0.016611987724900246, + 0.016111785545945168, + 0.010067010298371315, + 0.009081726893782616, + 0.007613690569996834, + 0.005971905309706926, + 0.007279164623469114, + 0.006593071389943361, + 0.006098279729485512, + 0.009238614700734615, + 0.007470055483281612, + 0.015854742377996445, + 0.014926625415682793, + 0.011400914750993252, + 0.010106422007083893, + 0.017282191663980484, + 0.012744475156068802, + 0.011210317723453045, + 0.012659209780395031, + 0.010687222704291344, + 0.008085624314844608, + 0.006868144031614065, + 0.008195213042199612, + 0.008754503913223743, + 0.01411418616771698, + 0.009637946262955666, + 0.011708496138453484, + 0.013903138227760792, + 0.009733142331242561, + 0.007161553483456373, + 0.008267704397439957, + 0.008801274001598358, + 0.01301509514451027, + 0.010165881365537643, + 0.010977682657539845, + 0.02011951617896557, + 0.01624608226120472, + 0.00788396131247282, + 0.012338840402662754, + 0.0120232617482543, + 0.013473757542669773, + 0.01682506501674652, + 0.011097042821347713, + 0.01402793824672699, + 0.014803948812186718, + 0.007751428056508303, + 0.008460145443677902, + 0.006973103620111942, + 0.009424407966434956, + 0.011481544002890587, + 0.009153040125966072, + 0.01841728389263153, + 0.026154324412345886, + 0.013641376979649067, + 0.009575623087584972, + 0.011129837483167648, + 0.023085149005055428, + 0.01052854023873806, + 0.016617633402347565, + 0.008943328633904457, + 0.009220466017723083, + 0.006291045807301998, + 0.0057831150479614735, + 0.005822060164064169, + 0.007330153603106737, + 0.00738384248688817, + 0.00640038400888443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.12704847753047943 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "strictly static, bird's-eye view, the shot is focused vertically on a set of weathered bluestone steps. The surface is etched with age, its edges smoothed by age, and wisps of tenacious dark green moss dot the crevices. Suddenly, a thick, dark brown terracotta jar plummets from above, its rough surface grazing the air as it falls. Upon contact with the solid stone steps, the jar disintegrates with a dull crack, shattering into countless terracotta fragments of varying sizes and sharp edges. Centered on the point of impact, these fragments, mixed with a small cloud of rising dust, fly in all directions at high speed, leaving a fleeting and chaotic scene of destruction on the cold stone surface. A harsh, sidelight slices through the scene, highlighting the ancient texture of the stone steps and capturing every fleeting reflection as the fragments fly, creating an atmosphere of tension, suddenness, and irreversibility.", + "seed": 11491, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_2/index_4113_seed_11491_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07371232509613038, + "vlm_logprob": -0.07371232509613038, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.07371232509613038, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a bird's-eye view focused vertically on weathered bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, dark brown terracotta jar visible falling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jar disintegrate upon impact, shattering into countless fragments of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments fly in all directions at high speed, mixed with a small cloud of rising dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a harsh sidelight that highlights the stone texture and captures reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.031844284385442734 + ], + "opposite_logprob_tokens": [ + -3.7818443775177 + ], + "logprob_reward": -0.031844284385442734, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.031844284385442734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20634658634662628, + 0.049290165305137634, + 0.02445022203028202, + 0.031427692621946335, + 0.38597142696380615, + 0.04232485219836235, + 0.047058917582035065, + 0.15888172388076782, + 0.03848344087600708, + 0.015764916315674782 + ], + "credit_map": [ + 0.015151592902839184, + 0.020783089101314545, + 0.06646297872066498, + 0.0198198352009058, + 0.009056626819074154, + 0.054493095725774765, + 0.010967728681862354, + 0.00961165688931942, + 0.007344099227339029, + 0.0051763830706477165, + 0.006194762885570526, + 0.0038988583255559206, + 0.006083667743951082, + 0.007583620026707649, + 0.004767388105392456, + 0.008241391740739346, + 0.0037346112076193094, + 0.003922071307897568, + 0.00338233751244843, + 0.002412833971902728, + 0.0030034035444259644, + 0.0025551016442477703, + 0.0028949042316526175, + 0.002544959308579564, + 0.0037345460150390863, + 0.00300966901704669, + 0.0042937081307172775, + 0.0036220825277268887, + 0.006487981881946325, + 0.003673798870295286, + 0.0037758087273687124, + 0.002830099081620574, + 0.044938765466213226, + 0.014024992473423481, + 0.13299213349819183, + 0.020819416269659996, + 0.023279963061213493, + 0.12237153202295303, + 0.01317610777914524, + 0.014368539676070213, + 0.004600994288921356, + 0.006105121690779924, + 0.0059715439565479755, + 0.00547386659309268, + 0.005772900767624378, + 0.0037593250162899494, + 0.006809877697378397, + 0.003831226145848632, + 0.004676834214478731, + 0.007100685965269804, + 0.006719676312059164, + 0.009042328223586082, + 0.004663112573325634, + 0.003634209744632244, + 0.0045822723768651485, + 0.006639802362769842, + 0.012759427540004253, + 0.014148136600852013, + 0.021161314100027084, + 0.04591379314661026, + 0.01242194976657629, + 0.012353036552667618, + 0.02243795059621334, + 0.01768612489104271, + 0.00424701115116477, + 0.004755057394504547, + 0.006492673419415951, + 0.005839578341692686, + 0.003500232705846429, + 0.0031798668205738068, + 0.005232410039752722, + 0.0052366130985319614, + 0.0012473331298679113, + 0.001239354140125215, + 0.0021985573694109917, + 0.0038419717457145452, + 0.0013699885457754135, + 0.001748735667206347, + 0.0020934876520186663, + 0.0020254908595234156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.031844284385442734 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a bird's-eye view focused vertically on weathered bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, dark brown terracotta jar visible falling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jar disintegrate upon impact, shattering into countless fragments of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments fly in all directions at high speed, mixed with a small cloud of rising dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a harsh sidelight that highlights the stone texture and captures reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011089714244008064 + ], + "opposite_logprob_tokens": [ + -4.51108980178833 + ], + "logprob_reward": -0.011089714244008064, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011089714244008064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14797702431678772, + 0.10255035758018494, + 0.04342040419578552, + 0.13959458470344543, + 0.3116592764854431, + 0.03690851852297783, + 0.028268827125430107, + 0.1197909489274025, + 0.04748713597655296, + 0.022342998534440994 + ], + "credit_map": [ + 0.010573191568255424, + 0.014149979688227177, + 0.047783009707927704, + 0.014740871265530586, + 0.005811147391796112, + 0.03979803994297981, + 0.007717366795986891, + 0.00740340631455183, + 0.01830953173339367, + 0.008704821579158306, + 0.013846063986420631, + 0.008402494713664055, + 0.014026369899511337, + 0.01317303255200386, + 0.01074597891420126, + 0.015342051163315773, + 0.004727765917778015, + 0.00432001706212759, + 0.005513964220881462, + 0.00582108786329627, + 0.006535202264785767, + 0.004578982014209032, + 0.006474703084677458, + 0.005448677111417055, + 0.01645882986485958, + 0.011476232670247555, + 0.024052003398537636, + 0.01745334453880787, + 0.02163502387702465, + 0.015326932072639465, + 0.017016539350152016, + 0.016175655648112297, + 0.03605924919247627, + 0.01023860089480877, + 0.10746113955974579, + 0.01675114780664444, + 0.018830789253115654, + 0.10013407468795776, + 0.010597101412713528, + 0.011587142013013363, + 0.004292364697903395, + 0.004300045780837536, + 0.004514834377914667, + 0.004574811551719904, + 0.00403080228716135, + 0.005138358101248741, + 0.005272414535284042, + 0.00478488439694047, + 0.00355572160333395, + 0.003519314108416438, + 0.003579199779778719, + 0.0037041392643004656, + 0.0032480985391885042, + 0.0030536428093910217, + 0.0031776903197169304, + 0.004431017674505711, + 0.009679583832621574, + 0.010927312076091766, + 0.01578453928232193, + 0.03340731933712959, + 0.009786979295313358, + 0.009369269013404846, + 0.01637425273656845, + 0.014461670070886612, + 0.003784971544519067, + 0.004514368716627359, + 0.005911334417760372, + 0.011655200272798538, + 0.004298253450542688, + 0.00505548482760787, + 0.005704246927052736, + 0.006563269533216953, + 0.0018851098138839006, + 0.0023806269746273756, + 0.002749738050624728, + 0.004849870223551989, + 0.0020988131873309612, + 0.0028451166581362486, + 0.0025458885356783867, + 0.0029878332279622555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.011089714244008064 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a bird's-eye view focused vertically on weathered bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, dark brown terracotta jar visible falling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jar disintegrate upon impact, shattering into countless fragments of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments fly in all directions at high speed, mixed with a small cloud of rising dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a harsh sidelight that highlights the stone texture and captures reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22553575038909912 + ], + "opposite_logprob_tokens": [ + -1.6005357503890991 + ], + "logprob_reward": -0.22553575038909912, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22553575038909912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13441646099090576, + 0.1148921549320221, + 0.029513968154788017, + 0.061747822910547256, + 0.5418136715888977, + 0.0304600540548563, + 0.013768565841019154, + 0.017824796959757805, + 0.014650437980890274, + 0.04091205447912216 + ], + "credit_map": [ + 0.009275862947106361, + 0.013816227205097675, + 0.04149414226412773, + 0.013627848587930202, + 0.006401108577847481, + 0.034384358674287796, + 0.008820564486086369, + 0.006596350576728582, + 0.022284071892499924, + 0.011406590230762959, + 0.013113516382873058, + 0.00848119892179966, + 0.012436658143997192, + 0.015626313164830208, + 0.01003405824303627, + 0.021509753540158272, + 0.003009483218193054, + 0.0028812384698539972, + 0.003960231319069862, + 0.004285205155611038, + 0.003169754985719919, + 0.0038827629759907722, + 0.0043085599318146706, + 0.004016734194010496, + 0.0078858258202672, + 0.011945409700274467, + 0.006151525303721428, + 0.005879296455532312, + 0.012066775001585484, + 0.006087891757488251, + 0.006557518150657415, + 0.005173582583665848, + 0.06258691102266312, + 0.016691748052835464, + 0.1869327425956726, + 0.028767304494976997, + 0.033015601336956024, + 0.1745913028717041, + 0.01879105530679226, + 0.020437000319361687, + 0.002314399229362607, + 0.0030166609212756157, + 0.0034609113354235888, + 0.005537505727261305, + 0.0038076371420174837, + 0.003957363776862621, + 0.004636092577129602, + 0.003729485673829913, + 0.0012410894269123673, + 0.0015611929120495915, + 0.001523389364592731, + 0.0027825557626783848, + 0.0013263198779895902, + 0.0014050041791051626, + 0.0014642359456047416, + 0.00246477872133255, + 0.0015350135508924723, + 0.0012788334861397743, + 0.0020700697787106037, + 0.004968875087797642, + 0.0014406212139874697, + 0.0016847981605678797, + 0.0025825840421020985, + 0.0022640018723905087, + 0.0018107607029378414, + 0.0014210683293640614, + 0.0021599172614514828, + 0.0014592139050364494, + 0.0016982657834887505, + 0.0017195214750245214, + 0.0015745186246931553, + 0.002807171083986759, + 0.003644971875473857, + 0.003970756661146879, + 0.005212237127125263, + 0.009862714447081089, + 0.003562602447345853, + 0.004929015878587961, + 0.0048571862280368805, + 0.00487257307395339 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.22553575038909912 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a bird's-eye view focused vertically on weathered bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, dark brown terracotta jar visible falling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jar disintegrate upon impact, shattering into countless fragments of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments fly in all directions at high speed, mixed with a small cloud of rising dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a harsh sidelight that highlights the stone texture and captures reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011073680594563484 + ], + "opposite_logprob_tokens": [ + -4.511073589324951 + ], + "logprob_reward": -0.011073680594563484, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011073680594563484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.228866845369339, + 0.057109080255031586, + 0.05353956297039986, + 0.08678889274597168, + 0.4011162519454956, + 0.03264473006129265, + 0.02837252803146839, + 0.050395652651786804, + 0.01313378382474184, + 0.04803268238902092 + ], + "credit_map": [ + 0.015547525137662888, + 0.022412283346056938, + 0.0745556652545929, + 0.021898917853832245, + 0.009779148735105991, + 0.06076614931225777, + 0.013487395830452442, + 0.010419754311442375, + 0.008806843310594559, + 0.00664076441898942, + 0.006699196062982082, + 0.005451842211186886, + 0.005521080922335386, + 0.00860997661948204, + 0.005263300146907568, + 0.010116077959537506, + 0.006350166629999876, + 0.005707382690161467, + 0.006776497699320316, + 0.008170291781425476, + 0.005603798199445009, + 0.007104646414518356, + 0.007080144714564085, + 0.006746634840965271, + 0.01306596677750349, + 0.017799286171793938, + 0.010257876478135586, + 0.011388694867491722, + 0.010237220674753189, + 0.00606832280755043, + 0.008980587124824524, + 0.008990936912596226, + 0.04649501293897629, + 0.01306631788611412, + 0.1378585696220398, + 0.02149546705186367, + 0.024770313873887062, + 0.12760181725025177, + 0.014525887556374073, + 0.015302862972021103, + 0.0038968354929238558, + 0.0036435644142329693, + 0.00414288928732276, + 0.004321584012359381, + 0.00369861233048141, + 0.0042509050108492374, + 0.0050955964252352715, + 0.0035947419237345457, + 0.002072888659313321, + 0.003839482320472598, + 0.0037813729140907526, + 0.00526838144287467, + 0.0033035422675311565, + 0.0034479456953704357, + 0.0031555730383843184, + 0.0035033419262617826, + 0.003962612245231867, + 0.0036647713277488947, + 0.006227755919098854, + 0.013109378516674042, + 0.004586886148899794, + 0.004241778049618006, + 0.0070587326772511005, + 0.007543736603111029, + 0.0010418788297101855, + 0.0010928084375336766, + 0.0016111342702060938, + 0.0024298494681715965, + 0.0013640302931889892, + 0.001632773200981319, + 0.001463408349081874, + 0.0024979007430374622, + 0.004232483915984631, + 0.004711055662482977, + 0.00624633114784956, + 0.01171973254531622, + 0.0042104218155145645, + 0.0055821300484240055, + 0.0057959239929914474, + 0.005534604657441378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.011073680594563484 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a bird's-eye view focused vertically on weathered bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, dark brown terracotta jar visible falling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jar disintegrate upon impact, shattering into countless fragments of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments fly in all directions at high speed, mixed with a small cloud of rising dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a harsh sidelight that highlights the stone texture and captures reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08901819586753845 + ], + "opposite_logprob_tokens": [ + -2.4640181064605713 + ], + "logprob_reward": -0.08901819586753845, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901819586753845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3209201991558075, + 0.10703882575035095, + 0.014784914441406727, + 0.027385810390114784, + 0.0827777087688446, + 0.012581771239638329, + 0.06179944798350334, + 0.2935566306114197, + 0.05678935721516609, + 0.022365359589457512 + ], + "credit_map": [ + 0.022098973393440247, + 0.03132016211748123, + 0.10576675087213516, + 0.0315878689289093, + 0.012177800759673119, + 0.08754085004329681, + 0.01726410910487175, + 0.01316368393599987, + 0.01643402874469757, + 0.010543321259319782, + 0.014206292107701302, + 0.009295319207012653, + 0.011216484010219574, + 0.022643379867076874, + 0.006531961262226105, + 0.016168029978871346, + 0.0018399013206362724, + 0.001371051766909659, + 0.002788109704852104, + 0.0018806009320542216, + 0.0019107857951894403, + 0.0013783504255115986, + 0.0020808489061892033, + 0.0015352640766650438, + 0.0027946496848016977, + 0.00299555785022676, + 0.003098326502367854, + 0.005604403093457222, + 0.002729266881942749, + 0.002161560347303748, + 0.005652694962918758, + 0.0023493492044508457, + 0.008320849388837814, + 0.00792618840932846, + 0.0210882555693388, + 0.005029681138694286, + 0.012490065768361092, + 0.018121633678674698, + 0.006187148857861757, + 0.0036138775758445263, + 0.0013210466131567955, + 0.0013945770915597677, + 0.0018163517815992236, + 0.0015085733029991388, + 0.0013594634365290403, + 0.0015918657882139087, + 0.0018517699791118503, + 0.0017381219659000635, + 0.00613003084436059, + 0.005432139616459608, + 0.007184586487710476, + 0.008851745165884495, + 0.00874782633036375, + 0.0072073787450790405, + 0.008018217980861664, + 0.010227518156170845, + 0.024599513038992882, + 0.025464599952101707, + 0.0375199094414711, + 0.06308231502771378, + 0.03329668194055557, + 0.03120238333940506, + 0.03662300109863281, + 0.04176819324493408, + 0.00469603855162859, + 0.004921984393149614, + 0.007151614874601364, + 0.014469134621322155, + 0.00501996511593461, + 0.005888795014470816, + 0.007046607322990894, + 0.007595209404826164, + 0.0022070985287427902, + 0.0025861780159175396, + 0.003078408306464553, + 0.005010651890188456, + 0.002124848309904337, + 0.0021769453305751085, + 0.0027241704519838095, + 0.0024570550303906202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.08901819586753845 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "strictly static, bird's-eye view, the shot is focused vertically on a set of weathered bluestone steps. The surface is etched with age, its edges smoothed by age, and wisps of tenacious dark green moss dot the crevices. Suddenly, a thick, dark brown terracotta jar plummets from above, its rough surface grazing the air as it falls. Upon contact with the solid stone steps, the jar disintegrates with a dull crack, shattering into countless terracotta fragments of varying sizes and sharp edges. Centered on the point of impact, these fragments, mixed with a small cloud of rising dust, fly in all directions at high speed, leaving a fleeting and chaotic scene of destruction on the cold stone surface. A harsh, sidelight slices through the scene, highlighting the ancient texture of the stone steps and capturing every fleeting reflection as the fragments fly, creating an atmosphere of tension, suddenness, and irreversibility.", + "seed": 11492, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_2/index_4113_seed_11492_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04199529606848955, + "vlm_logprob": -0.04199529606848955, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.04199529606848955, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a bird's-eye view focused vertically on weathered bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, dark brown terracotta jar visible falling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jar disintegrate upon impact, shattering into countless fragments of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments fly in all directions at high speed, mixed with a small cloud of rising dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a harsh sidelight that highlights the stone texture and captures reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.037358131259679794 + ], + "opposite_logprob_tokens": [ + -3.662358045578003 + ], + "logprob_reward": -0.037358131259679794, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.037358131259679794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3067763149738312, + 0.06963049620389938, + 0.06831707060337067, + 0.2033163607120514, + 0.06266500055789948, + 0.047197725623846054, + 0.022823890671133995, + 0.11081317067146301, + 0.04530250281095505, + 0.06315740942955017 + ], + "credit_map": [ + 0.026396755129098892, + 0.0159271452575922, + 0.1247512623667717, + 0.025479357689619064, + 0.016254503279924393, + 0.024228595197200775, + 0.016762956976890564, + 0.056975726038217545, + 0.006681715603917837, + 0.0069008986465632915, + 0.007649097591638565, + 0.005071981344372034, + 0.009618074633181095, + 0.018511749804019928, + 0.008892123587429523, + 0.006304855924099684, + 0.005844298750162125, + 0.006543025840073824, + 0.007886132225394249, + 0.0070650167763233185, + 0.00802136491984129, + 0.011006651446223259, + 0.011299991980195045, + 0.01065058633685112, + 0.02075648494064808, + 0.025643302127718925, + 0.03807961940765381, + 0.027013178914785385, + 0.03166241571307182, + 0.0158385019749403, + 0.015495509840548038, + 0.0288273673504591, + 0.006705792620778084, + 0.007864927873015404, + 0.006025434006005526, + 0.005071003455668688, + 0.009248274378478527, + 0.014039333909749985, + 0.009448565542697906, + 0.004261671099811792, + 0.005172578617930412, + 0.00615922873839736, + 0.0041412049904465675, + 0.01117255911231041, + 0.004094842355698347, + 0.004215245135128498, + 0.007227632682770491, + 0.005014434922486544, + 0.0031546843238174915, + 0.0030505929607897997, + 0.0018614459550008178, + 0.0014644929906353354, + 0.0030815729405730963, + 0.004715404473245144, + 0.0027118364814668894, + 0.0027838600799441338, + 0.014613397419452667, + 0.007352214772254229, + 0.003996892366558313, + 0.0033271554857492447, + 0.034376442432403564, + 0.037839457392692566, + 0.005220084451138973, + 0.004087524954229593, + 0.004457044880837202, + 0.004406724590808153, + 0.0038585655856877565, + 0.010789626277983189, + 0.004419144243001938, + 0.006855331361293793, + 0.006682735867798328, + 0.003833330236375332, + 0.0071564665995538235, + 0.007097024004906416, + 0.0037328205071389675, + 0.004436394665390253, + 0.015962028875947, + 0.015291121788322926, + 0.006109226960688829, + 0.0033723278902471066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.037358131259679794 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a bird's-eye view focused vertically on weathered bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, dark brown terracotta jar visible falling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jar disintegrate upon impact, shattering into countless fragments of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments fly in all directions at high speed, mixed with a small cloud of rising dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a harsh sidelight that highlights the stone texture and captures reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023310784250497818 + ], + "opposite_logprob_tokens": [ + -3.773310899734497 + ], + "logprob_reward": -0.023310784250497818, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023310784250497818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2610398530960083, + 0.23010453581809998, + 0.10186436027288437, + 0.19180725514888763, + 0.06152781471610069, + 0.0453522726893425, + 0.019359761849045753, + 0.015153059735894203, + 0.056655801832675934, + 0.017135243862867355 + ], + "credit_map": [ + 0.020091144368052483, + 0.022102778777480125, + 0.0705229863524437, + 0.020660849288105965, + 0.05628793686628342, + 0.020408881828188896, + 0.016710929572582245, + 0.034254368394613266, + 0.019807403907179832, + 0.018111662939190865, + 0.013974838890135288, + 0.022573571652173996, + 0.03919711336493492, + 0.06629440933465958, + 0.030473817139863968, + 0.01967170648276806, + 0.007756689563393593, + 0.017069218680262566, + 0.006171775050461292, + 0.00425004493445158, + 0.026633186265826225, + 0.017839912325143814, + 0.01030109915882349, + 0.011842426843941212, + 0.01799887977540493, + 0.02383621223270893, + 0.01699429750442505, + 0.014720090664923191, + 0.039216816425323486, + 0.019024886190891266, + 0.013562108390033245, + 0.04645396023988724, + 0.006334533914923668, + 0.007864825427532196, + 0.004706953186541796, + 0.004786261823028326, + 0.012339347042143345, + 0.011640333570539951, + 0.00971816387027502, + 0.004137397278100252, + 0.0038709035143256187, + 0.003927296027541161, + 0.002532605780288577, + 0.005788457114249468, + 0.00963787641376257, + 0.011552903801202774, + 0.005330015905201435, + 0.0027122115716338158, + 0.001972462749108672, + 0.0026879964862018824, + 0.0015287456335499883, + 0.002516878303140402, + 0.0030235445592552423, + 0.003966701682657003, + 0.00191062584053725, + 0.0017528078751638532, + 0.0018980290042236447, + 0.0016286583850160241, + 0.0009272106690332294, + 0.001002436620183289, + 0.0031396097037941217, + 0.0042307074181735516, + 0.0013123429380357265, + 0.0010140660451725125, + 0.005414037499576807, + 0.004952616058290005, + 0.0057301768101751804, + 0.013197818771004677, + 0.00725534139201045, + 0.008235731162130833, + 0.00647697364911437, + 0.005393108353018761, + 0.002837477484717965, + 0.0017583000008016825, + 0.0019555778708308935, + 0.0013516361359506845, + 0.002226181561127305, + 0.003054437693208456, + 0.002574760699644685, + 0.0013768711360171437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.023310784250497818 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a bird's-eye view focused vertically on weathered bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, dark brown terracotta jar visible falling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jar disintegrate upon impact, shattering into countless fragments of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments fly in all directions at high speed, mixed with a small cloud of rising dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a harsh sidelight that highlights the stone texture and captures reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043056320399045944 + ], + "opposite_logprob_tokens": [ + -3.1680562496185303 + ], + "logprob_reward": -0.043056320399045944, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043056320399045944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1342501938343048, + 0.06787888705730438, + 0.12011190503835678, + 0.084653340280056, + 0.16854999959468842, + 0.09434278309345245, + 0.0787237212061882, + 0.19301354885101318, + 0.04340357333421707, + 0.015072046779096127 + ], + "credit_map": [ + 0.01183405239135027, + 0.013397138565778732, + 0.012750939466059208, + 0.01584995537996292, + 0.034382302314043045, + 0.01932491548359394, + 0.012504958547651768, + 0.014205927029252052, + 0.006605693604797125, + 0.00776333175599575, + 0.007625427097082138, + 0.00899138581007719, + 0.009598949924111366, + 0.011935845948755741, + 0.009592286311089993, + 0.005765960551798344, + 0.016272705048322678, + 0.016877852380275726, + 0.013267862610518932, + 0.012076136656105518, + 0.015103614889085293, + 0.01458781585097313, + 0.01414248626679182, + 0.01778342016041279, + 0.006532380357384682, + 0.009809370152652264, + 0.007667844183743, + 0.007131034974008799, + 0.01742810569703579, + 0.018140966072678566, + 0.008016135543584824, + 0.009927494451403618, + 0.014737698249518871, + 0.02128421887755394, + 0.012318363413214684, + 0.011362608522176743, + 0.0363280214369297, + 0.03823039308190346, + 0.022981613874435425, + 0.011307066306471825, + 0.008589353412389755, + 0.007109825499355793, + 0.004354508128017187, + 0.009502744302153587, + 0.022872088477015495, + 0.028938433155417442, + 0.0072822729125618935, + 0.0056935446336865425, + 0.00952681340277195, + 0.005104192532598972, + 0.0037607226986438036, + 0.003363449824973941, + 0.019119692966341972, + 0.02604762092232704, + 0.006146524101495743, + 0.005654702428728342, + 0.024655507877469063, + 0.010719158686697483, + 0.00575666781514883, + 0.00457981089130044, + 0.06202366575598717, + 0.07025440782308578, + 0.00877253245562315, + 0.00625178636983037, + 0.006463061552494764, + 0.004101182334125042, + 0.003785981796681881, + 0.00805321242660284, + 0.005369285121560097, + 0.007310765795409679, + 0.004296989645808935, + 0.0040230900049209595, + 0.0015436731046065688, + 0.0021789006423205137, + 0.00135790032800287, + 0.0011165313189849257, + 0.003255021059885621, + 0.002673375653102994, + 0.0017815802711993456, + 0.0011650631204247475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.043056320399045944 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a bird's-eye view focused vertically on weathered bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, dark brown terracotta jar visible falling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jar disintegrate upon impact, shattering into countless fragments of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments fly in all directions at high speed, mixed with a small cloud of rising dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a harsh sidelight that highlights the stone texture and captures reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005948578007519245 + ], + "opposite_logprob_tokens": [ + -5.130948543548584 + ], + "logprob_reward": -0.005948578007519245, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005948578007519245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21573375165462494, + 0.03925436735153198, + 0.07134782522916794, + 0.18158750236034393, + 0.11355017870664597, + 0.11412311345338821, + 0.07719673961400986, + 0.13992232084274292, + 0.028871146962046623, + 0.018413055688142776 + ], + "credit_map": [ + 0.015814561396837234, + 0.01930614933371544, + 0.05398474261164665, + 0.0288788303732872, + 0.029135966673493385, + 0.028575055301189423, + 0.018319902941584587, + 0.021718522533774376, + 0.0037503798957914114, + 0.004644572269171476, + 0.0034577026963233948, + 0.005475419107824564, + 0.006034047808498144, + 0.005042609758675098, + 0.007569779176265001, + 0.0032798582687973976, + 0.005791823845356703, + 0.017616596072912216, + 0.006046738941222429, + 0.006868223659694195, + 0.014794251881539822, + 0.006560778245329857, + 0.0070221503265202045, + 0.006647262256592512, + 0.01781904324889183, + 0.02027570828795433, + 0.014376897364854813, + 0.00778309628367424, + 0.03518065810203552, + 0.0232606939971447, + 0.012461510486900806, + 0.050429899245500565, + 0.009448258206248283, + 0.013531808741390705, + 0.007757373154163361, + 0.008217552676796913, + 0.02446502260863781, + 0.030747683718800545, + 0.012609224766492844, + 0.006773257162421942, + 0.010412546806037426, + 0.008766943588852882, + 0.005900299176573753, + 0.012471766211092472, + 0.02724691852927208, + 0.03421466425061226, + 0.00917556881904602, + 0.005934401415288448, + 0.008291652426123619, + 0.005366330035030842, + 0.004010797943919897, + 0.0031853211112320423, + 0.019860511645674706, + 0.02616298757493496, + 0.005956355482339859, + 0.004362784326076508, + 0.017351554706692696, + 0.007839574478566647, + 0.004397445358335972, + 0.003574998350813985, + 0.04458858072757721, + 0.05111977830529213, + 0.006433392409235239, + 0.00461699441075325, + 0.002814660081639886, + 0.002686400432139635, + 0.0025091138668358326, + 0.00555462995544076, + 0.0037256854120641947, + 0.005111263133585453, + 0.0037058189045637846, + 0.002763577038422227, + 0.002650037407875061, + 0.0034136255271732807, + 0.0018290078733116388, + 0.0015845851739868522, + 0.002994055859744549, + 0.0019517970504239202, + 0.0024775597266852856, + 0.0015123876510187984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.005948578007519245 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a bird's-eye view focused vertically on weathered bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, dark brown terracotta jar visible falling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jar disintegrate upon impact, shattering into countless fragments of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments fly in all directions at high speed, mixed with a small cloud of rising dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a harsh sidelight that highlights the stone texture and captures reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030266642570496 + ], + "opposite_logprob_tokens": [ + -2.3503026962280273 + ], + "logprob_reward": -0.10030266642570496, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030266642570496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38528305292129517, + 0.05097994580864906, + 0.04789667949080467, + 0.06536637991666794, + 0.30021873116493225, + 0.02958088368177414, + 0.026141656562685966, + 0.04057438671588898, + 0.026531003415584564, + 0.027427269145846367 + ], + "credit_map": [ + 0.029671059921383858, + 0.012246973812580109, + 0.15283724665641785, + 0.026716137304902077, + 0.04781966283917427, + 0.018082480877637863, + 0.015159870497882366, + 0.0827496275305748, + 0.004230393096804619, + 0.005156649276614189, + 0.002799652051180601, + 0.004896063357591629, + 0.007539438083767891, + 0.0160149484872818, + 0.006079718004912138, + 0.004263082519173622, + 0.004695911426097155, + 0.007976577617228031, + 0.005905317608267069, + 0.006777159404009581, + 0.006380465347319841, + 0.00583861768245697, + 0.005351993255317211, + 0.004970637150108814, + 0.007045634090900421, + 0.006101437844336033, + 0.005023245699703693, + 0.0038373623974621296, + 0.00979091227054596, + 0.007343657314777374, + 0.006043966859579086, + 0.02018016390502453, + 0.031932711601257324, + 0.03798956051468849, + 0.02692815661430359, + 0.02728171832859516, + 0.0669705867767334, + 0.030595140531659126, + 0.05248837545514107, + 0.026032477617263794, + 0.002868574345484376, + 0.003915715496987104, + 0.002686224179342389, + 0.005200049374252558, + 0.004654945805668831, + 0.003984205424785614, + 0.003799302037805319, + 0.002471867250278592, + 0.003032383043318987, + 0.00412673270329833, + 0.00213741441257298, + 0.0015227519907057285, + 0.0046447874046862125, + 0.005834060255438089, + 0.002319280756637454, + 0.002524247393012047, + 0.004616664722561836, + 0.004000111483037472, + 0.0014113547513261437, + 0.0015678672352805734, + 0.012055953964591026, + 0.012837148271501064, + 0.0021828014869242907, + 0.0019024874782189727, + 0.0035732656251639128, + 0.002755071735009551, + 0.0018184591317549348, + 0.0013429672690108418, + 0.005582067184150219, + 0.00797185581177473, + 0.0016778612043708563, + 0.0018094565020874143, + 0.005131387617439032, + 0.004128832835704088, + 0.0021429467014968395, + 0.0017834753962233663, + 0.005922497715801001, + 0.0028318208642303944, + 0.0034816814586520195, + 0.0020046276040375233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.10030266642570496 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "strictly static, bird's-eye view, the shot is focused vertically on a set of weathered bluestone steps. The surface is etched with age, its edges smoothed by age, and wisps of tenacious dark green moss dot the crevices. Suddenly, a thick, dark brown terracotta jar plummets from above, its rough surface grazing the air as it falls. Upon contact with the solid stone steps, the jar disintegrates with a dull crack, shattering into countless terracotta fragments of varying sizes and sharp edges. Centered on the point of impact, these fragments, mixed with a small cloud of rising dust, fly in all directions at high speed, leaving a fleeting and chaotic scene of destruction on the cold stone surface. A harsh, sidelight slices through the scene, highlighting the ancient texture of the stone steps and capturing every fleeting reflection as the fragments fly, creating an atmosphere of tension, suddenness, and irreversibility.", + "seed": 11493, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_2/index_4113_seed_11493_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06521430294960737, + "vlm_logprob": -0.06521430294960737, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.06521430294960737, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a bird's-eye view focused vertically on weathered bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, dark brown terracotta jar visible falling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jar disintegrate upon impact, shattering into countless fragments of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments fly in all directions at high speed, mixed with a small cloud of rising dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a harsh sidelight that highlights the stone texture and captures reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0554058738052845 + ], + "opposite_logprob_tokens": [ + -3.180405855178833 + ], + "logprob_reward": -0.0554058738052845, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0554058738052845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1767343133687973, + 0.15988513827323914, + 0.09527460485696793, + 0.09631457179784775, + 0.2201903760433197, + 0.10574986040592194, + 0.03908011317253113, + 0.025173179805278778, + 0.026723669841885567, + 0.05487419664859772 + ], + "credit_map": [ + 0.013868812471628189, + 0.015098554082214832, + 0.02648979425430298, + 0.04069006070494652, + 0.02125317044556141, + 0.017664557322859764, + 0.028028439730405807, + 0.013640893623232841, + 0.018256746232509613, + 0.01755804382264614, + 0.024043288081884384, + 0.02848333865404129, + 0.018158162012696266, + 0.0136789595708251, + 0.024906100705266, + 0.014800489880144596, + 0.010857325047254562, + 0.021689189597964287, + 0.011500739492475986, + 0.014313635416328907, + 0.007186779752373695, + 0.00763581832870841, + 0.010869807563722134, + 0.011221298947930336, + 0.01822306588292122, + 0.010331272147595882, + 0.013096878305077553, + 0.012522870674729347, + 0.01469487976282835, + 0.009633705951273441, + 0.010540399700403214, + 0.007271486334502697, + 0.018843190744519234, + 0.03830305486917496, + 0.048653494566679, + 0.01761328987777233, + 0.0201629139482975, + 0.03130393475294113, + 0.021559227257966995, + 0.023751256987452507, + 0.015351178124547005, + 0.028838088735938072, + 0.008973903954029083, + 0.011460907757282257, + 0.010368323884904385, + 0.012908880598843098, + 0.007916776463389397, + 0.009931789711117744, + 0.005301040597259998, + 0.006733836606144905, + 0.0073165991343557835, + 0.004484525881707668, + 0.0035752567928284407, + 0.0061633833684027195, + 0.0035325775388628244, + 0.001972891390323639, + 0.002034811768680811, + 0.004102959763258696, + 0.006573228631168604, + 0.005389825440943241, + 0.0024606746155768633, + 0.0017102607525885105, + 0.0016131410375237465, + 0.0012882754672318697, + 0.0028543821536004543, + 0.00369067769497633, + 0.004472027532756329, + 0.005619854666292667, + 0.003160875290632248, + 0.0032075420022010803, + 0.0014486197615042329, + 0.002269688993692398, + 0.005443217698484659, + 0.011946595273911953, + 0.006181112956255674, + 0.01111864298582077, + 0.007352571468800306, + 0.005935761146247387, + 0.002679649507626891, + 0.004216642118990421 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.0554058738052845 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a bird's-eye view focused vertically on weathered bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, dark brown terracotta jar visible falling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jar disintegrate upon impact, shattering into countless fragments of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments fly in all directions at high speed, mixed with a small cloud of rising dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a harsh sidelight that highlights the stone texture and captures reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372414782643318 + ], + "opposite_logprob_tokens": [ + -3.408724069595337 + ], + "logprob_reward": -0.03372414782643318, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372414782643318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1375012844800949, + 0.20395998656749725, + 0.09791777282953262, + 0.3052040636539459, + 0.09530089050531387, + 0.07495122402906418, + 0.032446909695863724, + 0.02052181214094162, + 0.011217334307730198, + 0.02097872458398342 + ], + "credit_map": [ + 0.014500570483505726, + 0.01342166680842638, + 0.014692943543195724, + 0.02485732175409794, + 0.02033575251698494, + 0.026792006567120552, + 0.013665670529007912, + 0.009235340170562267, + 0.01794380694627762, + 0.03529121354222298, + 0.011894750408828259, + 0.022345583885908127, + 0.03743094578385353, + 0.036986593157052994, + 0.0275048166513443, + 0.014562263153493404, + 0.008968963287770748, + 0.008620025590062141, + 0.010566660203039646, + 0.021068500354886055, + 0.0078765032812953, + 0.009798912331461906, + 0.01665988750755787, + 0.014358309097588062, + 0.056939221918582916, + 0.03863481059670448, + 0.03191433101892471, + 0.03502773866057396, + 0.042579520493745804, + 0.04464571923017502, + 0.03699352964758873, + 0.0184691920876503, + 0.008278475143015385, + 0.01092139445245266, + 0.016946809366345406, + 0.01353307906538248, + 0.010641651228070259, + 0.008733071386814117, + 0.015437038615345955, + 0.010809367522597313, + 0.007391003891825676, + 0.02986065298318863, + 0.005623622331768274, + 0.010137608274817467, + 0.009763325564563274, + 0.004686355125159025, + 0.0038799152243882418, + 0.003608741331845522, + 0.004683168604969978, + 0.0032102009281516075, + 0.004809909034520388, + 0.003573937574401498, + 0.003210346447303891, + 0.00739798741415143, + 0.0034678077790886164, + 0.00209355098195374, + 0.001759521197527647, + 0.0013888953253626823, + 0.0049739256501197815, + 0.005278982222080231, + 0.00147050223313272, + 0.0030107656493782997, + 0.0017564079025760293, + 0.0008828106801956892, + 0.0014228671789169312, + 0.0014801179058849812, + 0.0015229611890390515, + 0.0008475763024762273, + 0.0019131150329485536, + 0.0015256417682394385, + 0.0011903462000191212, + 0.0013147087302058935, + 0.0027773315086960793, + 0.002288549905642867, + 0.0036017780657857656, + 0.0021211407147347927, + 0.00288430810905993, + 0.002983302576467395, + 0.001621433300897479, + 0.0027008799370378256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.03372414782643318 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a bird's-eye view focused vertically on weathered bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, dark brown terracotta jar visible falling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jar disintegrate upon impact, shattering into countless fragments of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments fly in all directions at high speed, mixed with a small cloud of rising dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a harsh sidelight that highlights the stone texture and captures reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292840540409088 + ], + "opposite_logprob_tokens": [ + -2.2379283905029297 + ], + "logprob_reward": -0.11292840540409088, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292840540409088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13433168828487396, + 0.10268061608076096, + 0.12795552611351013, + 0.2364979088306427, + 0.13326872885227203, + 0.07433471083641052, + 0.05339622125029564, + 0.02730526588857174, + 0.049048811197280884, + 0.06118045747280121 + ], + "credit_map": [ + 0.01416646596044302, + 0.01730821654200554, + 0.017977267503738403, + 0.01577567495405674, + 0.018808670341968536, + 0.0256035178899765, + 0.0143518578261137, + 0.010340015403926373, + 0.0082637844607234, + 0.009551205672323704, + 0.01110580749809742, + 0.013106731697916985, + 0.01903728023171425, + 0.015134328044950962, + 0.013693614862859249, + 0.012787871062755585, + 0.012911808677017689, + 0.016652142629027367, + 0.012859044596552849, + 0.018557999283075333, + 0.01499815471470356, + 0.013781298883259296, + 0.021769871935248375, + 0.01642521470785141, + 0.03924531117081642, + 0.03383481502532959, + 0.03048212267458439, + 0.029339391738176346, + 0.03123442642390728, + 0.026942359283566475, + 0.029839418828487396, + 0.015580067411065102, + 0.010741605423390865, + 0.017875224351882935, + 0.02674175798892975, + 0.011442803777754307, + 0.014052408747375011, + 0.01970086246728897, + 0.01542403269559145, + 0.01729002222418785, + 0.008745154365897179, + 0.02563709393143654, + 0.004947683308273554, + 0.012583030387759209, + 0.008366222493350506, + 0.006047472357749939, + 0.004613058641552925, + 0.003394994419068098, + 0.004562149755656719, + 0.0133755411952734, + 0.007445028517395258, + 0.008701750077307224, + 0.004748319275677204, + 0.0051759821362793446, + 0.003815168747678399, + 0.005572280380874872, + 0.002760442206636071, + 0.008791981264948845, + 0.004632904659956694, + 0.0020162290893495083, + 0.002747930586338043, + 0.0024104397743940353, + 0.0019750322680920362, + 0.0019703060388565063, + 0.006180983502417803, + 0.020318904891610146, + 0.0033182776533067226, + 0.007170693948864937, + 0.00511612044647336, + 0.002294882433488965, + 0.002374111209064722, + 0.0022748338524252176, + 0.008248387835919857, + 0.024113988503813744, + 0.0029613010119646788, + 0.009454373270273209, + 0.006651736795902252, + 0.003648584010079503, + 0.0030782180838286877, + 0.0030238726176321507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.11292840540409088 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a bird's-eye view focused vertically on weathered bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, dark brown terracotta jar visible falling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jar disintegrate upon impact, shattering into countless fragments of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments fly in all directions at high speed, mixed with a small cloud of rising dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a harsh sidelight that highlights the stone texture and captures reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011080636642873287 + ], + "opposite_logprob_tokens": [ + -4.511080741882324 + ], + "logprob_reward": -0.011080636642873287, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080636642873287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12781959772109985, + 0.05713210627436638, + 0.08644494414329529, + 0.38845768570899963, + 0.07219898700714111, + 0.13088016211986542, + 0.04555042088031769, + 0.026076342910528183, + 0.030690157786011696, + 0.034749586135149 + ], + "credit_map": [ + 0.01381293497979641, + 0.011883609928190708, + 0.016546299681067467, + 0.02411792241036892, + 0.015259220264852047, + 0.014705299399793148, + 0.01582414098083973, + 0.0156701672822237, + 0.005994231440126896, + 0.004226911347359419, + 0.006024974398314953, + 0.0077536143362522125, + 0.008976956829428673, + 0.009890589863061905, + 0.007741297595202923, + 0.00652353186160326, + 0.009984759613871574, + 0.008234492503106594, + 0.010097487829625607, + 0.014525912702083588, + 0.007413006853312254, + 0.008205673657357693, + 0.016313476487994194, + 0.011670134961605072, + 0.07090317457914352, + 0.053039953112602234, + 0.045552175492048264, + 0.04245589300990105, + 0.05670223385095596, + 0.04679495468735695, + 0.04873841628432274, + 0.024270914494991302, + 0.00728457560762763, + 0.013756886124610901, + 0.010803330689668655, + 0.00903877429664135, + 0.007511920295655727, + 0.010918102227151394, + 0.0073768990114331245, + 0.005508496426045895, + 0.015822531655430794, + 0.05343193933367729, + 0.006263756658881903, + 0.017381345853209496, + 0.016100266948342323, + 0.0066483295522630215, + 0.0065953414887189865, + 0.008636635728180408, + 0.004571284167468548, + 0.009319954551756382, + 0.006006502080708742, + 0.006460451986640692, + 0.003545372746884823, + 0.005956059321761131, + 0.0039635575376451015, + 0.005727238021790981, + 0.00342432945035398, + 0.007965780794620514, + 0.003971731755882502, + 0.001913763931952417, + 0.0023986208252608776, + 0.0020411224104464054, + 0.002334717195481062, + 0.002026276895776391, + 0.004142962861806154, + 0.012191488407552242, + 0.0024826705921441317, + 0.0038066781125962734, + 0.003075496293604374, + 0.0017217209096997976, + 0.0015653591835871339, + 0.0017037811921909451, + 0.00513743469491601, + 0.011920296587049961, + 0.0034045816864818335, + 0.0033814760390669107, + 0.003183321561664343, + 0.0026222229935228825, + 0.002370494417846203, + 0.002729754662141204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.011080636642873287 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a bird's-eye view focused vertically on weathered bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, dark brown terracotta jar visible falling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jar disintegrate upon impact, shattering into countless fragments of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments fly in all directions at high speed, mixed with a small cloud of rising dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a harsh sidelight that highlights the stone texture and captures reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293245106935501 + ], + "opposite_logprob_tokens": [ + -2.2379324436187744 + ], + "logprob_reward": -0.11293245106935501, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293245106935501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19904814660549164, + 0.1807631403207779, + 0.09027747064828873, + 0.16561554372310638, + 0.13325564563274384, + 0.09362484514713287, + 0.024452364072203636, + 0.026724740862846375, + 0.0466281957924366, + 0.03960997238755226 + ], + "credit_map": [ + 0.02255796454846859, + 0.018424993380904198, + 0.04130806773900986, + 0.020473498851060867, + 0.028536178171634674, + 0.02975071594119072, + 0.016315486282110214, + 0.021681221202015877, + 0.016183294355869293, + 0.02053050510585308, + 0.02851005643606186, + 0.023951273411512375, + 0.025278132408857346, + 0.019413692876696587, + 0.02417830191552639, + 0.022717870771884918, + 0.010681378655135632, + 0.008402290754020214, + 0.00991646945476532, + 0.015812648460268974, + 0.008428405970335007, + 0.008612134493887424, + 0.013569492846727371, + 0.01485464721918106, + 0.032797809690237045, + 0.02214455045759678, + 0.01911710575222969, + 0.020502230152487755, + 0.020618662238121033, + 0.022417131811380386, + 0.015408936887979507, + 0.012609096243977547, + 0.011930888518691063, + 0.01866338402032852, + 0.02990340255200863, + 0.010737604461610317, + 0.011770057491958141, + 0.01499785203486681, + 0.015801627188920975, + 0.01945081539452076, + 0.012136935256421566, + 0.037305109202861786, + 0.0056151594035327435, + 0.011278712190687656, + 0.010456233285367489, + 0.007545829750597477, + 0.005214807111769915, + 0.004072051495313644, + 0.0030855860095471144, + 0.005415941588580608, + 0.0023703661281615496, + 0.0038667432963848114, + 0.0024856645613908768, + 0.003272853558883071, + 0.0023671493399888277, + 0.0015880572609603405, + 0.0029154818039387465, + 0.008026620373129845, + 0.003669942496344447, + 0.002332316478714347, + 0.003169962204992771, + 0.0023690625093877316, + 0.0024247695691883564, + 0.0018165851943194866, + 0.005773864686489105, + 0.018899479880928993, + 0.002660830970853567, + 0.006910988595336676, + 0.005503471475094557, + 0.0028109983541071415, + 0.0019090408459305763, + 0.002159518888220191, + 0.004433112684637308, + 0.014862420037388802, + 0.003011393127962947, + 0.00625296775251627, + 0.0043846783228218555, + 0.002592361532151699, + 0.0022165204863995314, + 0.001856515766121447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.11293245106935501 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_3/index_tensor([3844, 3844])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_3/index_tensor([3844, 3844])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9345d22b6d29afa164cf8f8a6e682c2c7259aaab --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_3/index_tensor([3844, 3844])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, held at a slightly low angle, slowly pans right, following an elderly farmer inspecting his crops. His face is weathered, his eyes kind and contemplative. At dusk, he stands in a field of tall, lush corn, wearing a faded blue denim shirt with the sleeves rolled up to his forearms. As he moves, he deliberately extends his calloused, work-worn right hand, letting his fingertips gently and reverently glide across the broad, green surface of the corn leaves. The setting sun, just behind him and just out of frame, casts a magnificent, warm golden glow over the entire scene, creating a brilliant rim light on his silhouette and a fiery glow on the edges of the corn leaves. The sky transitions from a vibrant orange near the horizon to a deep twilight purple above. Long shadows stretch forward across the fertile fields, creating a profoundly serene and peaceful atmosphere, evoking a sense of a day's work and a deep connection to the land.", + "seed": 22672, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_3/index_3844_seed_22672_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.2506143980659544, + "vlm_logprob": -0.2506143980659544, + "vlm_yes_no_margin": 2.55, + "vlm_token_credit": -0.2506143980659544, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot of an elderly farmer inspecting crops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer wearing a faded blue denim shirt with sleeves rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the farmer extend his right hand to gently glide his fingertips across the corn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a field of tall, lush corn at dusk with a sky transitioning from orange to purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while the farmer moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17329058051109314 + ], + "opposite_logprob_tokens": [ + -1.9232906103134155 + ], + "logprob_reward": -0.17329058051109314, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.17329058051109314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3129428029060364, + 0.05604586377739906, + 0.06454475969076157, + 0.039971910417079926, + 0.16076640784740448, + 0.06208329647779465, + 0.04577462747693062, + 0.051837071776390076, + 0.1484224945306778, + 0.05761079490184784 + ], + "credit_map": [ + 0.015626443549990654, + 0.020705603063106537, + 0.08898858726024628, + 0.08758050948381424, + 0.015570199117064476, + 0.016225380823016167, + 0.022882549092173576, + 0.04536352679133415, + 0.004003174137324095, + 0.0054197609424591064, + 0.008530938997864723, + 0.009455996565520763, + 0.006508176214993, + 0.005967842414975166, + 0.010045172646641731, + 0.006114798132330179, + 0.004032564349472523, + 0.006250509526580572, + 0.011328322812914848, + 0.011922580190002918, + 0.004785282537341118, + 0.007376714609563351, + 0.011448192410171032, + 0.007400584872812033, + 0.0033726419787853956, + 0.004939420614391565, + 0.006822613999247551, + 0.006741377525031567, + 0.003180653089657426, + 0.0036831959150731564, + 0.006563599221408367, + 0.004668406676501036, + 0.013667089864611626, + 0.01857154071331024, + 0.030028682202100754, + 0.04493178427219391, + 0.008936568163335323, + 0.011243915185332298, + 0.018935207277536392, + 0.014451609924435616, + 0.005373030435293913, + 0.008196252398192883, + 0.00859109964221716, + 0.009989490732550621, + 0.0078986044973135, + 0.0049558281898498535, + 0.0102466382086277, + 0.006832346320152283, + 0.0036048192996531725, + 0.005499871913343668, + 0.00709012383595109, + 0.006322980392724276, + 0.004143072292208672, + 0.006036057602614164, + 0.007599841803312302, + 0.0054778545163571835, + 0.005419822875410318, + 0.006354390177875757, + 0.008110393770039082, + 0.006769142113626003, + 0.0037651483435183764, + 0.0058537921868264675, + 0.008824010379612446, + 0.006740369834005833, + 0.009208932518959045, + 0.011066165752708912, + 0.019082503393292427, + 0.01921769604086876, + 0.030507558956742287, + 0.013556909747421741, + 0.026900261640548706, + 0.018882455304265022, + 0.0035516980569809675, + 0.004963851999491453, + 0.007660031318664551, + 0.008827220648527145, + 0.010163399390876293, + 0.0040189712308347225, + 0.010380484163761139, + 0.008045127615332603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.17329058051109314 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot of an elderly farmer inspecting crops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer wearing a faded blue denim shirt with sleeves rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the farmer extend his right hand to gently glide his fingertips across the corn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a field of tall, lush corn at dusk with a sky transitioning from orange to purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while the farmer moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020603444427251816 + ], + "opposite_logprob_tokens": [ + -3.8956034183502197 + ], + "logprob_reward": -0.020603444427251816, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020603444427251816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30533504486083984, + 0.10247746109962463, + 0.04383380711078644, + 0.04903174936771393, + 0.09410694986581802, + 0.16261208057403564, + 0.0638834536075592, + 0.03503795340657234, + 0.044594574719667435, + 0.09908687323331833 + ], + "credit_map": [ + 0.021505653858184814, + 0.018357113003730774, + 0.04209865629673004, + 0.04870748519897461, + 0.023112045601010323, + 0.033830426633358, + 0.050371281802654266, + 0.06735238432884216, + 0.003135591046884656, + 0.007916092872619629, + 0.014279361814260483, + 0.014969997107982635, + 0.011982911266386509, + 0.006571012083441019, + 0.029464667662978172, + 0.014157832600176334, + 0.0027471433859318495, + 0.005920812021940947, + 0.006301986053586006, + 0.008877039887011051, + 0.004274055361747742, + 0.005475776735693216, + 0.005437490995973349, + 0.004799502901732922, + 0.0030840944964438677, + 0.0041115195490419865, + 0.007556242868304253, + 0.006830971222370863, + 0.004904596600681543, + 0.0063794683665037155, + 0.010054669342935085, + 0.006110187154263258, + 0.004572228528559208, + 0.005498677492141724, + 0.015253223478794098, + 0.013082858175039291, + 0.009842250496149063, + 0.013594155199825764, + 0.01885313168168068, + 0.013410419225692749, + 0.01052764616906643, + 0.016963692381978035, + 0.021648723632097244, + 0.027904635295271873, + 0.026957284659147263, + 0.012354370206594467, + 0.029485052451491356, + 0.016770677641034126, + 0.0037392142694443464, + 0.005515038035809994, + 0.008795270696282387, + 0.008952856063842773, + 0.008844421245157719, + 0.005964426789432764, + 0.014032434672117233, + 0.008039787411689758, + 0.003083365736529231, + 0.0036131476517766714, + 0.00472779106348753, + 0.0053891451098024845, + 0.0047455234453082085, + 0.003757632803171873, + 0.005410151090472937, + 0.004311197437345982, + 0.003660589689388871, + 0.0038593739736825228, + 0.006802894175052643, + 0.0066438280045986176, + 0.00531402975320816, + 0.003919134382158518, + 0.008571763522922993, + 0.00582295935600996, + 0.006668125279247761, + 0.007624468766152859, + 0.014482765458524227, + 0.014235787093639374, + 0.020685646682977676, + 0.0057537732645869255, + 0.017209118232131004, + 0.012427188456058502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.020603444427251816 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot of an elderly farmer inspecting crops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer wearing a faded blue denim shirt with sleeves rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the farmer extend his right hand to gently glide his fingertips across the corn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a field of tall, lush corn at dusk with a sky transitioning from orange to purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while the farmer moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789823904633522 + ], + "opposite_logprob_tokens": [ + -2.578982353210449 + ], + "logprob_reward": -0.0789823904633522, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789823904633522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15502773225307465, + 0.05706868693232536, + 0.03243184834718704, + 0.1895293891429901, + 0.13985490798950195, + 0.08605340123176575, + 0.08497559279203415, + 0.0462447889149189, + 0.13922016322612762, + 0.06959349662065506 + ], + "credit_map": [ + 0.010868050158023834, + 0.015741335228085518, + 0.023781776428222656, + 0.023835428059101105, + 0.01673775352537632, + 0.01489853486418724, + 0.028462087735533714, + 0.020702756941318512, + 0.0029272891115397215, + 0.0056410678662359715, + 0.01048427727073431, + 0.010556523688137531, + 0.004558679647743702, + 0.004327361937612295, + 0.010955926030874252, + 0.007617556024342775, + 0.003036069218069315, + 0.0038106453139334917, + 0.0045663220807909966, + 0.00469697592779994, + 0.002922678366303444, + 0.005533537827432156, + 0.003992660902440548, + 0.0038729587104171515, + 0.012955190613865852, + 0.014080938883125782, + 0.029882887378335, + 0.0227041132748127, + 0.019438648596405983, + 0.024081610143184662, + 0.042102739214897156, + 0.024283260107040405, + 0.013210454024374485, + 0.018146801739931107, + 0.022327845916152, + 0.036212872713804245, + 0.007909211330115795, + 0.012335809879004955, + 0.017572451382875443, + 0.012139462865889072, + 0.008419524878263474, + 0.013630641624331474, + 0.01131338533014059, + 0.016868405044078827, + 0.00675123231485486, + 0.007846783846616745, + 0.0107894791290164, + 0.01043394673615694, + 0.00466502970084548, + 0.007187692914158106, + 0.01345689594745636, + 0.012487887404859066, + 0.008772850967943668, + 0.007326844613999128, + 0.0191224105656147, + 0.011955979280173779, + 0.004956343676894903, + 0.0041193184442818165, + 0.0078882509842515, + 0.00689062150195241, + 0.004259863402694464, + 0.005879493430256844, + 0.005476128775626421, + 0.006774768698960543, + 0.008191154338419437, + 0.010937586426734924, + 0.01897132769227028, + 0.024574073031544685, + 0.02374662458896637, + 0.009929577820003033, + 0.027339650318026543, + 0.015530173666775227, + 0.007977774366736412, + 0.00463057542219758, + 0.009154096245765686, + 0.013481910340487957, + 0.008934869430959225, + 0.005099320784211159, + 0.011561834253370762, + 0.008753115311264992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.0789823904633522 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot of an elderly farmer inspecting crops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer wearing a faded blue denim shirt with sleeves rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the farmer extend his right hand to gently glide his fingertips across the corn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a field of tall, lush corn at dusk with a sky transitioning from orange to purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while the farmer moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00594549672678113 + ], + "opposite_logprob_tokens": [ + -5.130945682525635 + ], + "logprob_reward": -0.00594549672678113, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00594549672678113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15493017435073853, + 0.03925154358148575, + 0.11120249330997467, + 0.041857779026031494, + 0.19849660992622375, + 0.06990835070610046, + 0.10231619328260422, + 0.11554073542356491, + 0.05891134589910507, + 0.10758475959300995 + ], + "credit_map": [ + 0.012174496427178383, + 0.01817530021071434, + 0.025901611894369125, + 0.030667437240481377, + 0.010265538468956947, + 0.010534051805734634, + 0.0210537388920784, + 0.02615802362561226, + 0.0024561535101383924, + 0.0036463765427470207, + 0.006593186873942614, + 0.005735092330724001, + 0.004426725674420595, + 0.004401644226163626, + 0.007180097512900829, + 0.004812267608940601, + 0.0057489629834890366, + 0.009973259642720222, + 0.020215483382344246, + 0.021689383313059807, + 0.011670583859086037, + 0.012884288094937801, + 0.017464302480220795, + 0.011556234210729599, + 0.0029301748145371675, + 0.006093215197324753, + 0.007170484866946936, + 0.008635667152702808, + 0.0026140217669308186, + 0.003883716883137822, + 0.005210111849009991, + 0.005320391152054071, + 0.0163864828646183, + 0.0227015633136034, + 0.033531416207551956, + 0.04516199603676796, + 0.01639917679131031, + 0.021359676495194435, + 0.021891407668590546, + 0.021064886823296547, + 0.006989412009716034, + 0.0149439238011837, + 0.009261461906135082, + 0.013268235139548779, + 0.005383187904953957, + 0.0071122064255177975, + 0.0051744244992733, + 0.007775495294481516, + 0.010091808624565601, + 0.01562414038926363, + 0.016049576923251152, + 0.017139140516519547, + 0.007216759491711855, + 0.010799583047628403, + 0.016401536762714386, + 0.008993641473352909, + 0.013692466542124748, + 0.019177015870809555, + 0.014973021112382412, + 0.016754016280174255, + 0.008550768718123436, + 0.014609660021960735, + 0.013430324383080006, + 0.014353460632264614, + 0.007435220759361982, + 0.006837625056505203, + 0.007922961376607418, + 0.009153865277767181, + 0.005558309145271778, + 0.008471245877444744, + 0.006502234376966953, + 0.007029886357486248, + 0.017477046698331833, + 0.010179075412452221, + 0.012012317776679993, + 0.013650869950652122, + 0.009084898978471756, + 0.018008273094892502, + 0.01175930630415678, + 0.01541297696530819 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.00594549672678113 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot of an elderly farmer inspecting crops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer wearing a faded blue denim shirt with sleeves rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the farmer extend his right hand to gently glide his fingertips across the corn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a field of tall, lush corn at dusk with a sky transitioning from orange to purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while the farmer moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.974250078201294 + ], + "opposite_logprob_tokens": [ + -0.47425007820129395 + ], + "logprob_reward": -0.974250078201294, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.974250078201294, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.17448444664478302, + 0.042668089270591736, + 0.11889281123876572, + 0.23602531850337982, + 0.165317103266716, + 0.06401994824409485, + 0.06779249757528305, + 0.03585347533226013, + 0.0381806381046772, + 0.05676564574241638 + ], + "credit_map": [ + 0.00577678345143795, + 0.013348153792321682, + 0.040370505303144455, + 0.0433720126748085, + 0.01253987941890955, + 0.012950188480317593, + 0.024846136569976807, + 0.02128077670931816, + 0.0024761215317994356, + 0.003120630281046033, + 0.006148278247565031, + 0.0060175382532179356, + 0.003397878725081682, + 0.003966065123677254, + 0.009418508037924767, + 0.008123070001602173, + 0.00415309751406312, + 0.008494668640196323, + 0.017677713185548782, + 0.017069412395358086, + 0.015817007049918175, + 0.01597307063639164, + 0.024209417402744293, + 0.015498426742851734, + 0.01592310704290867, + 0.018550630658864975, + 0.035574015229940414, + 0.031023845076560974, + 0.022666621953248978, + 0.029646558687090874, + 0.04959271848201752, + 0.03304781764745712, + 0.007910885848104954, + 0.00810257624834776, + 0.024899084120988846, + 0.025335514917969704, + 0.02116590552031994, + 0.019473697990179062, + 0.035602085292339325, + 0.02282736450433731, + 0.005447047762572765, + 0.008855044841766357, + 0.00854636449366808, + 0.011893760412931442, + 0.007020357996225357, + 0.0063463603146374226, + 0.007960321381688118, + 0.007950692437589169, + 0.0043498678132891655, + 0.008270897902548313, + 0.010963845066726208, + 0.011422867886722088, + 0.0066468133591115475, + 0.006948795169591904, + 0.01190320122987032, + 0.007286212872713804, + 0.0039906008169054985, + 0.004364981781691313, + 0.005190620198845863, + 0.00453930301591754, + 0.002459186827763915, + 0.005525961518287659, + 0.004961761180311441, + 0.004821058828383684, + 0.003187024500221014, + 0.003265637904405594, + 0.005424878094345331, + 0.005612328182905912, + 0.0053148409351706505, + 0.0035902680829167366, + 0.007894566282629967, + 0.0038910922594368458, + 0.005224066320806742, + 0.004969348665326834, + 0.007444673217833042, + 0.010102836415171623, + 0.005438824649900198, + 0.0056746164336800575, + 0.009725648909807205, + 0.008185631595551968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.974250078201294 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "medium shot, held at a slightly low angle, slowly pans right, following an elderly farmer inspecting his crops. His face is weathered, his eyes kind and contemplative. At dusk, he stands in a field of tall, lush corn, wearing a faded blue denim shirt with the sleeves rolled up to his forearms. As he moves, he deliberately extends his calloused, work-worn right hand, letting his fingertips gently and reverently glide across the broad, green surface of the corn leaves. The setting sun, just behind him and just out of frame, casts a magnificent, warm golden glow over the entire scene, creating a brilliant rim light on his silhouette and a fiery glow on the edges of the corn leaves. The sky transitions from a vibrant orange near the horizon to a deep twilight purple above. Long shadows stretch forward across the fertile fields, creating a profoundly serene and peaceful atmosphere, evoking a sense of a day's work and a deep connection to the land.", + "seed": 22673, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_3/index_3844_seed_22673_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.14162662238813936, + "vlm_logprob": -0.14162662238813936, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.14162662238813936, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot of an elderly farmer inspecting crops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer wearing a faded blue denim shirt with sleeves rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the farmer extend his right hand to gently glide his fingertips across the corn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a field of tall, lush corn at dusk with a sky transitioning from orange to purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while the farmer moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07049930095672607 + ], + "opposite_logprob_tokens": [ + -2.8204994201660156 + ], + "logprob_reward": -0.07049930095672607, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07049930095672607, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12954503297805786, + 0.07336322218179703, + 0.05446399375796318, + 0.035769980400800705, + 0.05255798250436783, + 0.030612682923674583, + 0.08541212975978851, + 0.3164694011211395, + 0.2165716290473938, + 0.005233941599726677 + ], + "credit_map": [ + 0.014198853634297848, + 0.02993454411625862, + 0.006973035167902708, + 0.011603046208620071, + 0.03138095885515213, + 0.011244176886975765, + 0.014095542952418327, + 0.01011486817151308, + 0.004059783648699522, + 0.0211821049451828, + 0.003173082834109664, + 0.0061314902268350124, + 0.021475892513990402, + 0.0035518144723027945, + 0.010047071613371372, + 0.0037419823929667473, + 0.002710747765377164, + 0.014856892637908459, + 0.0023375835735350847, + 0.0026888654101639986, + 0.014861004427075386, + 0.00747664412483573, + 0.002976047107949853, + 0.006556210108101368, + 0.0035728062503039837, + 0.004489140119403601, + 0.005539200734347105, + 0.004578244872391224, + 0.004833253100514412, + 0.004688743501901627, + 0.004164355341345072, + 0.0039042343851178885, + 0.0026863371022045612, + 0.015042652375996113, + 0.003633843269199133, + 0.003152283141389489, + 0.014942290261387825, + 0.004079117439687252, + 0.003906977828592062, + 0.0051144822500646114, + 0.0028677782975137234, + 0.005076202098280191, + 0.003190268063917756, + 0.002592073753476143, + 0.005271722096949816, + 0.005340176168829203, + 0.0028805348556488752, + 0.0033939285203814507, + 0.008714779280126095, + 0.020007513463497162, + 0.004355281125754118, + 0.006699464749544859, + 0.023593466728925705, + 0.007822202518582344, + 0.006815479602664709, + 0.007403942756354809, + 0.03592861816287041, + 0.06603920459747314, + 0.026876945048570633, + 0.019642004743218422, + 0.07616116851568222, + 0.044721297919750214, + 0.021570635959506035, + 0.025529490783810616, + 0.035418249666690826, + 0.024894999340176582, + 0.019484393298625946, + 0.02431669645011425, + 0.029673323035240173, + 0.030484182760119438, + 0.02557661570608616, + 0.026723170652985573, + 0.0006815585657022893, + 0.0006843448500148952, + 0.0004002198693342507, + 0.0006067724316380918, + 0.0007667181780561805, + 0.0008410714217461646, + 0.0006197405746206641, + 0.0006335152429528534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.07049930095672607 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot of an elderly farmer inspecting crops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer wearing a faded blue denim shirt with sleeves rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the farmer extend his right hand to gently glide his fingertips across the corn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a field of tall, lush corn at dusk with a sky transitioning from orange to purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while the farmer moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020603561773896217 + ], + "opposite_logprob_tokens": [ + -3.895603656768799 + ], + "logprob_reward": -0.020603561773896217, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020603561773896217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10326119512319565, + 0.04187959060072899, + 0.07608173787593842, + 0.04606582969427109, + 0.1340041309595108, + 0.2854435443878174, + 0.05231296643614769, + 0.0952954962849617, + 0.11068781465291977, + 0.054967716336250305 + ], + "credit_map": [ + 0.01266483124345541, + 0.009905138984322548, + 0.011741705238819122, + 0.013773316517472267, + 0.011674604378640652, + 0.016756294295191765, + 0.015379806980490685, + 0.011365494690835476, + 0.004983291029930115, + 0.007108953315764666, + 0.002575822174549103, + 0.005108521785587072, + 0.0070480103604495525, + 0.004713431000709534, + 0.00576499430462718, + 0.0045765661634504795, + 0.00554437842220068, + 0.011652487330138683, + 0.008735133334994316, + 0.007742634043097496, + 0.01140665728598833, + 0.011334296315908432, + 0.012281541712582111, + 0.007384606171399355, + 0.005043265875428915, + 0.006060476414859295, + 0.003919488750398159, + 0.00443869037553668, + 0.0071608563885092735, + 0.007057120557874441, + 0.00664482032880187, + 0.0057411156594753265, + 0.007166593801230192, + 0.04190487414598465, + 0.007482666056603193, + 0.008347955532371998, + 0.04166286066174507, + 0.01357245072722435, + 0.006101840641349554, + 0.007764883805066347, + 0.01689888723194599, + 0.08579310774803162, + 0.01758459024131298, + 0.019773652777075768, + 0.08695428818464279, + 0.020756658166646957, + 0.017880478873848915, + 0.019801899790763855, + 0.007610830944031477, + 0.008307150565087795, + 0.0049211191944777966, + 0.0038651064969599247, + 0.009265348315238953, + 0.0073682405054569244, + 0.004876627586781979, + 0.006098541896790266, + 0.006899374071508646, + 0.020395267754793167, + 0.006112368777394295, + 0.006435731425881386, + 0.02467614971101284, + 0.016135966405272484, + 0.007597917690873146, + 0.0070427181199193, + 0.009959129616618156, + 0.029968783259391785, + 0.008007860742509365, + 0.0060689388774335384, + 0.030795397236943245, + 0.010026706382632256, + 0.007071000523865223, + 0.008790001273155212, + 0.005676473490893841, + 0.010308569297194481, + 0.0027390134055167437, + 0.0058257137425243855, + 0.011600986123085022, + 0.008903862908482552, + 0.004834329709410667, + 0.005078773945569992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 8 + ], + "selected_reward": -0.020603561773896217 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot of an elderly farmer inspecting crops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer wearing a faded blue denim shirt with sleeves rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the farmer extend his right hand to gently glide his fingertips across the corn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a field of tall, lush corn at dusk with a sky transitioning from orange to purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while the farmer moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038091447204351425 + ], + "opposite_logprob_tokens": [ + -3.2880914211273193 + ], + "logprob_reward": -0.038091447204351425, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038091447204351425, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24905619025230408, + 0.06138013303279877, + 0.12905888259410858, + 0.08958571404218674, + 0.03385134041309357, + 0.02941024675965309, + 0.07813838869333267, + 0.14075937867164612, + 0.11893251538276672, + 0.06982724368572235 + ], + "credit_map": [ + 0.013858052901923656, + 0.07772360742092133, + 0.009076704271137714, + 0.019145527854561806, + 0.08143652230501175, + 0.01249183714389801, + 0.023067228496074677, + 0.012256709858775139, + 0.008348512463271618, + 0.012166147120296955, + 0.004375898744910955, + 0.0055816988460719585, + 0.012290189042687416, + 0.008257072418928146, + 0.004692078568041325, + 0.0056685335002839565, + 0.004779845010489225, + 0.04494098573923111, + 0.004453576169908047, + 0.006773270666599274, + 0.043847937136888504, + 0.007815783843398094, + 0.010805675759911537, + 0.005641811527311802, + 0.005603439640253782, + 0.03299896791577339, + 0.005294422153383493, + 0.004510907921940088, + 0.03183961659669876, + 0.0032624215818941593, + 0.0032475111074745655, + 0.0028284203726798296, + 0.0032629554625600576, + 0.008068817667663097, + 0.003101724199950695, + 0.004024887923151255, + 0.00788615271449089, + 0.0031288480386137962, + 0.0021778636146336794, + 0.0022000905591994524, + 0.0037115563172847033, + 0.0043954430148005486, + 0.0029603797011077404, + 0.0029502934776246548, + 0.006190060172230005, + 0.0028533143922686577, + 0.0031112716533243656, + 0.0032379303593188524, + 0.010767161846160889, + 0.013459227979183197, + 0.005185991059988737, + 0.005078800022602081, + 0.015955191105604172, + 0.009589767083525658, + 0.009434104897081852, + 0.0086681442335248, + 0.018641473725438118, + 0.032789409160614014, + 0.01218610629439354, + 0.00872353371232748, + 0.03760446980595589, + 0.014937576837837696, + 0.00740963127464056, + 0.008467189967632294, + 0.014352498576045036, + 0.026814382523298264, + 0.007795850280672312, + 0.009454083628952503, + 0.027423551306128502, + 0.010767998173832893, + 0.010432890616357327, + 0.011891257017850876, + 0.007933392189443111, + 0.007743733935058117, + 0.005505686160176992, + 0.00769028952345252, + 0.009732632897794247, + 0.013817154802381992, + 0.007714713457971811, + 0.009689641185104847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.038091447204351425 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot of an elderly farmer inspecting crops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer wearing a faded blue denim shirt with sleeves rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the farmer extend his right hand to gently glide his fingertips across the corn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a field of tall, lush corn at dusk with a sky transitioning from orange to purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while the farmer moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028178293723613024 + ], + "opposite_logprob_tokens": [ + -5.877817630767822 + ], + "logprob_reward": -0.0028178293723613024, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028178293723613024, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1497989445924759, + 0.08879631012678146, + 0.26291245222091675, + 0.09263011813163757, + 0.11328738182783127, + 0.02723957970738411, + 0.0913853645324707, + 0.0722096860408783, + 0.07493875175714493, + 0.026801398023962975 + ], + "credit_map": [ + 0.007731155958026648, + 0.051732387393713, + 0.004830085672438145, + 0.008892139419913292, + 0.055704936385154724, + 0.007204940542578697, + 0.007759407162666321, + 0.005943900439888239, + 0.0053640748374164104, + 0.016948342323303223, + 0.006820376962423325, + 0.013443519361317158, + 0.017231062054634094, + 0.007050459738820791, + 0.015025759115815163, + 0.0069127073511481285, + 0.011390097439289093, + 0.08176032453775406, + 0.018105966970324516, + 0.012935924343764782, + 0.08201706409454346, + 0.0256436038762331, + 0.015305782668292522, + 0.015753667801618576, + 0.00610297592356801, + 0.028907891362905502, + 0.007002132944762707, + 0.007345838472247124, + 0.02905762381851673, + 0.0054586040787398815, + 0.005200343672186136, + 0.0035546959843486547, + 0.004921653773635626, + 0.03121124394237995, + 0.007010613102465868, + 0.007374640088528395, + 0.03143611550331116, + 0.011414822190999985, + 0.010438216850161552, + 0.009480081498622894, + 0.002214504638686776, + 0.005649695172905922, + 0.002556626917794347, + 0.0026142823044210672, + 0.006530143320560455, + 0.0031186325941234827, + 0.002047906629741192, + 0.002507788361981511, + 0.011569838039577007, + 0.024144263938069344, + 0.005009179934859276, + 0.007010660599917173, + 0.02896789275109768, + 0.0057445731945335865, + 0.005280321463942528, + 0.0036586353089660406, + 0.0086231529712677, + 0.016018973663449287, + 0.00618812395259738, + 0.004887920804321766, + 0.018322616815567017, + 0.008692333474755287, + 0.004357395693659782, + 0.005119169596582651, + 0.009475402534008026, + 0.015264999121427536, + 0.00539826974272728, + 0.005533173214644194, + 0.01590990275144577, + 0.008913001045584679, + 0.0070083788596093655, + 0.007435630541294813, + 0.0023814670275896788, + 0.005723692011088133, + 0.0013849296374246478, + 0.0026401907671242952, + 0.00651299674063921, + 0.003165092784911394, + 0.0026908807922154665, + 0.0023021467495709658 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.0028178293723613024 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot of an elderly farmer inspecting crops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer wearing a faded blue denim shirt with sleeves rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the farmer extend his right hand to gently glide his fingertips across the corn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a field of tall, lush corn at dusk with a sky transitioning from orange to purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while the farmer moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761209726333618 + ], + "opposite_logprob_tokens": [ + -0.8261209726333618 + ], + "logprob_reward": -0.5761209726333618, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761209726333618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2940692901611328, + 0.1536819487810135, + 0.031347449868917465, + 0.06367306411266327, + 0.04917536675930023, + 0.01975199766457081, + 0.1308448165655136, + 0.0690581426024437, + 0.10738526284694672, + 0.08101271837949753 + ], + "credit_map": [ + 0.012614747509360313, + 0.07775368541479111, + 0.012262774631381035, + 0.03016512840986252, + 0.08372119814157486, + 0.017406120896339417, + 0.041195545345544815, + 0.01895006000995636, + 0.006580706685781479, + 0.04481658339500427, + 0.006893626879900694, + 0.01339352410286665, + 0.046441011130809784, + 0.00836254097521305, + 0.01834222860634327, + 0.008851728402078152, + 0.0019097530748695135, + 0.008082372136414051, + 0.0013557280180975795, + 0.0021085997577756643, + 0.00817708857357502, + 0.0030534681864082813, + 0.003489574883133173, + 0.0031708627939224243, + 0.0033219854813069105, + 0.02228052355349064, + 0.0033984354231506586, + 0.003201924031600356, + 0.022028189152479172, + 0.003669429337605834, + 0.003587438492104411, + 0.0021851351484656334, + 0.0035409322008490562, + 0.012632619589567184, + 0.003420454915612936, + 0.005314900539815426, + 0.012531089596450329, + 0.004922636318951845, + 0.003291176864877343, + 0.003521550679579377, + 0.0014222889440134168, + 0.004502932075411081, + 0.0019308965420350432, + 0.0022727574687451124, + 0.004681624472141266, + 0.0021206196397542953, + 0.0013997819041833282, + 0.0014210952213034034, + 0.017632661387324333, + 0.032020196318626404, + 0.008170416578650475, + 0.007707257755100727, + 0.03910123556852341, + 0.010511758737266064, + 0.009431058540940285, + 0.006270229816436768, + 0.005068856291472912, + 0.009346522390842438, + 0.006373005453497171, + 0.008146905340254307, + 0.011952205561101437, + 0.011178732849657536, + 0.009495365433394909, + 0.0074965497478842735, + 0.00995735265314579, + 0.02297651208937168, + 0.0067246658727526665, + 0.00849435105919838, + 0.023511158302426338, + 0.011904739774763584, + 0.00958877895027399, + 0.01422770693898201, + 0.008309107273817062, + 0.009300760924816132, + 0.006356412544846535, + 0.00998669397085905, + 0.012118184007704258, + 0.015640826895833015, + 0.009842483326792717, + 0.009458248503506184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.5761209726333618 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, held at a slightly low angle, slowly pans right, following an elderly farmer inspecting his crops. His face is weathered, his eyes kind and contemplative. At dusk, he stands in a field of tall, lush corn, wearing a faded blue denim shirt with the sleeves rolled up to his forearms. As he moves, he deliberately extends his calloused, work-worn right hand, letting his fingertips gently and reverently glide across the broad, green surface of the corn leaves. The setting sun, just behind him and just out of frame, casts a magnificent, warm golden glow over the entire scene, creating a brilliant rim light on his silhouette and a fiery glow on the edges of the corn leaves. The sky transitions from a vibrant orange near the horizon to a deep twilight purple above. Long shadows stretch forward across the fertile fields, creating a profoundly serene and peaceful atmosphere, evoking a sense of a day's work and a deep connection to the land.", + "seed": 22674, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_3/index_3844_seed_22674_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.19607006181031467, + "vlm_logprob": -0.19607006181031467, + "vlm_yes_no_margin": 2.525, + "vlm_token_credit": -0.19607006181031467, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot of an elderly farmer inspecting crops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer wearing a faded blue denim shirt with sleeves rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the farmer extend his right hand to gently glide his fingertips across the corn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a field of tall, lush corn at dusk with a sky transitioning from orange to purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while the farmer moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10764394700527191 + ], + "opposite_logprob_tokens": [ + -2.3576438426971436 + ], + "logprob_reward": -0.10764394700527191, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10764394700527191, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09934588521718979, + 0.03835074231028557, + 0.05011821165680885, + 0.12265471369028091, + 0.08452191203832626, + 0.25018903613090515, + 0.03256230801343918, + 0.08068758249282837, + 0.06368721276521683, + 0.1778823733329773 + ], + "credit_map": [ + 0.010687756352126598, + 0.011006530374288559, + 0.0075338659808039665, + 0.00869692675769329, + 0.010250980965793133, + 0.027990663424134254, + 0.007857273332774639, + 0.015321888960897923, + 0.005016862414777279, + 0.00416555255651474, + 0.0028077401220798492, + 0.003715106286108494, + 0.004390025977045298, + 0.00678632827475667, + 0.006122603081166744, + 0.00534652266651392, + 0.007007034495472908, + 0.006641679909080267, + 0.006392564624547958, + 0.007314277347177267, + 0.004024942871183157, + 0.00798784103244543, + 0.005115403328090906, + 0.005634468514472246, + 0.015336744487285614, + 0.01487475074827671, + 0.010235277004539967, + 0.010733706876635551, + 0.009270314127206802, + 0.02367071807384491, + 0.01714320294559002, + 0.021390004083514214, + 0.00914059765636921, + 0.00958880316466093, + 0.004507542587816715, + 0.005311434622853994, + 0.007125220261514187, + 0.019592951983213425, + 0.013379836454987526, + 0.015875523909926414, + 0.07198310643434525, + 0.09009430557489395, + 0.008295384235680103, + 0.010235633701086044, + 0.013843808323144913, + 0.03117566742002964, + 0.010193555615842342, + 0.01436756830662489, + 0.0049383617006242275, + 0.004021675791591406, + 0.005042483098804951, + 0.004176388960331678, + 0.003438535612076521, + 0.004424060229212046, + 0.0033703558146953583, + 0.0031504484359174967, + 0.008020740002393723, + 0.00613793171942234, + 0.004748297855257988, + 0.007194143254309893, + 0.0058191753923892975, + 0.016958992928266525, + 0.009116584435105324, + 0.022691715508699417, + 0.0067630489356815815, + 0.00706973671913147, + 0.005131682381033897, + 0.0062669748440384865, + 0.004502065945416689, + 0.009475599974393845, + 0.010067375376820564, + 0.014410733245313168, + 0.03879985585808754, + 0.038793738931417465, + 0.007607335690408945, + 0.01030827034264803, + 0.012363137677311897, + 0.024725299328565598, + 0.008605956099927425, + 0.03667876869440079 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 3 + ], + "selected_reward": -0.10764394700527191 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot of an elderly farmer inspecting crops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer wearing a faded blue denim shirt with sleeves rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the farmer extend his right hand to gently glide his fingertips across the corn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a field of tall, lush corn at dusk with a sky transitioning from orange to purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while the farmer moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430886335670948 + ], + "opposite_logprob_tokens": [ + -3.168088674545288 + ], + "logprob_reward": -0.0430886335670948, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430886335670948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21285638213157654, + 0.09523409605026245, + 0.05707449093461037, + 0.14643065631389618, + 0.13812348246574402, + 0.14077715575695038, + 0.07604985684156418, + 0.044657737016677856, + 0.03393108770251274, + 0.05486508831381798 + ], + "credit_map": [ + 0.01913449726998806, + 0.024189282208681107, + 0.026075338944792747, + 0.023205861449241638, + 0.02365841716527939, + 0.055209968239068985, + 0.017420899122953415, + 0.02396211586892605, + 0.012268473394215107, + 0.01108715683221817, + 0.010319926775991917, + 0.012227727100253105, + 0.007143088616430759, + 0.018385766074061394, + 0.011203148402273655, + 0.012598804198205471, + 0.007646400481462479, + 0.006966155953705311, + 0.006912482436746359, + 0.008737262338399887, + 0.004715331364423037, + 0.0094844875857234, + 0.006261859089136124, + 0.006350507959723473, + 0.02487872913479805, + 0.02856265753507614, + 0.010224565863609314, + 0.009968467988073826, + 0.012243698351085186, + 0.019773755222558975, + 0.021939106285572052, + 0.018839659169316292, + 0.022478658705949783, + 0.022280938923358917, + 0.00823204591870308, + 0.008670070208609104, + 0.009333421476185322, + 0.028119122609496117, + 0.017639031633734703, + 0.021370185539126396, + 0.024335164576768875, + 0.030718935653567314, + 0.007291762623935938, + 0.007325338665395975, + 0.01594461500644684, + 0.022503549233078957, + 0.012967121787369251, + 0.019690660759806633, + 0.013901736587285995, + 0.01751011423766613, + 0.003914845176041126, + 0.004546500276774168, + 0.006912766955792904, + 0.011552711017429829, + 0.008310090750455856, + 0.009401090443134308, + 0.005125882104039192, + 0.007519870530813932, + 0.003416931489482522, + 0.003669748781248927, + 0.0032811074052006006, + 0.009935460053384304, + 0.0039055331144481897, + 0.007803203072398901, + 0.0030583087354898453, + 0.0032863784581422806, + 0.00412046629935503, + 0.004592715762555599, + 0.003334315260872245, + 0.004783243406563997, + 0.003971864935010672, + 0.006783793214708567, + 0.006534455809742212, + 0.006361464969813824, + 0.0039067394100129604, + 0.0034289283212274313, + 0.00835901964455843, + 0.00894023198634386, + 0.008161499164998531, + 0.00917274784296751 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.0430886335670948 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot of an elderly farmer inspecting crops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer wearing a faded blue denim shirt with sleeves rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the farmer extend his right hand to gently glide his fingertips across the corn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a field of tall, lush corn at dusk with a sky transitioning from orange to purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while the farmer moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204694136977196 + ], + "opposite_logprob_tokens": [ + -2.812047004699707 + ], + "logprob_reward": -0.06204694136977196, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204694136977196, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1045914962887764, + 0.09179846197366714, + 0.10264208167791367, + 0.233852356672287, + 0.10450208187103271, + 0.034926142543554306, + 0.05648933723568916, + 0.057864923030138016, + 0.08308439701795578, + 0.13024872541427612 + ], + "credit_map": [ + 0.016527945175766945, + 0.01326796505600214, + 0.010068793781101704, + 0.00956328958272934, + 0.0095893032848835, + 0.022989943623542786, + 0.008909950964152813, + 0.013674301095306873, + 0.015686551108956337, + 0.017277127131819725, + 0.009231816045939922, + 0.010774794034659863, + 0.006967454683035612, + 0.01349648553878069, + 0.008297037333250046, + 0.010067199356853962, + 0.012033454142510891, + 0.012006483972072601, + 0.01425182819366455, + 0.015979617834091187, + 0.008519082330167294, + 0.01556566171348095, + 0.011568273417651653, + 0.012717672623693943, + 0.04042181745171547, + 0.047880396246910095, + 0.013470095582306385, + 0.01210935227572918, + 0.013103906996548176, + 0.04600568488240242, + 0.031069882214069366, + 0.029791222885251045, + 0.019621871411800385, + 0.01920425519347191, + 0.007491068448871374, + 0.00590888038277626, + 0.00816784892231226, + 0.017371399328112602, + 0.014375627040863037, + 0.012361129745841026, + 0.00478754285722971, + 0.00617654575034976, + 0.0025727914180606604, + 0.0029138519894331694, + 0.004160082899034023, + 0.006707449443638325, + 0.004411409609019756, + 0.0031964692752808332, + 0.007015124429017305, + 0.00692092627286911, + 0.009013880044221878, + 0.008611396886408329, + 0.0045478492975234985, + 0.008409470319747925, + 0.005819827783852816, + 0.006150863133370876, + 0.007149338256567717, + 0.007840360514819622, + 0.007032237946987152, + 0.008095340803265572, + 0.00715439859777689, + 0.008807416073977947, + 0.005392618477344513, + 0.006393210496753454, + 0.01142874825745821, + 0.017773453146219254, + 0.003613334149122238, + 0.005192859563976526, + 0.007424784824252129, + 0.015382696874439716, + 0.009438025765120983, + 0.012830495834350586, + 0.03526010364294052, + 0.0382382869720459, + 0.004605210851877928, + 0.006688093300908804, + 0.006596820894628763, + 0.014349556528031826, + 0.005143553949892521, + 0.01936710998415947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.06204694136977196 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot of an elderly farmer inspecting crops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer wearing a faded blue denim shirt with sleeves rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the farmer extend his right hand to gently glide his fingertips across the corn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a field of tall, lush corn at dusk with a sky transitioning from orange to purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while the farmer moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780817665159702 + ], + "opposite_logprob_tokens": [ + -4.6347808837890625 + ], + "logprob_reward": -0.009780817665159702, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780817665159702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14737765491008759, + 0.10290532559156418, + 0.0701475590467453, + 0.10048115998506546, + 0.045143015682697296, + 0.0836246907711029, + 0.13130392134189606, + 0.25127166509628296, + 0.03361942991614342, + 0.03412555903196335 + ], + "credit_map": [ + 0.017656031996011734, + 0.01959163136780262, + 0.012920155189931393, + 0.012706448324024677, + 0.01328328624367714, + 0.0465276800096035, + 0.00951550342142582, + 0.015176926739513874, + 0.013835306279361248, + 0.015533321537077427, + 0.013429297134280205, + 0.01524844579398632, + 0.008486498147249222, + 0.01470454502850771, + 0.011932695284485817, + 0.009735222905874252, + 0.010302778333425522, + 0.00991988554596901, + 0.0074596405029296875, + 0.008736551739275455, + 0.008479416370391846, + 0.012884771451354027, + 0.006266636308282614, + 0.006097882054746151, + 0.024880358949303627, + 0.026890182867646217, + 0.00827815756201744, + 0.010877816937863827, + 0.006527812220156193, + 0.0096889054402709, + 0.007048330269753933, + 0.0062896013259887695, + 0.008586272597312927, + 0.010216743685305119, + 0.002855550730600953, + 0.0030420678667724133, + 0.0027460590936243534, + 0.008119501173496246, + 0.004050917457789183, + 0.005525902844965458, + 0.02316853404045105, + 0.028255140408873558, + 0.0067490702494978905, + 0.006321643479168415, + 0.003979699686169624, + 0.008244396187365055, + 0.002845205832272768, + 0.004061004146933556, + 0.034721553325653076, + 0.038563456386327744, + 0.006805457174777985, + 0.007836660370230675, + 0.013826338574290276, + 0.012559711001813412, + 0.008464396931231022, + 0.008526350371539593, + 0.027149230241775513, + 0.02407843805849552, + 0.022520694881677628, + 0.03592868149280548, + 0.02119891345500946, + 0.04410894587635994, + 0.02243124321103096, + 0.05385550111532211, + 0.004709844011813402, + 0.005008307285606861, + 0.0028115781024098396, + 0.003125633578747511, + 0.00226677767932415, + 0.0067839776165783405, + 0.0027016072999686003, + 0.006211703643202782, + 0.009104294702410698, + 0.009035240858793259, + 0.0015512360259890556, + 0.002006722614169121, + 0.0026697961147874594, + 0.0035949277225881815, + 0.0014688720693811774, + 0.004694467876106501 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.009780817665159702 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot of an elderly farmer inspecting crops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer wearing a faded blue denim shirt with sleeves rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the farmer extend his right hand to gently glide his fingertips across the corn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a field of tall, lush corn at dusk with a sky transitioning from orange to purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while the farmer moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7577899694442749 + ], + "opposite_logprob_tokens": [ + -0.6327899694442749 + ], + "logprob_reward": -0.7577899694442749, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7577899694442749, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.19308219850063324, + 0.027949323877692223, + 0.06867605447769165, + 0.14854778349399567, + 0.20328618586063385, + 0.11842013895511627, + 0.07561451196670532, + 0.050468262284994125, + 0.07070769369602203, + 0.04324788972735405 + ], + "credit_map": [ + 0.018272632732987404, + 0.021518971771001816, + 0.022559350356459618, + 0.015135393477976322, + 0.020264893770217896, + 0.06077175587415695, + 0.015454739332199097, + 0.01910446397960186, + 0.0046045538038015366, + 0.004308599978685379, + 0.0030377658549696207, + 0.0033476222306489944, + 0.0028046858496963978, + 0.004196764901280403, + 0.0023393479641526937, + 0.0033099849242717028, + 0.00911838747560978, + 0.008637458086013794, + 0.00844775140285492, + 0.010201742872595787, + 0.004893062636256218, + 0.011950305663049221, + 0.007706773467361927, + 0.007720576599240303, + 0.03896341472864151, + 0.04367906227707863, + 0.007949494756758213, + 0.00882759504020214, + 0.007303695194423199, + 0.019276821985840797, + 0.013888020068407059, + 0.008659681305289268, + 0.0409218929708004, + 0.044298093765974045, + 0.00852248352020979, + 0.009153326041996479, + 0.017149776220321655, + 0.0323464497923851, + 0.027017943561077118, + 0.023876236751675606, + 0.03275261074304581, + 0.039150189608335495, + 0.007932361215353012, + 0.008822125382721424, + 0.006490569096058607, + 0.013382598757743835, + 0.004559236112982035, + 0.00533045269548893, + 0.01751433126628399, + 0.02357514388859272, + 0.003833143273368478, + 0.005214022938162088, + 0.005583532154560089, + 0.011898765340447426, + 0.004433450289070606, + 0.0035621244460344315, + 0.004736395552754402, + 0.005077797453850508, + 0.00489741051569581, + 0.005441245157271624, + 0.006272032856941223, + 0.008396881632506847, + 0.005759959574788809, + 0.009886541403830051, + 0.00707192812114954, + 0.010608168318867683, + 0.005283463280647993, + 0.005867499392479658, + 0.00562512269243598, + 0.011278531514108181, + 0.009520020335912704, + 0.015452963300049305, + 0.007363658398389816, + 0.006749786902219057, + 0.001800422789528966, + 0.002774806460365653, + 0.002688087522983551, + 0.007627454586327076, + 0.002691777655854821, + 0.011551897041499615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.7577899694442749 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "medium shot, held at a slightly low angle, slowly pans right, following an elderly farmer inspecting his crops. His face is weathered, his eyes kind and contemplative. At dusk, he stands in a field of tall, lush corn, wearing a faded blue denim shirt with the sleeves rolled up to his forearms. As he moves, he deliberately extends his calloused, work-worn right hand, letting his fingertips gently and reverently glide across the broad, green surface of the corn leaves. The setting sun, just behind him and just out of frame, casts a magnificent, warm golden glow over the entire scene, creating a brilliant rim light on his silhouette and a fiery glow on the edges of the corn leaves. The sky transitions from a vibrant orange near the horizon to a deep twilight purple above. Long shadows stretch forward across the fertile fields, creating a profoundly serene and peaceful atmosphere, evoking a sense of a day's work and a deep connection to the land.", + "seed": 22675, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_3/index_3844_seed_22675_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0848793052136898, + "vlm_logprob": -0.0848793052136898, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.0848793052136898, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot of an elderly farmer inspecting crops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer wearing a faded blue denim shirt with sleeves rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the farmer extend his right hand to gently glide his fingertips across the corn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a field of tall, lush corn at dusk with a sky transitioning from orange to purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while the farmer moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03155648335814476 + ], + "opposite_logprob_tokens": [ + -3.6565563678741455 + ], + "logprob_reward": -0.03155648335814476, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03155648335814476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11895949393510818, + 0.1436377763748169, + 0.15799571573734283, + 0.19221581518650055, + 0.09990928322076797, + 0.1470891684293747, + 0.04646715521812439, + 0.020083045586943626, + 0.036102671176195145, + 0.03753988444805145 + ], + "credit_map": [ + 0.017181556671857834, + 0.008313327096402645, + 0.01724258065223694, + 0.018932949751615524, + 0.011975539848208427, + 0.018096977844834328, + 0.013961153104901314, + 0.013255414552986622, + 0.01672826148569584, + 0.015971031039953232, + 0.025593938305974007, + 0.014608554542064667, + 0.015588806942105293, + 0.011060579679906368, + 0.02450186014175415, + 0.019584747031331062, + 0.026821695268154144, + 0.014050540514290333, + 0.019770247861742973, + 0.0350344255566597, + 0.022012529894709587, + 0.009147875010967255, + 0.018619490787386894, + 0.012538912706077099, + 0.015993012115359306, + 0.018013427034020424, + 0.03827774152159691, + 0.011679721996188164, + 0.05897990241646767, + 0.023964818567037582, + 0.016907602548599243, + 0.008399568498134613, + 0.014572887681424618, + 0.01090223342180252, + 0.012833647429943085, + 0.016089895740151405, + 0.011642299592494965, + 0.006253127008676529, + 0.010810306295752525, + 0.016804886981844902, + 0.02574833668768406, + 0.023669542744755745, + 0.02449602633714676, + 0.01952025294303894, + 0.013940398581326008, + 0.009597172960639, + 0.012428120709955692, + 0.017689311876893044, + 0.0035358350723981857, + 0.007514643017202616, + 0.004887659102678299, + 0.006620784290134907, + 0.0051550427451729774, + 0.004660279955714941, + 0.007197421044111252, + 0.006895489990711212, + 0.0019629881717264652, + 0.0029355413280427456, + 0.0021237472537904978, + 0.0018580761970952153, + 0.0036690852139145136, + 0.002767427358776331, + 0.0020404241513460875, + 0.0027257574256509542, + 0.0019267159514129162, + 0.004288047552108765, + 0.003624560311436653, + 0.0038292002864181995, + 0.004294330254197121, + 0.00773798068985343, + 0.004356703720986843, + 0.006045130081474781, + 0.007306812796741724, + 0.003756824880838394, + 0.006666642148047686, + 0.0035943016409873962, + 0.004949134774506092, + 0.0036301733925938606, + 0.003109284909442067, + 0.004526710603386164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 5 + ], + "selected_reward": -0.03155648335814476 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot of an elderly farmer inspecting crops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer wearing a faded blue denim shirt with sleeves rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the farmer extend his right hand to gently glide his fingertips across the corn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a field of tall, lush corn at dusk with a sky transitioning from orange to purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while the farmer moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009792503900825977 + ], + "opposite_logprob_tokens": [ + -4.634792327880859 + ], + "logprob_reward": -0.009792503900825977, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009792503900825977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3110751211643219, + 0.27722907066345215, + 0.05432044714689255, + 0.0518609844148159, + 0.05596194788813591, + 0.09027612954378128, + 0.02707708068192005, + 0.022316085174679756, + 0.07382321357727051, + 0.036059893667697906 + ], + "credit_map": [ + 0.028763780370354652, + 0.02129121497273445, + 0.052729737013578415, + 0.04187646880745888, + 0.03565387800335884, + 0.05752165615558624, + 0.03318917006254196, + 0.040049221366643906, + 0.013252855278551579, + 0.02079925499856472, + 0.025643380358815193, + 0.01245332695543766, + 0.07735872268676758, + 0.036082688719034195, + 0.06283245235681534, + 0.028806397691369057, + 0.007291522808372974, + 0.004736100323498249, + 0.007754347287118435, + 0.015296543017029762, + 0.004251888487488031, + 0.0034010394010692835, + 0.00693742698058486, + 0.0046515800058841705, + 0.004326102789491415, + 0.004348870366811752, + 0.009093781001865864, + 0.00347268208861351, + 0.015243489295244217, + 0.007089346647262573, + 0.005655392073094845, + 0.0026313175912946463, + 0.006707742344588041, + 0.005763010587543249, + 0.007757311221212149, + 0.010246890597045422, + 0.005029510706663132, + 0.00456569530069828, + 0.006918379571288824, + 0.008973407559096813, + 0.01587628573179245, + 0.013125618919730186, + 0.016346026211977005, + 0.012459264136850834, + 0.009955891408026218, + 0.005624755285680294, + 0.0072246696799993515, + 0.00966362189501524, + 0.002744625322520733, + 0.0049476842395961285, + 0.0026583231519907713, + 0.002744086552411318, + 0.0032755748834460974, + 0.00471101701259613, + 0.0035748230293393135, + 0.002420947654172778, + 0.001976702595129609, + 0.0034631381276994944, + 0.0030079390853643417, + 0.0024755308404564857, + 0.0034180604852735996, + 0.002282052766531706, + 0.0027242449577897787, + 0.00296841561794281, + 0.0025057035963982344, + 0.007837751880288124, + 0.0069276695139706135, + 0.007599480915814638, + 0.006318125408142805, + 0.01766888052225113, + 0.008368450216948986, + 0.016597161069512367, + 0.003588687162846327, + 0.004188067279756069, + 0.006103911437094212, + 0.0029050654266029596, + 0.006722093094140291, + 0.003860177705064416, + 0.00388492364436388, + 0.004806966055184603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.009792503900825977 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot of an elderly farmer inspecting crops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer wearing a faded blue denim shirt with sleeves rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the farmer extend his right hand to gently glide his fingertips across the corn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a field of tall, lush corn at dusk with a sky transitioning from orange to purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while the farmer moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978871762752533 + ], + "opposite_logprob_tokens": [ + -3.5297887325286865 + ], + "logprob_reward": -0.02978871762752533, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978871762752533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08028652518987656, + 0.07256688177585602, + 0.08393097668886185, + 0.2599388360977173, + 0.11790204048156738, + 0.17098428308963776, + 0.05601269751787186, + 0.06355346739292145, + 0.05666187405586243, + 0.038162458688020706 + ], + "credit_map": [ + 0.010261583141982555, + 0.007690927479416132, + 0.012823621742427349, + 0.009229063987731934, + 0.00965039897710085, + 0.014047677628695965, + 0.007661463692784309, + 0.008921780623495579, + 0.004826965741813183, + 0.006755039095878601, + 0.01035887561738491, + 0.006156529299914837, + 0.014365274459123611, + 0.009206952527165413, + 0.0133897103369236, + 0.00750753004103899, + 0.008141497150063515, + 0.009880980476737022, + 0.011364995501935482, + 0.01891084387898445, + 0.007516166195273399, + 0.0043998416513204575, + 0.01483297161757946, + 0.008883674629032612, + 0.016056951135396957, + 0.025604281574487686, + 0.046392254531383514, + 0.009639852680265903, + 0.09100008010864258, + 0.04050777107477188, + 0.020819557830691338, + 0.009918068535625935, + 0.01381195243448019, + 0.011450041085481644, + 0.01725814864039421, + 0.01963670179247856, + 0.012788369320333004, + 0.007907025516033173, + 0.01221254002302885, + 0.02283724583685398, + 0.028833186253905296, + 0.02403625100851059, + 0.03080044873058796, + 0.023209434002637863, + 0.020454222336411476, + 0.010205553844571114, + 0.014341193251311779, + 0.019103992730379105, + 0.007022975478321314, + 0.004906516522169113, + 0.00681741489097476, + 0.007763330824673176, + 0.0048805647529661655, + 0.004224215634167194, + 0.012884520925581455, + 0.007513158489018679, + 0.005913245491683483, + 0.010594206862151623, + 0.00715508870780468, + 0.007022173143923283, + 0.011318526230752468, + 0.004741868935525417, + 0.009476177394390106, + 0.007332171779125929, + 0.0049719964154064655, + 0.007292449474334717, + 0.005166548769921064, + 0.013347895815968513, + 0.0078058382496237755, + 0.004475231282413006, + 0.007705206051468849, + 0.005896702874451876, + 0.004955533426254988, + 0.003250535111874342, + 0.004444582387804985, + 0.002582625253126025, + 0.007173219695687294, + 0.005810797680169344, + 0.00526023143902421, + 0.004684931132942438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.02978871762752533 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot of an elderly farmer inspecting crops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer wearing a faded blue denim shirt with sleeves rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the farmer extend his right hand to gently glide his fingertips across the corn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a field of tall, lush corn at dusk with a sky transitioning from orange to purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while the farmer moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004634472541511059 + ], + "opposite_logprob_tokens": [ + -5.379634380340576 + ], + "logprob_reward": -0.004634472541511059, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004634472541511059, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06922408938407898, + 0.4934018850326538, + 0.02996382862329483, + 0.04664299264550209, + 0.036457475274801254, + 0.040582843124866486, + 0.1290775090456009, + 0.08452912420034409, + 0.043697476387023926, + 0.026422753930091858 + ], + "credit_map": [ + 0.0072051286697387695, + 0.0045248037204146385, + 0.011796388775110245, + 0.007705901749432087, + 0.009695149958133698, + 0.009217804297804832, + 0.00831076130270958, + 0.010768154636025429, + 0.016479341313242912, + 0.03222004696726799, + 0.03672895208001137, + 0.014508373104035854, + 0.14687320590019226, + 0.07767906785011292, + 0.11463388800621033, + 0.05427904427051544, + 0.0027123023755848408, + 0.002336772857233882, + 0.0027241245843470097, + 0.005718465894460678, + 0.004880271852016449, + 0.0027671135030686855, + 0.003790932707488537, + 0.005033849738538265, + 0.004418957978487015, + 0.004513503983616829, + 0.004779953043907881, + 0.0021443215664476156, + 0.014982281252741814, + 0.008774327114224434, + 0.003552605863660574, + 0.0034770481288433075, + 0.0044669173657894135, + 0.0037121635396033525, + 0.004439144860953093, + 0.004188681486994028, + 0.004313441459089518, + 0.00639719981700182, + 0.0041468385607004166, + 0.004793092142790556, + 0.004287041258066893, + 0.005815023556351662, + 0.0070788017474114895, + 0.002758622169494629, + 0.006998690310865641, + 0.0050368839874863625, + 0.003075453219935298, + 0.005532323382794857, + 0.010162478312849998, + 0.028370674699544907, + 0.01318704430013895, + 0.010925551876425743, + 0.021213797852396965, + 0.017934177070856094, + 0.01281088124960661, + 0.014472921378910542, + 0.0074256607331335545, + 0.012678304687142372, + 0.009685215540230274, + 0.006318981293588877, + 0.007201039232313633, + 0.007252196781337261, + 0.011353462934494019, + 0.02261427417397499, + 0.0024694623425602913, + 0.005560120567679405, + 0.003792649367824197, + 0.006650606170296669, + 0.006508377846330404, + 0.008979009464383125, + 0.005296786781400442, + 0.004440467804670334, + 0.0029723746702075005, + 0.0032209649216383696, + 0.004742972087115049, + 0.0020304883364588022, + 0.005734067875891924, + 0.003009796142578125, + 0.00173455779440701, + 0.0029775346629321575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 7 + ], + "selected_reward": -0.004634472541511059 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot of an elderly farmer inspecting crops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the farmer wearing a faded blue denim shirt with sleeves rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the farmer extend his right hand to gently glide his fingertips across the corn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a field of tall, lush corn at dusk with a sky transitioning from orange to purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while the farmer moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486243486404419 + ], + "opposite_logprob_tokens": [ + -1.223624348640442 + ], + "logprob_reward": -0.3486243486404419, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486243486404419, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03997192531824112, + 0.09989163279533386, + 0.03779497742652893, + 0.14769238233566284, + 0.19078052043914795, + 0.0718114823102951, + 0.08439066261053085, + 0.0924995020031929, + 0.1778774857521057, + 0.05728943273425102 + ], + "credit_map": [ + 0.0036841914989054203, + 0.004350177012383938, + 0.0053258780390024185, + 0.004674267489463091, + 0.005986685864627361, + 0.004094534087926149, + 0.006927385926246643, + 0.004928804934024811, + 0.004605807829648256, + 0.006896700710058212, + 0.01147503312677145, + 0.007502296939492226, + 0.0272300336509943, + 0.01445007137954235, + 0.017519794404506683, + 0.010211886838078499, + 0.0042769829742610455, + 0.0034514563158154488, + 0.004085863940417767, + 0.007832220755517483, + 0.004853362683206797, + 0.002714223228394985, + 0.005256217904388905, + 0.005324651021510363, + 0.008655423298478127, + 0.014341614209115505, + 0.024225713685154915, + 0.007698498200625181, + 0.04870353639125824, + 0.025194616988301277, + 0.011193649843335152, + 0.007679321337491274, + 0.025473400950431824, + 0.021050797775387764, + 0.025874145328998566, + 0.028780704364180565, + 0.022694213315844536, + 0.01612972840666771, + 0.02025711163878441, + 0.03052041307091713, + 0.007247453555464745, + 0.009271466173231602, + 0.006314198020845652, + 0.005387127865105867, + 0.017934419214725494, + 0.010560751892626286, + 0.004638512618839741, + 0.010457552969455719, + 0.008440781384706497, + 0.010300176218152046, + 0.008732949383556843, + 0.008576680906116962, + 0.012617966160178185, + 0.008743228390812874, + 0.016632238402962685, + 0.010346638970077038, + 0.009069982916116714, + 0.010482368990778923, + 0.009974244982004166, + 0.008156269788742065, + 0.013214667327702045, + 0.009356594644486904, + 0.015184723772108555, + 0.017060644924640656, + 0.008812486194074154, + 0.019406568259000778, + 0.016684215515851974, + 0.03160014748573303, + 0.01594414934515953, + 0.021308692172169685, + 0.026137391105294228, + 0.037983834743499756, + 0.005898201372474432, + 0.007275014650076628, + 0.007529998663812876, + 0.003909956198185682, + 0.01026652567088604, + 0.00964537262916565, + 0.005269966553896666, + 0.007494398392736912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 3 + ], + "selected_reward": -0.3486243486404419 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_4/index_tensor([7862, 7862])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_4/index_tensor([7862, 7862])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b20f2746b22955fa5e6fb260d98054bef1239728 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_4/index_tensor([7862, 7862])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-angle close-up captures the brutal force of a sleek, black, metallic sports car as it violently collides with a cold, galvanized steel guardrail. The car's front fenders instantly collapse inward, sending a shower of brilliant orange and white sparks blasting into the pitch-black night, completely illuminating the twisted metal for a few seconds. The vehicle chaotically bounces off the obstacle, tires screeching on the wet asphalt, the chassis shaking, and the red glow of its taillights smearing across the frame. Immediately after the impact, the camera executes a swift and dramatic pullback, quickly zooming out to reveal the full, terrifying context: the stricken car, now small and isolated, its headlights casting a sharp, lonely beam against the darkness, trapped on the sharp curve of a winding mountain road that twists precariously around a sheer cliff face and disappears into the oppressive darkness of the surrounding wilderness beneath a starless sky, creating a tense, perilous atmosphere of sudden disaster and profound isolation.", + "seed": 21789, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_4/index_7862_seed_21789_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.20965617429465055, + "vlm_logprob": -0.20965617429465055, + "vlm_yes_no_margin": 2.45, + "vlm_token_credit": -0.20965617429465055, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, black, metallic sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car violently collide with a cold, galvanized steel guardrail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant orange and white sparks visible blasting from the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift pullback and zoom out immediately after the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final scene set on a winding mountain road with a sheer cliff face under a starless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05899003520607948 + ], + "opposite_logprob_tokens": [ + -2.933990001678467 + ], + "logprob_reward": -0.05899003520607948, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05899003520607948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15480802953243256, + 0.07074899226427078, + 0.05615871772170067, + 0.22191160917282104, + 0.19123023748397827, + 0.08552566915750504, + 0.11253450810909271, + 0.043889712542295456, + 0.026205549016594887, + 0.0369870625436306 + ], + "credit_map": [ + 0.012483416125178337, + 0.016209101304411888, + 0.015949996188282967, + 0.01510713342577219, + 0.024606242775917053, + 0.02227006107568741, + 0.022078275680541992, + 0.026103783398866653, + 0.008925230242311954, + 0.005506841465830803, + 0.00856844149529934, + 0.0082543371245265, + 0.012372997589409351, + 0.007023991551250219, + 0.008929690346121788, + 0.011167459189891815, + 0.0053659467957913876, + 0.00856843963265419, + 0.0043469881638884544, + 0.006255977787077427, + 0.006784989032894373, + 0.005111083388328552, + 0.012142220512032509, + 0.007583071943372488, + 0.01723729819059372, + 0.026781747117638588, + 0.01931159384548664, + 0.03780512511730194, + 0.02223789133131504, + 0.02926413156092167, + 0.028790254145860672, + 0.040483538061380386, + 0.01627223752439022, + 0.04282532259821892, + 0.010746113024652004, + 0.04166976362466812, + 0.01719617284834385, + 0.030776331201195717, + 0.012803088873624802, + 0.018941190093755722, + 0.007887092418968678, + 0.011846777983009815, + 0.006431786809116602, + 0.013797367922961712, + 0.011996853165328503, + 0.013948183506727219, + 0.008249500766396523, + 0.011368093080818653, + 0.01209103874862194, + 0.01188488956540823, + 0.010093911550939083, + 0.008653111755847931, + 0.019352899864315987, + 0.0161888487637043, + 0.008546417579054832, + 0.025723382830619812, + 0.003903389908373356, + 0.005421498324722052, + 0.008544370532035828, + 0.003339757677167654, + 0.004652880132198334, + 0.006394551135599613, + 0.0040756091475486755, + 0.007557650562375784, + 0.0026688978541642427, + 0.002921122359111905, + 0.0020715738646686077, + 0.0020267884247004986, + 0.004897308070212603, + 0.0049082571640610695, + 0.002081961138173938, + 0.004629637114703655, + 0.003440691391006112, + 0.003798339981585741, + 0.0034319967962801456, + 0.003352109808474779, + 0.004596085287630558, + 0.0051925512962043285, + 0.0035770474933087826, + 0.009598237462341785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.05899003520607948 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, black, metallic sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car violently collide with a cold, galvanized steel guardrail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant orange and white sparks visible blasting from the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift pullback and zoom out immediately after the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final scene set on a winding mountain road with a sheer cliff face under a starless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17986446619033813 + ], + "opposite_logprob_tokens": [ + -1.8048644065856934 + ], + "logprob_reward": -0.17986446619033813, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17986446619033813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14681628346443176, + 0.08317874372005463, + 0.08758382499217987, + 0.09092246741056442, + 0.11236429214477539, + 0.04233270883560181, + 0.08283302187919617, + 0.2543206810951233, + 0.05405113846063614, + 0.04559687525033951 + ], + "credit_map": [ + 0.012824793346226215, + 0.012175297364592552, + 0.012340563349425793, + 0.017475182190537453, + 0.020020555704832077, + 0.020989855751395226, + 0.027677034959197044, + 0.02331298403441906, + 0.01332687959074974, + 0.005443974398076534, + 0.005723309703171253, + 0.009824411943554878, + 0.013883499428629875, + 0.015024572610855103, + 0.010555206798017025, + 0.009396885521709919, + 0.008048731833696365, + 0.007936406880617142, + 0.006351237650960684, + 0.009170470759272575, + 0.011390481144189835, + 0.008977938443422318, + 0.02031050994992256, + 0.015398039482533932, + 0.005614241119474173, + 0.012827524915337563, + 0.008500207215547562, + 0.013261371292173862, + 0.011107963509857655, + 0.012987316586077213, + 0.013055253773927689, + 0.013568575493991375, + 0.00911105889827013, + 0.02597462758421898, + 0.005970984231680632, + 0.024338373914361, + 0.010035194456577301, + 0.01886129379272461, + 0.007043442223221064, + 0.011029304936528206, + 0.004569420125335455, + 0.006060224026441574, + 0.0028591961599886417, + 0.005748560186475515, + 0.00600295327603817, + 0.007370532490313053, + 0.004044156055897474, + 0.005677663721144199, + 0.009281881153583527, + 0.008271648548543453, + 0.01046262588351965, + 0.005973737221211195, + 0.010882150381803513, + 0.012354811653494835, + 0.006572258193045855, + 0.019033905118703842, + 0.025257963687181473, + 0.02780999056994915, + 0.04800596460700035, + 0.013684614561498165, + 0.02338455058634281, + 0.03768068552017212, + 0.014012426137924194, + 0.06448447704315186, + 0.005693824961781502, + 0.005921356379985809, + 0.0064364331774413586, + 0.005540845450013876, + 0.009703204967081547, + 0.008438724093139172, + 0.005626893136650324, + 0.0066898479126393795, + 0.0036646174266934395, + 0.007354688365012407, + 0.003185681300237775, + 0.003520284779369831, + 0.010006255470216274, + 0.007198751904070377, + 0.003169354749843478, + 0.007497239857912064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.17986446619033813 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, black, metallic sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car violently collide with a cold, galvanized steel guardrail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant orange and white sparks visible blasting from the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift pullback and zoom out immediately after the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final scene set on a winding mountain road with a sheer cliff face under a starless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636622689664364 + ], + "opposite_logprob_tokens": [ + -4.758636474609375 + ], + "logprob_reward": -0.008636622689664364, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636622689664364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09591533243656158, + 0.06827832758426666, + 0.031351737678050995, + 0.12878170609474182, + 0.06726973503828049, + 0.059975557029247284, + 0.11476434022188187, + 0.23666784167289734, + 0.1318145990371704, + 0.06518077850341797 + ], + "credit_map": [ + 0.013021847233176231, + 0.012332211248576641, + 0.008934498764574528, + 0.012989141047000885, + 0.012840214185416698, + 0.01237164530903101, + 0.0100016538053751, + 0.013424123637378216, + 0.011128210462629795, + 0.004765168763697147, + 0.006376315373927355, + 0.005307134240865707, + 0.0120630394667387, + 0.009102781303226948, + 0.007762005086988211, + 0.011773682199418545, + 0.003087046090513468, + 0.003664956893771887, + 0.0024647030513733625, + 0.0029940782114863396, + 0.0039988006465137005, + 0.003557893680408597, + 0.004908185917884111, + 0.006676069926470518, + 0.011334963142871857, + 0.015994906425476074, + 0.011463573202490807, + 0.024094702675938606, + 0.011193601414561272, + 0.016575679183006287, + 0.014987711794674397, + 0.023136574774980545, + 0.007320133503526449, + 0.015280229970812798, + 0.006352093070745468, + 0.008481690660119057, + 0.007398948539048433, + 0.010434223338961601, + 0.005659972783178091, + 0.006342442240566015, + 0.007723913993686438, + 0.008978012949228287, + 0.006876361556351185, + 0.006597355008125305, + 0.006256271619349718, + 0.00902547501027584, + 0.0057455203495919704, + 0.008772645145654678, + 0.011157440021634102, + 0.01610260270535946, + 0.009829654358327389, + 0.004902172368019819, + 0.027569705620408058, + 0.0236403439193964, + 0.009992367587983608, + 0.011570053175091743, + 0.024378234520554543, + 0.0440419539809227, + 0.02967783249914646, + 0.01771826297044754, + 0.028435876592993736, + 0.028738530352711678, + 0.02242639847099781, + 0.041250742971897125, + 0.01272903848439455, + 0.013158813118934631, + 0.021210262551903725, + 0.014740463346242905, + 0.01938539743423462, + 0.01985958032310009, + 0.012648502364754677, + 0.01808254048228264, + 0.006232588551938534, + 0.008629278279840946, + 0.005246242042630911, + 0.0089883329346776, + 0.010455705225467682, + 0.010003638453781605, + 0.005462265107780695, + 0.010162725113332272 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 3 + ], + "selected_reward": -0.008636622689664364 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, black, metallic sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car violently collide with a cold, galvanized steel guardrail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant orange and white sparks visible blasting from the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift pullback and zoom out immediately after the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final scene set on a winding mountain road with a sheer cliff face under a starless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304227605462074 + ], + "opposite_logprob_tokens": [ + -3.1680421829223633 + ], + "logprob_reward": -0.04304227605462074, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304227605462074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05582108348608017, + 0.06977406144142151, + 0.021306481212377548, + 0.020431824028491974, + 0.019321728497743607, + 0.012765178456902504, + 0.1831875741481781, + 0.5887354612350464, + 0.023386497050523758, + 0.0052701206877827644 + ], + "credit_map": [ + 0.00447268458083272, + 0.004697793163359165, + 0.0042471871711313725, + 0.008816058747470379, + 0.00815783441066742, + 0.007484504021704197, + 0.010992191731929779, + 0.006952834315598011, + 0.012219903990626335, + 0.005524574313312769, + 0.007151543628424406, + 0.007782666943967342, + 0.009380914270877838, + 0.009733778424561024, + 0.008102591149508953, + 0.009878088720142841, + 0.002124055754393339, + 0.003961655311286449, + 0.001967672724276781, + 0.0022617678623646498, + 0.002683238824829459, + 0.0024453303776681423, + 0.0034886219073086977, + 0.002374138217419386, + 0.001909410348162055, + 0.0030571413226425648, + 0.0017203678144142032, + 0.002556856721639633, + 0.0036327599082142115, + 0.002549748634919524, + 0.0022437460720539093, + 0.002761791693046689, + 0.0020359628833830357, + 0.0031946133822202682, + 0.0019257114036008716, + 0.0028847369831055403, + 0.002361222868785262, + 0.0034382783342152834, + 0.001284369151107967, + 0.002196833724156022, + 0.0016222810372710228, + 0.0020225276239216328, + 0.0010187876177951694, + 0.001447040238417685, + 0.0015573729760944843, + 0.0021972847171127796, + 0.0009663128876127303, + 0.0019335715332999825, + 0.0201112087816, + 0.019797924906015396, + 0.014492826536297798, + 0.00926622562110424, + 0.03133883699774742, + 0.029474785551428795, + 0.014240752905607224, + 0.04446501284837723, + 0.06339488178491592, + 0.0461118258535862, + 0.11039932072162628, + 0.03519080579280853, + 0.04886660352349281, + 0.08454412966966629, + 0.0434955433011055, + 0.15673235058784485, + 0.002141183940693736, + 0.002714704954996705, + 0.002629816997796297, + 0.001689159544184804, + 0.005240375641733408, + 0.004311577416956425, + 0.0022914367727935314, + 0.0023682431783527136, + 0.0005155495018698275, + 0.0008437555516138673, + 0.00046367256436496973, + 0.0004630059702321887, + 0.0008492807392030954, + 0.000745259050745517, + 0.00044378600432537496, + 0.000945811509154737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 1 + ], + "selected_reward": -0.04304227605462074 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, black, metallic sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car violently collide with a cold, galvanized steel guardrail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant orange and white sparks visible blasting from the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift pullback and zoom out immediately after the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final scene set on a winding mountain road with a sheer cliff face under a starless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.75774747133255 + ], + "opposite_logprob_tokens": [ + -0.63274747133255 + ], + "logprob_reward": -0.75774747133255, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.75774747133255, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.23620927333831787, + 0.17305988073349, + 0.1044192910194397, + 0.05139864236116409, + 0.10253652930259705, + 0.028246544301509857, + 0.10990721732378006, + 0.10951324552297592, + 0.056042999029159546, + 0.02866639755666256 + ], + "credit_map": [ + 0.022884946316480637, + 0.026321088895201683, + 0.020948806777596474, + 0.022715291008353233, + 0.03572637960314751, + 0.03259726241230965, + 0.040823984891176224, + 0.034191496670246124, + 0.03191179409623146, + 0.014600485563278198, + 0.02436862513422966, + 0.026057086884975433, + 0.023877687752246857, + 0.018771523609757423, + 0.01607067510485649, + 0.01740199886262417, + 0.014748763293027878, + 0.01664901152253151, + 0.008243250660598278, + 0.010778850875794888, + 0.014605935662984848, + 0.009718658402562141, + 0.01949348673224449, + 0.010181338526308537, + 0.00448073074221611, + 0.006551692262291908, + 0.004852256737649441, + 0.006302850786596537, + 0.007283526938408613, + 0.007294828072190285, + 0.008712098002433777, + 0.005920658353716135, + 0.008930562064051628, + 0.026119329035282135, + 0.008438404649496078, + 0.021111801266670227, + 0.007235079072415829, + 0.016524692997336388, + 0.0056792814284563065, + 0.008497381582856178, + 0.00301355984993279, + 0.0035616063978523016, + 0.0035572429187595844, + 0.0034966650418937206, + 0.0038626408204436302, + 0.0041746217757463455, + 0.002851792611181736, + 0.003728414187207818, + 0.010225360281765461, + 0.013159665279090405, + 0.010406293906271458, + 0.005906380712985992, + 0.02955632284283638, + 0.01648901216685772, + 0.009555254131555557, + 0.014608924277126789, + 0.01041905302554369, + 0.01210023183375597, + 0.020029008388519287, + 0.007074438035488129, + 0.011385118588805199, + 0.01679714024066925, + 0.005942082032561302, + 0.025766171514987946, + 0.004890040028840303, + 0.005262251012027264, + 0.0057729389518499374, + 0.0041539533995091915, + 0.010099908336997032, + 0.009249510243535042, + 0.005515589378774166, + 0.011098806746304035, + 0.003657869528979063, + 0.003838845994323492, + 0.0024601053446531296, + 0.0025155122857540846, + 0.005856936797499657, + 0.004370507318526506, + 0.0019453888526186347, + 0.004021232016384602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.75774747133255 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "dynamic, low-angle close-up captures the brutal force of a sleek, black, metallic sports car as it violently collides with a cold, galvanized steel guardrail. The car's front fenders instantly collapse inward, sending a shower of brilliant orange and white sparks blasting into the pitch-black night, completely illuminating the twisted metal for a few seconds. The vehicle chaotically bounces off the obstacle, tires screeching on the wet asphalt, the chassis shaking, and the red glow of its taillights smearing across the frame. Immediately after the impact, the camera executes a swift and dramatic pullback, quickly zooming out to reveal the full, terrifying context: the stricken car, now small and isolated, its headlights casting a sharp, lonely beam against the darkness, trapped on the sharp curve of a winding mountain road that twists precariously around a sheer cliff face and disappears into the oppressive darkness of the surrounding wilderness beneath a starless sky, creating a tense, perilous atmosphere of sudden disaster and profound isolation.", + "seed": 21790, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_4/index_7862_seed_21790_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1619884066749364, + "vlm_logprob": -0.1619884066749364, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.1619884066749364, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, black, metallic sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car violently collide with a cold, galvanized steel guardrail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant orange and white sparks visible blasting from the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift pullback and zoom out immediately after the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final scene set on a winding mountain road with a sheer cliff face under a starless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08170394599437714 + ], + "opposite_logprob_tokens": [ + -2.5817039012908936 + ], + "logprob_reward": -0.08170394599437714, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08170394599437714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23862870037555695, + 0.01948450691998005, + 0.03666522726416588, + 0.09974759817123413, + 0.3770673871040344, + 0.04206537827849388, + 0.038693901151418686, + 0.03620689734816551, + 0.05683751404285431, + 0.05460285395383835 + ], + "credit_map": [ + 0.024551749229431152, + 0.02868156135082245, + 0.0234463419765234, + 0.027197908610105515, + 0.032322075217962265, + 0.03276573121547699, + 0.03815058246254921, + 0.03151276335120201, + 0.0025112999137490988, + 0.002846792107447982, + 0.0017843992682173848, + 0.0029707588255405426, + 0.0027979156002402306, + 0.002235430059954524, + 0.001568642444908619, + 0.0027692709118127823, + 0.005642639007419348, + 0.0038884622044861317, + 0.002699774457141757, + 0.0020382090006023645, + 0.005384758580476046, + 0.006272708997130394, + 0.0058615501038730145, + 0.00487712724134326, + 0.007725037634372711, + 0.015643268823623657, + 0.008138354867696762, + 0.008279042318463326, + 0.02114233188331127, + 0.025851435959339142, + 0.00679338863119483, + 0.0061747427098453045, + 0.007735127583146095, + 0.01593492180109024, + 0.004869654309004545, + 0.006408631335943937, + 0.16971160471439362, + 0.16374756395816803, + 0.003800818929448724, + 0.004859057255089283, + 0.0031920517794787884, + 0.0067648692056536674, + 0.0040359944105148315, + 0.003634557593613863, + 0.00990451592952013, + 0.006172731053084135, + 0.005972238723188639, + 0.0023884261026978493, + 0.0030738108325749636, + 0.006949921138584614, + 0.0035340331960469484, + 0.002704292070120573, + 0.00898615550249815, + 0.006435105577111244, + 0.005102524068206549, + 0.0019080621423199773, + 0.0030601078178733587, + 0.005700691137462854, + 0.003544620471075177, + 0.0027585639618337154, + 0.008539468050003052, + 0.005347891245037317, + 0.0036786531563848257, + 0.0035769047681242228, + 0.004731951747089624, + 0.006877807900309563, + 0.00709425238892436, + 0.005969665013253689, + 0.009052532725036144, + 0.008169794455170631, + 0.009214336983859539, + 0.0057271746918559074, + 0.0051831346936523914, + 0.009672396816313267, + 0.005771520081907511, + 0.003598191309720278, + 0.010310549288988113, + 0.009574885480105877, + 0.006913207471370697, + 0.003578974399715662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.08170394599437714 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, black, metallic sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car violently collide with a cold, galvanized steel guardrail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant orange and white sparks visible blasting from the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift pullback and zoom out immediately after the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final scene set on a winding mountain road with a sheer cliff face under a starless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000703364610672 + ], + "opposite_logprob_tokens": [ + -2.695007085800171 + ], + "logprob_reward": -0.07000703364610672, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000703364610672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19079498946666718, + 0.07028216868638992, + 0.10944749414920807, + 0.11087608337402344, + 0.22779430449008942, + 0.0734131708741188, + 0.03217420354485512, + 0.055999454110860825, + 0.08087477087974548, + 0.04834333434700966 + ], + "credit_map": [ + 0.019340872764587402, + 0.025566238909959793, + 0.018768813461065292, + 0.02822103351354599, + 0.029194749891757965, + 0.018404746428132057, + 0.029899070039391518, + 0.021399466320872307, + 0.0068885856308043, + 0.010897943750023842, + 0.008757863193750381, + 0.010387348011136055, + 0.0076413387432694435, + 0.006231095176190138, + 0.009716924279928207, + 0.009761069901287556, + 0.008499937132000923, + 0.006259843707084656, + 0.0031607637647539377, + 0.003704190719872713, + 0.038241490721702576, + 0.03753231465816498, + 0.004772889893501997, + 0.0072760614566504955, + 0.009727523662149906, + 0.023092081770300865, + 0.00927136279642582, + 0.008427876979112625, + 0.03131911903619766, + 0.017206238582730293, + 0.006268166471272707, + 0.0055637178011238575, + 0.006025309674441814, + 0.014797277748584747, + 0.004728467203676701, + 0.005350098479539156, + 0.0899687260389328, + 0.09155454486608505, + 0.005084851756691933, + 0.010285028256475925, + 0.005995502695441246, + 0.01493525505065918, + 0.007796564605087042, + 0.0058195036835968494, + 0.014677544124424458, + 0.012016518041491508, + 0.007360583636909723, + 0.004811700899153948, + 0.003099235240370035, + 0.004937987774610519, + 0.003292258596047759, + 0.0022709739860147238, + 0.008117114193737507, + 0.005441256798803806, + 0.0030344806145876646, + 0.0019808984361588955, + 0.004602741450071335, + 0.008734949864447117, + 0.004921798128634691, + 0.005073958076536655, + 0.01073242723941803, + 0.008680282160639763, + 0.008546180091798306, + 0.0047071208246052265, + 0.006658171769231558, + 0.011330830864608288, + 0.008266020566225052, + 0.007413025479763746, + 0.014017436653375626, + 0.012378406710922718, + 0.014001174829900265, + 0.006809706799685955, + 0.004346003290265799, + 0.008326719515025616, + 0.00509287603199482, + 0.0029431532602757215, + 0.009484991431236267, + 0.008477253839373589, + 0.005923333577811718, + 0.003749007126316428 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.07000703364610672 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, black, metallic sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car violently collide with a cold, galvanized steel guardrail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant orange and white sparks visible blasting from the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift pullback and zoom out immediately after the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final scene set on a winding mountain road with a sheer cliff face under a starless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002198899397626519 + ], + "opposite_logprob_tokens": [ + -6.127198696136475 + ], + "logprob_reward": -0.002198899397626519, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002198899397626519, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16413235664367676, + 0.08252903074026108, + 0.10063862055540085, + 0.05097284913063049, + 0.3557738661766052, + 0.047557830810546875, + 0.0528569370508194, + 0.07258810102939606, + 0.04004418104887009, + 0.032906245440244675 + ], + "credit_map": [ + 0.015595030970871449, + 0.020000232383608818, + 0.019050322473049164, + 0.028012825176119804, + 0.018997611477971077, + 0.016033943742513657, + 0.026153264567255974, + 0.020289119333028793, + 0.013515332713723183, + 0.0124294338747859, + 0.009731350466609001, + 0.011709736660122871, + 0.008890924043953419, + 0.010235338471829891, + 0.0057120430283248425, + 0.010304867289960384, + 0.008791672065854073, + 0.009517020545899868, + 0.004745131824165583, + 0.007005021907389164, + 0.02528301626443863, + 0.02976546622812748, + 0.008792061358690262, + 0.006739235948771238, + 0.004008542280644178, + 0.006916698534041643, + 0.003060947870835662, + 0.0034150835126638412, + 0.011983494274318218, + 0.015677638351917267, + 0.003273568581789732, + 0.0026368757244199514, + 0.007923480123281479, + 0.016185037791728973, + 0.006448667962104082, + 0.006857333239167929, + 0.15738129615783691, + 0.1522035300731659, + 0.0049184407107532024, + 0.0038561236578971148, + 0.004168648272752762, + 0.007123919203877449, + 0.006823247764259577, + 0.004052026197314262, + 0.004999521654099226, + 0.005106932949274778, + 0.007082717958837748, + 0.008200815878808498, + 0.005316845141351223, + 0.008706466294825077, + 0.006546600256115198, + 0.0039344714023172855, + 0.009608829393982887, + 0.008201663382351398, + 0.007185596507042646, + 0.003356464207172394, + 0.005755052901804447, + 0.010691249743103981, + 0.008558240719139576, + 0.005903833545744419, + 0.015097330324351788, + 0.010210935957729816, + 0.008416957221925259, + 0.00795450247824192, + 0.0030111379455775023, + 0.004752144683152437, + 0.003762359032407403, + 0.003529891138896346, + 0.009506963193416595, + 0.0063589951023459435, + 0.005766452755779028, + 0.0033562344033271074, + 0.0026950454339385033, + 0.004062366206198931, + 0.0030980005394667387, + 0.003354005515575409, + 0.005139559041708708, + 0.005340686533600092, + 0.006626127287745476, + 0.002590454649180174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.002198899397626519 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, black, metallic sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car violently collide with a cold, galvanized steel guardrail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant orange and white sparks visible blasting from the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift pullback and zoom out immediately after the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final scene set on a winding mountain road with a sheer cliff face under a starless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023277709260582924 + ], + "opposite_logprob_tokens": [ + -3.773277759552002 + ], + "logprob_reward": -0.023277709260582924, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023277709260582924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10032352060079575, + 0.03470923751592636, + 0.052269723266363144, + 0.23588921129703522, + 0.2585601508617401, + 0.08202297985553741, + 0.1081821620464325, + 0.08123915642499924, + 0.024290936067700386, + 0.022512899711728096 + ], + "credit_map": [ + 0.010906713083386421, + 0.01330499816685915, + 0.008802817203104496, + 0.011877082288265228, + 0.013260702602565289, + 0.01210413035005331, + 0.017909429967403412, + 0.01215764507651329, + 0.0057421112433075905, + 0.005624131299555302, + 0.0023541792761534452, + 0.003933818079531193, + 0.0054403808899223804, + 0.003913120366632938, + 0.003956294618546963, + 0.0037452036049216986, + 0.008177503012120724, + 0.006277767941355705, + 0.003586735110729933, + 0.0036356421187520027, + 0.0068627274595201015, + 0.007926823571324348, + 0.007205655332654715, + 0.008596869185566902, + 0.013437537476420403, + 0.04200084134936333, + 0.009035359136760235, + 0.012485023587942123, + 0.07792921364307404, + 0.05984445661306381, + 0.010291795246303082, + 0.010864987038075924, + 0.007549964357167482, + 0.013583487831056118, + 0.0061152721755206585, + 0.005968135315924883, + 0.10728757828474045, + 0.10018916428089142, + 0.00462499912828207, + 0.013241548091173172, + 0.005179091356694698, + 0.01848815381526947, + 0.007341830991208553, + 0.003816753625869751, + 0.02074834145605564, + 0.01411353424191475, + 0.00649544270709157, + 0.005839826073497534, + 0.008191687054932117, + 0.016324013471603394, + 0.010533819906413555, + 0.00836444366723299, + 0.028934726491570473, + 0.017157768830657005, + 0.012790579348802567, + 0.005885132122784853, + 0.007722239010035992, + 0.015612161718308926, + 0.008634954690933228, + 0.005993377882987261, + 0.020503651350736618, + 0.011851872317492962, + 0.007013353984802961, + 0.003907544072717428, + 0.002329900860786438, + 0.0037205759435892105, + 0.0028436847496777773, + 0.0026135467924177647, + 0.0035004648379981518, + 0.002732655731961131, + 0.003660305868834257, + 0.00288979965262115, + 0.001981581561267376, + 0.003006531158462167, + 0.003575342707335949, + 0.002233447041362524, + 0.0033778278157114983, + 0.00301514845341444, + 0.0032412437722086906, + 0.002081778133288026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 6 + ], + "selected_reward": -0.023277709260582924 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, black, metallic sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car violently collide with a cold, galvanized steel guardrail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant orange and white sparks visible blasting from the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift pullback and zoom out immediately after the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final scene set on a winding mountain road with a sheer cliff face under a starless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6327544450759888 + ], + "opposite_logprob_tokens": [ + -0.7577544450759888 + ], + "logprob_reward": -0.6327544450759888, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6327544450759888, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0829102024435997, + 0.0412624254822731, + 0.06973055750131607, + 0.1150810569524765, + 0.16220952570438385, + 0.061713673174381256, + 0.24850869178771973, + 0.07269252836704254, + 0.0577898733317852, + 0.08810151368379593 + ], + "credit_map": [ + 0.008498689159750938, + 0.013940372504293919, + 0.007703952956944704, + 0.010875726118683815, + 0.01266312226653099, + 0.010551158338785172, + 0.01013003196567297, + 0.008547145873308182, + 0.005011256784200668, + 0.006541604176163673, + 0.0033178667072206736, + 0.004215127322822809, + 0.005629361141473055, + 0.006237047724425793, + 0.00498085655272007, + 0.00532930251210928, + 0.008360547013580799, + 0.007779877632856369, + 0.0040618604980409145, + 0.0036890802439302206, + 0.01456792838871479, + 0.012906689196825027, + 0.00807939562946558, + 0.010285167954862118, + 0.007734007202088833, + 0.015571988187730312, + 0.004336996469646692, + 0.0038716646376997232, + 0.036899786442518234, + 0.036481257528066635, + 0.005611666478216648, + 0.004573681391775608, + 0.004320322070270777, + 0.010643631219863892, + 0.005002943798899651, + 0.0061438786797225475, + 0.06316220015287399, + 0.0645759105682373, + 0.0031997717451304197, + 0.00516084861010313, + 0.005223932676017284, + 0.008787252008914948, + 0.0075867329724133015, + 0.004119625315070152, + 0.00812025647610426, + 0.00824799295514822, + 0.010183624923229218, + 0.009444251656532288, + 0.0187673419713974, + 0.039604488760232925, + 0.027027437463402748, + 0.021599140018224716, + 0.06360460072755814, + 0.03718887269496918, + 0.027722422033548355, + 0.012994378805160522, + 0.005696992855519056, + 0.011138005182147026, + 0.007806319277733564, + 0.005719443783164024, + 0.017189232632517815, + 0.010719538666307926, + 0.008966733701527119, + 0.005456259939819574, + 0.00866912305355072, + 0.011455083265900612, + 0.00787486881017685, + 0.0033191253896802664, + 0.008261140435934067, + 0.008386529050767422, + 0.005529520567506552, + 0.004294475074857473, + 0.008756476454436779, + 0.016926737502217293, + 0.012832189910113811, + 0.006103954743593931, + 0.01557877380400896, + 0.013636438176035881, + 0.006461051758378744, + 0.007805883418768644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 3 + ], + "selected_reward": -0.6327544450759888 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle close-up captures the brutal force of a sleek, black, metallic sports car as it violently collides with a cold, galvanized steel guardrail. The car's front fenders instantly collapse inward, sending a shower of brilliant orange and white sparks blasting into the pitch-black night, completely illuminating the twisted metal for a few seconds. The vehicle chaotically bounces off the obstacle, tires screeching on the wet asphalt, the chassis shaking, and the red glow of its taillights smearing across the frame. Immediately after the impact, the camera executes a swift and dramatic pullback, quickly zooming out to reveal the full, terrifying context: the stricken car, now small and isolated, its headlights casting a sharp, lonely beam against the darkness, trapped on the sharp curve of a winding mountain road that twists precariously around a sheer cliff face and disappears into the oppressive darkness of the surrounding wilderness beneath a starless sky, creating a tense, perilous atmosphere of sudden disaster and profound isolation.", + "seed": 21791, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_4/index_7862_seed_21791_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.22530359216034412, + "vlm_logprob": -0.22530359216034412, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.22530359216034412, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, black, metallic sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car violently collide with a cold, galvanized steel guardrail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant orange and white sparks visible blasting from the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift pullback and zoom out immediately after the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final scene set on a winding mountain road with a sheer cliff face under a starless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017008861526846886 + ], + "opposite_logprob_tokens": [ + -4.2670087814331055 + ], + "logprob_reward": -0.017008861526846886, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.017008861526846886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7876945734024048, + 0.012726465240120888, + 0.004948255605995655, + 0.01534848753362894, + 0.14192092418670654, + 0.00810540933161974, + 0.0026040158700197935, + 0.005789492279291153, + 0.012497765012085438, + 0.008364628069102764 + ], + "credit_map": [ + 0.0609271377325058, + 0.31932446360588074, + 0.2867966592311859, + 0.02957763895392418, + 0.023153230547904968, + 0.016477249562740326, + 0.02714649587869644, + 0.024291696026921272, + 0.0012064209440723062, + 0.0011921009281650186, + 0.0016689990879967809, + 0.0020646858029067516, + 0.001349866040982306, + 0.0017148647457361221, + 0.0018258068012073636, + 0.0017037206562235951, + 0.000667333195451647, + 0.0008181158918887377, + 0.0005718389875255525, + 0.0005067523452453315, + 0.0008153097005560994, + 0.0007397562148980796, + 0.0004024303052574396, + 0.0004267193144187331, + 0.0014291128609329462, + 0.0016768744681030512, + 0.0016245348379015923, + 0.0017924532294273376, + 0.0032860857900232077, + 0.002430400112643838, + 0.0020848687272518873, + 0.0010241573909297585, + 0.006255467887967825, + 0.04976636543869972, + 0.0034432513639330864, + 0.004206458106637001, + 0.011314552277326584, + 0.05915867164731026, + 0.004571340046823025, + 0.0032048169523477554, + 0.0006248524878174067, + 0.001085957046598196, + 0.000618020654655993, + 0.001136753591708839, + 0.0010042821522802114, + 0.0011970947962254286, + 0.0007778216386213899, + 0.001660626963712275, + 0.00022092564904596657, + 0.00042105314787477255, + 0.00022830879606772214, + 0.00018740991072263569, + 0.0006283909315243363, + 0.00041124148992821574, + 0.0002186830242862925, + 0.00028800280415453017, + 0.00036124358302913606, + 0.001351600862108171, + 0.0003412006190046668, + 0.0004558642685879022, + 0.0006799938855692744, + 0.0015567982336506248, + 0.00035449964343570173, + 0.0006882911548018456, + 0.0008310693665407598, + 0.0014401640510186553, + 0.0007776628481224179, + 0.0008261203183792531, + 0.0021270345896482468, + 0.0018403074936941266, + 0.0026969127357006073, + 0.001958494307473302, + 0.0007615754147991538, + 0.001929751131683588, + 0.0009467026102356613, + 0.0005000894889235497, + 0.0010383253684267402, + 0.0018794579664245248, + 0.0006543504423461854, + 0.0006543761119246483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.017008861526846886 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, black, metallic sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car violently collide with a cold, galvanized steel guardrail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant orange and white sparks visible blasting from the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift pullback and zoom out immediately after the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final scene set on a winding mountain road with a sheer cliff face under a starless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703514099121094 + ], + "opposite_logprob_tokens": [ + -2.127035140991211 + ], + "logprob_reward": -0.12703514099121094, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12703514099121094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8772074580192566, + 0.031126881018280983, + 0.009048206731677055, + 0.006038760766386986, + 0.03189663216471672, + 0.008208036422729492, + 0.005840160883963108, + 0.006093365605920553, + 0.0192773025482893, + 0.005263186059892178 + ], + "credit_map": [ + 0.06837847828865051, + 0.3514210879802704, + 0.32473641633987427, + 0.033389635384082794, + 0.02500850521028042, + 0.018209634348750114, + 0.029092665761709213, + 0.02697095088660717, + 0.0030058070551604033, + 0.003626645542681217, + 0.003369635436683893, + 0.00462490925565362, + 0.00596762215718627, + 0.004604678135365248, + 0.0031998984050005674, + 0.0027276864275336266, + 0.0012454786337912083, + 0.001242644852027297, + 0.0010226097656413913, + 0.000694998714607209, + 0.0014582985313609242, + 0.0013962405500933528, + 0.0009128126548603177, + 0.0010751229710876942, + 0.0007272393559105694, + 0.0006659997161477804, + 0.000620901642832905, + 0.0007367403595708311, + 0.001348815974779427, + 0.0007043849909678102, + 0.0007732735248282552, + 0.0004614049685187638, + 0.0014817067421972752, + 0.01034828182309866, + 0.0008677946170791984, + 0.0010918410262092948, + 0.0028923549689352512, + 0.012710576876997948, + 0.001103913877159357, + 0.0014001602539792657, + 0.0005733019206672907, + 0.0017336332239210606, + 0.0006498544244095683, + 0.0008186005870811641, + 0.0011229657102376223, + 0.001800592290237546, + 0.0006856074905954301, + 0.0008234810666181147, + 0.0005174165708012879, + 0.0008444136474281549, + 0.00038820429472252727, + 0.0003744697605725378, + 0.001852134126238525, + 0.0009598708129487932, + 0.0005014854250475764, + 0.0004021659551654011, + 0.0005927378078922629, + 0.0008264583884738386, + 0.0005372281302697957, + 0.0007391657563857734, + 0.0009089508093893528, + 0.0009316563373431563, + 0.0006200206698849797, + 0.0009371477644890547, + 0.0010632964549586177, + 0.0033203153870999813, + 0.0014160972787067294, + 0.0010136542841792107, + 0.002897315425798297, + 0.003278546268120408, + 0.003858091775327921, + 0.002429987071081996, + 0.0005003743572160602, + 0.0007263053557835519, + 0.000609214068390429, + 0.00038330774987116456, + 0.0010102191008627415, + 0.0007496732287108898, + 0.0006282584508880973, + 0.000655833980999887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.12703514099121094 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, black, metallic sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car violently collide with a cold, galvanized steel guardrail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant orange and white sparks visible blasting from the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift pullback and zoom out immediately after the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final scene set on a winding mountain road with a sheer cliff face under a starless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015150746330618858 + ], + "opposite_logprob_tokens": [ + -6.501514911651611 + ], + "logprob_reward": -0.0015150746330618858, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015150746330618858, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7561095356941223, + 0.15057790279388428, + 0.023553386330604553, + 0.011141002178192139, + 0.0085993567481637, + 0.009778805077075958, + 0.0068405307829380035, + 0.01221050601452589, + 0.010444224812090397, + 0.010744811967015266 + ], + "credit_map": [ + 0.059046246111392975, + 0.30781862139701843, + 0.2717475891113281, + 0.028103115037083626, + 0.022408880293369293, + 0.01661413349211216, + 0.02715556137263775, + 0.023215308785438538, + 0.012924587354063988, + 0.01375703327357769, + 0.01675558276474476, + 0.01968739740550518, + 0.028358493000268936, + 0.019514424726366997, + 0.01992536522448063, + 0.01965501345694065, + 0.0026387390680611134, + 0.0031133231241256, + 0.002093629678711295, + 0.0015939384466037154, + 0.005551163572818041, + 0.003783827181905508, + 0.002264991868287325, + 0.0025137709453701973, + 0.0012706548441201448, + 0.0011034922208637, + 0.0009029294014908373, + 0.0014017621288076043, + 0.0024477378465235233, + 0.0012223066296428442, + 0.0019507684046402574, + 0.0008413498871959746, + 0.0010643191635608673, + 0.0016555192414671183, + 0.0007696488173678517, + 0.0007722636801190674, + 0.001090516452677548, + 0.001086886622942984, + 0.0011896856594830751, + 0.0009705167612992227, + 0.0009173472644761205, + 0.0012406195746734738, + 0.0008197535062208772, + 0.0008975624805316329, + 0.0016603640979155898, + 0.0014994938392192125, + 0.0010151260066777468, + 0.0017285372596234083, + 0.0006568953394889832, + 0.000980545300990343, + 0.0007674348307773471, + 0.0006154307629913092, + 0.0010181822581216693, + 0.0008971770876087248, + 0.0006919075967743993, + 0.0012129563838243484, + 0.0005783852539025247, + 0.0026570193003863096, + 0.0008453452028334141, + 0.0011791375000029802, + 0.001222087419591844, + 0.003276465693488717, + 0.00086145696695894, + 0.001590608386322856, + 0.0007875050650909543, + 0.001768041285686195, + 0.0007023149519227445, + 0.0007321017328649759, + 0.0015962189063429832, + 0.0024679312482476234, + 0.0008801650837995112, + 0.0015099445590749383, + 0.001068557845428586, + 0.0019832830876111984, + 0.0008644465706311166, + 0.0006561279296875, + 0.0016906758537515998, + 0.0018353641498833895, + 0.0011133054504171014, + 0.0015330492751672864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0015150746330618858 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, black, metallic sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car violently collide with a cold, galvanized steel guardrail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant orange and white sparks visible blasting from the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift pullback and zoom out immediately after the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final scene set on a winding mountain road with a sheer cliff face under a starless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006731097586452961 + ], + "opposite_logprob_tokens": [ + -5.006731033325195 + ], + "logprob_reward": -0.006731097586452961, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006731097586452961, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42800626158714294, + 0.029804030433297157, + 0.020793264731764793, + 0.041584160178899765, + 0.26269394159317017, + 0.04271886870265007, + 0.015670783817768097, + 0.06602218747138977, + 0.0667400062084198, + 0.025966499000787735 + ], + "credit_map": [ + 0.032743457704782486, + 0.1668540984392166, + 0.15845002233982086, + 0.01716144010424614, + 0.01219893153756857, + 0.010924546048045158, + 0.01461396086961031, + 0.015059812925755978, + 0.0034951947163790464, + 0.004251214675605297, + 0.003278245683759451, + 0.004504273645579815, + 0.0040549649856984615, + 0.004363459534943104, + 0.002596305450424552, + 0.0032603717409074306, + 0.0023334429133683443, + 0.0026902302633970976, + 0.0017768172547221184, + 0.0017143413424491882, + 0.003967109136283398, + 0.0030957660637795925, + 0.0020719815511256456, + 0.003143576206639409, + 0.004394723102450371, + 0.004589515272527933, + 0.006160961464047432, + 0.007961605675518513, + 0.004306014161556959, + 0.004591083154082298, + 0.006311732344329357, + 0.0032685203477740288, + 0.011444870382547379, + 0.09078678488731384, + 0.007321739103645086, + 0.008536153472959995, + 0.01894002966582775, + 0.10548197478055954, + 0.008489852771162987, + 0.011692540720105171, + 0.003412538208067417, + 0.007006244733929634, + 0.003981670830398798, + 0.004990838468074799, + 0.004340743646025658, + 0.009100549854338169, + 0.0036994677502661943, + 0.00618681451305747, + 0.0012512891553342342, + 0.0025908532552421093, + 0.001594890607520938, + 0.0012850683415308595, + 0.0027047886978834867, + 0.002194472821429372, + 0.0025528869591653347, + 0.0014965346781536937, + 0.00480620376765728, + 0.011728378012776375, + 0.0065923938527703285, + 0.010117369703948498, + 0.006090916693210602, + 0.011304013431072235, + 0.006334849167615175, + 0.009048061445355415, + 0.004316200967878103, + 0.009334090165793896, + 0.00646986486390233, + 0.006048019044101238, + 0.006974346470087767, + 0.009908395819365978, + 0.014847970567643642, + 0.008841121569275856, + 0.002280355431139469, + 0.004917892161756754, + 0.0032941093668341637, + 0.0020068292506039143, + 0.003948437515646219, + 0.004332099575549364, + 0.002578204032033682, + 0.002608574228361249 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.006731097586452961 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, black, metallic sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car violently collide with a cold, galvanized steel guardrail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant orange and white sparks visible blasting from the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift pullback and zoom out immediately after the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final scene set on a winding mountain road with a sheer cliff face under a starless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.974227786064148 + ], + "opposite_logprob_tokens": [ + -0.47422775626182556 + ], + "logprob_reward": -0.974227786064148, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.974227786064148, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.6650691628456116, + 0.03723065182566643, + 0.038583170622587204, + 0.05545998364686966, + 0.033523187041282654, + 0.05218258500099182, + 0.014684496447443962, + 0.023854408413171768, + 0.05680878087878227, + 0.022603614255785942 + ], + "credit_map": [ + 0.05209102854132652, + 0.2690940499305725, + 0.24025732278823853, + 0.024176480248570442, + 0.019787542521953583, + 0.015109965577721596, + 0.023347405716776848, + 0.021205369383096695, + 0.003639610717073083, + 0.00446869432926178, + 0.0032895125914365053, + 0.005986014846712351, + 0.0059218029491603374, + 0.0057517229579389095, + 0.0031724560540169477, + 0.005000832956284285, + 0.00710988650098443, + 0.005093917716294527, + 0.006597795989364386, + 0.0020855118054896593, + 0.005334116518497467, + 0.00602547824382782, + 0.003015244612470269, + 0.0033212141133844852, + 0.005946800112724304, + 0.006379624363034964, + 0.0038367302622646093, + 0.005300708115100861, + 0.01715584471821785, + 0.008694279938936234, + 0.005380673334002495, + 0.0027653162833303213, + 0.0017372163711115718, + 0.010369693860411644, + 0.0015883267624303699, + 0.0015347925946116447, + 0.0026440289802849293, + 0.011783675290644169, + 0.0018831348279491067, + 0.0019823159091174603, + 0.005456049460917711, + 0.00785532034933567, + 0.004918009974062443, + 0.0033958505373448133, + 0.010303824208676815, + 0.007979510352015495, + 0.003651699284091592, + 0.008622316643595695, + 0.0011985208839178085, + 0.0021046383772045374, + 0.0008261990733444691, + 0.0007434783619828522, + 0.005598385352641344, + 0.0024810463655740023, + 0.000819289474748075, + 0.0009129371028393507, + 0.00160397426225245, + 0.004508843179792166, + 0.002382288919761777, + 0.002861729823052883, + 0.002734722336754203, + 0.004105314612388611, + 0.0021182617638260126, + 0.003539272118359804, + 0.003358932211995125, + 0.00875502172857523, + 0.0047293794341385365, + 0.00446451036259532, + 0.007622654549777508, + 0.009001697413623333, + 0.010986129753291607, + 0.007890450768172741, + 0.00224973913282156, + 0.003927506506443024, + 0.003392318496480584, + 0.0016345704207196832, + 0.003823357867076993, + 0.003335480112582445, + 0.0022588481660932302, + 0.001981794135645032 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.974227786064148 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "dynamic, low-angle close-up captures the brutal force of a sleek, black, metallic sports car as it violently collides with a cold, galvanized steel guardrail. The car's front fenders instantly collapse inward, sending a shower of brilliant orange and white sparks blasting into the pitch-black night, completely illuminating the twisted metal for a few seconds. The vehicle chaotically bounces off the obstacle, tires screeching on the wet asphalt, the chassis shaking, and the red glow of its taillights smearing across the frame. Immediately after the impact, the camera executes a swift and dramatic pullback, quickly zooming out to reveal the full, terrifying context: the stricken car, now small and isolated, its headlights casting a sharp, lonely beam against the darkness, trapped on the sharp curve of a winding mountain road that twists precariously around a sheer cliff face and disappears into the oppressive darkness of the surrounding wilderness beneath a starless sky, creating a tense, perilous atmosphere of sudden disaster and profound isolation.", + "seed": 21792, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_4/index_7862_seed_21792_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.12726611560210585, + "vlm_logprob": -0.12726611560210585, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.12726611560210585, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, black, metallic sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car violently collide with a cold, galvanized steel guardrail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant orange and white sparks visible blasting from the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift pullback and zoom out immediately after the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final scene set on a winding mountain road with a sheer cliff face under a starless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1832089126110077 + ], + "opposite_logprob_tokens": [ + -1.80820894241333 + ], + "logprob_reward": -0.1832089126110077, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1832089126110077, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08616848289966583, + 0.02967951074242592, + 0.04053056240081787, + 0.22792483866214752, + 0.020255815237760544, + 0.0907084196805954, + 0.020631179213523865, + 0.4152248501777649, + 0.03707563132047653, + 0.03180064633488655 + ], + "credit_map": [ + 0.005384568125009537, + 0.006411857903003693, + 0.009641332551836967, + 0.013732892461121082, + 0.006941983476281166, + 0.013223757967352867, + 0.010377260856330395, + 0.020454825833439827, + 0.0034309171605855227, + 0.0026528898160904646, + 0.0035046220291405916, + 0.004210765939205885, + 0.004530502483248711, + 0.0026821729261428118, + 0.00448131887242198, + 0.004186320584267378, + 0.00530517240986228, + 0.003472651122137904, + 0.0036967117339372635, + 0.004234198480844498, + 0.005963679403066635, + 0.006056281737983227, + 0.005370797589421272, + 0.006431072950363159, + 0.027793310582637787, + 0.014662439003586769, + 0.019929639995098114, + 0.021394170820713043, + 0.033130425959825516, + 0.032184991985559464, + 0.048095956444740295, + 0.03073388710618019, + 0.00175752944778651, + 0.004214980639517307, + 0.0029360600747168064, + 0.0018954038387164474, + 0.00343524431809783, + 0.001616816152818501, + 0.001961772795766592, + 0.0024380080867558718, + 0.003747418988496065, + 0.03304499015212059, + 0.007272763643413782, + 0.00379161792807281, + 0.028829146176576614, + 0.006496985908597708, + 0.004044536035507917, + 0.003480966668576002, + 0.0017366306856274605, + 0.002776344073936343, + 0.0033050402998924255, + 0.0026971730403602123, + 0.0028079182375222445, + 0.002670665504410863, + 0.0024856270756572485, + 0.0021517800632864237, + 0.012772953137755394, + 0.1742067188024521, + 0.006341693457216024, + 0.007752707693725824, + 0.02238004468381405, + 0.17649130523204803, + 0.006150717847049236, + 0.009128708392381668, + 0.0025596406776458025, + 0.01337158028036356, + 0.0023789198603481054, + 0.001898184884339571, + 0.005315363872796297, + 0.0067669604904949665, + 0.0013665971346199512, + 0.003418383654206991, + 0.0020330348052084446, + 0.0074711330235004425, + 0.0030598652083426714, + 0.0041193109937012196, + 0.0053841182962059975, + 0.003801100654527545, + 0.0031143114902079105, + 0.002817772561684251 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 5 + ], + "selected_reward": -0.1832089126110077 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, black, metallic sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car violently collide with a cold, galvanized steel guardrail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant orange and white sparks visible blasting from the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift pullback and zoom out immediately after the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final scene set on a winding mountain road with a sheer cliff face under a starless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496309697628021 + ], + "opposite_logprob_tokens": [ + -2.9299631118774414 + ], + "logprob_reward": -0.05496309697628021, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496309697628021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05527842044830322, + 0.03939426317811012, + 0.028186023235321045, + 0.04970363900065422, + 0.04004433751106262, + 0.07061733305454254, + 0.02934698574244976, + 0.6245786547660828, + 0.04357896372675896, + 0.019271405413746834 + ], + "credit_map": [ + 0.0036824755370616913, + 0.005742603447288275, + 0.00467540742829442, + 0.009548314847052097, + 0.008236547000706196, + 0.006247144658118486, + 0.007762865163385868, + 0.009383059106767178, + 0.004526252392679453, + 0.004872411955147982, + 0.0022437938023358583, + 0.0040227374993264675, + 0.006919289007782936, + 0.007119480054825544, + 0.004451475106179714, + 0.0052388194017112255, + 0.004041188862174749, + 0.003252958180382848, + 0.0032678593415766954, + 0.003037476446479559, + 0.0035802109632641077, + 0.004357788246124983, + 0.0030352137982845306, + 0.003613326232880354, + 0.009185460396111012, + 0.0073999566957354546, + 0.004232741426676512, + 0.004357920493930578, + 0.004954457748681307, + 0.0045539638958871365, + 0.00827667023986578, + 0.006742469500750303, + 0.003675971645861864, + 0.006383911240845919, + 0.005582854151725769, + 0.004529250785708427, + 0.005947144702076912, + 0.003320937277749181, + 0.005032066255807877, + 0.005572201684117317, + 0.0027854680083692074, + 0.02112746611237526, + 0.008537680841982365, + 0.0034958631731569767, + 0.022487055510282516, + 0.006239046808332205, + 0.003494547912850976, + 0.002450200030580163, + 0.0024937856942415237, + 0.00415516784414649, + 0.004567700903862715, + 0.004542520269751549, + 0.003585214726626873, + 0.003874805523082614, + 0.003594031324610114, + 0.0025337613187730312, + 0.019434798508882523, + 0.26355159282684326, + 0.00972515158355236, + 0.012226507999002934, + 0.03271030634641647, + 0.26398178935050964, + 0.007964594289660454, + 0.014983929693698883, + 0.0031566680409014225, + 0.013063246384263039, + 0.0035508594010025263, + 0.0033592735417187214, + 0.004872793331742287, + 0.008887500502169132, + 0.002549797995015979, + 0.004138822667300701, + 0.0012330439640209079, + 0.002913345582783222, + 0.0020355915185064077, + 0.0023715472780168056, + 0.002663919236510992, + 0.0033061779104173183, + 0.0021314651239663363, + 0.0026163156144320965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.05496309697628021 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, black, metallic sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car violently collide with a cold, galvanized steel guardrail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant orange and white sparks visible blasting from the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift pullback and zoom out immediately after the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final scene set on a winding mountain road with a sheer cliff face under a starless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024937265552580357 + ], + "opposite_logprob_tokens": [ + -6.002493858337402 + ], + "logprob_reward": -0.0024937265552580357, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024937265552580357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08878801017999649, + 0.03686865046620369, + 0.03598465770483017, + 0.16868135333061218, + 0.05824318900704384, + 0.1321997493505478, + 0.051482658833265305, + 0.32870620489120483, + 0.05230679363012314, + 0.046738676726818085 + ], + "credit_map": [ + 0.005536757875233889, + 0.010150391608476639, + 0.007136131636798382, + 0.012162243016064167, + 0.01214186754077673, + 0.011760853230953217, + 0.010336317121982574, + 0.019563443958759308, + 0.0044608390890061855, + 0.003978808410465717, + 0.002996917814016342, + 0.004637992009520531, + 0.004599513486027718, + 0.003803938627243042, + 0.006471632048487663, + 0.0059190113097429276, + 0.004770546220242977, + 0.0029589980840682983, + 0.003396786516532302, + 0.0026474439073354006, + 0.002735297428444028, + 0.0077594309113919735, + 0.005247351713478565, + 0.006468805950134993, + 0.03442292660474777, + 0.0267333947122097, + 0.01600661128759384, + 0.01781119965016842, + 0.018988510593771935, + 0.016666235402226448, + 0.01766994222998619, + 0.020382555201649666, + 0.004199606366455555, + 0.015013775788247585, + 0.00442279689013958, + 0.004211673978716135, + 0.015141339041292667, + 0.004275421146303415, + 0.005137480795383453, + 0.005841098725795746, + 0.0050776260904967785, + 0.04358182102441788, + 0.014404411427676678, + 0.006078062113374472, + 0.04171963036060333, + 0.01067321840673685, + 0.006520778872072697, + 0.004144212696701288, + 0.004251090809702873, + 0.0072634839452803135, + 0.011492560617625713, + 0.009733647108078003, + 0.004934667143970728, + 0.005110297817736864, + 0.004705314990133047, + 0.00399160198867321, + 0.009813699871301651, + 0.13318473100662231, + 0.007195386569947004, + 0.00783386081457138, + 0.01820985972881317, + 0.1373436152935028, + 0.006323888432234526, + 0.008801176212728024, + 0.004249572288244963, + 0.016789499670267105, + 0.004517082124948502, + 0.004668065346777439, + 0.00989330094307661, + 0.005904211197048426, + 0.003153778612613678, + 0.003131288569420576, + 0.0035897281486541033, + 0.008338541723787785, + 0.005224432330578566, + 0.005411977879703045, + 0.005038181319832802, + 0.009626379236578941, + 0.0036279151681810617, + 0.005881526507437229 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 5 + ], + "selected_reward": -0.0024937265552580357 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, black, metallic sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car violently collide with a cold, galvanized steel guardrail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant orange and white sparks visible blasting from the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift pullback and zoom out immediately after the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final scene set on a winding mountain road with a sheer cliff face under a starless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008632958866655827 + ], + "opposite_logprob_tokens": [ + -4.758633136749268 + ], + "logprob_reward": -0.008632958866655827, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008632958866655827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06006189435720444, + 0.06880422681570053, + 0.07526490837335587, + 0.05388525128364563, + 0.09952306747436523, + 0.09379436075687408, + 0.034827060997486115, + 0.4579288065433502, + 0.024501735344529152, + 0.03140866383910179 + ], + "credit_map": [ + 0.0052431439980864525, + 0.007184404414147139, + 0.003183072665706277, + 0.006158244796097279, + 0.012121114879846573, + 0.007759039290249348, + 0.007825779728591442, + 0.010587094351649284, + 0.005598908755928278, + 0.005958306137472391, + 0.00972642470151186, + 0.011912637390196323, + 0.007675246801227331, + 0.0059004477225244045, + 0.011153556406497955, + 0.010878702625632286, + 0.01036152709275484, + 0.008064854890108109, + 0.0051681469194591045, + 0.005143910180777311, + 0.017441343516111374, + 0.008187025785446167, + 0.00816921517252922, + 0.012728886678814888, + 0.008073599077761173, + 0.005636530928313732, + 0.005131169687956572, + 0.005997996311634779, + 0.006504679564386606, + 0.007477960083633661, + 0.008687248453497887, + 0.006376068107783794, + 0.007068261969834566, + 0.025139713659882545, + 0.008316160179674625, + 0.006996683776378632, + 0.026121262460947037, + 0.007979558780789375, + 0.008776175789535046, + 0.009125255048274994, + 0.00854151789098978, + 0.014561854302883148, + 0.019809972494840622, + 0.00888837967067957, + 0.016800204291939735, + 0.012097124941647053, + 0.007244457956403494, + 0.005850851070135832, + 0.0029198750853538513, + 0.004109835717827082, + 0.007301324512809515, + 0.007968955673277378, + 0.003264869097620249, + 0.0032505833078175783, + 0.003367017023265362, + 0.002644600346684456, + 0.013677227310836315, + 0.19339700043201447, + 0.007437456864863634, + 0.008051694370806217, + 0.024143004789948463, + 0.19514915347099304, + 0.005832053255289793, + 0.010241267271339893, + 0.001747348578646779, + 0.009456112049520016, + 0.0016767665510997176, + 0.0009327803272753954, + 0.003836458083242178, + 0.0045318384654819965, + 0.0006973984418436885, + 0.0016230319160968065, + 0.002184856915846467, + 0.007422527763992548, + 0.003187982365489006, + 0.004303362220525742, + 0.0054204268380999565, + 0.004126924090087414, + 0.002282407134771347, + 0.002480176743119955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 5 + ], + "selected_reward": -0.008632958866655827 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, black, metallic sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car violently collide with a cold, galvanized steel guardrail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant orange and white sparks visible blasting from the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift pullback and zoom out immediately after the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final scene set on a winding mountain road with a sheer cliff face under a starless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870318830013275 + ], + "opposite_logprob_tokens": [ + -1.13703191280365 + ], + "logprob_reward": -0.3870318830013275, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870318830013275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03367140144109726, + 0.02346576191484928, + 0.04037243872880936, + 0.07562859356403351, + 0.024307115003466606, + 0.0894785150885582, + 0.09197982400655746, + 0.5309827923774719, + 0.035085245966911316, + 0.055028412491083145 + ], + "credit_map": [ + 0.0028557153418660164, + 0.004720732104033232, + 0.002213718369603157, + 0.005069664213806391, + 0.004797105677425861, + 0.004122308921068907, + 0.003881049109622836, + 0.006011102348566055, + 0.003040048759430647, + 0.002181784715503454, + 0.00255964114330709, + 0.0028186317067593336, + 0.004390918184071779, + 0.0025389790534973145, + 0.0033721241634339094, + 0.002563630696386099, + 0.005543303210288286, + 0.003945455886423588, + 0.0023152772337198257, + 0.003447562688961625, + 0.0047224340960383415, + 0.008756370283663273, + 0.005653539672493935, + 0.005988487973809242, + 0.009654226712882519, + 0.0047971936874091625, + 0.007192989345639944, + 0.007880127057433128, + 0.010104900225996971, + 0.01025465875864029, + 0.015532441437244415, + 0.010212044231593609, + 0.0020016650669276714, + 0.0031423266045749187, + 0.002038394333794713, + 0.0024172274861484766, + 0.003615827765315771, + 0.0029914516489952803, + 0.0035990593023598194, + 0.004501159302890301, + 0.004210687708109617, + 0.02502625621855259, + 0.009350479580461979, + 0.005494562443345785, + 0.024854572489857674, + 0.008289296180009842, + 0.006381693296134472, + 0.005870952736586332, + 0.008664152584969997, + 0.011820056475698948, + 0.022191712632775307, + 0.018717432394623756, + 0.0074720256961882114, + 0.010406344197690487, + 0.005602932535111904, + 0.007105150260031223, + 0.015902850776910782, + 0.2219017893075943, + 0.009755484759807587, + 0.010347813367843628, + 0.027266545221209526, + 0.22619377076625824, + 0.0075072068721055984, + 0.012107294984161854, + 0.002689145039767027, + 0.012781494297087193, + 0.0026834411546587944, + 0.0015809382311999798, + 0.005556994117796421, + 0.006174078676849604, + 0.0013808212243020535, + 0.002238326705992222, + 0.0041292826645076275, + 0.008948288857936859, + 0.005895450245589018, + 0.003378052031621337, + 0.005660363007336855, + 0.013851498253643513, + 0.004809154197573662, + 0.008356316946446896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.3870318830013275 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_5/index_tensor([2128, 2128])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_5/index_tensor([2128, 2128])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fc9164727efd2fcd046b2690a75caf7e0bf3b7db --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_5/index_tensor([2128, 2128])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up slowly pans across the face of a woman, her features a mask of pure terror amidst a dark and intense horror scene. Her skin is deathly pale, almost translucent, providing a stark canvas for the fiery red hair that cascades across her forehead and cheeks in a tousled, damp mass. The scene is illuminated by a single, harsh overhead light, casting deep, dramatic shadows that sculpt her features and enhance the eerie, claustrophobic atmosphere. Her wide-open eyes, a startlingly clear blue, gaze at an unseen point beyond the camera, reflecting a profound terror. A stream of thick, crimson blood steadily and uncomfortably drips from her right nostril, tracing a path along her philtrum, while another stream of fresh blood flows from the corner of her slightly open mouth, which hangs open as she gasps shallowly and raggedly. Her jaw trembles uncontrollably, and the entire composition relentlessly focuses on her anguished facial expression and the slow, terrifying descent of blood, creating a deeply unsettling and intensely terrifying visual narrative.", + "seed": 15069, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_5/index_2128_seed_15069_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.33410713509656487, + "vlm_logprob": -0.33410713509656487, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.33410713509656487, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly across the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's skin deathly pale and almost translucent with fiery red hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single harsh overhead light casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a stream of blood dripping from her right nostril and flowing from the corner of her mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005573803558945656 + ], + "opposite_logprob_tokens": [ + -7.005573749542236 + ], + "logprob_reward": -0.005573803558945656, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.005573803558945656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10507955402135849, + 0.15789435803890228, + 0.15644745528697968, + 0.07708973437547684, + 0.07744624465703964, + 0.12676450610160828, + 0.13653779029846191, + 0.060699883848428726, + 0.056195519864559174, + 0.0458449125289917 + ], + "credit_map": [ + 0.009689896367490292, + 0.017979849129915237, + 0.009614656679332256, + 0.010992703959345818, + 0.011153150349855423, + 0.015662843361496925, + 0.015959378331899643, + 0.01402706652879715, + 0.020103873685002327, + 0.01683323085308075, + 0.016628995537757874, + 0.009491698816418648, + 0.01451893150806427, + 0.02886745147407055, + 0.029663465917110443, + 0.02178671397268772, + 0.008506734855473042, + 0.0126096336171031, + 0.013198806904256344, + 0.016381315886974335, + 0.019202740862965584, + 0.021043146029114723, + 0.053408585488796234, + 0.012096493504941463, + 0.007592189125716686, + 0.007358509115874767, + 0.014026227407157421, + 0.006022046785801649, + 0.00718827685341239, + 0.01154110487550497, + 0.013804007321596146, + 0.009557372890412807, + 0.012823144905269146, + 0.008081568405032158, + 0.017750607803463936, + 0.005693465005606413, + 0.008190673775970936, + 0.010376131162047386, + 0.008028553798794746, + 0.006502101663500071, + 0.010920110158622265, + 0.020592156797647476, + 0.013004221953451633, + 0.011035691946744919, + 0.023325154557824135, + 0.01579209975898266, + 0.016995377838611603, + 0.015099688433110714, + 0.01385554950684309, + 0.012781822122633457, + 0.02064800262451172, + 0.014362504705786705, + 0.01476496085524559, + 0.018961841240525246, + 0.027510853484272957, + 0.013652268797159195, + 0.005123376380652189, + 0.0081931222230196, + 0.007123615592718124, + 0.0069666663184762, + 0.007759764324873686, + 0.008654303848743439, + 0.011032572947442532, + 0.005846461746841669, + 0.0045178961008787155, + 0.00711440946906805, + 0.006878956686705351, + 0.005215582903474569, + 0.008572520688176155, + 0.007095123175531626, + 0.011417771689593792, + 0.005383254494518042, + 0.007143249735236168, + 0.00316468789242208, + 0.006568138021975756, + 0.00357644259929657, + 0.003744156565517187, + 0.00883034523576498, + 0.004381590988487005, + 0.008436302654445171 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 6 + ], + "selected_reward": -0.005573803558945656 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly across the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's skin deathly pale and almost translucent with fiery red hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single harsh overhead light casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a stream of blood dripping from her right nostril and flowing from the corner of her mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4063173532485962 + ], + "opposite_logprob_tokens": [ + -0.2813173830509186 + ], + "logprob_reward": -1.4063173532485962, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4063173532485962, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.12300100177526474, + 0.17795626819133759, + 0.1407974809408188, + 0.05569221079349518, + 0.0949268713593483, + 0.1519124060869217, + 0.05706801265478134, + 0.07255437970161438, + 0.024395786225795746, + 0.10169566422700882 + ], + "credit_map": [ + 0.009426803328096867, + 0.01771165244281292, + 0.01513068936765194, + 0.0072966706939041615, + 0.01160703506320715, + 0.021288558840751648, + 0.018032342195510864, + 0.02250724844634533, + 0.023078883066773415, + 0.012770399451255798, + 0.042359188199043274, + 0.0101679852232337, + 0.01098245196044445, + 0.027252446860074997, + 0.02966936305165291, + 0.021675538271665573, + 0.01170430053025484, + 0.022415973246097565, + 0.01630517654120922, + 0.012475000694394112, + 0.012726014479994774, + 0.01697416976094246, + 0.03401331976056099, + 0.01418352872133255, + 0.005982647649943829, + 0.006543723400682211, + 0.011256356723606586, + 0.005388335324823856, + 0.005624992307275534, + 0.007265659514814615, + 0.007244470529258251, + 0.006386025808751583, + 0.014197750017046928, + 0.009226629510521889, + 0.018687495961785316, + 0.006976778153330088, + 0.010612239129841328, + 0.014088507741689682, + 0.01387767307460308, + 0.007259794976562262, + 0.01566719077527523, + 0.018957020714879036, + 0.020732643082737923, + 0.010975362733006477, + 0.020522980019450188, + 0.023781972005963326, + 0.022754060104489326, + 0.018521171063184738, + 0.008655240759253502, + 0.0066145299933850765, + 0.009755908511579037, + 0.004281678702682257, + 0.005544833838939667, + 0.010032127611339092, + 0.006842100992798805, + 0.005341593641787767, + 0.0074144103564321995, + 0.010455428622663021, + 0.008309018798172474, + 0.006233637221157551, + 0.009328247047960758, + 0.010305696167051792, + 0.013348235748708248, + 0.007159702945500612, + 0.002305921632796526, + 0.0029675131663680077, + 0.003080613212659955, + 0.0024578613229095936, + 0.0027897495310753584, + 0.003246321575716138, + 0.00494944304227829, + 0.002598363906145096, + 0.01442202739417553, + 0.008871006779372692, + 0.017351584509015083, + 0.006163565907627344, + 0.013218004256486893, + 0.016365114599466324, + 0.009992862120270729, + 0.015311509370803833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 2 + ], + "selected_reward": -1.4063173532485962 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly across the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's skin deathly pale and almost translucent with fiery red hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single harsh overhead light casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a stream of blood dripping from her right nostril and flowing from the corner of her mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011897159274667501 + ], + "opposite_logprob_tokens": [ + -6.751189708709717 + ], + "logprob_reward": -0.0011897159274667501, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011897159274667501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07260165363550186, + 0.127489373087883, + 0.09689828008413315, + 0.1431509554386139, + 0.09836464375257492, + 0.31042972207069397, + 0.03825658559799194, + 0.03868735581636429, + 0.04068067669868469, + 0.0334407277405262 + ], + "credit_map": [ + 0.0078138317912817, + 0.005525726824998856, + 0.011506051756441593, + 0.004902323242276907, + 0.007225723005831242, + 0.011314469389617443, + 0.013279866427183151, + 0.011033664457499981, + 0.014025837182998657, + 0.013961516320705414, + 0.02488185279071331, + 0.00758718978613615, + 0.012723272666335106, + 0.017571434378623962, + 0.023670708760619164, + 0.013067563995718956, + 0.009300466626882553, + 0.015484103932976723, + 0.01397376786917448, + 0.009367824532091618, + 0.008727863430976868, + 0.012745438143610954, + 0.016580887138843536, + 0.010717927478253841, + 0.015504051931202412, + 0.016082074493169785, + 0.028907205909490585, + 0.013333105482161045, + 0.015912173315882683, + 0.017932182177901268, + 0.022990211844444275, + 0.012489945627748966, + 0.013976023532450199, + 0.00953237060457468, + 0.01755194365978241, + 0.00772870471701026, + 0.01255702879279852, + 0.014283744618296623, + 0.01454742718487978, + 0.008187402039766312, + 0.029941199347376823, + 0.046436894685029984, + 0.03735249117016792, + 0.028883909806609154, + 0.052379950881004333, + 0.04376711696386337, + 0.03690426051616669, + 0.034763921052217484, + 0.0037332705687731504, + 0.003367674769833684, + 0.006186121609061956, + 0.004820503760129213, + 0.003286557737737894, + 0.00498262420296669, + 0.007095418870449066, + 0.004784414544701576, + 0.004700401332229376, + 0.005632698070257902, + 0.006186256185173988, + 0.0038384420331567526, + 0.0038260528817772865, + 0.004684416111558676, + 0.0059369406662881374, + 0.003882147604599595, + 0.004296517930924892, + 0.004795458633452654, + 0.007305534090846777, + 0.0034199408255517483, + 0.004406370688229799, + 0.0054832990281283855, + 0.006929690949618816, + 0.004043868277221918, + 0.003316557500511408, + 0.002450040075927973, + 0.0040984563529491425, + 0.002384986961260438, + 0.002776772016659379, + 0.00665260199457407, + 0.005194919183850288, + 0.006566395051777363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 1 + ], + "selected_reward": -0.0011897159274667501 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly across the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's skin deathly pale and almost translucent with fiery red hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single harsh overhead light casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a stream of blood dripping from her right nostril and flowing from the corner of her mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25219863653182983 + ], + "opposite_logprob_tokens": [ + -1.5021986961364746 + ], + "logprob_reward": -0.25219863653182983, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25219863653182983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11763807386159897, + 0.07940255850553513, + 0.055320676416158676, + 0.15601719915866852, + 0.2551816403865814, + 0.14097169041633606, + 0.05727030336856842, + 0.03234762325882912, + 0.058394577354192734, + 0.047455690801143646 + ], + "credit_map": [ + 0.014730202965438366, + 0.013138929381966591, + 0.018421197310090065, + 0.007733498699963093, + 0.012096020393073559, + 0.019735338166356087, + 0.01258949376642704, + 0.019193394109606743, + 0.012586516328155994, + 0.007658501621335745, + 0.015806056559085846, + 0.005034320056438446, + 0.008952689357101917, + 0.012905651703476906, + 0.00867506954818964, + 0.007783760316669941, + 0.003935087006539106, + 0.005263802595436573, + 0.006172926165163517, + 0.006870512384921312, + 0.005695529282093048, + 0.007483475375920534, + 0.012682448141276836, + 0.007216898258775473, + 0.019194746389985085, + 0.01740945875644684, + 0.038133908063173294, + 0.008970444090664387, + 0.012607324868440628, + 0.022265756502747536, + 0.013058073818683624, + 0.024377478286623955, + 0.039040856063365936, + 0.025161484256386757, + 0.05290953442454338, + 0.01770300604403019, + 0.02895689196884632, + 0.038781579583883286, + 0.023062340915203094, + 0.02956593781709671, + 0.01022852398455143, + 0.022678306326270103, + 0.017083769664168358, + 0.013931781984865665, + 0.022727815434336662, + 0.018692947924137115, + 0.020493201911449432, + 0.015135357156395912, + 0.008187087252736092, + 0.004178404342383146, + 0.012464398518204689, + 0.0033827100414782763, + 0.005174619611352682, + 0.010874145664274693, + 0.006296481471508741, + 0.006712453905493021, + 0.0030201657209545374, + 0.0032500133384019136, + 0.004009455908089876, + 0.003267661202698946, + 0.004632560070604086, + 0.004650464281439781, + 0.0055358572863042355, + 0.003981445450335741, + 0.00574257830157876, + 0.008940380066633224, + 0.009001493453979492, + 0.007198782172054052, + 0.007003537844866514, + 0.0074298991821706295, + 0.007893020287156105, + 0.005184886511415243, + 0.006256673019379377, + 0.004865455906838179, + 0.005747686605900526, + 0.0037115709856152534, + 0.005851041525602341, + 0.007761865388602018, + 0.00718064047396183, + 0.006080755032598972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.25219863653182983 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly across the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's skin deathly pale and almost translucent with fiery red hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single harsh overhead light casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a stream of blood dripping from her right nostril and flowing from the corner of her mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005256166215986013 + ], + "opposite_logprob_tokens": [ + -5.255256175994873 + ], + "logprob_reward": -0.005256166215986013, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005256166215986013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09480803459882736, + 0.10770776867866516, + 0.098968006670475, + 0.0691729485988617, + 0.13023844361305237, + 0.10150859504938126, + 0.2341170459985733, + 0.02981003373861313, + 0.08404993265867233, + 0.04961921647191048 + ], + "credit_map": [ + 0.009140816517174244, + 0.018696144223213196, + 0.007834948599338531, + 0.0053957621566951275, + 0.010151186026632786, + 0.015095061622560024, + 0.013793481513857841, + 0.014700634405016899, + 0.0148384477943182, + 0.00856544729322195, + 0.023540418595075607, + 0.00516258692368865, + 0.008609623648226261, + 0.01735171303153038, + 0.015969671308994293, + 0.013669857755303383, + 0.008647847920656204, + 0.016721319407224655, + 0.009990168735384941, + 0.00731823593378067, + 0.009881255216896534, + 0.013099126517772675, + 0.01978200674057007, + 0.013528036884963512, + 0.0064858864061534405, + 0.007057743147015572, + 0.009421761147677898, + 0.007562490180134773, + 0.009428918361663818, + 0.008731820620596409, + 0.012231100350618362, + 0.008253222331404686, + 0.010752945207059383, + 0.017115361988544464, + 0.01860160380601883, + 0.016482030972838402, + 0.015141553245484829, + 0.018101664260029793, + 0.02277173474431038, + 0.011271537281572819, + 0.009480948559939861, + 0.01369399018585682, + 0.01378123089671135, + 0.007061456795781851, + 0.019418874755501747, + 0.015001408755779266, + 0.011651084758341312, + 0.011419598013162613, + 0.019801216199994087, + 0.0331590361893177, + 0.03306768462061882, + 0.02247604914009571, + 0.029932627454400063, + 0.02911757305264473, + 0.04439140111207962, + 0.022171439602971077, + 0.0032190787605941296, + 0.0037244011182338, + 0.0037289541214704514, + 0.0022167633287608624, + 0.004310593940317631, + 0.003883873810991645, + 0.005932551342993975, + 0.002793815452605486, + 0.004266042727977037, + 0.008786829188466072, + 0.01020702812820673, + 0.009425527416169643, + 0.01201961562037468, + 0.010660181753337383, + 0.021542904898524284, + 0.0071418010629713535, + 0.004647834226489067, + 0.004395923111587763, + 0.0047944579273462296, + 0.0055166371166706085, + 0.007734511513262987, + 0.006347717717289925, + 0.010413144715130329, + 0.005768990609794855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 1 + ], + "selected_reward": -0.005256166215986013 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up slowly pans across the face of a woman, her features a mask of pure terror amidst a dark and intense horror scene. Her skin is deathly pale, almost translucent, providing a stark canvas for the fiery red hair that cascades across her forehead and cheeks in a tousled, damp mass. The scene is illuminated by a single, harsh overhead light, casting deep, dramatic shadows that sculpt her features and enhance the eerie, claustrophobic atmosphere. Her wide-open eyes, a startlingly clear blue, gaze at an unseen point beyond the camera, reflecting a profound terror. A stream of thick, crimson blood steadily and uncomfortably drips from her right nostril, tracing a path along her philtrum, while another stream of fresh blood flows from the corner of her slightly open mouth, which hangs open as she gasps shallowly and raggedly. Her jaw trembles uncontrollably, and the entire composition relentlessly focuses on her anguished facial expression and the slow, terrifying descent of blood, creating a deeply unsettling and intensely terrifying visual narrative.", + "seed": 15070, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_5/index_2128_seed_15070_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.19874500455334781, + "vlm_logprob": -0.19874500455334781, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.19874500455334781, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly across the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's skin deathly pale and almost translucent with fiery red hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single harsh overhead light casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a stream of blood dripping from her right nostril and flowing from the corner of her mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005690802354365587 + ], + "opposite_logprob_tokens": [ + -6.880690574645996 + ], + "logprob_reward": -0.005690802354365587, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.005690802354365587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17474296689033508, + 0.0968698039650917, + 0.06276323646306992, + 0.047430697828531265, + 0.13334648311138153, + 0.05691355839371681, + 0.16592174768447876, + 0.07586495578289032, + 0.06077401340007782, + 0.1253725290298462 + ], + "credit_map": [ + 0.020298456773161888, + 0.021750081330537796, + 0.026105165481567383, + 0.02357175201177597, + 0.02214767411351204, + 0.02121397852897644, + 0.02030211314558983, + 0.019353769719600677, + 0.012273893691599369, + 0.01375569961965084, + 0.01697637140750885, + 0.008839347399771214, + 0.011790428310632706, + 0.012597660534083843, + 0.011198143474757671, + 0.009438267908990383, + 0.0058052451349794865, + 0.009777994826436043, + 0.006958089768886566, + 0.005980916786938906, + 0.008063876070082188, + 0.0061804489232599735, + 0.012194231152534485, + 0.00780244218185544, + 0.00422463146969676, + 0.005814302712678909, + 0.00589186605066061, + 0.006078613921999931, + 0.00599259790033102, + 0.0046754381619393826, + 0.008563152514398098, + 0.0061900983564555645, + 0.013230016455054283, + 0.013296005316078663, + 0.01860976032912731, + 0.01480918563902378, + 0.012368720956146717, + 0.02836231328547001, + 0.019193388521671295, + 0.013477100990712643, + 0.006432053633034229, + 0.006107548251748085, + 0.01303549762815237, + 0.004680932033807039, + 0.005082905292510986, + 0.010250207968056202, + 0.005643630865961313, + 0.005680791102349758, + 0.01417659129947424, + 0.019630253314971924, + 0.024119341745972633, + 0.027145441621541977, + 0.019078047946095467, + 0.021937839686870575, + 0.025116104632616043, + 0.014718162827193737, + 0.01027008518576622, + 0.008195738308131695, + 0.011040773242712021, + 0.00644685560837388, + 0.009868473745882511, + 0.008578500710427761, + 0.012418902479112148, + 0.009045635350048542, + 0.006455627735704184, + 0.00822720117866993, + 0.009094695560634136, + 0.005662230774760246, + 0.007121399510651827, + 0.009607967920601368, + 0.007821972481906414, + 0.006782921496778727, + 0.015548715367913246, + 0.0184900164604187, + 0.018999747931957245, + 0.013367329724133015, + 0.01106873620301485, + 0.025659525766968727, + 0.012007996439933777, + 0.010230470448732376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.005690802354365587 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly across the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's skin deathly pale and almost translucent with fiery red hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single harsh overhead light casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a stream of blood dripping from her right nostril and flowing from the corner of her mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261731266975403 + ], + "opposite_logprob_tokens": [ + -0.5761731266975403 + ], + "logprob_reward": -0.8261731266975403, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261731266975403, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10273010283708572, + 0.045873478055000305, + 0.039621226489543915, + 0.08531592041254044, + 0.056537821888923645, + 0.07451674342155457, + 0.48134100437164307, + 0.03560050576925278, + 0.027270836755633354, + 0.05119231343269348 + ], + "credit_map": [ + 0.012059892527759075, + 0.012771385721862316, + 0.01633690856397152, + 0.011747904121875763, + 0.009118379093706608, + 0.014661254361271858, + 0.014091148041188717, + 0.01194323692470789, + 0.0043600681237876415, + 0.005349926184862852, + 0.006258225999772549, + 0.004766224417835474, + 0.006115836091339588, + 0.0059348056092858315, + 0.008692450821399689, + 0.004395945463329554, + 0.0037367993500083685, + 0.0050400798209011555, + 0.005395935848355293, + 0.003293111687526107, + 0.004964934196323156, + 0.0037075721193104982, + 0.00751468213275075, + 0.005968115292489529, + 0.008172375150024891, + 0.0081963911652565, + 0.010998478159308434, + 0.009325958788394928, + 0.01168210618197918, + 0.007479499559849501, + 0.022586893290281296, + 0.006874220445752144, + 0.006610624957829714, + 0.006262437906116247, + 0.007430158089846373, + 0.006699030753225088, + 0.005839396268129349, + 0.00929658580571413, + 0.007186281494796276, + 0.0072133103385567665, + 0.0074362303130328655, + 0.007115681190043688, + 0.021525930613279343, + 0.005517452023923397, + 0.006301529239863157, + 0.014406212605535984, + 0.008213934488594532, + 0.003999775741249323, + 0.04240964725613594, + 0.05719252675771713, + 0.10106197744607925, + 0.05509957671165466, + 0.05004898086190224, + 0.08328527212142944, + 0.06098337098956108, + 0.031259696930646896, + 0.003362946677953005, + 0.004383367020636797, + 0.0054398891516029835, + 0.002859252505004406, + 0.004098725039511919, + 0.004545436706393957, + 0.005044276360422373, + 0.005866614170372486, + 0.003218151396140456, + 0.00484224408864975, + 0.003641839837655425, + 0.0026552339550107718, + 0.0027217965107411146, + 0.003332616528496146, + 0.0035270394291728735, + 0.0033319194335490465, + 0.004370018839836121, + 0.007436451502144337, + 0.005123029462993145, + 0.012307148426771164, + 0.0055472832173109055, + 0.006243606097996235, + 0.00595577247440815, + 0.004209007136523724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.8261731266975403 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly across the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's skin deathly pale and almost translucent with fiery red hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single harsh overhead light casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a stream of blood dripping from her right nostril and flowing from the corner of her mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014212576672434807 + ], + "opposite_logprob_tokens": [ + -4.264212608337402 + ], + "logprob_reward": -0.014212576672434807, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014212576672434807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.216652050614357, + 0.06979971379041672, + 0.05021797493100166, + 0.09833432734012604, + 0.17743901908397675, + 0.036594588309526443, + 0.21417254209518433, + 0.023665357381105423, + 0.057082489132881165, + 0.05604195594787598 + ], + "credit_map": [ + 0.019786929711699486, + 0.019816819578409195, + 0.0338779054582119, + 0.017708169296383858, + 0.02740326337516308, + 0.03148547559976578, + 0.029327871277928352, + 0.03724559023976326, + 0.0073998309671878815, + 0.00988641194999218, + 0.009640663862228394, + 0.007253774907439947, + 0.008123338222503662, + 0.008432172238826752, + 0.01400059461593628, + 0.005062925163656473, + 0.0047300118021667, + 0.007494241930544376, + 0.006842009257525206, + 0.004495860543102026, + 0.008847375400364399, + 0.005447502247989178, + 0.0076598613522946835, + 0.004701109603047371, + 0.009899619966745377, + 0.011477077379822731, + 0.01259884424507618, + 0.008938143029808998, + 0.016324810683727264, + 0.009991560131311417, + 0.019024722278118134, + 0.01007954590022564, + 0.017061706632375717, + 0.01618582382798195, + 0.02134387195110321, + 0.018023431301116943, + 0.02073279768228531, + 0.030002489686012268, + 0.031941864639520645, + 0.022147029638290405, + 0.003368122735992074, + 0.0034475435968488455, + 0.010337989777326584, + 0.002473822794854641, + 0.0033179018646478653, + 0.00711645744740963, + 0.003862111596390605, + 0.002670635934919119, + 0.018992962315678596, + 0.027683064341545105, + 0.02945614792406559, + 0.03661182150244713, + 0.026653775945305824, + 0.022323017939925194, + 0.035665325820446014, + 0.016786424443125725, + 0.0024295977782458067, + 0.002525950316339731, + 0.004248713608831167, + 0.002330826362594962, + 0.002671249210834503, + 0.0030660477932542562, + 0.0031706043519079685, + 0.003222366329282522, + 0.005608227103948593, + 0.00731669319793582, + 0.008627979084849358, + 0.005177563056349754, + 0.0068833064287900925, + 0.011502261273562908, + 0.007170968689024448, + 0.00479548517614603, + 0.005137667991220951, + 0.009076615795493126, + 0.005905458703637123, + 0.009490041993558407, + 0.006899215746670961, + 0.006408338900655508, + 0.008058857172727585, + 0.005065755918622017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.014212576672434807 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly across the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's skin deathly pale and almost translucent with fiery red hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single harsh overhead light casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a stream of blood dripping from her right nostril and flowing from the corner of her mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706831097602844 + ], + "opposite_logprob_tokens": [ + -2.127068281173706 + ], + "logprob_reward": -0.12706831097602844, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12706831097602844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16852878034114838, + 0.07477869093418121, + 0.06938132643699646, + 0.07148906588554382, + 0.1935373991727829, + 0.07622557133436203, + 0.07854722440242767, + 0.16910651326179504, + 0.057561520487070084, + 0.04084396734833717 + ], + "credit_map": [ + 0.018171800300478935, + 0.01481185108423233, + 0.02796054258942604, + 0.013245823793113232, + 0.01852395571768284, + 0.025053856894373894, + 0.023319698870182037, + 0.027441252022981644, + 0.008972764015197754, + 0.013333532027900219, + 0.012999781407415867, + 0.008231797255575657, + 0.007453154772520065, + 0.008456846699118614, + 0.00909371580928564, + 0.006237104535102844, + 0.010902095586061478, + 0.008859873749315739, + 0.010302645154297352, + 0.008439561352133751, + 0.007079174742102623, + 0.006654597818851471, + 0.009591097943484783, + 0.007552277762442827, + 0.0077989548444747925, + 0.009990544058382511, + 0.011044876649975777, + 0.006130416411906481, + 0.009729010052978992, + 0.008621281012892723, + 0.01122431643307209, + 0.006949662230908871, + 0.018965132534503937, + 0.020390383899211884, + 0.02571277879178524, + 0.012204261496663094, + 0.028063848614692688, + 0.03931278735399246, + 0.031843964010477066, + 0.01704423688352108, + 0.010584327392280102, + 0.00918656773865223, + 0.014762268401682377, + 0.006718072108924389, + 0.007690212223678827, + 0.010103815235197544, + 0.009700157679617405, + 0.007480149623006582, + 0.006466767750680447, + 0.008590406738221645, + 0.011921591125428677, + 0.012971237301826477, + 0.006702351849526167, + 0.013467893935739994, + 0.012227156199514866, + 0.006199820898473263, + 0.017375778406858444, + 0.021108638495206833, + 0.02544163167476654, + 0.01391417346894741, + 0.017741024494171143, + 0.04186490550637245, + 0.014015838503837585, + 0.017644530162215233, + 0.00819964800029993, + 0.008192679844796658, + 0.009532006457448006, + 0.00518671702593565, + 0.005040167830884457, + 0.008180870674550533, + 0.006479782052338123, + 0.006749643478542566, + 0.005059582646936178, + 0.005983980838209391, + 0.0057067666202783585, + 0.006768427789211273, + 0.003387532429769635, + 0.006067737936973572, + 0.004198045469820499, + 0.0036718903575092554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.12706831097602844 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly across the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's skin deathly pale and almost translucent with fiery red hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single harsh overhead light casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a stream of blood dripping from her right nostril and flowing from the corner of her mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058020606637001 + ], + "opposite_logprob_tokens": [ + -3.895580291748047 + ], + "logprob_reward": -0.02058020606637001, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058020606637001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1577266901731491, + 0.05090813338756561, + 0.037511471658945084, + 0.11332328617572784, + 0.16039717197418213, + 0.047680191695690155, + 0.24443106353282928, + 0.04091832414269447, + 0.0770636722445488, + 0.07003999501466751 + ], + "credit_map": [ + 0.017546381801366806, + 0.018788374960422516, + 0.026547910645604134, + 0.01589766889810562, + 0.01730136014521122, + 0.017831075936555862, + 0.02753744274377823, + 0.01627648063004017, + 0.0057367319241166115, + 0.005874823313206434, + 0.006195123307406902, + 0.007522398140281439, + 0.0044905636459589005, + 0.005330208223313093, + 0.010586192831397057, + 0.005172088276594877, + 0.004334714263677597, + 0.0053984420374035835, + 0.005300584249198437, + 0.003571179462596774, + 0.004227727651596069, + 0.004742015153169632, + 0.005505572073161602, + 0.004431233275681734, + 0.012123064137995243, + 0.011181997135281563, + 0.01617521420121193, + 0.012281039729714394, + 0.015242770314216614, + 0.010221781209111214, + 0.023870505392551422, + 0.012226908467710018, + 0.014516246505081654, + 0.014626814983785152, + 0.020006343722343445, + 0.013268311507999897, + 0.015693316236138344, + 0.022022400051355362, + 0.030506255105137825, + 0.029757492244243622, + 0.004237045533955097, + 0.006290887948125601, + 0.009376227855682373, + 0.004818374756723642, + 0.0040086726658046246, + 0.0068217371590435505, + 0.007807885762304068, + 0.0043193562887609005, + 0.02132522501051426, + 0.031877949833869934, + 0.032317083328962326, + 0.04281789809465408, + 0.031133035197854042, + 0.022890230640769005, + 0.0422833226621151, + 0.019786309450864792, + 0.003755912883207202, + 0.004003226291388273, + 0.007136007305234671, + 0.002995941089466214, + 0.003927200101315975, + 0.010071799159049988, + 0.00413177628070116, + 0.004896460101008415, + 0.006849450990557671, + 0.012109019793570042, + 0.007750133518129587, + 0.009860411286354065, + 0.008325213566422462, + 0.00955028086900711, + 0.016100557520985603, + 0.0065186056308448315, + 0.006446704734116793, + 0.010256187058985233, + 0.00775742344558239, + 0.017149580642580986, + 0.00738562410697341, + 0.006473702844232321, + 0.009297411888837814, + 0.005273355171084404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.02058020606637001 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up slowly pans across the face of a woman, her features a mask of pure terror amidst a dark and intense horror scene. Her skin is deathly pale, almost translucent, providing a stark canvas for the fiery red hair that cascades across her forehead and cheeks in a tousled, damp mass. The scene is illuminated by a single, harsh overhead light, casting deep, dramatic shadows that sculpt her features and enhance the eerie, claustrophobic atmosphere. Her wide-open eyes, a startlingly clear blue, gaze at an unseen point beyond the camera, reflecting a profound terror. A stream of thick, crimson blood steadily and uncomfortably drips from her right nostril, tracing a path along her philtrum, while another stream of fresh blood flows from the corner of her slightly open mouth, which hangs open as she gasps shallowly and raggedly. Her jaw trembles uncontrollably, and the entire composition relentlessly focuses on her anguished facial expression and the slow, terrifying descent of blood, creating a deeply unsettling and intensely terrifying visual narrative.", + "seed": 15071, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_5/index_2128_seed_15071_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.3833560635335743, + "vlm_logprob": -0.3833560635335743, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.3833560635335743, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly across the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's skin deathly pale and almost translucent with fiery red hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single harsh overhead light casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a stream of blood dripping from her right nostril and flowing from the corner of her mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00630958890542388 + ], + "opposite_logprob_tokens": [ + -6.881309509277344 + ], + "logprob_reward": -0.00630958890542388, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.00630958890542388, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17571856081485748, + 0.10304412245750427, + 0.2114679217338562, + 0.06529726833105087, + 0.08669599145650864, + 0.09417341649532318, + 0.09477008134126663, + 0.0667906254529953, + 0.052929800003767014, + 0.04911220446228981 + ], + "credit_map": [ + 0.014343755319714546, + 0.011500204913318157, + 0.01731574907898903, + 0.03835749998688698, + 0.013487862423062325, + 0.047561924904584885, + 0.015769505873322487, + 0.017382046207785606, + 0.005242308136075735, + 0.01075913105159998, + 0.007333383895456791, + 0.01723024621605873, + 0.0169379785656929, + 0.01568598859012127, + 0.021610375493764877, + 0.008244703523814678, + 0.013069621287286282, + 0.025586282834410667, + 0.017471544444561005, + 0.026123082265257835, + 0.032365456223487854, + 0.019115639850497246, + 0.051652003079652786, + 0.026084287092089653, + 0.005133462604135275, + 0.007226844318211079, + 0.008476078510284424, + 0.0067752948962152, + 0.011087493970990181, + 0.005831961054354906, + 0.013427414000034332, + 0.007338724099099636, + 0.006217632442712784, + 0.009571967646479607, + 0.008078151382505894, + 0.009628130123019218, + 0.01657298021018505, + 0.008391138166189194, + 0.016648458316922188, + 0.011587529443204403, + 0.005969332065433264, + 0.012512678280472755, + 0.01098833978176117, + 0.010626153089106083, + 0.013496358878910542, + 0.007805560249835253, + 0.02403285726904869, + 0.008742141537368298, + 0.004855509381741285, + 0.008324249647557735, + 0.012795135378837585, + 0.015634354203939438, + 0.009002750739455223, + 0.024910423904657364, + 0.009047665633261204, + 0.010199988260865211, + 0.003900837851688266, + 0.0043916297145187855, + 0.003877406707033515, + 0.007008266635239124, + 0.008274300023913383, + 0.015709443017840385, + 0.011922850273549557, + 0.011705885641276836, + 0.003153126919642091, + 0.0054628606885671616, + 0.004425492603331804, + 0.0066357748582959175, + 0.008250840939581394, + 0.005204192828387022, + 0.01528707705438137, + 0.004510435741394758, + 0.003136930288746953, + 0.0054100798442959785, + 0.005854550749063492, + 0.007583244238048792, + 0.0075075370259583, + 0.00583278713747859, + 0.00895595084875822, + 0.0048311250284314156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.00630958890542388 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly across the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's skin deathly pale and almost translucent with fiery red hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single harsh overhead light casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a stream of blood dripping from her right nostril and flowing from the corner of her mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.5020954608917236 + ], + "opposite_logprob_tokens": [ + -0.25209546089172363 + ], + "logprob_reward": -1.5020954608917236, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.5020954608917236, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.106593556702137, + 0.059157200157642365, + 0.12613503634929657, + 0.08235161751508713, + 0.06304589658975601, + 0.12685178220272064, + 0.26839256286621094, + 0.10050581395626068, + 0.0472298227250576, + 0.01973673887550831 + ], + "credit_map": [ + 0.011360698379576206, + 0.011351932771503925, + 0.013504121452569962, + 0.009612929075956345, + 0.01612425036728382, + 0.015881909057497978, + 0.015425613150000572, + 0.013332102447748184, + 0.005009020213037729, + 0.006933598313480616, + 0.0051344335079193115, + 0.006976862903684378, + 0.008060775697231293, + 0.00797852873802185, + 0.01146470382809639, + 0.0075992741622030735, + 0.008489145897328854, + 0.012683364562690258, + 0.010807568207383156, + 0.016320725902915, + 0.01609160751104355, + 0.018514836207032204, + 0.028357507660984993, + 0.014870285987854004, + 0.006103829015046358, + 0.009964882396161556, + 0.008725619874894619, + 0.007320086937397718, + 0.013625344261527061, + 0.0072752609848976135, + 0.021342061460018158, + 0.007994529791176319, + 0.004267302341759205, + 0.007442695088684559, + 0.007559861987829208, + 0.006737120449542999, + 0.011152672581374645, + 0.004758299328386784, + 0.015851160511374474, + 0.005276785697788, + 0.007832951843738556, + 0.012590175494551659, + 0.01063812430948019, + 0.013067361898720264, + 0.017178166657686234, + 0.014006865210831165, + 0.0339566208422184, + 0.0175815187394619, + 0.02339951880276203, + 0.027746766805648804, + 0.047325924038887024, + 0.026071446016430855, + 0.030657755210995674, + 0.04598623514175415, + 0.03257400542497635, + 0.0346309095621109, + 0.008105605840682983, + 0.009439704939723015, + 0.011022025719285011, + 0.015601558610796928, + 0.016708282753825188, + 0.01015510130673647, + 0.01913401484489441, + 0.010339515283703804, + 0.004182294011116028, + 0.00585776986554265, + 0.008177612908184528, + 0.004692587535828352, + 0.00511959008872509, + 0.007131684571504593, + 0.00548590486869216, + 0.006582377944141626, + 0.00140340905636549, + 0.002111717127263546, + 0.0017218445427715778, + 0.0026889590080827475, + 0.0030633839778602123, + 0.002294359030202031, + 0.004414967726916075, + 0.002038098406046629 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 2 + ], + "selected_reward": -1.5020954608917236 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly across the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's skin deathly pale and almost translucent with fiery red hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single harsh overhead light casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a stream of blood dripping from her right nostril and flowing from the corner of her mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012566078454256058 + ], + "opposite_logprob_tokens": [ + -4.387566089630127 + ], + "logprob_reward": -0.012566078454256058, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012566078454256058, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21161611378192902, + 0.19638368487358093, + 0.10048474371433258, + 0.061315279453992844, + 0.06331805884838104, + 0.11141347885131836, + 0.09173610061407089, + 0.0693877711892128, + 0.046227678656578064, + 0.048117052763700485 + ], + "credit_map": [ + 0.017827633768320084, + 0.0184137225151062, + 0.02385888807475567, + 0.03021782450377941, + 0.026492539793252945, + 0.02738168090581894, + 0.025191280990839005, + 0.04223255068063736, + 0.010409136302769184, + 0.018997346982359886, + 0.01453168224543333, + 0.02677021361887455, + 0.022204723209142685, + 0.045698776841163635, + 0.02848481386899948, + 0.02928699366748333, + 0.00789803359657526, + 0.013749350793659687, + 0.009537851437926292, + 0.012101998552680016, + 0.01439437922090292, + 0.010626694187521935, + 0.01943247765302658, + 0.01274397037923336, + 0.004934499505907297, + 0.006281829904764891, + 0.006234952248632908, + 0.005886131897568703, + 0.009698989801108837, + 0.005452829413115978, + 0.016694756224751472, + 0.006131293252110481, + 0.0049697342328727245, + 0.00792229175567627, + 0.006128051318228245, + 0.008427232503890991, + 0.01058382261544466, + 0.005575178191065788, + 0.010226474143564701, + 0.009485277347266674, + 0.007120272610336542, + 0.01187936495989561, + 0.012368646450340748, + 0.011627920903265476, + 0.013165744952857494, + 0.014508410356938839, + 0.02133321762084961, + 0.01940990798175335, + 0.007005802355706692, + 0.009790630079805851, + 0.013323976658284664, + 0.008286675438284874, + 0.017534488812088966, + 0.009290498681366444, + 0.0171895120292902, + 0.009314521215856075, + 0.003490913426503539, + 0.004581929184496403, + 0.0046105217188596725, + 0.008470403030514717, + 0.009862017817795277, + 0.013464801013469696, + 0.013483114540576935, + 0.011424075812101364, + 0.00312164518982172, + 0.004403704311698675, + 0.004196893889456987, + 0.00616300106048584, + 0.006463421043008566, + 0.005148680415004492, + 0.011874637566506863, + 0.004855695646256208, + 0.002555291401222348, + 0.004305734299123287, + 0.0031850826926529408, + 0.006327887065708637, + 0.007947481237351894, + 0.005604660138487816, + 0.011784142814576626, + 0.006406777538359165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.012566078454256058 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly across the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's skin deathly pale and almost translucent with fiery red hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single harsh overhead light casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a stream of blood dripping from her right nostril and flowing from the corner of her mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38716772198677063 + ], + "opposite_logprob_tokens": [ + -1.1371676921844482 + ], + "logprob_reward": -0.38716772198677063, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38716772198677063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1381494551897049, + 0.14073461294174194, + 0.15272264182567596, + 0.05338184908032417, + 0.22033312916755676, + 0.11570338904857635, + 0.09272217005491257, + 0.030600864440202713, + 0.02473309077322483, + 0.03091876208782196 + ], + "credit_map": [ + 0.014935360290110111, + 0.010662377811968327, + 0.016315840184688568, + 0.022641925141215324, + 0.013245039619505405, + 0.015880485996603966, + 0.015689769759774208, + 0.028778662905097008, + 0.010567069984972477, + 0.018631720915436745, + 0.021251363679766655, + 0.015150204300880432, + 0.02588798478245735, + 0.01882566511631012, + 0.01330909226089716, + 0.017111511901021004, + 0.0136437201872468, + 0.015342295169830322, + 0.01479090005159378, + 0.01649990677833557, + 0.02418413572013378, + 0.018067045137286186, + 0.026601888239383698, + 0.0235927551984787, + 0.004292222671210766, + 0.006410513538867235, + 0.008499173447489738, + 0.006051090080291033, + 0.006570085417479277, + 0.004832139704376459, + 0.009498809464275837, + 0.007227814756333828, + 0.01782563515007496, + 0.028437085449695587, + 0.030768776312470436, + 0.02707308530807495, + 0.03224736452102661, + 0.019854290410876274, + 0.03353288024663925, + 0.03059402108192444, + 0.007395374123007059, + 0.013405317440629005, + 0.014366118237376213, + 0.013596788980066776, + 0.016203945502638817, + 0.011926616542041302, + 0.024061447009444237, + 0.014747772365808487, + 0.009663183242082596, + 0.008603873662650585, + 0.014808489009737968, + 0.015409511514008045, + 0.008314779959619045, + 0.015986304730176926, + 0.008680375292897224, + 0.011255651712417603, + 0.0021943929605185986, + 0.0037359583657234907, + 0.0022898807656019926, + 0.003370336489751935, + 0.002912897150963545, + 0.00762527622282505, + 0.0025782340671867132, + 0.005893888883292675, + 0.0016290913335978985, + 0.0026199310086667538, + 0.0033077511470764875, + 0.0026773293502628803, + 0.0038184942677617073, + 0.0031085803639143705, + 0.0045019895769655704, + 0.003069923957809806, + 0.0016273960936814547, + 0.0026649958454072475, + 0.002446057740598917, + 0.004994300194084644, + 0.004867994226515293, + 0.003726587165147066, + 0.007394975516945124, + 0.003196453209966421 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.38716772198677063 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly across the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's skin deathly pale and almost translucent with fiery red hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single harsh overhead light casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a stream of blood dripping from her right nostril and flowing from the corner of her mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008641467429697514 + ], + "opposite_logprob_tokens": [ + -4.758641242980957 + ], + "logprob_reward": -0.008641467429697514, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008641467429697514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07369287312030792, + 0.0851617157459259, + 0.11217581480741501, + 0.032058633863925934, + 0.044980984181165695, + 0.2558254897594452, + 0.10169143229722977, + 0.0951296016573906, + 0.022230876609683037, + 0.17705252766609192 + ], + "credit_map": [ + 0.010187648236751556, + 0.00631899805739522, + 0.011364520527422428, + 0.007475558202713728, + 0.010435555130243301, + 0.010981950908899307, + 0.00999422650784254, + 0.006934415083378553, + 0.0040898933075368404, + 0.007920250296592712, + 0.00624811789020896, + 0.009565632790327072, + 0.01674947887659073, + 0.010459794662892818, + 0.021651150658726692, + 0.008477393537759781, + 0.0074601005762815475, + 0.013689305633306503, + 0.009805358946323395, + 0.013419054448604584, + 0.017218731343746185, + 0.010866482742130756, + 0.026212478056550026, + 0.013504303060472012, + 0.002762229647487402, + 0.004407181870192289, + 0.004208555445075035, + 0.0037316628731787205, + 0.0039050329942256212, + 0.003498282516375184, + 0.00514132808893919, + 0.004404359497129917, + 0.003455240512266755, + 0.005323206074535847, + 0.005465635564178228, + 0.004574066027998924, + 0.007422794587910175, + 0.003855515271425247, + 0.009376563131809235, + 0.005507967434823513, + 0.013412007130682468, + 0.021602779626846313, + 0.017981043085455894, + 0.027065817266702652, + 0.03667531907558441, + 0.01797005720436573, + 0.09214714914560318, + 0.02897130325436592, + 0.006706873420625925, + 0.008033238351345062, + 0.011667617596685886, + 0.011654069647192955, + 0.011451944708824158, + 0.026617512106895447, + 0.010865586809813976, + 0.01469458919018507, + 0.005059918388724327, + 0.005700668785721064, + 0.006004851311445236, + 0.01139836385846138, + 0.013031198643147945, + 0.01882539503276348, + 0.01885600946843624, + 0.016253188252449036, + 0.0014743563951924443, + 0.0018729991279542446, + 0.0026334526482969522, + 0.002333919983357191, + 0.003262850223109126, + 0.002511814935132861, + 0.005771940108388662, + 0.0023695440031588078, + 0.008529482409358025, + 0.019298117607831955, + 0.0107355872169137, + 0.025847746059298515, + 0.03517253324389458, + 0.019215604290366173, + 0.0467492900788784, + 0.01150416024029255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 2 + ], + "selected_reward": -0.008641467429697514 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up slowly pans across the face of a woman, her features a mask of pure terror amidst a dark and intense horror scene. Her skin is deathly pale, almost translucent, providing a stark canvas for the fiery red hair that cascades across her forehead and cheeks in a tousled, damp mass. The scene is illuminated by a single, harsh overhead light, casting deep, dramatic shadows that sculpt her features and enhance the eerie, claustrophobic atmosphere. Her wide-open eyes, a startlingly clear blue, gaze at an unseen point beyond the camera, reflecting a profound terror. A stream of thick, crimson blood steadily and uncomfortably drips from her right nostril, tracing a path along her philtrum, while another stream of fresh blood flows from the corner of her slightly open mouth, which hangs open as she gasps shallowly and raggedly. Her jaw trembles uncontrollably, and the entire composition relentlessly focuses on her anguished facial expression and the slow, terrifying descent of blood, creating a deeply unsettling and intensely terrifying visual narrative.", + "seed": 15072, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_5/index_2128_seed_15072_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.27539365775883196, + "vlm_logprob": -0.27539365775883196, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.27539365775883196, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly across the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's skin deathly pale and almost translucent with fiery red hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single harsh overhead light casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a stream of blood dripping from her right nostril and flowing from the corner of her mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006372370757162571 + ], + "opposite_logprob_tokens": [ + -6.381372451782227 + ], + "logprob_reward": -0.006372370757162571, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.006372370757162571, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11346829682588577, + 0.18506279587745667, + 0.0821964219212532, + 0.048823293298482895, + 0.09772185236215591, + 0.15117208659648895, + 0.06337045133113861, + 0.10109435021877289, + 0.04673238471150398, + 0.11035798490047455 + ], + "credit_map": [ + 0.01310364343225956, + 0.013895917683839798, + 0.014099973253905773, + 0.011429877020418644, + 0.015910739079117775, + 0.015297659672796726, + 0.015748566016554832, + 0.013981934636831284, + 0.01830657385289669, + 0.01881290040910244, + 0.01982211507856846, + 0.019616875797510147, + 0.03053716942667961, + 0.024381252005696297, + 0.033564839512109756, + 0.02002108097076416, + 0.011519116349518299, + 0.008411787450313568, + 0.011695735156536102, + 0.005556298885494471, + 0.013771764002740383, + 0.013467338867485523, + 0.008711472153663635, + 0.009062916971743107, + 0.0056360866874456406, + 0.00487771350890398, + 0.007976347580552101, + 0.005189212504774332, + 0.0050014229491353035, + 0.009707479737699032, + 0.004923476371914148, + 0.005511558149009943, + 0.007988495752215385, + 0.008584280498325825, + 0.01793384552001953, + 0.009290513582527637, + 0.015788588672876358, + 0.01887531578540802, + 0.009804001078009605, + 0.009456817992031574, + 0.012510583736002445, + 0.011052245274186134, + 0.023021535947918892, + 0.014389842748641968, + 0.03338238596916199, + 0.022790607064962387, + 0.020958615466952324, + 0.013066280633211136, + 0.008039706386625767, + 0.006625684909522533, + 0.007451040204614401, + 0.006908545270562172, + 0.008297116495668888, + 0.00878144707530737, + 0.007856110110878944, + 0.009410804137587547, + 0.010227426886558533, + 0.010212395340204239, + 0.014202190563082695, + 0.01253275852650404, + 0.00824398547410965, + 0.01697051338851452, + 0.009503461420536041, + 0.01920163817703724, + 0.005159722175449133, + 0.005091957747936249, + 0.005945455748587847, + 0.005382866133004427, + 0.0056974380277097225, + 0.006733400747179985, + 0.006093756761401892, + 0.0066277929581701756, + 0.010585436597466469, + 0.011122558265924454, + 0.012639994733035564, + 0.013469653204083443, + 0.010356683284044266, + 0.021327894181013107, + 0.009983494877815247, + 0.020872285589575768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.006372370757162571 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly across the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's skin deathly pale and almost translucent with fiery red hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single harsh overhead light casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a stream of blood dripping from her right nostril and flowing from the corner of her mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578274607658386 + ], + "opposite_logprob_tokens": [ + -0.6328274607658386 + ], + "logprob_reward": -0.7578274607658386, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578274607658386, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1709628403186798, + 0.09190520644187927, + 0.10596787184476852, + 0.06771894544363022, + 0.06400183588266373, + 0.08903343230485916, + 0.05345204845070839, + 0.06579580903053284, + 0.05559888109564781, + 0.23556314408779144 + ], + "credit_map": [ + 0.020119506865739822, + 0.023855706676840782, + 0.017722947522997856, + 0.02215459942817688, + 0.01820802129805088, + 0.01943039335310459, + 0.030207999050617218, + 0.019263671711087227, + 0.011151503771543503, + 0.007252187933772802, + 0.01449718326330185, + 0.01033586636185646, + 0.008948003873229027, + 0.016645420342683792, + 0.010797714814543724, + 0.012277323752641678, + 0.016544608399271965, + 0.00893728993833065, + 0.01624124124646187, + 0.009301562793552876, + 0.01806367002427578, + 0.014188949950039387, + 0.01348857395350933, + 0.009201974608004093, + 0.005473368801176548, + 0.007285051513463259, + 0.008112089708447456, + 0.005883699283003807, + 0.00862803217023611, + 0.009423967450857162, + 0.01143118366599083, + 0.01148154865950346, + 0.004759570118039846, + 0.004474008921533823, + 0.012689726427197456, + 0.005234681535512209, + 0.007653326727449894, + 0.015422308817505836, + 0.006243486888706684, + 0.007524731568992138, + 0.013576781377196312, + 0.007780265994369984, + 0.018836230039596558, + 0.01065931748598814, + 0.00899783056229353, + 0.012568049132823944, + 0.009484987705945969, + 0.007129972334951162, + 0.0057372888550162315, + 0.006150667555630207, + 0.00568404933437705, + 0.004389119800180197, + 0.006775443907827139, + 0.007657626178115606, + 0.007882459089159966, + 0.009175390005111694, + 0.008556031621992588, + 0.008203020319342613, + 0.00885655079036951, + 0.00818589422851801, + 0.006764831952750683, + 0.009538205340504646, + 0.006226366851478815, + 0.009464899078011513, + 0.006654097232967615, + 0.004776336718350649, + 0.01022142730653286, + 0.0045118448324501514, + 0.00570351118221879, + 0.0104824248701334, + 0.007301585283130407, + 0.005947652272880077, + 0.02578994445502758, + 0.02260644920170307, + 0.022367509081959724, + 0.023489343002438545, + 0.024421097710728645, + 0.03856438398361206, + 0.022652028128504753, + 0.055672381073236465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.7578274607658386 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly across the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's skin deathly pale and almost translucent with fiery red hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single harsh overhead light casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a stream of blood dripping from her right nostril and flowing from the corner of her mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006750398315489292 + ], + "opposite_logprob_tokens": [ + -5.006750583648682 + ], + "logprob_reward": -0.006750398315489292, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006750398315489292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1378430873155594, + 0.0753987580537796, + 0.12137611210346222, + 0.035923879593610764, + 0.07635781913995743, + 0.40581244230270386, + 0.013670118525624275, + 0.06461458653211594, + 0.02140953578054905, + 0.04759373143315315 + ], + "credit_map": [ + 0.016463415697216988, + 0.015818219631910324, + 0.015369023196399212, + 0.013194454833865166, + 0.017705600708723068, + 0.017557404935359955, + 0.02402633987367153, + 0.017708614468574524, + 0.006429111585021019, + 0.007461179047822952, + 0.008454898372292519, + 0.006745853926986456, + 0.012452860362827778, + 0.01029936969280243, + 0.012746338732540607, + 0.010809134691953659, + 0.007793539669364691, + 0.01358116790652275, + 0.01011327188462019, + 0.009396234527230263, + 0.03073643334209919, + 0.01551208458840847, + 0.018486477434635162, + 0.015756886452436447, + 0.0038417407777160406, + 0.004079171922057867, + 0.005078814923763275, + 0.004505734890699387, + 0.00354140717536211, + 0.004967282526195049, + 0.005642157979309559, + 0.004267565440386534, + 0.005508072208613157, + 0.004509238991886377, + 0.023309627547860146, + 0.004161793738603592, + 0.004745017271488905, + 0.02318212203681469, + 0.005590071901679039, + 0.0053518665954470634, + 0.056917302310466766, + 0.034565214067697525, + 0.07949113845825195, + 0.04277468100190163, + 0.05708632990717888, + 0.060359708964824677, + 0.0407613106071949, + 0.03385668620467186, + 0.0017477840883657336, + 0.001822265679948032, + 0.0019878409802913666, + 0.0013976104091852903, + 0.0014812833396717906, + 0.0019006493967026472, + 0.0016729364870116115, + 0.0016597456997260451, + 0.008400463499128819, + 0.0070464638993144035, + 0.008725986815989017, + 0.005779065191745758, + 0.009032477624714375, + 0.008888881653547287, + 0.008004678413271904, + 0.00873656291514635, + 0.0022695313673466444, + 0.002555029932409525, + 0.0024644092191010714, + 0.0022271587513387203, + 0.0032924313563853502, + 0.002744214376434684, + 0.0029930041637271643, + 0.002863755915313959, + 0.005273801274597645, + 0.004222388379275799, + 0.004341093357652426, + 0.0035773662384599447, + 0.0034931853879243135, + 0.010052491910755634, + 0.0033242141362279654, + 0.013309184461832047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.006750398315489292 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly across the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's skin deathly pale and almost translucent with fiery red hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single harsh overhead light casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a stream of blood dripping from her right nostril and flowing from the corner of her mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.576217770576477 + ], + "opposite_logprob_tokens": [ + -0.826217770576477 + ], + "logprob_reward": -0.576217770576477, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.576217770576477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14954151213169098, + 0.0746537446975708, + 0.0963781327009201, + 0.04845708608627319, + 0.07272452861070633, + 0.09404695779085159, + 0.04598892852663994, + 0.24796979129314423, + 0.035825904458761215, + 0.13441336154937744 + ], + "credit_map": [ + 0.013697179965674877, + 0.020694933831691742, + 0.016536153852939606, + 0.014666539616882801, + 0.022162578999996185, + 0.01862446777522564, + 0.021661775186657906, + 0.021497881039977074, + 0.008818450383841991, + 0.00869695469737053, + 0.007147865369915962, + 0.007625713013112545, + 0.010366866365075111, + 0.009502277709543705, + 0.013737566769123077, + 0.008758049458265305, + 0.009636087343096733, + 0.008857902139425278, + 0.012671430595219135, + 0.007173216436058283, + 0.016976021230220795, + 0.017276324331760406, + 0.013791963458061218, + 0.009995187632739544, + 0.006563170813024044, + 0.004638297017663717, + 0.007187692914158106, + 0.00452195992693305, + 0.0059483908116817474, + 0.006786664482206106, + 0.006292147096246481, + 0.0065187616273760796, + 0.007019040174782276, + 0.004159288946539164, + 0.021319974213838577, + 0.004661488812416792, + 0.006562128197401762, + 0.019171740859746933, + 0.004516796674579382, + 0.005314072128385305, + 0.013271566480398178, + 0.008920776657760143, + 0.016736024990677834, + 0.010050507262349129, + 0.014684949070215225, + 0.012415032833814621, + 0.01004050113260746, + 0.007927601225674152, + 0.004909909795969725, + 0.004202228505164385, + 0.0074468133971095085, + 0.00466188695281744, + 0.004431463312357664, + 0.008145644329488277, + 0.005285922437906265, + 0.00690506212413311, + 0.03957934305071831, + 0.032515328377485275, + 0.03605422377586365, + 0.022158607840538025, + 0.020637385547161102, + 0.029677579179406166, + 0.026386858895421028, + 0.04096045717597008, + 0.005080046597868204, + 0.0030782476533204317, + 0.00815178919583559, + 0.003301737830042839, + 0.003411956364288926, + 0.006366690620779991, + 0.0036051536444574594, + 0.0028302776627242565, + 0.018361946567893028, + 0.015704162418842316, + 0.01903568208217621, + 0.012449276633560658, + 0.011193737387657166, + 0.02149064838886261, + 0.01375025324523449, + 0.022427663207054138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.576217770576477 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly across the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's skin deathly pale and almost translucent with fiery red hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single harsh overhead light casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a stream of blood dripping from her right nostril and flowing from the corner of her mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029800288379192352 + ], + "opposite_logprob_tokens": [ + -3.5298001766204834 + ], + "logprob_reward": -0.029800288379192352, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029800288379192352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.270197331905365, + 0.08174905180931091, + 0.04144075885415077, + 0.03085794672369957, + 0.09702931344509125, + 0.25547268986701965, + 0.012464137747883797, + 0.0980437621474266, + 0.043051864951848984, + 0.0696931704878807 + ], + "credit_map": [ + 0.02575688436627388, + 0.019407063722610474, + 0.025257322937250137, + 0.022290006279945374, + 0.03789875656366348, + 0.03458328917622566, + 0.07610839605331421, + 0.028895599767565727, + 0.008932514116168022, + 0.008159814402461052, + 0.008101675659418106, + 0.007982676848769188, + 0.011616370640695095, + 0.011276350356638432, + 0.01621883362531662, + 0.009460813365876675, + 0.004845594987273216, + 0.003624262288212776, + 0.005746062379330397, + 0.003223937004804611, + 0.006080951076000929, + 0.008878365159034729, + 0.004548031836748123, + 0.004493554122745991, + 0.0036491744685918093, + 0.0032036923803389072, + 0.004881813190877438, + 0.0033764508552849293, + 0.0031974969897419214, + 0.003958749584853649, + 0.004541043192148209, + 0.004049524664878845, + 0.006311520468443632, + 0.00652909092605114, + 0.024771658703684807, + 0.008200122974812984, + 0.010236202739179134, + 0.025550778955221176, + 0.008246161043643951, + 0.007183771580457687, + 0.020291419699788094, + 0.01965194381773472, + 0.027962299063801765, + 0.029469942674040794, + 0.04272076487541199, + 0.03167840093374252, + 0.05359131097793579, + 0.030106602236628532, + 0.001519722631201148, + 0.0016496815951541066, + 0.001631875173188746, + 0.0012147548841312528, + 0.001220670877955854, + 0.001528722932562232, + 0.0017897060606628656, + 0.0019090035930275917, + 0.00948073249310255, + 0.010336735285818577, + 0.013013753108680248, + 0.010836636647582054, + 0.01596151851117611, + 0.014279932714998722, + 0.010953119024634361, + 0.01318132970482111, + 0.0036292304284870625, + 0.002728752326220274, + 0.0052615064196288586, + 0.003412719815969467, + 0.005645143799483776, + 0.006936573889106512, + 0.0098680779337883, + 0.005569859873503447, + 0.007007819600403309, + 0.006098806392401457, + 0.00796930119395256, + 0.007172682788223028, + 0.006404495798051357, + 0.012394415214657784, + 0.005919002927839756, + 0.016726642847061157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.029800288379192352 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_6/index_tensor([3958, 3958])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_6/index_tensor([3958, 3958])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6ba1e6a868bed38496c8d4d36c73b7d763457d5e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_6/index_tensor([3958, 3958])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot at eye level, with the camera moving in with an almost imperceptible slowness, captures a composition of exquisite elegance. To the left, a tall, sturdy bottle of smooth, dark glass, its classic shape almost black, rests on a polished, dark wood surface. The bottle's elegant, unpretentious paper label catches the soft light, revealing subtle, shifting reflections on its matte texture. To its right, a clear crystal goblet, half-filled with rich, crimson liquid, emanates a luscious, ruby-like translucency. A warm light source on the right sculpts the scene, creating brilliant highlights along the curve of the glass and casting long, soft shadows across the bottle, artfully creating a striking contrast between the bottle's profound opacity and the glass's vibrant transparency. The background blends into a warm, out-of-focus bokeh, enhancing this simple yet harmonious pairing and immersing the entire image in an atmosphere of quiet refinement and contemplative tranquility.", + "seed": 1158, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_6/index_3958_seed_1158_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.025445808481890708, + "vlm_logprob": -0.025445808481890708, + "vlm_yes_no_margin": 5.025, + "vlm_token_credit": -0.025445808481890708, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tall, sturdy bottle made of smooth, dark glass (almost black) on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a clear crystal goblet half-filled with rich, crimson liquid on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, out-of-focus bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in with an almost imperceptible slowness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0844912976026535 + ], + "opposite_logprob_tokens": [ + -3.58449125289917 + ], + "logprob_reward": -0.0844912976026535, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0844912976026535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05957411229610443, + 0.08805868774652481, + 0.1504574716091156, + 0.04369259998202324, + 0.09716574102640152, + 0.20805814862251282, + 0.10143053531646729, + 0.1346212923526764, + 0.048862677067518234, + 0.06807875633239746 + ], + "credit_map": [ + 0.008955905213952065, + 0.010215219110250473, + 0.007032691966742277, + 0.006546678487211466, + 0.00835421122610569, + 0.006373728159815073, + 0.006819522939622402, + 0.005276160314679146, + 0.012686848640441895, + 0.008973736315965652, + 0.00913762953132391, + 0.004981440491974354, + 0.011194795370101929, + 0.01337849348783493, + 0.01459541730582714, + 0.01311032846570015, + 0.0381968654692173, + 0.01401717308908701, + 0.04646553844213486, + 0.0073965913616120815, + 0.01771477796137333, + 0.009636144153773785, + 0.01058176439255476, + 0.006448610220104456, + 0.0058781057596206665, + 0.007018580101430416, + 0.0050431289710104465, + 0.005383104085922241, + 0.004405324812978506, + 0.003727457718923688, + 0.008851668797433376, + 0.0033852295018732548, + 0.012136412784457207, + 0.011931846849620342, + 0.008288267999887466, + 0.007954227738082409, + 0.010955838486552238, + 0.01713453233242035, + 0.01816217415034771, + 0.0106024369597435, + 0.0506899431347847, + 0.021226637065410614, + 0.03442677482962608, + 0.010537618771195412, + 0.04425374045968056, + 0.016553394496440887, + 0.020978471264243126, + 0.009391557425260544, + 0.017068248242139816, + 0.012222422286868095, + 0.02414516918361187, + 0.01083349995315075, + 0.014100012369453907, + 0.0091946916654706, + 0.007124849129468203, + 0.006741641089320183, + 0.020886307582259178, + 0.017259903252124786, + 0.03662094846367836, + 0.013251842930912971, + 0.01640806533396244, + 0.012872487306594849, + 0.010193370282649994, + 0.007128366734832525, + 0.007730438373982906, + 0.005465621594339609, + 0.011378680355846882, + 0.008743132464587688, + 0.004709291737526655, + 0.0032675787806510925, + 0.0041084750555455685, + 0.0034594552125781775, + 0.0084280576556921, + 0.007799198385328054, + 0.013716937974095345, + 0.00669877091422677, + 0.007923346012830734, + 0.007187402341514826, + 0.009714443236589432, + 0.006610607262700796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 7 + ], + "selected_reward": -0.0844912976026535 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tall, sturdy bottle made of smooth, dark glass (almost black) on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a clear crystal goblet half-filled with rich, crimson liquid on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, out-of-focus bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in with an almost imperceptible slowness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003654469968751073 + ], + "opposite_logprob_tokens": [ + -5.628654479980469 + ], + "logprob_reward": -0.003654469968751073, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003654469968751073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0819549411535263, + 0.08129379898309708, + 0.19923904538154602, + 0.10028880089521408, + 0.12241793423891068, + 0.08088070154190063, + 0.06492168456315994, + 0.09672083705663681, + 0.10080449283123016, + 0.0714777484536171 + ], + "credit_map": [ + 0.012596141546964645, + 0.0209423266351223, + 0.00734860310330987, + 0.005042442586272955, + 0.0096633555367589, + 0.010977189056575298, + 0.007755179423838854, + 0.007629704661667347, + 0.011968760751187801, + 0.011272904463112354, + 0.008407403714954853, + 0.007676415145397186, + 0.013932690024375916, + 0.010043504647910595, + 0.009983968921005726, + 0.00800815038383007, + 0.054882921278476715, + 0.019730517640709877, + 0.06275321543216705, + 0.010794770903885365, + 0.019957086071372032, + 0.011010206304490566, + 0.011634379625320435, + 0.008475948125123978, + 0.012301262468099594, + 0.014575870707631111, + 0.010158215649425983, + 0.011027521453797817, + 0.010236386209726334, + 0.011280544102191925, + 0.022311190143227577, + 0.008397810161113739, + 0.016889743506908417, + 0.022867385298013687, + 0.014675518497824669, + 0.013322506099939346, + 0.013292734511196613, + 0.013965492136776447, + 0.018539074808359146, + 0.008865478448569775, + 0.016803892329335213, + 0.010624479502439499, + 0.014474968425929546, + 0.004354128614068031, + 0.010304578579962254, + 0.010788828134536743, + 0.008014731109142303, + 0.00551509577780962, + 0.010799871757626534, + 0.009685276076197624, + 0.008819147944450378, + 0.004346485715359449, + 0.013580983504652977, + 0.00529647758230567, + 0.006976177450269461, + 0.005417269188910723, + 0.015251858159899712, + 0.014910105615854263, + 0.015590116381645203, + 0.009659851901233196, + 0.013555887155234814, + 0.010327714495360851, + 0.008817204274237156, + 0.00860810000449419, + 0.01520176324993372, + 0.011944461613893509, + 0.036492813378572464, + 0.010090349242091179, + 0.013583173975348473, + 0.005074250046163797, + 0.004398928955197334, + 0.004018755629658699, + 0.008870379999279976, + 0.011446364223957062, + 0.006257310975342989, + 0.004156738985329866, + 0.016539335250854492, + 0.01121573057025671, + 0.006487169768661261, + 0.006504711229354143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 8 + ], + "selected_reward": -0.003654469968751073 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tall, sturdy bottle made of smooth, dark glass (almost black) on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a clear crystal goblet half-filled with rich, crimson liquid on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, out-of-focus bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in with an almost imperceptible slowness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007294139941222966 + ], + "opposite_logprob_tokens": [ + -7.250729560852051 + ], + "logprob_reward": -0.0007294139941222966, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007294139941222966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10331431776285172, + 0.09108167141675949, + 0.2068772166967392, + 0.1295280158519745, + 0.04688672348856926, + 0.15810276567935944, + 0.034532494843006134, + 0.09395388513803482, + 0.05060824379324913, + 0.08511462807655334 + ], + "credit_map": [ + 0.016252735629677773, + 0.015067175030708313, + 0.011150912381708622, + 0.009535033255815506, + 0.016483841463923454, + 0.010891099460422993, + 0.01286030001938343, + 0.011073223315179348, + 0.015154123306274414, + 0.009848088957369328, + 0.010007685050368309, + 0.007856808602809906, + 0.015057618729770184, + 0.011986861936748028, + 0.011916961520910263, + 0.00925353728234768, + 0.05771957337856293, + 0.01755492389202118, + 0.06747391819953918, + 0.010307539254426956, + 0.02216893434524536, + 0.01361935306340456, + 0.01142670027911663, + 0.00660629291087389, + 0.019968941807746887, + 0.015623834915459156, + 0.010034671984612942, + 0.014737529680132866, + 0.011438168585300446, + 0.011630047112703323, + 0.033645786345005035, + 0.012449045665562153, + 0.006728651933372021, + 0.006324969232082367, + 0.004420612007379532, + 0.004056939855217934, + 0.005135225597769022, + 0.006600652355700731, + 0.008256746456027031, + 0.005362928379327059, + 0.03620495647192001, + 0.013679975643754005, + 0.018742531538009644, + 0.007283930201083422, + 0.038749706000089645, + 0.016624512150883675, + 0.018215730786323547, + 0.008601426146924496, + 0.005348002538084984, + 0.003605449805036187, + 0.006814958062022924, + 0.0030526791233569384, + 0.006197218783199787, + 0.0029940453823655844, + 0.003925603348761797, + 0.002594539662823081, + 0.01712481677532196, + 0.01067754440009594, + 0.03336364030838013, + 0.0068095047026872635, + 0.009262948296964169, + 0.0062722098082304, + 0.004392220173031092, + 0.006051006261259317, + 0.007550021167844534, + 0.005687998607754707, + 0.016242505982518196, + 0.006235282402485609, + 0.005913625005632639, + 0.0026721428148448467, + 0.0031466607470065355, + 0.0031600112561136484, + 0.011846665292978287, + 0.008096850477159023, + 0.020201435312628746, + 0.009043018333613873, + 0.010198481380939484, + 0.009077421389520168, + 0.010480141267180443, + 0.006170624867081642 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 3 + ], + "selected_reward": -0.0007294139941222966 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tall, sturdy bottle made of smooth, dark glass (almost black) on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a clear crystal goblet half-filled with rich, crimson liquid on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, out-of-focus bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in with an almost imperceptible slowness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004645389039069414 + ], + "opposite_logprob_tokens": [ + -5.379645347595215 + ], + "logprob_reward": -0.004645389039069414, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004645389039069414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04111826419830322, + 0.03293489292263985, + 0.7993664145469666, + 0.01985233649611473, + 0.00810454972088337, + 0.029699984937906265, + 0.018923770636320114, + 0.023669326677918434, + 0.011515398509800434, + 0.01481514610350132 + ], + "credit_map": [ + 0.00615179818123579, + 0.010207503102719784, + 0.003994572442024946, + 0.002849461743608117, + 0.0050517902709543705, + 0.0060925730504095554, + 0.0035443336237221956, + 0.0032262280583381653, + 0.004641343839466572, + 0.004339022561907768, + 0.0034786402247846127, + 0.002183322561904788, + 0.00661492021754384, + 0.004387094173580408, + 0.004597713239490986, + 0.0026928323786705732, + 0.21949432790279388, + 0.06673803180456161, + 0.2661351263523102, + 0.039612237364053726, + 0.09110232442617416, + 0.04109569266438484, + 0.04943260923027992, + 0.025755982846021652, + 0.0025319077540189028, + 0.0026888519059866667, + 0.0022842194885015488, + 0.002875089645385742, + 0.002151068300008774, + 0.0018669620621949434, + 0.003911927342414856, + 0.0015423099976032972, + 0.0010773692047223449, + 0.0008767890394665301, + 0.0008400673978030682, + 0.0006385697633959353, + 0.000937262608204037, + 0.00172270683106035, + 0.0012212813599035144, + 0.0007905028760433197, + 0.0073005761951208115, + 0.003065291792154312, + 0.004943239968270063, + 0.0014671130338683724, + 0.005986474454402924, + 0.00230988091789186, + 0.0031149734277278185, + 0.0015124335186555982, + 0.0028933375142514706, + 0.0023109489120543003, + 0.0037399630527943373, + 0.0018715057522058487, + 0.00326296198181808, + 0.0018126069335266948, + 0.0016546695260331035, + 0.0013777747517451644, + 0.004218664485961199, + 0.002545528579503298, + 0.007846360094845295, + 0.001893696840852499, + 0.002705034799873829, + 0.0018511541420593858, + 0.001192436320707202, + 0.0014164496678858995, + 0.0016082257498055696, + 0.001382231479510665, + 0.003925125114619732, + 0.001458453480154276, + 0.001452272292226553, + 0.0006038851570338011, + 0.0005615052068606019, + 0.0005236989818513393, + 0.001738241408020258, + 0.0017124289879575372, + 0.0014522421406581998, + 0.0010676226811483502, + 0.002910800976678729, + 0.0024408726021647453, + 0.0019232798367738724, + 0.0015696557238698006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.004645389039069414 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tall, sturdy bottle made of smooth, dark glass (almost black) on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a clear crystal goblet half-filled with rich, crimson liquid on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, out-of-focus bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in with an almost imperceptible slowness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033708471804857254 + ], + "opposite_logprob_tokens": [ + -3.4087085723876953 + ], + "logprob_reward": -0.033708471804857254, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033708471804857254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09334591031074524, + 0.08657706528902054, + 0.09770365804433823, + 0.032196447253227234, + 0.0665980651974678, + 0.08120059221982956, + 0.06643138080835342, + 0.05572370067238808, + 0.10277130454778671, + 0.3174518644809723 + ], + "credit_map": [ + 0.01687936671078205, + 0.01429335493594408, + 0.010557274334132671, + 0.005483175162225962, + 0.01444273255765438, + 0.014062866568565369, + 0.010451034642755985, + 0.007176112849265337, + 0.012254234403371811, + 0.011656559072434902, + 0.009718106128275394, + 0.005584124941378832, + 0.013646972365677357, + 0.013375313021242619, + 0.011819073930382729, + 0.008522680029273033, + 0.02204909175634384, + 0.010810778476297855, + 0.02530977502465248, + 0.005851739086210728, + 0.01377298217266798, + 0.00796310044825077, + 0.007228195201605558, + 0.0047179944813251495, + 0.004764277022331953, + 0.00428648991510272, + 0.0033719863276928663, + 0.00418964633718133, + 0.003400480840355158, + 0.0033222977072000504, + 0.006124195642769337, + 0.0027370741590857506, + 0.009106644429266453, + 0.010775174014270306, + 0.008418270386755466, + 0.005957470741122961, + 0.009607362560927868, + 0.00846752617508173, + 0.00996373500674963, + 0.004301890265196562, + 0.015704508870840073, + 0.00796109065413475, + 0.010318128392100334, + 0.006318933796137571, + 0.01725698821246624, + 0.01038715522736311, + 0.008557042106986046, + 0.0046967449598014355, + 0.012150523252785206, + 0.011475619859993458, + 0.011932323686778545, + 0.0067784134298563, + 0.009208261966705322, + 0.005092659965157509, + 0.005032120272517204, + 0.004761457908898592, + 0.011356326751410961, + 0.00579806650057435, + 0.019937464967370033, + 0.004041660577058792, + 0.005693281069397926, + 0.0033834544010460377, + 0.0027547599747776985, + 0.0027586868964135647, + 0.014673920348286629, + 0.011316809803247452, + 0.034447792917490005, + 0.01204821839928627, + 0.01340543758124113, + 0.005039695184677839, + 0.006587055511772633, + 0.005252384580671787, + 0.04082903638482094, + 0.030400004237890244, + 0.054392796009778976, + 0.029772993177175522, + 0.03997962176799774, + 0.05078553780913353, + 0.04327283427119255, + 0.028019027784466743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.033708471804857254 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot at eye level, with the camera moving in with an almost imperceptible slowness, captures a composition of exquisite elegance. To the left, a tall, sturdy bottle of smooth, dark glass, its classic shape almost black, rests on a polished, dark wood surface. The bottle's elegant, unpretentious paper label catches the soft light, revealing subtle, shifting reflections on its matte texture. To its right, a clear crystal goblet, half-filled with rich, crimson liquid, emanates a luscious, ruby-like translucency. A warm light source on the right sculpts the scene, creating brilliant highlights along the curve of the glass and casting long, soft shadows across the bottle, artfully creating a striking contrast between the bottle's profound opacity and the glass's vibrant transparency. The background blends into a warm, out-of-focus bokeh, enhancing this simple yet harmonious pairing and immersing the entire image in an atmosphere of quiet refinement and contemplative tranquility.", + "seed": 1159, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_6/index_3958_seed_1159_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03474679160863161, + "vlm_logprob": -0.03474679160863161, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.03474679160863161, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tall, sturdy bottle made of smooth, dark glass (almost black) on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a clear crystal goblet half-filled with rich, crimson liquid on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, out-of-focus bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in with an almost imperceptible slowness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09605303406715393 + ], + "opposite_logprob_tokens": [ + -2.721053123474121 + ], + "logprob_reward": -0.09605303406715393, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.09605303406715393, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1663101315498352, + 0.14087331295013428, + 0.05411006510257721, + 0.030082011595368385, + 0.03357701748609543, + 0.2721112370491028, + 0.16112081706523895, + 0.04527578875422478, + 0.07534687221050262, + 0.02119278907775879 + ], + "credit_map": [ + 0.014982427470386028, + 0.007195575628429651, + 0.025797734037041664, + 0.020989907905459404, + 0.03133954107761383, + 0.015903977677226067, + 0.03115115500986576, + 0.01894979737699032, + 0.007847984321415424, + 0.008714579045772552, + 0.015106791630387306, + 0.017043104395270348, + 0.01712464541196823, + 0.03128533810377121, + 0.028111634775996208, + 0.015639236196875572, + 0.003246918087825179, + 0.006683266721665859, + 0.005666652694344521, + 0.011135589331388474, + 0.009653374552726746, + 0.00383653212338686, + 0.010235621593892574, + 0.0036521118599921465, + 0.002331194467842579, + 0.0039072404615581036, + 0.004244506824761629, + 0.004276840016245842, + 0.004254648461937904, + 0.0034056948497891426, + 0.00505614560097456, + 0.002605740213766694, + 0.0026707318611443043, + 0.005975065752863884, + 0.0041367835365235806, + 0.0060231732204556465, + 0.0023558184038847685, + 0.0037879152223467827, + 0.004722400568425655, + 0.003905129386112094, + 0.012905292212963104, + 0.02058016136288643, + 0.10830468684434891, + 0.09439627826213837, + 0.007598786149173975, + 0.00872757751494646, + 0.014116091653704643, + 0.005482358857989311, + 0.01097922958433628, + 0.030889000743627548, + 0.028246795758605003, + 0.04169847071170807, + 0.00792006030678749, + 0.009516149759292603, + 0.017895132303237915, + 0.013975975103676319, + 0.0032739979214966297, + 0.0076300883665680885, + 0.008884144015610218, + 0.008828922174870968, + 0.0034903225023299456, + 0.002333413576707244, + 0.006904131732881069, + 0.003930766135454178, + 0.01248371135443449, + 0.009071582928299904, + 0.014823440462350845, + 0.013299593701958656, + 0.007894483394920826, + 0.006497223395854235, + 0.006392538547515869, + 0.004884297959506512, + 0.002731837099418044, + 0.0020056322682648897, + 0.003823100123554468, + 0.0036207265220582485, + 0.002803794341161847, + 0.002431006170809269, + 0.002020840300247073, + 0.001755854464136064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.09605303406715393 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tall, sturdy bottle made of smooth, dark glass (almost black) on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a clear crystal goblet half-filled with rich, crimson liquid on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, out-of-focus bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in with an almost imperceptible slowness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005293875467032194 + ], + "opposite_logprob_tokens": [ + -5.255293846130371 + ], + "logprob_reward": -0.005293875467032194, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005293875467032194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11442453414201736, + 0.35994139313697815, + 0.09889404475688934, + 0.07402163743972778, + 0.035678938031196594, + 0.08043553680181503, + 0.09935282170772552, + 0.05221040919423103, + 0.03758978843688965, + 0.04745091497898102 + ], + "credit_map": [ + 0.015931518748402596, + 0.008820591494441032, + 0.016161466017365456, + 0.01851690746843815, + 0.012798958458006382, + 0.011497125029563904, + 0.013441336341202259, + 0.017256634309887886, + 0.02114158309996128, + 0.023579392582178116, + 0.03711787983775139, + 0.0341363288462162, + 0.05076432228088379, + 0.09748036414384842, + 0.05121253803372383, + 0.044508982449769974, + 0.005556502845138311, + 0.010311582125723362, + 0.008893345482647419, + 0.03795626387000084, + 0.005738870706409216, + 0.0057080769911408424, + 0.01966993324458599, + 0.005059472285211086, + 0.005324790254235268, + 0.007630249485373497, + 0.011609171517193317, + 0.011775057762861252, + 0.005823276937007904, + 0.009113091975450516, + 0.013416726142168045, + 0.009329274296760559, + 0.0033484427258372307, + 0.004707900807261467, + 0.0065009682439267635, + 0.005307050887495279, + 0.002744929166510701, + 0.004672240931540728, + 0.005251466296613216, + 0.003145939204841852, + 0.007337205111980438, + 0.011301129125058651, + 0.02402827888727188, + 0.01574939489364624, + 0.004223153926432133, + 0.0030221811030060053, + 0.009849056601524353, + 0.004925142973661423, + 0.0062219444662332535, + 0.01788375899195671, + 0.02165602706372738, + 0.029212869703769684, + 0.004868312273174524, + 0.0037803035229444504, + 0.010378904640674591, + 0.0053507001139223576, + 0.005213103257119656, + 0.0066326274536550045, + 0.010069181211292744, + 0.008762316778302193, + 0.004077699966728687, + 0.004665070213377476, + 0.008238558657467365, + 0.004551849327981472, + 0.006468582432717085, + 0.005786635912954807, + 0.005880964919924736, + 0.005333909299224615, + 0.0035062741953879595, + 0.003693296806886792, + 0.0036094889510422945, + 0.0033106363844126463, + 0.008581222966313362, + 0.0040464079938828945, + 0.011318102478981018, + 0.011815537698566914, + 0.0024485981557518244, + 0.0026408068370074034, + 0.0035984793212264776, + 0.0030017595272511244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.005293875467032194 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tall, sturdy bottle made of smooth, dark glass (almost black) on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a clear crystal goblet half-filled with rich, crimson liquid on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, out-of-focus bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in with an almost imperceptible slowness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009797697886824608 + ], + "opposite_logprob_tokens": [ + -4.6347975730896 + ], + "logprob_reward": -0.009797697886824608, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009797697886824608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05057745799422264, + 0.48605823516845703, + 0.14748644828796387, + 0.041614942252635956, + 0.019400307908654213, + 0.13025255501270294, + 0.03890537843108177, + 0.033128004521131516, + 0.034185752272605896, + 0.01839098148047924 + ], + "credit_map": [ + 0.004869220312684774, + 0.0039124395698308945, + 0.005643030162900686, + 0.008223776705563068, + 0.008177957497537136, + 0.004912218544632196, + 0.008151749148964882, + 0.00668706139549613, + 0.028626389801502228, + 0.03210046887397766, + 0.04774470254778862, + 0.042021531611680984, + 0.06813880056142807, + 0.13712814450263977, + 0.0692533329129219, + 0.061044853180646896, + 0.005974908359348774, + 0.01483213622123003, + 0.016798503696918488, + 0.047158125787973404, + 0.013478409498929977, + 0.010765410959720612, + 0.030221059918403625, + 0.008257881738245487, + 0.0022645897697657347, + 0.0032043620012700558, + 0.004781149327754974, + 0.0063194516114890575, + 0.003197436686605215, + 0.00922007579356432, + 0.006566290277987719, + 0.006061585154384375, + 0.0016337934648618102, + 0.0021415946539491415, + 0.002258395543321967, + 0.002815034007653594, + 0.0017060990212485194, + 0.0034583411179482937, + 0.0025186894927173853, + 0.0028683606069535017, + 0.0061242906376719475, + 0.009933680295944214, + 0.05180907994508743, + 0.04659244790673256, + 0.003686464624479413, + 0.0029393320437520742, + 0.006620137486606836, + 0.002547122538089752, + 0.002610483206808567, + 0.007949170656502247, + 0.008209237828850746, + 0.013188451528549194, + 0.0017069721361622214, + 0.001354079693555832, + 0.002459992654621601, + 0.0014269915409386158, + 0.002082201186567545, + 0.0035401522181928158, + 0.006187467370182276, + 0.006257759407162666, + 0.002828221768140793, + 0.0033504192251712084, + 0.005065283738076687, + 0.0038164977449923754, + 0.004923844709992409, + 0.00579110300168395, + 0.0059096128679811954, + 0.005080576054751873, + 0.0037750904448330402, + 0.003395767416805029, + 0.002640344901010394, + 0.0026694133412092924, + 0.0032444989774376154, + 0.0014035975327715278, + 0.003956722095608711, + 0.0036791509483009577, + 0.0018136659637093544, + 0.0016903443029150367, + 0.0015412252396345139, + 0.0010617768857628107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 5 + ], + "selected_reward": -0.009797697886824608 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tall, sturdy bottle made of smooth, dark glass (almost black) on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a clear crystal goblet half-filled with rich, crimson liquid on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, out-of-focus bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in with an almost imperceptible slowness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00763652054592967 + ], + "opposite_logprob_tokens": [ + -4.882636547088623 + ], + "logprob_reward": -0.00763652054592967, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00763652054592967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09956181049346924, + 0.10570753365755081, + 0.3456844389438629, + 0.05909280478954315, + 0.033386409282684326, + 0.06378302723169327, + 0.05117633193731308, + 0.17106330394744873, + 0.02039429545402527, + 0.05015004053711891 + ], + "credit_map": [ + 0.013939857482910156, + 0.006535552907735109, + 0.015536641702055931, + 0.01209187600761652, + 0.011769458651542664, + 0.013667314313352108, + 0.012899097055196762, + 0.013122006319463253, + 0.007289297878742218, + 0.006953130476176739, + 0.008958018384873867, + 0.008049027994275093, + 0.016014935448765755, + 0.028674840927124023, + 0.016540754586458206, + 0.013227522373199463, + 0.016193898394703865, + 0.035433776676654816, + 0.038531042635440826, + 0.1079944297671318, + 0.035630371421575546, + 0.021933693438768387, + 0.06924751400947571, + 0.020719699561595917, + 0.0037553906440734863, + 0.006404463667422533, + 0.007989224046468735, + 0.009422589093446732, + 0.00457336101680994, + 0.00938475038856268, + 0.00920628197491169, + 0.008356740698218346, + 0.0027426593005657196, + 0.004565281793475151, + 0.00392442150041461, + 0.00569998100399971, + 0.0028783041052520275, + 0.006063355132937431, + 0.004142243880778551, + 0.0033701639622449875, + 0.0026798793114721775, + 0.005921693053096533, + 0.020847447216510773, + 0.020890342071652412, + 0.002245772397145629, + 0.0029568856116384268, + 0.0059805926866829395, + 0.0022604193072766066, + 0.004176834598183632, + 0.008405584841966629, + 0.009788395836949348, + 0.01142063271254301, + 0.004089216701686382, + 0.0028130956925451756, + 0.006311139091849327, + 0.004171433858573437, + 0.013857129961252213, + 0.01767450012266636, + 0.03304317593574524, + 0.03034912794828415, + 0.018966296687722206, + 0.015987439081072807, + 0.018641198053956032, + 0.02254442311823368, + 0.0022656889632344246, + 0.0020542857237160206, + 0.003964114934206009, + 0.003297188552096486, + 0.002469350816681981, + 0.00264708511531353, + 0.001557482872158289, + 0.002139098010957241, + 0.007654337212443352, + 0.0036433024797588587, + 0.007975281216204166, + 0.006895874161273241, + 0.007433661259710789, + 0.006551890168339014, + 0.0062675089575350285, + 0.0037281911354511976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 1 + ], + "selected_reward": -0.00763652054592967 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tall, sturdy bottle made of smooth, dark glass (almost black) on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a clear crystal goblet half-filled with rich, crimson liquid on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, out-of-focus bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in with an almost imperceptible slowness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495283007621765 + ], + "opposite_logprob_tokens": [ + -2.92995285987854 + ], + "logprob_reward": -0.05495283007621765, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495283007621765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3261474370956421, + 0.10372723639011383, + 0.18603810667991638, + 0.0469043143093586, + 0.0723273828625679, + 0.0696013867855072, + 0.09635866433382034, + 0.016715368255972862, + 0.04195144400000572, + 0.04022866114974022 + ], + "credit_map": [ + 0.04153595119714737, + 0.019688228145241737, + 0.0389484167098999, + 0.03766337037086487, + 0.042106952518224716, + 0.04528728500008583, + 0.03974544256925583, + 0.06117180734872818, + 0.007287546992301941, + 0.0080552464351058, + 0.008300838991999626, + 0.012191647663712502, + 0.01856967620551586, + 0.016033926978707314, + 0.01730230636894703, + 0.01598605141043663, + 0.010409664362668991, + 0.02045869082212448, + 0.023625778034329414, + 0.054909318685531616, + 0.018984124064445496, + 0.011615514755249023, + 0.035058069974184036, + 0.010976951569318771, + 0.0034561108332127333, + 0.007379866670817137, + 0.007502340711653233, + 0.007666325196623802, + 0.006055101752281189, + 0.004001904744654894, + 0.0071624405682086945, + 0.0036802201066166162, + 0.004816983826458454, + 0.016217976808547974, + 0.010225917212665081, + 0.01291064452379942, + 0.005502866115421057, + 0.005840093828737736, + 0.012108583003282547, + 0.004704315681010485, + 0.005079648923128843, + 0.009508199989795685, + 0.021727576851844788, + 0.013867339119315147, + 0.0038849508855491877, + 0.004110366106033325, + 0.007448696997016668, + 0.003974609076976776, + 0.005038853269070387, + 0.018357207998633385, + 0.02082163654267788, + 0.03435235843062401, + 0.003615032881498337, + 0.0030738613568246365, + 0.006970318034291267, + 0.004129390232264996, + 0.002089418936520815, + 0.0023365304805338383, + 0.0037925124634057283, + 0.00294664828106761, + 0.0015657332260161638, + 0.0015605908120051026, + 0.0010511446744203568, + 0.0013727907789871097, + 0.006651320029050112, + 0.005706303752958775, + 0.009288029745221138, + 0.008375379256904125, + 0.0027024687733501196, + 0.0033158811274915934, + 0.0034912589471787214, + 0.0024208028335124254, + 0.0036806799471378326, + 0.002803208539262414, + 0.006446603685617447, + 0.007392281200736761, + 0.005957904737442732, + 0.00550084188580513, + 0.005579069722443819, + 0.0028680735267698765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.05495283007621765 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot at eye level, with the camera moving in with an almost imperceptible slowness, captures a composition of exquisite elegance. To the left, a tall, sturdy bottle of smooth, dark glass, its classic shape almost black, rests on a polished, dark wood surface. The bottle's elegant, unpretentious paper label catches the soft light, revealing subtle, shifting reflections on its matte texture. To its right, a clear crystal goblet, half-filled with rich, crimson liquid, emanates a luscious, ruby-like translucency. A warm light source on the right sculpts the scene, creating brilliant highlights along the curve of the glass and casting long, soft shadows across the bottle, artfully creating a striking contrast between the bottle's profound opacity and the glass's vibrant transparency. The background blends into a warm, out-of-focus bokeh, enhancing this simple yet harmonious pairing and immersing the entire image in an atmosphere of quiet refinement and contemplative tranquility.", + "seed": 1160, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_6/index_3958_seed_1160_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03635293890838511, + "vlm_logprob": -0.03635293890838511, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.03635293890838511, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tall, sturdy bottle made of smooth, dark glass (almost black) on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a clear crystal goblet half-filled with rich, crimson liquid on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, out-of-focus bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in with an almost imperceptible slowness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10288708657026291 + ], + "opposite_logprob_tokens": [ + -3.1028871536254883 + ], + "logprob_reward": -0.10288708657026291, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.10288708657026291, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06450240314006805, + 0.2787919044494629, + 0.03041159175336361, + 0.09437504410743713, + 0.12008243799209595, + 0.19057850539684296, + 0.06073903664946556, + 0.04702672362327576, + 0.04150542616844177, + 0.0719868466258049 + ], + "credit_map": [ + 0.009581455960869789, + 0.004787608981132507, + 0.008866767399013042, + 0.009552668780088425, + 0.007745307870209217, + 0.007537683472037315, + 0.007795329205691814, + 0.008635581471025944, + 0.09320558607578278, + 0.008838573470711708, + 0.07870730012655258, + 0.008438401855528355, + 0.01176504883915186, + 0.020419925451278687, + 0.014304743148386478, + 0.04311229661107063, + 0.006066387519240379, + 0.0032160987611860037, + 0.007279016077518463, + 0.002815629355609417, + 0.0026028414722532034, + 0.0023757307790219784, + 0.0031408381182700396, + 0.002915049437433481, + 0.026682542636990547, + 0.006914354395121336, + 0.016833988949656487, + 0.005305813625454903, + 0.005700394511222839, + 0.0071388124488294125, + 0.010461443103849888, + 0.015337700955569744, + 0.023933645337820053, + 0.007819561287760735, + 0.03914893418550491, + 0.008434991352260113, + 0.005378496367484331, + 0.009374523535370827, + 0.005908983759582043, + 0.020083295181393623, + 0.024117130786180496, + 0.006490594241768122, + 0.07128459215164185, + 0.009805279783904552, + 0.00438704015687108, + 0.011281351558864117, + 0.008605294860899448, + 0.054607223719358444, + 0.007635735906660557, + 0.0037669287994503975, + 0.019168652594089508, + 0.004105805419385433, + 0.002968450076878071, + 0.00469948397949338, + 0.004329346120357513, + 0.01406463235616684, + 0.011070536449551582, + 0.005274745635688305, + 0.006760230287909508, + 0.004481855779886246, + 0.003148975083604455, + 0.003777743550017476, + 0.006192647852003574, + 0.006319989915937185, + 0.003970604855567217, + 0.0019099537748843431, + 0.011350352317094803, + 0.0018110232194885612, + 0.002657501492649317, + 0.004220134578645229, + 0.0031196835916489363, + 0.01246617641299963, + 0.00687974551692605, + 0.0026476532220840454, + 0.02278793416917324, + 0.0029360533226281404, + 0.003715943545103073, + 0.007033391389995813, + 0.005024225916713476, + 0.02096189744770527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 4 + ], + "selected_reward": -0.10288708657026291 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tall, sturdy bottle made of smooth, dark glass (almost black) on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a clear crystal goblet half-filled with rich, crimson liquid on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, out-of-focus bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in with an almost imperceptible slowness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022324419114738703 + ], + "opposite_logprob_tokens": [ + -6.127232551574707 + ], + "logprob_reward": -0.0022324419114738703, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022324419114738703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16745369136333466, + 0.16137480735778809, + 0.0639319196343422, + 0.07795359194278717, + 0.114079050719738, + 0.1411340832710266, + 0.09120415896177292, + 0.06011144071817398, + 0.0805477425456047, + 0.04220952093601227 + ], + "credit_map": [ + 0.0172798503190279, + 0.0110935615375638, + 0.034227412194013596, + 0.02459811046719551, + 0.016398724168539047, + 0.017241591587662697, + 0.015868885442614555, + 0.030745534226298332, + 0.046349648386240005, + 0.007538183592259884, + 0.031370870769023895, + 0.010806760750710964, + 0.009530307725071907, + 0.014071669429540634, + 0.00992690771818161, + 0.03178044408559799, + 0.009524841792881489, + 0.006163382902741432, + 0.013011861592531204, + 0.006810684222728014, + 0.007028982043266296, + 0.005407706834375858, + 0.0090211546048522, + 0.006963304243981838, + 0.020790087059140205, + 0.0074776941910386086, + 0.014501705765724182, + 0.005468736402690411, + 0.00602794298902154, + 0.00607477780431509, + 0.00865504052489996, + 0.008957607671618462, + 0.023445099592208862, + 0.008816669695079327, + 0.03394237533211708, + 0.010558740235865116, + 0.004436350893229246, + 0.00888852495700121, + 0.0065419855527579784, + 0.017449291422963142, + 0.027538355439901352, + 0.0070981415919959545, + 0.046187132596969604, + 0.009344805963337421, + 0.006604078225791454, + 0.009022639133036137, + 0.007643878925591707, + 0.02769504301249981, + 0.016778696328401566, + 0.0063546462915837765, + 0.02466849610209465, + 0.006490275729447603, + 0.004485792014747858, + 0.006421438883990049, + 0.0075769005343317986, + 0.01842789724469185, + 0.010720977559685707, + 0.004896031226962805, + 0.013398381881415844, + 0.004573635291308165, + 0.003567006206139922, + 0.0037124520167708397, + 0.00556954788044095, + 0.013673406094312668, + 0.012917808257043362, + 0.005780759267508984, + 0.02058902569115162, + 0.004981076344847679, + 0.004458003211766481, + 0.006652329117059708, + 0.004557180218398571, + 0.020611552521586418, + 0.011132392100989819, + 0.004694068338721991, + 0.007243820000439882, + 0.004075303673744202, + 0.0031882906332612038, + 0.0032943992409855127, + 0.003714495338499546, + 0.004866747185587883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.0022324419114738703 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tall, sturdy bottle made of smooth, dark glass (almost black) on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a clear crystal goblet half-filled with rich, crimson liquid on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, out-of-focus bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in with an almost imperceptible slowness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0003968881501350552 + ], + "opposite_logprob_tokens": [ + -7.875396728515625 + ], + "logprob_reward": -0.0003968881501350552, + "yes_no_margin_reward": 7.875, + "token_credit_reward": -0.0003968881501350552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12158205360174179, + 0.270269513130188, + 0.0603935569524765, + 0.0609595812857151, + 0.12065307796001434, + 0.09610427170991898, + 0.10515785217285156, + 0.06540489196777344, + 0.03252720087766647, + 0.06694801896810532 + ], + "credit_map": [ + 0.020093118771910667, + 0.008461224846541882, + 0.02394942007958889, + 0.012306139804422855, + 0.02030758373439312, + 0.009916014038026333, + 0.01025376096367836, + 0.016294794157147408, + 0.08330675959587097, + 0.011591716669499874, + 0.08184022456407547, + 0.01598403975367546, + 0.016357166692614555, + 0.018783513456583023, + 0.014919127337634563, + 0.027486948296427727, + 0.007410856895148754, + 0.005455028265714645, + 0.011108347214758396, + 0.00570491049438715, + 0.007615597918629646, + 0.005002404097467661, + 0.009528074413537979, + 0.008568339981138706, + 0.010270776227116585, + 0.005414546001702547, + 0.009555398486554623, + 0.0050012958236038685, + 0.008631780743598938, + 0.005574914626777172, + 0.009292559698224068, + 0.007218305487185717, + 0.021949410438537598, + 0.007577148731797934, + 0.03989923745393753, + 0.008035197854042053, + 0.005726146977394819, + 0.009337837807834148, + 0.007574628107249737, + 0.020553477108478546, + 0.012552016414701939, + 0.0041287485510110855, + 0.03443368896842003, + 0.006175480782985687, + 0.005131912883371115, + 0.0061787511222064495, + 0.005311047658324242, + 0.02219262905418873, + 0.018551256507635117, + 0.00711856922134757, + 0.029826216399669647, + 0.007559480145573616, + 0.003994202706962824, + 0.0082060182467103, + 0.007810093462467194, + 0.02209201268851757, + 0.012747837230563164, + 0.006841752678155899, + 0.011063920333981514, + 0.007185845170170069, + 0.004214674234390259, + 0.0050239539705216885, + 0.0075803399085998535, + 0.010746566578745842, + 0.008864739909768105, + 0.003963283728808165, + 0.005200983490794897, + 0.0028453597333282232, + 0.0028271798510104418, + 0.0028816210106015205, + 0.002651569666340947, + 0.003292463021352887, + 0.006700401194393635, + 0.0040754284709692, + 0.01039445586502552, + 0.004148212261497974, + 0.016103681176900864, + 0.009146243333816528, + 0.010129227302968502, + 0.006250370759516954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.0003968881501350552 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tall, sturdy bottle made of smooth, dark glass (almost black) on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a clear crystal goblet half-filled with rich, crimson liquid on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, out-of-focus bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in with an almost imperceptible slowness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014204584993422031 + ], + "opposite_logprob_tokens": [ + -4.264204502105713 + ], + "logprob_reward": -0.014204584993422031, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014204584993422031, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10714983940124512, + 0.5992997884750366, + 0.028526175767183304, + 0.03934513032436371, + 0.023320654407143593, + 0.012651092372834682, + 0.05644794553518295, + 0.040937792509794235, + 0.022498784586787224, + 0.0698227807879448 + ], + "credit_map": [ + 0.013916542753577232, + 0.011110776104032993, + 0.015599977225065231, + 0.01913018338382244, + 0.009346341714262962, + 0.015310673043131828, + 0.009145040065050125, + 0.013590303249657154, + 0.18795214593410492, + 0.024272654205560684, + 0.19381412863731384, + 0.02407500334084034, + 0.0363154374063015, + 0.04519965499639511, + 0.027635738253593445, + 0.06003505364060402, + 0.004798781592398882, + 0.0028775872197002172, + 0.004313965793699026, + 0.003735362319275737, + 0.002279086736962199, + 0.002781430259346962, + 0.0037470953539013863, + 0.0039928690530359745, + 0.007678877096623182, + 0.0038075968623161316, + 0.005945637822151184, + 0.0031796270050108433, + 0.0033598062582314014, + 0.005139884073287249, + 0.004241315182298422, + 0.005992384627461433, + 0.004518712870776653, + 0.0021767618600279093, + 0.005475059151649475, + 0.0018921898445114493, + 0.0018852291395887733, + 0.0014932153280824423, + 0.0018813065253198147, + 0.003998179454356432, + 0.0024716011248528957, + 0.0011830448638647795, + 0.0028051731642335653, + 0.001251077395863831, + 0.0011486690491437912, + 0.0009545916691422462, + 0.001329899183474481, + 0.0015070352237671614, + 0.006578180938959122, + 0.0022839405573904514, + 0.018009325489401817, + 0.0024862424470484257, + 0.0017793982988223433, + 0.003503740532323718, + 0.003066063392907381, + 0.018741056323051453, + 0.004133687354624271, + 0.0022682971321046352, + 0.009607584215700626, + 0.0023131633643060923, + 0.003426877548918128, + 0.004489542450755835, + 0.004271378740668297, + 0.010427262634038925, + 0.003667564131319523, + 0.0016716147074475884, + 0.005211499519646168, + 0.00165323237888515, + 0.001323326607234776, + 0.0021207835525274277, + 0.0014553435612469912, + 0.005395420361310244, + 0.006738976109772921, + 0.0028821814339607954, + 0.022990604862570763, + 0.0035565136931836605, + 0.0047986158169806, + 0.008165858685970306, + 0.00566293578594923, + 0.015027097426354885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.014204584993422031 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tall, sturdy bottle made of smooth, dark glass (almost black) on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a clear crystal goblet half-filled with rich, crimson liquid on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, out-of-focus bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in with an almost imperceptible slowness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0620436929166317 + ], + "opposite_logprob_tokens": [ + -2.8120436668395996 + ], + "logprob_reward": -0.0620436929166317, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0620436929166317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09883332252502441, + 0.1320749968290329, + 0.060569919645786285, + 0.14081913232803345, + 0.07386322319507599, + 0.08959656208753586, + 0.09524081647396088, + 0.05192161351442337, + 0.017737390473484993, + 0.23934300243854523 + ], + "credit_map": [ + 0.014075287617743015, + 0.008325926959514618, + 0.014604075811803341, + 0.010219612158834934, + 0.017288455739617348, + 0.011579684913158417, + 0.01035725325345993, + 0.01238302979618311, + 0.033547982573509216, + 0.005585205275565386, + 0.019558552652597427, + 0.006541971582919359, + 0.010164531879127026, + 0.013674325309693813, + 0.011171435937285423, + 0.03183100000023842, + 0.006691623013466597, + 0.0056425281800329685, + 0.009161497466266155, + 0.004932069219648838, + 0.00387911987490952, + 0.007222634274512529, + 0.007620986085385084, + 0.015419461764395237, + 0.034627798944711685, + 0.015061541460454464, + 0.024368995800614357, + 0.013849076814949512, + 0.009207610972225666, + 0.012801864184439182, + 0.013863537460565567, + 0.017038710415363312, + 0.0150943323969841, + 0.007000279147177935, + 0.016408784314990044, + 0.008607659488916397, + 0.005575668998062611, + 0.0054407003335654736, + 0.008090641349554062, + 0.007645157631486654, + 0.012144763953983784, + 0.004731826949864626, + 0.030364420264959335, + 0.006803994998335838, + 0.004890950862318277, + 0.005942509975284338, + 0.0058434526436030865, + 0.018874647095799446, + 0.015921374782919884, + 0.011157969012856483, + 0.013780536130070686, + 0.009131752885878086, + 0.009558272548019886, + 0.012184717692434788, + 0.010536971502006054, + 0.012969224713742733, + 0.008241530507802963, + 0.004618846345692873, + 0.010144810192286968, + 0.004774065222591162, + 0.003870770102366805, + 0.004942708648741245, + 0.004380738828331232, + 0.01094814669340849, + 0.0033877468667924404, + 0.0017125684535130858, + 0.0032922287937253714, + 0.0013684717705473304, + 0.0017654085531830788, + 0.0017409962601959705, + 0.0015332179609686136, + 0.0029367520473897457, + 0.016649801284074783, + 0.009170236997306347, + 0.0679863765835762, + 0.008422331884503365, + 0.019923772662878036, + 0.026171697303652763, + 0.017391301691532135, + 0.07362748682498932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 1 + ], + "selected_reward": -0.0620436929166317 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot at eye level, with the camera moving in with an almost imperceptible slowness, captures a composition of exquisite elegance. To the left, a tall, sturdy bottle of smooth, dark glass, its classic shape almost black, rests on a polished, dark wood surface. The bottle's elegant, unpretentious paper label catches the soft light, revealing subtle, shifting reflections on its matte texture. To its right, a clear crystal goblet, half-filled with rich, crimson liquid, emanates a luscious, ruby-like translucency. A warm light source on the right sculpts the scene, creating brilliant highlights along the curve of the glass and casting long, soft shadows across the bottle, artfully creating a striking contrast between the bottle's profound opacity and the glass's vibrant transparency. The background blends into a warm, out-of-focus bokeh, enhancing this simple yet harmonious pairing and immersing the entire image in an atmosphere of quiet refinement and contemplative tranquility.", + "seed": 1161, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_6/index_3958_seed_1161_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02618302045157179, + "vlm_logprob": -0.02618302045157179, + "vlm_yes_no_margin": 5.075, + "vlm_token_credit": -0.02618302045157179, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tall, sturdy bottle made of smooth, dark glass (almost black) on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a clear crystal goblet half-filled with rich, crimson liquid on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, out-of-focus bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in with an almost imperceptible slowness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08072464913129807 + ], + "opposite_logprob_tokens": [ + -3.3307247161865234 + ], + "logprob_reward": -0.08072464913129807, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.08072464913129807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42110633850097656, + 0.12759971618652344, + 0.08225994557142258, + 0.05168931931257248, + 0.054976947605609894, + 0.02535799890756607, + 0.12803946435451508, + 0.04357752576470375, + 0.055532537400722504, + 0.00986019242554903 + ], + "credit_map": [ + 0.025007542222738266, + 0.16015248000621796, + 0.10460146516561508, + 0.06763165444135666, + 0.016066623851656914, + 0.014965393580496311, + 0.015468166209757328, + 0.01721302978694439, + 0.014001364819705486, + 0.027343206107616425, + 0.00980143528431654, + 0.034161388874053955, + 0.013067160733044147, + 0.00760718435049057, + 0.012521418742835522, + 0.009096547961235046, + 0.005652222316712141, + 0.030021317303180695, + 0.022477835416793823, + 0.012616414576768875, + 0.0037900605238974094, + 0.0031367517076432705, + 0.0020834726747125387, + 0.002481874544173479, + 0.006366423796862364, + 0.005483501125127077, + 0.00649533374235034, + 0.00944812037050724, + 0.005035754758864641, + 0.006490238942205906, + 0.008040115237236023, + 0.004329830873757601, + 0.005247264634817839, + 0.008345267735421658, + 0.004971493035554886, + 0.007719841320067644, + 0.005217171274125576, + 0.009891228750348091, + 0.004842552822083235, + 0.008742131292819977, + 0.003678833832964301, + 0.005366197321563959, + 0.001940265647135675, + 0.004682608414441347, + 0.0033742713276296854, + 0.002189381280913949, + 0.002123776823282242, + 0.002002664376050234, + 0.008480587974190712, + 0.04241691902279854, + 0.009335318580269814, + 0.04740598797798157, + 0.0060804178938269615, + 0.005839190911501646, + 0.005058716516941786, + 0.0034223233815282583, + 0.004979693330824375, + 0.006192096509039402, + 0.006029215641319752, + 0.0072823697701096535, + 0.008111481554806232, + 0.004021128173917532, + 0.003932280000299215, + 0.0030292582232505083, + 0.006373890675604343, + 0.0081360237672925, + 0.0061055440455675125, + 0.00880676694214344, + 0.007592992857098579, + 0.0041717104613780975, + 0.0090111019089818, + 0.0053345030173659325, + 0.0010431252885609865, + 0.0017267054645344615, + 0.0010276783723384142, + 0.0012119091115891933, + 0.0013055471936240792, + 0.001208349596709013, + 0.0008991790236905217, + 0.001437697559595108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.08072464913129807 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tall, sturdy bottle made of smooth, dark glass (almost black) on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a clear crystal goblet half-filled with rich, crimson liquid on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, out-of-focus bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in with an almost imperceptible slowness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004673154558986425 + ], + "opposite_logprob_tokens": [ + -5.379673004150391 + ], + "logprob_reward": -0.004673154558986425, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004673154558986425, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5604905486106873, + 0.044348470866680145, + 0.028125008568167686, + 0.03402144834399223, + 0.018197715282440186, + 0.04640683904290199, + 0.13538889586925507, + 0.026379097253084183, + 0.07619649916887283, + 0.030445490032434464 + ], + "credit_map": [ + 0.04426257684826851, + 0.20881685614585876, + 0.08009211719036102, + 0.11304876953363419, + 0.02645101770758629, + 0.027937185019254684, + 0.02892371080815792, + 0.030958272516727448, + 0.004732166416943073, + 0.00883110798895359, + 0.004996531177312136, + 0.01240900345146656, + 0.0030007516033947468, + 0.0034571962896734476, + 0.0036142561584711075, + 0.0033074587117880583, + 0.005300144664943218, + 0.003647119738161564, + 0.005671308375895023, + 0.0033417653758078814, + 0.0024774863850325346, + 0.002224304946139455, + 0.00283083226531744, + 0.0026320461183786392, + 0.005848702508956194, + 0.0028475781437009573, + 0.007707743439823389, + 0.006187283433973789, + 0.0023772972635924816, + 0.0029700775630772114, + 0.003673453349620104, + 0.0024093131069093943, + 0.0019541862420737743, + 0.0025936015881597996, + 0.0016412831610068679, + 0.0020881453529000282, + 0.0019377628341317177, + 0.002166627673432231, + 0.002528669312596321, + 0.0032874392345547676, + 0.004019656218588352, + 0.016820499673485756, + 0.001902107847854495, + 0.013027055189013481, + 0.003575623268261552, + 0.0021081494633108377, + 0.002868283074349165, + 0.0020854624453932047, + 0.007846423424780369, + 0.035274963825941086, + 0.01781177520751953, + 0.04978584125638008, + 0.004778905771672726, + 0.006196849979460239, + 0.004416371695697308, + 0.009277774021029472, + 0.004538141656666994, + 0.0027786947321146727, + 0.006844701711088419, + 0.003465736983343959, + 0.002878502244129777, + 0.0024223930668085814, + 0.0015322276158258319, + 0.0019187009893357754, + 0.009341845288872719, + 0.012154604308307171, + 0.007311129476875067, + 0.012200586497783661, + 0.00884323287755251, + 0.008262557908892632, + 0.009692509658634663, + 0.008390032686293125, + 0.003828078508377075, + 0.004940485581755638, + 0.003250932553783059, + 0.004739425145089626, + 0.003151980461552739, + 0.004322015214711428, + 0.002929548267275095, + 0.003283025696873665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.004673154558986425 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tall, sturdy bottle made of smooth, dark glass (almost black) on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a clear crystal goblet half-filled with rich, crimson liquid on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, out-of-focus bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in with an almost imperceptible slowness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007281036232598126 + ], + "opposite_logprob_tokens": [ + -7.250728130340576 + ], + "logprob_reward": -0.0007281036232598126, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007281036232598126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12820805609226227, + 0.16674575209617615, + 0.15470024943351746, + 0.048315759748220444, + 0.03872288018465042, + 0.06172632426023483, + 0.29595476388931274, + 0.02405831217765808, + 0.035085488110780716, + 0.04648244380950928 + ], + "credit_map": [ + 0.014998257160186768, + 0.02057335525751114, + 0.023254267871379852, + 0.01400960423052311, + 0.006330026779323816, + 0.015810659155249596, + 0.011964656412601471, + 0.021267244592308998, + 0.017835751175880432, + 0.03824930265545845, + 0.012147530913352966, + 0.04815314710140228, + 0.01508830301463604, + 0.012180439196527004, + 0.012796841561794281, + 0.010294429957866669, + 0.015831690281629562, + 0.0361911877989769, + 0.04051640257239342, + 0.01817438192665577, + 0.008750918321311474, + 0.01197856292128563, + 0.009780516847968102, + 0.013476588763296604, + 0.007501133251935244, + 0.004234427120536566, + 0.008273222483694553, + 0.00833896454423666, + 0.0035052630119025707, + 0.00637260964140296, + 0.005125498399138451, + 0.004964641761034727, + 0.004554553888738155, + 0.0068977586925029755, + 0.0034607078414410353, + 0.005170947406440973, + 0.0034187836572527885, + 0.0037885073106735945, + 0.006502737756818533, + 0.004928885493427515, + 0.0054222759790718555, + 0.021038314327597618, + 0.00283485883846879, + 0.016142185777425766, + 0.00480624521151185, + 0.003741208929568529, + 0.004512522369623184, + 0.003228710498660803, + 0.012479318305850029, + 0.08531256020069122, + 0.03341158479452133, + 0.11839719861745834, + 0.011481383815407753, + 0.012478913180530071, + 0.01148947048932314, + 0.01090437825769186, + 0.004205194301903248, + 0.004677227232605219, + 0.0041811843402683735, + 0.00334743014536798, + 0.0016450954135507345, + 0.00240990798920393, + 0.0014126715250313282, + 0.0021796012297272682, + 0.005092844367027283, + 0.004953292198479176, + 0.003688480705022812, + 0.004815406631678343, + 0.0031424856279045343, + 0.004495964851230383, + 0.004750458057969809, + 0.004146555904299021, + 0.006141789723187685, + 0.005640391726046801, + 0.005842374172061682, + 0.005544506944715977, + 0.004730576649308205, + 0.00844916608184576, + 0.004469491075724363, + 0.005664150696247816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 2 + ], + "selected_reward": -0.0007281036232598126 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tall, sturdy bottle made of smooth, dark glass (almost black) on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a clear crystal goblet half-filled with rich, crimson liquid on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, out-of-focus bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in with an almost imperceptible slowness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017215682892128825 + ], + "opposite_logprob_tokens": [ + -6.376721382141113 + ], + "logprob_reward": -0.0017215682892128825, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017215682892128825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1937963217496872, + 0.3209303021430969, + 0.05989518761634827, + 0.03192054107785225, + 0.10117951035499573, + 0.0491488054394722, + 0.11304133385419846, + 0.041012756526470184, + 0.07080598175525665, + 0.018269285559654236 + ], + "credit_map": [ + 0.015652159228920937, + 0.051775455474853516, + 0.027842847630381584, + 0.027345934882760048, + 0.010458563454449177, + 0.01713877171278, + 0.015413179993629456, + 0.028169410303235054, + 0.036154963076114655, + 0.078914113342762, + 0.027044575661420822, + 0.10819420218467712, + 0.01827019266784191, + 0.01781662367284298, + 0.018019340932369232, + 0.016516294330358505, + 0.013974582776427269, + 0.011179880239069462, + 0.008100413717329502, + 0.00948571227490902, + 0.0028063058853149414, + 0.0037800618447363377, + 0.006422135047614574, + 0.0041460939683020115, + 0.005963431671261787, + 0.0042153592221438885, + 0.0032016097102314234, + 0.004788139835000038, + 0.002623444190248847, + 0.0037018104922026396, + 0.0038542824331671, + 0.0035724653862416744, + 0.011891116388142109, + 0.0316857248544693, + 0.005908170249313116, + 0.026795586571097374, + 0.004835657775402069, + 0.005859387572854757, + 0.007142342627048492, + 0.007061517331749201, + 0.005413332022726536, + 0.01631658710539341, + 0.0028231588657945395, + 0.012717231176793575, + 0.003525546984747052, + 0.002527285134419799, + 0.002872649347409606, + 0.002953016897663474, + 0.005339812953025103, + 0.03748105838894844, + 0.00755661353468895, + 0.04305321350693703, + 0.004409699235111475, + 0.004828586243093014, + 0.005163132678717375, + 0.005209218244999647, + 0.005273622460663319, + 0.006070943549275398, + 0.003676055697724223, + 0.005331446882337332, + 0.008570880629122257, + 0.00521493423730135, + 0.0036178568843752146, + 0.0032570164185017347, + 0.010324530303478241, + 0.010747725144028664, + 0.009125182405114174, + 0.011516343802213669, + 0.007944929413497448, + 0.0056378645822405815, + 0.009540035389363766, + 0.005969368387013674, + 0.002097868360579014, + 0.003666423726826906, + 0.002171129919588566, + 0.0032932250760495663, + 0.001553845126181841, + 0.0014972633216530085, + 0.0023648187052458525, + 0.0016247110906988382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.0017215682892128825 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tall, sturdy bottle made of smooth, dark glass (almost black) on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a clear crystal goblet half-filled with rich, crimson liquid on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, out-of-focus bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in with an almost imperceptible slowness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043067626655101776 + ], + "opposite_logprob_tokens": [ + -3.168067693710327 + ], + "logprob_reward": -0.043067626655101776, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043067626655101776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40546128153800964, + 0.10305200517177582, + 0.1429285705089569, + 0.0692761018872261, + 0.029006069526076317, + 0.022099245339632034, + 0.04243043437600136, + 0.03258376568555832, + 0.07074698805809021, + 0.08241550624370575 + ], + "credit_map": [ + 0.02869272232055664, + 0.15033145248889923, + 0.0676703155040741, + 0.0759962797164917, + 0.019692178815603256, + 0.021391499787569046, + 0.020131736993789673, + 0.021555129438638687, + 0.013465506955981255, + 0.02516067959368229, + 0.0076969088986516, + 0.030938731506466866, + 0.00697473855689168, + 0.0056168390437960625, + 0.006504318211227655, + 0.0066942861303687096, + 0.006977811921387911, + 0.05680375546216965, + 0.03608037158846855, + 0.027341600507497787, + 0.005456870421767235, + 0.0033550707157701254, + 0.003495843382552266, + 0.003417249070480466, + 0.011937915347516537, + 0.013820566236972809, + 0.007618342060595751, + 0.012663030065596104, + 0.004478392656892538, + 0.006733680609613657, + 0.006035428959876299, + 0.005988751072436571, + 0.00310217821970582, + 0.0049754297360777855, + 0.001978266052901745, + 0.0048989541828632355, + 0.0023401633370667696, + 0.00456964410841465, + 0.002993685659021139, + 0.004147747531533241, + 0.0019038082100450993, + 0.0035578878596425056, + 0.002160424832254648, + 0.0030963739845901728, + 0.0037563801743090153, + 0.0029281796887516975, + 0.002423989586532116, + 0.002272200770676136, + 0.009845019318163395, + 0.007720139808952808, + 0.007060749921947718, + 0.005753687117248774, + 0.003604614408686757, + 0.0032780810724943876, + 0.0026892994064837694, + 0.0024788440205156803, + 0.005352652166038752, + 0.003984468057751656, + 0.006730920635163784, + 0.004216859582811594, + 0.001982282381504774, + 0.0036392230540513992, + 0.0016186762368306518, + 0.005058683920651674, + 0.010611185804009438, + 0.008949454873800278, + 0.0108559001237154, + 0.010018647648394108, + 0.008357957005500793, + 0.006365432403981686, + 0.009568457491695881, + 0.0060199531726539135, + 0.009742524474859238, + 0.016451260074973106, + 0.010586310178041458, + 0.014250135980546474, + 0.006091776769608259, + 0.0075986324809491634, + 0.009032693691551685, + 0.008662172593176365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.043067626655101776 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_7/index_tensor([6858, 6858])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_7/index_tensor([6858, 6858])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1c8d054e9f647d57eabfe1c32cb5300cfefa9a42 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_7/index_tensor([6858, 6858])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "cinematic, low-angle, extreme wide shot begins, the camera tilting upward while simultaneously pulling back, gradually revealing the full, terrifying scale of a massive biomechanical structure silently wriggling in the upper atmosphere. Its form is a nightmarish tapestry reminiscent of H.R. Giger, a seamless fusion of polished obsidian exoskeleton and damp, gleaming sinew, with ribbed skeletal frameworks connecting pulsating organic sacs. The entire surface shimmers as if eternally damp in the distant, cold starlight, contrasting starkly with the faint curvature of the planet below and the absolute blackness of space. The massive entity undulates with a slow, eerie grace, its distinctive, funnel-shaped spouts flaring, steadily and deliberately drawing in swirling clouds of ethereal cosmic particles that glitter like diamond dust. As these particles are absorbed, the intricate network of translucent veins running beneath its black metal shell begins to glow with a sickly, fluctuating turquoise light, visually representing the horrific metamorphosis occurring within. Suddenly, with a convulsive shudder, the building released this trapped energy; violent bursts of jagged white-blue lightning erupted from the spires of its belly and struck downward with terrifying speed upon the unseen world below, illuminating the scene with brilliant, terrible flashes and creating an overwhelming atmosphere of silent cosmic terror.", + "seed": 6952, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_7/index_6858_seed_6952_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04437764310277999, + "vlm_logprob": -0.04437764310277999, + "vlm_yes_no_margin": 3.9999999523162844, + "vlm_token_credit": -0.04437764310277999, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that begins with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward while simultaneously pulling back to reveal a massive biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the structure composed of polished obsidian exoskeleton and damp, gleaming sinew with ribbed skeletal frameworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent veins beneath the black metal shell glow with a sickly, fluctuating turquoise light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do violent bursts of jagged white-blue lightning erupt from the structure's spires and strike downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16002492606639862 + ], + "opposite_logprob_tokens": [ + -2.035024881362915 + ], + "logprob_reward": -0.16002492606639862, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.16002492606639862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09667667001485825, + 0.05889402702450752, + 0.1175934374332428, + 0.1216709241271019, + 0.06489241123199463, + 0.11061131209135056, + 0.12011142075061798, + 0.12053520977497101, + 0.04473666101694107, + 0.14427794516086578 + ], + "credit_map": [ + 0.011476353742182255, + 0.00761719373986125, + 0.006871630437672138, + 0.012145171873271465, + 0.012570063583552837, + 0.015438493341207504, + 0.015122782438993454, + 0.015434976667165756, + 0.009496621787548065, + 0.007780304178595543, + 0.007003062404692173, + 0.006444155238568783, + 0.006020683329552412, + 0.008464982733130455, + 0.008051175624132156, + 0.005633043125271797, + 0.014544778503477573, + 0.014732453972101212, + 0.014419279992580414, + 0.015515618026256561, + 0.009576295502483845, + 0.02442375011742115, + 0.011025004088878632, + 0.013356258161365986, + 0.010790928266942501, + 0.02379150129854679, + 0.016726281493902206, + 0.012606995180249214, + 0.013712254352867603, + 0.016872962936758995, + 0.014269156381487846, + 0.01290084607899189, + 0.006161519326269627, + 0.008987809531390667, + 0.0109398839995265, + 0.0066396682523190975, + 0.007810930255800486, + 0.00955814030021429, + 0.006868438329547644, + 0.007926024496555328, + 0.015685399994254112, + 0.006321602035313845, + 0.021512748673558235, + 0.0353674441576004, + 0.0062482357025146484, + 0.005042340140789747, + 0.01289461925625801, + 0.007538916543126106, + 0.02687261439859867, + 0.007164446171373129, + 0.04337979853153229, + 0.016601305454969406, + 0.004165732767432928, + 0.003324277698993683, + 0.01190844178199768, + 0.006694804411381483, + 0.01404200866818428, + 0.012243042699992657, + 0.017315970733761787, + 0.01660747639834881, + 0.008162282407283783, + 0.008123229257762432, + 0.025536634027957916, + 0.01850457303225994, + 0.0037208194844424725, + 0.008490651845932007, + 0.004944952204823494, + 0.005318212788552046, + 0.004934439901262522, + 0.006193430628627539, + 0.006346350070089102, + 0.004787805490195751, + 0.007057297509163618, + 0.010836776345968246, + 0.021432112902402878, + 0.007857544347643852, + 0.025982536375522614, + 0.05684048309922218, + 0.007152112200856209, + 0.007119075860828161 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 7 + ], + "selected_reward": -0.16002492606639862 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that begins with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward while simultaneously pulling back to reveal a massive biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the structure composed of polished obsidian exoskeleton and damp, gleaming sinew with ribbed skeletal frameworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent veins beneath the black metal shell glow with a sickly, fluctuating turquoise light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do violent bursts of jagged white-blue lightning erupt from the structure's spires and strike downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067648435942828655 + ], + "opposite_logprob_tokens": [ + -5.006764888763428 + ], + "logprob_reward": -0.0067648435942828655, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067648435942828655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04626273736357689, + 0.05209608003497124, + 0.08271484822034836, + 0.03706974908709526, + 0.01930948719382286, + 0.06928087770938873, + 0.05751216411590576, + 0.046248823404312134, + 0.03622651472687721, + 0.5532786846160889 + ], + "credit_map": [ + 0.006266743876039982, + 0.0030451074708253145, + 0.003642147174105048, + 0.007598561234772205, + 0.005563888233155012, + 0.0054612974636256695, + 0.00839545764029026, + 0.006289533339440823, + 0.007414259482175112, + 0.007809247821569443, + 0.005884428508579731, + 0.0054992809891700745, + 0.007193588651716709, + 0.007184260990470648, + 0.005741199478507042, + 0.005369813181459904, + 0.010729864239692688, + 0.008658076636493206, + 0.010538864880800247, + 0.011418195441365242, + 0.005613153334707022, + 0.019595060497522354, + 0.006096793804317713, + 0.010064837522804737, + 0.0035854955203831196, + 0.006045275833457708, + 0.005030080210417509, + 0.005373953841626644, + 0.0035221027210354805, + 0.004813104867935181, + 0.004902199376374483, + 0.003797535551711917, + 0.0015686489641666412, + 0.0026382594369351864, + 0.0030219426844269037, + 0.002635103417560458, + 0.0022290453780442476, + 0.0019283471629023552, + 0.0028677869122475386, + 0.0024203527718782425, + 0.008685174398124218, + 0.00467982143163681, + 0.011194734834134579, + 0.022623850032687187, + 0.008190641179680824, + 0.003335252869874239, + 0.007100176531821489, + 0.0034712280612438917, + 0.007641841657459736, + 0.002954269992187619, + 0.019835958257317543, + 0.013413617387413979, + 0.0036720200441777706, + 0.0022946344688534737, + 0.005019859876483679, + 0.0026799615006893873, + 0.0053327386267483234, + 0.004760657902806997, + 0.006536908447742462, + 0.005691975355148315, + 0.00363775622099638, + 0.0039244722574949265, + 0.010830157436430454, + 0.005534160882234573, + 0.004614305216819048, + 0.006289470940828323, + 0.003980137873440981, + 0.004067669156938791, + 0.0033017753157764673, + 0.007162484340369701, + 0.0036618835292756557, + 0.003148791380226612, + 0.041291575878858566, + 0.047424785792827606, + 0.1577117145061493, + 0.030100274831056595, + 0.08381522446870804, + 0.141831636428833, + 0.027273954823613167, + 0.023829512298107147 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 5 + ], + "selected_reward": -0.0067648435942828655 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that begins with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward while simultaneously pulling back to reveal a massive biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the structure composed of polished obsidian exoskeleton and damp, gleaming sinew with ribbed skeletal frameworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent veins beneath the black metal shell glow with a sickly, fluctuating turquoise light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do violent bursts of jagged white-blue lightning erupt from the structure's spires and strike downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372899070382118 + ], + "opposite_logprob_tokens": [ + -3.408729076385498 + ], + "logprob_reward": -0.03372899070382118, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372899070382118, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15058135986328125, + 0.05867137759923935, + 0.2313026636838913, + 0.05801406875252724, + 0.07148604840040207, + 0.08836667984724045, + 0.14934757351875305, + 0.06465911865234375, + 0.05379217863082886, + 0.07377896457910538 + ], + "credit_map": [ + 0.01855405792593956, + 0.008326815441250801, + 0.009014487266540527, + 0.027347497642040253, + 0.012231271713972092, + 0.014436249621212482, + 0.039235252887010574, + 0.02143574133515358, + 0.007055202033370733, + 0.007287680171430111, + 0.007354462053626776, + 0.008596700616180897, + 0.0067354231141507626, + 0.0057573155499994755, + 0.009883351624011993, + 0.006001248024404049, + 0.026361651718616486, + 0.030595624819397926, + 0.025451267138123512, + 0.030027596279978752, + 0.01900598593056202, + 0.052008308470249176, + 0.020238617435097694, + 0.027613602578639984, + 0.005711037665605545, + 0.007710603065788746, + 0.00807934533804655, + 0.0054823048412799835, + 0.0055016991682350636, + 0.011120116338133812, + 0.007546075154095888, + 0.006862882059067488, + 0.009457704611122608, + 0.008877702988684177, + 0.01136716827750206, + 0.00920309778302908, + 0.007063771598041058, + 0.009425299242138863, + 0.008288384415209293, + 0.007802919019013643, + 0.008135360665619373, + 0.007928552106022835, + 0.009813493117690086, + 0.0180136039853096, + 0.018567919731140137, + 0.007511388044804335, + 0.012285922653973103, + 0.006110438611358404, + 0.030085168778896332, + 0.00898406095802784, + 0.05416551232337952, + 0.02481934428215027, + 0.006789785344153643, + 0.004212726838886738, + 0.013944124802947044, + 0.0063468413427472115, + 0.00676830904558301, + 0.007268769200891256, + 0.009379943832755089, + 0.007483948953449726, + 0.006780543830245733, + 0.009803393855690956, + 0.007125159725546837, + 0.010049048811197281, + 0.006716619711369276, + 0.008427798748016357, + 0.0046353694051504135, + 0.004814142361283302, + 0.005518057849258184, + 0.014714342541992664, + 0.004631825722754002, + 0.0043340218253433704, + 0.00663596298545599, + 0.007596036419272423, + 0.01765166036784649, + 0.005502647254616022, + 0.013052579946815968, + 0.014675932936370373, + 0.004665361251682043, + 0.003998781554400921 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.03372899070382118 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that begins with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward while simultaneously pulling back to reveal a massive biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the structure composed of polished obsidian exoskeleton and damp, gleaming sinew with ribbed skeletal frameworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent veins beneath the black metal shell glow with a sickly, fluctuating turquoise light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do violent bursts of jagged white-blue lightning erupt from the structure's spires and strike downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00319268717430532 + ], + "opposite_logprob_tokens": [ + -5.753192901611328 + ], + "logprob_reward": -0.00319268717430532, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00319268717430532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08594545722007751, + 0.05190281569957733, + 0.19475406408309937, + 0.044744811952114105, + 0.03429453819990158, + 0.11637242138385773, + 0.08361376076936722, + 0.04854168742895126, + 0.02294447459280491, + 0.3168860673904419 + ], + "credit_map": [ + 0.009684257209300995, + 0.00599868455901742, + 0.005240307189524174, + 0.02152472734451294, + 0.008048895746469498, + 0.008694078773260117, + 0.011539842933416367, + 0.015214650891721249, + 0.006098882760852575, + 0.006360972765833139, + 0.006317597348242998, + 0.005998671054840088, + 0.006433495786041021, + 0.007843181490898132, + 0.007629675790667534, + 0.005220334976911545, + 0.022764336317777634, + 0.021533939987421036, + 0.02851172164082527, + 0.02772502787411213, + 0.014835329726338387, + 0.04193655773997307, + 0.01649729534983635, + 0.020949849858880043, + 0.004353824537247419, + 0.007201864384114742, + 0.007130911573767662, + 0.005181239452213049, + 0.003930169157683849, + 0.007098210044205189, + 0.004712694324553013, + 0.005135894753038883, + 0.004124414641410112, + 0.004611903335899115, + 0.0054126642644405365, + 0.004306276328861713, + 0.0035592501517385244, + 0.00409079110249877, + 0.004692615009844303, + 0.00349661847576499, + 0.014163763262331486, + 0.009832408279180527, + 0.018657471984624863, + 0.034517012536525726, + 0.014992612414062023, + 0.005911213345825672, + 0.012243200093507767, + 0.006054722238332033, + 0.017399277538061142, + 0.006166449282318354, + 0.01930089294910431, + 0.011087297461926937, + 0.008781230077147484, + 0.0036570411175489426, + 0.011565589345991611, + 0.0056559741497039795, + 0.008559844456613064, + 0.00424161646515131, + 0.005038844887167215, + 0.005953450221568346, + 0.005026685073971748, + 0.0033452946227043867, + 0.011608011089265347, + 0.004767935257405043, + 0.00222253380343318, + 0.004304560832679272, + 0.0024019302800297737, + 0.0028583130333572626, + 0.002423617523163557, + 0.003389653516933322, + 0.0030330419540405273, + 0.0023108201567083597, + 0.023787157610058784, + 0.027263840660452843, + 0.08901084959506989, + 0.017836028710007668, + 0.05102325230836868, + 0.0779535248875618, + 0.016180694103240967, + 0.01383068598806858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 5 + ], + "selected_reward": -0.00319268717430532 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that begins with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward while simultaneously pulling back to reveal a massive biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the structure composed of polished obsidian exoskeleton and damp, gleaming sinew with ribbed skeletal frameworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent veins beneath the black metal shell glow with a sickly, fluctuating turquoise light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do violent bursts of jagged white-blue lightning erupt from the structure's spires and strike downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018176767975091934 + ], + "opposite_logprob_tokens": [ + -4.018176555633545 + ], + "logprob_reward": -0.018176767975091934, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018176767975091934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02414614148437977, + 0.006042626686394215, + 0.03428040072321892, + 0.008585566654801369, + 0.007223288528621197, + 0.013880486600100994, + 0.007931878790259361, + 0.007309794891625643, + 0.017447087913751602, + 0.8731527328491211 + ], + "credit_map": [ + 0.0014103864086791873, + 0.0010591393802314997, + 0.0008108679903671145, + 0.008465834893286228, + 0.0010721649741753936, + 0.001750883529894054, + 0.0024110774975270033, + 0.007165786810219288, + 0.0008576042600907385, + 0.0006996546871960163, + 0.000750992854591459, + 0.0005709931720048189, + 0.0007625984726473689, + 0.0009704272379167378, + 0.0008059732499532402, + 0.0006243825773708522, + 0.004032603930681944, + 0.0034547848626971245, + 0.004265912342816591, + 0.004704317543655634, + 0.002830485114827752, + 0.007986028678715229, + 0.0029315955471247435, + 0.004074671771377325, + 0.000885866757016629, + 0.0016886164667084813, + 0.0011702930787578225, + 0.0007392828119918704, + 0.0009249188005924225, + 0.001083495793864131, + 0.0009066120837815106, + 0.0011864813277497888, + 0.000621217826846987, + 0.0012677808990702033, + 0.0011881054379045963, + 0.0009720588568598032, + 0.0007466229144483805, + 0.00058312265900895, + 0.0010013016872107983, + 0.0008430785965174437, + 0.0017812501173466444, + 0.001042802701704204, + 0.002487137448042631, + 0.004427840001881123, + 0.0013584583066403866, + 0.0005601117154583335, + 0.0013815703568980098, + 0.0008413162431679666, + 0.0006354082725010812, + 0.0005053186323493719, + 0.0022418282460421324, + 0.0018752707401290536, + 0.0007488892879337072, + 0.00043217887287028134, + 0.0010041543282568455, + 0.000488830148242414, + 0.001050569349899888, + 0.0005754326703026891, + 0.0008673828560858965, + 0.0007394953863695264, + 0.0010972019517794251, + 0.0006803396390751004, + 0.0016194903291761875, + 0.0006798827089369297, + 0.001987035619094968, + 0.0030715004540979862, + 0.0019671879708766937, + 0.0019097907934337854, + 0.0016124859685078263, + 0.003417884698137641, + 0.0018619988113641739, + 0.001619202084839344, + 0.06328795850276947, + 0.07819537818431854, + 0.24470852315425873, + 0.04553355276584625, + 0.0960211306810379, + 0.26360735297203064, + 0.04262600839138031, + 0.03917280584573746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.018176767975091934 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, low-angle, extreme wide shot begins, the camera tilting upward while simultaneously pulling back, gradually revealing the full, terrifying scale of a massive biomechanical structure silently wriggling in the upper atmosphere. Its form is a nightmarish tapestry reminiscent of H.R. Giger, a seamless fusion of polished obsidian exoskeleton and damp, gleaming sinew, with ribbed skeletal frameworks connecting pulsating organic sacs. The entire surface shimmers as if eternally damp in the distant, cold starlight, contrasting starkly with the faint curvature of the planet below and the absolute blackness of space. The massive entity undulates with a slow, eerie grace, its distinctive, funnel-shaped spouts flaring, steadily and deliberately drawing in swirling clouds of ethereal cosmic particles that glitter like diamond dust. As these particles are absorbed, the intricate network of translucent veins running beneath its black metal shell begins to glow with a sickly, fluctuating turquoise light, visually representing the horrific metamorphosis occurring within. Suddenly, with a convulsive shudder, the building released this trapped energy; violent bursts of jagged white-blue lightning erupted from the spires of its belly and struck downward with terrifying speed upon the unseen world below, illuminating the scene with brilliant, terrible flashes and creating an overwhelming atmosphere of silent cosmic terror.", + "seed": 6953, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_7/index_6858_seed_6953_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.017454718099907042, + "vlm_logprob": -0.017454718099907042, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.017454718099907042, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that begins with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward while simultaneously pulling back to reveal a massive biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the structure composed of polished obsidian exoskeleton and damp, gleaming sinew with ribbed skeletal frameworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent veins beneath the black metal shell glow with a sickly, fluctuating turquoise light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do violent bursts of jagged white-blue lightning erupt from the structure's spires and strike downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.037788912653923035 + ], + "opposite_logprob_tokens": [ + -3.4127888679504395 + ], + "logprob_reward": -0.037788912653923035, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.037788912653923035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2236504703760147, + 0.06061724200844765, + 0.16565020382404327, + 0.08843297511339188, + 0.045882079750299454, + 0.0897594541311264, + 0.1300850212574005, + 0.12781482934951782, + 0.031112391501665115, + 0.03699544444680214 + ], + "credit_map": [ + 0.017250116914510727, + 0.04388986527919769, + 0.024660654366016388, + 0.03221132978796959, + 0.03685290738940239, + 0.026119207963347435, + 0.023838035762310028, + 0.018828323110938072, + 0.008268949575722218, + 0.007235890720039606, + 0.008529639802873135, + 0.007146869786083698, + 0.008447903208434582, + 0.008924726396799088, + 0.005964381620287895, + 0.006098875775933266, + 0.01774442195892334, + 0.024395795539021492, + 0.01271793432533741, + 0.030871232971549034, + 0.03155529126524925, + 0.021568261086940765, + 0.016935894265770912, + 0.009861360304057598, + 0.008746439591050148, + 0.0109221450984478, + 0.01458998303860426, + 0.00942948367446661, + 0.012151731178164482, + 0.010965896770358086, + 0.011716975830495358, + 0.009910309687256813, + 0.006612683646380901, + 0.006120480131357908, + 0.004915934056043625, + 0.0051857600919902325, + 0.004862657748162746, + 0.006043989211320877, + 0.005918506067246199, + 0.006222062278538942, + 0.007229791488498449, + 0.008843828924000263, + 0.012793219648301601, + 0.009112249128520489, + 0.007267167326062918, + 0.01567533239722252, + 0.01189788430929184, + 0.01693997159600258, + 0.011180061846971512, + 0.012037547305226326, + 0.016184894368052483, + 0.018433813005685806, + 0.011719366535544395, + 0.019599316641688347, + 0.01894432120025158, + 0.02198568917810917, + 0.010762877762317657, + 0.015246993862092495, + 0.01494310051202774, + 0.015044435858726501, + 0.01850997470319271, + 0.01899215765297413, + 0.02008420042693615, + 0.014231068082153797, + 0.004175170790404081, + 0.00454953545704484, + 0.003591937478631735, + 0.003897392190992832, + 0.004035183694213629, + 0.002977964701130986, + 0.0037518402095884085, + 0.00413336418569088, + 0.004545699805021286, + 0.00479454081505537, + 0.0045892116613686085, + 0.0039561595767736435, + 0.006631040014326572, + 0.0042641195468604565, + 0.003582149278372526, + 0.004632522817701101 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.037788912653923035 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that begins with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward while simultaneously pulling back to reveal a massive biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the structure composed of polished obsidian exoskeleton and damp, gleaming sinew with ribbed skeletal frameworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent veins beneath the black metal shell glow with a sickly, fluctuating turquoise light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do violent bursts of jagged white-blue lightning erupt from the structure's spires and strike downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012553953565657139 + ], + "opposite_logprob_tokens": [ + -4.387554168701172 + ], + "logprob_reward": -0.012553953565657139, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012553953565657139, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15542857348918915, + 0.07890893518924713, + 0.17359930276870728, + 0.04018624126911163, + 0.0790480524301529, + 0.2041994035243988, + 0.06385917216539383, + 0.07180733233690262, + 0.0398758202791214, + 0.09308717399835587 + ], + "credit_map": [ + 0.013083214871585369, + 0.02722153440117836, + 0.019397901371121407, + 0.02002895623445511, + 0.029455188661813736, + 0.01317146047949791, + 0.014705484732985497, + 0.018364833667874336, + 0.00786156952381134, + 0.014821653254330158, + 0.004222637042403221, + 0.008066413924098015, + 0.020606977865099907, + 0.011263201013207436, + 0.005476763006299734, + 0.006589718163013458, + 0.019910559058189392, + 0.021279992535710335, + 0.018691856414079666, + 0.02859555557370186, + 0.030130358412861824, + 0.022465474903583527, + 0.01883789524435997, + 0.013687608763575554, + 0.0038874202873557806, + 0.005478648003190756, + 0.004628214053809643, + 0.005896363873034716, + 0.006562036462128162, + 0.005259231198579073, + 0.004898434970527887, + 0.003575891489163041, + 0.010012632235884666, + 0.006604067049920559, + 0.009420393966138363, + 0.009935101494193077, + 0.006998701021075249, + 0.011410311795771122, + 0.012745656073093414, + 0.011921189725399017, + 0.018639585003256798, + 0.029725346714258194, + 0.03480987623333931, + 0.02095642127096653, + 0.021746398881077766, + 0.029727589339017868, + 0.024504059925675392, + 0.02409011870622635, + 0.005974560510367155, + 0.010110273025929928, + 0.006810568738728762, + 0.006038912106305361, + 0.008755427785217762, + 0.009619089774787426, + 0.005845368839800358, + 0.01070497091859579, + 0.008111447095870972, + 0.01111836452037096, + 0.0068036396987736225, + 0.007590596564114094, + 0.010093902237713337, + 0.010573463514447212, + 0.007568202447146177, + 0.00994771346449852, + 0.00463073281571269, + 0.004043249413371086, + 0.004050034563988447, + 0.0046646567061543465, + 0.004900401458144188, + 0.004135661758482456, + 0.0051452359184622765, + 0.008305848576128483, + 0.009900309145450592, + 0.010817306116223335, + 0.011509493924677372, + 0.01231155451387167, + 0.00944665726274252, + 0.011628448963165283, + 0.01064014621078968, + 0.01683325320482254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.012553953565657139 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that begins with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward while simultaneously pulling back to reveal a massive biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the structure composed of polished obsidian exoskeleton and damp, gleaming sinew with ribbed skeletal frameworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent veins beneath the black metal shell glow with a sickly, fluctuating turquoise light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do violent bursts of jagged white-blue lightning erupt from the structure's spires and strike downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029809430241584778 + ], + "opposite_logprob_tokens": [ + -3.5298094749450684 + ], + "logprob_reward": -0.029809430241584778, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029809430241584778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16326168179512024, + 0.07117792218923569, + 0.08667000383138657, + 0.13631847500801086, + 0.23505699634552002, + 0.12367498129606247, + 0.08062677830457687, + 0.0353870652616024, + 0.028086809441447258, + 0.03973930701613426 + ], + "credit_map": [ + 0.016099058091640472, + 0.022559164091944695, + 0.01732269488275051, + 0.031200407072901726, + 0.02104085311293602, + 0.016554944217205048, + 0.0219782255589962, + 0.016506334766745567, + 0.007328195963054895, + 0.012070591561496258, + 0.007054935675114393, + 0.009720935486257076, + 0.014825363643467426, + 0.008219430223107338, + 0.006989805493503809, + 0.004968664143234491, + 0.007364173885434866, + 0.010707389563322067, + 0.01024530827999115, + 0.011803894303739071, + 0.013356171548366547, + 0.010386199690401554, + 0.011337417177855968, + 0.011469447985291481, + 0.014860589988529682, + 0.021250568330287933, + 0.017158405855298042, + 0.015692124143242836, + 0.02458927594125271, + 0.01820652186870575, + 0.01376180537045002, + 0.01079918909817934, + 0.026014866307377815, + 0.022794868797063828, + 0.029090935364365578, + 0.03142930567264557, + 0.030167724937200546, + 0.03145701438188553, + 0.03631683439016342, + 0.027785442769527435, + 0.013223807327449322, + 0.01930181495845318, + 0.01100943237543106, + 0.009552602656185627, + 0.016438443213701248, + 0.021875279024243355, + 0.011411057785153389, + 0.02086254395544529, + 0.007077542133629322, + 0.007422810420393944, + 0.007413784973323345, + 0.01194764394313097, + 0.009314505383372307, + 0.013011627830564976, + 0.011217976920306683, + 0.013220887631177902, + 0.0034508011303842068, + 0.004718339070677757, + 0.0034878868609666824, + 0.004625422414392233, + 0.0061106448993086815, + 0.0044490257278084755, + 0.004364571068435907, + 0.004180372226983309, + 0.0027271222788840532, + 0.0038526663556694984, + 0.0028152146842330694, + 0.00366817694157362, + 0.003333928296342492, + 0.004368592519313097, + 0.003430495271459222, + 0.003890612395480275, + 0.005398062989115715, + 0.004615351557731628, + 0.004989492241293192, + 0.005013641901314259, + 0.003921349532902241, + 0.004147827159613371, + 0.004400461912155151, + 0.007253117859363556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.029809430241584778 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that begins with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward while simultaneously pulling back to reveal a massive biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the structure composed of polished obsidian exoskeleton and damp, gleaming sinew with ribbed skeletal frameworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent veins beneath the black metal shell glow with a sickly, fluctuating turquoise light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do violent bursts of jagged white-blue lightning erupt from the structure's spires and strike downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011784045491367579 + ], + "opposite_logprob_tokens": [ + -6.75117826461792 + ], + "logprob_reward": -0.0011784045491367579, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011784045491367579, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10837174952030182, + 0.05705349147319794, + 0.08309809863567352, + 0.1083725094795227, + 0.22310742735862732, + 0.08570551127195358, + 0.14983302354812622, + 0.04048176854848862, + 0.04893851280212402, + 0.09503792226314545 + ], + "credit_map": [ + 0.008520961739122868, + 0.015465132892131805, + 0.017080536112189293, + 0.016659125685691833, + 0.013750291429460049, + 0.007431976962834597, + 0.014156702905893326, + 0.015307016670703888, + 0.005839432589709759, + 0.010108515620231628, + 0.004995951894670725, + 0.007772450335323811, + 0.012335200794041157, + 0.006652086973190308, + 0.005724055226892233, + 0.0036257952451705933, + 0.006764872465282679, + 0.012603259645402431, + 0.009303204715251923, + 0.012431585229933262, + 0.012894385494291782, + 0.010031835176050663, + 0.008988049812614918, + 0.010080906562507153, + 0.012250293046236038, + 0.018189938738942146, + 0.011777539737522602, + 0.014409372583031654, + 0.01958937756717205, + 0.014057386666536331, + 0.010704469867050648, + 0.007394121028482914, + 0.025139302015304565, + 0.01983271911740303, + 0.02361312322318554, + 0.029352225363254547, + 0.025229698047041893, + 0.03121916577219963, + 0.0368448868393898, + 0.03187631070613861, + 0.01148267276585102, + 0.013725439086556435, + 0.009173177182674408, + 0.008226903155446053, + 0.013423486612737179, + 0.01110763382166624, + 0.008805192075669765, + 0.009761005640029907, + 0.009195066057145596, + 0.016968833282589912, + 0.01941841095685959, + 0.023387309163808823, + 0.017649799585342407, + 0.02157205156981945, + 0.02539709582924843, + 0.01624445617198944, + 0.003922421019524336, + 0.005478049162775278, + 0.00496119586750865, + 0.0040410710498690605, + 0.006337469909340143, + 0.005379049107432365, + 0.0046447208151221275, + 0.005717791151255369, + 0.007401198148727417, + 0.007736345287412405, + 0.0062036034651100636, + 0.005939546972513199, + 0.006652866490185261, + 0.0042702616192400455, + 0.006217371206730604, + 0.004517322406172752, + 0.00829875934869051, + 0.01581258699297905, + 0.010609613731503487, + 0.011487758718430996, + 0.011676454916596413, + 0.011041573248803616, + 0.011306097730994225, + 0.014805078506469727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 3 + ], + "selected_reward": -0.0011784045491367579 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that begins with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward while simultaneously pulling back to reveal a massive biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the structure composed of polished obsidian exoskeleton and damp, gleaming sinew with ribbed skeletal frameworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent veins beneath the black metal shell glow with a sickly, fluctuating turquoise light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do violent bursts of jagged white-blue lightning erupt from the structure's spires and strike downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005942889489233494 + ], + "opposite_logprob_tokens": [ + -5.1309428215026855 + ], + "logprob_reward": -0.005942889489233494, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005942889489233494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11109800636768341, + 0.043171778321266174, + 0.1440337747335434, + 0.03904348239302635, + 0.07190170884132385, + 0.07144241034984589, + 0.05124707520008087, + 0.06120285391807556, + 0.09371219575405121, + 0.3131466209888458 + ], + "credit_map": [ + 0.008318861946463585, + 0.016444077715277672, + 0.01631343364715576, + 0.013478800654411316, + 0.017107708379626274, + 0.01546426210552454, + 0.01043744757771492, + 0.013533424586057663, + 0.0043466100469231606, + 0.005801472347229719, + 0.007335723843425512, + 0.005969304591417313, + 0.006560806650668383, + 0.005114923231303692, + 0.004475723952054977, + 0.003567218314856291, + 0.014533039182424545, + 0.01645062118768692, + 0.018038086593151093, + 0.02205231785774231, + 0.02384345233440399, + 0.017266632989048958, + 0.016266223043203354, + 0.015583422966301441, + 0.003456322243437171, + 0.006203002296388149, + 0.005494805984199047, + 0.004111387766897678, + 0.006719538941979408, + 0.005276585929095745, + 0.004243331495672464, + 0.0035385119263082743, + 0.005399197805672884, + 0.007290743291378021, + 0.00830749049782753, + 0.009274802170693874, + 0.008478495292365551, + 0.00976383313536644, + 0.013910611160099506, + 0.00947654340416193, + 0.00757952407002449, + 0.00838479958474636, + 0.009736070409417152, + 0.008494640700519085, + 0.008950015529990196, + 0.01076606847345829, + 0.007743038237094879, + 0.009788263589143753, + 0.006531507708132267, + 0.007329958491027355, + 0.00790887139737606, + 0.005814923904836178, + 0.005255779251456261, + 0.007321953307837248, + 0.005516842007637024, + 0.005567240994423628, + 0.006846041418612003, + 0.00734325498342514, + 0.0067435940727591515, + 0.00739276921376586, + 0.008625365793704987, + 0.007994163781404495, + 0.006926465313881636, + 0.009331201203167439, + 0.012438708916306496, + 0.015174144878983498, + 0.007787854876369238, + 0.009049912914633751, + 0.016217393800616264, + 0.011325218714773655, + 0.008749897591769695, + 0.012969065457582474, + 0.03657529503107071, + 0.03569522500038147, + 0.03524748980998993, + 0.03283025324344635, + 0.0452040359377861, + 0.04670359194278717, + 0.028765391558408737, + 0.05212535709142685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.005942889489233494 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, low-angle, extreme wide shot begins, the camera tilting upward while simultaneously pulling back, gradually revealing the full, terrifying scale of a massive biomechanical structure silently wriggling in the upper atmosphere. Its form is a nightmarish tapestry reminiscent of H.R. Giger, a seamless fusion of polished obsidian exoskeleton and damp, gleaming sinew, with ribbed skeletal frameworks connecting pulsating organic sacs. The entire surface shimmers as if eternally damp in the distant, cold starlight, contrasting starkly with the faint curvature of the planet below and the absolute blackness of space. The massive entity undulates with a slow, eerie grace, its distinctive, funnel-shaped spouts flaring, steadily and deliberately drawing in swirling clouds of ethereal cosmic particles that glitter like diamond dust. As these particles are absorbed, the intricate network of translucent veins running beneath its black metal shell begins to glow with a sickly, fluctuating turquoise light, visually representing the horrific metamorphosis occurring within. Suddenly, with a convulsive shudder, the building released this trapped energy; violent bursts of jagged white-blue lightning erupted from the spires of its belly and struck downward with terrifying speed upon the unseen world below, illuminating the scene with brilliant, terrible flashes and creating an overwhelming atmosphere of silent cosmic terror.", + "seed": 6954, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_7/index_6858_seed_6954_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05992337395437062, + "vlm_logprob": -0.05992337395437062, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.05992337395437062, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that begins with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward while simultaneously pulling back to reveal a massive biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the structure composed of polished obsidian exoskeleton and damp, gleaming sinew with ribbed skeletal frameworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent veins beneath the black metal shell glow with a sickly, fluctuating turquoise light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do violent bursts of jagged white-blue lightning erupt from the structure's spires and strike downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.24484319984912872 + ], + "opposite_logprob_tokens": [ + -1.6198432445526123 + ], + "logprob_reward": -0.24484319984912872, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.24484319984912872, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1310088187456131, + 0.2358938604593277, + 0.06322609633207321, + 0.09892716258764267, + 0.27836406230926514, + 0.02822396531701088, + 0.022438032552599907, + 0.06746204197406769, + 0.0320175476372242, + 0.04243838042020798 + ], + "credit_map": [ + 0.014403591863811016, + 0.02289859764277935, + 0.012706165201961994, + 0.020851997658610344, + 0.01471303403377533, + 0.016891498118638992, + 0.017199523746967316, + 0.011344421654939651, + 0.01584937795996666, + 0.02228543348610401, + 0.0151442876085639, + 0.019780555739998817, + 0.020177528262138367, + 0.06702422350645065, + 0.05575040355324745, + 0.0198820848017931, + 0.00632797135040164, + 0.002841150388121605, + 0.020543554797768593, + 0.005274622701108456, + 0.005583218298852444, + 0.00682965200394392, + 0.005569394212216139, + 0.01025653537362814, + 0.00963548943400383, + 0.009705840609967709, + 0.011988366954028606, + 0.010797650553286076, + 0.011063278652727604, + 0.018164608627557755, + 0.016662325710058212, + 0.010909607633948326, + 0.021465297788381577, + 0.036959074437618256, + 0.028739964589476585, + 0.0570022352039814, + 0.023071572184562683, + 0.03505668416619301, + 0.04405106604099274, + 0.03201819211244583, + 0.0024316974449902773, + 0.004036810249090195, + 0.002967200009152293, + 0.004047048278152943, + 0.003650170983746648, + 0.004161901772022247, + 0.003635900095105171, + 0.003293238580226898, + 0.002132855821400881, + 0.003197300946339965, + 0.003195091150701046, + 0.0029825123492628336, + 0.0021824361756443977, + 0.0031415231060236692, + 0.002727633574977517, + 0.0028786812908947468, + 0.0068051498383283615, + 0.009651025757193565, + 0.010082927532494068, + 0.006631436292082071, + 0.006017348263412714, + 0.009727336466312408, + 0.009887215681374073, + 0.008659608662128448, + 0.0027657835744321346, + 0.003596188733354211, + 0.005077174864709377, + 0.005867192056030035, + 0.003613358596339822, + 0.003523812862113118, + 0.0038474653847515583, + 0.0037265766877681017, + 0.003096860134974122, + 0.006871234159916639, + 0.0065925996750593185, + 0.006109038833528757, + 0.004559805616736412, + 0.004786436911672354, + 0.005443011410534382, + 0.004979400429874659 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.24484319984912872 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that begins with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward while simultaneously pulling back to reveal a massive biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the structure composed of polished obsidian exoskeleton and damp, gleaming sinew with ribbed skeletal frameworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent veins beneath the black metal shell glow with a sickly, fluctuating turquoise light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do violent bursts of jagged white-blue lightning erupt from the structure's spires and strike downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004670662805438042 + ], + "opposite_logprob_tokens": [ + -5.3796706199646 + ], + "logprob_reward": -0.004670662805438042, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004670662805438042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0808180496096611, + 0.06476465612649918, + 0.03921513631939888, + 0.03023703396320343, + 0.046762801706790924, + 0.09049634635448456, + 0.07172960788011551, + 0.12460760027170181, + 0.23321056365966797, + 0.21815824508666992 + ], + "credit_map": [ + 0.01317090354859829, + 0.013272477313876152, + 0.008913090452551842, + 0.006390339694917202, + 0.01116690132766962, + 0.008911403827369213, + 0.012465185485780239, + 0.0065277437679469585, + 0.004098559729754925, + 0.007849336601793766, + 0.00741198193281889, + 0.00743960402905941, + 0.005500716157257557, + 0.011525939218699932, + 0.011632570996880531, + 0.00930594652891159, + 0.004885363858193159, + 0.0037671560421586037, + 0.004237762186676264, + 0.004332162439823151, + 0.007831309922039509, + 0.004820419475436211, + 0.00413875188678503, + 0.005202210042625666, + 0.0027677768375724554, + 0.002947318134829402, + 0.006140889599919319, + 0.0029855105094611645, + 0.003649459918960929, + 0.002905718283727765, + 0.003933597356081009, + 0.004906762391328812, + 0.003708151401951909, + 0.0058483243919909, + 0.005699174478650093, + 0.007750438991934061, + 0.004665607586503029, + 0.005309800151735544, + 0.008242442272603512, + 0.00553886266425252, + 0.007897586561739445, + 0.015379749238491058, + 0.007700100541114807, + 0.013001827523112297, + 0.012579004280269146, + 0.00812468584626913, + 0.01542384922504425, + 0.010389539413154125, + 0.007578568533062935, + 0.009672407060861588, + 0.008463562466204166, + 0.012556585483253002, + 0.006375967524945736, + 0.007611228618770838, + 0.010191272012889385, + 0.009280024096369743, + 0.012728477828204632, + 0.01581139676272869, + 0.017401332035660744, + 0.013122482225298882, + 0.01120448112487793, + 0.021286169067025185, + 0.015049930661916733, + 0.018003322184085846, + 0.021972570568323135, + 0.0303163044154644, + 0.04258717969059944, + 0.028929436579346657, + 0.026344453915953636, + 0.028948064893484116, + 0.027881810441613197, + 0.02623073384165764, + 0.01969916932284832, + 0.029851432889699936, + 0.03551021218299866, + 0.026334619149565697, + 0.021052461117506027, + 0.027294941246509552, + 0.038354404270648956, + 0.02006099931895733 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.004670662805438042 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that begins with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward while simultaneously pulling back to reveal a massive biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the structure composed of polished obsidian exoskeleton and damp, gleaming sinew with ribbed skeletal frameworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent veins beneath the black metal shell glow with a sickly, fluctuating turquoise light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do violent bursts of jagged white-blue lightning erupt from the structure's spires and strike downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038125645369291306 + ], + "opposite_logprob_tokens": [ + -3.28812575340271 + ], + "logprob_reward": -0.038125645369291306, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038125645369291306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10231375694274902, + 0.20850956439971924, + 0.04123438149690628, + 0.05724857375025749, + 0.07811488211154938, + 0.02637900412082672, + 0.04801363870501518, + 0.03367404267191887, + 0.19514158368110657, + 0.20937064290046692 + ], + "credit_map": [ + 0.012676145881414413, + 0.01613071747124195, + 0.014361991547048092, + 0.011552836745977402, + 0.012329486198723316, + 0.010885855183005333, + 0.01436823420226574, + 0.010008489713072777, + 0.012623537331819534, + 0.009276537224650383, + 0.010558558627963066, + 0.013552024029195309, + 0.012842746451497078, + 0.07253129780292511, + 0.0631132647395134, + 0.01401157770305872, + 0.0039675417356193066, + 0.0023741163313388824, + 0.008970898576080799, + 0.002831356367096305, + 0.0066832974553108215, + 0.005590817425400019, + 0.0036556641571223736, + 0.007160686422139406, + 0.005897276569157839, + 0.00555617269128561, + 0.008141383528709412, + 0.006729059386998415, + 0.006104302126914263, + 0.008604619652032852, + 0.009863664396107197, + 0.00635208934545517, + 0.007037552539259195, + 0.013087640516459942, + 0.007826827466487885, + 0.007137429900467396, + 0.008145377971231937, + 0.013083870522677898, + 0.014050043188035488, + 0.007746129762381315, + 0.0020613819360733032, + 0.004995740484446287, + 0.0026700638700276613, + 0.0034355255775153637, + 0.003735379781574011, + 0.003189394250512123, + 0.0035020571667701006, + 0.002789456630125642, + 0.004311831202358007, + 0.007333599030971527, + 0.006133832968771458, + 0.007652337197214365, + 0.004269636236131191, + 0.004261289723217487, + 0.00816381722688675, + 0.00588728953152895, + 0.0031762649305164814, + 0.003631273750215769, + 0.005029451102018356, + 0.00397894810885191, + 0.003311876440420747, + 0.006035570986568928, + 0.004077525809407234, + 0.004433128517121077, + 0.01379256509244442, + 0.016506528481841087, + 0.03479589521884918, + 0.06081376224756241, + 0.014757558703422546, + 0.01669268123805523, + 0.022400589659810066, + 0.01538198720663786, + 0.018405715003609657, + 0.03737584128975868, + 0.023976048454642296, + 0.030102333053946495, + 0.020680584013462067, + 0.028445927426218987, + 0.026758598163723946, + 0.023625561967492104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 8 + ], + "selected_reward": -0.038125645369291306 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that begins with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward while simultaneously pulling back to reveal a massive biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the structure composed of polished obsidian exoskeleton and damp, gleaming sinew with ribbed skeletal frameworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent veins beneath the black metal shell glow with a sickly, fluctuating turquoise light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do violent bursts of jagged white-blue lightning erupt from the structure's spires and strike downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021997320000082254 + ], + "opposite_logprob_tokens": [ + -6.127199649810791 + ], + "logprob_reward": -0.0021997320000082254, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021997320000082254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08324196189641953, + 0.048134252429008484, + 0.05243242159485817, + 0.04899728298187256, + 0.02858121693134308, + 0.025043662637472153, + 0.0720275491476059, + 0.0427667498588562, + 0.10795672237873077, + 0.49081817269325256 + ], + "credit_map": [ + 0.010711614973843098, + 0.01056591421365738, + 0.010154129937291145, + 0.013323402963578701, + 0.00879210326820612, + 0.008863021619617939, + 0.013047515414655209, + 0.007784253917634487, + 0.0032679899595677853, + 0.003741192165762186, + 0.0036321927327662706, + 0.006168232299387455, + 0.003324147081002593, + 0.011660913936793804, + 0.009937150403857231, + 0.006402429658919573, + 0.004934243392199278, + 0.002673377748578787, + 0.017639504745602608, + 0.0032303626649081707, + 0.003594358451664448, + 0.004754745867103338, + 0.0036576897837221622, + 0.01194814220070839, + 0.0052565112709999084, + 0.004150270018726587, + 0.008019737899303436, + 0.004880903754383326, + 0.00568220904096961, + 0.007227005437016487, + 0.007526914123445749, + 0.006253732368350029, + 0.0026731989346444607, + 0.003358211601153016, + 0.0033301946241408587, + 0.003355576889589429, + 0.0032473611645400524, + 0.005182713735848665, + 0.004105734638869762, + 0.0033282251097261906, + 0.001769133610650897, + 0.0037859610747545958, + 0.0025378218851983547, + 0.004195573274046183, + 0.002924449509009719, + 0.0027387484442442656, + 0.004509487189352512, + 0.002582486951723695, + 0.006710006855428219, + 0.010546958073973656, + 0.008761529810726643, + 0.011064221151173115, + 0.006182996556162834, + 0.007935810834169388, + 0.01242911908775568, + 0.008396913297474384, + 0.0031815769616514444, + 0.005705695599317551, + 0.004781270865350962, + 0.007408989127725363, + 0.004251956008374691, + 0.005515771917998791, + 0.007159545551985502, + 0.00476194079965353, + 0.00790186133235693, + 0.00813042838126421, + 0.017840659245848656, + 0.03988463431596756, + 0.007444050628691912, + 0.005359548144042492, + 0.014208660461008549, + 0.007186873350292444, + 0.042046044021844864, + 0.09226636588573456, + 0.062073107808828354, + 0.06694347411394119, + 0.0487385019659996, + 0.06171417608857155, + 0.05838626250624657, + 0.058650244027376175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.0021997320000082254 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that begins with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward while simultaneously pulling back to reveal a massive biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the structure composed of polished obsidian exoskeleton and damp, gleaming sinew with ribbed skeletal frameworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent veins beneath the black metal shell glow with a sickly, fluctuating turquoise light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do violent bursts of jagged white-blue lightning erupt from the structure's spires and strike downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009777629747986794 + ], + "opposite_logprob_tokens": [ + -4.634777545928955 + ], + "logprob_reward": -0.009777629747986794, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009777629747986794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.012478986755013466, + 0.00966412853449583, + 0.018093597143888474, + 0.005187137983739376, + 0.016573268920183182, + 0.007074838038533926, + 0.009075703099370003, + 0.019132332876324654, + 0.10357539355754852, + 0.7991446852684021 + ], + "credit_map": [ + 0.0015596551820635796, + 0.0016020556213334203, + 0.0014718499733135104, + 0.001583175384439528, + 0.0014058068627491593, + 0.0017520457040518522, + 0.001696583814918995, + 0.0014078138628974557, + 0.0007432089769281447, + 0.0007740411092527211, + 0.0008138050325214863, + 0.0011052497429773211, + 0.0011696716537699103, + 0.0021362563129514456, + 0.0021103061735630035, + 0.0008115895325317979, + 0.0013168464647606015, + 0.0006631715223193169, + 0.005269830115139484, + 0.0011881248792633414, + 0.0015618056058883667, + 0.0013052548747509718, + 0.001331157865934074, + 0.005457405932247639, + 0.0004887831164523959, + 0.000579517858568579, + 0.0007205232977867126, + 0.0007990311714820564, + 0.0006400105776265264, + 0.0006948515656404197, + 0.0007782284519635141, + 0.0004861917404923588, + 0.0014608470955863595, + 0.0025311100762337446, + 0.001721706474199891, + 0.0030723263043910265, + 0.0017579463310539722, + 0.0018042135052382946, + 0.002627610694617033, + 0.001597509253770113, + 0.001011532498523593, + 0.0011280473554506898, + 0.0007006715168245137, + 0.0009089592495001853, + 0.0008589437929913402, + 0.000829362019430846, + 0.0007459046901203692, + 0.000891416915692389, + 0.001322984928265214, + 0.0016461675986647606, + 0.0008905300055630505, + 0.001442431821487844, + 0.0007691857172176242, + 0.000713435816578567, + 0.001100698602385819, + 0.0011902680853381753, + 0.0021536287385970354, + 0.0039411550387740135, + 0.0021106356289237738, + 0.002156174508854747, + 0.0017512633930891752, + 0.002217338653281331, + 0.0025052290875464678, + 0.0022969089914113283, + 0.009990932419896126, + 0.01425442285835743, + 0.014513503760099411, + 0.012586778961122036, + 0.012646153569221497, + 0.013775192201137543, + 0.012235116213560104, + 0.013573290780186653, + 0.0665343850851059, + 0.14961466193199158, + 0.09218199551105499, + 0.11141126602888107, + 0.07895314693450928, + 0.10102280974388123, + 0.09810006618499756, + 0.10132630169391632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.009777629747986794 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, low-angle, extreme wide shot begins, the camera tilting upward while simultaneously pulling back, gradually revealing the full, terrifying scale of a massive biomechanical structure silently wriggling in the upper atmosphere. Its form is a nightmarish tapestry reminiscent of H.R. Giger, a seamless fusion of polished obsidian exoskeleton and damp, gleaming sinew, with ribbed skeletal frameworks connecting pulsating organic sacs. The entire surface shimmers as if eternally damp in the distant, cold starlight, contrasting starkly with the faint curvature of the planet below and the absolute blackness of space. The massive entity undulates with a slow, eerie grace, its distinctive, funnel-shaped spouts flaring, steadily and deliberately drawing in swirling clouds of ethereal cosmic particles that glitter like diamond dust. As these particles are absorbed, the intricate network of translucent veins running beneath its black metal shell begins to glow with a sickly, fluctuating turquoise light, visually representing the horrific metamorphosis occurring within. Suddenly, with a convulsive shudder, the building released this trapped energy; violent bursts of jagged white-blue lightning erupted from the spires of its belly and struck downward with terrifying speed upon the unseen world below, illuminating the scene with brilliant, terrible flashes and creating an overwhelming atmosphere of silent cosmic terror.", + "seed": 6955, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_7/index_6858_seed_6955_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03430282974150032, + "vlm_logprob": -0.03430282974150032, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.03430282974150032, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that begins with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward while simultaneously pulling back to reveal a massive biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the structure composed of polished obsidian exoskeleton and damp, gleaming sinew with ribbed skeletal frameworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent veins beneath the black metal shell glow with a sickly, fluctuating turquoise light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do violent bursts of jagged white-blue lightning erupt from the structure's spires and strike downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07664600759744644 + ], + "opposite_logprob_tokens": [ + -2.701646089553833 + ], + "logprob_reward": -0.07664600759744644, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07664600759744644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0949946790933609, + 0.04343339055776596, + 0.0732455626130104, + 0.08489250391721725, + 0.18174119293689728, + 0.11766735464334488, + 0.14330929517745972, + 0.17554275691509247, + 0.06845629960298538, + 0.016717005521059036 + ], + "credit_map": [ + 0.010187022387981415, + 0.012400994077324867, + 0.01109171099960804, + 0.008927819319069386, + 0.01180177740752697, + 0.011641647666692734, + 0.015331581234931946, + 0.01361212320625782, + 0.007255543023347855, + 0.004402967169880867, + 0.005510457791388035, + 0.007792260032147169, + 0.00473201647400856, + 0.005654969718307257, + 0.0041616689413785934, + 0.003923505544662476, + 0.008519702591001987, + 0.008786126971244812, + 0.007850234396755695, + 0.01025836355984211, + 0.011887181550264359, + 0.008211848326027393, + 0.009455800987780094, + 0.00827630516141653, + 0.007462414912879467, + 0.006332389544695616, + 0.01660080812871456, + 0.008955962024629116, + 0.0070863692089915276, + 0.02384183555841446, + 0.006388487294316292, + 0.008224238641560078, + 0.014706895686686039, + 0.014991981908679008, + 0.027302728965878487, + 0.0456792414188385, + 0.01613171212375164, + 0.0117877759039402, + 0.03242392838001251, + 0.01871691271662712, + 0.01927887462079525, + 0.011549551971256733, + 0.016454437747597694, + 0.013018683530390263, + 0.015450762584805489, + 0.008806390687823296, + 0.015189021825790405, + 0.017919626086950302, + 0.014754346571862698, + 0.014630189165472984, + 0.019918566569685936, + 0.020367417484521866, + 0.018287362530827522, + 0.016424119472503662, + 0.025849493220448494, + 0.013077785260975361, + 0.01948302984237671, + 0.019586661830544472, + 0.027087517082691193, + 0.02506360411643982, + 0.02190481126308441, + 0.014807334169745445, + 0.03045281209051609, + 0.01715698093175888, + 0.007932379841804504, + 0.008895676583051682, + 0.010177687741816044, + 0.007453594822436571, + 0.01023406907916069, + 0.009570300579071045, + 0.008970088325440884, + 0.005222502164542675, + 0.0014098432147875428, + 0.0009387132595293224, + 0.0026120927650481462, + 0.0016134834149852395, + 0.004916701000183821, + 0.0012429371709004045, + 0.002583881374448538, + 0.0013993537286296487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 6 + ], + "selected_reward": -0.07664600759744644 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that begins with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward while simultaneously pulling back to reveal a massive biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the structure composed of polished obsidian exoskeleton and damp, gleaming sinew with ribbed skeletal frameworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent veins beneath the black metal shell glow with a sickly, fluctuating turquoise light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do violent bursts of jagged white-blue lightning erupt from the structure's spires and strike downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006759989075362682 + ], + "opposite_logprob_tokens": [ + -5.006760120391846 + ], + "logprob_reward": -0.006759989075362682, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006759989075362682, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10180962085723877, + 0.21579933166503906, + 0.0364515446126461, + 0.12094160914421082, + 0.11007466912269592, + 0.08663851022720337, + 0.0885540172457695, + 0.10533513873815536, + 0.05434219539165497, + 0.08005333691835403 + ], + "credit_map": [ + 0.009597808122634888, + 0.008072931319475174, + 0.010541082359850407, + 0.008118696510791779, + 0.008101552724838257, + 0.023164086043834686, + 0.008573947474360466, + 0.02563951537013054, + 0.025681793689727783, + 0.031850460916757584, + 0.026588790118694305, + 0.08107571303844452, + 0.01177311223000288, + 0.014030717313289642, + 0.014760185964405537, + 0.010038560256361961, + 0.005265741143375635, + 0.004377488046884537, + 0.003009512322023511, + 0.009219895116984844, + 0.0040550921112298965, + 0.004348597954958677, + 0.0034573518205434084, + 0.002717864466831088, + 0.01996018923819065, + 0.00861417967826128, + 0.015468966215848923, + 0.028354506939649582, + 0.00954433623701334, + 0.02174276113510132, + 0.009730174206197262, + 0.007526488043367863, + 0.013312103226780891, + 0.012157515622675419, + 0.014962650835514069, + 0.021810511127114296, + 0.012747244909405708, + 0.007553696632385254, + 0.017853502184152603, + 0.009677443653345108, + 0.01114162988960743, + 0.008378377184271812, + 0.011893405579030514, + 0.009337373077869415, + 0.01200944371521473, + 0.007669494487345219, + 0.011411788873374462, + 0.014796997420489788, + 0.009629909880459309, + 0.007948500104248524, + 0.012484022416174412, + 0.011344525031745434, + 0.013156501576304436, + 0.009572603739798069, + 0.015463479794561863, + 0.008954467251896858, + 0.011802353896200657, + 0.011601335369050503, + 0.016346372663974762, + 0.01585984230041504, + 0.01216469518840313, + 0.009285385720431805, + 0.017577767372131348, + 0.010697382502257824, + 0.004920565523207188, + 0.007770448923110962, + 0.00796419195830822, + 0.0063472106121480465, + 0.0067040338180959225, + 0.006253784988075495, + 0.01020185835659504, + 0.0041801040060818195, + 0.006560156121850014, + 0.003910104278475046, + 0.01138203777372837, + 0.00795869529247284, + 0.022708607837557793, + 0.008070307783782482, + 0.01227103266865015, + 0.007192395627498627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.006759989075362682 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that begins with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward while simultaneously pulling back to reveal a massive biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the structure composed of polished obsidian exoskeleton and damp, gleaming sinew with ribbed skeletal frameworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent veins beneath the black metal shell glow with a sickly, fluctuating turquoise light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do violent bursts of jagged white-blue lightning erupt from the structure's spires and strike downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813023492693901 + ], + "opposite_logprob_tokens": [ + -3.288130283355713 + ], + "logprob_reward": -0.03813023492693901, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813023492693901, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08798126131296158, + 0.05765441060066223, + 0.027671534568071365, + 0.17915105819702148, + 0.40371382236480713, + 0.04443838819861412, + 0.04569112882018089, + 0.05260413885116577, + 0.024541441351175308, + 0.07655277848243713 + ], + "credit_map": [ + 0.011784255504608154, + 0.01330468337982893, + 0.013269901275634766, + 0.012016159482300282, + 0.009953168220818043, + 0.010222267359495163, + 0.010968353599309921, + 0.006462473422288895, + 0.0096519123762846, + 0.006680600345134735, + 0.010743554681539536, + 0.01586948335170746, + 0.0036856273654848337, + 0.0027721174992620945, + 0.004512016195803881, + 0.0037390985526144505, + 0.0036579249426722527, + 0.0031959388870745897, + 0.0035336483269929886, + 0.004827751312404871, + 0.0038728038780391216, + 0.002327319933101535, + 0.003604304511100054, + 0.002651840914040804, + 0.01955413445830345, + 0.011471865698695183, + 0.016506943851709366, + 0.01819024048745632, + 0.014805172570049763, + 0.058009933680295944, + 0.009151529520750046, + 0.03146124631166458, + 0.030458780005574226, + 0.031463801860809326, + 0.06024696305394173, + 0.10654222965240479, + 0.023952962830662727, + 0.018655309453606606, + 0.09027925878763199, + 0.04211452975869179, + 0.005399958696216345, + 0.0048746210522949696, + 0.005680859088897705, + 0.0055050416849553585, + 0.006257445085793734, + 0.0037628819700330496, + 0.006690263282507658, + 0.006267317105084658, + 0.004399228375405073, + 0.005935500841587782, + 0.006276184227317572, + 0.00569329597055912, + 0.005924977362155914, + 0.006647764705121517, + 0.006338983308523893, + 0.00447519076988101, + 0.005174402147531509, + 0.007486672606319189, + 0.007980793714523315, + 0.006911674980074167, + 0.007444145157933235, + 0.004669891204684973, + 0.007249016314744949, + 0.005687539931386709, + 0.001993433339521289, + 0.002608837094157934, + 0.003821238875389099, + 0.0027252191212028265, + 0.0038000945933163166, + 0.0034149079583585262, + 0.003546575317159295, + 0.0026311350520700216, + 0.006622783839702606, + 0.0033200017642229795, + 0.011492112651467323, + 0.005861937999725342, + 0.024330316111445427, + 0.006272596772760153, + 0.011560698039829731, + 0.00709233433008194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.03813023492693901 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that begins with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward while simultaneously pulling back to reveal a massive biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the structure composed of polished obsidian exoskeleton and damp, gleaming sinew with ribbed skeletal frameworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent veins beneath the black metal shell glow with a sickly, fluctuating turquoise light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do violent bursts of jagged white-blue lightning erupt from the structure's spires and strike downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013369916705414653 + ], + "opposite_logprob_tokens": [ + -6.626337051391602 + ], + "logprob_reward": -0.0013369916705414653, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013369916705414653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10534843802452087, + 0.054909598082304, + 0.03146359696984291, + 0.08814182877540588, + 0.30539295077323914, + 0.04478656128048897, + 0.040280599147081375, + 0.07372947037220001, + 0.02788974530994892, + 0.2280571311712265 + ], + "credit_map": [ + 0.01422332413494587, + 0.013364552520215511, + 0.0194996427744627, + 0.01354293990880251, + 0.008843483403325081, + 0.00944455061107874, + 0.015069887042045593, + 0.011360055766999722, + 0.011616997420787811, + 0.005558409728109837, + 0.008930118754506111, + 0.009776141494512558, + 0.004314770456403494, + 0.006277808919548988, + 0.0037374142557382584, + 0.004697937518358231, + 0.003496692981570959, + 0.004209127277135849, + 0.003230629488825798, + 0.0051839109510183334, + 0.005262524355202913, + 0.00249926233664155, + 0.004417933989316225, + 0.0031635172199457884, + 0.007697757799178362, + 0.004689209628850222, + 0.008369498886168003, + 0.01615264266729355, + 0.008024447597563267, + 0.02809992805123329, + 0.005488595459610224, + 0.009619749151170254, + 0.021628523245453835, + 0.02447797730565071, + 0.04547429084777832, + 0.08061530441045761, + 0.020320026203989983, + 0.014497791416943073, + 0.06992009282112122, + 0.028458964079618454, + 0.005218117497861385, + 0.0039192973636090755, + 0.006390151102095842, + 0.005405503790825605, + 0.006406937725841999, + 0.003813488408923149, + 0.007676389068365097, + 0.005956678185611963, + 0.0045407782308757305, + 0.006244709715247154, + 0.005748760886490345, + 0.004155653528869152, + 0.0040380628779530525, + 0.00627875467762351, + 0.00505851348862052, + 0.004215369466692209, + 0.0073076230473816395, + 0.009989593178033829, + 0.011515311896800995, + 0.009356619790196419, + 0.008941883221268654, + 0.0070612989366054535, + 0.012369646690785885, + 0.007187497336417437, + 0.002500114729627967, + 0.003277479438111186, + 0.004785226657986641, + 0.0027236295863986015, + 0.004395677242428064, + 0.003441068110987544, + 0.004082242492586374, + 0.002684306586161256, + 0.020786209031939507, + 0.011215395294129848, + 0.031213872134685516, + 0.02024497091770172, + 0.06651865690946579, + 0.02280380390584469, + 0.03482966497540474, + 0.020444558933377266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.0013369916705414653 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that begins with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward while simultaneously pulling back to reveal a massive biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the structure composed of polished obsidian exoskeleton and damp, gleaming sinew with ribbed skeletal frameworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent veins beneath the black metal shell glow with a sickly, fluctuating turquoise light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do violent bursts of jagged white-blue lightning erupt from the structure's spires and strike downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864092543721199 + ], + "opposite_logprob_tokens": [ + -3.0486409664154053 + ], + "logprob_reward": -0.04864092543721199, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864092543721199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06103269383311272, + 0.0648728683590889, + 0.03432198241353035, + 0.045782968401908875, + 0.2531820833683014, + 0.026179319247603416, + 0.06742367893457413, + 0.09084056317806244, + 0.05522832274436951, + 0.30113545060157776 + ], + "credit_map": [ + 0.00809445045888424, + 0.010411135852336884, + 0.006801332347095013, + 0.006997109390795231, + 0.00694101769477129, + 0.008332718163728714, + 0.004958714824169874, + 0.008496219292283058, + 0.007207646034657955, + 0.009631742723286152, + 0.00686969980597496, + 0.02502433955669403, + 0.0033350882586091757, + 0.005963997449725866, + 0.0043809786438941956, + 0.0024593817070126534, + 0.004745753016322851, + 0.004338324069976807, + 0.0026168902404606342, + 0.010492230765521526, + 0.0030881070997565985, + 0.003773188916966319, + 0.0035286517813801765, + 0.0017388375708833337, + 0.00479454081505537, + 0.0037095218431204557, + 0.0075454749166965485, + 0.006051064468920231, + 0.0043221828527748585, + 0.01154402643442154, + 0.0034060098696500063, + 0.0044101532548666, + 0.017409203574061394, + 0.019864238798618317, + 0.038830049335956573, + 0.06867609918117523, + 0.015694430097937584, + 0.01201540045440197, + 0.058290038257837296, + 0.022402672097086906, + 0.003546379739418626, + 0.00270997267216444, + 0.0034078918397426605, + 0.0034083796199411154, + 0.0036446629092097282, + 0.0020990492776036263, + 0.004177086520940065, + 0.003185901092365384, + 0.006308324635028839, + 0.007168433163315058, + 0.00926930084824562, + 0.010350741446018219, + 0.007017545402050018, + 0.00752981286495924, + 0.012861528433859348, + 0.006917996797710657, + 0.010642210952937603, + 0.009642014279961586, + 0.013788179494440556, + 0.00930833164602518, + 0.011514654383063316, + 0.009570562280714512, + 0.014214194379746914, + 0.01216043159365654, + 0.005064943339675665, + 0.0072547076269984245, + 0.009191302582621574, + 0.006531288381665945, + 0.007334588561207056, + 0.00840521976351738, + 0.007031465880572796, + 0.004414813127368689, + 0.026166696101427078, + 0.015456986613571644, + 0.04035860672593117, + 0.02849975787103176, + 0.08238916099071503, + 0.034045424312353134, + 0.046553581953048706, + 0.027665266767144203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 7 + ], + "selected_reward": -0.04864092543721199 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_8/index_tensor([6894, 6894])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_8/index_tensor([6894, 6894])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..27cd2630d134fd594d72973c1054dfbeba1141d9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_8/index_tensor([6894, 6894])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "steady, static medium shot immerses the viewer in an animated watercolor world, capturing a tender moment of courtship against a warm, creamy, textured paper background. At the center of the frame stand two anthropomorphic forget-me-not flowers, their delicate, soft green stems atop elegant blossoms. Each flower features a simple, expressive black-dot eye and a subtle pink watercolor blush, delicately blossoming at the base of the petals. The petals themselves are a sky-blue, with the characteristic soft, bleeding edges of the wet-on-wet watercolor technique, making the translucent color appear to radiate an inner, ethereal glow. The flower on the left tentatively begins the movement, its stem slowly bending in a deliberate arc. The second flower responds with shy hesitation, its stem gently bending inward until its blue petals are just a breath away from touching. The entire painting, set against a dark green background and enhanced by the luminous effect of the paint, exudes a profound atmosphere of tenderness, innocence, and romance.", + "seed": 19189, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_8/index_6894_seed_19189_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03150347131304443, + "vlm_logprob": -0.03150347131304443, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.03150347131304443, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects anthropomorphic forget-me-not flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flowers have sky-blue petals with soft, bleeding watercolor edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower on the left begin moving by bending its stem in a deliberate arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second flower respond by bending its stem inward until the petals are nearly touching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05839075520634651 + ], + "opposite_logprob_tokens": [ + -3.3083908557891846 + ], + "logprob_reward": -0.05839075520634651, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.05839075520634651, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18948309123516083, + 0.1860160529613495, + 0.028143679723143578, + 0.02857377752661705, + 0.028582286089658737, + 0.1408674120903015, + 0.18116697669029236, + 0.11359769850969315, + 0.08277186006307602, + 0.020797189325094223 + ], + "credit_map": [ + 0.01931130886077881, + 0.017286747694015503, + 0.02680857479572296, + 0.015775566920638084, + 0.023089325055480003, + 0.00977487862110138, + 0.06780236214399338, + 0.009634331800043583, + 0.014463810250163078, + 0.008609329350292683, + 0.06645089387893677, + 0.013512598350644112, + 0.05515338480472565, + 0.008534214459359646, + 0.008233441971242428, + 0.011058381758630276, + 0.0031349514611065388, + 0.005211019888520241, + 0.003858521580696106, + 0.0037379369605332613, + 0.003168431343510747, + 0.0036153914406895638, + 0.0031012268736958504, + 0.0023162018042057753, + 0.0035247905179858208, + 0.0031345493625849485, + 0.00513286329805851, + 0.0029438838828355074, + 0.003711861791089177, + 0.00493910675868392, + 0.0025981110520660877, + 0.00258861412294209, + 0.0028329165652394295, + 0.0034715619403868914, + 0.0034097167663276196, + 0.004001358989626169, + 0.003970937803387642, + 0.004527230281382799, + 0.0039011500775814056, + 0.002467416925355792, + 0.011334439739584923, + 0.022276727482676506, + 0.013389933854341507, + 0.01283667329698801, + 0.03129720687866211, + 0.01849195547401905, + 0.0212191641330719, + 0.010021322406828403, + 0.013376693241298199, + 0.018002888187766075, + 0.04016779735684395, + 0.01524398848414421, + 0.020987147465348244, + 0.04350512474775314, + 0.017264466732740402, + 0.012618865817785263, + 0.009176873601973057, + 0.012204420752823353, + 0.016831325367093086, + 0.015206851065158844, + 0.020150767639279366, + 0.01577048748731613, + 0.014534462243318558, + 0.0097225122153759, + 0.008512875065207481, + 0.008252507075667381, + 0.012013932690024376, + 0.008939314633607864, + 0.016928505152463913, + 0.011199350468814373, + 0.009042681194841862, + 0.00788269191980362, + 0.002425626153126359, + 0.0027700213249772787, + 0.002492583589628339, + 0.002508452394977212, + 0.003503806423395872, + 0.0028517129831016064, + 0.0024186228401958942, + 0.0018263638485223055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.05839075520634651 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects anthropomorphic forget-me-not flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flowers have sky-blue petals with soft, bleeding watercolor edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower on the left begin moving by bending its stem in a deliberate arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second flower respond by bending its stem inward until the petals are nearly touching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002512633567675948 + ], + "opposite_logprob_tokens": [ + -6.002512454986572 + ], + "logprob_reward": -0.002512633567675948, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002512633567675948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32562366127967834, + 0.47097983956336975, + 0.03540504723787308, + 0.01594657078385353, + 0.015435570850968361, + 0.021011406555771828, + 0.045007817447185516, + 0.04703583940863609, + 0.012339194305241108, + 0.011215080507099628 + ], + "credit_map": [ + 0.014225508086383343, + 0.05919032543897629, + 0.043437059968709946, + 0.031148632988333702, + 0.04303070157766342, + 0.008629984222352505, + 0.11271785199642181, + 0.013243619352579117, + 0.03191383555531502, + 0.021476758643984795, + 0.17983031272888184, + 0.038860902190208435, + 0.12733586132526398, + 0.019907638430595398, + 0.019945090636610985, + 0.03170941025018692, + 0.0037388294003903866, + 0.004974220879375935, + 0.0038402124773710966, + 0.004641067702323198, + 0.005010216496884823, + 0.005135065875947475, + 0.0045173135586082935, + 0.0035481236409395933, + 0.002130479086190462, + 0.003367179073393345, + 0.0018183517968282104, + 0.0020436595659703016, + 0.001903629512526095, + 0.0019496126333251595, + 0.001270900946110487, + 0.0014627586351707578, + 0.0015067593194544315, + 0.0021754135377705097, + 0.0016215884825214744, + 0.0022508129477500916, + 0.0029437222983688116, + 0.0016574776964262128, + 0.002260442590340972, + 0.0010193546768277884, + 0.0016585428966209292, + 0.0033282043877989054, + 0.002435513073578477, + 0.0023166604805737734, + 0.00426754355430603, + 0.0026815179735422134, + 0.0027010715566575527, + 0.0016223517013713717, + 0.0030063875019550323, + 0.007956478744745255, + 0.007664941251277924, + 0.003796441713348031, + 0.0077006942592561245, + 0.0091816121712327, + 0.0033818071242421865, + 0.002319451654329896, + 0.004445085767656565, + 0.0050063529051840305, + 0.006708936300128698, + 0.005815420765429735, + 0.008413968607783318, + 0.006716557778418064, + 0.005727946758270264, + 0.004201567266136408, + 0.0013589077861979604, + 0.0011367470724508166, + 0.0017468250589445233, + 0.0011577745899558067, + 0.0021767420694231987, + 0.0018843046855181456, + 0.0013388023944571614, + 0.0015390896005555987, + 0.0013685196172446012, + 0.0010660304687917233, + 0.0014936740044504404, + 0.001237956341356039, + 0.0016479199985042214, + 0.0017416554037481546, + 0.00120603630784899, + 0.0014532883651554585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.002512633567675948 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects anthropomorphic forget-me-not flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flowers have sky-blue petals with soft, bleeding watercolor edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower on the left begin moving by bending its stem in a deliberate arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second flower respond by bending its stem inward until the petals are nearly touching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029793230816721916 + ], + "opposite_logprob_tokens": [ + -3.5297932624816895 + ], + "logprob_reward": -0.029793230816721916, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029793230816721916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2575688660144806, + 0.45228421688079834, + 0.03312060236930847, + 0.028853727504611015, + 0.05284934490919113, + 0.04292915388941765, + 0.0461990050971508, + 0.024331605061888695, + 0.0409819521009922, + 0.020881474018096924 + ], + "credit_map": [ + 0.017399422824382782, + 0.061333704739809036, + 0.01483209989964962, + 0.0199800543487072, + 0.01977338083088398, + 0.009818360209465027, + 0.1010175570845604, + 0.013414306566119194, + 0.031692322343587875, + 0.02238272875547409, + 0.16955333948135376, + 0.03646797314286232, + 0.12243965268135071, + 0.020113248378038406, + 0.01890345849096775, + 0.030731558799743652, + 0.0034101021010428667, + 0.0040375059470534325, + 0.006022279616445303, + 0.00521148182451725, + 0.003835513722151518, + 0.0033224194776266813, + 0.004124754574149847, + 0.0031565509270876646, + 0.002908099675551057, + 0.003924396354705095, + 0.003549249842762947, + 0.004250079859048128, + 0.00382993440143764, + 0.003749881638213992, + 0.004205025732517242, + 0.0024370618630200624, + 0.004229499027132988, + 0.006912988144904375, + 0.008751633577048779, + 0.004353734198957682, + 0.006314628291875124, + 0.01400480605661869, + 0.004970477428287268, + 0.003311585634946823, + 0.0031570959836244583, + 0.006377936340868473, + 0.004494144581258297, + 0.004407379776239395, + 0.009693056344985962, + 0.006480855401605368, + 0.005572346970438957, + 0.002746344543993473, + 0.002899629296734929, + 0.0065734051167964935, + 0.005696077365428209, + 0.006079761311411858, + 0.008714238181710243, + 0.007558281533420086, + 0.005435475613921881, + 0.003242145525291562, + 0.0026016579940915108, + 0.0027635684236884117, + 0.0034245681017637253, + 0.0035246533807367086, + 0.003517706645652652, + 0.0031013023108243942, + 0.0033439381513744593, + 0.002054212149232626, + 0.003729779040440917, + 0.0037522383499890566, + 0.006187618710100651, + 0.0066832005977630615, + 0.006230133585631847, + 0.005394536070525646, + 0.005878908559679985, + 0.0031255397479981184, + 0.00334137468598783, + 0.0018025842728093266, + 0.0034134243614971638, + 0.002040678635239601, + 0.002787280362099409, + 0.00316599034704268, + 0.002496970584616065, + 0.0018331744940951467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.029793230816721916 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects anthropomorphic forget-me-not flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flowers have sky-blue petals with soft, bleeding watercolor edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower on the left begin moving by bending its stem in a deliberate arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second flower respond by bending its stem inward until the petals are nearly touching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018179811537265778 + ], + "opposite_logprob_tokens": [ + -4.018179893493652 + ], + "logprob_reward": -0.018179811537265778, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018179811537265778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19953036308288574, + 0.3499002158641815, + 0.045336078852415085, + 0.05358114838600159, + 0.057950377464294434, + 0.053836461156606674, + 0.13053031265735626, + 0.03966759145259857, + 0.03953632712364197, + 0.030131103470921516 + ], + "credit_map": [ + 0.020736301317811012, + 0.04747582972049713, + 0.04531559720635414, + 0.017226191237568855, + 0.027622733265161514, + 0.012521597556769848, + 0.014039565809071064, + 0.014592565596103668, + 0.024817248806357384, + 0.01825772225856781, + 0.1300058215856552, + 0.02837359718978405, + 0.09374191612005234, + 0.015893856063485146, + 0.014414834789931774, + 0.02439524047076702, + 0.004596605431288481, + 0.005565109197050333, + 0.006153132766485214, + 0.006826762575656176, + 0.006272828206419945, + 0.005128702614456415, + 0.0058456966653466225, + 0.004947244189679623, + 0.006019840482622385, + 0.008378232829272747, + 0.005941967014223337, + 0.005868429318070412, + 0.009083438664674759, + 0.007143788505345583, + 0.0063803051598370075, + 0.004765148274600506, + 0.004714552778750658, + 0.0096146734431386, + 0.008683695457875729, + 0.0056532458402216434, + 0.006672980263829231, + 0.013639704324305058, + 0.005510388873517513, + 0.0034611360169947147, + 0.004771793726831675, + 0.007534768898040056, + 0.005186212714761496, + 0.0052685728296637535, + 0.01341207418590784, + 0.007367107085883617, + 0.006480678450316191, + 0.003815256292000413, + 0.009461133740842342, + 0.022423431277275085, + 0.01400400884449482, + 0.014905661344528198, + 0.02166767418384552, + 0.02082664892077446, + 0.017558667808771133, + 0.009683087468147278, + 0.0037878714501857758, + 0.004581742454320192, + 0.005659089889377356, + 0.004418834578245878, + 0.007362421602010727, + 0.00633408036082983, + 0.005442507565021515, + 0.002081046812236309, + 0.003020544070750475, + 0.003851718734949827, + 0.005847929511219263, + 0.006186133716255426, + 0.006805073004215956, + 0.0051214806735515594, + 0.006047356873750687, + 0.0026560891419649124, + 0.002439850475639105, + 0.0035201734863221645, + 0.00488805677741766, + 0.0034049227833747864, + 0.003870207816362381, + 0.004738344345241785, + 0.00440680468454957, + 0.0028627451974898577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.018179811537265778 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects anthropomorphic forget-me-not flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flowers have sky-blue petals with soft, bleeding watercolor edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower on the left begin moving by bending its stem in a deliberate arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second flower respond by bending its stem inward until the petals are nearly touching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864092543721199 + ], + "opposite_logprob_tokens": [ + -3.0486409664154053 + ], + "logprob_reward": -0.04864092543721199, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864092543721199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3404533565044403, + 0.2526668608188629, + 0.05210570991039276, + 0.042198244482278824, + 0.07162533700466156, + 0.07651771605014801, + 0.0637708529829979, + 0.0346989631652832, + 0.03514308109879494, + 0.03081979788839817 + ], + "credit_map": [ + 0.02214214764535427, + 0.03436994552612305, + 0.06171624734997749, + 0.032566506415605545, + 0.06550778448581696, + 0.01097080484032631, + 0.09698006510734558, + 0.016199897974729538, + 0.01982187107205391, + 0.014133128337562084, + 0.09882382303476334, + 0.023537764325737953, + 0.054040342569351196, + 0.011864733882248402, + 0.012031521648168564, + 0.018413713201880455, + 0.006156786810606718, + 0.005222439765930176, + 0.00937082152813673, + 0.010781874880194664, + 0.004674961790442467, + 0.0037516301963478327, + 0.007397882640361786, + 0.004749319050461054, + 0.004857573192566633, + 0.00758754089474678, + 0.005334876477718353, + 0.005215596407651901, + 0.005697908811271191, + 0.004994310904294252, + 0.005207783076912165, + 0.0033026610035449266, + 0.005985700059682131, + 0.010678757913410664, + 0.006988088134676218, + 0.009922842495143414, + 0.014453638345003128, + 0.008663520216941833, + 0.009371723048388958, + 0.005561074707657099, + 0.005998372100293636, + 0.01185534056276083, + 0.008330773562192917, + 0.007702738977968693, + 0.015331701375544071, + 0.011542992666363716, + 0.010255902074277401, + 0.005499905440956354, + 0.004497944377362728, + 0.006289507262408733, + 0.014272812753915787, + 0.005784417502582073, + 0.005586063954979181, + 0.015714257955551147, + 0.005654019303619862, + 0.005971839651465416, + 0.0031746174208819866, + 0.0033515910618007183, + 0.004550569225102663, + 0.004773185588419437, + 0.006010136567056179, + 0.004694498144090176, + 0.004844230599701405, + 0.003300135023891926, + 0.004237255547195673, + 0.0028160696383565664, + 0.004581039305776358, + 0.00478655518963933, + 0.004603623412549496, + 0.004812056664377451, + 0.00554726691916585, + 0.0037592200096696615, + 0.0036795984487980604, + 0.003250017063692212, + 0.005777824204415083, + 0.003525692503899336, + 0.004378659185022116, + 0.004736226983368397, + 0.003040320472791791, + 0.0024314604233950377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.04864092543721199 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, static medium shot immerses the viewer in an animated watercolor world, capturing a tender moment of courtship against a warm, creamy, textured paper background. At the center of the frame stand two anthropomorphic forget-me-not flowers, their delicate, soft green stems atop elegant blossoms. Each flower features a simple, expressive black-dot eye and a subtle pink watercolor blush, delicately blossoming at the base of the petals. The petals themselves are a sky-blue, with the characteristic soft, bleeding edges of the wet-on-wet watercolor technique, making the translucent color appear to radiate an inner, ethereal glow. The flower on the left tentatively begins the movement, its stem slowly bending in a deliberate arc. The second flower responds with shy hesitation, its stem gently bending inward until its blue petals are just a breath away from touching. The entire painting, set against a dark green background and enhanced by the luminous effect of the paint, exudes a profound atmosphere of tenderness, innocence, and romance.", + "seed": 19190, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_8/index_6894_seed_19190_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05955014042556286, + "vlm_logprob": -0.05955014042556286, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.05955014042556286, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects anthropomorphic forget-me-not flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flowers have sky-blue petals with soft, bleeding watercolor edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower on the left begin moving by bending its stem in a deliberate arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second flower respond by bending its stem inward until the petals are nearly touching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20453330874443054 + ], + "opposite_logprob_tokens": [ + -1.829533338546753 + ], + "logprob_reward": -0.20453330874443054, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.20453330874443054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09517992287874222, + 0.18456321954727173, + 0.04888830706477165, + 0.152931809425354, + 0.028857365250587463, + 0.08271007239818573, + 0.08163260668516159, + 0.14807501435279846, + 0.030553674325346947, + 0.14660796523094177 + ], + "credit_map": [ + 0.014186776243150234, + 0.007064038887619972, + 0.014156408607959747, + 0.009167680516839027, + 0.00900992751121521, + 0.010510478168725967, + 0.013653201051056385, + 0.017431417480111122, + 0.032477471977472305, + 0.014184803701937199, + 0.029117854312062263, + 0.018774408847093582, + 0.01604396663606167, + 0.031250499188899994, + 0.017033331096172333, + 0.025680892169475555, + 0.0066633522510528564, + 0.00634091766551137, + 0.0045823692344129086, + 0.0049919914454221725, + 0.009817147627472878, + 0.007258478552103043, + 0.005053466651588678, + 0.004180582240223885, + 0.01982998475432396, + 0.022119062021374702, + 0.012555389665067196, + 0.012209492735564709, + 0.03770919516682625, + 0.018978310748934746, + 0.012235713191330433, + 0.017294663935899734, + 0.004456477239727974, + 0.003043258795514703, + 0.00326713640242815, + 0.0028573162853717804, + 0.003957326989620924, + 0.004511354956775904, + 0.003484841203317046, + 0.0032796526793390512, + 0.017049232497811317, + 0.003745320253074169, + 0.011461208574473858, + 0.0050869295373559, + 0.0036907673347741365, + 0.011119607836008072, + 0.0054860832169651985, + 0.025070924311876297, + 0.014566008001565933, + 0.013776648789644241, + 0.0068286629393696785, + 0.004562898073345423, + 0.014058375731110573, + 0.007082813885062933, + 0.0056355614215135574, + 0.015121636912226677, + 0.049190714955329895, + 0.008268469013273716, + 0.011149651370942593, + 0.007286985404789448, + 0.0048316544853150845, + 0.013679943047463894, + 0.007763158064335585, + 0.04590443894267082, + 0.0054868473671376705, + 0.0038368464447557926, + 0.0025730743072927, + 0.0018447734182700515, + 0.006643976550549269, + 0.0036777991335839033, + 0.0019232167396694422, + 0.0045671421103179455, + 0.04220784455537796, + 0.011895019561052322, + 0.00873754732310772, + 0.006626071874052286, + 0.016199111938476562, + 0.010190799832344055, + 0.008104830048978329, + 0.04264673963189125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 7 + ], + "selected_reward": -0.20453330874443054 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects anthropomorphic forget-me-not flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flowers have sky-blue petals with soft, bleeding watercolor edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower on the left begin moving by bending its stem in a deliberate arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second flower respond by bending its stem inward until the petals are nearly touching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014237492345273495 + ], + "opposite_logprob_tokens": [ + -4.264237403869629 + ], + "logprob_reward": -0.014237492345273495, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014237492345273495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02973097935318947, + 0.22564370930194855, + 0.03132540360093117, + 0.08281846344470978, + 0.059588223695755005, + 0.03534289449453354, + 0.08179274946451187, + 0.37354519963264465, + 0.04957672953605652, + 0.030635662376880646 + ], + "credit_map": [ + 0.003963039256632328, + 0.002795260865241289, + 0.0041743735782802105, + 0.0031944597139954567, + 0.004906190559267998, + 0.0037281119730323553, + 0.0029791018459945917, + 0.003990438766777515, + 0.03432135656476021, + 0.015011253766715527, + 0.04334953799843788, + 0.02868209034204483, + 0.020249223336577415, + 0.02779964916408062, + 0.02058996446430683, + 0.03564060851931572, + 0.004875333048403263, + 0.0034809282515197992, + 0.004089890513569117, + 0.00378023530356586, + 0.004363427869975567, + 0.003937718458473682, + 0.0037295599468052387, + 0.003068304155021906, + 0.012053939513862133, + 0.012282179668545723, + 0.006975881289690733, + 0.006157953757792711, + 0.02057352289557457, + 0.010117292404174805, + 0.006428270600736141, + 0.00822941679507494, + 0.008346601389348507, + 0.006151226814836264, + 0.0067884954623878, + 0.005212519783526659, + 0.0093305092304945, + 0.007744639180600643, + 0.005703001283109188, + 0.010311223566532135, + 0.00548563152551651, + 0.0024038117844611406, + 0.005086192395538092, + 0.002862024586647749, + 0.0033646621741354465, + 0.004604050423949957, + 0.003049049759283662, + 0.008487465791404247, + 0.019918829202651978, + 0.0071115014143288136, + 0.0065438211895525455, + 0.0042349230498075485, + 0.007599721197038889, + 0.007183996960520744, + 0.004794407170265913, + 0.02440553903579712, + 0.12252822518348694, + 0.026089509949088097, + 0.026200268417596817, + 0.016805563122034073, + 0.024426106363534927, + 0.034966841340065, + 0.016678303480148315, + 0.10585037618875504, + 0.011622588150203228, + 0.004441668279469013, + 0.0039231800474226475, + 0.00322930165566504, + 0.007323121186345816, + 0.006775014568120241, + 0.002803742652758956, + 0.009458106942474842, + 0.007976708933711052, + 0.0034653141628950834, + 0.0019549026619642973, + 0.001536110183224082, + 0.003688504919409752, + 0.0019213779596611857, + 0.001775969984009862, + 0.008316769264638424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 3 + ], + "selected_reward": -0.014237492345273495 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects anthropomorphic forget-me-not flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flowers have sky-blue petals with soft, bleeding watercolor edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower on the left begin moving by bending its stem in a deliberate arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second flower respond by bending its stem inward until the petals are nearly touching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029807578772306442 + ], + "opposite_logprob_tokens": [ + -3.5298075675964355 + ], + "logprob_reward": -0.029807578772306442, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029807578772306442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05598977953195572, + 0.1400802582502365, + 0.03530866652727127, + 0.052350159734487534, + 0.042483873665332794, + 0.03177974373102188, + 0.09076943248510361, + 0.039396267384290695, + 0.0409020259976387, + 0.4709398150444031 + ], + "credit_map": [ + 0.008920230902731419, + 0.005275070667266846, + 0.007611303590238094, + 0.004898815881460905, + 0.007693457882851362, + 0.0068891472183167934, + 0.00565136456862092, + 0.00905038695782423, + 0.021373512223362923, + 0.011218897998332977, + 0.018668541684746742, + 0.014793618582189083, + 0.01789083145558834, + 0.020312052220106125, + 0.012791708111763, + 0.02303108014166355, + 0.006229381076991558, + 0.0037311268970370293, + 0.0040681189857423306, + 0.0032504999544471502, + 0.005169549956917763, + 0.00483060535043478, + 0.003562837140634656, + 0.00446654437109828, + 0.009873801842331886, + 0.0058337063528597355, + 0.004914328921586275, + 0.00397880095988512, + 0.00984109751880169, + 0.005980730056762695, + 0.005108451005071402, + 0.00681923795491457, + 0.006426156964153051, + 0.0031780675053596497, + 0.00663173571228981, + 0.004401603247970343, + 0.004005264490842819, + 0.004610841162502766, + 0.004922753199934959, + 0.008307447656989098, + 0.00584354717284441, + 0.0026694482658058405, + 0.004070109222084284, + 0.0028678569942712784, + 0.0036919903941452503, + 0.004802824929356575, + 0.0037533175200223923, + 0.004080647137016058, + 0.016039032489061356, + 0.011697575449943542, + 0.006002908572554588, + 0.005653695203363895, + 0.01630742847919464, + 0.011016963049769402, + 0.006294058170169592, + 0.01775776594877243, + 0.009162741713225842, + 0.004923161584883928, + 0.004132953472435474, + 0.003277552779763937, + 0.004876420833170414, + 0.0048039318062365055, + 0.0022032244596630335, + 0.0060162791050970554, + 0.008056233637034893, + 0.005253519397228956, + 0.004458296578377485, + 0.002852260833606124, + 0.006580025888979435, + 0.004984804429113865, + 0.0026111137121915817, + 0.006105768959969282, + 0.14678961038589478, + 0.03283232823014259, + 0.02652992121875286, + 0.02121368795633316, + 0.040664613246917725, + 0.03204764053225517, + 0.02427317202091217, + 0.14658884704113007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 6 + ], + "selected_reward": -0.029807578772306442 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects anthropomorphic forget-me-not flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flowers have sky-blue petals with soft, bleeding watercolor edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower on the left begin moving by bending its stem in a deliberate arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second flower respond by bending its stem inward until the petals are nearly touching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076628230512142 + ], + "opposite_logprob_tokens": [ + -4.5110764503479 + ], + "logprob_reward": -0.011076628230512142, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076628230512142, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04840179532766342, + 0.15617799758911133, + 0.031436797231435776, + 0.11616609990596771, + 0.020751547068357468, + 0.0439145527780056, + 0.04963301494717598, + 0.07385329902172089, + 0.02313336171209812, + 0.43653154373168945 + ], + "credit_map": [ + 0.0075391121208667755, + 0.0031114299781620502, + 0.007086972706019878, + 0.00425995048135519, + 0.0035528100561350584, + 0.005225364584475756, + 0.006451049819588661, + 0.011175104416906834, + 0.027043644338846207, + 0.009829316288232803, + 0.027917107567191124, + 0.019062688574194908, + 0.012757085263729095, + 0.021529745310544968, + 0.013722024857997894, + 0.024316390976309776, + 0.005770182702690363, + 0.003204049775376916, + 0.003766312962397933, + 0.0031049379613250494, + 0.004952718038111925, + 0.0034351462963968515, + 0.0029710486996918917, + 0.004232401493936777, + 0.01525166630744934, + 0.016229810193181038, + 0.009861552156507969, + 0.008932916447520256, + 0.028034809976816177, + 0.014177649281919003, + 0.008797948248684406, + 0.014879748225212097, + 0.003482786240056157, + 0.0011934252688661218, + 0.003363141557201743, + 0.0022470715921372175, + 0.0018963487818837166, + 0.002249273005872965, + 0.002945305546745658, + 0.0033741947263479233, + 0.00863576214760542, + 0.0018405653536319733, + 0.005866553634405136, + 0.0031703764107078314, + 0.0026198418345302343, + 0.0056701176799833775, + 0.00402381457388401, + 0.012087520211935043, + 0.007071014493703842, + 0.009386533871293068, + 0.004167673643678427, + 0.004245189484208822, + 0.008987889625132084, + 0.00692481454461813, + 0.0038719603326171637, + 0.00497793685644865, + 0.02300158701837063, + 0.006522448267787695, + 0.0052058398723602295, + 0.003225102089345455, + 0.00626874016597867, + 0.006802815943956375, + 0.0032640043646097183, + 0.01956275850534439, + 0.005724285263568163, + 0.002184197073802352, + 0.0024770074523985386, + 0.0016266594175249338, + 0.0016435690922662616, + 0.00284980284050107, + 0.0011913137277588248, + 0.005436527542769909, + 0.14828377962112427, + 0.022937292233109474, + 0.0258291345089674, + 0.019500572234392166, + 0.017969181761145592, + 0.031016914173960686, + 0.02098231203854084, + 0.1500123292207718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 3 + ], + "selected_reward": -0.011076628230512142 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects anthropomorphic forget-me-not flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flowers have sky-blue petals with soft, bleeding watercolor edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower on the left begin moving by bending its stem in a deliberate arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second flower respond by bending its stem inward until the petals are nearly touching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809569403529167 + ], + "opposite_logprob_tokens": [ + -3.288095712661743 + ], + "logprob_reward": -0.03809569403529167, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809569403529167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07179740071296692, + 0.06445327401161194, + 0.05286899954080582, + 0.21847595274448395, + 0.02723504602909088, + 0.036421407014131546, + 0.02228757180273533, + 0.37885981798171997, + 0.04138626158237457, + 0.08621421456336975 + ], + "credit_map": [ + 0.009528356604278088, + 0.006172195542603731, + 0.009481828659772873, + 0.009335899725556374, + 0.008130614645779133, + 0.007674580439925194, + 0.010295547544956207, + 0.011178377084434032, + 0.011025087907910347, + 0.004208874423056841, + 0.009104968048632145, + 0.007935208268463612, + 0.00491553358733654, + 0.00978785753250122, + 0.00826407503336668, + 0.009211666882038116, + 0.0074835242703557014, + 0.005599559750407934, + 0.006100949365645647, + 0.0075395358726382256, + 0.00953496154397726, + 0.00631854310631752, + 0.005530304741114378, + 0.004761621821671724, + 0.02727437950670719, + 0.028762836009263992, + 0.020832205191254616, + 0.01854494959115982, + 0.049042221158742905, + 0.025961723178625107, + 0.01909552700817585, + 0.02896210551261902, + 0.004609598312526941, + 0.002904272871091962, + 0.00328391557559371, + 0.0022369439247995615, + 0.004150029271841049, + 0.0031752698123455048, + 0.0028483013156801462, + 0.004026714246720076, + 0.007205995265394449, + 0.001978790620341897, + 0.0048441714607179165, + 0.0025553482118993998, + 0.002627371344715357, + 0.004445548634976149, + 0.003045111894607544, + 0.009719070978462696, + 0.003734010271728039, + 0.0023281266912817955, + 0.0031584484968334436, + 0.0018825067672878504, + 0.0025304697919636965, + 0.002681957557797432, + 0.001699793734587729, + 0.004272259306162596, + 0.12813760340213776, + 0.02394583635032177, + 0.0273340605199337, + 0.01741345040500164, + 0.018176276236772537, + 0.034555584192276, + 0.017690643668174744, + 0.11160638928413391, + 0.010981295257806778, + 0.0025985834654420614, + 0.007597639691084623, + 0.0026707840152084827, + 0.0028886504005640745, + 0.003569088177755475, + 0.001968949567526579, + 0.009111268445849419, + 0.029259037226438522, + 0.004857535008341074, + 0.004569221753627062, + 0.003803481347858906, + 0.005304028280079365, + 0.005912069231271744, + 0.003923304378986359, + 0.028585534542798996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 9 + ], + "selected_reward": -0.03809569403529167 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, static medium shot immerses the viewer in an animated watercolor world, capturing a tender moment of courtship against a warm, creamy, textured paper background. At the center of the frame stand two anthropomorphic forget-me-not flowers, their delicate, soft green stems atop elegant blossoms. Each flower features a simple, expressive black-dot eye and a subtle pink watercolor blush, delicately blossoming at the base of the petals. The petals themselves are a sky-blue, with the characteristic soft, bleeding edges of the wet-on-wet watercolor technique, making the translucent color appear to radiate an inner, ethereal glow. The flower on the left tentatively begins the movement, its stem slowly bending in a deliberate arc. The second flower responds with shy hesitation, its stem gently bending inward until its blue petals are just a breath away from touching. The entire painting, set against a dark green background and enhanced by the luminous effect of the paint, exudes a profound atmosphere of tenderness, innocence, and romance.", + "seed": 19191, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_8/index_6894_seed_19191_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.027193028200417756, + "vlm_logprob": -0.027193028200417756, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.027193028200417756, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects anthropomorphic forget-me-not flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flowers have sky-blue petals with soft, bleeding watercolor edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower on the left begin moving by bending its stem in a deliberate arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second flower respond by bending its stem inward until the petals are nearly touching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.032304685562849045 + ], + "opposite_logprob_tokens": [ + -4.282304763793945 + ], + "logprob_reward": -0.032304685562849045, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.032304685562849045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10669536143541336, + 0.06010652706027031, + 0.06067432835698128, + 0.06266360729932785, + 0.07651399821043015, + 0.15418830513954163, + 0.1709761619567871, + 0.21537333726882935, + 0.048579324036836624, + 0.04422902315855026 + ], + "credit_map": [ + 0.015567189082503319, + 0.012568581849336624, + 0.013526063412427902, + 0.017839781939983368, + 0.011189941316843033, + 0.009700282476842403, + 0.016814887523651123, + 0.009488641284406185, + 0.008283759467303753, + 0.007867483422160149, + 0.007961414754390717, + 0.009543592110276222, + 0.005184731911867857, + 0.0073777963407337666, + 0.007752625271677971, + 0.006135126110166311, + 0.008353427983820438, + 0.007301978766918182, + 0.004985842853784561, + 0.015775902196764946, + 0.005906874313950539, + 0.006505953148007393, + 0.008083616383373737, + 0.003760727820917964, + 0.014322999864816666, + 0.0083154933527112, + 0.004014357924461365, + 0.007605067919939756, + 0.005776622798293829, + 0.013678953051567078, + 0.005267593543976545, + 0.0036825165152549744, + 0.016339097172021866, + 0.009216331876814365, + 0.006119847297668457, + 0.00975709781050682, + 0.006951179355382919, + 0.013513104058802128, + 0.009350487031042576, + 0.005266852676868439, + 0.03661658242344856, + 0.01635778322815895, + 0.010919347405433655, + 0.020154623314738274, + 0.013318111188709736, + 0.03586011379957199, + 0.013362846337258816, + 0.007598885800689459, + 0.021986935287714005, + 0.022152338176965714, + 0.01930682733654976, + 0.0328499972820282, + 0.018635988235473633, + 0.021683838218450546, + 0.022266460582613945, + 0.012093786150217056, + 0.06584767997264862, + 0.015565529465675354, + 0.013563345186412334, + 0.009663710370659828, + 0.010610345751047134, + 0.0790872871875763, + 0.009903828613460064, + 0.011131622828543186, + 0.007852130569517612, + 0.0034293709322810173, + 0.006809834856539965, + 0.005561788100749254, + 0.002910017967224121, + 0.010783471167087555, + 0.004918496590107679, + 0.006314214784651995, + 0.007602388970553875, + 0.003143072361126542, + 0.004375843331217766, + 0.004226666409522295, + 0.005166799295693636, + 0.00981938187032938, + 0.004010635428130627, + 0.005884239450097084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.032304685562849045 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects anthropomorphic forget-me-not flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flowers have sky-blue petals with soft, bleeding watercolor edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower on the left begin moving by bending its stem in a deliberate arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second flower respond by bending its stem inward until the petals are nearly touching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005271582398563623 + ], + "opposite_logprob_tokens": [ + -5.2552714347839355 + ], + "logprob_reward": -0.005271582398563623, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005271582398563623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04607207328081131, + 0.02998657152056694, + 0.030688507482409477, + 0.013960018754005432, + 0.05024926736950874, + 0.06714913249015808, + 0.0817674770951271, + 0.5966193675994873, + 0.06674209982156754, + 0.0167655311524868 + ], + "credit_map": [ + 0.00838526152074337, + 0.0061417133547365665, + 0.003405763301998377, + 0.009337359108030796, + 0.004508850164711475, + 0.004762150347232819, + 0.006210289429873228, + 0.003320680232718587, + 0.005441945046186447, + 0.0036179465241730213, + 0.0025699075777083635, + 0.006378308869898319, + 0.0024778973311185837, + 0.0036711019929498434, + 0.003989106975495815, + 0.0018403533613309264, + 0.009890582412481308, + 0.0023357877507805824, + 0.0021934746764600277, + 0.002023695968091488, + 0.0018659866182133555, + 0.009119904600083828, + 0.0017056041397154331, + 0.0015534680569544435, + 0.003040170995518565, + 0.0015219252090901136, + 0.0009404941229149699, + 0.0020368609111756086, + 0.0013009783579036593, + 0.003111873287707567, + 0.0013220988912507892, + 0.0006856158142909408, + 0.012918689288198948, + 0.004917484242469072, + 0.0037439526058733463, + 0.0048035867512226105, + 0.003990186844021082, + 0.012389847077429295, + 0.0043118721805512905, + 0.0031736446544528008, + 0.01769173890352249, + 0.006470737978816032, + 0.003998309373855591, + 0.006769661325961351, + 0.005951808299869299, + 0.018123066052794456, + 0.004669243469834328, + 0.0034745645243674517, + 0.010279549285769463, + 0.01114137563854456, + 0.010201210156083107, + 0.013126706704497337, + 0.011123941279947758, + 0.00884720217436552, + 0.009132003411650658, + 0.007915478199720383, + 0.18756680190563202, + 0.038803406059741974, + 0.04183980077505112, + 0.020317567512392998, + 0.023398112505674362, + 0.2332843542098999, + 0.019277814775705338, + 0.03213149309158325, + 0.013768002390861511, + 0.005844383500516415, + 0.005533618852496147, + 0.006074483040720224, + 0.005262476392090321, + 0.019212061539292336, + 0.00687666330486536, + 0.004170403815805912, + 0.002288449788466096, + 0.0015313589246943593, + 0.0022621348034590483, + 0.0021229966077953577, + 0.001913185347802937, + 0.002824756084010005, + 0.0015866252360865474, + 0.0022360223811119795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.005271582398563623 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects anthropomorphic forget-me-not flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flowers have sky-blue petals with soft, bleeding watercolor edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower on the left begin moving by bending its stem in a deliberate arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second flower respond by bending its stem inward until the petals are nearly touching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863570258021355 + ], + "opposite_logprob_tokens": [ + -3.048635721206665 + ], + "logprob_reward": -0.04863570258021355, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863570258021355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2781949043273926, + 0.037025969475507736, + 0.09217808395624161, + 0.06118131801486015, + 0.05937015265226364, + 0.16526676714420319, + 0.13885536789894104, + 0.08690670877695084, + 0.03126518428325653, + 0.049755509942770004 + ], + "credit_map": [ + 0.03509256988763809, + 0.029404591768980026, + 0.025610992684960365, + 0.07854442298412323, + 0.028258798643946648, + 0.02720119245350361, + 0.036310113966464996, + 0.01777222752571106, + 0.00540956761687994, + 0.0061651188880205154, + 0.0030369276646524668, + 0.0063026235438883305, + 0.00489381980150938, + 0.00406448170542717, + 0.004646122921258211, + 0.0025073059368878603, + 0.016366947442293167, + 0.009665880352258682, + 0.005951251834630966, + 0.018803779035806656, + 0.008465287275612354, + 0.01665682904422283, + 0.010654376819729805, + 0.005613731220364571, + 0.007689171005040407, + 0.008267635479569435, + 0.00628940062597394, + 0.009303320199251175, + 0.009121726267039776, + 0.005260647274553776, + 0.010151887312531471, + 0.005097538232803345, + 0.01185467466711998, + 0.008213221095502377, + 0.005296765826642513, + 0.00832856260240078, + 0.005595754366368055, + 0.009591727517545223, + 0.006979419384151697, + 0.0035100311506539583, + 0.02384870871901512, + 0.02134755067527294, + 0.014819824136793613, + 0.032345596700906754, + 0.022443581372499466, + 0.01901869662106037, + 0.02079165168106556, + 0.010651161894202232, + 0.022098585963249207, + 0.017749032005667686, + 0.016105206683278084, + 0.02070651948451996, + 0.015902234241366386, + 0.02083093486726284, + 0.013451186008751392, + 0.012011665850877762, + 0.02587711438536644, + 0.006456801667809486, + 0.006232377141714096, + 0.0038587464950978756, + 0.00494073610752821, + 0.030160097405314445, + 0.003818987403064966, + 0.005561863072216511, + 0.007350965868681669, + 0.0022140529472380877, + 0.003024884732440114, + 0.00243255659006536, + 0.0015547833172604442, + 0.010318813845515251, + 0.0017119119875133038, + 0.0026572186034172773, + 0.010703746229410172, + 0.003928049001842737, + 0.004288290627300739, + 0.004081254824995995, + 0.004296133294701576, + 0.01500549167394638, + 0.00360357528552413, + 0.0038489745929837227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.04863570258021355 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects anthropomorphic forget-me-not flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flowers have sky-blue petals with soft, bleeding watercolor edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower on the left begin moving by bending its stem in a deliberate arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second flower respond by bending its stem inward until the petals are nearly touching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606578752398491 + ], + "opposite_logprob_tokens": [ + -4.14106559753418 + ], + "logprob_reward": -0.01606578752398491, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606578752398491, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24218159914016724, + 0.029983792454004288, + 0.05439457297325134, + 0.05250178277492523, + 0.05360505357384682, + 0.2957490086555481, + 0.16997797787189484, + 0.043833423405885696, + 0.02799607440829277, + 0.029776766896247864 + ], + "credit_map": [ + 0.02976379171013832, + 0.024897407740354538, + 0.022773927077651024, + 0.06848983466625214, + 0.024173075333237648, + 0.02408863790333271, + 0.03145569935441017, + 0.016539203003048897, + 0.003914669621735811, + 0.0033983837347477674, + 0.0021623580250889063, + 0.00719196442514658, + 0.0027719405479729176, + 0.0030192381236702204, + 0.0046738432720303535, + 0.0028513933066278696, + 0.011231490410864353, + 0.005542157217860222, + 0.0045582614839077, + 0.008684035390615463, + 0.004457598552107811, + 0.01146157830953598, + 0.004811722785234451, + 0.003647719044238329, + 0.00830073468387127, + 0.007823703810572624, + 0.005282262340188026, + 0.006364766974002123, + 0.007752852980047464, + 0.005678546614944935, + 0.007350177504122257, + 0.003948731813579798, + 0.008259929716587067, + 0.00862671434879303, + 0.004841996356844902, + 0.006867261603474617, + 0.007303152699023485, + 0.00749544519931078, + 0.006074920296669006, + 0.00413562823086977, + 0.05278841406106949, + 0.034483298659324646, + 0.022991366684436798, + 0.04724644124507904, + 0.03676890954375267, + 0.05277450382709503, + 0.03203738480806351, + 0.016658643260598183, + 0.028912093490362167, + 0.021262239664793015, + 0.018350867554545403, + 0.02410035952925682, + 0.019277488812804222, + 0.027832752093672752, + 0.01664797216653824, + 0.013594182208180428, + 0.009102713316679, + 0.0044777002185583115, + 0.004246919881552458, + 0.0036843623965978622, + 0.003946271724998951, + 0.011588611640036106, + 0.0027904075104743242, + 0.003996431827545166, + 0.00650781812146306, + 0.0020685431081801653, + 0.002406574320048094, + 0.0024688795674592257, + 0.0015222771326079965, + 0.009118402376770973, + 0.0018258093623444438, + 0.0020777673926204443, + 0.004687263630330563, + 0.002581268083304167, + 0.002543781418353319, + 0.0032423962838947773, + 0.003396702231839299, + 0.0068905530497431755, + 0.0039377459324896336, + 0.002497055334970355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.01606578752398491 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects anthropomorphic forget-me-not flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flowers have sky-blue petals with soft, bleeding watercolor edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower on the left begin moving by bending its stem in a deliberate arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second flower respond by bending its stem inward until the petals are nearly touching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368738293647766 + ], + "opposite_logprob_tokens": [ + -3.4086873531341553 + ], + "logprob_reward": -0.03368738293647766, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368738293647766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20365209877490997, + 0.051474470645189285, + 0.05085466802120209, + 0.045344360172748566, + 0.04148838296532631, + 0.20235611498355865, + 0.16057594120502472, + 0.18009234964847565, + 0.03419577702879906, + 0.029965760186314583 + ], + "credit_map": [ + 0.02808345854282379, + 0.02171296998858452, + 0.015263560228049755, + 0.060753751546144485, + 0.018221886828541756, + 0.01997547410428524, + 0.026101283729076385, + 0.013539724983274937, + 0.006941452622413635, + 0.005883027799427509, + 0.004593588411808014, + 0.011445029638707638, + 0.0062944344244897366, + 0.005189866758882999, + 0.007237390149384737, + 0.003889685496687889, + 0.009200146421790123, + 0.007157854735851288, + 0.0043645440600812435, + 0.00908523891121149, + 0.005208921153098345, + 0.0070005436427891254, + 0.005644258111715317, + 0.003193166572600603, + 0.011425206437706947, + 0.003608797676861286, + 0.003924104385077953, + 0.004913926590234041, + 0.0033043301664292812, + 0.011413821950554848, + 0.0036022071726620197, + 0.0031519727781414986, + 0.005818420555442572, + 0.004777149297297001, + 0.003305627964437008, + 0.008021693676710129, + 0.00519506586715579, + 0.00484682759270072, + 0.006164823193103075, + 0.0033587771467864513, + 0.049060482531785965, + 0.019087983295321465, + 0.014639778062701225, + 0.023361731320619583, + 0.01859738677740097, + 0.05181222781538963, + 0.014833404682576656, + 0.010963134467601776, + 0.04985727742314339, + 0.012423842214047909, + 0.010576246306300163, + 0.010669674724340439, + 0.008967909030616283, + 0.05206892639398575, + 0.007879267446696758, + 0.00813281163573265, + 0.056020401418209076, + 0.013003051280975342, + 0.01315323542803526, + 0.006993554066866636, + 0.0071189263835549355, + 0.06836719065904617, + 0.006028669886291027, + 0.009407346136868, + 0.00459514232352376, + 0.0030315693002194166, + 0.005645196884870529, + 0.004749656189233065, + 0.002496244153007865, + 0.005275745876133442, + 0.002887638984248042, + 0.005514590535312891, + 0.003210114547982812, + 0.002652982249855995, + 0.003315279260277748, + 0.003797039622440934, + 0.00431775813922286, + 0.00468333438038826, + 0.00538678839802742, + 0.00260246847756207 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.03368738293647766 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, static medium shot immerses the viewer in an animated watercolor world, capturing a tender moment of courtship against a warm, creamy, textured paper background. At the center of the frame stand two anthropomorphic forget-me-not flowers, their delicate, soft green stems atop elegant blossoms. Each flower features a simple, expressive black-dot eye and a subtle pink watercolor blush, delicately blossoming at the base of the petals. The petals themselves are a sky-blue, with the characteristic soft, bleeding edges of the wet-on-wet watercolor technique, making the translucent color appear to radiate an inner, ethereal glow. The flower on the left tentatively begins the movement, its stem slowly bending in a deliberate arc. The second flower responds with shy hesitation, its stem gently bending inward until its blue petals are just a breath away from touching. The entire painting, set against a dark green background and enhanced by the luminous effect of the paint, exudes a profound atmosphere of tenderness, innocence, and romance.", + "seed": 19192, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_8/index_6894_seed_19192_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07433819179423154, + "vlm_logprob": -0.07433819179423154, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.07433819179423154, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects anthropomorphic forget-me-not flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flowers have sky-blue petals with soft, bleeding watercolor edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower on the left begin moving by bending its stem in a deliberate arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second flower respond by bending its stem inward until the petals are nearly touching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2231811285018921 + ], + "opposite_logprob_tokens": [ + -1.723181128501892 + ], + "logprob_reward": -0.2231811285018921, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2231811285018921, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23936259746551514, + 0.17190386354923248, + 0.12294114381074905, + 0.2025420069694519, + 0.0536138080060482, + 0.04543274641036987, + 0.047028955072164536, + 0.0466311015188694, + 0.029551388695836067, + 0.04099240154027939 + ], + "credit_map": [ + 0.02885599248111248, + 0.024929119274020195, + 0.02462192438542843, + 0.03992536664009094, + 0.029506342485547066, + 0.022370589897036552, + 0.04512226954102516, + 0.02403099276125431, + 0.03831729665398598, + 0.07459650933742523, + 0.0066002472303807735, + 0.015512597747147083, + 0.010294958017766476, + 0.008943816646933556, + 0.008127354085445404, + 0.009511088021099567, + 0.02667640522122383, + 0.014578078873455524, + 0.009618446230888367, + 0.016476403921842575, + 0.017353355884552002, + 0.009504012763500214, + 0.010558542795479298, + 0.01817590370774269, + 0.02697540447115898, + 0.07367920875549316, + 0.011037471704185009, + 0.015513543039560318, + 0.019389713183045387, + 0.01173659972846508, + 0.012625998817384243, + 0.03158406540751457, + 0.006347385235130787, + 0.004681839607656002, + 0.005658757407218218, + 0.008840374648571014, + 0.005342225078493357, + 0.006926949135959148, + 0.006824753247201443, + 0.00899152085185051, + 0.008691582828760147, + 0.004639252554625273, + 0.0039043675642460585, + 0.006437190342694521, + 0.005528255831450224, + 0.004090198781341314, + 0.005237691104412079, + 0.006904208101332188, + 0.007418167777359486, + 0.005785134620964527, + 0.005234336946159601, + 0.004248062614351511, + 0.006771855056285858, + 0.0060144346207380295, + 0.0058699012733995914, + 0.0056870561093091965, + 0.00611557113006711, + 0.0045845480635762215, + 0.005404649768024683, + 0.006508848164230585, + 0.004692979156970978, + 0.00737124914303422, + 0.006532170344144106, + 0.005421084351837635, + 0.00506133446469903, + 0.0029481234960258007, + 0.003786705434322357, + 0.0033606502693146467, + 0.0028007940854877234, + 0.004878513980656862, + 0.003025691257789731, + 0.0036895761732012033, + 0.007729788776487112, + 0.0051798452623188496, + 0.005117285996675491, + 0.004044434055685997, + 0.004481997340917587, + 0.00408934336155653, + 0.00572814978659153, + 0.004621557891368866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.2231811285018921 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects anthropomorphic forget-me-not flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flowers have sky-blue petals with soft, bleeding watercolor edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower on the left begin moving by bending its stem in a deliberate arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second flower respond by bending its stem inward until the petals are nearly touching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032209681812673807 + ], + "opposite_logprob_tokens": [ + -5.753221035003662 + ], + "logprob_reward": -0.0032209681812673807, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032209681812673807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15524132549762726, + 0.21457770466804504, + 0.14882856607437134, + 0.11235957592725754, + 0.24764499068260193, + 0.022997502237558365, + 0.027082517743110657, + 0.025068044662475586, + 0.021237310022115707, + 0.024962536990642548 + ], + "credit_map": [ + 0.019115550443530083, + 0.01857812888920307, + 0.01755932718515396, + 0.026729252189397812, + 0.016736237332224846, + 0.012768185697495937, + 0.02693565934896469, + 0.016818983480334282, + 0.03491557389497757, + 0.09614010155200958, + 0.007156927138566971, + 0.019928159192204475, + 0.014510816894471645, + 0.010739440098404884, + 0.008627735078334808, + 0.02255893498659134, + 0.016275780275464058, + 0.029082193970680237, + 0.01046194788068533, + 0.017236506566405296, + 0.02605673298239708, + 0.013923789374530315, + 0.014201921410858631, + 0.021589694544672966, + 0.01370618212968111, + 0.029019348323345184, + 0.013735160231590271, + 0.010767703875899315, + 0.013756933622062206, + 0.008300279267132282, + 0.010367611423134804, + 0.012706358917057514, + 0.027019886299967766, + 0.024657242000102997, + 0.021695930510759354, + 0.02282710187137127, + 0.02863181382417679, + 0.06297014653682709, + 0.02138826623558998, + 0.038454603403806686, + 0.004499586299061775, + 0.0023808928672224283, + 0.0024445499293506145, + 0.002977158408612013, + 0.002438463969156146, + 0.00234685349278152, + 0.003019432770088315, + 0.0028905633371323347, + 0.004675146192312241, + 0.0029926993884146214, + 0.0030137381982058287, + 0.0029591985512524843, + 0.0028977051842957735, + 0.00440826965495944, + 0.0027600314933806658, + 0.003375727916136384, + 0.003008143277838826, + 0.0026268165092915297, + 0.0026866949629038572, + 0.002795298583805561, + 0.0026810565032064915, + 0.004818151239305735, + 0.0030816763173788786, + 0.0033702063374221325, + 0.0030125149060040712, + 0.0020329360850155354, + 0.0029257945716381073, + 0.002151002874597907, + 0.001692165620625019, + 0.003724833717569709, + 0.002844119444489479, + 0.002853941870853305, + 0.0036772717721760273, + 0.0030625604558736086, + 0.0035547595471143723, + 0.0031478614546358585, + 0.002644844353199005, + 0.0027673589065670967, + 0.002806417178362608, + 0.0033014637883752584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.0032209681812673807 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects anthropomorphic forget-me-not flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flowers have sky-blue petals with soft, bleeding watercolor edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower on the left begin moving by bending its stem in a deliberate arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second flower respond by bending its stem inward until the petals are nearly touching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789603590965271 + ], + "opposite_logprob_tokens": [ + -2.578960418701172 + ], + "logprob_reward": -0.0789603590965271, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789603590965271, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19877275824546814, + 0.32443365454673767, + 0.07605025917291641, + 0.16584454476833344, + 0.10148759931325912, + 0.016309555619955063, + 0.02322356402873993, + 0.019178807735443115, + 0.017875440418720245, + 0.05682383477687836 + ], + "credit_map": [ + 0.02467252127826214, + 0.02090069279074669, + 0.017749439924955368, + 0.03106481023132801, + 0.023630214855074883, + 0.01763337105512619, + 0.04045715183019638, + 0.02266455814242363, + 0.06275498121976852, + 0.14596869051456451, + 0.011112716980278492, + 0.027378074824810028, + 0.020679324865341187, + 0.01744319684803486, + 0.012904388830065727, + 0.026192286983132362, + 0.02269223891198635, + 0.013738223351538181, + 0.00509639922529459, + 0.006882477086037397, + 0.007181023713201284, + 0.004925263114273548, + 0.004382958170026541, + 0.011151672340929508, + 0.019719505682587624, + 0.0524759478867054, + 0.015422710217535496, + 0.014570971019566059, + 0.016560757532715797, + 0.009894340299069881, + 0.013226552866399288, + 0.02397376298904419, + 0.01255021058022976, + 0.010834155604243279, + 0.009387589059770107, + 0.010962000116705894, + 0.01192676555365324, + 0.022395458072423935, + 0.010348957031965256, + 0.013082453981041908, + 0.002488659927621484, + 0.0022805174812674522, + 0.0013072753790766, + 0.0020264415070414543, + 0.002272899029776454, + 0.0016238879179582, + 0.0015722491079941392, + 0.0027376250363886356, + 0.005562237463891506, + 0.0026171852368861437, + 0.003007651073858142, + 0.002482717391103506, + 0.0018471364164724946, + 0.0019033821299672127, + 0.0027794153429567814, + 0.0030238383915275335, + 0.002889132359996438, + 0.001829145010560751, + 0.0023408117704093456, + 0.0029876194894313812, + 0.0015941348392516375, + 0.0024987373035401106, + 0.003086067968979478, + 0.0019531596917659044, + 0.0034300019033253193, + 0.0017817122861742973, + 0.0036678477190434933, + 0.0016691847704350948, + 0.0015710025327280164, + 0.001943067298270762, + 0.0013594655320048332, + 0.002453158376738429, + 0.006167153362184763, + 0.00863729603588581, + 0.007016184739768505, + 0.006017748266458511, + 0.008348338305950165, + 0.007140791043639183, + 0.006539497524499893, + 0.006956824101507664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.0789603590965271 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects anthropomorphic forget-me-not flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flowers have sky-blue petals with soft, bleeding watercolor edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower on the left begin moving by bending its stem in a deliberate arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second flower respond by bending its stem inward until the petals are nearly touching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023281201720237732 + ], + "opposite_logprob_tokens": [ + -3.7732810974121094 + ], + "logprob_reward": -0.023281201720237732, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023281201720237732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1411195993423462, + 0.2789861857891083, + 0.0730944275856018, + 0.1936771124601364, + 0.08308892697095871, + 0.025982515886425972, + 0.04283514246344566, + 0.02625788003206253, + 0.02644205279648304, + 0.1085161417722702 + ], + "credit_map": [ + 0.016492392867803574, + 0.014921302907168865, + 0.011787108145654202, + 0.022942177951335907, + 0.01671075075864792, + 0.012783566489815712, + 0.028986798599362373, + 0.01649549789726734, + 0.05362968519330025, + 0.12345153838396072, + 0.010433126240968704, + 0.024221021682024002, + 0.017833374440670013, + 0.0152225811034441, + 0.01211625523865223, + 0.022078586742281914, + 0.020756131038069725, + 0.013145570643246174, + 0.004739963449537754, + 0.007611915934830904, + 0.007407200988382101, + 0.005085076671093702, + 0.0046163746155798435, + 0.009732195176184177, + 0.023062270134687424, + 0.056229401379823685, + 0.021331362426280975, + 0.01814214698970318, + 0.021543866023421288, + 0.01237551774829626, + 0.016279026865959167, + 0.024713529273867607, + 0.010378516279160976, + 0.009201740846037865, + 0.007162658963352442, + 0.00914359837770462, + 0.010359629988670349, + 0.01681162603199482, + 0.008709659799933434, + 0.011321494355797768, + 0.0026761912740767, + 0.00311528448946774, + 0.002096928656101227, + 0.0034830300137400627, + 0.003239187877625227, + 0.002730191685259342, + 0.0028727445751428604, + 0.005768956616520882, + 0.009609587490558624, + 0.004613745026290417, + 0.0054393126629292965, + 0.004455886781215668, + 0.004025428090244532, + 0.004506043158471584, + 0.005145712289959192, + 0.005039429757744074, + 0.0032074442133307457, + 0.0025069057010114193, + 0.003148294985294342, + 0.003696407889947295, + 0.0025140808429569006, + 0.003579014213755727, + 0.004469358362257481, + 0.003136372659355402, + 0.004603723995387554, + 0.00243781809695065, + 0.005010257475078106, + 0.0023964501451700926, + 0.0025984495878219604, + 0.0040687150321900845, + 0.0018466091714799404, + 0.0034800302237272263, + 0.010001515038311481, + 0.01735687628388405, + 0.012598294764757156, + 0.012191670946776867, + 0.01707707904279232, + 0.013538781553506851, + 0.013139155693352222, + 0.01261276938021183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.023281201720237732 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects anthropomorphic forget-me-not flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flowers have sky-blue petals with soft, bleeding watercolor edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower on the left begin moving by bending its stem in a deliberate arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second flower respond by bending its stem inward until the petals are nearly touching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304730147123337 + ], + "opposite_logprob_tokens": [ + -3.1680471897125244 + ], + "logprob_reward": -0.04304730147123337, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304730147123337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1915857344865799, + 0.3961453437805176, + 0.07393312454223633, + 0.1487666517496109, + 0.051804762333631516, + 0.03465529531240463, + 0.031973544508218765, + 0.023138850927352905, + 0.025092700496315956, + 0.02290397509932518 + ], + "credit_map": [ + 0.023171978071331978, + 0.020986689254641533, + 0.019694533199071884, + 0.029896780848503113, + 0.023015830665826797, + 0.018451940268278122, + 0.03768450766801834, + 0.018683474510908127, + 0.08620607107877731, + 0.18162254989147186, + 0.012511293403804302, + 0.03278614953160286, + 0.02353881672024727, + 0.018166808411478996, + 0.015819111838936806, + 0.02549452893435955, + 0.016970742493867874, + 0.012709736824035645, + 0.005154510959982872, + 0.00952198076993227, + 0.007697065360844135, + 0.005387372802942991, + 0.007334735710173845, + 0.009156972169876099, + 0.017800213769078255, + 0.04756046086549759, + 0.012696064077317715, + 0.013081821613013744, + 0.014872939325869083, + 0.009562376886606216, + 0.010346745140850544, + 0.022846035659313202, + 0.005410005338490009, + 0.004062667489051819, + 0.004261079244315624, + 0.006784559693187475, + 0.00523937214165926, + 0.012461467646062374, + 0.00590189965441823, + 0.007683706935495138, + 0.00443815253674984, + 0.003908548038452864, + 0.003439021296799183, + 0.004888926167041063, + 0.0036606157664209604, + 0.004217460751533508, + 0.0038326377980411053, + 0.006269929464906454, + 0.003058526897802949, + 0.0024188938550651073, + 0.005140929948538542, + 0.00581539049744606, + 0.0021624674554914236, + 0.0030015334486961365, + 0.005982839968055487, + 0.0043929582461714745, + 0.00289111933670938, + 0.0023506053257733583, + 0.0031777455005794764, + 0.0032935691997408867, + 0.0022854702547192574, + 0.0030758550856262445, + 0.003280225908383727, + 0.0027842598501592875, + 0.0038102183025330305, + 0.002299064537510276, + 0.0030017865356057882, + 0.0034493489656597376, + 0.002086081076413393, + 0.003565520979464054, + 0.0037334307562559843, + 0.0031472472473978996, + 0.004225257318466902, + 0.00263787480071187, + 0.0025646507274359465, + 0.0019448798848316073, + 0.0027754686307162046, + 0.0030261839274317026, + 0.0029240299481898546, + 0.0028056283481419086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.04304730147123337 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_9/index_tensor([6131, 6131])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_9/index_tensor([6131, 6131])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..97c4c1ca3e84cbc57d990b9d6163e9adc7f5a09d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_9/index_tensor([6131, 6131])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In the cozy style of Ghibli animation, an extreme close-up, held static in the foreground, captures the moment a teardrop-shaped candle flame suddenly swirls into existence. It gracefully spirals upward from a wick in an antique brass holder, which features soft, aged copper paste and a delicate, hand-etched floral pattern. The flame itself is magical; its core radiates a soft, ethereal blue, while a warm golden glow envelops its periphery. As the flame swirls with fluid, balletic grace, it leaves a trail of gentle, glowing pinpoints of light that detach and drift lazily through the air like tiny, flickering fireflies. These particles cast a soft, dancing glow on their surroundings before fading into the soft-focus background of the dimly lit, tranquil room, creating an overall atmosphere of tranquility and captivating warmth.", + "seed": 8972, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_9/index_6131_seed_8972_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06666611172258854, + "vlm_logprob": -0.06666611172258854, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.06666611172258854, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teardrop-shaped candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle flame swirling into existence and spiraling upward from a wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle held in an antique brass holder with soft aged copper paste and a hand-etched floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame have a soft ethereal blue core with a warm golden glow on its periphery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing pinpoints of light detach from the flame and drift like fireflies in a dimly lit, tranquil room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03164888545870781 + ], + "opposite_logprob_tokens": [ + -3.656648874282837 + ], + "logprob_reward": -0.03164888545870781, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03164888545870781, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11903174966573715, + 0.10147430747747421, + 0.16016095876693726, + 0.04500199481844902, + 0.06363075226545334, + 0.19674170017242432, + 0.1521543711423874, + 0.060163527727127075, + 0.05021831765770912, + 0.05142228305339813 + ], + "credit_map": [ + 0.022221624851226807, + 0.013138096779584885, + 0.021456044167280197, + 0.0206137802451849, + 0.006894222926348448, + 0.00695456750690937, + 0.015786880627274513, + 0.011966526508331299, + 0.013235015794634819, + 0.009883238933980465, + 0.015473362058401108, + 0.022550251334905624, + 0.006273255217820406, + 0.004448652267456055, + 0.019584786146879196, + 0.010025747120380402, + 0.03207818791270256, + 0.023238375782966614, + 0.027403531596064568, + 0.022253571078181267, + 0.012613261118531227, + 0.010481941513717175, + 0.018377777189016342, + 0.013714322820305824, + 0.006978622637689114, + 0.005725377704948187, + 0.008470454253256321, + 0.00607152096927166, + 0.004373969975858927, + 0.0022823535837233067, + 0.006063796114176512, + 0.005035901442170143, + 0.007775684352964163, + 0.009688579477369785, + 0.009848834946751595, + 0.011428184807300568, + 0.006819719448685646, + 0.003813226008787751, + 0.008748306892812252, + 0.005508215166628361, + 0.03787081316113472, + 0.026785260066390038, + 0.042255982756614685, + 0.030559347942471504, + 0.015026270411908627, + 0.007078718859702349, + 0.021147051826119423, + 0.016018273308873177, + 0.018951987847685814, + 0.02401150017976761, + 0.022314680740237236, + 0.019694233313202858, + 0.023862309753894806, + 0.010573905892670155, + 0.019727418199181557, + 0.013018338941037655, + 0.008835149928927422, + 0.0058931997045874596, + 0.00941956415772438, + 0.006624532397836447, + 0.009867237880825996, + 0.005056299269199371, + 0.009476623497903347, + 0.004990920424461365, + 0.006563558243215084, + 0.006265290081501007, + 0.00815186183899641, + 0.010950498282909393, + 0.0030808004084974527, + 0.002537095220759511, + 0.007656753063201904, + 0.005012459121644497, + 0.005895783193409443, + 0.011563941836357117, + 0.0069584352895617485, + 0.007365962956100702, + 0.007600878365337849, + 0.0028290871996432543, + 0.005848415661603212, + 0.0033597731962800026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 6 + ], + "selected_reward": -0.03164888545870781 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teardrop-shaped candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle flame swirling into existence and spiraling upward from a wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle held in an antique brass holder with soft aged copper paste and a hand-etched floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame have a soft ethereal blue core with a warm golden glow on its periphery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing pinpoints of light detach from the flame and drift like fireflies in a dimly lit, tranquil room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054980020970106125 + ], + "opposite_logprob_tokens": [ + -2.9299800395965576 + ], + "logprob_reward": -0.054980020970106125, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054980020970106125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0967836007475853, + 0.04845732823014259, + 0.1505962312221527, + 0.049314070492982864, + 0.1436683088541031, + 0.17700748145580292, + 0.04899578541517258, + 0.027826521545648575, + 0.2281886786222458, + 0.0291619710624218 + ], + "credit_map": [ + 0.01709916442632675, + 0.0094776451587677, + 0.025098802521824837, + 0.013589042238891125, + 0.007300414610654116, + 0.006000684108585119, + 0.010144407860934734, + 0.00807343702763319, + 0.009475205093622208, + 0.0053115240298211575, + 0.010033524595201015, + 0.007692074403166771, + 0.003458541352301836, + 0.001996344653889537, + 0.007297174073755741, + 0.00319294142536819, + 0.031224098056554794, + 0.014598785899579525, + 0.03235853835940361, + 0.024662554264068604, + 0.007485019043087959, + 0.006934025324881077, + 0.022304635494947433, + 0.011028564535081387, + 0.006713657174259424, + 0.00745334941893816, + 0.007906932383775711, + 0.007599526550620794, + 0.006355911493301392, + 0.002549064112827182, + 0.006299022119492292, + 0.0044366056099534035, + 0.01739964634180069, + 0.016607923433184624, + 0.023190846666693687, + 0.024486657232046127, + 0.010569160804152489, + 0.010850056074559689, + 0.02575785294175148, + 0.014806179329752922, + 0.03536732494831085, + 0.022976959124207497, + 0.03654862940311432, + 0.029313528910279274, + 0.012401171959936619, + 0.008038598112761974, + 0.018856966868042946, + 0.013504311442375183, + 0.006713134236633778, + 0.009375522844493389, + 0.006105651147663593, + 0.007559585385024548, + 0.005346688907593489, + 0.003552926704287529, + 0.005946599878370762, + 0.004395680036395788, + 0.0028496074955910444, + 0.003612678963690996, + 0.0033003902062773705, + 0.0047037941403687, + 0.0034589956048876047, + 0.002301236381754279, + 0.004473217763006687, + 0.0031266009900718927, + 0.03953713923692703, + 0.03650914505124092, + 0.03353436663746834, + 0.03571588918566704, + 0.015032941475510597, + 0.01456846296787262, + 0.02857007458806038, + 0.024720655754208565, + 0.004109370522201061, + 0.006408789660781622, + 0.004285484552383423, + 0.0047906250692903996, + 0.002228210447356105, + 0.0011948359897360206, + 0.003549192799255252, + 0.0025954623706638813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 2 + ], + "selected_reward": -0.054980020970106125 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teardrop-shaped candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle flame swirling into existence and spiraling upward from a wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle held in an antique brass holder with soft aged copper paste and a hand-etched floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame have a soft ethereal blue core with a warm golden glow on its periphery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing pinpoints of light detach from the flame and drift like fireflies in a dimly lit, tranquil room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16060274839401245 + ], + "opposite_logprob_tokens": [ + -1.9106028079986572 + ], + "logprob_reward": -0.16060274839401245, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16060274839401245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07208273559808731, + 0.05995442345738411, + 0.11362344026565552, + 0.0239239651709795, + 0.08896195888519287, + 0.1674773395061493, + 0.23339228332042694, + 0.12166456878185272, + 0.06784731894731522, + 0.05107196047902107 + ], + "credit_map": [ + 0.01451107021421194, + 0.007091935258358717, + 0.01338901836425066, + 0.011764013208448887, + 0.0039413729682564735, + 0.004365177825093269, + 0.007996072061359882, + 0.00902407243847847, + 0.009215282276272774, + 0.00563320005312562, + 0.00830700621008873, + 0.01238213386386633, + 0.004476024769246578, + 0.003484127577394247, + 0.009199532680213451, + 0.007257111836224794, + 0.021099798381328583, + 0.015563461929559708, + 0.01857919618487358, + 0.01827118918299675, + 0.008048850111663342, + 0.007205907721072435, + 0.015423953533172607, + 0.0094310836866498, + 0.0028782314620912075, + 0.0024253265000879765, + 0.0035925686825066805, + 0.005243534222245216, + 0.001418250147253275, + 0.0018612490966916084, + 0.0033036288805305958, + 0.0032011757139116526, + 0.009833261370658875, + 0.018069230020046234, + 0.011574381962418556, + 0.012539367191493511, + 0.011867746710777283, + 0.005098122172057629, + 0.012778513133525848, + 0.007201335858553648, + 0.030871082097291946, + 0.018540337681770325, + 0.03578823059797287, + 0.026381289586424828, + 0.01147162914276123, + 0.006832609884440899, + 0.02345455251634121, + 0.014137603342533112, + 0.026472488418221474, + 0.04597077518701553, + 0.027837546542286873, + 0.03564669191837311, + 0.03235746547579765, + 0.015455028973519802, + 0.02649029530584812, + 0.023162011057138443, + 0.0170038640499115, + 0.019201334565877914, + 0.016164448112249374, + 0.01965034008026123, + 0.009562079794704914, + 0.008182202465832233, + 0.017706256359815598, + 0.014194052666425705, + 0.010405146516859531, + 0.007320646662265062, + 0.009245116263628006, + 0.014024091884493828, + 0.0057205380871891975, + 0.003766589565202594, + 0.009671379812061787, + 0.007693813648074865, + 0.007882392033934593, + 0.009978793561458588, + 0.006856628693640232, + 0.007850074209272861, + 0.005918208509683609, + 0.002308416645973921, + 0.006702079903334379, + 0.0035753645934164524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.16060274839401245 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teardrop-shaped candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle flame swirling into existence and spiraling upward from a wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle held in an antique brass holder with soft aged copper paste and a hand-etched floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame have a soft ethereal blue core with a warm golden glow on its periphery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing pinpoints of light detach from the flame and drift like fireflies in a dimly lit, tranquil room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003170996904373 + ], + "opposite_logprob_tokens": [ + -2.6950316429138184 + ], + "logprob_reward": -0.07003170996904373, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003170996904373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11491511017084122, + 0.09370481967926025, + 0.3666463792324066, + 0.016938557848334312, + 0.0860893726348877, + 0.07153470069169998, + 0.08237416297197342, + 0.05299380421638489, + 0.10133688896894455, + 0.01346615795046091 + ], + "credit_map": [ + 0.015621316619217396, + 0.010906655341386795, + 0.016105297952890396, + 0.023781035095453262, + 0.01063036359846592, + 0.008836652152240276, + 0.01959611289203167, + 0.009437676519155502, + 0.012047920376062393, + 0.010113244876265526, + 0.01175395306199789, + 0.01890670880675316, + 0.008410941809415817, + 0.006154749076813459, + 0.017904626205563545, + 0.008412674069404602, + 0.0801231637597084, + 0.04850217327475548, + 0.07867036014795303, + 0.047247156500816345, + 0.025515854358673096, + 0.019755996763706207, + 0.03582971170544624, + 0.031001953408122063, + 0.002173030050471425, + 0.001648031291551888, + 0.0022748014889657497, + 0.0032297498546540737, + 0.0015713755274191499, + 0.001152176409959793, + 0.0030970643274486065, + 0.0017923302948474884, + 0.00937534961849451, + 0.011495755985379219, + 0.01136615127325058, + 0.01586485654115677, + 0.009361468255519867, + 0.006777443923056126, + 0.013467136770486832, + 0.008381214924156666, + 0.012756379321217537, + 0.008255038410425186, + 0.01454476360231638, + 0.011384338140487671, + 0.005879566073417664, + 0.0034380550496280193, + 0.009131127968430519, + 0.006145433057099581, + 0.01112369354814291, + 0.016068022698163986, + 0.011136811226606369, + 0.0120640704408288, + 0.008773495443165302, + 0.005472790915518999, + 0.00898469053208828, + 0.008750587701797485, + 0.005895065143704414, + 0.004748769570142031, + 0.007105435244739056, + 0.006546346470713615, + 0.00842453446239233, + 0.004829813726246357, + 0.009365592151880264, + 0.006078249774873257, + 0.014896106906235218, + 0.017363477498292923, + 0.013854249380528927, + 0.018842458724975586, + 0.006929360330104828, + 0.0057384176179766655, + 0.01572558283805847, + 0.00798723939806223, + 0.0017231069505214691, + 0.0021830026526004076, + 0.0018037798581644893, + 0.0021386316511780024, + 0.0018030238570645452, + 0.0006892840028740466, + 0.001878521521575749, + 0.0012468057684600353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.07003170996904373 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teardrop-shaped candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle flame swirling into existence and spiraling upward from a wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle held in an antique brass holder with soft aged copper paste and a hand-etched floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame have a soft ethereal blue core with a warm golden glow on its periphery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing pinpoints of light detach from the flame and drift like fireflies in a dimly lit, tranquil room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606719382107258 + ], + "opposite_logprob_tokens": [ + -4.141067028045654 + ], + "logprob_reward": -0.01606719382107258, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606719382107258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05604330077767372, + 0.033626120537519455, + 0.022602615877985954, + 0.0305364690721035, + 0.06512222439050674, + 0.08282563835382462, + 0.08672401309013367, + 0.06810981035232544, + 0.43848976492881775, + 0.11592014878988266 + ], + "credit_map": [ + 0.007068125531077385, + 0.004641640465706587, + 0.009009839966893196, + 0.012327324599027634, + 0.0033757847268134356, + 0.0027919597923755646, + 0.01134287379682064, + 0.00548574049025774, + 0.0052248695865273476, + 0.0038896610494703054, + 0.004763739183545113, + 0.006972752045840025, + 0.002582745160907507, + 0.0024745543487370014, + 0.004630346782505512, + 0.003087446792051196, + 0.0033467502798885107, + 0.0020257055293768644, + 0.0033314714673906565, + 0.0033185824286192656, + 0.002310192445293069, + 0.002669725799933076, + 0.002845070092007518, + 0.002755114110186696, + 0.005065203178673983, + 0.0035315535496920347, + 0.005898774601519108, + 0.004518868401646614, + 0.0028523136861622334, + 0.001547685475088656, + 0.0042463685385882854, + 0.0028756950050592422, + 0.00748573848977685, + 0.011385586112737656, + 0.009154497645795345, + 0.01064646802842617, + 0.007885495200753212, + 0.004079327452927828, + 0.009468434378504753, + 0.005016663111746311, + 0.01523878239095211, + 0.009449208155274391, + 0.016733186319470406, + 0.013180747628211975, + 0.007063918747007847, + 0.003116087056696415, + 0.009928843006491661, + 0.008114846423268318, + 0.013104435056447983, + 0.013013758696615696, + 0.01278924010694027, + 0.01077218260616064, + 0.01013286504894495, + 0.006242577452212572, + 0.009547297842800617, + 0.011121636256575584, + 0.00814522709697485, + 0.006586425472050905, + 0.008442813530564308, + 0.011936026625335217, + 0.009538117796182632, + 0.005599886644631624, + 0.011638620868325233, + 0.006222677417099476, + 0.07417260855436325, + 0.08779411017894745, + 0.06578389555215836, + 0.0625278502702713, + 0.03362295404076576, + 0.029909640550613403, + 0.041322361677885056, + 0.043356265872716904, + 0.0152349341660738, + 0.022388434037566185, + 0.01743626594543457, + 0.019214672967791557, + 0.012642459012567997, + 0.006061558146029711, + 0.01391205657273531, + 0.009029749780893326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.01606719382107258 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the cozy style of Ghibli animation, an extreme close-up, held static in the foreground, captures the moment a teardrop-shaped candle flame suddenly swirls into existence. It gracefully spirals upward from a wick in an antique brass holder, which features soft, aged copper paste and a delicate, hand-etched floral pattern. The flame itself is magical; its core radiates a soft, ethereal blue, while a warm golden glow envelops its periphery. As the flame swirls with fluid, balletic grace, it leaves a trail of gentle, glowing pinpoints of light that detach and drift lazily through the air like tiny, flickering fireflies. These particles cast a soft, dancing glow on their surroundings before fading into the soft-focus background of the dimly lit, tranquil room, creating an overall atmosphere of tranquility and captivating warmth.", + "seed": 8973, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_9/index_6131_seed_8973_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.18839506320655347, + "vlm_logprob": -0.18839506320655347, + "vlm_yes_no_margin": 2.35, + "vlm_token_credit": -0.18839506320655347, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teardrop-shaped candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle flame swirling into existence and spiraling upward from a wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle held in an antique brass holder with soft aged copper paste and a hand-etched floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame have a soft ethereal blue core with a warm golden glow on its periphery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing pinpoints of light detach from the flame and drift like fireflies in a dimly lit, tranquil room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038962967693805695 + ], + "opposite_logprob_tokens": [ + -3.4139630794525146 + ], + "logprob_reward": -0.038962967693805695, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.038962967693805695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2443503886461258, + 0.07250989973545074, + 0.05340088903903961, + 0.09624144434928894, + 0.22215153276920319, + 0.05865773558616638, + 0.07050462067127228, + 0.06761979311704636, + 0.029785064980387688, + 0.08477868139743805 + ], + "credit_map": [ + 0.05191858857870102, + 0.020342878997325897, + 0.03451451659202576, + 0.02740512043237686, + 0.015572317875921726, + 0.026874396950006485, + 0.03293338045477867, + 0.03478914871811867, + 0.00923268310725689, + 0.008459712378680706, + 0.008361267857253551, + 0.005783279426395893, + 0.00789562426507473, + 0.011819541454315186, + 0.009698979556560516, + 0.011258808895945549, + 0.005885542836040258, + 0.004484714008867741, + 0.007440394256263971, + 0.004378383047878742, + 0.004275546874850988, + 0.011852987110614777, + 0.005605919286608696, + 0.009477396495640278, + 0.008479563519358635, + 0.008623062632977962, + 0.015367340296506882, + 0.01087215542793274, + 0.008513049222528934, + 0.014838297851383686, + 0.015123442746698856, + 0.014424524269998074, + 0.018492931500077248, + 0.031051382422447205, + 0.026647698134183884, + 0.01866476982831955, + 0.022751543670892715, + 0.03194533288478851, + 0.029321374371647835, + 0.043276477605104446, + 0.009563660249114037, + 0.006761002354323864, + 0.008133949711918831, + 0.00457990076392889, + 0.005978609435260296, + 0.009544873610138893, + 0.00617590406909585, + 0.00791983213275671, + 0.005841654725372791, + 0.005766944494098425, + 0.012374159879982471, + 0.0065598455257713795, + 0.005652434658259153, + 0.015862826257944107, + 0.007104883436113596, + 0.011341866105794907, + 0.009259541518986225, + 0.00732929864898324, + 0.007071910426020622, + 0.0042573921382427216, + 0.007783304899930954, + 0.012754328548908234, + 0.006183162797242403, + 0.012980847619473934, + 0.003723548259586096, + 0.0026153638027608395, + 0.0032805665396153927, + 0.0032636341638863087, + 0.0033285601530224085, + 0.0035511208698153496, + 0.00517183355987072, + 0.004850434605032206, + 0.01076811645179987, + 0.008004138246178627, + 0.013316886499524117, + 0.009174094535410404, + 0.006848850753158331, + 0.011016083881258965, + 0.015154429711401463, + 0.010496078059077263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.038962967693805695 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teardrop-shaped candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle flame swirling into existence and spiraling upward from a wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle held in an antique brass holder with soft aged copper paste and a hand-etched floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame have a soft ethereal blue core with a warm golden glow on its periphery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing pinpoints of light detach from the flame and drift like fireflies in a dimly lit, tranquil room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07006071507930756 + ], + "opposite_logprob_tokens": [ + -2.6950607299804688 + ], + "logprob_reward": -0.07006071507930756, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07006071507930756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21829766035079956, + 0.049682967364788055, + 0.14984889328479767, + 0.22418300807476044, + 0.061282236129045486, + 0.056190814822912216, + 0.05104909837245941, + 0.09222222864627838, + 0.07442936301231384, + 0.02281368523836136 + ], + "credit_map": [ + 0.031168511137366295, + 0.014964764006435871, + 0.02321934700012207, + 0.016736876219511032, + 0.011202260851860046, + 0.057250916957855225, + 0.013798226602375507, + 0.04995676875114441, + 0.005907212384045124, + 0.005127245094627142, + 0.007266016211360693, + 0.005512290634214878, + 0.006066065281629562, + 0.006348087452352047, + 0.007612162735313177, + 0.005843894090503454, + 0.019224897027015686, + 0.026141665875911713, + 0.016133710741996765, + 0.010179891251027584, + 0.010826990008354187, + 0.028223492205142975, + 0.011439923197031021, + 0.02767832763493061, + 0.021899539977312088, + 0.0273448396474123, + 0.0266105979681015, + 0.02315843105316162, + 0.04259895905852318, + 0.02498900517821312, + 0.03151184320449829, + 0.02606980688869953, + 0.009244118817150593, + 0.005599816795438528, + 0.009088495746254921, + 0.004654225893318653, + 0.0049973721615970135, + 0.010792630724608898, + 0.007503553759306669, + 0.00940202921628952, + 0.004752554930746555, + 0.0070570181123912334, + 0.0075917551293969154, + 0.003287946106866002, + 0.0038794130086898804, + 0.01453955564647913, + 0.003996089566498995, + 0.011086484417319298, + 0.006735099479556084, + 0.006831498816609383, + 0.008501035161316395, + 0.004475614055991173, + 0.005150818731635809, + 0.007272979244589806, + 0.0064659263007342815, + 0.005616130772978067, + 0.012618784792721272, + 0.007021758239716291, + 0.013109562918543816, + 0.008297155611217022, + 0.010446388274431229, + 0.011922462843358517, + 0.013146131299436092, + 0.015659991651773453, + 0.01051742397248745, + 0.006172510329633951, + 0.009775332175195217, + 0.009219926781952381, + 0.006515663582831621, + 0.008612356148660183, + 0.013226831331849098, + 0.010389328002929688, + 0.0026872060261666775, + 0.0017487016739323735, + 0.003203901695087552, + 0.003650539554655552, + 0.0020891849417239428, + 0.002754582790657878, + 0.004261063877493143, + 0.002418505260720849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.07006071507930756 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teardrop-shaped candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle flame swirling into existence and spiraling upward from a wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle held in an antique brass holder with soft aged copper paste and a hand-etched floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame have a soft ethereal blue core with a warm golden glow on its periphery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing pinpoints of light detach from the flame and drift like fireflies in a dimly lit, tranquil room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6936310529708862 + ], + "opposite_logprob_tokens": [ + -0.6936310529708862 + ], + "logprob_reward": -0.6936310529708862, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6936310529708862, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.28007620573043823, + 0.05519303306937218, + 0.051979172974824905, + 0.347096711397171, + 0.06870661675930023, + 0.034426670521497726, + 0.021052127704024315, + 0.05855809152126312, + 0.03505665436387062, + 0.04785474017262459 + ], + "credit_map": [ + 0.06031142175197601, + 0.018823964521288872, + 0.03218651935458183, + 0.023747872561216354, + 0.011950097046792507, + 0.05764273181557655, + 0.021737752482295036, + 0.05367584899067879, + 0.009337983094155788, + 0.00512476684525609, + 0.009989350102841854, + 0.005369453225284815, + 0.004706717561930418, + 0.007537812925875187, + 0.006947546266019344, + 0.006179397460073233, + 0.005758380983024836, + 0.004235419910401106, + 0.008585615083575249, + 0.0044842190109193325, + 0.00509747164323926, + 0.00920990388840437, + 0.005282553844153881, + 0.009325603023171425, + 0.03382636606693268, + 0.03721602261066437, + 0.04297192767262459, + 0.035206139087677, + 0.05257735773921013, + 0.04973897337913513, + 0.03165537118911743, + 0.0639045313000679, + 0.006716765463352203, + 0.01163695752620697, + 0.00794089026749134, + 0.0063210888765752316, + 0.007650747429579496, + 0.007249974645674229, + 0.011025837622582912, + 0.010164344683289528, + 0.0029044190887361765, + 0.0036867542658001184, + 0.004061772488057613, + 0.003413714934140444, + 0.0028345438186079264, + 0.00716561172157526, + 0.0047593857161700726, + 0.005600464530289173, + 0.002312258817255497, + 0.003154275706037879, + 0.002364732325077057, + 0.0025539593771100044, + 0.002983781974762678, + 0.0020557083189487457, + 0.0026173328515142202, + 0.0030100762378424406, + 0.007105959113687277, + 0.006844803225249052, + 0.007414961699396372, + 0.005817543249577284, + 0.006249492987990379, + 0.007439142558723688, + 0.007210217881947756, + 0.010475967079401016, + 0.004475474823266268, + 0.0029629215132445097, + 0.005630678031593561, + 0.003423861460760236, + 0.0032165301963686943, + 0.005827273707836866, + 0.004602470435202122, + 0.004917442332953215, + 0.004887687508016825, + 0.0028150666039437056, + 0.005385782569646835, + 0.007747844327241182, + 0.0033142592292279005, + 0.00628749979659915, + 0.01192738488316536, + 0.005489211529493332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.6936310529708862 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teardrop-shaped candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle flame swirling into existence and spiraling upward from a wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle held in an antique brass holder with soft aged copper paste and a hand-etched floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame have a soft ethereal blue core with a warm golden glow on its periphery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing pinpoints of light detach from the flame and drift like fireflies in a dimly lit, tranquil room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635362558066845 + ], + "opposite_logprob_tokens": [ + -3.651353597640991 + ], + "logprob_reward": -0.02635362558066845, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635362558066845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2601298987865448, + 0.12993977963924408, + 0.10914649069309235, + 0.11412093043327332, + 0.08873092383146286, + 0.04750958830118179, + 0.04012918472290039, + 0.0740097314119339, + 0.06362182646989822, + 0.07266166806221008 + ], + "credit_map": [ + 0.04747316241264343, + 0.019726302474737167, + 0.03285606577992439, + 0.022580191493034363, + 0.014684617519378662, + 0.04984830319881439, + 0.021986139938235283, + 0.050975099205970764, + 0.01898360624909401, + 0.012109534814953804, + 0.02238265424966812, + 0.013777470216155052, + 0.011538119986653328, + 0.016756199300289154, + 0.018773438408970833, + 0.015618756413459778, + 0.012022478505969048, + 0.020904352888464928, + 0.013350153341889381, + 0.008914506994187832, + 0.016624627634882927, + 0.013671871274709702, + 0.012339859269559383, + 0.011318636126816273, + 0.011419802904129028, + 0.016126317903399467, + 0.015576938167214394, + 0.011049965396523476, + 0.021018672734498978, + 0.01241303887218237, + 0.01370159536600113, + 0.012814595364034176, + 0.009299232624471188, + 0.0075860340148210526, + 0.011429345235228539, + 0.007048571016639471, + 0.007565691601485014, + 0.01783190295100212, + 0.009916618466377258, + 0.018053531646728516, + 0.00834592618048191, + 0.005830508656799793, + 0.006918288767337799, + 0.0034734357614070177, + 0.004770659375935793, + 0.006967248860746622, + 0.004240110050886869, + 0.006963411346077919, + 0.004690119996666908, + 0.004768948536366224, + 0.004722381941974163, + 0.004782930016517639, + 0.005116159096360207, + 0.004324235953390598, + 0.006482111290097237, + 0.005242297425866127, + 0.009653124958276749, + 0.005763678811490536, + 0.0109219616279006, + 0.006058544386178255, + 0.007760209497064352, + 0.009672802872955799, + 0.010576787404716015, + 0.013602626509964466, + 0.012742340564727783, + 0.0058393399231135845, + 0.00944781769067049, + 0.006039412226527929, + 0.00687079131603241, + 0.007068999577313662, + 0.008968927897512913, + 0.006644199602305889, + 0.01224824320524931, + 0.005471652373671532, + 0.010803760960698128, + 0.008123640902340412, + 0.005089652258902788, + 0.00946976337581873, + 0.011731522157788277, + 0.009723438881337643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.02635362558066845 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teardrop-shaped candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle flame swirling into existence and spiraling upward from a wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle held in an antique brass holder with soft aged copper paste and a hand-etched floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame have a soft ethereal blue core with a warm golden glow on its periphery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing pinpoints of light detach from the flame and drift like fireflies in a dimly lit, tranquil room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296695470809937 + ], + "opposite_logprob_tokens": [ + -2.237967014312744 + ], + "logprob_reward": -0.11296695470809937, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296695470809937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14968280494213104, + 0.06957799196243286, + 0.15634779632091522, + 0.18257535994052887, + 0.08863165229558945, + 0.18011511862277985, + 0.026602545753121376, + 0.06949711591005325, + 0.05641714856028557, + 0.020552529022097588 + ], + "credit_map": [ + 0.029022542759776115, + 0.010691375471651554, + 0.020733648911118507, + 0.013615915551781654, + 0.006307198666036129, + 0.02710677869617939, + 0.011084071360528469, + 0.031121257692575455, + 0.015400239266455173, + 0.004585144575685263, + 0.012118577025830746, + 0.006171379238367081, + 0.005035433452576399, + 0.011188279837369919, + 0.0060591865330934525, + 0.009019741788506508, + 0.028530972078442574, + 0.03288232907652855, + 0.018864352256059647, + 0.010896935127675533, + 0.009763532318174839, + 0.024322625249624252, + 0.010312855243682861, + 0.02077416703104973, + 0.019140902906656265, + 0.01984674669802189, + 0.02016744390130043, + 0.017745807766914368, + 0.034776657819747925, + 0.021427428349852562, + 0.025965332984924316, + 0.023505020886659622, + 0.013450419530272484, + 0.00604657270014286, + 0.01063437107950449, + 0.005901603028178215, + 0.004583185538649559, + 0.021749569103121758, + 0.0057137166149914265, + 0.02055220492184162, + 0.025495480746030807, + 0.023043189197778702, + 0.024455031380057335, + 0.010628541931509972, + 0.014342106878757477, + 0.03732381761074066, + 0.012066801078617573, + 0.03276011720299721, + 0.0017780743073672056, + 0.001925764256156981, + 0.002626607660204172, + 0.0022093774750828743, + 0.0021807956509292126, + 0.005982379894703627, + 0.003756545018404722, + 0.006142998114228249, + 0.009031171910464764, + 0.0053157988004386425, + 0.013069660402834415, + 0.005433361046016216, + 0.005920817144215107, + 0.011112398467957973, + 0.008184765465557575, + 0.011429132893681526, + 0.01109995599836111, + 0.004405952990055084, + 0.007701221853494644, + 0.004075968638062477, + 0.006680953316390514, + 0.00845866184681654, + 0.0062463246285915375, + 0.007748101372271776, + 0.001945365802384913, + 0.0015629201661795378, + 0.002959650009870529, + 0.0029727225191891193, + 0.002118755131959915, + 0.0023833843879401684, + 0.004596385173499584, + 0.002013342920690775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.11296695470809937 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the cozy style of Ghibli animation, an extreme close-up, held static in the foreground, captures the moment a teardrop-shaped candle flame suddenly swirls into existence. It gracefully spirals upward from a wick in an antique brass holder, which features soft, aged copper paste and a delicate, hand-etched floral pattern. The flame itself is magical; its core radiates a soft, ethereal blue, while a warm golden glow envelops its periphery. As the flame swirls with fluid, balletic grace, it leaves a trail of gentle, glowing pinpoints of light that detach and drift lazily through the air like tiny, flickering fireflies. These particles cast a soft, dancing glow on their surroundings before fading into the soft-focus background of the dimly lit, tranquil room, creating an overall atmosphere of tranquility and captivating warmth.", + "seed": 8974, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_9/index_6131_seed_8974_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.14269724301993847, + "vlm_logprob": -0.14269724301993847, + "vlm_yes_no_margin": 2.575, + "vlm_token_credit": -0.14269724301993847, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teardrop-shaped candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle flame swirling into existence and spiraling upward from a wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle held in an antique brass holder with soft aged copper paste and a hand-etched floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame have a soft ethereal blue core with a warm golden glow on its periphery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing pinpoints of light detach from the flame and drift like fireflies in a dimly lit, tranquil room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04714913293719292 + ], + "opposite_logprob_tokens": [ + -3.172149181365967 + ], + "logprob_reward": -0.04714913293719292, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04714913293719292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12285169214010239, + 0.08406490087509155, + 0.2537992596626282, + 0.042618878185749054, + 0.05977543443441391, + 0.05905553326010704, + 0.06047665327787399, + 0.12572099268436432, + 0.1616753190755844, + 0.029961325228214264 + ], + "credit_map": [ + 0.010591489262878895, + 0.012515068985521793, + 0.03716618940234184, + 0.017595656216144562, + 0.007506061811000109, + 0.011060340330004692, + 0.015306271612644196, + 0.01111063826829195, + 0.008178764022886753, + 0.009829777292907238, + 0.02127540484070778, + 0.0092605110257864, + 0.005718591623008251, + 0.007588434033095837, + 0.012884913012385368, + 0.009328508749604225, + 0.021684028208255768, + 0.02858717180788517, + 0.06822092831134796, + 0.03862743079662323, + 0.011285176500678062, + 0.03698432445526123, + 0.021303797140717506, + 0.02710641361773014, + 0.003551532281562686, + 0.004830766934901476, + 0.007265227381139994, + 0.004855937324464321, + 0.003284217556938529, + 0.004595112055540085, + 0.009049094282090664, + 0.0051869917660951614, + 0.00709540443494916, + 0.012158319354057312, + 0.009102440439164639, + 0.010676632635295391, + 0.0026920989621430635, + 0.006805382668972015, + 0.005052879918366671, + 0.006192281376570463, + 0.005093646235764027, + 0.007929741404950619, + 0.009116889908909798, + 0.007882042787969112, + 0.006322908215224743, + 0.007152367848902941, + 0.007708746939897537, + 0.007849193178117275, + 0.005516081117093563, + 0.014462437480688095, + 0.007781394757330418, + 0.011944184079766273, + 0.0035888012498617172, + 0.007442940957844257, + 0.004066461697220802, + 0.005674355663359165, + 0.012433541938662529, + 0.023467572405934334, + 0.01888173818588257, + 0.015493734739720821, + 0.009763289242982864, + 0.02188532054424286, + 0.013639528304338455, + 0.010156277567148209, + 0.012671858072280884, + 0.030152911320328712, + 0.0325215645134449, + 0.0332879014313221, + 0.0073591857217252254, + 0.01872730627655983, + 0.00963004119694233, + 0.017324555665254593, + 0.0022358321584761143, + 0.0035056902561336756, + 0.006667773704975843, + 0.004199100658297539, + 0.003068238031119108, + 0.003723223228007555, + 0.0037753803189843893, + 0.002786090364679694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 7 + ], + "selected_reward": -0.04714913293719292 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teardrop-shaped candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle flame swirling into existence and spiraling upward from a wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle held in an antique brass holder with soft aged copper paste and a hand-etched floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame have a soft ethereal blue core with a warm golden glow on its periphery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing pinpoints of light detach from the flame and drift like fireflies in a dimly lit, tranquil room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0790163204073906 + ], + "opposite_logprob_tokens": [ + -2.5790162086486816 + ], + "logprob_reward": -0.0790163204073906, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0790163204073906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07812319695949554, + 0.09732536971569061, + 0.17732013761997223, + 0.10959314554929733, + 0.04099898785352707, + 0.108282670378685, + 0.09762416779994965, + 0.16548362374305725, + 0.06885126978158951, + 0.0563974529504776 + ], + "credit_map": [ + 0.0068361712619662285, + 0.008399865590035915, + 0.01895434595644474, + 0.011357326060533524, + 0.009404068812727928, + 0.00722033204510808, + 0.009132912382483482, + 0.006818172987550497, + 0.008451872505247593, + 0.009591225534677505, + 0.011876619420945644, + 0.014446905814111233, + 0.009995212778449059, + 0.012477708980441093, + 0.013657482340931892, + 0.01682833582162857, + 0.01489234808832407, + 0.020245520398020744, + 0.049595050513744354, + 0.02740398608148098, + 0.007431476842612028, + 0.02517237327992916, + 0.016157379373908043, + 0.016422007232904434, + 0.008035445585846901, + 0.017133381217718124, + 0.016400689259171486, + 0.01656157337129116, + 0.004923900123685598, + 0.015500513836741447, + 0.01771130971610546, + 0.013326331973075867, + 0.0038105188868939877, + 0.005967369768768549, + 0.005973611027002335, + 0.008209790103137493, + 0.0027161408215761185, + 0.005822767037898302, + 0.004646300803869963, + 0.0038524908013641834, + 0.009269457310438156, + 0.01396402157843113, + 0.022370923310518265, + 0.01662120781838894, + 0.013386912643909454, + 0.010507686994969845, + 0.014324643649160862, + 0.007837818004190922, + 0.009951712563633919, + 0.027802567929029465, + 0.010912829078733921, + 0.016843527555465698, + 0.005920879077166319, + 0.012821843847632408, + 0.005632034502923489, + 0.007738769985735416, + 0.013690754771232605, + 0.028591696172952652, + 0.02662690356373787, + 0.02230922505259514, + 0.012467452324926853, + 0.030216924846172333, + 0.015935104340314865, + 0.015645554289221764, + 0.005845753476023674, + 0.010246366262435913, + 0.01084454357624054, + 0.010858364403247833, + 0.0032247621566057205, + 0.00962943583726883, + 0.005409929435700178, + 0.012792118825018406, + 0.004149778746068478, + 0.007454887498170137, + 0.008861202746629715, + 0.013672402128577232, + 0.003584609366953373, + 0.006534038111567497, + 0.005831798072904348, + 0.006308731157332659 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 3 + ], + "selected_reward": -0.0790163204073906 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teardrop-shaped candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle flame swirling into existence and spiraling upward from a wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle held in an antique brass holder with soft aged copper paste and a hand-etched floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame have a soft ethereal blue core with a warm golden glow on its periphery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing pinpoints of light detach from the flame and drift like fireflies in a dimly lit, tranquil room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5237984657287598 + ], + "opposite_logprob_tokens": [ + -0.8987984657287598 + ], + "logprob_reward": -0.5237984657287598, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5237984657287598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14407041668891907, + 0.0880475714802742, + 0.523079514503479, + 0.03739669919013977, + 0.05803047493100166, + 0.02570282854139805, + 0.03341435268521309, + 0.026954423636198044, + 0.029454629868268967, + 0.03384910896420479 + ], + "credit_map": [ + 0.01712188683450222, + 0.024843167513608932, + 0.020623469725251198, + 0.019707091152668, + 0.006282373797148466, + 0.019292671233415604, + 0.009411006234586239, + 0.026788752526044846, + 0.00956962164491415, + 0.01178925670683384, + 0.024251682683825493, + 0.010731739923357964, + 0.006029240787029266, + 0.00971712451428175, + 0.007418734487146139, + 0.008540171198546886, + 0.0477980375289917, + 0.059388842433691025, + 0.16215653717517853, + 0.07945461571216583, + 0.01936757192015648, + 0.06782449781894684, + 0.04075533524155617, + 0.046334102749824524, + 0.003233000636100769, + 0.00648542633280158, + 0.004585247952491045, + 0.0054810489527881145, + 0.0017629122594371438, + 0.0063201505690813065, + 0.004475661087781191, + 0.005053252913057804, + 0.006697461009025574, + 0.01322967279702425, + 0.007861271500587463, + 0.011300581507384777, + 0.0033607417717576027, + 0.005678432993590832, + 0.005046512931585312, + 0.0048558032140135765, + 0.0019594107288867235, + 0.002685351762920618, + 0.003435624297708273, + 0.0036010874900966883, + 0.0036298250779509544, + 0.002876541344448924, + 0.004555332474410534, + 0.0029596551321446896, + 0.0023297383449971676, + 0.003922143019735813, + 0.005007273983210325, + 0.008700082078576088, + 0.0025369636714458466, + 0.0030036496464163065, + 0.004364495165646076, + 0.003550008637830615, + 0.0020801417995244265, + 0.005482178647071123, + 0.0049879467114806175, + 0.004202656447887421, + 0.0016997489146888256, + 0.004155187401920557, + 0.002026340691372752, + 0.0023202234879136086, + 0.002525694901123643, + 0.007421474438160658, + 0.004247159697115421, + 0.004025575704872608, + 0.0016308431513607502, + 0.003839473007246852, + 0.00150469527579844, + 0.004259712528437376, + 0.0030924687162041664, + 0.006243494339287281, + 0.008977076970040798, + 0.005410246085375547, + 0.0020600880961865187, + 0.0026894148904830217, + 0.002791264094412327, + 0.0025850553065538406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.5237984657287598 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teardrop-shaped candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle flame swirling into existence and spiraling upward from a wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle held in an antique brass holder with soft aged copper paste and a hand-etched floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame have a soft ethereal blue core with a warm golden glow on its periphery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing pinpoints of light detach from the flame and drift like fireflies in a dimly lit, tranquil room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370928019285202 + ], + "opposite_logprob_tokens": [ + -3.4087092876434326 + ], + "logprob_reward": -0.03370928019285202, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370928019285202, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09163384884595871, + 0.36908552050590515, + 0.20899882912635803, + 0.04576616361737251, + 0.02663496509194374, + 0.05345813184976578, + 0.03175961226224899, + 0.08186718821525574, + 0.07064002007246017, + 0.020155714824795723 + ], + "credit_map": [ + 0.01057848148047924, + 0.013103482313454151, + 0.013303268700838089, + 0.011341813020408154, + 0.0052457512356340885, + 0.012930313125252724, + 0.009082037955522537, + 0.01604870706796646, + 0.02911175973713398, + 0.04591180011630058, + 0.09419999271631241, + 0.05083591490983963, + 0.016737302765250206, + 0.048792511224746704, + 0.023494651541113853, + 0.06000160425901413, + 0.019778361544013023, + 0.024194594472646713, + 0.06350278854370117, + 0.030683958902955055, + 0.007815745659172535, + 0.026033738628029823, + 0.01833026297390461, + 0.018659401684999466, + 0.0041347770020365715, + 0.008827781304717064, + 0.006531167309731245, + 0.006946471054106951, + 0.002253912854939699, + 0.00687234615907073, + 0.003770348383113742, + 0.006429360713809729, + 0.002011500298976898, + 0.003287248546257615, + 0.0035208645276725292, + 0.004403796512633562, + 0.0032667883206158876, + 0.003956216853111982, + 0.0033511065412312746, + 0.002837445819750428, + 0.004935247357934713, + 0.0074548544362187386, + 0.010779502801597118, + 0.006383400876075029, + 0.00617156783118844, + 0.005526312161237001, + 0.006862711161375046, + 0.0053445384837687016, + 0.002231842139735818, + 0.0051267207600176334, + 0.004794437438249588, + 0.005433159880340099, + 0.0026022500824183226, + 0.005689975805580616, + 0.0025038989260792732, + 0.0033773304894566536, + 0.007000760640949011, + 0.009611123241484165, + 0.009595059789717197, + 0.010494847781956196, + 0.007204849738627672, + 0.02299417369067669, + 0.00674852542579174, + 0.008217851631343365, + 0.005511600058525801, + 0.018871275708079338, + 0.009343286976218224, + 0.01065453514456749, + 0.004427253734320402, + 0.009117597714066505, + 0.004325784742832184, + 0.008388692513108253, + 0.0015521704917773604, + 0.0034514639992266893, + 0.004064972512423992, + 0.002988033462315798, + 0.001818695804104209, + 0.002174744848161936, + 0.0022571429144591093, + 0.0018484903266653419 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.03370928019285202 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teardrop-shaped candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle flame swirling into existence and spiraling upward from a wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle held in an antique brass holder with soft aged copper paste and a hand-etched floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame have a soft ethereal blue core with a warm golden glow on its periphery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing pinpoints of light detach from the flame and drift like fireflies in a dimly lit, tranquil room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029813015833497047 + ], + "opposite_logprob_tokens": [ + -3.529813051223755 + ], + "logprob_reward": -0.029813015833497047, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029813015833497047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04824899137020111, + 0.16957126557826996, + 0.11199118942022324, + 0.09434628486633301, + 0.04936060681939125, + 0.11575645953416824, + 0.12435510754585266, + 0.2183118313550949, + 0.048622600734233856, + 0.01943564973771572 + ], + "credit_map": [ + 0.004386346321552992, + 0.005261565558612347, + 0.01090897899121046, + 0.00544024258852005, + 0.0033377145882695913, + 0.006011442746967077, + 0.00462360680103302, + 0.00827909354120493, + 0.01275530643761158, + 0.02146366611123085, + 0.04147328436374664, + 0.023062782362103462, + 0.008654894307255745, + 0.022364895790815353, + 0.01140674389898777, + 0.02838969975709915, + 0.00974334217607975, + 0.01511840894818306, + 0.028409192338585854, + 0.01891077682375908, + 0.004657890181988478, + 0.015136856585741043, + 0.010431759059429169, + 0.00958295539021492, + 0.007960734888911247, + 0.014144635759294033, + 0.02020004391670227, + 0.0136483833193779, + 0.0038796295411884785, + 0.01598796248435974, + 0.006706838496029377, + 0.011818060651421547, + 0.00528462091460824, + 0.008318188600242138, + 0.008403057232499123, + 0.007699055131524801, + 0.003330711741000414, + 0.006517454981803894, + 0.005540687590837479, + 0.004266828764230013, + 0.008340103551745415, + 0.016393881291151047, + 0.01798255369067192, + 0.01942671835422516, + 0.014836770482361317, + 0.018411429598927498, + 0.010594796389341354, + 0.009770194999873638, + 0.008530369028449059, + 0.013757613487541676, + 0.014716412872076035, + 0.02512369491159916, + 0.009158764034509659, + 0.028892841190099716, + 0.00790672842413187, + 0.016268689185380936, + 0.016250036656856537, + 0.029249632731080055, + 0.02887417934834957, + 0.02343256212770939, + 0.017056019976735115, + 0.05823443830013275, + 0.025130413472652435, + 0.020084545016288757, + 0.0038749652449041605, + 0.007113240193575621, + 0.008687552064657211, + 0.008233561180531979, + 0.002340201288461685, + 0.007153934799134731, + 0.00305059808306396, + 0.008168547414243221, + 0.0021706700790673494, + 0.002384905470535159, + 0.0030344906263053417, + 0.004325646907091141, + 0.0013587330467998981, + 0.0026933415792882442, + 0.0017757314490154386, + 0.0016921315109357238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 6 + ], + "selected_reward": -0.029813015833497047 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the cozy style of Ghibli animation, an extreme close-up, held static in the foreground, captures the moment a teardrop-shaped candle flame suddenly swirls into existence. It gracefully spirals upward from a wick in an antique brass holder, which features soft, aged copper paste and a delicate, hand-etched floral pattern. The flame itself is magical; its core radiates a soft, ethereal blue, while a warm golden glow envelops its periphery. As the flame swirls with fluid, balletic grace, it leaves a trail of gentle, glowing pinpoints of light that detach and drift lazily through the air like tiny, flickering fireflies. These particles cast a soft, dancing glow on their surroundings before fading into the soft-focus background of the dimly lit, tranquil room, creating an overall atmosphere of tranquility and captivating warmth.", + "seed": 8975, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_9/index_6131_seed_8975_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1629682168364525, + "vlm_logprob": -0.1629682168364525, + "vlm_yes_no_margin": 2.225, + "vlm_token_credit": -0.1629682168364525, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teardrop-shaped candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle flame swirling into existence and spiraling upward from a wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle held in an antique brass holder with soft aged copper paste and a hand-etched floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame have a soft ethereal blue core with a warm golden glow on its periphery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing pinpoints of light detach from the flame and drift like fireflies in a dimly lit, tranquil room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12046001851558685 + ], + "opposite_logprob_tokens": [ + -2.245460033416748 + ], + "logprob_reward": -0.12046001851558685, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.12046001851558685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08824627101421356, + 0.06744661927223206, + 0.1758785992860794, + 0.04808122292160988, + 0.04171476885676384, + 0.022191904485225677, + 0.23723271489143372, + 0.22084234654903412, + 0.041673023253679276, + 0.0566924549639225 + ], + "credit_map": [ + 0.01162718515843153, + 0.010643058456480503, + 0.013912814669311047, + 0.009121797978878021, + 0.006664714310318232, + 0.011076279915869236, + 0.010148903355002403, + 0.015051526948809624, + 0.007154929451644421, + 0.009025096893310547, + 0.010597475804388523, + 0.006417489144951105, + 0.005648325197398663, + 0.01228433009237051, + 0.0068956702016294, + 0.009423309937119484, + 0.017964141443371773, + 0.03932705894112587, + 0.016793910413980484, + 0.014386308379471302, + 0.028981758281588554, + 0.019016901031136513, + 0.018117917701601982, + 0.02129061333835125, + 0.006587884854525328, + 0.006999570410698652, + 0.005422710906714201, + 0.004635527729988098, + 0.005853571929037571, + 0.005290182773023844, + 0.004873495548963547, + 0.008418278768658638, + 0.005845365580171347, + 0.01039873342961073, + 0.004502513445913792, + 0.003302086377516389, + 0.00379967805929482, + 0.005858955904841423, + 0.004100625403225422, + 0.0039068106561899185, + 0.0020359144546091557, + 0.003546161577105522, + 0.003094922751188278, + 0.002471767831593752, + 0.00275225518271327, + 0.003805861109867692, + 0.0025445667561143637, + 0.0019404554041102529, + 0.02588994801044464, + 0.0497257336974144, + 0.026201894506812096, + 0.017418526113033295, + 0.038448624312877655, + 0.02763521671295166, + 0.02200389839708805, + 0.029908888041973114, + 0.027873922139406204, + 0.04361279308795929, + 0.027645276859402657, + 0.012561175040900707, + 0.031186241656541824, + 0.028644220903515816, + 0.020755628123879433, + 0.028563102707266808, + 0.004977511242032051, + 0.007252020761370659, + 0.0062674591317772865, + 0.002937993500381708, + 0.004292900208383799, + 0.007047455757856369, + 0.003578945528715849, + 0.00531874131411314, + 0.005532937590032816, + 0.007746296934783459, + 0.007581070531159639, + 0.0064694820903241634, + 0.006190601736307144, + 0.008350007236003876, + 0.006706887856125832, + 0.008115176111459732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 2 + ], + "selected_reward": -0.12046001851558685 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teardrop-shaped candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle flame swirling into existence and spiraling upward from a wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle held in an antique brass holder with soft aged copper paste and a hand-etched floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame have a soft ethereal blue core with a warm golden glow on its periphery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing pinpoints of light detach from the flame and drift like fireflies in a dimly lit, tranquil room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07902711629867554 + ], + "opposite_logprob_tokens": [ + -2.5790271759033203 + ], + "logprob_reward": -0.07902711629867554, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07902711629867554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08130259811878204, + 0.02500925399363041, + 0.18680235743522644, + 0.08493363112211227, + 0.0543997548520565, + 0.05674543231725693, + 0.32489508390426636, + 0.1292399764060974, + 0.021977422758936882, + 0.034694500267505646 + ], + "credit_map": [ + 0.010639185085892677, + 0.006077507510781288, + 0.011040864512324333, + 0.00771294953301549, + 0.00581233948469162, + 0.014574279077351093, + 0.008806637488305569, + 0.01663883775472641, + 0.0024848375469446182, + 0.004013509023934603, + 0.003371576312929392, + 0.0029101017862558365, + 0.0021662709768861532, + 0.0029505151323974133, + 0.004773201886564493, + 0.002339240862056613, + 0.018449431285262108, + 0.046840935945510864, + 0.019297543913125992, + 0.015466703101992607, + 0.03174879774451256, + 0.021006522700190544, + 0.01739436201751232, + 0.01659807562828064, + 0.009159693494439125, + 0.018419688567519188, + 0.008879671804606915, + 0.009376349858939648, + 0.013328316621482372, + 0.009749861434102058, + 0.008954438380897045, + 0.0070656160824000835, + 0.005473804660141468, + 0.010094682686030865, + 0.0063917082734405994, + 0.004209166392683983, + 0.009109189733862877, + 0.005913980770856142, + 0.005970154423266649, + 0.007237068843096495, + 0.007499121595174074, + 0.008832978084683418, + 0.008032022044062614, + 0.0038684681057929993, + 0.008179835043847561, + 0.009431344456970692, + 0.004883185960352421, + 0.006018477492034435, + 0.03327513113617897, + 0.0711141973733902, + 0.038554511964321136, + 0.024518465623259544, + 0.0444173626601696, + 0.05546240508556366, + 0.03259500116109848, + 0.02495799958705902, + 0.016528775915503502, + 0.024836216121912003, + 0.016307581216096878, + 0.00756431557238102, + 0.01782364770770073, + 0.016770849004387856, + 0.01315795537084341, + 0.01625063084065914, + 0.002676567295566201, + 0.0051845829002559185, + 0.002264951588585973, + 0.0015934306429699063, + 0.0030042659491300583, + 0.00252495682798326, + 0.0018148759845644236, + 0.0029137921519577503, + 0.005065718200057745, + 0.004594766069203615, + 0.00467959139496088, + 0.0024881574790924788, + 0.002730644540861249, + 0.005859047640115023, + 0.0032786023803055286, + 0.005997973959892988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 7 + ], + "selected_reward": -0.07902711629867554 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teardrop-shaped candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle flame swirling into existence and spiraling upward from a wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle held in an antique brass holder with soft aged copper paste and a hand-etched floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame have a soft ethereal blue core with a warm golden glow on its periphery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing pinpoints of light detach from the flame and drift like fireflies in a dimly lit, tranquil room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5236147046089172 + ], + "opposite_logprob_tokens": [ + -0.8986147046089172 + ], + "logprob_reward": -0.5236147046089172, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5236147046089172, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12687452137470245, + 0.04643303155899048, + 0.10988719016313553, + 0.02992120571434498, + 0.20696966350078583, + 0.05523047596216202, + 0.3427140712738037, + 0.03127247467637062, + 0.026442600414156914, + 0.024254845455288887 + ], + "credit_map": [ + 0.015094438567757607, + 0.008096267469227314, + 0.018211305141448975, + 0.01577519066631794, + 0.010382762178778648, + 0.014244163408875465, + 0.023123392835259438, + 0.021947013214230537, + 0.0061150421388447285, + 0.0038555797655135393, + 0.006565062794834375, + 0.004036756698042154, + 0.004268139600753784, + 0.00501194316893816, + 0.006290620192885399, + 0.010289883241057396, + 0.013114255852997303, + 0.026703182607889175, + 0.010148582048714161, + 0.007779471110552549, + 0.019256632775068283, + 0.01199374720454216, + 0.009921901859343052, + 0.010969418101012707, + 0.002927553141489625, + 0.006893116515129805, + 0.0033272872678935528, + 0.0019325183238834143, + 0.0045928386971354485, + 0.003263500751927495, + 0.003603848395869136, + 0.003380543552339077, + 0.02422110177576542, + 0.05013004690408707, + 0.023252207785844803, + 0.016500692814588547, + 0.020842203870415688, + 0.03717824071645737, + 0.016680454835295677, + 0.018164725974202156, + 0.004906876944005489, + 0.008742283098399639, + 0.008314608596265316, + 0.0039531029760837555, + 0.006015188526362181, + 0.013084820471704006, + 0.004978254437446594, + 0.005235334392637014, + 0.037400227040052414, + 0.07403645664453506, + 0.035963334143161774, + 0.02369152195751667, + 0.0575273260474205, + 0.04024885222315788, + 0.0359644815325737, + 0.03788182511925697, + 0.004054415505379438, + 0.0067651295103132725, + 0.003999689593911171, + 0.00214503426104784, + 0.005106485914438963, + 0.003271824447438121, + 0.0031928245443850756, + 0.002737069735303521, + 0.00318131479434669, + 0.007047200575470924, + 0.002748221857473254, + 0.001871339394710958, + 0.003830516943708062, + 0.0025592627935111523, + 0.002320903353393078, + 0.0028838403522968292, + 0.002918204991146922, + 0.0023072005715221167, + 0.003706587478518486, + 0.00229846709407866, + 0.0023726015351712704, + 0.0033378570806235075, + 0.0029694645199924707, + 0.004344462417066097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.5236147046089172 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teardrop-shaped candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle flame swirling into existence and spiraling upward from a wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle held in an antique brass holder with soft aged copper paste and a hand-etched floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame have a soft ethereal blue core with a warm golden glow on its periphery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing pinpoints of light detach from the flame and drift like fireflies in a dimly lit, tranquil room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048665110021829605 + ], + "opposite_logprob_tokens": [ + -3.0486650466918945 + ], + "logprob_reward": -0.048665110021829605, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048665110021829605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08824601769447327, + 0.1240622028708458, + 0.17133450508117676, + 0.0532478392124176, + 0.0702279731631279, + 0.07388123869895935, + 0.18654479086399078, + 0.11203267425298691, + 0.054252445697784424, + 0.06617032736539841 + ], + "credit_map": [ + 0.008404728025197983, + 0.011377409100532532, + 0.01245891209691763, + 0.00979667529463768, + 0.008943643420934677, + 0.01649373583495617, + 0.008171902038156986, + 0.012599010020494461, + 0.01520322822034359, + 0.012954079546034336, + 0.020816389471292496, + 0.009017370641231537, + 0.009743602015078068, + 0.021427134051918983, + 0.010089416056871414, + 0.024810990318655968, + 0.015216892585158348, + 0.04252909496426582, + 0.015811022371053696, + 0.015341328456997871, + 0.03164922446012497, + 0.017679229378700256, + 0.01698714680969715, + 0.016120558604598045, + 0.004568887408822775, + 0.011600416153669357, + 0.005101413000375032, + 0.005991781130433083, + 0.007787076756358147, + 0.0045659807510674, + 0.008864380419254303, + 0.004767902195453644, + 0.006958896294236183, + 0.013119826093316078, + 0.008550870232284069, + 0.004220865201205015, + 0.009399130940437317, + 0.012286962941288948, + 0.007108695339411497, + 0.008582726120948792, + 0.010468022897839546, + 0.013214790262281895, + 0.008805687539279461, + 0.004976233467459679, + 0.011961005628108978, + 0.010537487454712391, + 0.005964653100818396, + 0.007953359745442867, + 0.020245982334017754, + 0.03872518613934517, + 0.021318038925528526, + 0.012426306493580341, + 0.03346489742398262, + 0.01915218122303486, + 0.017277322709560394, + 0.023934876546263695, + 0.012102566659450531, + 0.019433127716183662, + 0.015173603780567646, + 0.008049564436078072, + 0.015130224637687206, + 0.018149198964238167, + 0.01209681760519743, + 0.011897564865648746, + 0.0062212818302214146, + 0.011050083674490452, + 0.0065977443009614944, + 0.004610816482454538, + 0.007368334103375673, + 0.007120166439563036, + 0.0040704975835978985, + 0.007213517092168331, + 0.008022344671189785, + 0.007806223817169666, + 0.009502955712378025, + 0.008573477156460285, + 0.005627989768981934, + 0.009862435981631279, + 0.0064513483084738255, + 0.010323550552129745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 1 + ], + "selected_reward": -0.048665110021829605 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teardrop-shaped candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle flame swirling into existence and spiraling upward from a wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle held in an antique brass holder with soft aged copper paste and a hand-etched floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame have a soft ethereal blue core with a warm golden glow on its periphery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing pinpoints of light detach from the flame and drift like fireflies in a dimly lit, tranquil room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307413473725319 + ], + "opposite_logprob_tokens": [ + -3.168074131011963 + ], + "logprob_reward": -0.04307413473725319, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307413473725319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08254782855510712, + 0.06971849501132965, + 0.29890337586402893, + 0.035764098167419434, + 0.04057725891470909, + 0.05147004500031471, + 0.2967011034488678, + 0.06486501544713974, + 0.01670938916504383, + 0.04274337366223335 + ], + "credit_map": [ + 0.01167084276676178, + 0.00861621554940939, + 0.012767472304403782, + 0.006320972926914692, + 0.006875925697386265, + 0.013687563128769398, + 0.008000865578651428, + 0.014607972465455532, + 0.006846780423074961, + 0.013617358170449734, + 0.009094213135540485, + 0.006250876933336258, + 0.008290440775454044, + 0.010405946522951126, + 0.008394737727940083, + 0.006818144116550684, + 0.02615109458565712, + 0.08065148442983627, + 0.03079221583902836, + 0.02743249014019966, + 0.0544680617749691, + 0.02753349579870701, + 0.02811102196574211, + 0.02376350201666355, + 0.004115072079002857, + 0.007117293309420347, + 0.0038232754450291395, + 0.0034883136395365, + 0.00474256556481123, + 0.0033510876819491386, + 0.005282492842525244, + 0.0038439957424998283, + 0.004506266210228205, + 0.0064483825117349625, + 0.003703269176185131, + 0.0038165294099599123, + 0.006498699542135, + 0.004056503530591726, + 0.005947600584477186, + 0.005600009113550186, + 0.00635463697835803, + 0.011213632300496101, + 0.005439169239252806, + 0.0038332785479724407, + 0.010420538485050201, + 0.005233703646808863, + 0.004669083282351494, + 0.00430600019171834, + 0.03107968531548977, + 0.07048983871936798, + 0.03562214598059654, + 0.02646050415933132, + 0.0364193469285965, + 0.046492017805576324, + 0.02632952108979225, + 0.023808056488633156, + 0.008401506580412388, + 0.012751509435474873, + 0.008450131863355637, + 0.004317146725952625, + 0.008685648441314697, + 0.007773835211992264, + 0.006276008207350969, + 0.008209225721657276, + 0.0021238422486931086, + 0.0022336726542562246, + 0.0021863076835870743, + 0.0010617817752063274, + 0.001661120681092143, + 0.0027432022616267204, + 0.0018395690713077784, + 0.0028598932549357414, + 0.0058340164832770824, + 0.004366008099168539, + 0.006090685725212097, + 0.005412745289504528, + 0.0035750975366681814, + 0.005935784429311752, + 0.005661093629896641, + 0.0058679417707026005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.04307413473725319 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_0/index_tensor([6838, 6838])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_0/index_tensor([6838, 6838])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6b2ac9a05a8bea35e74ab143742779cfffa20516 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_0/index_tensor([6838, 6838])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, slowly tilting upward shot begins, gradually revealing a dense jungle vine cascading down a steep, damp rock face. The ancient stone, dark gray, damp, and smooth, is almost entirely covered in a vibrant, velvety green moss that clings to its rough surface. The vine itself is a complex, tangled web of gnarled, woody vines and softer, leafy climbers, their dark green and earthy brown forms intertwined, hanging heavily from an unseen point above. Water perpetually seeps from the rock, flowing in sparkling streams along the moss and vines. As warm, filtered sunlight penetrates the dense jungle canopy, it illuminates the scene, causing countless crystal droplets clinging to every leaf and tendril to shimmer like tiny suspended diamonds. The deliberate upward movement of the camera enhances the scale and majesty of the natural formations, creating a serene and magical atmosphere of pristine, untouched beauty.", + "seed": 31328, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_0/index_6838_seed_31328_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05418249531649053, + "vlm_logprob": -0.05418249531649053, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.05418249531649053, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a steep rock face visible that is dark gray, damp, and smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rock face covered in vibrant, velvety green moss and a complex web of dark green and earthy brown vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparkling streams of water flowing along the moss and vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, filtered sunlight causing crystal droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06791669130325317 + ], + "opposite_logprob_tokens": [ + -2.8179166316986084 + ], + "logprob_reward": -0.06791669130325317, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06791669130325317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13934527337551117, + 0.10915544629096985, + 0.13460399210453033, + 0.11866962164640427, + 0.15793898701667786, + 0.08135025948286057, + 0.11447342485189438, + 0.05014229938387871, + 0.03627699241042137, + 0.058043740689754486 + ], + "credit_map": [ + 0.011511017568409443, + 0.009451465681195259, + 0.012203408405184746, + 0.013043594546616077, + 0.02116534858942032, + 0.026438619941473007, + 0.02143123745918274, + 0.024100570008158684, + 0.008778863586485386, + 0.008282775990664959, + 0.008600609377026558, + 0.009493947960436344, + 0.01427181251347065, + 0.025621801614761353, + 0.01605546474456787, + 0.018050171434879303, + 0.009319843724370003, + 0.010512066073715687, + 0.013237499631941319, + 0.017923228442668915, + 0.02296784706413746, + 0.024291057139635086, + 0.018696188926696777, + 0.01765625737607479, + 0.005042233504354954, + 0.0062858047895133495, + 0.008982956409454346, + 0.00978236086666584, + 0.010060866363346577, + 0.033353742212057114, + 0.010164476931095123, + 0.034997180104255676, + 0.011665730737149715, + 0.01509038731455803, + 0.03293609619140625, + 0.015413490124046803, + 0.02229108288884163, + 0.024669576436281204, + 0.018728865310549736, + 0.01714375987648964, + 0.00562688335776329, + 0.006736343260854483, + 0.013066352345049381, + 0.009643098339438438, + 0.008419488556683064, + 0.013050556182861328, + 0.012141597457230091, + 0.01266593299806118, + 0.004693980328738689, + 0.008271715603768826, + 0.00873278733342886, + 0.011097419075667858, + 0.02390894666314125, + 0.01771385595202446, + 0.02348902076482773, + 0.016565704718232155, + 0.006388556212186813, + 0.003914061933755875, + 0.01155855506658554, + 0.004531474784016609, + 0.005187997594475746, + 0.005689350422471762, + 0.005338072311133146, + 0.0075342305935919285, + 0.002600808162242174, + 0.003104446455836296, + 0.003013454843312502, + 0.004972045309841633, + 0.004884624388068914, + 0.004843206610530615, + 0.00790021289139986, + 0.004958195611834526, + 0.006203662138432264, + 0.005115767475217581, + 0.005562739912420511, + 0.005819992627948523, + 0.006535250227898359, + 0.008072949014604092, + 0.011064917780458927, + 0.009668458253145218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.06791669130325317 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a steep rock face visible that is dark gray, damp, and smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rock face covered in vibrant, velvety green moss and a complex web of dark green and earthy brown vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparkling streams of water flowing along the moss and vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, filtered sunlight causing crystal droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270606517791748 + ], + "opposite_logprob_tokens": [ + -2.127060651779175 + ], + "logprob_reward": -0.1270606517791748, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270606517791748, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11822718381881714, + 0.10074420273303986, + 0.17977607250213623, + 0.18787921965122223, + 0.11972096562385559, + 0.11010602116584778, + 0.037115588784217834, + 0.0802706778049469, + 0.02355170249938965, + 0.042608343064785004 + ], + "credit_map": [ + 0.011401216499507427, + 0.010187809355556965, + 0.011288834735751152, + 0.010188576765358448, + 0.009665172547101974, + 0.02238616906106472, + 0.01376927550882101, + 0.029340116307139397, + 0.009749582037329674, + 0.008349377661943436, + 0.01098535768687725, + 0.009343700483441353, + 0.012289394624531269, + 0.019644837826490402, + 0.014034146443009377, + 0.016347788274288177, + 0.012696363031864166, + 0.015018531121313572, + 0.018601424992084503, + 0.0233319029211998, + 0.03184714913368225, + 0.0296284481883049, + 0.0245332233607769, + 0.024119028821587563, + 0.006626579910516739, + 0.009709826670587063, + 0.01430613175034523, + 0.01824156381189823, + 0.026622150093317032, + 0.047817882150411606, + 0.01776782050728798, + 0.04678725451231003, + 0.005555048584938049, + 0.012387228198349476, + 0.008624779060482979, + 0.014242589473724365, + 0.0286078043282032, + 0.01822403073310852, + 0.01566077582538128, + 0.01641869731247425, + 0.006911615841090679, + 0.010396608151495457, + 0.012950685806572437, + 0.012369116768240929, + 0.01839294470846653, + 0.016467055305838585, + 0.016139719635248184, + 0.016478268429636955, + 0.0027962839230895042, + 0.002463954035192728, + 0.0036153856199234724, + 0.005257914308458567, + 0.006438429467380047, + 0.005805772729218006, + 0.0056603942066431046, + 0.005077450070530176, + 0.007969195023179054, + 0.006254877429455519, + 0.010485601611435413, + 0.009749682620167732, + 0.013419842347502708, + 0.01197004783898592, + 0.010736992582678795, + 0.009684433229267597, + 0.0025007950607687235, + 0.0017028739675879478, + 0.0017759406473487616, + 0.003223980776965618, + 0.0036018446553498507, + 0.002978406846523285, + 0.004746563732624054, + 0.0030212949495762587, + 0.00279813539236784, + 0.003071455517783761, + 0.0036039368715137243, + 0.0044334204867482185, + 0.0067267464473843575, + 0.007505612447857857, + 0.006545586045831442, + 0.007923447526991367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 4 + ], + "selected_reward": -0.1270606517791748 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a steep rock face visible that is dark gray, damp, and smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rock face covered in vibrant, velvety green moss and a complex web of dark green and earthy brown vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparkling streams of water flowing along the moss and vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, filtered sunlight causing crystal droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013412775006145239 + ], + "opposite_logprob_tokens": [ + -6.626341342926025 + ], + "logprob_reward": -0.0013412775006145239, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013412775006145239, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14748823642730713, + 0.06939582526683807, + 0.1314489096403122, + 0.18809537589550018, + 0.06788085401058197, + 0.1168065294623375, + 0.09366309642791748, + 0.09028751403093338, + 0.03182555362582207, + 0.06310809403657913 + ], + "credit_map": [ + 0.014488637447357178, + 0.015695543959736824, + 0.014338002540171146, + 0.01937650702893734, + 0.014887077733874321, + 0.026476966217160225, + 0.018992114812135696, + 0.023233387619256973, + 0.005283478181809187, + 0.00665223877876997, + 0.008270595222711563, + 0.006288664881139994, + 0.007419810630381107, + 0.014086802490055561, + 0.00861933920532465, + 0.012774896807968616, + 0.01027897372841835, + 0.011941567994654179, + 0.013593181036412716, + 0.014169570989906788, + 0.01881279982626438, + 0.022253550589084625, + 0.021858517080545425, + 0.018540741875767708, + 0.009469403885304928, + 0.007423019967973232, + 0.016471022740006447, + 0.009810386225581169, + 0.01312361191958189, + 0.05890905484557152, + 0.013321041129529476, + 0.059567827731370926, + 0.0037798720877617598, + 0.005071565043181181, + 0.0070266276597976685, + 0.008237743750214577, + 0.014024997130036354, + 0.011343162506818771, + 0.008861870504915714, + 0.009535016492009163, + 0.007046164013445377, + 0.010852961800992489, + 0.010766967199742794, + 0.012811137363314629, + 0.021880317479372025, + 0.018641632050275803, + 0.016801781952381134, + 0.018005570396780968, + 0.004558046348392963, + 0.007937518879771233, + 0.009616033174097538, + 0.011181337758898735, + 0.019695837050676346, + 0.01207874994724989, + 0.016724886372685432, + 0.011870680376887321, + 0.00767147820442915, + 0.007978956215083599, + 0.009402110241353512, + 0.00996239110827446, + 0.015268821269273758, + 0.01346780639141798, + 0.013436478562653065, + 0.013099472038447857, + 0.002955605974420905, + 0.0029893219470977783, + 0.002338420134037733, + 0.004939745645970106, + 0.004149476531893015, + 0.004158176947385073, + 0.005799976643174887, + 0.004494829569011927, + 0.004389354959130287, + 0.005191562697291374, + 0.004209443926811218, + 0.00661878427490592, + 0.00989683996886015, + 0.008728770539164543, + 0.013112682849168777, + 0.010960652492940426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.0013412775006145239 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a steep rock face visible that is dark gray, damp, and smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rock face covered in vibrant, velvety green moss and a complex web of dark green and earthy brown vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparkling streams of water flowing along the moss and vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, filtered sunlight causing crystal droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205803155899048 + ], + "opposite_logprob_tokens": [ + -2.8120579719543457 + ], + "logprob_reward": -0.06205803155899048, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205803155899048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19524767994880676, + 0.09868480265140533, + 0.04336947575211525, + 0.20119944214820862, + 0.06272494792938232, + 0.056679196655750275, + 0.07093197852373123, + 0.07800762355327606, + 0.033722441643476486, + 0.1594323366880417 + ], + "credit_map": [ + 0.013994860462844372, + 0.017832836136221886, + 0.015158692374825478, + 0.024901993572711945, + 0.015654563903808594, + 0.0409347340464592, + 0.027605321258306503, + 0.0391647145152092, + 0.0063719856552779675, + 0.009171546436846256, + 0.011002602986991405, + 0.015903625637292862, + 0.010819156654179096, + 0.01734459027647972, + 0.01188307348638773, + 0.016188235953450203, + 0.004235716536641121, + 0.0036083348095417023, + 0.004737658891826868, + 0.005480166990309954, + 0.00800659041851759, + 0.006331135984510183, + 0.005257768090814352, + 0.005712105426937342, + 0.009111572988331318, + 0.006478624884039164, + 0.012969049625098705, + 0.007610345724970102, + 0.00968710146844387, + 0.07158366590738297, + 0.009676794521510601, + 0.07408232241868973, + 0.0039961314760148525, + 0.0053795697167515755, + 0.00874670222401619, + 0.00904274545609951, + 0.009039375931024551, + 0.009018311277031898, + 0.008492760360240936, + 0.00900935661047697, + 0.003644471988081932, + 0.005483543034642935, + 0.005600246600806713, + 0.005828072316944599, + 0.011675460264086723, + 0.009259194135665894, + 0.008556410670280457, + 0.006631803698837757, + 0.004433616064488888, + 0.004962426144629717, + 0.00843722652643919, + 0.008525104261934757, + 0.01264879759401083, + 0.00979513768106699, + 0.012341765686869621, + 0.009787914343178272, + 0.006225118413567543, + 0.007714178878813982, + 0.007518052589148283, + 0.00915050320327282, + 0.015137838199734688, + 0.010727956891059875, + 0.011018251068890095, + 0.010515732690691948, + 0.002885362133383751, + 0.003192141419276595, + 0.0030562207102775574, + 0.00540880486369133, + 0.0042401389218866825, + 0.005330882966518402, + 0.005283151753246784, + 0.004325742367655039, + 0.014371532015502453, + 0.01380493026226759, + 0.013367979787290096, + 0.010438522323966026, + 0.012717907316982746, + 0.03852321207523346, + 0.012161956168711185, + 0.044046301394701004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.06205803155899048 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a steep rock face visible that is dark gray, damp, and smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rock face covered in vibrant, velvety green moss and a complex web of dark green and earthy brown vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparkling streams of water flowing along the moss and vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, filtered sunlight causing crystal droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012535824440419674 + ], + "opposite_logprob_tokens": [ + -4.38753604888916 + ], + "logprob_reward": -0.012535824440419674, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012535824440419674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1313578188419342, + 0.0630388855934143, + 0.04625310003757477, + 0.2114182859659195, + 0.11009359359741211, + 0.06757394969463348, + 0.05144037678837776, + 0.1429070085287094, + 0.06691951304674149, + 0.10899750888347626 + ], + "credit_map": [ + 0.009895062074065208, + 0.011810410767793655, + 0.011167885735630989, + 0.017474673688411713, + 0.014509661123156548, + 0.02334098145365715, + 0.022025298327207565, + 0.02113383635878563, + 0.003812612034380436, + 0.007284626830369234, + 0.006669227033853531, + 0.008872255682945251, + 0.005475811660289764, + 0.01214270107448101, + 0.007441495079547167, + 0.011340148746967316, + 0.004755811765789986, + 0.003584686666727066, + 0.008819076232612133, + 0.004037213511765003, + 0.006961686536669731, + 0.006625039037317038, + 0.005976495333015919, + 0.00549308629706502, + 0.014192849397659302, + 0.009306029416620731, + 0.02706022746860981, + 0.013116628862917423, + 0.01039559580385685, + 0.06238608807325363, + 0.010659316554665565, + 0.06430154293775558, + 0.010564285330474377, + 0.009366784244775772, + 0.020030803978443146, + 0.014638490974903107, + 0.013485655188560486, + 0.01746751181781292, + 0.012644138187170029, + 0.01189591083675623, + 0.0038567581214010715, + 0.005740938242524862, + 0.007070484571158886, + 0.009477900341153145, + 0.010227197781205177, + 0.01070004515349865, + 0.010532760992646217, + 0.00996785331517458, + 0.0034807906486094, + 0.0036600029561668634, + 0.006360993254929781, + 0.00587544497102499, + 0.007313269190490246, + 0.009128740057349205, + 0.006872900296002626, + 0.008748234249651432, + 0.02488875389099121, + 0.01045121718198061, + 0.03821396827697754, + 0.010443436913192272, + 0.009333758614957333, + 0.017597317695617676, + 0.009640234522521496, + 0.022338304668664932, + 0.006146514322608709, + 0.006561212707310915, + 0.007073771208524704, + 0.007907133549451828, + 0.007343298755586147, + 0.009421654045581818, + 0.011996927671134472, + 0.01046899612993002, + 0.01174340769648552, + 0.009048868902027607, + 0.015415417030453682, + 0.007819718681275845, + 0.011609098874032497, + 0.01952679082751274, + 0.012674919329583645, + 0.021159280091524124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.012535824440419674 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, slowly tilting upward shot begins, gradually revealing a dense jungle vine cascading down a steep, damp rock face. The ancient stone, dark gray, damp, and smooth, is almost entirely covered in a vibrant, velvety green moss that clings to its rough surface. The vine itself is a complex, tangled web of gnarled, woody vines and softer, leafy climbers, their dark green and earthy brown forms intertwined, hanging heavily from an unseen point above. Water perpetually seeps from the rock, flowing in sparkling streams along the moss and vines. As warm, filtered sunlight penetrates the dense jungle canopy, it illuminates the scene, causing countless crystal droplets clinging to every leaf and tendril to shimmer like tiny suspended diamonds. The deliberate upward movement of the camera enhances the scale and majesty of the natural formations, creating a serene and magical atmosphere of pristine, untouched beauty.", + "seed": 31329, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_0/index_6838_seed_31329_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04598796600475907, + "vlm_logprob": -0.04598796600475907, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.04598796600475907, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a steep rock face visible that is dark gray, damp, and smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rock face covered in vibrant, velvety green moss and a complex web of dark green and earthy brown vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparkling streams of water flowing along the moss and vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, filtered sunlight causing crystal droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09856073558330536 + ], + "opposite_logprob_tokens": [ + -2.4735608100891113 + ], + "logprob_reward": -0.09856073558330536, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09856073558330536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02522352524101734, + 0.022700175642967224, + 0.050817910581827164, + 0.015290649607777596, + 0.027410607784986496, + 0.013439038768410683, + 0.04296791926026344, + 0.05929027125239372, + 0.4576222002506256, + 0.28523769974708557 + ], + "credit_map": [ + 0.0023008675780147314, + 0.002637673169374466, + 0.0018363541457802057, + 0.0030967395287007093, + 0.0036132163368165493, + 0.004931224975734949, + 0.0029103411361575127, + 0.0038971113972365856, + 0.0017148904735222459, + 0.0028502114582806826, + 0.003526114160194993, + 0.002545889001339674, + 0.004289853852242231, + 0.002906510839238763, + 0.0028226058930158615, + 0.0020441010128706694, + 0.0021369056776165962, + 0.001712862285785377, + 0.0018017884576693177, + 0.0017653838731348515, + 0.003632067935541272, + 0.020309120416641235, + 0.0024862519931048155, + 0.016973532736301422, + 0.000913147465325892, + 0.0017210827209055424, + 0.0011061885161325336, + 0.002114274073392153, + 0.003002682700753212, + 0.002294191624969244, + 0.0017574538942426443, + 0.00238163024187088, + 0.0009540646569803357, + 0.0011906620347872376, + 0.0016089919954538345, + 0.0015968518564477563, + 0.0031878824811428785, + 0.008940635249018669, + 0.0026124638970941305, + 0.007319057360291481, + 0.0011183099122717977, + 0.0011063509155064821, + 0.0013524321839213371, + 0.0015181238995864987, + 0.0024382462725043297, + 0.0023805396631360054, + 0.002051444724202156, + 0.0014735921286046505, + 0.002946107881143689, + 0.003840337973088026, + 0.0033786462154239416, + 0.007102780044078827, + 0.007473263423889875, + 0.006542835850268602, + 0.0065415240824222565, + 0.0051424261182546616, + 0.007152010686695576, + 0.004028382711112499, + 0.006395820062607527, + 0.0083724120631814, + 0.010789832100272179, + 0.007817242294549942, + 0.009066881611943245, + 0.005667691584676504, + 0.013478074222803116, + 0.004300350788980722, + 0.005261934362351894, + 0.006337041966617107, + 0.011346475221216679, + 0.18776924908161163, + 0.02624444104731083, + 0.20288465917110443, + 0.008398358710110188, + 0.0027019584085792303, + 0.003589523257687688, + 0.0034839704167097807, + 0.008065671660006046, + 0.11414819955825806, + 0.01655343733727932, + 0.1282965987920761 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.09856073558330536 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a steep rock face visible that is dark gray, damp, and smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rock face covered in vibrant, velvety green moss and a complex web of dark green and earthy brown vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparkling streams of water flowing along the moss and vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, filtered sunlight causing crystal droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004237920045853 + ], + "opposite_logprob_tokens": [ + -2.695042371749878 + ], + "logprob_reward": -0.07004237920045853, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004237920045853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05727372318506241, + 0.029867205768823624, + 0.08316269516944885, + 0.015706388279795647, + 0.03897452726960182, + 0.009014767594635487, + 0.02579328417778015, + 0.013786799274384975, + 0.23086299002170563, + 0.495557576417923 + ], + "credit_map": [ + 0.004652669653296471, + 0.005608227103948593, + 0.005892005283385515, + 0.004565652925521135, + 0.006835848558694124, + 0.009747111238539219, + 0.008415153250098228, + 0.011557064019143581, + 0.002956969430670142, + 0.0024933386594057083, + 0.006094090174883604, + 0.002651637652888894, + 0.003732686396688223, + 0.004229761194437742, + 0.003735798178240657, + 0.00397292897105217, + 0.002601283835247159, + 0.0033235098235309124, + 0.002742060460150242, + 0.0022950610145926476, + 0.006670704577118158, + 0.03247269615530968, + 0.006244566757231951, + 0.026812823489308357, + 0.0009547859663143754, + 0.0015945143532007933, + 0.0018212530994787812, + 0.0017174107488244772, + 0.0031457778532058, + 0.002471616491675377, + 0.0020728050731122494, + 0.0019282260909676552, + 0.001368028693832457, + 0.00201614061370492, + 0.0016376698622480035, + 0.00200263480655849, + 0.0034268160816282034, + 0.013855447992682457, + 0.002453255234286189, + 0.0122145377099514, + 0.0007432875572703779, + 0.0007651544292457402, + 0.0006316319922916591, + 0.0010869826655834913, + 0.0018354248022660613, + 0.0014015204505994916, + 0.0015140334144234657, + 0.0010367329232394695, + 0.0015180237824097276, + 0.0020672811660915613, + 0.0020251101814210415, + 0.005053243134170771, + 0.00352464965544641, + 0.0036739297211170197, + 0.004786366131156683, + 0.003144682152196765, + 0.001262793317437172, + 0.0008762615034356713, + 0.0013034805888310075, + 0.0017531231278553605, + 0.0029331818222999573, + 0.002004473004490137, + 0.0022358442656695843, + 0.0014176430413499475, + 0.006714898161590099, + 0.0018947063945233822, + 0.002813442377373576, + 0.0031312915962189436, + 0.005830328445881605, + 0.09555704146623611, + 0.01253290195018053, + 0.10238838195800781, + 0.013666308484971523, + 0.004667277447879314, + 0.005787775851786137, + 0.005600243341177702, + 0.013563612475991249, + 0.19975246489048004, + 0.028697544708848, + 0.22382241487503052 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.07004237920045853 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a steep rock face visible that is dark gray, damp, and smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rock face covered in vibrant, velvety green moss and a complex web of dark green and earthy brown vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparkling streams of water flowing along the moss and vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, filtered sunlight causing crystal droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002203062642365694 + ], + "opposite_logprob_tokens": [ + -6.127202987670898 + ], + "logprob_reward": -0.002203062642365694, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002203062642365694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05560624971985817, + 0.03755717724561691, + 0.18213504552841187, + 0.011355441063642502, + 0.024847690016031265, + 0.023223992437124252, + 0.030535325407981873, + 0.0424845851957798, + 0.2408289909362793, + 0.351425439119339 + ], + "credit_map": [ + 0.004130189772695303, + 0.004785570316016674, + 0.005742227658629417, + 0.004641629755496979, + 0.005250366870313883, + 0.01335963886231184, + 0.007878209464251995, + 0.009818424470722675, + 0.0038155480287969112, + 0.004909766372293234, + 0.0059006852097809315, + 0.0052552977576851845, + 0.004664557054638863, + 0.005663188640028238, + 0.004129296634346247, + 0.003218841040506959, + 0.005220911931246519, + 0.004682125058025122, + 0.0037415113765746355, + 0.003915234003216028, + 0.012603183276951313, + 0.07746902853250504, + 0.009694140404462814, + 0.0648089125752449, + 0.000664835621137172, + 0.001448938506655395, + 0.0012257442576810718, + 0.0012700704392045736, + 0.002075720112770796, + 0.0019846023060381413, + 0.0013970212312415242, + 0.0012885089963674545, + 0.0010043762158602476, + 0.0017163536977022886, + 0.0020994313526898623, + 0.002191354287788272, + 0.005290014669299126, + 0.005120002198964357, + 0.003981039393693209, + 0.0034451191313564777, + 0.0017885534325614572, + 0.0018888103077188134, + 0.0014636070700362325, + 0.0023674091789871454, + 0.0033767546992748976, + 0.004651544615626335, + 0.003553648479282856, + 0.0041336663998663425, + 0.001921399380080402, + 0.0025716146919876337, + 0.002210373291745782, + 0.0032464696560055017, + 0.006560744717717171, + 0.005382165312767029, + 0.0049401866272091866, + 0.0037023737095296383, + 0.004890277050435543, + 0.0028722556307911873, + 0.00456638541072607, + 0.006205413024872541, + 0.0075651295483112335, + 0.005980414338409901, + 0.006201239302754402, + 0.004203470423817635, + 0.0071051400154829025, + 0.002639584941789508, + 0.0028476621955633163, + 0.0038251816295087337, + 0.0062489076517522335, + 0.09813813865184784, + 0.014040312729775906, + 0.10598409175872803, + 0.009735103696584702, + 0.003172649536281824, + 0.0032550464384257793, + 0.003801005193963647, + 0.009371698834002018, + 0.14230740070343018, + 0.020143894478678703, + 0.1596386432647705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.002203062642365694 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a steep rock face visible that is dark gray, damp, and smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rock face covered in vibrant, velvety green moss and a complex web of dark green and earthy brown vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparkling streams of water flowing along the moss and vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, filtered sunlight causing crystal droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306739568710327 + ], + "opposite_logprob_tokens": [ + -3.168067455291748 + ], + "logprob_reward": -0.04306739568710327, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306739568710327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06624617427587509, + 0.032935891300439835, + 0.05214129388332367, + 0.02953595109283924, + 0.054666999727487564, + 0.04512576386332512, + 0.017597798258066177, + 0.018468251451849937, + 0.07866418361663818, + 0.6046176552772522 + ], + "credit_map": [ + 0.006695830728858709, + 0.008897957392036915, + 0.008898016065359116, + 0.007649954874068499, + 0.00639105960726738, + 0.009644580073654652, + 0.008242904208600521, + 0.00982587318867445, + 0.003165309550240636, + 0.004582495428621769, + 0.00408321525901556, + 0.004089754540473223, + 0.004579827189445496, + 0.004993197973817587, + 0.0038452527951449156, + 0.003596839029341936, + 0.002515216823667288, + 0.0045126648619771, + 0.002773356856778264, + 0.0024837504606693983, + 0.005556439980864525, + 0.016817936673760414, + 0.004812194500118494, + 0.012669731862843037, + 0.001661960268393159, + 0.004306301940232515, + 0.0023264281917363405, + 0.004347752779722214, + 0.006214562803506851, + 0.004890972748398781, + 0.0025599864311516285, + 0.0032279861625283957, + 0.0013642910635098815, + 0.0017272236291319132, + 0.0011973140062764287, + 0.002297736704349518, + 0.00764483492821455, + 0.021100236102938652, + 0.0034008468501269817, + 0.01593451388180256, + 0.0037723437417298555, + 0.0031526866368949413, + 0.003253357019275427, + 0.005949762184172869, + 0.00532199302688241, + 0.009761862456798553, + 0.004706589039415121, + 0.009207170456647873, + 0.0009715703781694174, + 0.0014900972601026297, + 0.0018045699689537287, + 0.0037076135631650686, + 0.002324592787772417, + 0.0026349476538598537, + 0.0025049494579434395, + 0.0021594571880996227, + 0.00239533931016922, + 0.0016453906428068876, + 0.002467947080731392, + 0.0025444019120186567, + 0.002297281054779887, + 0.0027445238083601, + 0.002502553863450885, + 0.0018708126153796911, + 0.0034516695886850357, + 0.002385458443313837, + 0.004022340755909681, + 0.003119973698630929, + 0.003358064452186227, + 0.026932967826724052, + 0.006014907732605934, + 0.029378794133663177, + 0.017076877877116203, + 0.005246222950518131, + 0.005600429605692625, + 0.006774440873414278, + 0.016319574788212776, + 0.24346883594989777, + 0.03512758016586304, + 0.2750037610530853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.04306739568710327 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a steep rock face visible that is dark gray, damp, and smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rock face covered in vibrant, velvety green moss and a complex web of dark green and earthy brown vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparkling streams of water flowing along the moss and vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, filtered sunlight causing crystal droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016066256910562515 + ], + "opposite_logprob_tokens": [ + -4.141066074371338 + ], + "logprob_reward": -0.016066256910562515, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016066256910562515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06115125119686127, + 0.07228008657693863, + 0.04376637563109398, + 0.020094340667128563, + 0.07620356976985931, + 0.08146873861551285, + 0.02080315165221691, + 0.02796310931444168, + 0.13870620727539062, + 0.45756325125694275 + ], + "credit_map": [ + 0.004208381753414869, + 0.006386549212038517, + 0.005141718778759241, + 0.0057834177277982235, + 0.0062299566343426704, + 0.01460016705095768, + 0.007959776557981968, + 0.010841277427971363, + 0.007141922600567341, + 0.007104430813342333, + 0.011853116564452648, + 0.006225304212421179, + 0.010413316078484058, + 0.013647963292896748, + 0.00767888268455863, + 0.008215135894715786, + 0.002104594372212887, + 0.0031956227030605078, + 0.002284984104335308, + 0.0025832431856542826, + 0.005414468236267567, + 0.013389297761023045, + 0.004648498259484768, + 0.010145656764507294, + 0.0011398030910640955, + 0.002443842124193907, + 0.0018036389956250787, + 0.002798843663185835, + 0.0037535480223596096, + 0.002934205112978816, + 0.002331756753847003, + 0.002888698363676667, + 0.002198091708123684, + 0.0026185824535787106, + 0.001998418476432562, + 0.003693921258673072, + 0.009251166135072708, + 0.030207883566617966, + 0.003609243081882596, + 0.022626254707574844, + 0.006612271070480347, + 0.004919441416859627, + 0.006125302519649267, + 0.010635648854076862, + 0.010404661297798157, + 0.01791827194392681, + 0.008513735607266426, + 0.016339393332600594, + 0.0016523251542821527, + 0.0019835953135043383, + 0.0017975722439587116, + 0.0031993447337299585, + 0.003180928761139512, + 0.002975679934024811, + 0.0035861495416611433, + 0.0024275528267025948, + 0.004750346299260855, + 0.001972345868125558, + 0.003976528998464346, + 0.0025089106056839228, + 0.0037348512560129166, + 0.0037617890629917383, + 0.003831481095403433, + 0.0034268510062247515, + 0.004729663487523794, + 0.0044310432858765125, + 0.004726751707494259, + 0.006799317896366119, + 0.005839355289936066, + 0.048007532954216, + 0.010892623104155064, + 0.05327989533543587, + 0.01342145074158907, + 0.004725949838757515, + 0.005823906045407057, + 0.005738356150686741, + 0.012539046816527843, + 0.18319934606552124, + 0.026116108521819115, + 0.20599903166294098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.016066256910562515 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, slowly tilting upward shot begins, gradually revealing a dense jungle vine cascading down a steep, damp rock face. The ancient stone, dark gray, damp, and smooth, is almost entirely covered in a vibrant, velvety green moss that clings to its rough surface. The vine itself is a complex, tangled web of gnarled, woody vines and softer, leafy climbers, their dark green and earthy brown forms intertwined, hanging heavily from an unseen point above. Water perpetually seeps from the rock, flowing in sparkling streams along the moss and vines. As warm, filtered sunlight penetrates the dense jungle canopy, it illuminates the scene, causing countless crystal droplets clinging to every leaf and tendril to shimmer like tiny suspended diamonds. The deliberate upward movement of the camera enhances the scale and majesty of the natural formations, creating a serene and magical atmosphere of pristine, untouched beauty.", + "seed": 31330, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_0/index_6838_seed_31330_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02428623295854777, + "vlm_logprob": -0.02428623295854777, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.02428623295854777, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a steep rock face visible that is dark gray, damp, and smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rock face covered in vibrant, velvety green moss and a complex web of dark green and earthy brown vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparkling streams of water flowing along the moss and vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, filtered sunlight causing crystal droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013592664152383804 + ], + "opposite_logprob_tokens": [ + -5.138592720031738 + ], + "logprob_reward": -0.013592664152383804, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.013592664152383804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2707459032535553, + 0.04011145606637001, + 0.04577656835317612, + 0.04352259635925293, + 0.07926144450902939, + 0.09976425021886826, + 0.19327758252620697, + 0.04246486350893974, + 0.03105873055756092, + 0.1540166437625885 + ], + "credit_map": [ + 0.026674969121813774, + 0.04407865181565285, + 0.04681301489472389, + 0.05249769240617752, + 0.023612022399902344, + 0.020163564011454582, + 0.033100951462984085, + 0.023805035278201103, + 0.0035095689818263054, + 0.004911291878670454, + 0.007490050978958607, + 0.006361042615026236, + 0.002741356613114476, + 0.004202603362500668, + 0.007802311331033707, + 0.0030932296067476273, + 0.003548468230292201, + 0.006155263166874647, + 0.006894764024764299, + 0.00947981234639883, + 0.00383398518897593, + 0.0021455816458910704, + 0.009605428203940392, + 0.004113263450562954, + 0.003403563518077135, + 0.006620277184993029, + 0.005433766636997461, + 0.006281506270170212, + 0.006971477065235376, + 0.004400420468300581, + 0.005780678242444992, + 0.0046309069730341434, + 0.005993915256112814, + 0.013711743988096714, + 0.00837409682571888, + 0.012595459818840027, + 0.0054573602974414825, + 0.013611567206680775, + 0.014734428375959396, + 0.0047828727401793, + 0.008714186027646065, + 0.011069603264331818, + 0.010871957056224346, + 0.01230719592422247, + 0.010298737324774265, + 0.021171828731894493, + 0.010678079910576344, + 0.014652660116553307, + 0.010423634201288223, + 0.020821236073970795, + 0.012254766188561916, + 0.047500140964984894, + 0.007612986955791712, + 0.061515022069215775, + 0.01725209318101406, + 0.015897713601589203, + 0.0030542442109435797, + 0.00620099063962698, + 0.0052358536049723625, + 0.005622170865535736, + 0.004016481805592775, + 0.008018210530281067, + 0.00445889588445425, + 0.0058580138720571995, + 0.003105457639321685, + 0.0037988396361470222, + 0.0037625597324222326, + 0.0040526678785681725, + 0.0030831231269985437, + 0.00540150934830308, + 0.0028598341159522533, + 0.004994737915694714, + 0.029380157589912415, + 0.026922715827822685, + 0.015198757871985435, + 0.018357649445533752, + 0.013185999356210232, + 0.019111623987555504, + 0.010538605973124504, + 0.0213211327791214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.013592664152383804 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a steep rock face visible that is dark gray, damp, and smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rock face covered in vibrant, velvety green moss and a complex web of dark green and earthy brown vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparkling streams of water flowing along the moss and vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, filtered sunlight causing crystal droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062065985053777695 + ], + "opposite_logprob_tokens": [ + -2.812066078186035 + ], + "logprob_reward": -0.062065985053777695, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062065985053777695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11455509811639786, + 0.063568115234375, + 0.04442829266190529, + 0.03323453292250633, + 0.07470829784870148, + 0.07716436684131622, + 0.21306347846984863, + 0.1752287596464157, + 0.04750373959541321, + 0.15654532611370087 + ], + "credit_map": [ + 0.008167307823896408, + 0.01191610936075449, + 0.023865312337875366, + 0.024919627234339714, + 0.010230368003249168, + 0.008760714903473854, + 0.015847915783524513, + 0.010847743600606918, + 0.0056509277783334255, + 0.010671025142073631, + 0.007616459857672453, + 0.011379921808838844, + 0.007477188482880592, + 0.005454342346638441, + 0.007717439439147711, + 0.007600812707096338, + 0.002823353512212634, + 0.004562615882605314, + 0.005530805326998234, + 0.007579373195767403, + 0.0037157125771045685, + 0.002779228612780571, + 0.012605239637196064, + 0.004831962287425995, + 0.002641084371134639, + 0.004440763033926487, + 0.004507200792431831, + 0.007243345025926828, + 0.0033850555773824453, + 0.002635606564581394, + 0.004422435071319342, + 0.003959041554480791, + 0.006096872501075268, + 0.013164506293833256, + 0.00792393647134304, + 0.01997974142432213, + 0.005235814489424229, + 0.005512914154678583, + 0.013084443286061287, + 0.0037100643385201693, + 0.006487169768661261, + 0.005685510113835335, + 0.006071703042834997, + 0.01124495267868042, + 0.008104072883725166, + 0.019445985555648804, + 0.009547264315187931, + 0.010577713139355183, + 0.007246505003422499, + 0.006959977559745312, + 0.010513931512832642, + 0.08284097909927368, + 0.005541763734072447, + 0.07537031918764114, + 0.011819876730442047, + 0.012770119123160839, + 0.009429261088371277, + 0.01114241685718298, + 0.014027765020728111, + 0.0388214997947216, + 0.016105463728308678, + 0.04268941283226013, + 0.020031534135341644, + 0.022981401532888412, + 0.0056700678542256355, + 0.0030731044244021177, + 0.003060298739001155, + 0.002571991877630353, + 0.0034290130715817213, + 0.013445922173559666, + 0.002557862550020218, + 0.013695484027266502, + 0.014485703781247139, + 0.019719215109944344, + 0.011821714229881763, + 0.01814287342131138, + 0.01618262566626072, + 0.03749227151274681, + 0.016125425696372986, + 0.022575486451387405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.062065985053777695 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a steep rock face visible that is dark gray, damp, and smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rock face covered in vibrant, velvety green moss and a complex web of dark green and earthy brown vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparkling streams of water flowing along the moss and vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, filtered sunlight causing crystal droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001719426247291267 + ], + "opposite_logprob_tokens": [ + -6.3767194747924805 + ], + "logprob_reward": -0.001719426247291267, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001719426247291267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12044122070074081, + 0.05922525003552437, + 0.05093899741768837, + 0.03246476873755455, + 0.3622090518474579, + 0.10971339046955109, + 0.15025359392166138, + 0.028042711317539215, + 0.039498910307884216, + 0.04721206799149513 + ], + "credit_map": [ + 0.010626628994941711, + 0.01630055531859398, + 0.016224756836891174, + 0.02926316298544407, + 0.011539686471223831, + 0.009589294902980328, + 0.01576317846775055, + 0.011133967898786068, + 0.005458583123981953, + 0.012118262238800526, + 0.007503561209887266, + 0.012680474668741226, + 0.006342154461890459, + 0.00468301260843873, + 0.006656531244516373, + 0.003782674204558134, + 0.0031739582773298025, + 0.006911208387464285, + 0.006048732902854681, + 0.007592125330120325, + 0.0037934475112706423, + 0.003932336810976267, + 0.010993137024343014, + 0.008494056761264801, + 0.0022799966391175985, + 0.005294033791869879, + 0.0035754279233515263, + 0.008365158922970295, + 0.002672207076102495, + 0.0019535727333277464, + 0.004944642540067434, + 0.003379727713763714, + 0.017158230766654015, + 0.009966595098376274, + 0.013637175783514977, + 0.131894052028656, + 0.013564051128923893, + 0.02933332696557045, + 0.13112197816371918, + 0.01553365122526884, + 0.008920932188630104, + 0.006370088551193476, + 0.008412945084273815, + 0.015566805377602577, + 0.013765810988843441, + 0.025419719517230988, + 0.01334469672292471, + 0.01791239343583584, + 0.008841732516884804, + 0.008353701792657375, + 0.011947168968617916, + 0.030064720660448074, + 0.010893573984503746, + 0.043694689869880676, + 0.01990245282649994, + 0.016555551439523697, + 0.002029112307354808, + 0.0017953228671103716, + 0.0033514571841806173, + 0.0052334824576973915, + 0.003334895009174943, + 0.0056705535389482975, + 0.003940054215490818, + 0.0026878321077674627, + 0.002243523020297289, + 0.0025450594257563353, + 0.0021392942871898413, + 0.002118548611178994, + 0.0015829633921384811, + 0.013848481699824333, + 0.0021792356856167316, + 0.012841805815696716, + 0.00580907566472888, + 0.004507188219577074, + 0.0051478976383805275, + 0.006359927821904421, + 0.0046922327019274235, + 0.008161114528775215, + 0.004629519302397966, + 0.007905114442110062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.001719426247291267 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a steep rock face visible that is dark gray, damp, and smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rock face covered in vibrant, velvety green moss and a complex web of dark green and earthy brown vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparkling streams of water flowing along the moss and vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, filtered sunlight causing crystal droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810510411858559 + ], + "opposite_logprob_tokens": [ + -3.288105010986328 + ], + "logprob_reward": -0.03810510411858559, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810510411858559, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09694686532020569, + 0.09265443682670593, + 0.08979444205760956, + 0.021151790395379066, + 0.2863686978816986, + 0.21835385262966156, + 0.06632569432258606, + 0.06355228275060654, + 0.01981222629547119, + 0.045039739459753036 + ], + "credit_map": [ + 0.006861167494207621, + 0.013517914339900017, + 0.02651779167354107, + 0.016247982159256935, + 0.008361118845641613, + 0.008450272493064404, + 0.009636864997446537, + 0.007353750988841057, + 0.007645415607839823, + 0.019845489412546158, + 0.013274818658828735, + 0.020952552556991577, + 0.008365195244550705, + 0.00804977398365736, + 0.008516820147633553, + 0.0060043660923838615, + 0.005351894069463015, + 0.012162024155259132, + 0.011324286460876465, + 0.01431016344577074, + 0.006685299798846245, + 0.007901862263679504, + 0.016538092866539955, + 0.015520819462835789, + 0.0019193035550415516, + 0.002645408269017935, + 0.0030685043893754482, + 0.004863747861236334, + 0.0018572547705844045, + 0.0014770252164453268, + 0.003200455801561475, + 0.002120089717209339, + 0.01248130202293396, + 0.010542104952037334, + 0.008052016608417034, + 0.10335401445627213, + 0.011149158701300621, + 0.023569738492369652, + 0.10457576811313629, + 0.012644580565392971, + 0.015061664395034313, + 0.011681446805596352, + 0.013422525487840176, + 0.039760708808898926, + 0.028243349865078926, + 0.050263021141290665, + 0.02752743661403656, + 0.032393679022789, + 0.0026253454852849245, + 0.004513478372246027, + 0.003858329262584448, + 0.024234751239418983, + 0.0023835510946810246, + 0.02176896296441555, + 0.0031171587761491537, + 0.0038241154979914427, + 0.00343247945420444, + 0.004151825327426195, + 0.0042243110947310925, + 0.020359260961413383, + 0.0032994814682751894, + 0.017097700387239456, + 0.004669892601668835, + 0.006317325867712498, + 0.0018905175384134054, + 0.0011570792412385345, + 0.0011738934554159641, + 0.0018345051212236285, + 0.0013767455238848925, + 0.005549055989831686, + 0.0013764400500804186, + 0.005453989375382662, + 0.007647502701729536, + 0.005829536356031895, + 0.004665305372327566, + 0.0053736004047095776, + 0.004689246416091919, + 0.005655304528772831, + 0.0039295717142522335, + 0.0072496687062084675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.03810510411858559 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a steep rock face visible that is dark gray, damp, and smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rock face covered in vibrant, velvety green moss and a complex web of dark green and earthy brown vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparkling streams of water flowing along the moss and vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, filtered sunlight causing crystal droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947985220700502 + ], + "opposite_logprob_tokens": [ + -5.130948066711426 + ], + "logprob_reward": -0.005947985220700502, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947985220700502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20552410185337067, + 0.07716107368469238, + 0.06281130760908127, + 0.019101178273558617, + 0.169168621301651, + 0.09023143351078033, + 0.06584509462118149, + 0.12054096907377243, + 0.03670359030365944, + 0.15291260182857513 + ], + "credit_map": [ + 0.015265016816556454, + 0.03869657590985298, + 0.022220700979232788, + 0.044277798384428024, + 0.02686552330851555, + 0.01571391336619854, + 0.024629831314086914, + 0.017854729667305946, + 0.007612080778926611, + 0.011850266717374325, + 0.011499714106321335, + 0.015842372551560402, + 0.006190082523971796, + 0.006573589518666267, + 0.010846422053873539, + 0.00674653984606266, + 0.004869250115007162, + 0.008179282769560814, + 0.006407343316823244, + 0.012196404859423637, + 0.004530945792794228, + 0.00529964966699481, + 0.009285220876336098, + 0.01204320415854454, + 0.0013763346942141652, + 0.0032102991826832294, + 0.0022105493117123842, + 0.004134013783186674, + 0.0019683698192238808, + 0.0012325843563303351, + 0.0030623176135122776, + 0.0019067093962803483, + 0.005983761977404356, + 0.005659414455294609, + 0.006252624560147524, + 0.06275025010108948, + 0.008545354939997196, + 0.013336007483303547, + 0.058652617037296295, + 0.00798859540373087, + 0.008941723965108395, + 0.00653748819604516, + 0.008905050344765186, + 0.018066370859742165, + 0.009028950706124306, + 0.01846226491034031, + 0.008912557736039162, + 0.011377028189599514, + 0.0060479543171823025, + 0.01469181664288044, + 0.004355518613010645, + 0.019525520503520966, + 0.0029321424663066864, + 0.009749026969075203, + 0.004411775153130293, + 0.004131346475332975, + 0.007839163765311241, + 0.00844285823404789, + 0.009522674605250359, + 0.03917565941810608, + 0.0055129691027104855, + 0.030978403985500336, + 0.010387690737843513, + 0.008681554347276688, + 0.0034401873126626015, + 0.00403836602345109, + 0.003396102925762534, + 0.004001670051366091, + 0.002746080979704857, + 0.009289281442761421, + 0.002486696233972907, + 0.007305205333977938, + 0.015638144686818123, + 0.029453841969370842, + 0.01651402749121189, + 0.02072899043560028, + 0.014951233752071857, + 0.02241932787001133, + 0.014782742597162724, + 0.018424291163682938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.005947985220700502 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, slowly tilting upward shot begins, gradually revealing a dense jungle vine cascading down a steep, damp rock face. The ancient stone, dark gray, damp, and smooth, is almost entirely covered in a vibrant, velvety green moss that clings to its rough surface. The vine itself is a complex, tangled web of gnarled, woody vines and softer, leafy climbers, their dark green and earthy brown forms intertwined, hanging heavily from an unseen point above. Water perpetually seeps from the rock, flowing in sparkling streams along the moss and vines. As warm, filtered sunlight penetrates the dense jungle canopy, it illuminates the scene, causing countless crystal droplets clinging to every leaf and tendril to shimmer like tiny suspended diamonds. The deliberate upward movement of the camera enhances the scale and majesty of the natural formations, creating a serene and magical atmosphere of pristine, untouched beauty.", + "seed": 31331, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_0/index_6838_seed_31331_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04836209521163255, + "vlm_logprob": -0.04836209521163255, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.04836209521163255, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a steep rock face visible that is dark gray, damp, and smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rock face covered in vibrant, velvety green moss and a complex web of dark green and earthy brown vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparkling streams of water flowing along the moss and vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, filtered sunlight causing crystal droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06963450461626053 + ], + "opposite_logprob_tokens": [ + -2.819634437561035 + ], + "logprob_reward": -0.06963450461626053, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06963450461626053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06210353225469589, + 0.03530498966574669, + 0.12261306494474411, + 0.12569016218185425, + 0.04030342027544975, + 0.03137435391545296, + 0.0254175066947937, + 0.0235762782394886, + 0.023799562826752663, + 0.5098171830177307 + ], + "credit_map": [ + 0.0054307421669363976, + 0.005543661303818226, + 0.004891707096248865, + 0.008892486803233624, + 0.013356884010136127, + 0.005945079959928989, + 0.009487585164606571, + 0.00855538621544838, + 0.0019618216902017593, + 0.0030787985306233168, + 0.0021118582226336002, + 0.008459405042231083, + 0.005107322707772255, + 0.004668242298066616, + 0.006669288035482168, + 0.003248253371566534, + 0.013881763443350792, + 0.005831214599311352, + 0.030110623687505722, + 0.014133610762655735, + 0.01595526747405529, + 0.01058910135179758, + 0.0265927966684103, + 0.00551869161427021, + 0.004721821751445532, + 0.012587998062372208, + 0.010257093235850334, + 0.04496118426322937, + 0.009613103233277798, + 0.00628964276984334, + 0.0319027379155159, + 0.005356583744287491, + 0.002014108467847109, + 0.0026615168899297714, + 0.002675463678315282, + 0.002364098560065031, + 0.013472072780132294, + 0.002844812348484993, + 0.012133777141571045, + 0.00213757180608809, + 0.0030798744410276413, + 0.004564580041915178, + 0.0031950618140399456, + 0.005014301743358374, + 0.004651268478482962, + 0.004260579124093056, + 0.0036711543798446655, + 0.002937535522505641, + 0.0018683840753510594, + 0.0037660717498511076, + 0.002784937620162964, + 0.0026795733720064163, + 0.004600249696522951, + 0.0031450404785573483, + 0.004415579605847597, + 0.0021576699800789356, + 0.0018418970284983516, + 0.00406737020239234, + 0.001808211556635797, + 0.0026792027056217194, + 0.0053226337768137455, + 0.002729569561779499, + 0.0025190969463437796, + 0.0026082955300807953, + 0.002140130614861846, + 0.0032783239148557186, + 0.002803538227453828, + 0.0026857664342969656, + 0.003119952278211713, + 0.003395679872483015, + 0.0035530433524399996, + 0.002823128132149577, + 0.02147122286260128, + 0.17143894731998444, + 0.016914740204811096, + 0.18529096245765686, + 0.03412986919283867, + 0.0176081545650959, + 0.03867608681321144, + 0.024287167936563492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 2 + ], + "selected_reward": -0.06963450461626053 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a steep rock face visible that is dark gray, damp, and smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rock face covered in vibrant, velvety green moss and a complex web of dark green and earthy brown vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparkling streams of water flowing along the moss and vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, filtered sunlight causing crystal droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043079957365989685 + ], + "opposite_logprob_tokens": [ + -3.1680798530578613 + ], + "logprob_reward": -0.043079957365989685, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043079957365989685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11184217780828476, + 0.4272097945213318, + 0.07529635727405548, + 0.0574946366250515, + 0.07206987589597702, + 0.045152951031923294, + 0.04093886911869049, + 0.09109571576118469, + 0.03282205015420914, + 0.04607757180929184 + ], + "credit_map": [ + 0.008844442665576935, + 0.010520604439079762, + 0.008899945765733719, + 0.009947038255631924, + 0.02504931204020977, + 0.011288868263363838, + 0.023894084617495537, + 0.013397889211773872, + 0.018467498943209648, + 0.035016730427742004, + 0.021504541859030724, + 0.15196014940738678, + 0.06596902757883072, + 0.028185218572616577, + 0.08178634196519852, + 0.024320276454091072, + 0.00734752370044589, + 0.00477832043543458, + 0.00396624393761158, + 0.005043079145252705, + 0.015553302131593227, + 0.01023317314684391, + 0.022704103961586952, + 0.005670612677931786, + 0.004629950504750013, + 0.003835130948573351, + 0.014846724458038807, + 0.012453089468181133, + 0.006004393566399813, + 0.0040545957162976265, + 0.007986531592905521, + 0.0036842196714133024, + 0.0025918972678482533, + 0.0043329400941729546, + 0.0030041029676795006, + 0.0033403849229216576, + 0.02674465999007225, + 0.0050268955528736115, + 0.021871957927942276, + 0.005157035309821367, + 0.003807686036452651, + 0.00491911917924881, + 0.003411119105294347, + 0.00538048567250371, + 0.00600022915750742, + 0.009247696958482265, + 0.006772998254746199, + 0.0056136175990104675, + 0.002399017568677664, + 0.006192907691001892, + 0.0032142556738108397, + 0.004494148306548595, + 0.008260917849838734, + 0.0066743153147399426, + 0.006093251518905163, + 0.0036100547295063734, + 0.005259828642010689, + 0.014407460577785969, + 0.00647022295743227, + 0.009190008975565434, + 0.026699641719460487, + 0.009877104312181473, + 0.010712703689932823, + 0.008478748612105846, + 0.0022194040939211845, + 0.003214049618691206, + 0.003715211758390069, + 0.003834361443296075, + 0.006363393273204565, + 0.004416631534695625, + 0.004113813396543264, + 0.004945185501128435, + 0.0029037431813776493, + 0.004182691685855389, + 0.0025593354366719723, + 0.013460330665111542, + 0.0042058625258505344, + 0.004512912128120661, + 0.006819053553044796, + 0.0074336412362754345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.043079957365989685 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a steep rock face visible that is dark gray, damp, and smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rock face covered in vibrant, velvety green moss and a complex web of dark green and earthy brown vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparkling streams of water flowing along the moss and vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, filtered sunlight causing crystal droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019455092260614038 + ], + "opposite_logprob_tokens": [ + -6.251945495605469 + ], + "logprob_reward": -0.0019455092260614038, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019455092260614038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05467220023274422, + 0.137452632188797, + 0.07509761303663254, + 0.3046489655971527, + 0.12002389132976532, + 0.018020251765847206, + 0.0816379114985466, + 0.04034852609038353, + 0.02456357516348362, + 0.14353437721729279 + ], + "credit_map": [ + 0.004700821824371815, + 0.004824806936085224, + 0.005774094723165035, + 0.006311322562396526, + 0.010346543043851852, + 0.004629133734852076, + 0.012696607038378716, + 0.005388876888900995, + 0.006335321813821793, + 0.010855640284717083, + 0.008827058598399162, + 0.04784293845295906, + 0.01658802106976509, + 0.00870262086391449, + 0.030063653364777565, + 0.008237391710281372, + 0.008377531543374062, + 0.0065558538772165775, + 0.01813580095767975, + 0.012682780623435974, + 0.00861439760774374, + 0.006128711625933647, + 0.008871415629982948, + 0.005731124430894852, + 0.017374539747834206, + 0.016676275059580803, + 0.07595168054103851, + 0.11348133534193039, + 0.03234072029590607, + 0.014288983307778835, + 0.02415246143937111, + 0.010382984764873981, + 0.004745416343212128, + 0.024577446281909943, + 0.0035821530036628246, + 0.003814367577433586, + 0.04847988858819008, + 0.009177405387163162, + 0.020143533125519753, + 0.005503689404577017, + 0.001626263139769435, + 0.0025787639897316694, + 0.0015958501026034355, + 0.0020295679569244385, + 0.0034421642776578665, + 0.0019795489497482777, + 0.00254509667865932, + 0.0022229987662285566, + 0.005137250758707523, + 0.014131159521639347, + 0.005515731871128082, + 0.007147280499339104, + 0.019271232187747955, + 0.0075503382831811905, + 0.015566404908895493, + 0.007318520452827215, + 0.0022116792388260365, + 0.00593180488795042, + 0.003731913398951292, + 0.004236709326505661, + 0.008377387188374996, + 0.004481976386159658, + 0.00572906294837594, + 0.00564799876883626, + 0.0018813565839082003, + 0.002377485390752554, + 0.0018229406559839845, + 0.0036965536419302225, + 0.003407877404242754, + 0.003915498498827219, + 0.0037676156498491764, + 0.003694248851388693, + 0.005547213368117809, + 0.043049655854701996, + 0.004998151212930679, + 0.04272056370973587, + 0.009836513549089432, + 0.0059722959995269775, + 0.019495682790875435, + 0.01191431563347578 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 1 + ], + "selected_reward": -0.0019455092260614038 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a steep rock face visible that is dark gray, damp, and smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rock face covered in vibrant, velvety green moss and a complex web of dark green and earthy brown vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparkling streams of water flowing along the moss and vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, filtered sunlight causing crystal droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295673251152039 + ], + "opposite_logprob_tokens": [ + -2.2379567623138428 + ], + "logprob_reward": -0.11295673251152039, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295673251152039, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03986908495426178, + 0.052682336419820786, + 0.13658766448497772, + 0.35867688059806824, + 0.08190210163593292, + 0.033590637147426605, + 0.1127852201461792, + 0.0450461246073246, + 0.030942482873797417, + 0.10791753232479095 + ], + "credit_map": [ + 0.0033572593238204718, + 0.002619820646941662, + 0.0035190433263778687, + 0.0034841597080230713, + 0.007237015292048454, + 0.004347727168351412, + 0.010872541926801205, + 0.004431520588696003, + 0.002791321137920022, + 0.0063164313323795795, + 0.005276104435324669, + 0.015723690390586853, + 0.007021354045718908, + 0.004327546339482069, + 0.008210311643779278, + 0.0030155745334923267, + 0.01613018848001957, + 0.0069456747733056545, + 0.03510122373700142, + 0.01621418073773384, + 0.016818590462207794, + 0.010995539836585522, + 0.02813691459596157, + 0.006245346274226904, + 0.020110219717025757, + 0.020586658269166946, + 0.08832564949989319, + 0.13326288759708405, + 0.03856559470295906, + 0.017059577628970146, + 0.028325241059064865, + 0.012441014871001244, + 0.0032487947028130293, + 0.015403165481984615, + 0.003091517835855484, + 0.0036906497552990913, + 0.032220400869846344, + 0.0060149189084768295, + 0.014080176129937172, + 0.004152478650212288, + 0.0032823204528540373, + 0.003965705167502165, + 0.003907239064574242, + 0.003592710942029953, + 0.005621104035526514, + 0.003607522463425994, + 0.00550787802785635, + 0.004106156527996063, + 0.006547837983816862, + 0.019289448857307434, + 0.0053271823562681675, + 0.011304969899356365, + 0.028508244082331657, + 0.010117175057530403, + 0.022612325847148895, + 0.00907803699374199, + 0.003490152768790722, + 0.008390975184738636, + 0.0040924386121332645, + 0.004133365582674742, + 0.010760994628071785, + 0.004389641340821981, + 0.004897525068372488, + 0.004891031887382269, + 0.002813148545101285, + 0.0037669837474823, + 0.0034684399142861366, + 0.003989500924944878, + 0.004383305553346872, + 0.0046040876768529415, + 0.0041982876136898994, + 0.003718728432431817, + 0.003879702650010586, + 0.03277864307165146, + 0.0033115753903985023, + 0.041012655943632126, + 0.006459466181695461, + 0.0037743274588137865, + 0.00842767208814621, + 0.008273492567241192 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 6 + ], + "selected_reward": -0.11295673251152039 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a steep rock face visible that is dark gray, damp, and smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rock face covered in vibrant, velvety green moss and a complex web of dark green and earthy brown vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparkling streams of water flowing along the moss and vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, filtered sunlight causing crystal droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193772338330746 + ], + "opposite_logprob_tokens": [ + -4.264193534851074 + ], + "logprob_reward": -0.014193772338330746, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193772338330746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08427052199840546, + 0.11702404916286469, + 0.10739200562238693, + 0.28732895851135254, + 0.09436284750699997, + 0.030678387731313705, + 0.07237671315670013, + 0.04104779660701752, + 0.052520494908094406, + 0.11299819499254227 + ], + "credit_map": [ + 0.007401943206787109, + 0.005500494036823511, + 0.008074871264398098, + 0.006682715844362974, + 0.016613703221082687, + 0.006881908047944307, + 0.025810550898313522, + 0.007304334081709385, + 0.0040795691311359406, + 0.007404946722090244, + 0.004718604031950235, + 0.017310604453086853, + 0.030227674171328545, + 0.005614417605102062, + 0.04279729351401329, + 0.0048709437251091, + 0.012194657698273659, + 0.005297766998410225, + 0.026257721707224846, + 0.012599845416843891, + 0.014040638692677021, + 0.009217672049999237, + 0.02284974977374077, + 0.004933943506330252, + 0.015158052556216717, + 0.018823249265551567, + 0.06281188875436783, + 0.11152303963899612, + 0.029542431235313416, + 0.013143764808773994, + 0.02546040713787079, + 0.010866119526326656, + 0.0033535475376993418, + 0.01480712927877903, + 0.002582476008683443, + 0.0032532652840018272, + 0.03877871483564377, + 0.006462721154093742, + 0.021041380241513252, + 0.004083616193383932, + 0.0026500653475522995, + 0.0037313406355679035, + 0.00340484781190753, + 0.0037879846058785915, + 0.0058877659030258656, + 0.0035430281423032284, + 0.004595434293150902, + 0.003077923785895109, + 0.004630677402019501, + 0.014901980757713318, + 0.005712148733437061, + 0.006488252431154251, + 0.012821822427213192, + 0.007945728488266468, + 0.013070197775959969, + 0.006805902812629938, + 0.004070370923727751, + 0.006618514657020569, + 0.0041648331098258495, + 0.003676321590319276, + 0.009724951349198818, + 0.004478353075683117, + 0.0044490378350019455, + 0.003865412902086973, + 0.004643728956580162, + 0.007215527351945639, + 0.007063413504511118, + 0.009810726158320904, + 0.00614899443462491, + 0.006543987430632114, + 0.0067779128439724445, + 0.004316210746765137, + 0.003632199252024293, + 0.031089626252651215, + 0.004310308024287224, + 0.03653084114193916, + 0.009087844751775265, + 0.004114257171750069, + 0.017360921949148178, + 0.006872194353491068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 9 + ], + "selected_reward": -0.014193772338330746 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_1/index_tensor([9702, 9702])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_1/index_tensor([9702, 9702])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..cef2a1b6253e7a7ebf0852ff2e5a41305fc4feae --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_1/index_tensor([9702, 9702])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a scene stylized like a Ukiyo-e woodblock print, the camera steadily ascends, beginning with a low-angle shot of the stone base of a tall, ornate pagoda. The structure features intense black outlines and flat planes of deep vermilion and dark-varnished wood, set against a backdrop of gnarled pine trees silhouetted against the night sky. As the camera ascends, revealing the tower's multi-tiered structure and elegantly curved roof, a powerful, jagged flash of brilliant white lightning suddenly streaks across the sky, connecting violently and cracklingly with the tower's tallest spire. For a brief moment, the intense flash whitens colors and casts stark, simplified shadows, mimicking the high-contrast effect of a print. The camera continues its steady upward journey, passing through the spire to reveal the full celestial spectacle above: the deep indigo night sky is filled with vibrant, flowing, dancing ribbons of auroras, its undulating draperies of emerald green, ethereal violet, and soft magenta slowly and gracefully swirling. The entire composition maintains a flat perspective and a subtle paper texture, creating a dramatic and stunning atmosphere that blends the raw power of nature with the serene, decorative beauty of traditional Japanese art.", + "seed": 5267, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_1/index_9702_seed_5267_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05395264569669962, + "vlm_logprob": -0.05395264569669962, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.05395264569669962, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene stylized like a Ukiyo-e woodblock print with flat perspective and paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily ascend from a low-angle shot of a stone pagoda base to reveal the full tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a powerful, jagged flash of brilliant white lightning connecting to the tower's tallest spire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the night sky display vibrant, flowing ribbons of auroras in emerald green, ethereal violet, and soft magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pagoda structure depicted with intense black outlines and flat planes of deep vermilion and dark-varnished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2033257633447647 + ], + "opposite_logprob_tokens": [ + -1.7033257484436035 + ], + "logprob_reward": -0.2033257633447647, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2033257633447647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08383647352457047, + 0.03258473798632622, + 0.046012409031391144, + 0.07150392234325409, + 0.03728295490145683, + 0.435510516166687, + 0.0336843878030777, + 0.17464031279087067, + 0.028457006439566612, + 0.05648728832602501 + ], + "credit_map": [ + 0.01229114830493927, + 0.010263263247907162, + 0.00977072399109602, + 0.00561402877792716, + 0.011681945063173771, + 0.010353557765483856, + 0.010856918059289455, + 0.013004885986447334, + 0.0038878389168530703, + 0.003835385898128152, + 0.0033846022561192513, + 0.003019755706191063, + 0.005889078602194786, + 0.00522777670994401, + 0.0028837372083216906, + 0.004456561524420977, + 0.007413679268211126, + 0.005360831506550312, + 0.005478071980178356, + 0.004186372272670269, + 0.006480303127318621, + 0.006535915657877922, + 0.005276881158351898, + 0.005280354060232639, + 0.007217880338430405, + 0.009419338777661324, + 0.009776224382221699, + 0.012649967335164547, + 0.010541950352489948, + 0.006267742719501257, + 0.00824101734906435, + 0.0073897959664464, + 0.003448435105383396, + 0.005604912992566824, + 0.003934540785849094, + 0.00859960075467825, + 0.004389026667922735, + 0.004250090103596449, + 0.0029540890827775, + 0.004102259874343872, + 0.02862914837896824, + 0.02027761936187744, + 0.01924847438931465, + 0.020336907356977463, + 0.020199907943606377, + 0.13409967720508575, + 0.027670346200466156, + 0.16504840552806854, + 0.004088402725756168, + 0.003102674847468734, + 0.005089810583740473, + 0.003569877240806818, + 0.003915159963071346, + 0.005826857872307301, + 0.004193492233753204, + 0.0038981155958026648, + 0.016365712508559227, + 0.010388156399130821, + 0.01410653255879879, + 0.01468194741755724, + 0.012879151850938797, + 0.033121269196271896, + 0.040560461580753326, + 0.03253709524869919, + 0.003657144960016012, + 0.0030825226567685604, + 0.004011528100818396, + 0.0039027458988130093, + 0.0031285360455513, + 0.0035771860275417566, + 0.003953543491661549, + 0.0031437997240573168, + 0.004766527097672224, + 0.008374745026230812, + 0.0048143272288143635, + 0.006128610577434301, + 0.00785896647721529, + 0.010291090235114098, + 0.00531448470428586, + 0.008938534185290337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.2033257633447647 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene stylized like a Ukiyo-e woodblock print with flat perspective and paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily ascend from a low-angle shot of a stone pagoda base to reveal the full tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a powerful, jagged flash of brilliant white lightning connecting to the tower's tallest spire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the night sky display vibrant, flowing ribbons of auroras in emerald green, ethereal violet, and soft magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pagoda structure depicted with intense black outlines and flat planes of deep vermilion and dark-varnished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038117725402116776 + ], + "opposite_logprob_tokens": [ + -3.2881176471710205 + ], + "logprob_reward": -0.038117725402116776, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038117725402116776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1441529542207718, + 0.028188036754727364, + 0.0629296824336052, + 0.08571428060531616, + 0.11106128990650177, + 0.08173982799053192, + 0.05380959436297417, + 0.11172282695770264, + 0.05438058078289032, + 0.26630091667175293 + ], + "credit_map": [ + 0.015530318021774292, + 0.03335763141512871, + 0.014424978755414486, + 0.01126261055469513, + 0.012637152336537838, + 0.01130207534879446, + 0.013154576532542706, + 0.03248360753059387, + 0.0031227751169353724, + 0.0038943272083997726, + 0.0033910637721419334, + 0.0040775444358587265, + 0.0032195819076150656, + 0.003425424452871084, + 0.0033269240520894527, + 0.0037303962744772434, + 0.008079647086560726, + 0.008037611842155457, + 0.007690090220421553, + 0.006995516363531351, + 0.008046133443713188, + 0.007941313087940216, + 0.007078368216753006, + 0.009061002172529697, + 0.007115423679351807, + 0.011666342616081238, + 0.014558115974068642, + 0.01338511798530817, + 0.012183516286313534, + 0.008178485557436943, + 0.00961729884147644, + 0.009009984321892262, + 0.008118324913084507, + 0.021261343732476234, + 0.010638334788382053, + 0.02778051607310772, + 0.011143287643790245, + 0.013759688474237919, + 0.009087782353162766, + 0.009272015653550625, + 0.0065515125170350075, + 0.0049770064651966095, + 0.004675835371017456, + 0.005131816491484642, + 0.004198319744318724, + 0.024674970656633377, + 0.005457480438053608, + 0.02607288397848606, + 0.005890282802283764, + 0.0052778893150389194, + 0.006650748196989298, + 0.0054231067188084126, + 0.0043832045048475266, + 0.009313601069152355, + 0.010099709033966064, + 0.0067710550501942635, + 0.009136076085269451, + 0.00715464074164629, + 0.008849837817251682, + 0.010177417658269405, + 0.008725722320377827, + 0.02312433533370495, + 0.02404956705868244, + 0.020505225285887718, + 0.00626401836052537, + 0.005308989901095629, + 0.006747686769813299, + 0.0073937708511948586, + 0.00658105593174696, + 0.007648476865142584, + 0.007520576473325491, + 0.006916005164384842, + 0.01111422199755907, + 0.021520350128412247, + 0.01583760976791382, + 0.02424420602619648, + 0.01514541544020176, + 0.0796264111995697, + 0.018320441246032715, + 0.08049227297306061 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.038117725402116776 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene stylized like a Ukiyo-e woodblock print with flat perspective and paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily ascend from a low-angle shot of a stone pagoda base to reveal the full tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a powerful, jagged flash of brilliant white lightning connecting to the tower's tallest spire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the night sky display vibrant, flowing ribbons of auroras in emerald green, ethereal violet, and soft magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pagoda structure depicted with intense black outlines and flat planes of deep vermilion and dark-varnished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581606775522232 + ], + "opposite_logprob_tokens": [ + -3.8955817222595215 + ], + "logprob_reward": -0.020581606775522232, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581606775522232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06549891084432602, + 0.016175752505660057, + 0.04195621609687805, + 0.03632642328739166, + 0.02473529428243637, + 0.03024626523256302, + 0.05282654985785484, + 0.06361908465623856, + 0.09605561941862106, + 0.5725598931312561 + ], + "credit_map": [ + 0.004544566385447979, + 0.01871328428387642, + 0.006859562359750271, + 0.004209867678582668, + 0.005008874461054802, + 0.004445902071893215, + 0.006116010714322329, + 0.0156008405610919, + 0.002101166872307658, + 0.00217408686876297, + 0.002394304145127535, + 0.0024812601041048765, + 0.0012088000075891614, + 0.0018657397013157606, + 0.0017953133210539818, + 0.00215507997199893, + 0.003994886297732592, + 0.005753279197961092, + 0.0069865211844444275, + 0.004832603968679905, + 0.004302366636693478, + 0.005480467341840267, + 0.0044344207271933556, + 0.006171665154397488, + 0.003530468326061964, + 0.005222045350819826, + 0.0046426597982645035, + 0.006553113926202059, + 0.004247213248163462, + 0.0038592496421188116, + 0.00396858062595129, + 0.004303089343011379, + 0.003444162430241704, + 0.003071428742259741, + 0.003404609626159072, + 0.0034962270874530077, + 0.0030721905641257763, + 0.0022244008723646402, + 0.0028021703474223614, + 0.0032201018184423447, + 0.0029645112808793783, + 0.0021767213474959135, + 0.0026553082279860973, + 0.0022970805875957012, + 0.001998674124479294, + 0.007124071009457111, + 0.0019511401187628508, + 0.009078753180801868, + 0.004694296978414059, + 0.00398674514144659, + 0.0064999558962881565, + 0.003978324588388205, + 0.004373836796730757, + 0.010750832036137581, + 0.007663904223591089, + 0.010878651402890682, + 0.006210463587194681, + 0.004473388660699129, + 0.005016301292926073, + 0.005444423295557499, + 0.008763585239648819, + 0.01051633246243, + 0.013225517235696316, + 0.009969073347747326, + 0.010484525002539158, + 0.011757653206586838, + 0.011042220517992973, + 0.012234941124916077, + 0.01372185256332159, + 0.013984555378556252, + 0.006886905524879694, + 0.01594296470284462, + 0.03766535967588425, + 0.09175349771976471, + 0.037065789103507996, + 0.0560065396130085, + 0.05103522911667824, + 0.13641922175884247, + 0.039360567927360535, + 0.12325361371040344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.020581606775522232 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene stylized like a Ukiyo-e woodblock print with flat perspective and paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily ascend from a low-angle shot of a stone pagoda base to reveal the full tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a powerful, jagged flash of brilliant white lightning connecting to the tower's tallest spire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the night sky display vibrant, flowing ribbons of auroras in emerald green, ethereal violet, and soft magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pagoda structure depicted with intense black outlines and flat planes of deep vermilion and dark-varnished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002486472949385643 + ], + "opposite_logprob_tokens": [ + -6.002486705780029 + ], + "logprob_reward": -0.002486472949385643, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002486472949385643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07188475877046585, + 0.052656933665275574, + 0.03917238116264343, + 0.039812494069337845, + 0.045434366911649704, + 0.5004191398620605, + 0.05947759374976158, + 0.0643169954419136, + 0.054136574268341064, + 0.07268881797790527 + ], + "credit_map": [ + 0.008364475332200527, + 0.009685576893389225, + 0.00625436520203948, + 0.003680567489936948, + 0.012408413924276829, + 0.010431261733174324, + 0.00885234959423542, + 0.012207751162350178, + 0.0057436092756688595, + 0.008997291326522827, + 0.0064686983823776245, + 0.006160386372357607, + 0.006525280885398388, + 0.007514362223446369, + 0.004680854734033346, + 0.0065664504654705524, + 0.004269126337021589, + 0.004406672902405262, + 0.004640869330614805, + 0.005279473960399628, + 0.005370926111936569, + 0.005403472110629082, + 0.0041479552164673805, + 0.005653884261846542, + 0.004543038550764322, + 0.004171821754425764, + 0.00527927977964282, + 0.004372098483145237, + 0.008214648813009262, + 0.003689577803015709, + 0.004943949170410633, + 0.00459807924926281, + 0.004322667606174946, + 0.005768837407231331, + 0.006213082931935787, + 0.005665785167366266, + 0.007072296924889088, + 0.004382615443319082, + 0.005172367207705975, + 0.006836712826043367, + 0.02804422751069069, + 0.023777440190315247, + 0.021300211548805237, + 0.021772293373942375, + 0.01855294033885002, + 0.16359049081802368, + 0.030444646254181862, + 0.19293688237667084, + 0.006362855434417725, + 0.0044789486564695835, + 0.007818363606929779, + 0.004424305632710457, + 0.004462669603526592, + 0.010489065200090408, + 0.007822325453162193, + 0.013619058765470982, + 0.0062975105829536915, + 0.004512057173997164, + 0.005484073422849178, + 0.005899978801608086, + 0.005180820357054472, + 0.010934780351817608, + 0.015229938551783562, + 0.010777832008898258, + 0.007426420692354441, + 0.004474477376788855, + 0.005141058936715126, + 0.0041493759490549564, + 0.015286559239029884, + 0.0078011685982346535, + 0.0037432594690471888, + 0.006114249117672443, + 0.007506915833801031, + 0.010810984298586845, + 0.005893413443118334, + 0.006392552517354488, + 0.008506842888891697, + 0.014608427882194519, + 0.006054970435798168, + 0.012914708815515041 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.002486472949385643 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene stylized like a Ukiyo-e woodblock print with flat perspective and paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily ascend from a low-angle shot of a stone pagoda base to reveal the full tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a powerful, jagged flash of brilliant white lightning connecting to the tower's tallest spire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the night sky display vibrant, flowing ribbons of auroras in emerald green, ethereal violet, and soft magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pagoda structure depicted with intense black outlines and flat planes of deep vermilion and dark-varnished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251660011708736 + ], + "opposite_logprob_tokens": [ + -5.255251884460449 + ], + "logprob_reward": -0.005251660011708736, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251660011708736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.160604327917099, + 0.05033891275525093, + 0.06362752616405487, + 0.11485987901687622, + 0.09105649590492249, + 0.26586970686912537, + 0.028138697147369385, + 0.09689193964004517, + 0.03548293188214302, + 0.09312956035137177 + ], + "credit_map": [ + 0.017868366092443466, + 0.020731965079903603, + 0.016194578260183334, + 0.012167012318968773, + 0.032715003937482834, + 0.024493224918842316, + 0.01505950652062893, + 0.021374674513936043, + 0.00559192756190896, + 0.008797439746558666, + 0.006844513583928347, + 0.005018844734877348, + 0.006185024511069059, + 0.006420440506190062, + 0.0046533020213246346, + 0.006827421486377716, + 0.006568835116922855, + 0.007474961224943399, + 0.00781365018337965, + 0.00814789067953825, + 0.006797569803893566, + 0.00925251841545105, + 0.00792901311069727, + 0.009643083438277245, + 0.01141586247831583, + 0.017111212015151978, + 0.015062970109283924, + 0.02384955808520317, + 0.014703280292451382, + 0.009953903965651989, + 0.010705073364078999, + 0.012058014050126076, + 0.00921639148145914, + 0.011096998117864132, + 0.011532610282301903, + 0.010848315432667732, + 0.01372811570763588, + 0.012413802556693554, + 0.009009934030473232, + 0.013210327364504337, + 0.0210853461176157, + 0.01686316542327404, + 0.019130585715174675, + 0.01584295742213726, + 0.014127805829048157, + 0.0784933939576149, + 0.015472667291760445, + 0.08485377579927444, + 0.0026517855003476143, + 0.004042208194732666, + 0.004475654102861881, + 0.0035931819584220648, + 0.002693223999813199, + 0.0033869240432977676, + 0.004150311462581158, + 0.0031454076524823904, + 0.009283989667892456, + 0.006177618633955717, + 0.007480161730200052, + 0.009146662428975105, + 0.008051375858485699, + 0.01948510855436325, + 0.02094133384525776, + 0.01632569171488285, + 0.004443359561264515, + 0.003744968678802252, + 0.0037545920349657536, + 0.004533444996923208, + 0.007307814434170723, + 0.004914345685392618, + 0.002709654625505209, + 0.004074753727763891, + 0.005874737165868282, + 0.012118042446672916, + 0.007523068226873875, + 0.0077199069783091545, + 0.010608387179672718, + 0.021804040297865868, + 0.0077261184342205524, + 0.019755257293581963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.005251660011708736 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a scene stylized like a Ukiyo-e woodblock print, the camera steadily ascends, beginning with a low-angle shot of the stone base of a tall, ornate pagoda. The structure features intense black outlines and flat planes of deep vermilion and dark-varnished wood, set against a backdrop of gnarled pine trees silhouetted against the night sky. As the camera ascends, revealing the tower's multi-tiered structure and elegantly curved roof, a powerful, jagged flash of brilliant white lightning suddenly streaks across the sky, connecting violently and cracklingly with the tower's tallest spire. For a brief moment, the intense flash whitens colors and casts stark, simplified shadows, mimicking the high-contrast effect of a print. The camera continues its steady upward journey, passing through the spire to reveal the full celestial spectacle above: the deep indigo night sky is filled with vibrant, flowing, dancing ribbons of auroras, its undulating draperies of emerald green, ethereal violet, and soft magenta slowly and gracefully swirling. The entire composition maintains a flat perspective and a subtle paper texture, creating a dramatic and stunning atmosphere that blends the raw power of nature with the serene, decorative beauty of traditional Japanese art.", + "seed": 5268, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_1/index_9702_seed_5268_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04294773968867958, + "vlm_logprob": -0.04294773968867958, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.04294773968867958, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene stylized like a Ukiyo-e woodblock print with flat perspective and paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily ascend from a low-angle shot of a stone pagoda base to reveal the full tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a powerful, jagged flash of brilliant white lightning connecting to the tower's tallest spire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the night sky display vibrant, flowing ribbons of auroras in emerald green, ethereal violet, and soft magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pagoda structure depicted with intense black outlines and flat planes of deep vermilion and dark-varnished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16121463477611542 + ], + "opposite_logprob_tokens": [ + -1.9112145900726318 + ], + "logprob_reward": -0.16121463477611542, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16121463477611542, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14200831949710846, + 0.03630209341645241, + 0.04566604271531105, + 0.05721873417496681, + 0.11800958216190338, + 0.05461116135120392, + 0.020853087306022644, + 0.27373483777046204, + 0.029774431139230728, + 0.2218216508626938 + ], + "credit_map": [ + 0.014856851659715176, + 0.02193828485906124, + 0.016186127439141273, + 0.010966671630740166, + 0.0357041172683239, + 0.017132000997662544, + 0.012804582715034485, + 0.012419666163623333, + 0.0036577151622623205, + 0.004627369809895754, + 0.0027561797760427, + 0.004024438094347715, + 0.005092069040983915, + 0.008458592928946018, + 0.003583092475309968, + 0.004102634731680155, + 0.0033170958049595356, + 0.007390834391117096, + 0.0036224694922566414, + 0.012057791464030743, + 0.006038471590727568, + 0.004512771964073181, + 0.004320692270994186, + 0.0044059124775230885, + 0.003624309552833438, + 0.0054160067811608315, + 0.0038607106544077396, + 0.00584693718701601, + 0.016747482120990753, + 0.005158173386007547, + 0.01136036030948162, + 0.005204756278544664, + 0.019738834351301193, + 0.009543176740407944, + 0.027952564880251884, + 0.01699790172278881, + 0.0072977482341229916, + 0.0107204494997859, + 0.013920077122747898, + 0.01183883287012577, + 0.005087618250399828, + 0.002848295960575342, + 0.013514908961951733, + 0.003294429276138544, + 0.0029979590326547623, + 0.016200080513954163, + 0.006100102327764034, + 0.0045677656307816505, + 0.0021274855826050043, + 0.0027546759229153395, + 0.0025056754238903522, + 0.002199635375291109, + 0.002385661005973816, + 0.0029190839268267155, + 0.002644788008183241, + 0.0033160827588289976, + 0.08440544456243515, + 0.016804996877908707, + 0.028440989553928375, + 0.015195230953395367, + 0.011412424966692924, + 0.09622305631637573, + 0.0055409688502550125, + 0.01571175456047058, + 0.004614875651896, + 0.0028025037609040737, + 0.002656453987583518, + 0.003834611037746072, + 0.003615351626649499, + 0.006036277394741774, + 0.0018726104171946645, + 0.004341749008744955, + 0.0436241552233696, + 0.010729975998401642, + 0.01372272428125143, + 0.006220255512744188, + 0.010704601183533669, + 0.04432898014783859, + 0.0073878890834748745, + 0.08510306477546692 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.16121463477611542 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene stylized like a Ukiyo-e woodblock print with flat perspective and paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily ascend from a low-angle shot of a stone pagoda base to reveal the full tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a powerful, jagged flash of brilliant white lightning connecting to the tower's tallest spire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the night sky display vibrant, flowing ribbons of auroras in emerald green, ethereal violet, and soft magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pagoda structure depicted with intense black outlines and flat planes of deep vermilion and dark-varnished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043077558279037476 + ], + "opposite_logprob_tokens": [ + -3.1680774688720703 + ], + "logprob_reward": -0.043077558279037476, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043077558279037476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21329465508460999, + 0.07676777243614197, + 0.062353648245334625, + 0.17520752549171448, + 0.08771480619907379, + 0.10440988093614578, + 0.07115332782268524, + 0.09443733096122742, + 0.03177827224135399, + 0.08288277685642242 + ], + "credit_map": [ + 0.02315479889512062, + 0.04538051784038544, + 0.028778957203030586, + 0.019757801666855812, + 0.04010315611958504, + 0.017839189618825912, + 0.0182372834533453, + 0.020042940974235535, + 0.0065901679918169975, + 0.010719796642661095, + 0.009019587188959122, + 0.00994983408600092, + 0.010553994216024876, + 0.013638867996633053, + 0.008495502173900604, + 0.007800019346177578, + 0.0041326191276311874, + 0.009692142717540264, + 0.004370313603430986, + 0.01674395427107811, + 0.00994639378041029, + 0.004322350025177002, + 0.008368485607206821, + 0.004777388647198677, + 0.01411623228341341, + 0.016798807308077812, + 0.013007820583879948, + 0.02297292836010456, + 0.048706844449043274, + 0.015020790509879589, + 0.025666268542408943, + 0.018917834386229515, + 0.014943351969122887, + 0.007552412338554859, + 0.021937377750873566, + 0.011350391432642937, + 0.005914979614317417, + 0.00781136192381382, + 0.009911003522574902, + 0.00829392485320568, + 0.013406279496848583, + 0.011602215468883514, + 0.012645751237869263, + 0.010710104368627071, + 0.006762051023542881, + 0.014356077648699284, + 0.020379777997732162, + 0.014547625556588173, + 0.007590464781969786, + 0.01013642456382513, + 0.006358705461025238, + 0.009565684013068676, + 0.008375936187803745, + 0.007624058518558741, + 0.01044664066284895, + 0.011055415496230125, + 0.01272959727793932, + 0.013327107764780521, + 0.012843881733715534, + 0.010733992792665958, + 0.015436984598636627, + 0.010827265679836273, + 0.006916798185557127, + 0.011621708981692791, + 0.004433517809957266, + 0.003880386706441641, + 0.0030983146280050278, + 0.0033327355049550533, + 0.006249450147151947, + 0.004203129559755325, + 0.0028936266899108887, + 0.0036871128249913454, + 0.015628274530172348, + 0.004412231966853142, + 0.005303435027599335, + 0.0028098139446228743, + 0.004516103770583868, + 0.01606292836368084, + 0.003297888906672597, + 0.030852099880576134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.043077558279037476 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene stylized like a Ukiyo-e woodblock print with flat perspective and paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily ascend from a low-angle shot of a stone pagoda base to reveal the full tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a powerful, jagged flash of brilliant white lightning connecting to the tower's tallest spire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the night sky display vibrant, flowing ribbons of auroras in emerald green, ethereal violet, and soft magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pagoda structure depicted with intense black outlines and flat planes of deep vermilion and dark-varnished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036171742249280214 + ], + "opposite_logprob_tokens": [ + -5.628617286682129 + ], + "logprob_reward": -0.0036171742249280214, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036171742249280214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0735059380531311, + 0.04711484536528587, + 0.025208406150341034, + 0.028123818337917328, + 0.07800894975662231, + 0.07411615550518036, + 0.027199657633900642, + 0.11473984271287918, + 0.15782947838306427, + 0.37415292859077454 + ], + "credit_map": [ + 0.010926421731710434, + 0.015540712513029575, + 0.012491763569414616, + 0.004446919541805983, + 0.009575216099619865, + 0.007131656631827354, + 0.004518133122473955, + 0.008875105530023575, + 0.004329410381615162, + 0.006968808826059103, + 0.005026655271649361, + 0.005453582853078842, + 0.006330150179564953, + 0.008291667327284813, + 0.004369448870420456, + 0.006345117464661598, + 0.002242796588689089, + 0.00301249953918159, + 0.003711820114403963, + 0.0034078052267432213, + 0.002915137680247426, + 0.0023944757413119078, + 0.002936250763013959, + 0.004587619099766016, + 0.002333340235054493, + 0.002501382725313306, + 0.003166710026562214, + 0.0032365398947149515, + 0.0060631996020674706, + 0.0029198734555393457, + 0.00496786180883646, + 0.002934908727183938, + 0.012116583064198494, + 0.0065660797990858555, + 0.018884198740124702, + 0.012274079024791718, + 0.005006226245313883, + 0.006839379668235779, + 0.009219801984727383, + 0.007102594710886478, + 0.006310788448899984, + 0.0034139605704694986, + 0.018823595717549324, + 0.004024355672299862, + 0.0034456439316272736, + 0.02405501902103424, + 0.007524386048316956, + 0.006518397480249405, + 0.0030147454235702753, + 0.003223929787054658, + 0.002921558916568756, + 0.0025607587303966284, + 0.004038102459162474, + 0.003414109582081437, + 0.002951114671304822, + 0.005075334571301937, + 0.03094492480158806, + 0.010450627654790878, + 0.013544537127017975, + 0.007212718483060598, + 0.006956751458346844, + 0.03516269847750664, + 0.003292334033176303, + 0.007175246719270945, + 0.01997292786836624, + 0.013428202830255032, + 0.015324820764362812, + 0.015725109726190567, + 0.03543943911790848, + 0.013823991641402245, + 0.013930268585681915, + 0.030184704810380936, + 0.06694870442152023, + 0.02051905356347561, + 0.021029675379395485, + 0.014623656868934631, + 0.030164245516061783, + 0.06899939477443695, + 0.021079635247588158, + 0.13078853487968445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.0036171742249280214 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene stylized like a Ukiyo-e woodblock print with flat perspective and paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily ascend from a low-angle shot of a stone pagoda base to reveal the full tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a powerful, jagged flash of brilliant white lightning connecting to the tower's tallest spire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the night sky display vibrant, flowing ribbons of auroras in emerald green, ethereal violet, and soft magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pagoda structure depicted with intense black outlines and flat planes of deep vermilion and dark-varnished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004633048549294472 + ], + "opposite_logprob_tokens": [ + -5.379632949829102 + ], + "logprob_reward": -0.004633048549294472, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004633048549294472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2604271471500397, + 0.05164237320423126, + 0.047416917979717255, + 0.06584050506353378, + 0.09249256551265717, + 0.07779040187597275, + 0.058575134724378586, + 0.09051991254091263, + 0.0549619235098362, + 0.20033317804336548 + ], + "credit_map": [ + 0.026346756145358086, + 0.04679142311215401, + 0.02457767352461815, + 0.019381912425160408, + 0.06873758882284164, + 0.023972678929567337, + 0.02597055956721306, + 0.024648530408740044, + 0.00466005876660347, + 0.008906696923077106, + 0.004860034678131342, + 0.00474918307736516, + 0.005381920840591192, + 0.010491983033716679, + 0.0038594903890043497, + 0.008733000606298447, + 0.0042096097022295, + 0.007433946244418621, + 0.003426959505304694, + 0.0067523205652832985, + 0.007417650427669287, + 0.00635281577706337, + 0.005767363123595715, + 0.006056247279047966, + 0.007724470924586058, + 0.0058745332062244415, + 0.012623474933207035, + 0.006653614342212677, + 0.012374951504170895, + 0.00836973637342453, + 0.005836401134729385, + 0.006383320316672325, + 0.015759063884615898, + 0.007766708265990019, + 0.023015260696411133, + 0.01386982947587967, + 0.006221984047442675, + 0.007525312714278698, + 0.009715541265904903, + 0.008618860505521297, + 0.009516953490674496, + 0.004156413488090038, + 0.017510907724499702, + 0.005288871005177498, + 0.00362406182102859, + 0.0205734521150589, + 0.009667204692959785, + 0.007452534977346659, + 0.009892893955111504, + 0.00602206913754344, + 0.009356530383229256, + 0.005712587386369705, + 0.0031577751506119967, + 0.013420578092336655, + 0.004465719219297171, + 0.006546978373080492, + 0.01287107914686203, + 0.013838756829500198, + 0.01293872855603695, + 0.008896265178918839, + 0.013978361152112484, + 0.009181256406009197, + 0.007444765418767929, + 0.011370694264769554, + 0.008545775897800922, + 0.008381105959415436, + 0.004700665362179279, + 0.00656465208157897, + 0.007421736605465412, + 0.009861544705927372, + 0.004243622999638319, + 0.0052428157068789005, + 0.034427300095558167, + 0.011377940885722637, + 0.01792445033788681, + 0.012235961854457855, + 0.010912309400737286, + 0.0380302332341671, + 0.008510426618158817, + 0.06691453605890274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.004633048549294472 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene stylized like a Ukiyo-e woodblock print with flat perspective and paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily ascend from a low-angle shot of a stone pagoda base to reveal the full tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a powerful, jagged flash of brilliant white lightning connecting to the tower's tallest spire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the night sky display vibrant, flowing ribbons of auroras in emerald green, ethereal violet, and soft magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pagoda structure depicted with intense black outlines and flat planes of deep vermilion and dark-varnished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021962826140224934 + ], + "opposite_logprob_tokens": [ + -6.127196311950684 + ], + "logprob_reward": -0.0021962826140224934, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021962826140224934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22407011687755585, + 0.05733458697795868, + 0.05517292395234108, + 0.06053286790847778, + 0.2338879406452179, + 0.12662647664546967, + 0.033200930804014206, + 0.047048889100551605, + 0.040640659630298615, + 0.12148457020521164 + ], + "credit_map": [ + 0.02740488387644291, + 0.05783380940556526, + 0.02928415685892105, + 0.012861275114119053, + 0.034878991544246674, + 0.027528559789061546, + 0.011196069419384003, + 0.023082373663783073, + 0.004499069415032864, + 0.00659581832587719, + 0.006001038011163473, + 0.006644262932240963, + 0.00554449949413538, + 0.014796086587011814, + 0.0044366419315338135, + 0.008817173540592194, + 0.004590640310198069, + 0.012179146520793438, + 0.0042976499535143375, + 0.009069853462278843, + 0.009053274989128113, + 0.004840713459998369, + 0.006216162815690041, + 0.004925480578094721, + 0.005827745422720909, + 0.007911080494523048, + 0.010445771738886833, + 0.005198522005230188, + 0.011377424001693726, + 0.005265618208795786, + 0.007140664383769035, + 0.007366041652858257, + 0.03870408982038498, + 0.01881387270987034, + 0.058420632034540176, + 0.035624727606773376, + 0.013743425719439983, + 0.020041145384311676, + 0.02775602973997593, + 0.02078399434685707, + 0.010922065004706383, + 0.004782886244356632, + 0.03781410679221153, + 0.006135213654488325, + 0.0055548748932778835, + 0.0447668619453907, + 0.006794850807636976, + 0.009855612181127071, + 0.0035656881518661976, + 0.005392057355493307, + 0.004500665236264467, + 0.0028670444153249264, + 0.002961362712085247, + 0.004935353994369507, + 0.0036249388940632343, + 0.005353819113224745, + 0.008646230213344097, + 0.004133447539061308, + 0.007225600071251392, + 0.0037460385356098413, + 0.0041837687604129314, + 0.010336545296013355, + 0.0035995475482195616, + 0.005177709739655256, + 0.0067595369182527065, + 0.003966428339481354, + 0.004175407811999321, + 0.0036687557585537434, + 0.0065054441802203655, + 0.005858418066054583, + 0.002847161842510104, + 0.006859505083411932, + 0.020824532955884933, + 0.008319966495037079, + 0.007759899832308292, + 0.0061674704775214195, + 0.010860367678105831, + 0.02177819423377514, + 0.00536385690793395, + 0.04041029140353203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.0021962826140224934 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a scene stylized like a Ukiyo-e woodblock print, the camera steadily ascends, beginning with a low-angle shot of the stone base of a tall, ornate pagoda. The structure features intense black outlines and flat planes of deep vermilion and dark-varnished wood, set against a backdrop of gnarled pine trees silhouetted against the night sky. As the camera ascends, revealing the tower's multi-tiered structure and elegantly curved roof, a powerful, jagged flash of brilliant white lightning suddenly streaks across the sky, connecting violently and cracklingly with the tower's tallest spire. For a brief moment, the intense flash whitens colors and casts stark, simplified shadows, mimicking the high-contrast effect of a print. The camera continues its steady upward journey, passing through the spire to reveal the full celestial spectacle above: the deep indigo night sky is filled with vibrant, flowing, dancing ribbons of auroras, its undulating draperies of emerald green, ethereal violet, and soft magenta slowly and gracefully swirling. The entire composition maintains a flat perspective and a subtle paper texture, creating a dramatic and stunning atmosphere that blends the raw power of nature with the serene, decorative beauty of traditional Japanese art.", + "seed": 5269, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_1/index_9702_seed_5269_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03258183798752725, + "vlm_logprob": -0.03258183798752725, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.03258183798752725, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene stylized like a Ukiyo-e woodblock print with flat perspective and paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily ascend from a low-angle shot of a stone pagoda base to reveal the full tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a powerful, jagged flash of brilliant white lightning connecting to the tower's tallest spire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the night sky display vibrant, flowing ribbons of auroras in emerald green, ethereal violet, and soft magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pagoda structure depicted with intense black outlines and flat planes of deep vermilion and dark-varnished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10175421833992004 + ], + "opposite_logprob_tokens": [ + -2.3517541885375977 + ], + "logprob_reward": -0.10175421833992004, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10175421833992004, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11816317588090897, + 0.03944246843457222, + 0.02327951416373253, + 0.14195866882801056, + 0.46148550510406494, + 0.0420168861746788, + 0.03388034552335739, + 0.05176062136888504, + 0.04498005658388138, + 0.043032798916101456 + ], + "credit_map": [ + 0.01723715104162693, + 0.017582550644874573, + 0.011939993128180504, + 0.010220936499536037, + 0.022480962797999382, + 0.016056835651397705, + 0.012530488893389702, + 0.010114256292581558, + 0.00441126199439168, + 0.005892409943044186, + 0.0029692635871469975, + 0.00458668265491724, + 0.009485443122684956, + 0.005030483473092318, + 0.004435548558831215, + 0.002631375566124916, + 0.0018282439559698105, + 0.004022552631795406, + 0.0025890853721648455, + 0.0030911387875676155, + 0.003974789287894964, + 0.002051876625046134, + 0.003490178380161524, + 0.002231650287285447, + 0.014421548694372177, + 0.011970887891948223, + 0.008508166298270226, + 0.029946107417345047, + 0.011007474735379219, + 0.009603533893823624, + 0.045490264892578125, + 0.011010691523551941, + 0.06669870018959045, + 0.07244411110877991, + 0.04239778593182564, + 0.12736746668815613, + 0.024090200662612915, + 0.027132561430335045, + 0.08165497332811356, + 0.019699709489941597, + 0.0052291653119027615, + 0.007072730455547571, + 0.0037819917779415846, + 0.007511488161981106, + 0.0030727055855095387, + 0.00427324278280139, + 0.007355361245572567, + 0.0037202066741883755, + 0.0027832905761897564, + 0.005366084631532431, + 0.0028189101722091436, + 0.006682606413960457, + 0.004280728287994862, + 0.0025233414489775896, + 0.00702751474454999, + 0.002397866453975439, + 0.008790585212409496, + 0.006470936816185713, + 0.006193060893565416, + 0.006284521427005529, + 0.0031683549750596285, + 0.011055034585297108, + 0.005840292200446129, + 0.003957832232117653, + 0.0063265045173466206, + 0.0058101085014641285, + 0.00738309882581234, + 0.007097389083355665, + 0.004403540398925543, + 0.005436468869447708, + 0.005343939643353224, + 0.0031790086068212986, + 0.005700779613107443, + 0.004653675481677055, + 0.008292373269796371, + 0.005238283891230822, + 0.004125837702304125, + 0.006689175963401794, + 0.0036420985125005245, + 0.004690574016422033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.10175421833992004 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene stylized like a Ukiyo-e woodblock print with flat perspective and paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily ascend from a low-angle shot of a stone pagoda base to reveal the full tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a powerful, jagged flash of brilliant white lightning connecting to the tower's tallest spire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the night sky display vibrant, flowing ribbons of auroras in emerald green, ethereal violet, and soft magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pagoda structure depicted with intense black outlines and flat planes of deep vermilion and dark-varnished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495880916714668 + ], + "opposite_logprob_tokens": [ + -2.9299588203430176 + ], + "logprob_reward": -0.05495880916714668, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495880916714668, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07435228675603867, + 0.029422853142023087, + 0.05407031998038292, + 0.19274114072322845, + 0.18733808398246765, + 0.06570422649383545, + 0.1501140296459198, + 0.04035566374659538, + 0.06812266260385513, + 0.13777871429920197 + ], + "credit_map": [ + 0.008816835470497608, + 0.00932258553802967, + 0.011619159020483494, + 0.007012596819549799, + 0.011593720875680447, + 0.01135845948010683, + 0.008270075544714928, + 0.006358847487717867, + 0.003910633735358715, + 0.004496302455663681, + 0.0026309958193451166, + 0.0033717125188559294, + 0.007117487955838442, + 0.002758253365755081, + 0.0033653026912361383, + 0.0017721669282764196, + 0.004934117663651705, + 0.007985035888850689, + 0.004750384483486414, + 0.008333344012498856, + 0.008900294080376625, + 0.0070470524951815605, + 0.004277104511857033, + 0.007842982187867165, + 0.020125756040215492, + 0.019891193136572838, + 0.012570470571517944, + 0.04031490162014961, + 0.01965918205678463, + 0.015922117978334427, + 0.04984857887029648, + 0.014408943243324757, + 0.025885358452796936, + 0.03023480251431465, + 0.01829487830400467, + 0.050751280039548874, + 0.009698275476694107, + 0.010540266521275043, + 0.03268476203083992, + 0.009248464368283749, + 0.008909850381314754, + 0.010408635251224041, + 0.00771736353635788, + 0.010104237124323845, + 0.004879167769104242, + 0.006600735243409872, + 0.010271546430885792, + 0.006812686566263437, + 0.033964358270168304, + 0.009162473492324352, + 0.013803326524794102, + 0.012125907465815544, + 0.010791451670229435, + 0.04854485020041466, + 0.012212787754833698, + 0.009508867748081684, + 0.009474528022110462, + 0.005582812242209911, + 0.00360289728268981, + 0.0035836915485560894, + 0.004302499815821648, + 0.007007049862295389, + 0.0032085380516946316, + 0.003593646688386798, + 0.014254248701035976, + 0.005970343016088009, + 0.006132862996309996, + 0.00556514598429203, + 0.006599514279514551, + 0.019986731931567192, + 0.005566352978348732, + 0.004047457594424486, + 0.03343299776315689, + 0.025048106908798218, + 0.034662287682294846, + 0.013366509228944778, + 0.005963904783129692, + 0.012499256990849972, + 0.004112649708986282, + 0.008693015202879906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 6 + ], + "selected_reward": -0.05495880916714668 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene stylized like a Ukiyo-e woodblock print with flat perspective and paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily ascend from a low-angle shot of a stone pagoda base to reveal the full tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a powerful, jagged flash of brilliant white lightning connecting to the tower's tallest spire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the night sky display vibrant, flowing ribbons of auroras in emerald green, ethereal violet, and soft magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pagoda structure depicted with intense black outlines and flat planes of deep vermilion and dark-varnished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002198661444708705 + ], + "opposite_logprob_tokens": [ + -6.127198696136475 + ], + "logprob_reward": -0.002198661444708705, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002198661444708705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046466581523418427, + 0.029457902535796165, + 0.03494764864444733, + 0.12662893533706665, + 0.3301815986633301, + 0.15750959515571594, + 0.05392974615097046, + 0.03389474377036095, + 0.043538279831409454, + 0.1434449851512909 + ], + "credit_map": [ + 0.00492952112108469, + 0.007418733090162277, + 0.0043658106587827206, + 0.003938945010304451, + 0.011037112213671207, + 0.007500809617340565, + 0.003530337708070874, + 0.0037453114055097103, + 0.004778706002980471, + 0.004035812336951494, + 0.0026340046897530556, + 0.004519687965512276, + 0.0031307190656661987, + 0.0036517642438411713, + 0.0035048711579293013, + 0.003202335676178336, + 0.0021083811298012733, + 0.004834357183426619, + 0.002934590680524707, + 0.007392981555312872, + 0.00656095938757062, + 0.0022184085100889206, + 0.005482978653162718, + 0.0034149927087128162, + 0.010823477990925312, + 0.008868061006069183, + 0.011521521024405956, + 0.027389805763959885, + 0.008149323984980583, + 0.008220895193517208, + 0.040353357791900635, + 0.011302483268082142, + 0.05042649060487747, + 0.03978494927287102, + 0.033267516642808914, + 0.09986694902181625, + 0.013021096587181091, + 0.01701660268008709, + 0.06408575177192688, + 0.012712241150438786, + 0.019137468189001083, + 0.02626763842999935, + 0.016310520470142365, + 0.027108751237392426, + 0.01666904240846634, + 0.014521721750497818, + 0.026873836293816566, + 0.01062061358243227, + 0.008078191429376602, + 0.005803743377327919, + 0.003625575453042984, + 0.007075350731611252, + 0.00520998565480113, + 0.010870561935007572, + 0.007607677485793829, + 0.005658660549670458, + 0.004376768600195646, + 0.005355458240956068, + 0.003548258915543556, + 0.0035037838388234377, + 0.0018510435475036502, + 0.008292377926409245, + 0.0041509829461574554, + 0.0028160689398646355, + 0.008432311937212944, + 0.0038752430118620396, + 0.00567118963226676, + 0.00477626034989953, + 0.0032997753005474806, + 0.010574758984148502, + 0.003982387483119965, + 0.0029263533651828766, + 0.03379972651600838, + 0.020790092647075653, + 0.04048209264874458, + 0.013412745669484138, + 0.005951044149696827, + 0.013027808628976345, + 0.005061119329184294, + 0.010920348577201366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 9 + ], + "selected_reward": -0.002198661444708705 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene stylized like a Ukiyo-e woodblock print with flat perspective and paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily ascend from a low-angle shot of a stone pagoda base to reveal the full tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a powerful, jagged flash of brilliant white lightning connecting to the tower's tallest spire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the night sky display vibrant, flowing ribbons of auroras in emerald green, ethereal violet, and soft magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pagoda structure depicted with intense black outlines and flat planes of deep vermilion and dark-varnished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024861162528395653 + ], + "opposite_logprob_tokens": [ + -6.002486228942871 + ], + "logprob_reward": -0.0024861162528395653, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024861162528395653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053442709147930145, + 0.0249873548746109, + 0.027941372245550156, + 0.0316493920981884, + 0.3086051046848297, + 0.05606134235858917, + 0.09304735064506531, + 0.12346500903367996, + 0.08655377477407455, + 0.19424663484096527 + ], + "credit_map": [ + 0.005361407063901424, + 0.011702426709234715, + 0.005119211971759796, + 0.003418641397729516, + 0.011871930211782455, + 0.007199551910161972, + 0.00446575740352273, + 0.004303778987377882, + 0.004647584166377783, + 0.0034631588496267796, + 0.0030351774767041206, + 0.002704555867239833, + 0.0036085706669837236, + 0.0029464776162058115, + 0.002781596267595887, + 0.001800233731046319, + 0.002342094900086522, + 0.004033973906189203, + 0.0030477866530418396, + 0.004201016388833523, + 0.004793586675077677, + 0.0026857678312808275, + 0.0036663436330854893, + 0.0031707999296486378, + 0.0028045892249792814, + 0.0036990793887525797, + 0.00474059022963047, + 0.004835792817175388, + 0.0042837937362492085, + 0.00428860355168581, + 0.0035066159907728434, + 0.003490326227620244, + 0.046163830906152725, + 0.04076515883207321, + 0.030223337933421135, + 0.09072930365800858, + 0.013247492723166943, + 0.01627763733267784, + 0.05787832289934158, + 0.01331998873502016, + 0.008312651887536049, + 0.01296319905668497, + 0.0046807266771793365, + 0.0071048615500330925, + 0.0062173050828278065, + 0.005942054092884064, + 0.006596527528017759, + 0.00424401368945837, + 0.008741911500692368, + 0.007707515265792608, + 0.00732570281252265, + 0.023586144670844078, + 0.006336220074445009, + 0.005820472724735737, + 0.025670846924185753, + 0.00785854086279869, + 0.01853625290095806, + 0.026170214638113976, + 0.014214638620615005, + 0.01011457946151495, + 0.00899647455662489, + 0.021958956494927406, + 0.012470192275941372, + 0.011003694497048855, + 0.018068086355924606, + 0.006718738935887814, + 0.010169335640966892, + 0.006696369498968124, + 0.006485587917268276, + 0.02627161145210266, + 0.006821851711720228, + 0.0053221858106553555, + 0.045837778598070145, + 0.030676905065774918, + 0.05124415084719658, + 0.020641962066292763, + 0.00930073019117117, + 0.017695464193820953, + 0.006598310545086861, + 0.012251321226358414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 7 + ], + "selected_reward": -0.0024861162528395653 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene stylized like a Ukiyo-e woodblock print with flat perspective and paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily ascend from a low-angle shot of a stone pagoda base to reveal the full tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a powerful, jagged flash of brilliant white lightning connecting to the tower's tallest spire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the night sky display vibrant, flowing ribbons of auroras in emerald green, ethereal violet, and soft magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pagoda structure depicted with intense black outlines and flat planes of deep vermilion and dark-varnished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015113847330212593 + ], + "opposite_logprob_tokens": [ + -6.501511573791504 + ], + "logprob_reward": -0.0015113847330212593, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015113847330212593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16542233526706696, + 0.03751756250858307, + 0.07570228725671768, + 0.18712806701660156, + 0.20918843150138855, + 0.028299842029809952, + 0.0503280833363533, + 0.0709574893116951, + 0.04524582624435425, + 0.13021013140678406 + ], + "credit_map": [ + 0.021221674978733063, + 0.03804558143019676, + 0.013201694004237652, + 0.011666232720017433, + 0.0302704069763422, + 0.02849438227713108, + 0.009993004612624645, + 0.012529355473816395, + 0.004251495469361544, + 0.004653008654713631, + 0.005321730859577656, + 0.0056459023617208, + 0.006417887285351753, + 0.0035286773927509785, + 0.005006414372473955, + 0.0026924426201730967, + 0.0054756049066782, + 0.009473645128309727, + 0.008784033358097076, + 0.012750870548188686, + 0.010872340761125088, + 0.00765713257715106, + 0.012770024128258228, + 0.007918629795312881, + 0.01720154844224453, + 0.014165098778903484, + 0.01935798116028309, + 0.035918328911066055, + 0.01301377359777689, + 0.013188104145228863, + 0.05650629848241806, + 0.017776919528841972, + 0.03408355265855789, + 0.016844790428876877, + 0.023546461015939713, + 0.06723295152187347, + 0.0067207058891654015, + 0.009642181918025017, + 0.04394742473959923, + 0.007170339114964008, + 0.0035350609105080366, + 0.004303774796426296, + 0.002653175499290228, + 0.0048033008351922035, + 0.0024169650860130787, + 0.0034371246583759785, + 0.004641115665435791, + 0.0025093231815844774, + 0.008318602107465267, + 0.004621679428964853, + 0.004144586157053709, + 0.00488730613142252, + 0.003910400904715061, + 0.011311199516057968, + 0.006398036610335112, + 0.006736266892403364, + 0.015594618394970894, + 0.010319136083126068, + 0.007419024594128132, + 0.005880793556571007, + 0.005405965726822615, + 0.012215884402394295, + 0.007770591881126165, + 0.0063514718785882, + 0.005463475827127695, + 0.004456996452063322, + 0.008618198335170746, + 0.006983991712331772, + 0.0033481281716376543, + 0.005858476273715496, + 0.006815256085246801, + 0.0037013019900768995, + 0.030294286087155342, + 0.016109952703118324, + 0.04006372019648552, + 0.011850356124341488, + 0.006494974717497826, + 0.011368587613105774, + 0.004550271201878786, + 0.009477966465055943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.0015113847330212593 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a scene stylized like a Ukiyo-e woodblock print, the camera steadily ascends, beginning with a low-angle shot of the stone base of a tall, ornate pagoda. The structure features intense black outlines and flat planes of deep vermilion and dark-varnished wood, set against a backdrop of gnarled pine trees silhouetted against the night sky. As the camera ascends, revealing the tower's multi-tiered structure and elegantly curved roof, a powerful, jagged flash of brilliant white lightning suddenly streaks across the sky, connecting violently and cracklingly with the tower's tallest spire. For a brief moment, the intense flash whitens colors and casts stark, simplified shadows, mimicking the high-contrast effect of a print. The camera continues its steady upward journey, passing through the spire to reveal the full celestial spectacle above: the deep indigo night sky is filled with vibrant, flowing, dancing ribbons of auroras, its undulating draperies of emerald green, ethereal violet, and soft magenta slowly and gracefully swirling. The entire composition maintains a flat perspective and a subtle paper texture, creating a dramatic and stunning atmosphere that blends the raw power of nature with the serene, decorative beauty of traditional Japanese art.", + "seed": 5270, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_1/index_9702_seed_5270_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08492465612944215, + "vlm_logprob": -0.08492465612944215, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.08492465612944215, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene stylized like a Ukiyo-e woodblock print with flat perspective and paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily ascend from a low-angle shot of a stone pagoda base to reveal the full tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a powerful, jagged flash of brilliant white lightning connecting to the tower's tallest spire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the night sky display vibrant, flowing ribbons of auroras in emerald green, ethereal violet, and soft magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pagoda structure depicted with intense black outlines and flat planes of deep vermilion and dark-varnished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31410297751426697 + ], + "opposite_logprob_tokens": [ + -1.3141030073165894 + ], + "logprob_reward": -0.31410297751426697, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31410297751426697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05743907019495964, + 0.05635879561305046, + 0.15408900380134583, + 0.0814274325966835, + 0.12065741419792175, + 0.05534277856349945, + 0.044949762523174286, + 0.12032480537891388, + 0.2556625008583069, + 0.05374845489859581 + ], + "credit_map": [ + 0.00515016121789813, + 0.007023512851446867, + 0.008839874528348446, + 0.006503617390990257, + 0.007026681210845709, + 0.010048693045973778, + 0.005902964621782303, + 0.006943566258996725, + 0.00607644859701395, + 0.011889291927218437, + 0.005190697964280844, + 0.005878762807697058, + 0.008197113871574402, + 0.009674369357526302, + 0.004956445656716824, + 0.004495663568377495, + 0.014646506868302822, + 0.06103682890534401, + 0.00801489781588316, + 0.02595084346830845, + 0.023008443415164948, + 0.005676991771906614, + 0.008983475156128407, + 0.006771009415388107, + 0.00701649347320199, + 0.0077280206605792046, + 0.006691416259855032, + 0.010768509469926357, + 0.01980302855372429, + 0.011877238750457764, + 0.012160729616880417, + 0.0053819953463971615, + 0.01658983901143074, + 0.009008187800645828, + 0.013926582410931587, + 0.011356838978827, + 0.01211355160921812, + 0.03144080936908722, + 0.010891400277614594, + 0.015330199152231216, + 0.007585538551211357, + 0.0076311794109642506, + 0.010208896361291409, + 0.006659437902271748, + 0.005501061212271452, + 0.007923277094960213, + 0.004659970756620169, + 0.005173419136554003, + 0.005393047817051411, + 0.006523731164634228, + 0.007956248708069324, + 0.003432445228099823, + 0.004016012419015169, + 0.008151223883032799, + 0.00430811382830143, + 0.005168944131582975, + 0.0123404860496521, + 0.021921049803495407, + 0.016768988221883774, + 0.014425914734601974, + 0.01054542325437069, + 0.01787763088941574, + 0.013256370089948177, + 0.013188942335546017, + 0.014448567293584347, + 0.006008924450725317, + 0.09032464772462845, + 0.007524010259658098, + 0.014875320717692375, + 0.10449489206075668, + 0.00806246418505907, + 0.009923667646944523, + 0.007223706226795912, + 0.0064983973279595375, + 0.006144977640360594, + 0.0048652770929038525, + 0.007469498552381992, + 0.00947326049208641, + 0.0063123251311481, + 0.005761014297604561 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 4 + ], + "selected_reward": -0.31410297751426697 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene stylized like a Ukiyo-e woodblock print with flat perspective and paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily ascend from a low-angle shot of a stone pagoda base to reveal the full tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a powerful, jagged flash of brilliant white lightning connecting to the tower's tallest spire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the night sky display vibrant, flowing ribbons of auroras in emerald green, ethereal violet, and soft magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pagoda structure depicted with intense black outlines and flat planes of deep vermilion and dark-varnished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033135116100311 + ], + "opposite_logprob_tokens": [ + -2.3503313064575195 + ], + "logprob_reward": -0.10033135116100311, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033135116100311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10107966512441635, + 0.13381843268871307, + 0.15889951586723328, + 0.08434050530195236, + 0.1544879972934723, + 0.04972263425588608, + 0.04074960574507713, + 0.0751887783408165, + 0.15600812435150146, + 0.04570472985506058 + ], + "credit_map": [ + 0.008503153920173645, + 0.018979905173182487, + 0.013089380227029324, + 0.008291671983897686, + 0.01576543226838112, + 0.017679328098893166, + 0.007356569636613131, + 0.011414225213229656, + 0.014133102260529995, + 0.03414859622716904, + 0.015885990113019943, + 0.013545281253755093, + 0.017620770260691643, + 0.01895947940647602, + 0.010594923049211502, + 0.008930288255214691, + 0.01756291277706623, + 0.06177695840597153, + 0.008861519396305084, + 0.024090731516480446, + 0.02386295050382614, + 0.006595077458769083, + 0.009769369848072529, + 0.006379989441484213, + 0.0068638199009001255, + 0.0077871172688901424, + 0.007575586438179016, + 0.009775160811841488, + 0.017622685059905052, + 0.012248496524989605, + 0.01546416711062193, + 0.007003467995673418, + 0.01551682036370039, + 0.011557352729141712, + 0.03276648744940758, + 0.015501574613153934, + 0.016671890392899513, + 0.024913102388381958, + 0.01587224006652832, + 0.021688537672162056, + 0.0063477130606770515, + 0.006268121302127838, + 0.00882734265178442, + 0.006535869557410479, + 0.005705207586288452, + 0.004900544416159391, + 0.007403355557471514, + 0.003734481753781438, + 0.0051228879019618034, + 0.005827290937304497, + 0.005385837983340025, + 0.005496920086443424, + 0.005346986465156078, + 0.004525280557572842, + 0.005286816041916609, + 0.0037575862370431423, + 0.006900290492922068, + 0.007908198051154613, + 0.011905201710760593, + 0.008854174986481667, + 0.0096439765766263, + 0.011340536177158356, + 0.00969413761049509, + 0.00894225388765335, + 0.011830608360469341, + 0.010737678967416286, + 0.039430126547813416, + 0.014871940948069096, + 0.012895496562123299, + 0.04398506507277489, + 0.014420329593122005, + 0.00783688947558403, + 0.0041206213645637035, + 0.003992286045104265, + 0.0065024918876588345, + 0.004824609495699406, + 0.0057802870869636536, + 0.007843410596251488, + 0.006585611961781979, + 0.006055411882698536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 4 + ], + "selected_reward": -0.10033135116100311 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene stylized like a Ukiyo-e woodblock print with flat perspective and paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily ascend from a low-angle shot of a stone pagoda base to reveal the full tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a powerful, jagged flash of brilliant white lightning connecting to the tower's tallest spire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the night sky display vibrant, flowing ribbons of auroras in emerald green, ethereal violet, and soft magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pagoda structure depicted with intense black outlines and flat planes of deep vermilion and dark-varnished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001940988120622933 + ], + "opposite_logprob_tokens": [ + -6.251941204071045 + ], + "logprob_reward": -0.001940988120622933, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001940988120622933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07481744140386581, + 0.065409354865551, + 0.11178289353847504, + 0.09841271489858627, + 0.09744151681661606, + 0.049411170184612274, + 0.02677478827536106, + 0.08521401137113571, + 0.2876977026462555, + 0.10303840041160583 + ], + "credit_map": [ + 0.00803278386592865, + 0.012220646254718304, + 0.010141436941921711, + 0.007369956932961941, + 0.010840981267392635, + 0.011149815283715725, + 0.006357299629598856, + 0.008704518899321556, + 0.006891554221510887, + 0.02149192988872528, + 0.009208675473928452, + 0.006756865419447422, + 0.008544723503291607, + 0.005053169559687376, + 0.0036976917181164026, + 0.003764745546504855, + 0.012099255807697773, + 0.04000651836395264, + 0.010364944115281105, + 0.018058964982628822, + 0.01503433845937252, + 0.004446082282811403, + 0.006547324825078249, + 0.005225460510700941, + 0.006664544343948364, + 0.007908739149570465, + 0.006211626343429089, + 0.01519076619297266, + 0.01548665389418602, + 0.011863483116030693, + 0.02885439619421959, + 0.006232515908777714, + 0.010538819245994091, + 0.007114939857274294, + 0.01181968953460455, + 0.009280079044401646, + 0.010640322230756283, + 0.024747295305132866, + 0.00935619231313467, + 0.013944178819656372, + 0.00703763123601675, + 0.00531404372304678, + 0.007039338815957308, + 0.0065536885522305965, + 0.005738026462495327, + 0.0075514367781579494, + 0.006796015426516533, + 0.003380989423021674, + 0.004017219413071871, + 0.0028526054229587317, + 0.005703575909137726, + 0.002569616073742509, + 0.0020058448426425457, + 0.0037636549677699804, + 0.0026220050640404224, + 0.0032402686774730682, + 0.007140615023672581, + 0.013044877909123898, + 0.014307327568531036, + 0.00952767301350832, + 0.009295981377363205, + 0.011290802620351315, + 0.00874220672994852, + 0.011864523403346539, + 0.02023162506520748, + 0.02427399531006813, + 0.0687863752245903, + 0.03394626826047897, + 0.030636362731456757, + 0.057683613151311874, + 0.03145645186305046, + 0.02068299427628517, + 0.01467225793749094, + 0.009967721998691559, + 0.011346811428666115, + 0.010406849905848503, + 0.01219126395881176, + 0.016755474731326103, + 0.01801561564207077, + 0.00968240574002266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 9 + ], + "selected_reward": -0.001940988120622933 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene stylized like a Ukiyo-e woodblock print with flat perspective and paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily ascend from a low-angle shot of a stone pagoda base to reveal the full tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a powerful, jagged flash of brilliant white lightning connecting to the tower's tallest spire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the night sky display vibrant, flowing ribbons of auroras in emerald green, ethereal violet, and soft magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pagoda structure depicted with intense black outlines and flat planes of deep vermilion and dark-varnished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036128980573266745 + ], + "opposite_logprob_tokens": [ + -5.628612995147705 + ], + "logprob_reward": -0.0036128980573266745, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036128980573266745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06829191744327545, + 0.2318323254585266, + 0.05893766134977341, + 0.07414034754037857, + 0.19841642677783966, + 0.11421404033899307, + 0.05713867396116257, + 0.06297441571950912, + 0.07692188024520874, + 0.05713231861591339 + ], + "credit_map": [ + 0.005899850279092789, + 0.011744249612092972, + 0.007565822917968035, + 0.007126576732844114, + 0.011004670523107052, + 0.011803744360804558, + 0.005663229152560234, + 0.007483776193112135, + 0.02741178497672081, + 0.08405076712369919, + 0.027949543669819832, + 0.01667158491909504, + 0.03478781133890152, + 0.02040780894458294, + 0.010206439532339573, + 0.010346593335270882, + 0.005840093828737736, + 0.007790001109242439, + 0.013067541643977165, + 0.007892418652772903, + 0.005951805040240288, + 0.006339557468891144, + 0.006973992567509413, + 0.005082253832370043, + 0.008486812934279442, + 0.00585188390687108, + 0.007357827387750149, + 0.009380190633237362, + 0.01816375181078911, + 0.009491734206676483, + 0.009844671003520489, + 0.005563470534980297, + 0.02514740452170372, + 0.012559140101075172, + 0.024905875325202942, + 0.020089374855160713, + 0.021169835701584816, + 0.050011128187179565, + 0.018470901995897293, + 0.02606276236474514, + 0.016241107136011124, + 0.010646967217326164, + 0.017377126961946487, + 0.011698423884809017, + 0.013345065526664257, + 0.02341814897954464, + 0.010301120579242706, + 0.01118608471006155, + 0.008197532966732979, + 0.007670367136597633, + 0.008904348127543926, + 0.0046577854081988335, + 0.004400399047881365, + 0.010406287387013435, + 0.006161402445286512, + 0.0067405495792627335, + 0.008497059345245361, + 0.015117928385734558, + 0.00737149640917778, + 0.006530109327286482, + 0.005892679560929537, + 0.008697627112269402, + 0.0054363589733839035, + 0.005431155674159527, + 0.009631238877773285, + 0.008608490228652954, + 0.01767972856760025, + 0.007780205924063921, + 0.006054181605577469, + 0.012400156818330288, + 0.0074294996447861195, + 0.007338379975408316, + 0.005954757798463106, + 0.008067402057349682, + 0.006451547611504793, + 0.0052553764544427395, + 0.00769686559215188, + 0.007337984628975391, + 0.005804876331239939, + 0.010563510470092297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 5 + ], + "selected_reward": -0.0036128980573266745 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene stylized like a Ukiyo-e woodblock print with flat perspective and paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily ascend from a low-angle shot of a stone pagoda base to reveal the full tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a powerful, jagged flash of brilliant white lightning connecting to the tower's tallest spire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the night sky display vibrant, flowing ribbons of auroras in emerald green, ethereal violet, and soft magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pagoda structure depicted with intense black outlines and flat planes of deep vermilion and dark-varnished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004635065793991089 + ], + "opposite_logprob_tokens": [ + -5.379634857177734 + ], + "logprob_reward": -0.004635065793991089, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004635065793991089, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08185116201639175, + 0.15931405127048492, + 0.1545892208814621, + 0.0519794337451458, + 0.13735058903694153, + 0.05869568884372711, + 0.04484692960977554, + 0.07607477903366089, + 0.1929260641336441, + 0.04237211123108864 + ], + "credit_map": [ + 0.007295273710042238, + 0.016630033031105995, + 0.007760605774819851, + 0.005744046997278929, + 0.01816606894135475, + 0.006562274880707264, + 0.009062668308615685, + 0.010630186647176743, + 0.023265279829502106, + 0.05424763262271881, + 0.012243370525538921, + 0.015341374091804028, + 0.024421678856015205, + 0.012723863124847412, + 0.008796803653240204, + 0.008274032734334469, + 0.012248431332409382, + 0.059575166553258896, + 0.006469324231147766, + 0.02927693910896778, + 0.0227208249270916, + 0.007755313068628311, + 0.010499355383217335, + 0.00604384858161211, + 0.0037629948928952217, + 0.005170444957911968, + 0.0038827185053378344, + 0.007561387959867716, + 0.01235043816268444, + 0.005615083500742912, + 0.00915673840790987, + 0.004479624330997467, + 0.017671894282102585, + 0.009011803194880486, + 0.016375893726944923, + 0.01359716709703207, + 0.013757681474089622, + 0.03630882501602173, + 0.012837451882660389, + 0.01778985746204853, + 0.006961125414818525, + 0.006197818089276552, + 0.008014186285436153, + 0.0076699163764715195, + 0.007835615426301956, + 0.004956930410116911, + 0.012305250391364098, + 0.0047548385336995125, + 0.0051194061525166035, + 0.006490237079560757, + 0.007825399748980999, + 0.00454973429441452, + 0.0042128427885472775, + 0.006391160190105438, + 0.004907239228487015, + 0.0053509073331952095, + 0.007301063742488623, + 0.012466710060834885, + 0.011000925675034523, + 0.00927965622395277, + 0.007332515437155962, + 0.010636309161782265, + 0.008584209717810154, + 0.009473384357988834, + 0.010989359579980373, + 0.006079105660319328, + 0.055917732417583466, + 0.012568622827529907, + 0.013385309837758541, + 0.07085350900888443, + 0.013891112990677357, + 0.009241304360330105, + 0.0066885193809866905, + 0.005115492735058069, + 0.004613537807017565, + 0.0035094721242785454, + 0.005555686075240374, + 0.005693314131349325, + 0.0054709287360310555, + 0.005725160241127014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 2 + ], + "selected_reward": -0.004635065793991089 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_10/index_tensor([2761, 2761])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_10/index_tensor([2761, 2761])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8191484c858237b672ce13b8deea44d7b155e056 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_10/index_tensor([2761, 2761])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-frame-rate, horizontal medium shot gently draws the viewer's attention to a white-haired grandfather, seated peacefully in a dark brown wooden rocking chair. Warm afternoon sunlight, like molten honey, streams through the large side windows, illuminating the cozy, bookish room. The grandfather, with a head full of silver hair and a face marked by the kindly wrinkles of time, wears a soft beige wool cardigan over a light-colored shirt. He hunches his back slightly, pressing a dry, age-freckled hand to his mouth as he lets out a soft, muffled cough. The high-frame-rate lens plays an extremely effective role in this moment: his thin shoulders tremble slightly but clearly, and every muscle contraction is slowed down into poetic detail; at the same time, in the golden beam of light that penetrates the room, countless tiny dust particles that were originally invisible to the naked eye are captured clearly. They are like disturbed golden elves, dancing, spinning and flickering slowly and gracefully in the air, forming a contrast between movement and stillness, grandeur and microcosm with the slight tremor of the old man's body. The whole picture is filled with a peaceful, nostalgic and warm atmosphere with a hint of the fragility of time under the ultimate clarity.", + "seed": 31389, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_10/index_2761_seed_31389_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03563338173553347, + "vlm_logprob": -0.03563338173553347, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.03563338173553347, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-haired grandfather seated in a dark brown wooden rocking chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a soft beige wool cardigan over a light-colored shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the grandfather perform the action of coughing with a hand pressed to his mouth while his shoulders tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible dancing and spinning in a golden beam of sunlight streaming through large side windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a high-frame-rate horizontal medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01572905294597149 + ], + "opposite_logprob_tokens": [ + -4.515728950500488 + ], + "logprob_reward": -0.01572905294597149, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01572905294597149, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15910205245018005, + 0.12954232096672058, + 0.04352336749434471, + 0.35300302505493164, + 0.030353132635354996, + 0.04572565481066704, + 0.04925629496574402, + 0.04472234845161438, + 0.10365775972604752, + 0.04111410677433014 + ], + "credit_map": [ + 0.021475138142704964, + 0.019710974767804146, + 0.021603848785161972, + 0.01693783514201641, + 0.01536718662828207, + 0.0191192664206028, + 0.026016004383563995, + 0.018871786072850227, + 0.007217305712401867, + 0.008548527956008911, + 0.016986044123768806, + 0.017435548827052116, + 0.0216621533036232, + 0.008591973222792149, + 0.031715381890535355, + 0.017385371029376984, + 0.0029974798671901226, + 0.0038711191155016422, + 0.004992883186787367, + 0.005986668169498444, + 0.0067718480713665485, + 0.005021080840379, + 0.00882805697619915, + 0.005054228939116001, + 0.03163013234734535, + 0.026186658069491386, + 0.036553557962179184, + 0.07689856737852097, + 0.05142362788319588, + 0.04268911853432655, + 0.06127414107322693, + 0.026347201317548752, + 0.00313516054302454, + 0.002743849763646722, + 0.003413514234125614, + 0.0035392395220696926, + 0.0035211367066949606, + 0.004351933021098375, + 0.006360800936818123, + 0.0032874951139092445, + 0.0036947070620954037, + 0.002990728709846735, + 0.005412809085100889, + 0.0050043645314872265, + 0.006919659208506346, + 0.004929666873067617, + 0.011722933501005173, + 0.0050507825799286366, + 0.004731296561658382, + 0.004244149662554264, + 0.0062890625558793545, + 0.006217045243829489, + 0.00689702620729804, + 0.00579709792509675, + 0.00968482531607151, + 0.005395790096372366, + 0.004657099489122629, + 0.003946964163333178, + 0.005743703804910183, + 0.0074025727808475494, + 0.005209588911384344, + 0.004037003964185715, + 0.007579689845442772, + 0.006145719904452562, + 0.006391922943294048, + 0.006864813156425953, + 0.015608707442879677, + 0.010545482859015465, + 0.019826311618089676, + 0.008879031985998154, + 0.026915688067674637, + 0.00862579233944416, + 0.004189420957118273, + 0.003412567311897874, + 0.005303848069161177, + 0.004501106683164835, + 0.008212306536734104, + 0.004337589256465435, + 0.007703480776399374, + 0.0034537827596068382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.01572905294597149 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-haired grandfather seated in a dark brown wooden rocking chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a soft beige wool cardigan over a light-colored shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the grandfather perform the action of coughing with a hand pressed to his mouth while his shoulders tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible dancing and spinning in a golden beam of sunlight streaming through large side windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a high-frame-rate horizontal medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076450384221971035 + ], + "opposite_logprob_tokens": [ + -4.882645130157471 + ], + "logprob_reward": -0.0076450384221971035, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076450384221971035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11316560953855515, + 0.050832781940698624, + 0.02439926378428936, + 0.6332669854164124, + 0.02140984684228897, + 0.01780656725168228, + 0.030146142467856407, + 0.054221563041210175, + 0.0349557101726532, + 0.01979556493461132 + ], + "credit_map": [ + 0.013700717128813267, + 0.012010961771011353, + 0.013047188520431519, + 0.012785096652805805, + 0.013414571061730385, + 0.012890568934381008, + 0.019030386582016945, + 0.016286112368106842, + 0.004119955003261566, + 0.004487460944801569, + 0.005658538080751896, + 0.007854403927922249, + 0.00841631181538105, + 0.004499578382819891, + 0.009049024432897568, + 0.006747500970959663, + 0.002357536694034934, + 0.0015366108855232596, + 0.0030017264652997255, + 0.003647529985755682, + 0.003497686702758074, + 0.0017740317853167653, + 0.00525488518178463, + 0.003329252591356635, + 0.056783657521009445, + 0.04702774062752724, + 0.07001639157533646, + 0.12258852273225784, + 0.08823779970407486, + 0.0808136910200119, + 0.11941798031330109, + 0.04838118329644203, + 0.0017192161176353693, + 0.0017532012425363064, + 0.002548667136579752, + 0.003021601354703307, + 0.0031205713748931885, + 0.0019064084626734257, + 0.004564570263028145, + 0.0027756094932556152, + 0.0014184769243001938, + 0.0011524513829499483, + 0.0023657663259655237, + 0.002358874538913369, + 0.0027417100500315428, + 0.0015352381160482764, + 0.004275118932127953, + 0.0019589283037930727, + 0.002308890223503113, + 0.002009963383898139, + 0.0044689602218568325, + 0.003531636670231819, + 0.0031342576257884502, + 0.0038801827467978, + 0.005166881717741489, + 0.005645365919917822, + 0.00417522806674242, + 0.0045132823288440704, + 0.0067947376519441605, + 0.005977704655379057, + 0.007463785354048014, + 0.006980120670050383, + 0.012711046263575554, + 0.005605657119303942, + 0.002366222208365798, + 0.002865055575966835, + 0.004947064910084009, + 0.004619095008820295, + 0.005341179668903351, + 0.003080782014876604, + 0.007517718244343996, + 0.004218589048832655, + 0.0018576513975858688, + 0.0018758850637823343, + 0.0024371123872697353, + 0.0021731897722929716, + 0.0030685216188430786, + 0.0026446040719747543, + 0.00351984822191298, + 0.002218750538304448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.0076450384221971035 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-haired grandfather seated in a dark brown wooden rocking chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a soft beige wool cardigan over a light-colored shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the grandfather perform the action of coughing with a hand pressed to his mouth while his shoulders tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible dancing and spinning in a golden beam of sunlight streaming through large side windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a high-frame-rate horizontal medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979334630072117 + ], + "opposite_logprob_tokens": [ + -3.5297932624816895 + ], + "logprob_reward": -0.02979334630072117, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979334630072117, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2540196180343628, + 0.056330230087041855, + 0.06885069608688354, + 0.10612128674983978, + 0.09594737738370895, + 0.09078126400709152, + 0.04902385175228119, + 0.09843508899211884, + 0.06958623975515366, + 0.11090431362390518 + ], + "credit_map": [ + 0.021565740928053856, + 0.020433884114027023, + 0.030037356540560722, + 0.02976609207689762, + 0.035412345081567764, + 0.02222139947116375, + 0.05459048971533775, + 0.03999228775501251, + 0.0042845718562603, + 0.004319511819630861, + 0.006649975664913654, + 0.008089111186563969, + 0.007569645065814257, + 0.006304111797362566, + 0.01069163903594017, + 0.008421659469604492, + 0.005037523340433836, + 0.005812016315758228, + 0.007673085201531649, + 0.010887990705668926, + 0.012285384349524975, + 0.005986548960208893, + 0.014024627394974232, + 0.007143521681427956, + 0.008016340434551239, + 0.008662541396915913, + 0.011155523359775543, + 0.0224524587392807, + 0.016248824074864388, + 0.01142532005906105, + 0.019398322328925133, + 0.008761962875723839, + 0.006090021226555109, + 0.006852715276181698, + 0.013226253911852837, + 0.013932566158473492, + 0.0116162383928895, + 0.00910673476755619, + 0.023557592183351517, + 0.011565258726477623, + 0.005173926707357168, + 0.006710032932460308, + 0.01113019697368145, + 0.012286865152418613, + 0.01346504595130682, + 0.00980196800082922, + 0.02276848442852497, + 0.009444746188819408, + 0.0037755260709673166, + 0.003950208891183138, + 0.005932226777076721, + 0.004870990291237831, + 0.008507371880114079, + 0.005357345100492239, + 0.008448773063719273, + 0.008181409910321236, + 0.010425955057144165, + 0.007211070042103529, + 0.012478564865887165, + 0.012093592435121536, + 0.010193418711423874, + 0.013323484919965267, + 0.017563389614224434, + 0.015145612880587578, + 0.004118998534977436, + 0.005803622771054506, + 0.009841462597250938, + 0.009301265701651573, + 0.009612616151571274, + 0.006276756525039673, + 0.017299113795161247, + 0.007332403678447008, + 0.009428028017282486, + 0.009177117608487606, + 0.014459450729191303, + 0.019902506843209267, + 0.014261752367019653, + 0.012852419167757034, + 0.02039588987827301, + 0.010427156463265419 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.02979334630072117 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-haired grandfather seated in a dark brown wooden rocking chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a soft beige wool cardigan over a light-colored shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the grandfather perform the action of coughing with a hand pressed to his mouth while his shoulders tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible dancing and spinning in a golden beam of sunlight streaming through large side windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a high-frame-rate horizontal medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700293779373169 + ], + "opposite_logprob_tokens": [ + -2.6950292587280273 + ], + "logprob_reward": -0.0700293779373169, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700293779373169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2599886357784271, + 0.13885460793972015, + 0.10671521723270416, + 0.16810977458953857, + 0.054023224860429764, + 0.08288754522800446, + 0.06875244528055191, + 0.0541502870619297, + 0.03271230310201645, + 0.033806007355451584 + ], + "credit_map": [ + 0.028858104720711708, + 0.02521676942706108, + 0.02733360044658184, + 0.02881130576133728, + 0.024697666987776756, + 0.0385783351957798, + 0.04409279301762581, + 0.04240003600716591, + 0.011337347328662872, + 0.017735745757818222, + 0.014787174761295319, + 0.01665930449962616, + 0.017419027164578438, + 0.020975498482584953, + 0.021236756816506386, + 0.018703753128647804, + 0.009479869157075882, + 0.01716071553528309, + 0.010734513401985168, + 0.011548265814781189, + 0.012503718957304955, + 0.020297225564718246, + 0.014421106316149235, + 0.0105697987601161, + 0.018356803804636, + 0.016612650826573372, + 0.02031167969107628, + 0.02003451995551586, + 0.02176855318248272, + 0.024715567007660866, + 0.03251044079661369, + 0.013799558393657207, + 0.006059635430574417, + 0.0072384728118777275, + 0.006440028082579374, + 0.004586418159306049, + 0.0058076209388673306, + 0.00890110619366169, + 0.010211054235696793, + 0.004778887610882521, + 0.008105036802589893, + 0.0062571121379733086, + 0.010140400379896164, + 0.011680196039378643, + 0.009490120224654675, + 0.011483623646199703, + 0.017030317336320877, + 0.008700736798346043, + 0.006379884667694569, + 0.005492120049893856, + 0.008054615929722786, + 0.007065996993333101, + 0.010143049992620945, + 0.010371622629463673, + 0.011234375648200512, + 0.010010778903961182, + 0.004459481220692396, + 0.004295687656849623, + 0.006098982412368059, + 0.007593888323754072, + 0.007980341091752052, + 0.007859586738049984, + 0.009661111049354076, + 0.00620120856910944, + 0.0028027959633618593, + 0.003256917232647538, + 0.004685511346906424, + 0.0033736072946339846, + 0.0038790800608694553, + 0.0041743372566998005, + 0.007817390374839306, + 0.0027226621750742197, + 0.003386313561350107, + 0.002527075121179223, + 0.0045416224747896194, + 0.005597750656306744, + 0.0040433998219668865, + 0.003702351823449135, + 0.005680736619979143, + 0.004326756577938795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.0700293779373169 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-haired grandfather seated in a dark brown wooden rocking chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a soft beige wool cardigan over a light-colored shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the grandfather perform the action of coughing with a hand pressed to his mouth while his shoulders tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible dancing and spinning in a golden beam of sunlight streaming through large side windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a high-frame-rate horizontal medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054970093071460724 + ], + "opposite_logprob_tokens": [ + -2.9299700260162354 + ], + "logprob_reward": -0.054970093071460724, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054970093071460724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23074014484882355, + 0.07035024464130402, + 0.045667000114917755, + 0.34723883867263794, + 0.08215603977441788, + 0.03844195976853371, + 0.04123793542385101, + 0.08012712746858597, + 0.03673553094267845, + 0.02730513922870159 + ], + "credit_map": [ + 0.0190202035009861, + 0.02202301286160946, + 0.026608876883983612, + 0.024795936420559883, + 0.03064718283712864, + 0.027255069464445114, + 0.037710025906562805, + 0.04267982393503189, + 0.005627111066132784, + 0.00595514290034771, + 0.012361233122646809, + 0.006882926449179649, + 0.010710913687944412, + 0.00716911256313324, + 0.013228324241936207, + 0.00841547641903162, + 0.0037987756077200174, + 0.004514236934483051, + 0.005814863368868828, + 0.005125300958752632, + 0.0073442100547254086, + 0.008014870807528496, + 0.006814164575189352, + 0.004240576643496752, + 0.03129854053258896, + 0.027070609852671623, + 0.0393105074763298, + 0.06524769216775894, + 0.049458764493465424, + 0.04639444872736931, + 0.06305626779794693, + 0.0254020094871521, + 0.005896132905036211, + 0.005932585336267948, + 0.009748799726366997, + 0.01245913840830326, + 0.012308744713664055, + 0.009947095066308975, + 0.017399262636899948, + 0.008464285172522068, + 0.002914705313742161, + 0.0027836773078888655, + 0.004367744084447622, + 0.00647298339754343, + 0.005346597172319889, + 0.005640230141580105, + 0.007569202687591314, + 0.003346820129081607, + 0.005346633028239012, + 0.0024107620120048523, + 0.004655718337744474, + 0.0034764865413308144, + 0.006514017004519701, + 0.005434323102235794, + 0.007843551225960255, + 0.00555644603446126, + 0.007037040777504444, + 0.00565342465415597, + 0.01102794986218214, + 0.011015777476131916, + 0.008344605565071106, + 0.01025933213531971, + 0.017325453460216522, + 0.009463543072342873, + 0.0021849232725799084, + 0.0028002194594591856, + 0.005405216012150049, + 0.0058025820180773735, + 0.005712389480322599, + 0.002375795738771558, + 0.009174211882054806, + 0.0032801947090774775, + 0.0026375597808510065, + 0.0018237653421238065, + 0.003758806036785245, + 0.003078117733821273, + 0.0046478016301989555, + 0.0026376231107860804, + 0.006125811953097582, + 0.0025956525932997465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.054970093071460724 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-frame-rate, horizontal medium shot gently draws the viewer's attention to a white-haired grandfather, seated peacefully in a dark brown wooden rocking chair. Warm afternoon sunlight, like molten honey, streams through the large side windows, illuminating the cozy, bookish room. The grandfather, with a head full of silver hair and a face marked by the kindly wrinkles of time, wears a soft beige wool cardigan over a light-colored shirt. He hunches his back slightly, pressing a dry, age-freckled hand to his mouth as he lets out a soft, muffled cough. The high-frame-rate lens plays an extremely effective role in this moment: his thin shoulders tremble slightly but clearly, and every muscle contraction is slowed down into poetic detail; at the same time, in the golden beam of light that penetrates the room, countless tiny dust particles that were originally invisible to the naked eye are captured clearly. They are like disturbed golden elves, dancing, spinning and flickering slowly and gracefully in the air, forming a contrast between movement and stillness, grandeur and microcosm with the slight tremor of the old man's body. The whole picture is filled with a peaceful, nostalgic and warm atmosphere with a hint of the fragility of time under the ultimate clarity.", + "seed": 31390, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_10/index_2761_seed_31390_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0469693474471569, + "vlm_logprob": -0.0469693474471569, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.0469693474471569, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-haired grandfather seated in a dark brown wooden rocking chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a soft beige wool cardigan over a light-colored shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the grandfather perform the action of coughing with a hand pressed to his mouth while his shoulders tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible dancing and spinning in a golden beam of sunlight streaming through large side windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a high-frame-rate horizontal medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02855031192302704 + ], + "opposite_logprob_tokens": [ + -3.653550386428833 + ], + "logprob_reward": -0.02855031192302704, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02855031192302704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2780360281467438, + 0.1648125797510147, + 0.11275991797447205, + 0.08470788598060608, + 0.10061280429363251, + 0.05217806249856949, + 0.11571300029754639, + 0.03330962732434273, + 0.03784237056970596, + 0.020027806982398033 + ], + "credit_map": [ + 0.022531650960445404, + 0.023145873099565506, + 0.021788857877254486, + 0.02316603809595108, + 0.05805347114801407, + 0.03307528421282768, + 0.04141905903816223, + 0.054855771362781525, + 0.012341652065515518, + 0.013220658525824547, + 0.017255879938602448, + 0.018991367891430855, + 0.03658052906394005, + 0.014065605588257313, + 0.02860094979405403, + 0.02375592291355133, + 0.006812140811234713, + 0.009246783331036568, + 0.007482856512069702, + 0.01375569123774767, + 0.031177565455436707, + 0.008335056714713573, + 0.016432125121355057, + 0.019517695531249046, + 0.00591124314814806, + 0.008032460696995258, + 0.007490865420550108, + 0.006804137025028467, + 0.02293648198246956, + 0.00843098759651184, + 0.012751441448926926, + 0.012350261211395264, + 0.006121557671576738, + 0.010021097026765347, + 0.007845979183912277, + 0.01196463406085968, + 0.025415712967514992, + 0.009890967048704624, + 0.015955688431859016, + 0.0133971581235528, + 0.004554553423076868, + 0.004374289885163307, + 0.003773574950173497, + 0.004293182399123907, + 0.010579130612313747, + 0.007754928432404995, + 0.008050207048654556, + 0.008798194117844105, + 0.010112550109624863, + 0.01250947080552578, + 0.010703655891120434, + 0.009883222170174122, + 0.021438991650938988, + 0.017222629860043526, + 0.014353187754750252, + 0.019489282742142677, + 0.002384376246482134, + 0.0040009417571127415, + 0.0018426249735057354, + 0.003165544243529439, + 0.008251973427832127, + 0.0033855908550322056, + 0.004749988671392202, + 0.005528584588319063, + 0.0029254420660436153, + 0.0031166106928139925, + 0.002596671925857663, + 0.004903951659798622, + 0.009273400530219078, + 0.004417661111801863, + 0.005678233224898577, + 0.004930398426949978, + 0.0014612337108701468, + 0.002133346162736416, + 0.0014831260778009892, + 0.0021198331378400326, + 0.005470318719744682, + 0.0018933172104880214, + 0.002545940224081278, + 0.0029206895269453526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.02855031192302704 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-haired grandfather seated in a dark brown wooden rocking chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a soft beige wool cardigan over a light-colored shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the grandfather perform the action of coughing with a hand pressed to his mouth while his shoulders tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible dancing and spinning in a golden beam of sunlight streaming through large side windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a high-frame-rate horizontal medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014211166650056839 + ], + "opposite_logprob_tokens": [ + -4.264211177825928 + ], + "logprob_reward": -0.014211166650056839, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014211166650056839, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33742663264274597, + 0.2791871726512909, + 0.07943660765886307, + 0.0646066889166832, + 0.05534873530268669, + 0.02932034246623516, + 0.06112303584814072, + 0.02817942388355732, + 0.03445613384246826, + 0.030915185809135437 + ], + "credit_map": [ + 0.022161519154906273, + 0.027250006794929504, + 0.02686770260334015, + 0.03992059454321861, + 0.07330863177776337, + 0.03395461291074753, + 0.06635365635156631, + 0.04760989546775818, + 0.014820429496467113, + 0.02211618795990944, + 0.021095119416713715, + 0.026588071137666702, + 0.06773117184638977, + 0.02786453627049923, + 0.05773391202092171, + 0.04123775288462639, + 0.004748817067593336, + 0.006808525417000055, + 0.004978226963430643, + 0.008947559632360935, + 0.023076744750142097, + 0.007999868132174015, + 0.01139957644045353, + 0.011477288790047169, + 0.004212648142129183, + 0.007814920507371426, + 0.00440653832629323, + 0.004935269244015217, + 0.019169989973306656, + 0.006493648048490286, + 0.007937081158161163, + 0.009636602364480495, + 0.003199866274371743, + 0.00460686394944787, + 0.003961448557674885, + 0.005987996701151133, + 0.012914680875837803, + 0.004768849816173315, + 0.010809497907757759, + 0.009099524468183517, + 0.0018959053559228778, + 0.002955131931230426, + 0.002945450833067298, + 0.003375515341758728, + 0.005014603491872549, + 0.003913021646440029, + 0.005104713141918182, + 0.0041160015389323235, + 0.005847799126058817, + 0.007685599848628044, + 0.004966087639331818, + 0.006692009512335062, + 0.011948510073125362, + 0.007143884431570768, + 0.006233252584934235, + 0.010605891235172749, + 0.0019284177105873823, + 0.0025371096562594175, + 0.0019582731183618307, + 0.0030243494547903538, + 0.006710898131132126, + 0.003013518638908863, + 0.004422743804752827, + 0.004584111738950014, + 0.002352718263864517, + 0.003208716167137027, + 0.002271652454510331, + 0.0043091922998428345, + 0.007985266856849194, + 0.0024299868382513523, + 0.0070505947805941105, + 0.004848005250096321, + 0.0024653926957398653, + 0.003693913109600544, + 0.0033394235651940107, + 0.004201458301395178, + 0.006186759565025568, + 0.0024104295298457146, + 0.004900471307337284, + 0.0037173351738601923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.014211166650056839 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-haired grandfather seated in a dark brown wooden rocking chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a soft beige wool cardigan over a light-colored shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the grandfather perform the action of coughing with a hand pressed to his mouth while his shoulders tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible dancing and spinning in a golden beam of sunlight streaming through large side windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a high-frame-rate horizontal medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896961271762848 + ], + "opposite_logprob_tokens": [ + -2.578969717025757 + ], + "logprob_reward": -0.07896961271762848, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896961271762848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5233893394470215, + 0.0819549486041069, + 0.0890900045633316, + 0.05287390574812889, + 0.053644999861717224, + 0.06767584383487701, + 0.026500791311264038, + 0.030304832383990288, + 0.043489109724760056, + 0.03107619658112526 + ], + "credit_map": [ + 0.02690165676176548, + 0.03399459272623062, + 0.024396881461143494, + 0.03554534539580345, + 0.10549280047416687, + 0.050717681646347046, + 0.16013343632221222, + 0.08620699495077133, + 0.006244180258363485, + 0.005990119650959969, + 0.009952640160918236, + 0.009222070686519146, + 0.012718687765300274, + 0.00895826704800129, + 0.0142195550724864, + 0.014649423770606518, + 0.005806813482195139, + 0.0074187759310007095, + 0.004998891148716211, + 0.007014330010861158, + 0.02586151659488678, + 0.006911260075867176, + 0.014879019930958748, + 0.016199396923184395, + 0.0034948985558003187, + 0.0033993427641689777, + 0.003926766570657492, + 0.005008082836866379, + 0.015634668990969658, + 0.004903040360659361, + 0.008966929279267788, + 0.007540175691246986, + 0.0032260522712022066, + 0.005004849750548601, + 0.004498602356761694, + 0.00978589616715908, + 0.012492887675762177, + 0.004230395425111055, + 0.008259495720267296, + 0.006146816536784172, + 0.0043112738057971, + 0.004731707740575075, + 0.004769753664731979, + 0.006832574028521776, + 0.011986777186393738, + 0.006412393879145384, + 0.017997102811932564, + 0.010634257458150387, + 0.0017717907903715968, + 0.002497300738468766, + 0.00262096943333745, + 0.002319532446563244, + 0.007359109353274107, + 0.0032119525130838156, + 0.002928768051788211, + 0.003791367169469595, + 0.00177229312248528, + 0.002730827545747161, + 0.0029213668312877417, + 0.004210269544273615, + 0.006996167358011007, + 0.003879472380504012, + 0.004007580690085888, + 0.0037868537474423647, + 0.0023029332514852285, + 0.00334570137783885, + 0.0037667332217097282, + 0.003782353363931179, + 0.012404526583850384, + 0.003488747403025627, + 0.00755400862544775, + 0.006844100076705217, + 0.001989554613828659, + 0.0035623402800410986, + 0.002266600960865617, + 0.002602411899715662, + 0.00853576697409153, + 0.003305508056655526, + 0.004653455223888159, + 0.004160561598837376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.07896961271762848 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-haired grandfather seated in a dark brown wooden rocking chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a soft beige wool cardigan over a light-colored shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the grandfather perform the action of coughing with a hand pressed to his mouth while his shoulders tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible dancing and spinning in a golden beam of sunlight streaming through large side windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a high-frame-rate horizontal medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043070822954177856 + ], + "opposite_logprob_tokens": [ + -3.1680707931518555 + ], + "logprob_reward": -0.043070822954177856, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043070822954177856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34008538722991943, + 0.146430104970932, + 0.13212649524211884, + 0.10194972902536392, + 0.050087980926036835, + 0.0527578629553318, + 0.06402047723531723, + 0.04887980967760086, + 0.036359068006277084, + 0.027303067967295647 + ], + "credit_map": [ + 0.025997355580329895, + 0.022985072806477547, + 0.03536423668265343, + 0.029213543981313705, + 0.0555989146232605, + 0.05429809167981148, + 0.068144291639328, + 0.04848390445113182, + 0.01677798666059971, + 0.013343204744160175, + 0.014294116757810116, + 0.00914673786610365, + 0.031507350504398346, + 0.022002235054969788, + 0.016161376610398293, + 0.023197103291749954, + 0.008738832548260689, + 0.008588666096329689, + 0.012090733274817467, + 0.014194892719388008, + 0.03728649020195007, + 0.01376536674797535, + 0.019772972911596298, + 0.017688535153865814, + 0.0059749651700258255, + 0.009628510102629662, + 0.008348041214048862, + 0.017981819808483124, + 0.028521757572889328, + 0.006615315563976765, + 0.01429825834929943, + 0.010581071488559246, + 0.003676875028759241, + 0.004238573834300041, + 0.0038133941125124693, + 0.007837419398128986, + 0.012599957175552845, + 0.005373342428356409, + 0.006998617667704821, + 0.005549801047891378, + 0.004404143430292606, + 0.005368353333324194, + 0.003975184168666601, + 0.0047407569363713264, + 0.00965155940502882, + 0.010602359659969807, + 0.00623310636729002, + 0.007782398723065853, + 0.006404837127774954, + 0.008682236075401306, + 0.0055977534502744675, + 0.005459723994135857, + 0.011991769075393677, + 0.010399429127573967, + 0.006956220604479313, + 0.008528504520654678, + 0.005019018426537514, + 0.01003992184996605, + 0.003107200376689434, + 0.004186953417956829, + 0.0104142464697361, + 0.005587016232311726, + 0.004702730569988489, + 0.005822724662721157, + 0.003159372601658106, + 0.0037549161352217197, + 0.004508808720856905, + 0.002754960674792528, + 0.006072512827813625, + 0.006029849406331778, + 0.005873890593647957, + 0.004204757511615753, + 0.0025397171266376972, + 0.0026055125053972006, + 0.0032503074035048485, + 0.0027391284238547087, + 0.005768588278442621, + 0.0033232688438147306, + 0.0034345577005296946, + 0.0036419874522835016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.043070822954177856 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-haired grandfather seated in a dark brown wooden rocking chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a soft beige wool cardigan over a light-colored shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the grandfather perform the action of coughing with a hand pressed to his mouth while his shoulders tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible dancing and spinning in a golden beam of sunlight streaming through large side windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a high-frame-rate horizontal medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004482299089432 + ], + "opposite_logprob_tokens": [ + -2.695044755935669 + ], + "logprob_reward": -0.07004482299089432, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004482299089432, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22362808883190155, + 0.131084144115448, + 0.41923555731773376, + 0.055738747119903564, + 0.029270755127072334, + 0.032499950379133224, + 0.04663505032658577, + 0.021233217790722847, + 0.02610764279961586, + 0.014566836878657341 + ], + "credit_map": [ + 0.019403288140892982, + 0.02028820663690567, + 0.01868535950779915, + 0.019934216514229774, + 0.06247290223836899, + 0.027073336765170097, + 0.02559034898877144, + 0.0301804356276989, + 0.010768100619316101, + 0.012749938294291496, + 0.014909614808857441, + 0.013598911464214325, + 0.026819851249456406, + 0.012907066382467747, + 0.021292991936206818, + 0.01803765445947647, + 0.025791110470891, + 0.03395099565386772, + 0.022827019914984703, + 0.03349636122584343, + 0.12591774761676788, + 0.03700832277536392, + 0.06761243939399719, + 0.07263153046369553, + 0.00389820896089077, + 0.004715082701295614, + 0.004039891064167023, + 0.004125171806663275, + 0.016996238380670547, + 0.005316951312124729, + 0.008127032779157162, + 0.008520172908902168, + 0.0019622270483523607, + 0.002316566649824381, + 0.002633508061990142, + 0.0035720868036150932, + 0.005729160737246275, + 0.0024317363277077675, + 0.006179040297865868, + 0.004446430131793022, + 0.002576112048700452, + 0.0026195046957582235, + 0.0031703258864581585, + 0.0021279659122228622, + 0.00714979600161314, + 0.0028596806805580854, + 0.007053987588733435, + 0.004942579660564661, + 0.003983814734965563, + 0.0044014030136168, + 0.0046640560030937195, + 0.00467884074896574, + 0.009479455649852753, + 0.005633217748254538, + 0.0065983934327960014, + 0.007195872254669666, + 0.002336813835427165, + 0.0027371360920369625, + 0.002048647729679942, + 0.0020879539661109447, + 0.0034799540881067514, + 0.0025147199630737305, + 0.0028535209130495787, + 0.0031744714360684156, + 0.001789767644368112, + 0.001737553859129548, + 0.002304202178493142, + 0.0021969189401715994, + 0.00769017543643713, + 0.0026076529175043106, + 0.003903026692569256, + 0.0038783433847129345, + 0.001170616247691214, + 0.0014745539519935846, + 0.0013433368876576424, + 0.0012273119064047933, + 0.0036676626186817884, + 0.0015618763864040375, + 0.0018957598367705941, + 0.002225718693807721 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.07004482299089432 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-frame-rate, horizontal medium shot gently draws the viewer's attention to a white-haired grandfather, seated peacefully in a dark brown wooden rocking chair. Warm afternoon sunlight, like molten honey, streams through the large side windows, illuminating the cozy, bookish room. The grandfather, with a head full of silver hair and a face marked by the kindly wrinkles of time, wears a soft beige wool cardigan over a light-colored shirt. He hunches his back slightly, pressing a dry, age-freckled hand to his mouth as he lets out a soft, muffled cough. The high-frame-rate lens plays an extremely effective role in this moment: his thin shoulders tremble slightly but clearly, and every muscle contraction is slowed down into poetic detail; at the same time, in the golden beam of light that penetrates the room, countless tiny dust particles that were originally invisible to the naked eye are captured clearly. They are like disturbed golden elves, dancing, spinning and flickering slowly and gracefully in the air, forming a contrast between movement and stillness, grandeur and microcosm with the slight tremor of the old man's body. The whole picture is filled with a peaceful, nostalgic and warm atmosphere with a hint of the fragility of time under the ultimate clarity.", + "seed": 31391, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_10/index_2761_seed_31391_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05604009050875902, + "vlm_logprob": -0.05604009050875902, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.05604009050875902, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-haired grandfather seated in a dark brown wooden rocking chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a soft beige wool cardigan over a light-colored shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the grandfather perform the action of coughing with a hand pressed to his mouth while his shoulders tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible dancing and spinning in a golden beam of sunlight streaming through large side windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a high-frame-rate horizontal medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02792775072157383 + ], + "opposite_logprob_tokens": [ + -3.7779276371002197 + ], + "logprob_reward": -0.02792775072157383, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02792775072157383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4307050406932831, + 0.03975161910057068, + 0.06226036697626114, + 0.07385975867509842, + 0.05761684849858284, + 0.03537127375602722, + 0.06554407626390457, + 0.09175267070531845, + 0.0361977145075798, + 0.10694072395563126 + ], + "credit_map": [ + 0.05689577758312225, + 0.07520191371440887, + 0.07091101258993149, + 0.07358313351869583, + 0.03750405088067055, + 0.026991579681634903, + 0.060442857444286346, + 0.029174679890275, + 0.004531709011644125, + 0.006819780915975571, + 0.004808367229998112, + 0.007103356532752514, + 0.00521561224013567, + 0.0026269692461937666, + 0.005296028219163418, + 0.003349791746586561, + 0.005717495921999216, + 0.008701416663825512, + 0.0071312240324914455, + 0.00765606202185154, + 0.00910766888409853, + 0.010620334185659885, + 0.00728997029364109, + 0.006036188453435898, + 0.0063176448456943035, + 0.014347493648529053, + 0.00920806359499693, + 0.010933906771242619, + 0.010288381017744541, + 0.005760048050433397, + 0.01014938298612833, + 0.006854833569377661, + 0.0059946393594145775, + 0.005793943535536528, + 0.011743558570742607, + 0.006341500207781792, + 0.012532249093055725, + 0.004118063487112522, + 0.0072010294534265995, + 0.0038918566424399614, + 0.0030985292978584766, + 0.004355359822511673, + 0.004814509302377701, + 0.006466640625149012, + 0.00344120804220438, + 0.002783567411825061, + 0.006404474377632141, + 0.0040069809183478355, + 0.00505564920604229, + 0.008863403461873531, + 0.008613375946879387, + 0.006175701506435871, + 0.01652686297893524, + 0.008771644905209541, + 0.005875056609511375, + 0.005662374198436737, + 0.008199231699109077, + 0.014236886985599995, + 0.011178984306752682, + 0.011938563548028469, + 0.016611343249678612, + 0.009325260296463966, + 0.010903069749474525, + 0.009359322488307953, + 0.0029842176008969545, + 0.004331469535827637, + 0.00452829385176301, + 0.005279561970382929, + 0.006019299849867821, + 0.004510986618697643, + 0.0032597833778709173, + 0.005284098908305168, + 0.008811980485916138, + 0.01522739976644516, + 0.01076798141002655, + 0.014666019007563591, + 0.017157185822725296, + 0.010366958566009998, + 0.015387749299407005, + 0.01455544214695692 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.02792775072157383 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-haired grandfather seated in a dark brown wooden rocking chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a soft beige wool cardigan over a light-colored shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the grandfather perform the action of coughing with a hand pressed to his mouth while his shoulders tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible dancing and spinning in a golden beam of sunlight streaming through large side windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a high-frame-rate horizontal medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014216690324246883 + ], + "opposite_logprob_tokens": [ + -4.264216899871826 + ], + "logprob_reward": -0.014216690324246883, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014216690324246883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2999933063983917, + 0.08537288010120392, + 0.05174766853451729, + 0.03824526071548462, + 0.05201296508312225, + 0.03328406810760498, + 0.06532757729291916, + 0.1350407600402832, + 0.053653400391340256, + 0.185322105884552 + ], + "credit_map": [ + 0.03337535634636879, + 0.04491933435201645, + 0.040720488876104355, + 0.05212239548563957, + 0.03380332887172699, + 0.019422639161348343, + 0.04873151332139969, + 0.026898229494690895, + 0.008473960682749748, + 0.012579254806041718, + 0.010726623237133026, + 0.01351148821413517, + 0.01172090508043766, + 0.008056998252868652, + 0.011907362379133701, + 0.008396278135478497, + 0.0050932723097503185, + 0.008972618728876114, + 0.005742656998336315, + 0.007258772850036621, + 0.007608029525727034, + 0.00784192606806755, + 0.004338737111538649, + 0.004891650285571814, + 0.0038484544493258, + 0.007213136181235313, + 0.005297005642205477, + 0.004480566829442978, + 0.006036281120032072, + 0.0030585145577788353, + 0.004994853399693966, + 0.003316444344818592, + 0.004463239572942257, + 0.007490949239581823, + 0.009343363344669342, + 0.0088877584785223, + 0.008971108123660088, + 0.0034888694062829018, + 0.005878771189600229, + 0.0034889024682343006, + 0.003632179694250226, + 0.0042112660594284534, + 0.004718692973256111, + 0.006108265370130539, + 0.004248784389346838, + 0.0027123705949634314, + 0.0044455090537667274, + 0.0032069983426481485, + 0.004836253356188536, + 0.007738399785012007, + 0.008909027092158794, + 0.006701501086354256, + 0.019772442057728767, + 0.006936284247785807, + 0.0051286728121340275, + 0.005304989870637655, + 0.012576112523674965, + 0.02170618809759617, + 0.015649165958166122, + 0.018203550949692726, + 0.022970683872699738, + 0.01268073171377182, + 0.017070725560188293, + 0.01418357901275158, + 0.004282431676983833, + 0.009446153417229652, + 0.0054357717745006084, + 0.007150407414883375, + 0.011117775924503803, + 0.004876567050814629, + 0.006836727261543274, + 0.004507559351623058, + 0.01472405344247818, + 0.031456466764211655, + 0.021320637315511703, + 0.030944425612688065, + 0.023932917043566704, + 0.01502762921154499, + 0.028022967278957367, + 0.01989300362765789 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.014216690324246883 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-haired grandfather seated in a dark brown wooden rocking chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a soft beige wool cardigan over a light-colored shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the grandfather perform the action of coughing with a hand pressed to his mouth while his shoulders tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible dancing and spinning in a golden beam of sunlight streaming through large side windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a high-frame-rate horizontal medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789749026298523 + ], + "opposite_logprob_tokens": [ + -2.578974962234497 + ], + "logprob_reward": -0.0789749026298523, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789749026298523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2701234221458435, + 0.03264534845948219, + 0.05313187465071678, + 0.034525204449892044, + 0.2649746239185333, + 0.04359447956085205, + 0.06236649304628372, + 0.08760092407464981, + 0.03801513463258743, + 0.11302243918180466 + ], + "credit_map": [ + 0.027329277247190475, + 0.05421486124396324, + 0.03242117539048195, + 0.04584059491753578, + 0.030607592314481735, + 0.01842399686574936, + 0.03378179669380188, + 0.027504142373800278, + 0.0029072416946291924, + 0.005091405939310789, + 0.0039264545775949955, + 0.00534436572343111, + 0.004618161357939243, + 0.003003971418365836, + 0.004795756656676531, + 0.0029579901602119207, + 0.005033199209719896, + 0.00806739367544651, + 0.007134688086807728, + 0.0073828608728945255, + 0.006764673162251711, + 0.006200778763741255, + 0.00784700084477663, + 0.004701283760368824, + 0.0028354572132229805, + 0.004329127725213766, + 0.004577833693474531, + 0.004683006089180708, + 0.006534972693771124, + 0.0028403878677636385, + 0.005399174522608519, + 0.0033252458088099957, + 0.017482638359069824, + 0.014476219192147255, + 0.06269004940986633, + 0.030819768086075783, + 0.08152319490909576, + 0.0189223475754261, + 0.026743894442915916, + 0.012316553853452206, + 0.0028285279404371977, + 0.004310378804802895, + 0.008471314795315266, + 0.004967234563082457, + 0.012977595441043377, + 0.00335404253564775, + 0.003683751914650202, + 0.0030016370583325624, + 0.0036494655068963766, + 0.006643055938184261, + 0.008839459158480167, + 0.012034472078084946, + 0.01618931069970131, + 0.005508909933269024, + 0.005450508091598749, + 0.004051310010254383, + 0.00886162556707859, + 0.013298310339450836, + 0.013373120687901974, + 0.017657605931162834, + 0.009314698167145252, + 0.0054929512552917, + 0.011166640557348728, + 0.008435972966253757, + 0.0029924598056823015, + 0.0047104754485189915, + 0.005741894710808992, + 0.006274820771068335, + 0.005262309685349464, + 0.003952882252633572, + 0.005221907049417496, + 0.0038583821151405573, + 0.008074265904724598, + 0.0169204194098711, + 0.012210399843752384, + 0.019924907013773918, + 0.01606353372335434, + 0.01236595120280981, + 0.018077313899993896, + 0.009385648183524609 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.0789749026298523 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-haired grandfather seated in a dark brown wooden rocking chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a soft beige wool cardigan over a light-colored shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the grandfather perform the action of coughing with a hand pressed to his mouth while his shoulders tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible dancing and spinning in a golden beam of sunlight streaming through large side windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a high-frame-rate horizontal medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905462175607681 + ], + "opposite_logprob_tokens": [ + -2.464054584503174 + ], + "logprob_reward": -0.08905462175607681, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905462175607681, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19221773743629456, + 0.047448281198740005, + 0.026203319430351257, + 0.03163604810833931, + 0.3340167999267578, + 0.031203433871269226, + 0.04746413230895996, + 0.09056949615478516, + 0.07960238307714462, + 0.1196383461356163 + ], + "credit_map": [ + 0.015829335898160934, + 0.03295504301786423, + 0.023361317813396454, + 0.04040582850575447, + 0.024794917553663254, + 0.014741083607077599, + 0.02492186427116394, + 0.015208344906568527, + 0.004005904775112867, + 0.00858220737427473, + 0.005579141899943352, + 0.0069883414544165134, + 0.010174386203289032, + 0.0031491995323449373, + 0.005162122193723917, + 0.0038069789297878742, + 0.0020954927895218134, + 0.004601834341883659, + 0.0034013560507446527, + 0.003844168735668063, + 0.0036749355494976044, + 0.003317206632345915, + 0.0023843133822083473, + 0.002884011482819915, + 0.002417207695543766, + 0.005554347299039364, + 0.0037405036855489016, + 0.0043274718336761, + 0.006319023203104734, + 0.002703705569729209, + 0.003935177810490131, + 0.0026386098470538855, + 0.02382507734000683, + 0.018640853464603424, + 0.08779193460941315, + 0.027797507122159004, + 0.11293543130159378, + 0.02630806528031826, + 0.021776992827653885, + 0.014940978027880192, + 0.002545050112530589, + 0.0032412575092166662, + 0.005486875306814909, + 0.0027479128912091255, + 0.008617181330919266, + 0.0031305449083447456, + 0.0021218478213995695, + 0.003312764223664999, + 0.003913072403520346, + 0.0050549739971756935, + 0.005962089169770479, + 0.0046661668457090855, + 0.009376396425068378, + 0.008513589389622211, + 0.004593977704644203, + 0.005383861716836691, + 0.006582531612366438, + 0.009300397709012032, + 0.011000462807714939, + 0.012805987149477005, + 0.0190583523362875, + 0.012079037725925446, + 0.009952479973435402, + 0.009790243580937386, + 0.0054389904253184795, + 0.009366207756102085, + 0.009997106157243252, + 0.007763259578496218, + 0.022860433906316757, + 0.010326554998755455, + 0.006378805730491877, + 0.007471022196114063, + 0.009811271913349628, + 0.010162992402911186, + 0.014289457350969315, + 0.012146973982453346, + 0.01738426834344864, + 0.01826828345656395, + 0.01187816634774208, + 0.025696927681565285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.08905462175607681 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-haired grandfather seated in a dark brown wooden rocking chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a soft beige wool cardigan over a light-colored shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the grandfather perform the action of coughing with a hand pressed to his mouth while his shoulders tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible dancing and spinning in a golden beam of sunlight streaming through large side windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a high-frame-rate horizontal medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002648711204529 + ], + "opposite_logprob_tokens": [ + -2.695026397705078 + ], + "logprob_reward": -0.07002648711204529, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002648711204529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1574930101633072, + 0.035102494060993195, + 0.017907436937093735, + 0.03417941927909851, + 0.41992050409317017, + 0.03777403384447098, + 0.0491638146340847, + 0.10527019947767258, + 0.11819998174905777, + 0.024989202618598938 + ], + "credit_map": [ + 0.018976125866174698, + 0.031377993524074554, + 0.01965564861893654, + 0.023965369910001755, + 0.01584957353770733, + 0.013064284808933735, + 0.020769823342561722, + 0.013834184966981411, + 0.003534862305969, + 0.0072193443775177, + 0.004499819129705429, + 0.0047530969604849815, + 0.0041319639421999454, + 0.0029400333296507597, + 0.004734680987894535, + 0.0032886932604014874, + 0.0020282287150621414, + 0.0026231687515974045, + 0.0027202439960092306, + 0.00233229692094028, + 0.0028762011788785458, + 0.0018214454175904393, + 0.0017946988809853792, + 0.001711149699985981, + 0.002752832602709532, + 0.006392287090420723, + 0.005389686208218336, + 0.0061463480815291405, + 0.004479305353015661, + 0.0025184552650898695, + 0.0030767356511205435, + 0.0034237669315189123, + 0.02751161716878414, + 0.023609522730112076, + 0.10005994141101837, + 0.05002569779753685, + 0.13054749369621277, + 0.029863612726330757, + 0.03854048252105713, + 0.01976209506392479, + 0.00254046986810863, + 0.0034606289118528366, + 0.00876672938466072, + 0.0033309615682810545, + 0.01173493918031454, + 0.002867621835321188, + 0.0028742742724716663, + 0.002198402537032962, + 0.004217954818159342, + 0.007292838301509619, + 0.005153615027666092, + 0.007140805944800377, + 0.010418609715998173, + 0.005443014204502106, + 0.005981918890029192, + 0.0035150537732988596, + 0.007792789489030838, + 0.013721792958676815, + 0.012490898370742798, + 0.012666076421737671, + 0.019257018342614174, + 0.014727852307260036, + 0.012918991036713123, + 0.011694767512381077, + 0.008634417317807674, + 0.01591457985341549, + 0.014063898473978043, + 0.011506910435855389, + 0.022766726091504097, + 0.02072986587882042, + 0.011072476394474506, + 0.013511085882782936, + 0.0022201458923518658, + 0.0026591806672513485, + 0.0033636700827628374, + 0.003426579525694251, + 0.0036970844957977533, + 0.004054195247590542, + 0.002672593342140317, + 0.0028957503382116556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.07002648711204529 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-frame-rate, horizontal medium shot gently draws the viewer's attention to a white-haired grandfather, seated peacefully in a dark brown wooden rocking chair. Warm afternoon sunlight, like molten honey, streams through the large side windows, illuminating the cozy, bookish room. The grandfather, with a head full of silver hair and a face marked by the kindly wrinkles of time, wears a soft beige wool cardigan over a light-colored shirt. He hunches his back slightly, pressing a dry, age-freckled hand to his mouth as he lets out a soft, muffled cough. The high-frame-rate lens plays an extremely effective role in this moment: his thin shoulders tremble slightly but clearly, and every muscle contraction is slowed down into poetic detail; at the same time, in the golden beam of light that penetrates the room, countless tiny dust particles that were originally invisible to the naked eye are captured clearly. They are like disturbed golden elves, dancing, spinning and flickering slowly and gracefully in the air, forming a contrast between movement and stillness, grandeur and microcosm with the slight tremor of the old man's body. The whole picture is filled with a peaceful, nostalgic and warm atmosphere with a hint of the fragility of time under the ultimate clarity.", + "seed": 31392, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_10/index_2761_seed_31392_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06130531951785088, + "vlm_logprob": -0.06130531951785088, + "vlm_yes_no_margin": 3.074999976158142, + "vlm_token_credit": -0.06130531951785088, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-haired grandfather seated in a dark brown wooden rocking chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a soft beige wool cardigan over a light-colored shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the grandfather perform the action of coughing with a hand pressed to his mouth while his shoulders tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible dancing and spinning in a golden beam of sunlight streaming through large side windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a high-frame-rate horizontal medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017825404182076454 + ], + "opposite_logprob_tokens": [ + -4.267825603485107 + ], + "logprob_reward": -0.017825404182076454, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.017825404182076454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3074803650379181, + 0.12422377616167068, + 0.11947080492973328, + 0.1081506684422493, + 0.07324901223182678, + 0.06129651516675949, + 0.03693518415093422, + 0.10204795002937317, + 0.031578585505485535, + 0.035567138344049454 + ], + "credit_map": [ + 0.050592027604579926, + 0.03718852251768112, + 0.036279596388339996, + 0.029304515570402145, + 0.023960055783391, + 0.03052213415503502, + 0.042915455996990204, + 0.056718070060014725, + 0.018621256574988365, + 0.023769835010170937, + 0.015709685161709785, + 0.015395360998809338, + 0.017081502825021744, + 0.01142736617475748, + 0.01450272835791111, + 0.007716034073382616, + 0.02543952874839306, + 0.01332490611821413, + 0.022335214540362358, + 0.014225298538804054, + 0.015811139717698097, + 0.008893853984773159, + 0.013664940372109413, + 0.00577591871842742, + 0.012565318495035172, + 0.013481044210493565, + 0.012610542587935925, + 0.017229950055480003, + 0.00968224834650755, + 0.007334407418966293, + 0.025000909343361855, + 0.010246249847114086, + 0.020600438117980957, + 0.007127552758902311, + 0.01448361948132515, + 0.006925085559487343, + 0.006716961972415447, + 0.006368334870785475, + 0.006994643248617649, + 0.004032377153635025, + 0.012226242572069168, + 0.005949461832642555, + 0.008531743660569191, + 0.006949468981474638, + 0.00655650207772851, + 0.007052548695355654, + 0.007142832037061453, + 0.006887717638164759, + 0.006034009624272585, + 0.004502429161220789, + 0.00632536131888628, + 0.004292559344321489, + 0.004258777946233749, + 0.002463576616719365, + 0.005479672458022833, + 0.0035787976812571287, + 0.014051925390958786, + 0.02434069849550724, + 0.011738755740225315, + 0.011000428348779678, + 0.01216845028102398, + 0.00815118569880724, + 0.008467587642371655, + 0.012128912843763828, + 0.0050703720189630985, + 0.006215947680175304, + 0.004261687397956848, + 0.004202262498438358, + 0.00392336118966341, + 0.0017027304274961352, + 0.0037984505761414766, + 0.002403774531558156, + 0.008042478933930397, + 0.005741616245359182, + 0.005733955651521683, + 0.003792478935793042, + 0.004248894285410643, + 0.0020247530192136765, + 0.003991211764514446, + 0.001991753000766039 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.017825404182076454 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-haired grandfather seated in a dark brown wooden rocking chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a soft beige wool cardigan over a light-colored shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the grandfather perform the action of coughing with a hand pressed to his mouth while his shoulders tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible dancing and spinning in a golden beam of sunlight streaming through large side windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a high-frame-rate horizontal medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060099132359028 + ], + "opposite_logprob_tokens": [ + -3.8956010341644287 + ], + "logprob_reward": -0.02060099132359028, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060099132359028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2204442024230957, + 0.31498730182647705, + 0.12750692665576935, + 0.04903484135866165, + 0.04569879546761513, + 0.0781913548707962, + 0.041319865733385086, + 0.05658011883497238, + 0.044340066611766815, + 0.02189655788242817 + ], + "credit_map": [ + 0.030731936916708946, + 0.023175330832600594, + 0.020342668518424034, + 0.024595243856310844, + 0.02509649097919464, + 0.019215639680624008, + 0.03979449346661568, + 0.03749240189790726, + 0.039909958839416504, + 0.071327805519104, + 0.03801492229104042, + 0.027245406061410904, + 0.05076613649725914, + 0.029418669641017914, + 0.038488034158945084, + 0.019816359505057335, + 0.0217086561024189, + 0.01192379742860794, + 0.020151259377598763, + 0.017452500760555267, + 0.016122225672006607, + 0.013024652376770973, + 0.01962500438094139, + 0.0074988361448049545, + 0.008281363174319267, + 0.008054529316723347, + 0.006455122958868742, + 0.007457990199327469, + 0.004219050984829664, + 0.003363746451213956, + 0.0062696547247469425, + 0.004933384712785482, + 0.00895044393837452, + 0.004624826833605766, + 0.007545111235231161, + 0.0069565842859447, + 0.004667014814913273, + 0.003414108185097575, + 0.006769821047782898, + 0.0027708846610039473, + 0.010898761451244354, + 0.010017646476626396, + 0.0091701690107584, + 0.007945414632558823, + 0.007973798550665379, + 0.006900918669998646, + 0.011033264920115471, + 0.01425138209015131, + 0.006295413710176945, + 0.00621167803183198, + 0.006059394683688879, + 0.004230407066643238, + 0.004945746157318354, + 0.00316254491917789, + 0.005602954421192408, + 0.004811728373169899, + 0.01056740153580904, + 0.009595743380486965, + 0.007592122070491314, + 0.006307418458163738, + 0.006884625181555748, + 0.004883578047156334, + 0.007004380691796541, + 0.003744850168004632, + 0.007676164153963327, + 0.009676466695964336, + 0.006063570734113455, + 0.005826003383845091, + 0.004322809167206287, + 0.002664235420525074, + 0.005088151432573795, + 0.0030226660892367363, + 0.003318990347906947, + 0.004204487893730402, + 0.002844695933163166, + 0.0029246758203953505, + 0.0026902267709374428, + 0.0013221566332504153, + 0.0029244213365018368, + 0.0016669038450345397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.02060099132359028 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-haired grandfather seated in a dark brown wooden rocking chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a soft beige wool cardigan over a light-colored shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the grandfather perform the action of coughing with a hand pressed to his mouth while his shoulders tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible dancing and spinning in a golden beam of sunlight streaming through large side windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a high-frame-rate horizontal medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789678543806076 + ], + "opposite_logprob_tokens": [ + -2.578967809677124 + ], + "logprob_reward": -0.0789678543806076, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789678543806076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20326265692710876, + 0.406762957572937, + 0.042562659829854965, + 0.048491474241018295, + 0.04116243124008179, + 0.0527929812669754, + 0.025339048355817795, + 0.09909531474113464, + 0.06022065505385399, + 0.02030976302921772 + ], + "credit_map": [ + 0.03343074396252632, + 0.023145711049437523, + 0.026597492396831512, + 0.021846365183591843, + 0.019837701693177223, + 0.02627621777355671, + 0.025032740086317062, + 0.02709571085870266, + 0.05914287269115448, + 0.09345928579568863, + 0.04707951843738556, + 0.0396510548889637, + 0.06077134609222412, + 0.03383686766028404, + 0.042980387806892395, + 0.029841654002666473, + 0.008744755759835243, + 0.006728721782565117, + 0.006566001567989588, + 0.004766070283949375, + 0.0057042851112782955, + 0.0030354848131537437, + 0.004283626098185778, + 0.002733715809881687, + 0.009668761864304543, + 0.006261042784899473, + 0.006434036418795586, + 0.005414078943431377, + 0.005559274461120367, + 0.0031112772412598133, + 0.00689069414511323, + 0.00515231117606163, + 0.008034944534301758, + 0.0056227114982903, + 0.006479012314230204, + 0.004139549098908901, + 0.0031218768563121557, + 0.00325257470831275, + 0.004972415044903755, + 0.005539349280297756, + 0.00873300526291132, + 0.005037663038820028, + 0.006991602946072817, + 0.005151230376213789, + 0.0059283822774887085, + 0.004126638639718294, + 0.008560498245060444, + 0.008263959549367428, + 0.0037206325214356184, + 0.002675564493983984, + 0.0032455488108098507, + 0.0030524602625519037, + 0.0020706518553197384, + 0.0031455366406589746, + 0.0038625055458396673, + 0.0035661489237099886, + 0.013599625788629055, + 0.019637133926153183, + 0.010546253062784672, + 0.015915637835860252, + 0.010255161672830582, + 0.007524400018155575, + 0.01142420805990696, + 0.010192902758717537, + 0.008164502680301666, + 0.010955245234072208, + 0.007855319418013096, + 0.008001486770808697, + 0.00880223885178566, + 0.005093801766633987, + 0.005819614045321941, + 0.005528449080884457, + 0.0034886987414211035, + 0.0034234635531902313, + 0.0022521214559674263, + 0.002923529827967286, + 0.003161609172821045, + 0.001192969037219882, + 0.002844704780727625, + 0.0010226666927337646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.0789678543806076 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-haired grandfather seated in a dark brown wooden rocking chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a soft beige wool cardigan over a light-colored shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the grandfather perform the action of coughing with a hand pressed to his mouth while his shoulders tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible dancing and spinning in a golden beam of sunlight streaming through large side windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a high-frame-rate horizontal medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12707409262657166 + ], + "opposite_logprob_tokens": [ + -2.1270740032196045 + ], + "logprob_reward": -0.12707409262657166, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12707409262657166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10763128846883774, + 0.320060133934021, + 0.1246253103017807, + 0.06789108365774155, + 0.10793328285217285, + 0.1003703698515892, + 0.060265567153692245, + 0.03471827879548073, + 0.020856274291872978, + 0.05564841628074646 + ], + "credit_map": [ + 0.016318123787641525, + 0.01471271924674511, + 0.013794281519949436, + 0.015407069586217403, + 0.012812715955078602, + 0.007710512261837721, + 0.01852153241634369, + 0.008354335092008114, + 0.052993301302194595, + 0.07326965779066086, + 0.038449279963970184, + 0.03136056661605835, + 0.04561708867549896, + 0.027394914999604225, + 0.02669226936995983, + 0.024283038452267647, + 0.03080901876091957, + 0.01585877500474453, + 0.02427966520190239, + 0.008895608596503735, + 0.011423792690038681, + 0.010567327961325645, + 0.011704633012413979, + 0.011086488142609596, + 0.009687704034149647, + 0.01340534444898367, + 0.007822857238352299, + 0.00802789255976677, + 0.006688260938972235, + 0.0055739437229931355, + 0.006959210615605116, + 0.009725860320031643, + 0.03261682391166687, + 0.013721796683967113, + 0.022065142169594765, + 0.007026864681392908, + 0.007323502097278833, + 0.009002015925943851, + 0.009728264063596725, + 0.0064488667994737625, + 0.0205224696546793, + 0.011420929804444313, + 0.014660872519016266, + 0.013149804435670376, + 0.010538368485867977, + 0.008541901595890522, + 0.011865010485053062, + 0.009671000763773918, + 0.01152641698718071, + 0.008298578672111034, + 0.01188906840980053, + 0.0060472646728158, + 0.00553865684196353, + 0.006318842060863972, + 0.0066525135189294815, + 0.003994225058704615, + 0.006581212859600782, + 0.005636079702526331, + 0.004900546744465828, + 0.004134014714509249, + 0.004030733369290829, + 0.0031302915886044502, + 0.003671943908557296, + 0.0026334498543292284, + 0.0037649197038263083, + 0.0036786373239010572, + 0.0031609677243977785, + 0.0021284697577357292, + 0.002611549338325858, + 0.0015297962818294764, + 0.0022643019910901785, + 0.0017176319379359484, + 0.013148288242518902, + 0.0074278321117162704, + 0.009629630483686924, + 0.005617647897452116, + 0.006303544156253338, + 0.0038882216904312372, + 0.006227925885468721, + 0.0034053227864205837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.12707409262657166 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-haired grandfather seated in a dark brown wooden rocking chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a soft beige wool cardigan over a light-colored shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the grandfather perform the action of coughing with a hand pressed to his mouth while his shoulders tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible dancing and spinning in a golden beam of sunlight streaming through large side windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a high-frame-rate horizontal medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062058255076408386 + ], + "opposite_logprob_tokens": [ + -2.812058210372925 + ], + "logprob_reward": -0.062058255076408386, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062058255076408386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21537712216377258, + 0.24592553079128265, + 0.05756557360291481, + 0.08022582530975342, + 0.09814096242189407, + 0.10290693491697311, + 0.04252280294895172, + 0.08895618468523026, + 0.0285409577190876, + 0.039838168770074844 + ], + "credit_map": [ + 0.029420793056488037, + 0.025114919990301132, + 0.023355809971690178, + 0.021367954090237617, + 0.020763276144862175, + 0.024043943732976913, + 0.033904850482940674, + 0.03740553557872772, + 0.03894635662436485, + 0.05772306025028229, + 0.027277419343590736, + 0.031689662486314774, + 0.03401622176170349, + 0.01612946204841137, + 0.02214874140918255, + 0.017994578927755356, + 0.011301436461508274, + 0.009062706492841244, + 0.007770709227770567, + 0.006524954456835985, + 0.0072572799399495125, + 0.0036854534409940243, + 0.006145321764051914, + 0.005817706696689129, + 0.011263915337622166, + 0.009979289956390858, + 0.009348020888864994, + 0.009905158542096615, + 0.006758985575288534, + 0.008198652416467667, + 0.016479047015309334, + 0.008292751386761665, + 0.027483602985739708, + 0.010974609293043613, + 0.021385421976447105, + 0.006763725075870752, + 0.00994777213782072, + 0.007026588544249535, + 0.009185194037854671, + 0.005374034401029348, + 0.03137046843767166, + 0.0107051907107234, + 0.018046313896775246, + 0.009519543498754501, + 0.007019859738647938, + 0.00925465114414692, + 0.009201939217746258, + 0.0077889529056847095, + 0.008134257048368454, + 0.004224533215165138, + 0.0055094133131206036, + 0.005123299080878496, + 0.003999075852334499, + 0.004660292062908411, + 0.007116780150681734, + 0.003755145240575075, + 0.012177247554063797, + 0.016234613955020905, + 0.010097622871398926, + 0.011707297526299953, + 0.008424586616456509, + 0.009898108430206776, + 0.011677674017846584, + 0.008739026263356209, + 0.004307039547711611, + 0.004103575833141804, + 0.004552383907139301, + 0.0035486856941133738, + 0.0033571079839020967, + 0.001996890641748905, + 0.004449514206498861, + 0.002225757110863924, + 0.010016004554927349, + 0.006044853013008833, + 0.007239021826535463, + 0.003677078289911151, + 0.003946315962821245, + 0.002507551573216915, + 0.003827809588983655, + 0.0025795316323637962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.062058255076408386 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_11/index_tensor([3169, 3169])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_11/index_tensor([3169, 3169])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..cc6cc8657954e3b8f48013c3aedf99b3532fed89 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_11/index_tensor([3169, 3169])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, slightly high-angle shot allows the camera to smoothly track the precise movements of the 3D printer's nozzle. The nozzle, a polished brass cone attached to a complex assembly of wires and fans, glides delicately just millimeters above the dark, textured print bed. From its tiny, heated tip, a thin, continuous thread of molten, vibrant red plastic steadily extrudes, gleaming under the bright, focused LED light. The nozzle moves with a swift and decisive automation, executing a series of linear and curvilinear paths as it meticulously deposits the material. The lustrous red filament cools and solidifies almost instantly upon contact, fusing with the layers below to form a complex geometric part whose surface displays a fine, striated texture layered in successive layers. The entire scene is bathed in a cool, clinical light, casting sharp highlights across the gleaming surfaces of the metal parts and the newly printed plastic, creating deep shadows and accentuating the object's complex, ever-growing form. The background is a soft, out-of-focus blur, suggesting a workshop environment where the methodical act of creation is the centerpiece, creating an atmosphere of focus, hypnosis, and technical precision.", + "seed": 27038, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_11/index_3169_seed_27038_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.024457338271895425, + "vlm_logprob": -0.024457338271895425, + "vlm_yes_no_margin": 4.875, + "vlm_token_credit": -0.024457338271895425, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a 3D printer's nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle a polished brass cone attached to wires and fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thin, continuous thread of molten vibrant red plastic extruding from the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the nozzle move with swift automation along linear and curvilinear paths over a dark textured print bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright focused LED light creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.049767497926950455 + ], + "opposite_logprob_tokens": [ + -3.17476749420166 + ], + "logprob_reward": -0.049767497926950455, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.049767497926950455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10455112159252167, + 0.17083758115768433, + 0.03005790337920189, + 0.11930567771196365, + 0.11521743983030319, + 0.09733248502016068, + 0.024891575798392296, + 0.05267559736967087, + 0.2243475615978241, + 0.06078312546014786 + ], + "credit_map": [ + 0.015838565304875374, + 0.018270032480359077, + 0.01103244349360466, + 0.015147848054766655, + 0.012815458700060844, + 0.007713311817497015, + 0.0107096116989851, + 0.01302385050803423, + 0.039983268827199936, + 0.0120270736515522, + 0.013437098823487759, + 0.014028523117303848, + 0.008827517740428448, + 0.013425266370177269, + 0.016017591580748558, + 0.05309121310710907, + 0.0026376948226243258, + 0.0026508576702326536, + 0.0034691046457737684, + 0.005648930091410875, + 0.0036546846386045218, + 0.004648405592888594, + 0.0032120198011398315, + 0.00413620425388217, + 0.01246294192969799, + 0.006384743843227625, + 0.008949567563831806, + 0.006896208971738815, + 0.0062879095785319805, + 0.04114390164613724, + 0.009669669903814793, + 0.02751072868704796, + 0.010805950500071049, + 0.016035912558436394, + 0.012247336097061634, + 0.009166034869849682, + 0.011686383746564388, + 0.021997377276420593, + 0.017522010952234268, + 0.01575641706585884, + 0.008196539245545864, + 0.009364999830722809, + 0.010862572118639946, + 0.013822871260344982, + 0.010792603716254234, + 0.009669875726103783, + 0.019734391942620277, + 0.014888623729348183, + 0.0026900016237050295, + 0.0024958972353488207, + 0.003094205167144537, + 0.0031665614806115627, + 0.003061498748138547, + 0.0038186879828572273, + 0.003110957331955433, + 0.0034537629690021276, + 0.0029439551290124655, + 0.002501258160918951, + 0.002329915529116988, + 0.0032220890279859304, + 0.003330637002363801, + 0.014424074441194534, + 0.005288533866405487, + 0.018635131418704987, + 0.02305692620575428, + 0.0068842824548482895, + 0.009504095651209354, + 0.01947956718504429, + 0.01129370741546154, + 0.07905455678701401, + 0.022501708939671516, + 0.05257271230220795, + 0.0057314308360219, + 0.007066486869007349, + 0.008230549283325672, + 0.011517548933625221, + 0.005991141311824322, + 0.006054598838090897, + 0.008067922666668892, + 0.008123443461954594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 3 + ], + "selected_reward": -0.049767497926950455 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a 3D printer's nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle a polished brass cone attached to wires and fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thin, continuous thread of molten vibrant red plastic extruding from the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the nozzle move with swift automation along linear and curvilinear paths over a dark textured print bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright focused LED light creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06208536773920059 + ], + "opposite_logprob_tokens": [ + -2.8120853900909424 + ], + "logprob_reward": -0.06208536773920059, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06208536773920059, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08728501945734024, + 0.5075148344039917, + 0.015131976455450058, + 0.01835877075791359, + 0.02925829589366913, + 0.02042163908481598, + 0.030995773151516914, + 0.1495775729417801, + 0.0939888060092926, + 0.04746730998158455 + ], + "credit_map": [ + 0.011590023525059223, + 0.01584109105169773, + 0.009646941907703876, + 0.011382774449884892, + 0.007906964048743248, + 0.008542532101273537, + 0.010250435210764408, + 0.012124260887503624, + 0.12783320248126984, + 0.02299303002655506, + 0.03326418250799179, + 0.024846920743584633, + 0.029473546892404556, + 0.034118421375751495, + 0.05332157760858536, + 0.18166394531726837, + 0.0017553868237882853, + 0.001356207299977541, + 0.0021834983490407467, + 0.0023038319777697325, + 0.0018478817073628306, + 0.0016778920544311404, + 0.0017585050081834197, + 0.0022487726528197527, + 0.0019577890634536743, + 0.002523573813959956, + 0.003007098101079464, + 0.0026438706554472446, + 0.0019264006987214088, + 0.0021674083545804024, + 0.0020091375336050987, + 0.0021234918385744095, + 0.0026499603409320116, + 0.003147970885038376, + 0.0037879596929997206, + 0.0032579749822616577, + 0.0029976621735841036, + 0.0028952693101018667, + 0.0066978661343455315, + 0.003823632374405861, + 0.001654149848036468, + 0.001789212808944285, + 0.0023359593469649553, + 0.002819073386490345, + 0.0035170349292457104, + 0.0029235545080155134, + 0.0029265875928103924, + 0.0024560659658163786, + 0.0031465778592973948, + 0.00855921395123005, + 0.0021128247026354074, + 0.0021626786328852177, + 0.0033531945664435625, + 0.0036816189531236887, + 0.0026047350838780403, + 0.0053749289363622665, + 0.008680525235831738, + 0.006168032065033913, + 0.004870353266596794, + 0.005928566213697195, + 0.00573553703725338, + 0.05063007026910782, + 0.013337459415197372, + 0.05422702804207802, + 0.009404261596500874, + 0.0034583257511258125, + 0.004345759749412537, + 0.010475734248757362, + 0.005302051547914743, + 0.0303904190659523, + 0.010371800512075424, + 0.020240457728505135, + 0.004049239680171013, + 0.004206890240311623, + 0.005597682204097509, + 0.006621128413826227, + 0.005194293800741434, + 0.008043486624956131, + 0.00694525521248579, + 0.006809332873672247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 8 + ], + "selected_reward": -0.06208536773920059 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a 3D printer's nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle a polished brass cone attached to wires and fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thin, continuous thread of molten vibrant red plastic extruding from the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the nozzle move with swift automation along linear and curvilinear paths over a dark textured print bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright focused LED light creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0003906917118001729 + ], + "opposite_logprob_tokens": [ + -7.875390529632568 + ], + "logprob_reward": -0.0003906917118001729, + "yes_no_margin_reward": 7.875, + "token_credit_reward": -0.0003906917118001729, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0312887467443943, + 0.6228381395339966, + 0.010639679618179798, + 0.02075076475739479, + 0.022427810356020927, + 0.058793533593416214, + 0.013430619612336159, + 0.09342983365058899, + 0.10410632193088531, + 0.02229454182088375 + ], + "credit_map": [ + 0.004062414634972811, + 0.0056176334619522095, + 0.0035721820313483477, + 0.0030827561859041452, + 0.004157882649451494, + 0.004179672338068485, + 0.0029627131298184395, + 0.003653491847217083, + 0.15655338764190674, + 0.02819538302719593, + 0.04081156477332115, + 0.028130892664194107, + 0.03707016259431839, + 0.041038673371076584, + 0.06738990545272827, + 0.22364813089370728, + 0.001381249283440411, + 0.0012321186950430274, + 0.0011074396315962076, + 0.002187118399888277, + 0.001025363802909851, + 0.0012369001051411033, + 0.0012040500296279788, + 0.0012654394377022982, + 0.0018551521934568882, + 0.0023617660626769066, + 0.002191864885389805, + 0.0036746219266206026, + 0.0027017472311854362, + 0.0018356539076194167, + 0.0035649908240884542, + 0.002564966445788741, + 0.0018524175975471735, + 0.0034767722245305777, + 0.0027397219091653824, + 0.0019953858572989702, + 0.002607238944619894, + 0.0030815619975328445, + 0.00384654407389462, + 0.0028281686827540398, + 0.004546753596514463, + 0.006214556749910116, + 0.006299206987023354, + 0.007214716169983149, + 0.010195333510637283, + 0.007126118987798691, + 0.008337379433214664, + 0.008859467692673206, + 0.0014742787461727858, + 0.001518908771686256, + 0.0012947495561093092, + 0.001904378063045442, + 0.002068187342956662, + 0.0016762169543653727, + 0.0018487833440303802, + 0.001645116601139307, + 0.00802344735711813, + 0.0031491080299019814, + 0.0033649865072220564, + 0.0032973492052406073, + 0.004427355248481035, + 0.03722864016890526, + 0.006343213841319084, + 0.02759573422372341, + 0.011146130040287971, + 0.0028105482924729586, + 0.004249198362231255, + 0.007944722659885883, + 0.004421229008585215, + 0.03819289058446884, + 0.009832971729338169, + 0.025508632883429527, + 0.001938547007739544, + 0.003306915983557701, + 0.003550905268639326, + 0.003740880871191621, + 0.0017133060609921813, + 0.0023365793749690056, + 0.0028358488343656063, + 0.002871559001505375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 7 + ], + "selected_reward": -0.0003906917118001729 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a 3D printer's nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle a polished brass cone attached to wires and fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thin, continuous thread of molten vibrant red plastic extruding from the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the nozzle move with swift automation along linear and curvilinear paths over a dark textured print bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright focused LED light creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004097281955182552 + ], + "opposite_logprob_tokens": [ + -5.5040974617004395 + ], + "logprob_reward": -0.004097281955182552, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004097281955182552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09298920631408691, + 0.23222367465496063, + 0.02154391258955002, + 0.02612767554819584, + 0.01953541673719883, + 0.0554986447095871, + 0.03967178985476494, + 0.14208631217479706, + 0.31681323051452637, + 0.053510162979364395 + ], + "credit_map": [ + 0.013160563074052334, + 0.016201075166463852, + 0.010033832862973213, + 0.013783987611532211, + 0.008701886050403118, + 0.010030570439994335, + 0.008892483077943325, + 0.012184804305434227, + 0.05652834475040436, + 0.010684299282729626, + 0.015817886218428612, + 0.025043193250894547, + 0.011776003986597061, + 0.015775857493281364, + 0.02216719090938568, + 0.07443089783191681, + 0.0021512212697416544, + 0.0018648722907528281, + 0.003514753421768546, + 0.004257827997207642, + 0.001822578371502459, + 0.0026303997728973627, + 0.0033362428657710552, + 0.0019660142716020346, + 0.002269165823236108, + 0.0032952630426734686, + 0.003868258558213711, + 0.003743821056559682, + 0.0029039226938039064, + 0.003412460209801793, + 0.003523357678204775, + 0.003111425554379821, + 0.001974977320060134, + 0.002300530206412077, + 0.003349052043631673, + 0.0020351146813482046, + 0.0022015238646417856, + 0.002287514042109251, + 0.0030965288169682026, + 0.0022901755291968584, + 0.0038561730179935694, + 0.00603094045072794, + 0.006656673736870289, + 0.007805520202964544, + 0.006270932499319315, + 0.00478573190048337, + 0.011338592506945133, + 0.008754082955420017, + 0.005260522942990065, + 0.010417836718261242, + 0.003415390383452177, + 0.004593082703649998, + 0.0037657013162970543, + 0.0031068231910467148, + 0.004610768519341946, + 0.00450166268274188, + 0.013640077784657478, + 0.0054083517752587795, + 0.0055143870413303375, + 0.005356983747333288, + 0.006860440596938133, + 0.056009259074926376, + 0.009562311694025993, + 0.03973449394106865, + 0.032913874834775925, + 0.009248806163668633, + 0.014328873716294765, + 0.02820511721074581, + 0.014911124482750893, + 0.11025869101285934, + 0.03291231021285057, + 0.0740344375371933, + 0.005403795745223761, + 0.0043476540595293045, + 0.006045069079846144, + 0.005914233159273863, + 0.004545810166746378, + 0.0118169654160738, + 0.008736232295632362, + 0.006700408179312944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 7 + ], + "selected_reward": -0.004097281955182552 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a 3D printer's nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle a polished brass cone attached to wires and fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thin, continuous thread of molten vibrant red plastic extruding from the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the nozzle move with swift automation along linear and curvilinear paths over a dark textured print bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright focused LED light creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005945852026343346 + ], + "opposite_logprob_tokens": [ + -5.130945682525635 + ], + "logprob_reward": -0.005945852026343346, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005945852026343346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04296840727329254, + 0.671675443649292, + 0.009924029000103474, + 0.023708529770374298, + 0.0259987935423851, + 0.06379301846027374, + 0.012947418726980686, + 0.08604102581739426, + 0.026028696447610855, + 0.03691454976797104 + ], + "credit_map": [ + 0.004387014079838991, + 0.010688254609704018, + 0.0034246661234647036, + 0.007521943189203739, + 0.004240940324962139, + 0.004600093699991703, + 0.0037307303864508867, + 0.00437476672232151, + 0.16678109765052795, + 0.029828088358044624, + 0.04260464012622833, + 0.030817802995443344, + 0.04106304422020912, + 0.04378402233123779, + 0.07334071397781372, + 0.24345606565475464, + 0.0011483097914606333, + 0.0007407923694700003, + 0.0018722069216892123, + 0.0011729898396879435, + 0.0011042102705687284, + 0.0016998897772282362, + 0.0010883681243285537, + 0.0010972622549161315, + 0.0026978682726621628, + 0.0013091748114675283, + 0.0016621991526335478, + 0.0012459666468203068, + 0.001563123776577413, + 0.008256341330707073, + 0.00216915481723845, + 0.004804702475667, + 0.0018672514706850052, + 0.0031748206820338964, + 0.0030461414717137814, + 0.0032425192184746265, + 0.003094435203820467, + 0.0030321499798446894, + 0.005405254662036896, + 0.003136221319437027, + 0.005214049015194178, + 0.006943278014659882, + 0.0069834026508033276, + 0.008898579515516758, + 0.011784334667026997, + 0.006545908749103546, + 0.0093346256762743, + 0.008088838309049606, + 0.0017431171145290136, + 0.001606628531590104, + 0.0017538754036650062, + 0.0015526973875239491, + 0.0014305649092420936, + 0.0015238592168316245, + 0.001702008070424199, + 0.001634667394682765, + 0.006575141567736864, + 0.0029355562292039394, + 0.0030770250596106052, + 0.0030429025646299124, + 0.004602352622896433, + 0.033928949385881424, + 0.005884512793272734, + 0.025994587689638138, + 0.002789018675684929, + 0.0010050622513517737, + 0.0016994737088680267, + 0.002432097215205431, + 0.001219144556671381, + 0.009221321903169155, + 0.002551804529502988, + 0.005110773723572493, + 0.002727955812588334, + 0.004332222975790501, + 0.004327517468482256, + 0.0039151874370872974, + 0.004880300723016262, + 0.0065955244936048985, + 0.005646189674735069, + 0.0044896528124809265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 5 + ], + "selected_reward": -0.005945852026343346 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, slightly high-angle shot allows the camera to smoothly track the precise movements of the 3D printer's nozzle. The nozzle, a polished brass cone attached to a complex assembly of wires and fans, glides delicately just millimeters above the dark, textured print bed. From its tiny, heated tip, a thin, continuous thread of molten, vibrant red plastic steadily extrudes, gleaming under the bright, focused LED light. The nozzle moves with a swift and decisive automation, executing a series of linear and curvilinear paths as it meticulously deposits the material. The lustrous red filament cools and solidifies almost instantly upon contact, fusing with the layers below to form a complex geometric part whose surface displays a fine, striated texture layered in successive layers. The entire scene is bathed in a cool, clinical light, casting sharp highlights across the gleaming surfaces of the metal parts and the newly printed plastic, creating deep shadows and accentuating the object's complex, ever-growing form. The background is a soft, out-of-focus blur, suggesting a workshop environment where the methodical act of creation is the centerpiece, creating an atmosphere of focus, hypnosis, and technical precision.", + "seed": 27039, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_11/index_3169_seed_27039_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03376535194693133, + "vlm_logprob": -0.03376535194693133, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.03376535194693133, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a 3D printer's nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle a polished brass cone attached to wires and fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thin, continuous thread of molten vibrant red plastic extruding from the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the nozzle move with swift automation along linear and curvilinear paths over a dark textured print bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright focused LED light creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04395489767193794 + ], + "opposite_logprob_tokens": [ + -3.543954849243164 + ], + "logprob_reward": -0.04395489767193794, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.04395489767193794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37122485041618347, + 0.05013742670416832, + 0.03211057186126709, + 0.10247326642274857, + 0.11160637438297272, + 0.026723777875304222, + 0.07469560950994492, + 0.021016867831349373, + 0.1494421809911728, + 0.060569021850824356 + ], + "credit_map": [ + 0.044643573462963104, + 0.0494827963411808, + 0.044302087277173996, + 0.03527132049202919, + 0.055557284504175186, + 0.059682998806238174, + 0.04404638707637787, + 0.03823840990662575, + 0.005933528300374746, + 0.006610517390072346, + 0.009036542847752571, + 0.006862435955554247, + 0.004858356434851885, + 0.005271318834275007, + 0.00688752019777894, + 0.00467721140012145, + 0.004218012560158968, + 0.005626125261187553, + 0.00389364012517035, + 0.0032245947513729334, + 0.0035800491459667683, + 0.003241846337914467, + 0.004392756149172783, + 0.003933548461645842, + 0.019084567204117775, + 0.011068034917116165, + 0.01136779971420765, + 0.01633131131529808, + 0.01207726076245308, + 0.015491010621190071, + 0.008880464360117912, + 0.008172824047505856, + 0.015236007049679756, + 0.015467596240341663, + 0.013301854953169823, + 0.010518552735447884, + 0.01241350919008255, + 0.028493845835328102, + 0.008094155229628086, + 0.008080868981778622, + 0.0037127910181879997, + 0.004984970670193434, + 0.0034584472887218, + 0.0023355334997177124, + 0.004125410225242376, + 0.002082056598737836, + 0.0031036175787448883, + 0.0029209519270807505, + 0.022634373977780342, + 0.008234566077589989, + 0.004758525639772415, + 0.01613140106201172, + 0.00319698010571301, + 0.01023530587553978, + 0.007339016534388065, + 0.002165451878681779, + 0.0023737470619380474, + 0.0021369592286646366, + 0.004119109828025103, + 0.0035793257411569357, + 0.0024601719342172146, + 0.002203283365815878, + 0.002762418007478118, + 0.0013818537117913365, + 0.018427247181534767, + 0.019915783777832985, + 0.0331590436398983, + 0.020779086276888847, + 0.017684923484921455, + 0.014109211042523384, + 0.013521495275199413, + 0.01184541080147028, + 0.017250308766961098, + 0.009299617260694504, + 0.006380642298609018, + 0.01258064340800047, + 0.0020689861848950386, + 0.004981331527233124, + 0.005343519616872072, + 0.002663980470970273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.04395489767193794 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a 3D printer's nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle a polished brass cone attached to wires and fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thin, continuous thread of molten vibrant red plastic extruding from the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the nozzle move with swift automation along linear and curvilinear paths over a dark textured print bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright focused LED light creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08907370269298553 + ], + "opposite_logprob_tokens": [ + -2.464073657989502 + ], + "logprob_reward": -0.08907370269298553, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08907370269298553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3332606554031372, + 0.10920751094818115, + 0.062497545033693314, + 0.06933817267417908, + 0.025996806100010872, + 0.050191957503557205, + 0.15757150948047638, + 0.04161450266838074, + 0.08987200260162354, + 0.060449402779340744 + ], + "credit_map": [ + 0.03751625865697861, + 0.042726509273052216, + 0.040105681866407394, + 0.04929864779114723, + 0.048392295837402344, + 0.03993159160017967, + 0.043427735567092896, + 0.03186190128326416, + 0.021834388375282288, + 0.020319974049925804, + 0.01033251266926527, + 0.008166841231286526, + 0.013454798609018326, + 0.01758177950978279, + 0.008917069062590599, + 0.008600139990448952, + 0.009488629177212715, + 0.010614915750920773, + 0.0074508413672447205, + 0.005986853502690792, + 0.008192182518541813, + 0.0064741321839392185, + 0.008587908931076527, + 0.005702075082808733, + 0.007436209823936224, + 0.006896720267832279, + 0.01163391675800085, + 0.008358492515981197, + 0.012424692511558533, + 0.0084193404763937, + 0.0069154961965978146, + 0.007253298070281744, + 0.004081820137798786, + 0.00403020391240716, + 0.0029980612453073263, + 0.0028608671855181456, + 0.0024026203900575638, + 0.004368409980088472, + 0.002566681709140539, + 0.002688140142709017, + 0.007302728947252035, + 0.007676982320845127, + 0.01087195985019207, + 0.007054528687149286, + 0.004925113171339035, + 0.0030596284195780754, + 0.0049978457391262054, + 0.004303161054849625, + 0.04759635403752327, + 0.014400805346667767, + 0.009833297692239285, + 0.025222688913345337, + 0.008769732899963856, + 0.029695600271224976, + 0.012854629196226597, + 0.009198387153446674, + 0.007103453390300274, + 0.007445039227604866, + 0.005506130866706371, + 0.005593434441834688, + 0.004138536285609007, + 0.00495570432394743, + 0.003988101612776518, + 0.002884095534682274, + 0.014739048667252064, + 0.01908237859606743, + 0.011787593364715576, + 0.006904916372150183, + 0.011052336543798447, + 0.011363854631781578, + 0.006832865998148918, + 0.008109004236757755, + 0.0073304506950080395, + 0.006398400291800499, + 0.012451482936739922, + 0.010506796650588512, + 0.004573362413793802, + 0.009113671258091927, + 0.005675602238625288, + 0.004399632569402456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.08907370269298553 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a 3D printer's nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle a polished brass cone attached to wires and fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thin, continuous thread of molten vibrant red plastic extruding from the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the nozzle move with swift automation along linear and curvilinear paths over a dark textured print bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright focused LED light creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008176797418855131 + ], + "opposite_logprob_tokens": [ + -7.125817775726318 + ], + "logprob_reward": -0.0008176797418855131, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008176797418855131, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2900020182132721, + 0.12621983885765076, + 0.08616772294044495, + 0.15155518054962158, + 0.05741916596889496, + 0.059828612953424454, + 0.04781459644436836, + 0.02523527480661869, + 0.07874248176813126, + 0.07701508700847626 + ], + "credit_map": [ + 0.034892693161964417, + 0.03485694155097008, + 0.043794434517621994, + 0.03532318025827408, + 0.04186732694506645, + 0.0261932872235775, + 0.047563351690769196, + 0.025510789826512337, + 0.020959457382559776, + 0.01926860585808754, + 0.014883358962833881, + 0.012222958728671074, + 0.01761249452829361, + 0.02018246427178383, + 0.0105928098782897, + 0.010497679933905602, + 0.01336923148483038, + 0.014976026490330696, + 0.00966371688991785, + 0.007697370834648609, + 0.012002890929579735, + 0.008740684948861599, + 0.01161852478981018, + 0.008099280297756195, + 0.02855844423174858, + 0.013387054204940796, + 0.01930002123117447, + 0.031095171347260475, + 0.01623031124472618, + 0.01435892190784216, + 0.016843101009726524, + 0.011782154440879822, + 0.007305692415684462, + 0.00845649465918541, + 0.008776109665632248, + 0.007151088677346706, + 0.006909605581313372, + 0.00990282278507948, + 0.004063068423420191, + 0.004854283295571804, + 0.008482006378471851, + 0.007788169663399458, + 0.010819670744240284, + 0.00802172627300024, + 0.007102383766323328, + 0.004288572818040848, + 0.0069306110963225365, + 0.006395469419658184, + 0.009480935521423817, + 0.004232252016663551, + 0.005336207803338766, + 0.004346458241343498, + 0.006114318501204252, + 0.008068081922829151, + 0.005250279325991869, + 0.0049860598519444466, + 0.0033126359339803457, + 0.0039384625852108, + 0.00282257329672575, + 0.0034179859794676304, + 0.003159686690196395, + 0.00349590671248734, + 0.0031068758107721806, + 0.0019811498932540417, + 0.011743755079805851, + 0.017345095053315163, + 0.012476588599383831, + 0.008875715546309948, + 0.009247083216905594, + 0.006502977106720209, + 0.006763033103197813, + 0.005788236390799284, + 0.009832773357629776, + 0.00909532979130745, + 0.01406437810510397, + 0.012578535825014114, + 0.008626880124211311, + 0.01048779021948576, + 0.0071214777417480946, + 0.005207926966249943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.0008176797418855131 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a 3D printer's nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle a polished brass cone attached to wires and fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thin, continuous thread of molten vibrant red plastic extruding from the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the nozzle move with swift automation along linear and curvilinear paths over a dark textured print bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright focused LED light creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008642295375466347 + ], + "opposite_logprob_tokens": [ + -4.758642196655273 + ], + "logprob_reward": -0.008642295375466347, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008642295375466347, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3298589587211609, + 0.1542089730501175, + 0.04135602340102196, + 0.12393035739660263, + 0.03254319727420807, + 0.04687650129199028, + 0.1034260019659996, + 0.03443051874637604, + 0.06863819807767868, + 0.06473124772310257 + ], + "credit_map": [ + 0.031944986432790756, + 0.04759768024086952, + 0.0357954241335392, + 0.04311033710837364, + 0.06690357625484467, + 0.038389746099710464, + 0.04435383528470993, + 0.02176336944103241, + 0.02559330128133297, + 0.027750398963689804, + 0.01095326617360115, + 0.010875795967876911, + 0.024657540023326874, + 0.028818456456065178, + 0.013702879659831524, + 0.011857341974973679, + 0.005214531905949116, + 0.006070649717003107, + 0.0063612330704927444, + 0.004609923809766769, + 0.005992322228848934, + 0.005390434060245752, + 0.003957283683121204, + 0.0037596451584249735, + 0.018573299050331116, + 0.013130644336342812, + 0.013383502140641212, + 0.025142131373286247, + 0.02137736976146698, + 0.011100095696747303, + 0.012987468391656876, + 0.008235846646130085, + 0.004988201893866062, + 0.0064028906635940075, + 0.0034959360491484404, + 0.00437596719712019, + 0.0031791392248123884, + 0.0042870244942605495, + 0.003117053769528866, + 0.0026969818864017725, + 0.005338483955711126, + 0.005300878547132015, + 0.012928874231874943, + 0.00837259367108345, + 0.004070137161761522, + 0.0021822666749358177, + 0.005029451567679644, + 0.00365381664596498, + 0.024774139747023582, + 0.011422757059335709, + 0.008556223474442959, + 0.018294231966137886, + 0.009096495807170868, + 0.013854939490556717, + 0.010980500839650631, + 0.006446709856390953, + 0.0054548694752156734, + 0.006204764824360609, + 0.004463881254196167, + 0.0035336248110979795, + 0.004022390116006136, + 0.004971922840923071, + 0.003727405797690153, + 0.002051659394055605, + 0.01093331165611744, + 0.015109291300177574, + 0.009449170902371407, + 0.005630940198898315, + 0.00675257807597518, + 0.007532660383731127, + 0.006385361775755882, + 0.0068448870442807674, + 0.010185545310378075, + 0.008262606337666512, + 0.01074578519910574, + 0.010403500869870186, + 0.006977700628340244, + 0.00776753481477499, + 0.005873288493603468, + 0.00451528700068593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.008642295375466347 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a 3D printer's nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle a polished brass cone attached to wires and fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thin, continuous thread of molten vibrant red plastic extruding from the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the nozzle move with swift automation along linear and curvilinear paths over a dark textured print bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright focused LED light creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026338184252381325 + ], + "opposite_logprob_tokens": [ + -3.6513381004333496 + ], + "logprob_reward": -0.026338184252381325, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026338184252381325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23908376693725586, + 0.156017005443573, + 0.04733423516154289, + 0.048042092472314835, + 0.04607154056429863, + 0.07337785512208939, + 0.06895726174116135, + 0.10245659202337265, + 0.14695866405963898, + 0.07170095294713974 + ], + "credit_map": [ + 0.026661885902285576, + 0.03216056153178215, + 0.03610118106007576, + 0.02769067883491516, + 0.04496489465236664, + 0.02479291707277298, + 0.026518598198890686, + 0.02019304595887661, + 0.025550484657287598, + 0.025852076709270477, + 0.01618727296590805, + 0.016733424738049507, + 0.01202651672065258, + 0.023014718666672707, + 0.018272966146469116, + 0.018379535526037216, + 0.006752002984285355, + 0.007529826834797859, + 0.006360514555126429, + 0.00499681755900383, + 0.004338419996201992, + 0.00454370304942131, + 0.006735098548233509, + 0.006077848840504885, + 0.005907145794481039, + 0.006289252080023289, + 0.004486024379730225, + 0.005092742387205362, + 0.010407022200524807, + 0.007743421010673046, + 0.00422489270567894, + 0.0038915907498449087, + 0.0064243897795677185, + 0.006242123898118734, + 0.005682244431227446, + 0.005033251363784075, + 0.004256180487573147, + 0.011252176016569138, + 0.0038561627734452486, + 0.003325012279674411, + 0.009425459429621696, + 0.011604559607803822, + 0.01437793392688036, + 0.010326583869755268, + 0.008825596421957016, + 0.004433572292327881, + 0.00728314695879817, + 0.007100997492671013, + 0.020605022087693214, + 0.007273491472005844, + 0.006690043956041336, + 0.004548955243080854, + 0.004572008736431599, + 0.016702774912118912, + 0.004746783059090376, + 0.0038181799463927746, + 0.02919662930071354, + 0.02014984004199505, + 0.01021044235676527, + 0.015302890911698341, + 0.004116677679121494, + 0.007319062948226929, + 0.009875708259642124, + 0.006285334005951881, + 0.018589749932289124, + 0.019698118790984154, + 0.02698027528822422, + 0.016311176121234894, + 0.016443544998764992, + 0.015995929017663002, + 0.015145975165069103, + 0.017793875187635422, + 0.009826011955738068, + 0.00907881185412407, + 0.012312760576605797, + 0.012649249285459518, + 0.006421222351491451, + 0.009576953947544098, + 0.006783891469240189, + 0.005052043590694666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.026338184252381325 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, slightly high-angle shot allows the camera to smoothly track the precise movements of the 3D printer's nozzle. The nozzle, a polished brass cone attached to a complex assembly of wires and fans, glides delicately just millimeters above the dark, textured print bed. From its tiny, heated tip, a thin, continuous thread of molten, vibrant red plastic steadily extrudes, gleaming under the bright, focused LED light. The nozzle moves with a swift and decisive automation, executing a series of linear and curvilinear paths as it meticulously deposits the material. The lustrous red filament cools and solidifies almost instantly upon contact, fusing with the layers below to form a complex geometric part whose surface displays a fine, striated texture layered in successive layers. The entire scene is bathed in a cool, clinical light, casting sharp highlights across the gleaming surfaces of the metal parts and the newly printed plastic, creating deep shadows and accentuating the object's complex, ever-growing form. The background is a soft, out-of-focus blur, suggesting a workshop environment where the methodical act of creation is the centerpiece, creating an atmosphere of focus, hypnosis, and technical precision.", + "seed": 27040, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_11/index_3169_seed_27040_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.014756002661306411, + "vlm_logprob": -0.014756002661306411, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.014756002661306411, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a 3D printer's nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle a polished brass cone attached to wires and fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thin, continuous thread of molten vibrant red plastic extruding from the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the nozzle move with swift automation along linear and curvilinear paths over a dark textured print bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright focused LED light creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027325564995408058 + ], + "opposite_logprob_tokens": [ + -3.9023256301879883 + ], + "logprob_reward": -0.027325564995408058, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.027325564995408058, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13482818007469177, + 0.05521371215581894, + 0.0700676366686821, + 0.07323476672172546, + 0.21208195388317108, + 0.038944147527217865, + 0.0994078665971756, + 0.05031571164727211, + 0.04312780499458313, + 0.22277824580669403 + ], + "credit_map": [ + 0.020375575870275497, + 0.015868602320551872, + 0.014423697255551815, + 0.012867726385593414, + 0.023606600239872932, + 0.01588021218776703, + 0.013269052840769291, + 0.018536720424890518, + 0.007688832003623247, + 0.0074118697084486485, + 0.005509660113602877, + 0.008891446515917778, + 0.00401042215526104, + 0.005648401565849781, + 0.008935360237956047, + 0.0071177175268530846, + 0.023047080263495445, + 0.007275391835719347, + 0.0037968617398291826, + 0.0026140939444303513, + 0.00591058935970068, + 0.005605054087936878, + 0.003384840674698353, + 0.01843372732400894, + 0.011096086353063583, + 0.01922939158976078, + 0.005620935466140509, + 0.006097892764955759, + 0.007198945619165897, + 0.008755593560636044, + 0.006284110248088837, + 0.008951808325946331, + 0.06459283828735352, + 0.019357670098543167, + 0.01019308902323246, + 0.007247649133205414, + 0.014474460855126381, + 0.028454303741455078, + 0.010763299651443958, + 0.056998640298843384, + 0.005504359025508165, + 0.0038331507239490747, + 0.004533057566732168, + 0.007358996197581291, + 0.002365975407883525, + 0.005357636604458094, + 0.004951108247041702, + 0.0050398665480315685, + 0.014702360145747662, + 0.016350157558918, + 0.006986929103732109, + 0.007903313264250755, + 0.018846748396754265, + 0.013087849132716656, + 0.010958889499306679, + 0.010571620427072048, + 0.007038225419819355, + 0.006205395329743624, + 0.006982991471886635, + 0.005350358318537474, + 0.003935322165489197, + 0.004480606876313686, + 0.0072852070443332195, + 0.009037604555487633, + 0.013556642457842827, + 0.0034778520930558443, + 0.002907167887315154, + 0.0027704478707164526, + 0.00289699318818748, + 0.004290927667170763, + 0.0025473120622336864, + 0.010680459439754486, + 0.02343142405152321, + 0.02670612744987011, + 0.02403583750128746, + 0.020178403705358505, + 0.03519798442721367, + 0.036933328956365585, + 0.030572261661291122, + 0.02572288177907467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.027325564995408058 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a 3D printer's nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle a polished brass cone attached to wires and fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thin, continuous thread of molten vibrant red plastic extruding from the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the nozzle move with swift automation along linear and curvilinear paths over a dark textured print bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright focused LED light creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060554549098015 + ], + "opposite_logprob_tokens": [ + -3.8956055641174316 + ], + "logprob_reward": -0.02060554549098015, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060554549098015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13427114486694336, + 0.02962944470345974, + 0.17488841712474823, + 0.04057817533612251, + 0.2888508141040802, + 0.0303203072398901, + 0.05896073952317238, + 0.1002582311630249, + 0.04090683162212372, + 0.10133589059114456 + ], + "credit_map": [ + 0.02258705534040928, + 0.030826510861516, + 0.011844463646411896, + 0.010493263602256775, + 0.016594085842370987, + 0.014960476197302341, + 0.00983262900263071, + 0.01713266782462597, + 0.0034315010998398066, + 0.002889091381803155, + 0.0032325047068297863, + 0.004860774613916874, + 0.00295367743819952, + 0.00404742918908596, + 0.004985197447240353, + 0.0032292683608829975, + 0.06820440292358398, + 0.010069597512483597, + 0.010429923422634602, + 0.005536661017686129, + 0.006588731426745653, + 0.008532021194696426, + 0.008251620456576347, + 0.057275477796792984, + 0.005150833632797003, + 0.005828092340379953, + 0.003149314783513546, + 0.003279786091297865, + 0.007844086736440659, + 0.006236847955733538, + 0.0044557154178619385, + 0.004633496981114149, + 0.0974428653717041, + 0.02003108710050583, + 0.01457265205681324, + 0.008554123342037201, + 0.012639476917684078, + 0.03649882599711418, + 0.012960847467184067, + 0.08615093678236008, + 0.003972494974732399, + 0.004773525521159172, + 0.002145889215171337, + 0.0032166303135454655, + 0.0032801306806504726, + 0.005456889979541302, + 0.0036953578237444162, + 0.0037793894298374653, + 0.010164965875446796, + 0.011841334402561188, + 0.004675616510212421, + 0.004153993912041187, + 0.007765672169625759, + 0.009182125329971313, + 0.0054645235650241375, + 0.00571250356733799, + 0.01398614700883627, + 0.019361522048711777, + 0.0049118720926344395, + 0.006213558372110128, + 0.026669606566429138, + 0.011686589568853378, + 0.00804693903774023, + 0.009381991811096668, + 0.008553595282137394, + 0.0029570600017905235, + 0.003022202057763934, + 0.0032129206229001284, + 0.003937177825719118, + 0.008222628384828568, + 0.003597790841013193, + 0.007403457537293434, + 0.010434532538056374, + 0.01270949374884367, + 0.00893300399184227, + 0.006733887363225222, + 0.01657666265964508, + 0.022840941324830055, + 0.010998995043337345, + 0.012108376249670982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.02060554549098015 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a 3D printer's nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle a polished brass cone attached to wires and fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thin, continuous thread of molten vibrant red plastic extruding from the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the nozzle move with swift automation along linear and curvilinear paths over a dark textured print bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright focused LED light creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009239218779839575 + ], + "opposite_logprob_tokens": [ + -7.000924110412598 + ], + "logprob_reward": -0.0009239218779839575, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009239218779839575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10069748014211655, + 0.058606263250112534, + 0.13173304498195648, + 0.1824924498796463, + 0.21871262788772583, + 0.029425714164972305, + 0.12164238095283508, + 0.051999542862176895, + 0.0378648117184639, + 0.06682562828063965 + ], + "credit_map": [ + 0.01473222952336073, + 0.027140740305185318, + 0.012775192968547344, + 0.00721832737326622, + 0.00841566827148199, + 0.009343277662992477, + 0.0075253709219396114, + 0.013546676374971867, + 0.007107457611709833, + 0.007606549188494682, + 0.006311364471912384, + 0.006883020047098398, + 0.005445459391921759, + 0.00905643217265606, + 0.00841546431183815, + 0.007780515123158693, + 0.04580594599246979, + 0.012911158613860607, + 0.009326532483100891, + 0.005001808051019907, + 0.009048979729413986, + 0.006411333102732897, + 0.005368740297853947, + 0.03785856068134308, + 0.028879323974251747, + 0.029779935255646706, + 0.018584270030260086, + 0.02466551400721073, + 0.012177561409771442, + 0.017801957204937935, + 0.02359667979180813, + 0.0270072091370821, + 0.04465586319565773, + 0.027804842218756676, + 0.012032250873744488, + 0.00946106482297182, + 0.022314617410302162, + 0.04539750516414642, + 0.015907056629657745, + 0.041139423847198486, + 0.004134413320571184, + 0.0032780838664621115, + 0.0022685055155307055, + 0.00241227843798697, + 0.003009287640452385, + 0.005947365425527096, + 0.0038563876878470182, + 0.004519392736256123, + 0.017402615398168564, + 0.016684515401721, + 0.010118584148585796, + 0.007298319134861231, + 0.018750518560409546, + 0.008963245898485184, + 0.013202506117522717, + 0.029222074896097183, + 0.007946993224322796, + 0.0059284972958266735, + 0.006719114724546671, + 0.005748587194830179, + 0.007472385186702013, + 0.0048406790010631084, + 0.007071868982166052, + 0.0062714191153645515, + 0.0055471621453762054, + 0.00281350314617157, + 0.0042032888159155846, + 0.0019626605790108442, + 0.00417112885043025, + 0.006491268519312143, + 0.0049620866775512695, + 0.007713713683187962, + 0.007292499300092459, + 0.006929120514541864, + 0.007090377621352673, + 0.006703113205730915, + 0.010606838390231133, + 0.00919393077492714, + 0.011050241999328136, + 0.007959505543112755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.0009239218779839575 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a 3D printer's nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle a polished brass cone attached to wires and fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thin, continuous thread of molten vibrant red plastic extruding from the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the nozzle move with swift automation along linear and curvilinear paths over a dark textured print bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright focused LED light creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0181918703019619 + ], + "opposite_logprob_tokens": [ + -4.018191814422607 + ], + "logprob_reward": -0.0181918703019619, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.0181918703019619, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19295017421245575, + 0.05684350058436394, + 0.12652581930160522, + 0.06298443675041199, + 0.18916858732700348, + 0.03701295331120491, + 0.15579377114772797, + 0.07907035946846008, + 0.05321316048502922, + 0.04643721505999565 + ], + "credit_map": [ + 0.03257855027914047, + 0.06035972386598587, + 0.016002485528588295, + 0.01267859898507595, + 0.019190263003110886, + 0.014318411238491535, + 0.014042424038052559, + 0.023779720067977905, + 0.008039948530495167, + 0.008118621073663235, + 0.004864659160375595, + 0.00818677432835102, + 0.00543313380330801, + 0.007116436492651701, + 0.006364095024764538, + 0.00871983077377081, + 0.042028460651636124, + 0.015041830949485302, + 0.006581731140613556, + 0.00560792675241828, + 0.008663050830364227, + 0.007108028512448072, + 0.0068415747955441475, + 0.034653205424547195, + 0.010309714823961258, + 0.011922880075871944, + 0.005543503910303116, + 0.004716836381703615, + 0.007598293479532003, + 0.008563097566366196, + 0.006515129469335079, + 0.007814977318048477, + 0.06302328407764435, + 0.013880372978746891, + 0.009079484269022942, + 0.005263331346213818, + 0.009712236002087593, + 0.02139425277709961, + 0.01044443715363741, + 0.05637119710445404, + 0.004536216147243977, + 0.006207164376974106, + 0.002637607976794243, + 0.0056210728362202644, + 0.003933382220566273, + 0.0057160439901053905, + 0.004207535181194544, + 0.004153931047767401, + 0.022533535957336426, + 0.024544505402445793, + 0.012268317863345146, + 0.008721647784113884, + 0.02403334155678749, + 0.011288988403975964, + 0.016332227736711502, + 0.03607119619846344, + 0.01301965955644846, + 0.015163918025791645, + 0.00821072980761528, + 0.008079647086560726, + 0.008054026402533054, + 0.0065268767066299915, + 0.008943237364292145, + 0.011072264052927494, + 0.007471096236258745, + 0.00473678158596158, + 0.005546428728848696, + 0.0034446767531335354, + 0.006044046021997929, + 0.008630172349512577, + 0.007128231693059206, + 0.010211726650595665, + 0.005651144310832024, + 0.0062049562111496925, + 0.0041281962767243385, + 0.004300555679947138, + 0.0066063725389540195, + 0.008271651342511177, + 0.006082647480070591, + 0.00519169308245182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.0181918703019619 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a 3D printer's nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle a polished brass cone attached to wires and fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thin, continuous thread of molten vibrant red plastic extruding from the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the nozzle move with swift automation along linear and curvilinear paths over a dark textured print bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright focused LED light creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006733110640197992 + ], + "opposite_logprob_tokens": [ + -5.006732940673828 + ], + "logprob_reward": -0.006733110640197992, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006733110640197992, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24770212173461914, + 0.031301770359277725, + 0.08931490778923035, + 0.049861811101436615, + 0.3817377984523773, + 0.01823139749467373, + 0.03856664523482323, + 0.05608277767896652, + 0.050682492554187775, + 0.03651825338602066 + ], + "credit_map": [ + 0.03513028472661972, + 0.05756130814552307, + 0.02939104288816452, + 0.021134981885552406, + 0.01909179799258709, + 0.02442776970565319, + 0.02304142527282238, + 0.03792354092001915, + 0.0040396456606686115, + 0.0027991533279418945, + 0.004105294588953257, + 0.005325736477971077, + 0.0025008718948811293, + 0.0029638856649398804, + 0.005435834173113108, + 0.004131351597607136, + 0.021829454228281975, + 0.014282699674367905, + 0.005645222030580044, + 0.007265884429216385, + 0.007048625033348799, + 0.006925460882484913, + 0.007773065473884344, + 0.018544508144259453, + 0.007534352596849203, + 0.009704272262752056, + 0.0042956494726240635, + 0.004913248587399721, + 0.005214587319642305, + 0.007008557207882404, + 0.0051255603320896626, + 0.006065587513148785, + 0.1256263703107834, + 0.025650888681411743, + 0.025938071310520172, + 0.022245611995458603, + 0.01588411256670952, + 0.035555899143218994, + 0.02227001264691353, + 0.10856687277555466, + 0.0019981383811682463, + 0.0022489302791655064, + 0.001977001316845417, + 0.002743830904364586, + 0.0017576656537130475, + 0.002451388631016016, + 0.00263003958389163, + 0.002424403792247176, + 0.004768324550241232, + 0.004710448440164328, + 0.00411715405061841, + 0.003738876199349761, + 0.005179936997592449, + 0.00598502391949296, + 0.005337257403880358, + 0.004729628097265959, + 0.009317805990576744, + 0.009657595306634903, + 0.004597033374011517, + 0.0045507620088756084, + 0.00985365454107523, + 0.004169262945652008, + 0.006727559957653284, + 0.007209105417132378, + 0.008294530212879181, + 0.004337762948125601, + 0.003403427079319954, + 0.0032519595697522163, + 0.005939881317317486, + 0.012833250686526299, + 0.006160348188132048, + 0.0064613413996994495, + 0.004275584127753973, + 0.0046677659265697, + 0.0030522653833031654, + 0.002099287696182728, + 0.004897975362837315, + 0.009338690899312496, + 0.0029572194907814264, + 0.005229467526078224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.006733110640197992 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, slightly high-angle shot allows the camera to smoothly track the precise movements of the 3D printer's nozzle. The nozzle, a polished brass cone attached to a complex assembly of wires and fans, glides delicately just millimeters above the dark, textured print bed. From its tiny, heated tip, a thin, continuous thread of molten, vibrant red plastic steadily extrudes, gleaming under the bright, focused LED light. The nozzle moves with a swift and decisive automation, executing a series of linear and curvilinear paths as it meticulously deposits the material. The lustrous red filament cools and solidifies almost instantly upon contact, fusing with the layers below to form a complex geometric part whose surface displays a fine, striated texture layered in successive layers. The entire scene is bathed in a cool, clinical light, casting sharp highlights across the gleaming surfaces of the metal parts and the newly printed plastic, creating deep shadows and accentuating the object's complex, ever-growing form. The background is a soft, out-of-focus blur, suggesting a workshop environment where the methodical act of creation is the centerpiece, creating an atmosphere of focus, hypnosis, and technical precision.", + "seed": 27041, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_11/index_3169_seed_27041_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04147485033608973, + "vlm_logprob": -0.04147485033608973, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.04147485033608973, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a 3D printer's nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle a polished brass cone attached to wires and fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thin, continuous thread of molten vibrant red plastic extruding from the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the nozzle move with swift automation along linear and curvilinear paths over a dark textured print bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright focused LED light creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1346518099308014 + ], + "opposite_logprob_tokens": [ + -2.1346518993377686 + ], + "logprob_reward": -0.1346518099308014, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1346518099308014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19231009483337402, + 0.08583783358335495, + 0.07848092913627625, + 0.1127866879105568, + 0.06344428658485413, + 0.11083903908729553, + 0.06265565752983093, + 0.03731824457645416, + 0.1357738971710205, + 0.1205533817410469 + ], + "credit_map": [ + 0.022217733785510063, + 0.017416341230273247, + 0.027146153151988983, + 0.03686375170946121, + 0.01616491936147213, + 0.024228425696492195, + 0.02825562097132206, + 0.020017141476273537, + 0.006403774488717318, + 0.009893226437270641, + 0.013898166827857494, + 0.01001948956400156, + 0.013547735288739204, + 0.011538384482264519, + 0.009341754950582981, + 0.011195302940905094, + 0.0075410837307572365, + 0.008422262035310268, + 0.011126016266644001, + 0.010585284791886806, + 0.008336804807186127, + 0.013657801784574986, + 0.008317296393215656, + 0.010494374670088291, + 0.010527726262807846, + 0.018064355477690697, + 0.016154423356056213, + 0.014943602494895458, + 0.02224828116595745, + 0.008549482561647892, + 0.01246719341725111, + 0.009831628762185574, + 0.006529850885272026, + 0.008262011222541332, + 0.008714127354323864, + 0.007379724644124508, + 0.007044561207294464, + 0.009792888537049294, + 0.006398697849363089, + 0.00932242814451456, + 0.012507946230471134, + 0.012857566587626934, + 0.010004756040871143, + 0.011275122873485088, + 0.012727112509310246, + 0.022688601166009903, + 0.0094703184440732, + 0.019307617098093033, + 0.006446976214647293, + 0.006668795831501484, + 0.007330077234655619, + 0.009816261939704418, + 0.009281129576265812, + 0.006083574146032333, + 0.00934511050581932, + 0.007683734875172377, + 0.004935449920594692, + 0.0032260403968393803, + 0.005545245483517647, + 0.003152265679091215, + 0.0035804978106170893, + 0.005958517547696829, + 0.004909164737910032, + 0.006011066492646933, + 0.019078999757766724, + 0.010281190276145935, + 0.015912970528006554, + 0.014937540516257286, + 0.017719220370054245, + 0.018581734970211983, + 0.025079986080527306, + 0.014182251878082752, + 0.016320912167429924, + 0.016964808106422424, + 0.014743730425834656, + 0.010942567139863968, + 0.022652242332696915, + 0.014124216511845589, + 0.013654238544404507, + 0.0111506637185812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.1346518099308014 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a 3D printer's nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle a polished brass cone attached to wires and fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thin, continuous thread of molten vibrant red plastic extruding from the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the nozzle move with swift automation along linear and curvilinear paths over a dark textured print bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright focused LED light creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0620780847966671 + ], + "opposite_logprob_tokens": [ + -2.8120779991149902 + ], + "logprob_reward": -0.0620780847966671, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0620780847966671, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15985924005508423, + 0.03965414687991142, + 0.04790875315666199, + 0.5553637146949768, + 0.04044417664408684, + 0.033024027943611145, + 0.026472611352801323, + 0.03167799115180969, + 0.030601870268583298, + 0.034993428736925125 + ], + "credit_map": [ + 0.020628711208701134, + 0.01828817091882229, + 0.02431437559425831, + 0.023699307814240456, + 0.015361718833446503, + 0.017952192574739456, + 0.022784465923905373, + 0.01683029904961586, + 0.0033609054517000914, + 0.004138533025979996, + 0.005338368937373161, + 0.006888995412737131, + 0.00535519840195775, + 0.0045061856508255005, + 0.005496156867593527, + 0.0045698038302361965, + 0.005056021735072136, + 0.0073353201150894165, + 0.004597770515829325, + 0.0055716028437018394, + 0.005465694237500429, + 0.010016879066824913, + 0.004379544872790575, + 0.00548592209815979, + 0.06024831160902977, + 0.0712273120880127, + 0.06509775668382645, + 0.14116401970386505, + 0.07610481232404709, + 0.03208847716450691, + 0.0660046711564064, + 0.04342835396528244, + 0.0045334408059716225, + 0.006137242075055838, + 0.005708083510398865, + 0.0049307020381093025, + 0.004598473664373159, + 0.004606819245964289, + 0.004479089751839638, + 0.005450327415019274, + 0.004647860303521156, + 0.003708240808919072, + 0.003084647236391902, + 0.0036885912995785475, + 0.0030781226232647896, + 0.005801859777420759, + 0.0029757407028228045, + 0.00603896751999855, + 0.0023726075887680054, + 0.002762898337095976, + 0.0026761149056255817, + 0.004276118706911802, + 0.004921570885926485, + 0.0036048891488462687, + 0.002827106975018978, + 0.0030313057359308004, + 0.0036010940093547106, + 0.003115658415481448, + 0.0031057402957230806, + 0.003059269627556205, + 0.0032327601220458746, + 0.005581013858318329, + 0.005286442581564188, + 0.00469601433724165, + 0.004138103686273098, + 0.002564564347267151, + 0.003655382664874196, + 0.004498525522649288, + 0.0035156053490936756, + 0.003485943889245391, + 0.00422705290839076, + 0.0045166900381445885, + 0.00783534999936819, + 0.006719392258673906, + 0.003958496730774641, + 0.0031931826379150152, + 0.004508094862103462, + 0.002869816729798913, + 0.003376088570803404, + 0.0025330083444714546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.0620780847966671 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a 3D printer's nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle a polished brass cone attached to wires and fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thin, continuous thread of molten vibrant red plastic extruding from the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the nozzle move with swift automation along linear and curvilinear paths over a dark textured print bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright focused LED light creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008173224050551653 + ], + "opposite_logprob_tokens": [ + -7.12581729888916 + ], + "logprob_reward": -0.0008173224050551653, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008173224050551653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12494686990976334, + 0.079445980489254, + 0.1151980385184288, + 0.4093811511993408, + 0.044400639832019806, + 0.06071556359529495, + 0.035112664103507996, + 0.03412356227636337, + 0.046318501234054565, + 0.05035703629255295 + ], + "credit_map": [ + 0.014108510687947273, + 0.012214387767016888, + 0.015595982782542706, + 0.0223867017775774, + 0.012019965797662735, + 0.01674330048263073, + 0.018938373774290085, + 0.012939640320837498, + 0.009053999558091164, + 0.00894375704228878, + 0.012382696382701397, + 0.015317688696086407, + 0.006303071044385433, + 0.007708449382334948, + 0.010127485729753971, + 0.009608825668692589, + 0.013064776547253132, + 0.01354611199349165, + 0.01835184171795845, + 0.01720120571553707, + 0.010712402872741222, + 0.014053788036108017, + 0.015047683380544186, + 0.013220232911407948, + 0.0449296310544014, + 0.053966157138347626, + 0.04972724989056587, + 0.10710447281599045, + 0.051918018609285355, + 0.021305782720446587, + 0.050065185874700546, + 0.030364613980054855, + 0.004775325767695904, + 0.00517027173191309, + 0.006216454319655895, + 0.00710271717980504, + 0.005239819176495075, + 0.00594111206009984, + 0.004444437567144632, + 0.005510504357516766, + 0.010519701056182384, + 0.0063423351384699345, + 0.005887738894671202, + 0.006936178542673588, + 0.004059480037540197, + 0.006574992090463638, + 0.005179711151868105, + 0.015215427614748478, + 0.004235052969306707, + 0.004433697089552879, + 0.0047322604805231094, + 0.006564327050000429, + 0.0031639838125556707, + 0.0029331219848245382, + 0.0048819975927472115, + 0.004168224520981312, + 0.003996503073722124, + 0.003263874677941203, + 0.006551051512360573, + 0.00356793194077909, + 0.004204398021101952, + 0.004410316236317158, + 0.004343200009316206, + 0.0037862882018089294, + 0.0057074339129030704, + 0.005974054802209139, + 0.008088260889053345, + 0.004730517044663429, + 0.005986405536532402, + 0.006865211762487888, + 0.00482918182387948, + 0.004137435927987099, + 0.005950384307652712, + 0.0053392392583191395, + 0.00595821114256978, + 0.004849397111684084, + 0.007971705868840218, + 0.009442935697734356, + 0.006300225388258696, + 0.004544935654848814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.0008173224050551653 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a 3D printer's nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle a polished brass cone attached to wires and fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thin, continuous thread of molten vibrant red plastic extruding from the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the nozzle move with swift automation along linear and curvilinear paths over a dark textured print bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright focused LED light creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002201516181230545 + ], + "opposite_logprob_tokens": [ + -6.127201557159424 + ], + "logprob_reward": -0.002201516181230545, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002201516181230545, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09583211690187454, + 0.0773831233382225, + 0.07341291010379791, + 0.3644399344921112, + 0.06688395142555237, + 0.09279078990221024, + 0.04615072160959244, + 0.04601092264056206, + 0.038935866206884384, + 0.09815963357686996 + ], + "credit_map": [ + 0.010945038869976997, + 0.010117761790752411, + 0.012589254416525364, + 0.014333158731460571, + 0.011387856677174568, + 0.013242492452263832, + 0.01330912671983242, + 0.009907430037856102, + 0.006516918074339628, + 0.009873381815850735, + 0.011128348298370838, + 0.014041692018508911, + 0.0086509445682168, + 0.007767838891595602, + 0.010433229617774487, + 0.008970768190920353, + 0.008520035073161125, + 0.007368326187133789, + 0.012985016219317913, + 0.013483449816703796, + 0.006038821302354336, + 0.00818883627653122, + 0.009482664056122303, + 0.0073457625694572926, + 0.04085102677345276, + 0.04798465967178345, + 0.04527980834245682, + 0.09525186568498611, + 0.04344431683421135, + 0.01848362758755684, + 0.048096850514411926, + 0.02504778653383255, + 0.007421604823321104, + 0.011669390834867954, + 0.007098946254700422, + 0.008328293450176716, + 0.008259751833975315, + 0.010189945809543133, + 0.007098798640072346, + 0.006817216984927654, + 0.012098625302314758, + 0.009873788803815842, + 0.00820117350667715, + 0.00967335607856512, + 0.008784218691289425, + 0.015184406191110611, + 0.009210755117237568, + 0.01976446807384491, + 0.0057194228284060955, + 0.004691743291914463, + 0.0071343607269227505, + 0.010235404595732689, + 0.004749472253024578, + 0.004747715778648853, + 0.004752620123326778, + 0.00411998201161623, + 0.005779251456260681, + 0.00823934469372034, + 0.00851224735379219, + 0.005319152493029833, + 0.0034792316146194935, + 0.005671207327395678, + 0.004377829376608133, + 0.004632661584764719, + 0.005000815726816654, + 0.00576605461537838, + 0.0042501469142735004, + 0.00411600898951292, + 0.006398143712431192, + 0.0038395787123590708, + 0.005602198652923107, + 0.003962916787713766, + 0.013043729588389397, + 0.01155299972742796, + 0.022156264632940292, + 0.010114536620676517, + 0.012248240411281586, + 0.011457499116659164, + 0.00935377087444067, + 0.008232594467699528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.002201516181230545 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a 3D printer's nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle a polished brass cone attached to wires and fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thin, continuous thread of molten vibrant red plastic extruding from the nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the nozzle move with swift automation along linear and curvilinear paths over a dark textured print bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright focused LED light creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00762551836669445 + ], + "opposite_logprob_tokens": [ + -4.882625579833984 + ], + "logprob_reward": -0.00762551836669445, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00762551836669445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08060318231582642, + 0.04317164793610573, + 0.05010820925235748, + 0.5507116913795471, + 0.03562801331281662, + 0.06676268577575684, + 0.058458808809518814, + 0.017135953530669212, + 0.045880917459726334, + 0.05153892561793327 + ], + "credit_map": [ + 0.009370018728077412, + 0.007613642606884241, + 0.010986126959323883, + 0.012314694933593273, + 0.00795740820467472, + 0.009885555133223534, + 0.01353590376675129, + 0.008939831517636776, + 0.0038535487838089466, + 0.00512967724353075, + 0.006410038098692894, + 0.005610746331512928, + 0.006042242981493473, + 0.0049609048292040825, + 0.005070096347481012, + 0.006094398908317089, + 0.004871691111475229, + 0.0063326358795166016, + 0.00662827817723155, + 0.006487521808594465, + 0.005630447529256344, + 0.009833630174398422, + 0.0044661941938102245, + 0.00585781829431653, + 0.06105872988700867, + 0.07888512313365936, + 0.06517378985881805, + 0.1423211693763733, + 0.06428734958171844, + 0.03216841071844101, + 0.06734835356473923, + 0.039468806236982346, + 0.0029504583217203617, + 0.004898448940366507, + 0.004283268470317125, + 0.004540777765214443, + 0.004782993812114, + 0.006288495380431414, + 0.004343317821621895, + 0.0035402539651840925, + 0.007893359288573265, + 0.008856222033500671, + 0.004739813972264528, + 0.007778762839734554, + 0.007881787605583668, + 0.010097973980009556, + 0.007588332984596491, + 0.011926441453397274, + 0.0069006821140646935, + 0.006756631191819906, + 0.0073470380157232285, + 0.013373715803027153, + 0.005788339767605066, + 0.005861589219421148, + 0.0069117676466703415, + 0.0055190506391227245, + 0.0019115917384624481, + 0.001691857585683465, + 0.0026864870451390743, + 0.002141538541764021, + 0.0011579360580071807, + 0.0026259315200150013, + 0.002714148024097085, + 0.002206464996561408, + 0.006660310551524162, + 0.0036068495828658342, + 0.006234365981072187, + 0.005854963790625334, + 0.0075038718059659, + 0.004810776561498642, + 0.005002486985176802, + 0.006207297556102276, + 0.009590615518391132, + 0.007981783710420132, + 0.006182312034070492, + 0.004344261717051268, + 0.009956013411283493, + 0.00533750606700778, + 0.004562645219266415, + 0.0035837930627167225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.00762551836669445 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_12/index_tensor([7487, 7487])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_12/index_tensor([7487, 7487])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..95b1fcd2e40b611158d57f6821f74c9415a57d58 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_12/index_tensor([7487, 7487])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "long shot, shot from a high angle, slowly pulls back to reveal a vast landscape shrouded in torrential rain. Thick, gloomy clouds hang low, almost touching the vast, turbulent surface of the gray water. Raindrops cascade down like thousands of silver needles, stirring up countless broken, rapid ripples on the turbulent surface. In the center of this vast expanse, the only thing in sight is a small, simple wooden rowboat, its faded blue paint peeling and revealing the weathered wood beneath. It is helplessly tossed up and down by the wind and waves, rocking violently with each surge, its bow sometimes lifted high, pointing toward the leaden sky, sometimes plunging into the trough of the waves, its bottom filled with icy rainwater. The entire picture is dominated by a cold, almost monochromatic gray tone. Under the even and oppressive diffuse light, the scene is filled with an extreme loneliness, desolation and a sense of insignificance in the face of natural forces, creating a tranquil yet tense atmosphere.", + "seed": 3197, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_12/index_7487_seed_3197_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.5629380336496979, + "vlm_logprob": -0.5629380336496979, + "vlm_yes_no_margin": 2.55, + "vlm_token_credit": -0.5629380336496979, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal a vast landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, simple wooden rowboat with faded blue paint visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat rocking violently up and down in turbulent gray water under torrential rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cold, monochromatic gray tone with low, thick clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08394119888544083 + ], + "opposite_logprob_tokens": [ + -2.9589412212371826 + ], + "logprob_reward": -0.08394119888544083, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.08394119888544083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1353968381881714, + 0.09898877143859863, + 0.05413459613919258, + 0.05803994834423065, + 0.3173507750034332, + 0.04863441735506058, + 0.04862947016954422, + 0.03688196837902069, + 0.13879352807998657, + 0.06314978003501892 + ], + "credit_map": [ + 0.020457858219742775, + 0.024661535397171974, + 0.01565447263419628, + 0.01123455073684454, + 0.016677264124155045, + 0.016827678307890892, + 0.013133842498064041, + 0.01674962043762207, + 0.009202529676258564, + 0.014122597873210907, + 0.012681896798312664, + 0.010712414048612118, + 0.009360707364976406, + 0.01271054893732071, + 0.018082713708281517, + 0.012115356512367725, + 0.005475230980664492, + 0.006500071845948696, + 0.00759106082841754, + 0.009027594700455666, + 0.004821115639060736, + 0.006904774345457554, + 0.007467918563634157, + 0.006346816662698984, + 0.006725287530571222, + 0.005792025942355394, + 0.007830510847270489, + 0.006853428203612566, + 0.004358671139925718, + 0.00759810721501708, + 0.01068873330950737, + 0.008193176239728928, + 0.029576189815998077, + 0.043801553547382355, + 0.04839659854769707, + 0.03171970695257187, + 0.03211893513798714, + 0.06758050620555878, + 0.024923529475927353, + 0.03923371434211731, + 0.005960458889603615, + 0.006445161532610655, + 0.007248931098729372, + 0.005543256178498268, + 0.004689127206802368, + 0.006624892354011536, + 0.006369845010340214, + 0.005752738565206528, + 0.004886017180979252, + 0.0027584072668105364, + 0.012852544896304607, + 0.00492109265178442, + 0.002909071044996381, + 0.010548285208642483, + 0.003253287635743618, + 0.006500757299363613, + 0.006399816367775202, + 0.0036662581842392683, + 0.00546664372086525, + 0.004287386778742075, + 0.0023295015562325716, + 0.006430422887206078, + 0.0030748380813747644, + 0.005227097775787115, + 0.011800169013440609, + 0.009770713746547699, + 0.03496279940009117, + 0.012110919691622257, + 0.008964408189058304, + 0.01612415723502636, + 0.008018665947020054, + 0.03704167529940605, + 0.010112746618688107, + 0.007223718799650669, + 0.007711587939411402, + 0.009752600453794003, + 0.004743324592709541, + 0.008187868632376194, + 0.0065253726206719875, + 0.00889255665242672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.08394119888544083 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal a vast landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, simple wooden rowboat with faded blue paint visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat rocking violently up and down in turbulent gray water under torrential rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cold, monochromatic gray tone with low, thick clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2236675024032593 + ], + "opposite_logprob_tokens": [ + -0.3486675024032593 + ], + "logprob_reward": -1.2236675024032593, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.2236675024032593, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09770037978887558, + 0.13975147902965546, + 0.04964481294155121, + 0.058707691729068756, + 0.11299294233322144, + 0.18267443776130676, + 0.0425264872610569, + 0.09401247650384903, + 0.16307972371578217, + 0.05890955775976181 + ], + "credit_map": [ + 0.008921902626752853, + 0.013125628232955933, + 0.013499356806278229, + 0.011060069315135479, + 0.009542825631797314, + 0.011011814698576927, + 0.014570721425116062, + 0.01596805453300476, + 0.020677775144577026, + 0.010786047205328941, + 0.026082618162035942, + 0.011831029318273067, + 0.014444965869188309, + 0.018892379477620125, + 0.015428281389176846, + 0.02160838432610035, + 0.00718320719897747, + 0.008697161450982094, + 0.005623830482363701, + 0.004893640521913767, + 0.004413113463670015, + 0.006373619195073843, + 0.004674089141190052, + 0.007786151487380266, + 0.005996689200401306, + 0.009955759160220623, + 0.006848359014838934, + 0.007108519319444895, + 0.006941457744687796, + 0.0073201097548007965, + 0.007282373029738665, + 0.007254428695887327, + 0.01524021103978157, + 0.015253881923854351, + 0.014124720357358456, + 0.009981865994632244, + 0.013316982425749302, + 0.020773479714989662, + 0.010185953229665756, + 0.014115848578512669, + 0.028202611953020096, + 0.01758302003145218, + 0.025449693202972412, + 0.020871268585324287, + 0.018595244735479355, + 0.02656605839729309, + 0.0241196658462286, + 0.02128687873482704, + 0.007988627068698406, + 0.0035103652626276016, + 0.007229714188724756, + 0.00490502268075943, + 0.0034540314227342606, + 0.005663331598043442, + 0.003768022870644927, + 0.006007371470332146, + 0.01344218384474516, + 0.01778620295226574, + 0.01712755858898163, + 0.011637559160590172, + 0.004898549988865852, + 0.007831013761460781, + 0.007862421683967113, + 0.013426992110908031, + 0.009794873185455799, + 0.006974530406296253, + 0.05310520529747009, + 0.009998200461268425, + 0.005933643784373999, + 0.009759629145264626, + 0.009293420240283012, + 0.058220215141773224, + 0.008374830707907677, + 0.004707803018391132, + 0.0084874602034688, + 0.01012322399765253, + 0.005057556089013815, + 0.0064273374155163765, + 0.006161530967801809, + 0.00956981722265482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 1 + ], + "selected_reward": -1.2236675024032593 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal a vast landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, simple wooden rowboat with faded blue paint visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat rocking violently up and down in turbulent gray water under torrential rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cold, monochromatic gray tone with low, thick clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002501931507140398 + ], + "opposite_logprob_tokens": [ + -6.002501964569092 + ], + "logprob_reward": -0.002501931507140398, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002501931507140398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16952912509441376, + 0.11716447025537491, + 0.04444734752178192, + 0.03676946088671684, + 0.11401675641536713, + 0.134042888879776, + 0.09698932617902756, + 0.07846248894929886, + 0.06620260328054428, + 0.14237557351589203 + ], + "credit_map": [ + 0.013131028972566128, + 0.030217768624424934, + 0.019054578617215157, + 0.01360491942614317, + 0.021168841049075127, + 0.020463120192289352, + 0.02780916914343834, + 0.02407967671751976, + 0.010203177109360695, + 0.018172187730669975, + 0.017264004796743393, + 0.014588880352675915, + 0.009280216880142689, + 0.012999245896935463, + 0.01820523850619793, + 0.016451504081487656, + 0.00505566131323576, + 0.005399900022894144, + 0.005885268561542034, + 0.0054110195487737656, + 0.004254693631082773, + 0.005844419822096825, + 0.0070733665488660336, + 0.005523016210645437, + 0.0031458954326808453, + 0.003957059234380722, + 0.00518430769443512, + 0.005166969727724791, + 0.003629149403423071, + 0.005777440033853054, + 0.004492802079766989, + 0.005415833089500666, + 0.009438207373023033, + 0.0194268636405468, + 0.015382633544504642, + 0.012418925762176514, + 0.014043617993593216, + 0.018236292526125908, + 0.012162099592387676, + 0.012908113189041615, + 0.016678456217050552, + 0.0126876849681139, + 0.016810381785035133, + 0.0147706875577569, + 0.015887299552559853, + 0.021830275654792786, + 0.018094094470143318, + 0.01728399470448494, + 0.012505916878581047, + 0.0084605161100626, + 0.016443893313407898, + 0.015845593065023422, + 0.009233640506863594, + 0.012367979623377323, + 0.011297033168375492, + 0.010834749788045883, + 0.009827232919633389, + 0.012363777495920658, + 0.017161954194307327, + 0.011450467631220818, + 0.0035463133826851845, + 0.006031807977706194, + 0.00583831500262022, + 0.012242612428963184, + 0.005744331981986761, + 0.006568683311343193, + 0.014775593765079975, + 0.009155584499239922, + 0.006698666140437126, + 0.005449303891509771, + 0.005784658249467611, + 0.012025773525238037, + 0.014442339539527893, + 0.018832005560398102, + 0.01594703271985054, + 0.015326992608606815, + 0.028166649863123894, + 0.016165077686309814, + 0.018684927374124527, + 0.014810539782047272 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.002501931507140398 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal a vast landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, simple wooden rowboat with faded blue paint visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat rocking violently up and down in turbulent gray water under torrential rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cold, monochromatic gray tone with low, thick clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.5020873546600342 + ], + "opposite_logprob_tokens": [ + -0.25208738446235657 + ], + "logprob_reward": -1.5020873546600342, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.5020873546600342, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3340521454811096, + 0.0914185643196106, + 0.06980293244123459, + 0.04846150800585747, + 0.11189733445644379, + 0.048607949167490005, + 0.06228914111852646, + 0.06556805968284607, + 0.05816780403256416, + 0.10973455011844635 + ], + "credit_map": [ + 0.052695732563734055, + 0.05398537591099739, + 0.03768221661448479, + 0.03190234676003456, + 0.03873910382390022, + 0.0417952686548233, + 0.04363878816366196, + 0.03361329808831215, + 0.008378077298402786, + 0.009593037888407707, + 0.018193190917372704, + 0.0089586665853858, + 0.006933321710675955, + 0.010600920766592026, + 0.013710537925362587, + 0.015050812624394894, + 0.008465006947517395, + 0.009730097837746143, + 0.008392532356083393, + 0.00788157619535923, + 0.005045413505285978, + 0.009778957813978195, + 0.010093914344906807, + 0.010415437631309032, + 0.005346855614334345, + 0.003897793823853135, + 0.007440692745149136, + 0.005573933012783527, + 0.003352668136358261, + 0.005708850454539061, + 0.009975613094866276, + 0.007165100425481796, + 0.016645537689328194, + 0.009544051252305508, + 0.019735893234610558, + 0.007256084121763706, + 0.008886619471013546, + 0.027458293363451958, + 0.006980228703469038, + 0.015390622429549694, + 0.00596966315060854, + 0.003446359420195222, + 0.008125842548906803, + 0.005384786054491997, + 0.004603242035955191, + 0.008958827704191208, + 0.005062153562903404, + 0.007057076785713434, + 0.01090478990226984, + 0.004217290319502354, + 0.012917378917336464, + 0.007465003058314323, + 0.0032463292591273785, + 0.010892919264733791, + 0.004352585878223181, + 0.008292845450341702, + 0.008957739919424057, + 0.012125344015657902, + 0.013256956823170185, + 0.008762349374592304, + 0.0027202486526221037, + 0.004550405777990818, + 0.0052650305442512035, + 0.009929982014000416, + 0.0035931544844061136, + 0.0024064225144684315, + 0.019201768562197685, + 0.00377800315618515, + 0.00253349170088768, + 0.004211076535284519, + 0.00253326166421175, + 0.019910624250769615, + 0.023382078856229782, + 0.01383991725742817, + 0.0158433485776186, + 0.01445556990802288, + 0.006359564606100321, + 0.014267328195273876, + 0.008935105055570602, + 0.012651629745960236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -1.5020873546600342 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal a vast landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, simple wooden rowboat with faded blue paint visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat rocking violently up and down in turbulent gray water under torrential rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cold, monochromatic gray tone with low, thick clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024921807926148176 + ], + "opposite_logprob_tokens": [ + -6.0024919509887695 + ], + "logprob_reward": -0.0024921807926148176, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024921807926148176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21796491742134094, + 0.05605470389127731, + 0.07612085342407227, + 0.04717858508229256, + 0.19374489784240723, + 0.06830891966819763, + 0.04572661221027374, + 0.04595435783267021, + 0.13491733372211456, + 0.11402887850999832 + ], + "credit_map": [ + 0.03568631783127785, + 0.03531740605831146, + 0.026328569278120995, + 0.022169414907693863, + 0.023178042843937874, + 0.027368247509002686, + 0.027454275637865067, + 0.020462656393647194, + 0.006629535928368568, + 0.004651913419365883, + 0.00914322305470705, + 0.004378353711217642, + 0.004617779050022364, + 0.007721554022282362, + 0.009474670514464378, + 0.009437678381800652, + 0.012651975266635418, + 0.011832318268716335, + 0.007567062973976135, + 0.008284653536975384, + 0.00589881232008338, + 0.011416424997150898, + 0.011392198503017426, + 0.007077411748468876, + 0.005640197545289993, + 0.003989755641669035, + 0.006448359694331884, + 0.005107331555336714, + 0.0037884737830609083, + 0.006597837433218956, + 0.009996200911700726, + 0.005610430613160133, + 0.028119191527366638, + 0.01596697047352791, + 0.03540824353694916, + 0.014204210601747036, + 0.015795910730957985, + 0.04539942368865013, + 0.011129037477076054, + 0.027721907943487167, + 0.010978064499795437, + 0.006418219767510891, + 0.00923879537731409, + 0.007996766828000546, + 0.005218698177486658, + 0.012130534276366234, + 0.004071623086929321, + 0.012256218120455742, + 0.00893382914364338, + 0.0027729300782084465, + 0.00765865808352828, + 0.006603880785405636, + 0.0025497181341052055, + 0.008620894514024258, + 0.003358482616022229, + 0.0052282195538282394, + 0.0074655949138104916, + 0.0032981436233967543, + 0.006218781694769859, + 0.005623810924589634, + 0.0028066597878932953, + 0.010099825449287891, + 0.004187359474599361, + 0.006254180334508419, + 0.008184549398720264, + 0.006956183817237616, + 0.04395353049039841, + 0.009573020040988922, + 0.006043865345418453, + 0.009405902586877346, + 0.005907710641622543, + 0.0448925644159317, + 0.025211350992321968, + 0.01376370806246996, + 0.01674581877887249, + 0.013656140305101871, + 0.007851038128137589, + 0.016124743968248367, + 0.008344202302396297, + 0.012331877835094929 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.0024921807926148176 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "long shot, shot from a high angle, slowly pulls back to reveal a vast landscape shrouded in torrential rain. Thick, gloomy clouds hang low, almost touching the vast, turbulent surface of the gray water. Raindrops cascade down like thousands of silver needles, stirring up countless broken, rapid ripples on the turbulent surface. In the center of this vast expanse, the only thing in sight is a small, simple wooden rowboat, its faded blue paint peeling and revealing the weathered wood beneath. It is helplessly tossed up and down by the wind and waves, rocking violently with each surge, its bow sometimes lifted high, pointing toward the leaden sky, sometimes plunging into the trough of the waves, its bottom filled with icy rainwater. The entire picture is dominated by a cold, almost monochromatic gray tone. Under the even and oppressive diffuse light, the scene is filled with an extreme loneliness, desolation and a sense of insignificance in the face of natural forces, creating a tranquil yet tense atmosphere.", + "seed": 3198, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_12/index_7487_seed_3198_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.4281055086059496, + "vlm_logprob": -0.4281055086059496, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.4281055086059496, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal a vast landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, simple wooden rowboat with faded blue paint visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat rocking violently up and down in turbulent gray water under torrential rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cold, monochromatic gray tone with low, thick clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04774802550673485 + ], + "opposite_logprob_tokens": [ + -3.54774808883667 + ], + "logprob_reward": -0.04774802550673485, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.04774802550673485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12414202094078064, + 0.06279148161411285, + 0.04494418948888779, + 0.12427928298711777, + 0.13034777343273163, + 0.07682943344116211, + 0.18497838079929352, + 0.09598884731531143, + 0.11657025665044785, + 0.03912826627492905 + ], + "credit_map": [ + 0.020985156297683716, + 0.010149143636226654, + 0.025855185464024544, + 0.012137806974351406, + 0.012694870121777058, + 0.01914217509329319, + 0.008327396586537361, + 0.014850301668047905, + 0.010635941289365292, + 0.01740015298128128, + 0.005883040837943554, + 0.005661061033606529, + 0.005874426104128361, + 0.007348301354795694, + 0.005504337605088949, + 0.004484216216951609, + 0.004405598156154156, + 0.005327160470187664, + 0.006814201828092337, + 0.006419340148568153, + 0.0035744670312851667, + 0.009091103449463844, + 0.004720563534647226, + 0.004591757897287607, + 0.01579650118947029, + 0.02417038381099701, + 0.016242561861872673, + 0.009773975238204002, + 0.011514279060065746, + 0.02137894555926323, + 0.012209181673824787, + 0.013193449005484581, + 0.01335057057440281, + 0.016955344006419182, + 0.019515695050358772, + 0.017511624842882156, + 0.013205239549279213, + 0.024035640060901642, + 0.011477790772914886, + 0.014295878820121288, + 0.009614839218556881, + 0.005644248332828283, + 0.005287279840558767, + 0.003996170125901699, + 0.026160547509789467, + 0.015381809324026108, + 0.005361189134418964, + 0.005383347161114216, + 0.050449199974536896, + 0.005725065246224403, + 0.019380807876586914, + 0.01051448192447424, + 0.015720408409833908, + 0.06353836506605148, + 0.006104732863605022, + 0.01354531291872263, + 0.015116711147129536, + 0.005020607262849808, + 0.014670121483504772, + 0.00701715424656868, + 0.011899897828698158, + 0.022462069988250732, + 0.004805002827197313, + 0.014997282065451145, + 0.01076692808419466, + 0.011573546566069126, + 0.01464781817048788, + 0.01685905270278454, + 0.013326977379620075, + 0.013272801414132118, + 0.018926164135336876, + 0.017196975648403168, + 0.004301136359572411, + 0.004824927542358637, + 0.006365133915096521, + 0.0037782134022563696, + 0.003978025633841753, + 0.006320097949355841, + 0.0043328809551894665, + 0.005227853078395128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 3 + ], + "selected_reward": -0.04774802550673485 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal a vast landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, simple wooden rowboat with faded blue paint visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat rocking violently up and down in turbulent gray water under torrential rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cold, monochromatic gray tone with low, thick clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38710513710975647 + ], + "opposite_logprob_tokens": [ + -1.137105107307434 + ], + "logprob_reward": -0.38710513710975647, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38710513710975647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14158068597316742, + 0.1290828436613083, + 0.07665995508432388, + 0.057826776057481766, + 0.1125873327255249, + 0.12147244065999985, + 0.1323796510696411, + 0.08789637684822083, + 0.12596823275089264, + 0.014545667916536331 + ], + "credit_map": [ + 0.030661270022392273, + 0.01122861634939909, + 0.019257569685578346, + 0.009603002108633518, + 0.012759581208229065, + 0.03234768286347389, + 0.00893264077603817, + 0.016790315508842468, + 0.013475134037435055, + 0.03278233855962753, + 0.014996709302067757, + 0.011353938840329647, + 0.011080190539360046, + 0.025787096470594406, + 0.010846257209777832, + 0.008761189877986908, + 0.006966947577893734, + 0.00934581644833088, + 0.010312716476619244, + 0.013064753264188766, + 0.005995552055537701, + 0.014485403895378113, + 0.009290417656302452, + 0.007198347244411707, + 0.003909460734575987, + 0.006428409367799759, + 0.007765060756355524, + 0.005575146526098251, + 0.00670595420524478, + 0.007695706095546484, + 0.008208032697439194, + 0.011539007537066936, + 0.008101207204163074, + 0.0158047117292881, + 0.016909100115299225, + 0.013893959112465382, + 0.014233832247555256, + 0.015767494216561317, + 0.012719358317553997, + 0.015157666057348251, + 0.009901508688926697, + 0.013025794178247452, + 0.013962547294795513, + 0.015000980347394943, + 0.025090733543038368, + 0.021343709900975227, + 0.010521541349589825, + 0.012625621631741524, + 0.03469610959291458, + 0.005148435477167368, + 0.014463747851550579, + 0.009673859924077988, + 0.011982073076069355, + 0.04047315940260887, + 0.006225403863936663, + 0.009716862812638283, + 0.014297607354819775, + 0.0028125890530645847, + 0.009818601422011852, + 0.005051831249147654, + 0.013307195156812668, + 0.02971905842423439, + 0.0033093036618083715, + 0.009580188430845737, + 0.01197521947324276, + 0.014036846347153187, + 0.015398915857076645, + 0.01692456379532814, + 0.017881199717521667, + 0.018097050487995148, + 0.016729505732655525, + 0.014924930408596992, + 0.0013198282103985548, + 0.0016862839693203568, + 0.00198354572057724, + 0.0015651570865884423, + 0.0018236105097457767, + 0.0018585744546726346, + 0.002041821600869298, + 0.0022668465971946716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.38710513710975647 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal a vast landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, simple wooden rowboat with faded blue paint visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat rocking violently up and down in turbulent gray water under torrential rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cold, monochromatic gray tone with low, thick clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019481267081573606 + ], + "opposite_logprob_tokens": [ + -6.251948356628418 + ], + "logprob_reward": -0.0019481267081573606, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019481267081573606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18638813495635986, + 0.11138360947370529, + 0.04423453286290169, + 0.03788166493177414, + 0.054327335208654404, + 0.07094484567642212, + 0.17883656919002533, + 0.22840794920921326, + 0.0545937716960907, + 0.0330016203224659 + ], + "credit_map": [ + 0.03705357387661934, + 0.01245657354593277, + 0.024326492100954056, + 0.012288274243474007, + 0.01777685433626175, + 0.04703062027692795, + 0.011923374608159065, + 0.023532360792160034, + 0.00710904598236084, + 0.03285238519310951, + 0.006975342985242605, + 0.004919710569083691, + 0.008699130266904831, + 0.03724636510014534, + 0.008163684979081154, + 0.005417938809841871, + 0.0033110473304986954, + 0.005542458035051823, + 0.004984984174370766, + 0.006466899067163467, + 0.0033984328620135784, + 0.0077720703557133675, + 0.0059745945036411285, + 0.006784046534448862, + 0.003263435559347272, + 0.0052854386158287525, + 0.0043015326373279095, + 0.0036293528974056244, + 0.003656362649053335, + 0.005290850065648556, + 0.005528387147933245, + 0.006926302798092365, + 0.0043594492599368095, + 0.006553693674504757, + 0.007297595497220755, + 0.007821832783520222, + 0.0048370216973125935, + 0.008698214776813984, + 0.0073945787735283375, + 0.007364948280155659, + 0.008426787331700325, + 0.00655224546790123, + 0.006305601913481951, + 0.005008529406040907, + 0.02220511995255947, + 0.013142023235559464, + 0.0037574099842458963, + 0.0055471304804086685, + 0.05127178877592087, + 0.007675644475966692, + 0.00975659117102623, + 0.006587045732885599, + 0.017512744292616844, + 0.07087292522192001, + 0.006456730421632528, + 0.00870310515165329, + 0.06311080604791641, + 0.007601890247315168, + 0.01132870838046074, + 0.006290379911661148, + 0.02786076068878174, + 0.09516626596450806, + 0.007523795589804649, + 0.009525339119136333, + 0.005237526725977659, + 0.006345108617097139, + 0.006583850830793381, + 0.006705205887556076, + 0.007735099643468857, + 0.00813053548336029, + 0.007422879338264465, + 0.006433561909943819, + 0.0034709894098341465, + 0.003959125839173794, + 0.004229147452861071, + 0.003323171753436327, + 0.0038512179162353277, + 0.004580146167427301, + 0.005399252288043499, + 0.004188567865639925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.0019481267081573606 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal a vast landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, simple wooden rowboat with faded blue paint visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat rocking violently up and down in turbulent gray water under torrential rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cold, monochromatic gray tone with low, thick clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.7015260457992554 + ], + "opposite_logprob_tokens": [ + -0.2015260010957718 + ], + "logprob_reward": -1.7015260457992554, + "yes_no_margin_reward": -1.5, + "token_credit_reward": -1.7015260457992554, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1021515354514122, + 0.24523259699344635, + 0.02215530350804329, + 0.026458192616701126, + 0.06822903454303741, + 0.06271928548812866, + 0.27644193172454834, + 0.12756206095218658, + 0.024020537734031677, + 0.045029520988464355 + ], + "credit_map": [ + 0.02466854266822338, + 0.005054428707808256, + 0.009809831157326698, + 0.006041139364242554, + 0.0067738196812570095, + 0.03170200437307358, + 0.006541297771036625, + 0.011560481041669846, + 0.02351491153240204, + 0.083467036485672, + 0.015150124207139015, + 0.010436180047690868, + 0.019154537469148636, + 0.07521682977676392, + 0.009319324977695942, + 0.008973675779998302, + 0.0034539715852588415, + 0.004569641780108213, + 0.002257703337818384, + 0.0026377979665994644, + 0.0014522845158353448, + 0.003518622135743499, + 0.001846078783273697, + 0.002419204916805029, + 0.0023692871909588575, + 0.0031193941831588745, + 0.004076316952705383, + 0.002240436850115657, + 0.0029034342151135206, + 0.005634333938360214, + 0.0031509711407124996, + 0.002964018378406763, + 0.006662388332188129, + 0.007790979463607073, + 0.014252937398850918, + 0.008002975955605507, + 0.007870105095207691, + 0.008450884371995926, + 0.005742266774177551, + 0.009456498548388481, + 0.006729515735059977, + 0.007355332374572754, + 0.008616958744823933, + 0.008762937970459461, + 0.007249215617775917, + 0.00799336563795805, + 0.009122745133936405, + 0.006889214273542166, + 0.0847715362906456, + 0.005861932411789894, + 0.015803296118974686, + 0.00655145850032568, + 0.02796156331896782, + 0.11661804467439651, + 0.0050199213437736034, + 0.013854176737368107, + 0.030998457223176956, + 0.004462513606995344, + 0.013437829911708832, + 0.006612370256334543, + 0.01115765143185854, + 0.04435264319181442, + 0.003948465455323458, + 0.012592121958732605, + 0.0026124012656509876, + 0.004461736883968115, + 0.002867165021598339, + 0.002623105188831687, + 0.0028841288294643164, + 0.0031093081925064325, + 0.002579870168119669, + 0.0028828224167227745, + 0.004202413372695446, + 0.004482975695282221, + 0.0086866719648242, + 0.004166020546108484, + 0.003972710110247135, + 0.006224762182682753, + 0.005319458898156881, + 0.00797451101243496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 7 + ], + "selected_reward": -1.7015260457992554 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal a vast landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, simple wooden rowboat with faded blue paint visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat rocking violently up and down in turbulent gray water under torrential rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cold, monochromatic gray tone with low, thick clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002200207905843854 + ], + "opposite_logprob_tokens": [ + -6.127200126647949 + ], + "logprob_reward": -0.002200207905843854, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002200207905843854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11209304630756378, + 0.3393743336200714, + 0.026547517627477646, + 0.03562711179256439, + 0.1188059002161026, + 0.0659632682800293, + 0.11147405952215195, + 0.08548873662948608, + 0.04969799891114235, + 0.054927997291088104 + ], + "credit_map": [ + 0.021065304055809975, + 0.006422333884984255, + 0.013826428912580013, + 0.00856201071292162, + 0.00784008577466011, + 0.024921447038650513, + 0.010148157365620136, + 0.019307279959321022, + 0.03267987072467804, + 0.11950550228357315, + 0.01916315034031868, + 0.01348357554525137, + 0.024413950741291046, + 0.10678253322839737, + 0.011043023318052292, + 0.01230276096612215, + 0.003042205236852169, + 0.0031094984151422977, + 0.004786375444382429, + 0.003711309749633074, + 0.0018199763726443052, + 0.0036787751596421003, + 0.002812290331348777, + 0.0035870857536792755, + 0.0033694463782012463, + 0.0038473468739539385, + 0.0052406927570700645, + 0.003423345973715186, + 0.003914828412234783, + 0.006385237909853458, + 0.0047096447087824345, + 0.0047365641221404076, + 0.011774541810154915, + 0.014569956809282303, + 0.025793761014938354, + 0.012877516448497772, + 0.014123943634331226, + 0.014614559710025787, + 0.008693451061844826, + 0.01635817438364029, + 0.006686318200081587, + 0.007082759402692318, + 0.008693231269717216, + 0.008143731392920017, + 0.00909597147256136, + 0.01073356345295906, + 0.008835515938699245, + 0.006692173890769482, + 0.032236676663160324, + 0.002600224455818534, + 0.007790424861013889, + 0.003828129731118679, + 0.011067886836826801, + 0.04495847225189209, + 0.0024055615067481995, + 0.006586691830307245, + 0.019842198118567467, + 0.0036967829801142216, + 0.007160732056945562, + 0.004626455716788769, + 0.010840343311429024, + 0.030189042910933495, + 0.002806804608553648, + 0.006326381582766771, + 0.004499162081629038, + 0.00672516506165266, + 0.004523489158600569, + 0.006159448530524969, + 0.009608645923435688, + 0.006232875864952803, + 0.0072546410374343395, + 0.004694570321589708, + 0.005506730638444424, + 0.005676314700394869, + 0.00957691390067339, + 0.00478052394464612, + 0.004956998396664858, + 0.007944419048726559, + 0.006284120958298445, + 0.010201975703239441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.002200207905843854 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "long shot, shot from a high angle, slowly pulls back to reveal a vast landscape shrouded in torrential rain. Thick, gloomy clouds hang low, almost touching the vast, turbulent surface of the gray water. Raindrops cascade down like thousands of silver needles, stirring up countless broken, rapid ripples on the turbulent surface. In the center of this vast expanse, the only thing in sight is a small, simple wooden rowboat, its faded blue paint peeling and revealing the weathered wood beneath. It is helplessly tossed up and down by the wind and waves, rocking violently with each surge, its bow sometimes lifted high, pointing toward the leaden sky, sometimes plunging into the trough of the waves, its bottom filled with icy rainwater. The entire picture is dominated by a cold, almost monochromatic gray tone. Under the even and oppressive diffuse light, the scene is filled with an extreme loneliness, desolation and a sense of insignificance in the face of natural forces, creating a tranquil yet tense atmosphere.", + "seed": 3199, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_12/index_7487_seed_3199_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2432161617442034, + "vlm_logprob": -0.2432161617442034, + "vlm_yes_no_margin": 3.549999988079071, + "vlm_token_credit": -0.2432161617442034, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal a vast landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, simple wooden rowboat with faded blue paint visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat rocking violently up and down in turbulent gray water under torrential rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cold, monochromatic gray tone with low, thick clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.068906269967556 + ], + "opposite_logprob_tokens": [ + -3.193906307220459 + ], + "logprob_reward": -0.068906269967556, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.068906269967556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14956365525722504, + 0.05672590062022209, + 0.039253927767276764, + 0.13895779848098755, + 0.021542608737945557, + 0.08163435012102127, + 0.05036884546279907, + 0.0945444330573082, + 0.2233067750930786, + 0.14410170912742615 + ], + "credit_map": [ + 0.00888278428465128, + 0.021664835512638092, + 0.013507122173905373, + 0.007768701296299696, + 0.03328888490796089, + 0.0395873598754406, + 0.010999063961207867, + 0.013864904642105103, + 0.003743716748431325, + 0.009950222447514534, + 0.0042876871302723885, + 0.00404834421351552, + 0.013947495259344578, + 0.010413304902613163, + 0.0045253559947013855, + 0.005809776484966278, + 0.002418204676359892, + 0.008390679955482483, + 0.004322106949985027, + 0.0018203514628112316, + 0.00821745116263628, + 0.007096522022038698, + 0.0030549063812941313, + 0.003933703526854515, + 0.01107969880104065, + 0.015074338763952255, + 0.013909400440752506, + 0.009777662344276905, + 0.024014882743358612, + 0.0421975776553154, + 0.00803898461163044, + 0.01486525684595108, + 0.0021636353339999914, + 0.0018034414388239384, + 0.0027586270589381456, + 0.0032139967661350965, + 0.0029529412277042866, + 0.0030180474277585745, + 0.002020476618781686, + 0.003611444029957056, + 0.0074537950567901134, + 0.010254843160510063, + 0.008986320346593857, + 0.010816341266036034, + 0.01654975861310959, + 0.012497315183281898, + 0.007612084969878197, + 0.007463896181434393, + 0.0045183273032307625, + 0.006656469777226448, + 0.004121960140764713, + 0.004439276643097401, + 0.010188035666942596, + 0.008028445765376091, + 0.004716524388641119, + 0.007699806243181229, + 0.008861435577273369, + 0.015521815977990627, + 0.01238070148974657, + 0.009098869748413563, + 0.011548995971679688, + 0.014550791122019291, + 0.010754795745015144, + 0.011827025562524796, + 0.023106908425688744, + 0.03935081511735916, + 0.02055221050977707, + 0.015195326879620552, + 0.03987453505396843, + 0.060232121497392654, + 0.010116508230566978, + 0.014878335408866405, + 0.013523288071155548, + 0.019517160952091217, + 0.0202624648809433, + 0.01595378667116165, + 0.017286168411374092, + 0.020966915413737297, + 0.014521918259561062, + 0.02206999994814396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.068906269967556 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal a vast landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, simple wooden rowboat with faded blue paint visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat rocking violently up and down in turbulent gray water under torrential rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cold, monochromatic gray tone with low, thick clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38710886240005493 + ], + "opposite_logprob_tokens": [ + -1.1371088027954102 + ], + "logprob_reward": -0.38710886240005493, + "yes_no_margin_reward": 0.7499999403953552, + "token_credit_reward": -0.38710886240005493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05658678710460663, + 0.33983322978019714, + 0.015997156500816345, + 0.025395892560482025, + 0.05495164915919304, + 0.03315337747335434, + 0.09406369179487228, + 0.278896301984787, + 0.08283091336488724, + 0.01829109713435173 + ], + "credit_map": [ + 0.004309036768972874, + 0.008179952390491962, + 0.002733382163569331, + 0.0023890172597020864, + 0.011344649828970432, + 0.017949819564819336, + 0.003148718737065792, + 0.006532203406095505, + 0.009160162881016731, + 0.13263742625713348, + 0.004535925108939409, + 0.004234427120536566, + 0.15547239780426025, + 0.02299269661307335, + 0.0056231762282550335, + 0.005176962353289127, + 0.001263390644453466, + 0.0017914599739015102, + 0.003178373444825411, + 0.0011619203723967075, + 0.0025478086899966, + 0.0024153224658221006, + 0.0016820495948195457, + 0.0019568302668631077, + 0.0018916045082733035, + 0.004619081970304251, + 0.0028485083021223545, + 0.001708847121335566, + 0.0060666464269161224, + 0.0037590188439935446, + 0.0019518821500241756, + 0.002550299046561122, + 0.0033508287742733955, + 0.0061728195287287235, + 0.006432652473449707, + 0.008130504749715328, + 0.006473866757005453, + 0.007565688341856003, + 0.011127250269055367, + 0.005698027089238167, + 0.003383530071005225, + 0.0026627795305103064, + 0.0034935742150992155, + 0.004506674129515886, + 0.00378181179985404, + 0.0056940410286188126, + 0.004346254747360945, + 0.00528470566496253, + 0.011221729218959808, + 0.014684956520795822, + 0.007478329353034496, + 0.005881528370082378, + 0.020545171573758125, + 0.018194958567619324, + 0.005770559888333082, + 0.010286444798111916, + 0.02453422360122204, + 0.039585571736097336, + 0.016208460554480553, + 0.02264503575861454, + 0.0694689154624939, + 0.0479547418653965, + 0.02526709996163845, + 0.033232226967811584, + 0.008768859319388866, + 0.012954316101968288, + 0.008242415264248848, + 0.00523033831268549, + 0.01706293411552906, + 0.020378321409225464, + 0.003950593993067741, + 0.006243137177079916, + 0.001444299123249948, + 0.0025077443569898605, + 0.0019675809890031815, + 0.0017047591973096132, + 0.004168252926319838, + 0.0023523359559476376, + 0.00243376474827528, + 0.0017123586731031537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 6 + ], + "selected_reward": -0.38710886240005493 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal a vast landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, simple wooden rowboat with faded blue paint visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat rocking violently up and down in turbulent gray water under torrential rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cold, monochromatic gray tone with low, thick clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017229963559657335 + ], + "opposite_logprob_tokens": [ + -6.376722812652588 + ], + "logprob_reward": -0.0017229963559657335, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017229963559657335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060236282646656036, + 0.3467314541339874, + 0.033208221197128296, + 0.045858994126319885, + 0.09672172367572784, + 0.03311995044350624, + 0.09047529101371765, + 0.17651435732841492, + 0.07318592816591263, + 0.04394775629043579 + ], + "credit_map": [ + 0.004422198049724102, + 0.0072062755934894085, + 0.004505516029894352, + 0.0033607359509915113, + 0.0127700911834836, + 0.015966078266501427, + 0.003914271481335163, + 0.008091118186712265, + 0.009391190484166145, + 0.13420537114143372, + 0.004767369478940964, + 0.004829851910471916, + 0.15848885476589203, + 0.02417435683310032, + 0.005848052445799112, + 0.005026441998779774, + 0.002332195872440934, + 0.005333649460226297, + 0.004016298335045576, + 0.002427044091746211, + 0.007704649120569229, + 0.004509711172431707, + 0.0032663450110703707, + 0.0036183283664286137, + 0.0029950395692139864, + 0.010023077018558979, + 0.0040492406114935875, + 0.0030290165450423956, + 0.012227174825966358, + 0.0077512688003480434, + 0.0020309756509959698, + 0.0037531994748860598, + 0.006061984691768885, + 0.009999561123549938, + 0.010316413827240467, + 0.016820436343550682, + 0.010612854734063148, + 0.010640867054462433, + 0.021858779713511467, + 0.010410825721919537, + 0.0031001241877675056, + 0.00344887119717896, + 0.003622347256168723, + 0.004683355800807476, + 0.005127802025526762, + 0.004989007953554392, + 0.0037308349274098873, + 0.004417606629431248, + 0.01087269652634859, + 0.014392097480595112, + 0.007403305731713772, + 0.005680972710251808, + 0.019314618781208992, + 0.018035434186458588, + 0.005427376367151737, + 0.009348791092634201, + 0.013470135629177094, + 0.023994067683815956, + 0.013655142858624458, + 0.015819016844034195, + 0.04316424950957298, + 0.0231022946536541, + 0.019651900976896286, + 0.023657552897930145, + 0.007470529060810804, + 0.01159270852804184, + 0.006886870600283146, + 0.004466061946004629, + 0.014619634486734867, + 0.017472580075263977, + 0.004330047871917486, + 0.006347497459501028, + 0.0030125053599476814, + 0.006655220873653889, + 0.0046338909305632114, + 0.0043313452042639256, + 0.009870230220258236, + 0.005188978277146816, + 0.006219937466084957, + 0.004035646095871925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 4 + ], + "selected_reward": -0.0017229963559657335 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal a vast landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, simple wooden rowboat with faded blue paint visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat rocking violently up and down in turbulent gray water under torrential rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cold, monochromatic gray tone with low, thick clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578477263450623 + ], + "opposite_logprob_tokens": [ + -0.6328477263450623 + ], + "logprob_reward": -0.7578477263450623, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578477263450623, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.21883152425289154, + 0.08529612421989441, + 0.05213943123817444, + 0.04424213245511055, + 0.06853272765874863, + 0.1716643124818802, + 0.04455037787556648, + 0.21993878483772278, + 0.052893999963998795, + 0.04191054403781891 + ], + "credit_map": [ + 0.012965177185833454, + 0.04395285248756409, + 0.014959198422729969, + 0.014371390454471111, + 0.05431993678212166, + 0.03912336751818657, + 0.015265018679201603, + 0.02387457899749279, + 0.004358354955911636, + 0.021535873413085938, + 0.005551383830606937, + 0.0035873232409358025, + 0.02543608285486698, + 0.011194393038749695, + 0.00601227767765522, + 0.007620436139404774, + 0.003585339756682515, + 0.0072674863040447235, + 0.008398856967687607, + 0.0027540544979274273, + 0.009785967878997326, + 0.00939156673848629, + 0.0036758719943463802, + 0.007280291989445686, + 0.003109110752120614, + 0.007043869234621525, + 0.003464090870693326, + 0.00341469282284379, + 0.011691133491694927, + 0.009641620330512524, + 0.002552023157477379, + 0.0033255943562835455, + 0.0056754290126264095, + 0.005336874630302191, + 0.011526520363986492, + 0.009333958849310875, + 0.00785645842552185, + 0.01230941154062748, + 0.007168634328991175, + 0.009325448423624039, + 0.016115134581923485, + 0.014926048927009106, + 0.02175002172589302, + 0.025032488629221916, + 0.01905316486954689, + 0.02654237300157547, + 0.018010864034295082, + 0.03023422881960869, + 0.004718686453998089, + 0.005985605996102095, + 0.006038750521838665, + 0.004391247406601906, + 0.008443298749625683, + 0.007967681623995304, + 0.003167093265801668, + 0.0038380222395062447, + 0.021705815568566322, + 0.03804488107562065, + 0.013170932419598103, + 0.014160865917801857, + 0.050619009882211685, + 0.051927801221609116, + 0.010408715344965458, + 0.01990078017115593, + 0.005739448592066765, + 0.010235180146992207, + 0.004400238394737244, + 0.003896949579939246, + 0.007180907763540745, + 0.013022860512137413, + 0.004354447592049837, + 0.004063968546688557, + 0.002807410666719079, + 0.009613948874175549, + 0.005130118224769831, + 0.002721036784350872, + 0.010207844898104668, + 0.004526179749518633, + 0.003511897288262844, + 0.0033921110443770885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.7578477263450623 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal a vast landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, simple wooden rowboat with faded blue paint visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat rocking violently up and down in turbulent gray water under torrential rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cold, monochromatic gray tone with low, thick clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0004949536523781717 + ], + "opposite_logprob_tokens": [ + -7.625494956970215 + ], + "logprob_reward": -0.0004949536523781717, + "yes_no_margin_reward": 7.625, + "token_credit_reward": -0.0004949536523781717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11265597492456436, + 0.06075672432780266, + 0.020057933405041695, + 0.03431343287229538, + 0.21662908792495728, + 0.16129712760448456, + 0.07239037752151489, + 0.18662768602371216, + 0.09956567734479904, + 0.0357060506939888 + ], + "credit_map": [ + 0.006898177787661552, + 0.024432143196463585, + 0.007374201435595751, + 0.005573207978159189, + 0.029520926997065544, + 0.020905597135424614, + 0.007560187950730324, + 0.010391510091722012, + 0.0032620306592434645, + 0.015846271067857742, + 0.0036739634815603495, + 0.0024106758646667004, + 0.01931513473391533, + 0.007504422217607498, + 0.0037869010120630264, + 0.004957315046340227, + 0.001385407173074782, + 0.0022018745075911283, + 0.0019920440390706062, + 0.001466588000766933, + 0.0026114408392459154, + 0.005422303918749094, + 0.0018698659259825945, + 0.003108406439423561, + 0.0025591980665922165, + 0.008033042773604393, + 0.0027685260865837336, + 0.0024471350479871035, + 0.009665018878877163, + 0.005121839698404074, + 0.0017654667608439922, + 0.001953200437128544, + 0.013639979995787144, + 0.02105128951370716, + 0.023949792608618736, + 0.03928685933351517, + 0.021982276812195778, + 0.024428723379969597, + 0.04770283401012421, + 0.024587292224168777, + 0.014343904331326485, + 0.013139505870640278, + 0.02005947194993496, + 0.02567228488624096, + 0.01764873042702675, + 0.022211505100131035, + 0.019942201673984528, + 0.028279511258006096, + 0.007945609278976917, + 0.009025001898407936, + 0.006412766873836517, + 0.00545768067240715, + 0.01558206882327795, + 0.012828920036554337, + 0.005775511730462313, + 0.009362808428704739, + 0.019083045423030853, + 0.03938119858503342, + 0.017890077084302902, + 0.015697095543146133, + 0.02721448987722397, + 0.04244224354624748, + 0.010441367514431477, + 0.014478117227554321, + 0.00978353526443243, + 0.01649542897939682, + 0.008430356159806252, + 0.00660183047875762, + 0.017755962908267975, + 0.02394161745905876, + 0.007123960647732019, + 0.009432973340153694, + 0.0034168707206845284, + 0.005730662494897842, + 0.003493349999189377, + 0.003391286125406623, + 0.006879022810608149, + 0.004621583968400955, + 0.0032729245722293854, + 0.004900345578789711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 5 + ], + "selected_reward": -0.0004949536523781717 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "long shot, shot from a high angle, slowly pulls back to reveal a vast landscape shrouded in torrential rain. Thick, gloomy clouds hang low, almost touching the vast, turbulent surface of the gray water. Raindrops cascade down like thousands of silver needles, stirring up countless broken, rapid ripples on the turbulent surface. In the center of this vast expanse, the only thing in sight is a small, simple wooden rowboat, its faded blue paint peeling and revealing the weathered wood beneath. It is helplessly tossed up and down by the wind and waves, rocking violently with each surge, its bow sometimes lifted high, pointing toward the leaden sky, sometimes plunging into the trough of the waves, its bottom filled with icy rainwater. The entire picture is dominated by a cold, almost monochromatic gray tone. Under the even and oppressive diffuse light, the scene is filled with an extreme loneliness, desolation and a sense of insignificance in the face of natural forces, creating a tranquil yet tense atmosphere.", + "seed": 3200, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_12/index_7487_seed_3200_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.24085387339582667, + "vlm_logprob": -0.24085387339582667, + "vlm_yes_no_margin": 3.674999988079071, + "vlm_token_credit": -0.24085387339582667, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal a vast landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, simple wooden rowboat with faded blue paint visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat rocking violently up and down in turbulent gray water under torrential rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cold, monochromatic gray tone with low, thick clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10430965572595596 + ], + "opposite_logprob_tokens": [ + -2.604309558868408 + ], + "logprob_reward": -0.10430965572595596, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.10430965572595596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5590134263038635, + 0.06480218470096588, + 0.0387328676879406, + 0.06889251619577408, + 0.04337503761053085, + 0.0345873162150383, + 0.038075339049100876, + 0.05226541683077812, + 0.030359145253896713, + 0.06989677995443344 + ], + "credit_map": [ + 0.07000332325696945, + 0.057472530752420425, + 0.11629776656627655, + 0.04725386202335358, + 0.07551470398902893, + 0.08052998781204224, + 0.03226286917924881, + 0.07967840135097504, + 0.016352377831935883, + 0.004927959758788347, + 0.006153848022222519, + 0.004404974635690451, + 0.01783723384141922, + 0.007230535615235567, + 0.004076815210282803, + 0.003818442579358816, + 0.007161728572100401, + 0.004144030623137951, + 0.00678648054599762, + 0.0029522287659347057, + 0.00564983393996954, + 0.0036851842887699604, + 0.0036026430316269398, + 0.004750738851726055, + 0.007159475237131119, + 0.008928640745580196, + 0.01720743253827095, + 0.006695631425827742, + 0.007280851248651743, + 0.009092760272324085, + 0.006665021646767855, + 0.005862704012542963, + 0.00361105939373374, + 0.005499822087585926, + 0.005319438874721527, + 0.006966000888496637, + 0.0033990293741226196, + 0.009186327457427979, + 0.004743504337966442, + 0.004649855196475983, + 0.003718489548191428, + 0.004145374987274408, + 0.00554930092766881, + 0.003132932586595416, + 0.004555833525955677, + 0.005959892179816961, + 0.003804056905210018, + 0.0037214357871562243, + 0.003783632768318057, + 0.008049308322370052, + 0.0037363104056566954, + 0.002265898510813713, + 0.010290217585861683, + 0.005173022858798504, + 0.0022068899124860764, + 0.0025700582191348076, + 0.00814942829310894, + 0.006263840012252331, + 0.009455746971070766, + 0.004255854990333319, + 0.007183748763054609, + 0.007023552432656288, + 0.0046501425094902515, + 0.005283105652779341, + 0.0038382376078516245, + 0.004251805599778891, + 0.005360482260584831, + 0.002217034576460719, + 0.0040728808380663395, + 0.0058895861729979515, + 0.0023772050626575947, + 0.002351911272853613, + 0.0075191836804151535, + 0.007807007525116205, + 0.01338556595146656, + 0.007054965011775494, + 0.009280680678784847, + 0.009937040507793427, + 0.004556307103484869, + 0.010356024838984013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.10430965572595596 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal a vast landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, simple wooden rowboat with faded blue paint visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat rocking violently up and down in turbulent gray water under torrential rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cold, monochromatic gray tone with low, thick clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0539443492889404 + ], + "opposite_logprob_tokens": [ + -0.42894428968429565 + ], + "logprob_reward": -1.0539443492889404, + "yes_no_margin_reward": -0.6250000596046448, + "token_credit_reward": -1.0539443492889404, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.4820791184902191, + 0.17760244011878967, + 0.03858136758208275, + 0.07902909815311432, + 0.022886084392666817, + 0.018054312095046043, + 0.062082137912511826, + 0.021328816190361977, + 0.029180945828557014, + 0.06917567551136017 + ], + "credit_map": [ + 0.06215324252843857, + 0.05602605268359184, + 0.0893012061715126, + 0.042457256466150284, + 0.06422977894544601, + 0.07498384267091751, + 0.029813220724463463, + 0.06311453133821487, + 0.0520247220993042, + 0.00770027469843626, + 0.015763897448778152, + 0.007093096151947975, + 0.0590067133307457, + 0.014958812855184078, + 0.00696009024977684, + 0.01409483514726162, + 0.006328603718429804, + 0.0043825386092066765, + 0.008019611239433289, + 0.0031745422165840864, + 0.0030615879222750664, + 0.005035254172980785, + 0.0037048214580863714, + 0.00487440824508667, + 0.007531925104558468, + 0.010824596509337425, + 0.023393534123897552, + 0.009237128309905529, + 0.006358593236654997, + 0.007956997491419315, + 0.00580585328862071, + 0.007920477539300919, + 0.0024289933498948812, + 0.003125451970845461, + 0.0033276542089879513, + 0.0021971147507429123, + 0.004076032433658838, + 0.0031662017572671175, + 0.0022690286859869957, + 0.0022956086322665215, + 0.0014984969748184085, + 0.0022490103729069233, + 0.0029967729933559895, + 0.0018807698506861925, + 0.002327860565856099, + 0.003578983712941408, + 0.0017223008908331394, + 0.0018001177813857794, + 0.003428687574341893, + 0.01980852521955967, + 0.0032724356278777122, + 0.0027134360279887915, + 0.022415507584810257, + 0.005581780802458525, + 0.002736984519287944, + 0.0021247852127999067, + 0.0018619694747030735, + 0.002466685837134719, + 0.003932005260139704, + 0.0030368659645318985, + 0.002774324733763933, + 0.001916771288961172, + 0.0022423823829740286, + 0.0030978145077824593, + 0.004099437966942787, + 0.0027170453686267138, + 0.004493910353630781, + 0.0025642646942287683, + 0.0034869625233113766, + 0.005008176434785128, + 0.0025524883531033993, + 0.00425866199657321, + 0.007377437315881252, + 0.008296546526253223, + 0.01312287524342537, + 0.005946460645645857, + 0.00883014127612114, + 0.011161942966282368, + 0.004987368825823069, + 0.00945290457457304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -1.0539443492889404 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal a vast landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, simple wooden rowboat with faded blue paint visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat rocking violently up and down in turbulent gray water under torrential rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cold, monochromatic gray tone with low, thick clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002209723461419344 + ], + "opposite_logprob_tokens": [ + -6.127209663391113 + ], + "logprob_reward": -0.002209723461419344, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002209723461419344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32727643847465515, + 0.33960798382759094, + 0.02816682495176792, + 0.0445195697247982, + 0.013603592291474342, + 0.06900617480278015, + 0.01987180858850479, + 0.05240807682275772, + 0.04063597694039345, + 0.06490357965230942 + ], + "credit_map": [ + 0.03962176293134689, + 0.03916718438267708, + 0.061916280537843704, + 0.029005328193306923, + 0.04476345703005791, + 0.04934269189834595, + 0.020294293761253357, + 0.04316543787717819, + 0.08865903317928314, + 0.02480303682386875, + 0.039401184767484665, + 0.016618050634860992, + 0.08936680853366852, + 0.03359174728393555, + 0.013070421293377876, + 0.03409769758582115, + 0.00710357166826725, + 0.004103297833353281, + 0.003411336103454232, + 0.001900757197290659, + 0.0038074911572039127, + 0.0038238924462348223, + 0.0019613541662693024, + 0.0020551232155412436, + 0.003625942626968026, + 0.005370565224438906, + 0.0072817811742424965, + 0.005482920911163092, + 0.00471998518332839, + 0.006462595891207457, + 0.006996049080044031, + 0.004579730331897736, + 0.001167011447250843, + 0.0017036052886396646, + 0.002660126192495227, + 0.001679897541180253, + 0.0016278151888400316, + 0.0019014232093468308, + 0.0015211179852485657, + 0.001342595904134214, + 0.006612773519009352, + 0.009476813487708569, + 0.010674314573407173, + 0.006180922966450453, + 0.01200045831501484, + 0.00844068918377161, + 0.008413110859692097, + 0.007207090966403484, + 0.0018033548258244991, + 0.0032891584560275078, + 0.002029680646955967, + 0.001906048390083015, + 0.003375954693183303, + 0.004083127249032259, + 0.0016573464963585138, + 0.0017271372489631176, + 0.006963294465094805, + 0.006070677191019058, + 0.00925359595566988, + 0.0050808428786695, + 0.006875670049339533, + 0.0062149870209395885, + 0.004660381004214287, + 0.007288627326488495, + 0.0031154691241681576, + 0.005207468289881945, + 0.004277173429727554, + 0.0036077271215617657, + 0.005048699676990509, + 0.008749505504965782, + 0.005428572650998831, + 0.005201361142098904, + 0.003605514531955123, + 0.007165664341300726, + 0.008620296604931355, + 0.010750756599009037, + 0.0069743734784424305, + 0.012506837956607342, + 0.00720119196921587, + 0.008078944869339466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.002209723461419344 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal a vast landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, simple wooden rowboat with faded blue paint visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat rocking violently up and down in turbulent gray water under torrential rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cold, monochromatic gray tone with low, thick clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043086349964141846 + ], + "opposite_logprob_tokens": [ + -3.168086290359497 + ], + "logprob_reward": -0.043086349964141846, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043086349964141846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21589063107967377, + 0.34811723232269287, + 0.04124215245246887, + 0.02129070833325386, + 0.03131766989827156, + 0.024167725816369057, + 0.08005517721176147, + 0.17810717225074768, + 0.028806913644075394, + 0.03100457228720188 + ], + "credit_map": [ + 0.026677532121539116, + 0.024099934846162796, + 0.041224438697099686, + 0.01871301233768463, + 0.027544908225536346, + 0.03617260977625847, + 0.013566526584327221, + 0.02789166383445263, + 0.10543021559715271, + 0.017029723152518272, + 0.030754800885915756, + 0.009634057059884071, + 0.12161456793546677, + 0.025806495919823647, + 0.012405294924974442, + 0.02544206939637661, + 0.009417559951543808, + 0.006492571905255318, + 0.004733023699373007, + 0.0024717289488762617, + 0.005660208873450756, + 0.005464533809572458, + 0.00281151314266026, + 0.004191013518720865, + 0.0026353553403168917, + 0.002587219001725316, + 0.0043470896780490875, + 0.0025396915152668953, + 0.001993973273783922, + 0.0025949771516025066, + 0.0026849170681089163, + 0.001907484489493072, + 0.003308056155219674, + 0.003930089995265007, + 0.005475873127579689, + 0.003927568905055523, + 0.0032706724014133215, + 0.004197101108729839, + 0.003423453541472554, + 0.0037848548963665962, + 0.0019502914510667324, + 0.0027091640513390303, + 0.004487867467105389, + 0.003365366719663143, + 0.001865703146904707, + 0.004505620803683996, + 0.0021700067445635796, + 0.0031137075275182724, + 0.005360590759664774, + 0.0229540653526783, + 0.006407028529793024, + 0.006287745200097561, + 0.02268099971115589, + 0.007834392599761486, + 0.004778383765369654, + 0.003751974320039153, + 0.02576640620827675, + 0.020623233169317245, + 0.03142305091023445, + 0.016183508560061455, + 0.02297860197722912, + 0.024936659261584282, + 0.01291804388165474, + 0.02327767200767994, + 0.0032776673324406147, + 0.0037708384916186333, + 0.004582010209560394, + 0.002479665447026491, + 0.004081189166754484, + 0.003813759656623006, + 0.0019988485146313906, + 0.004802933894097805, + 0.003212236799299717, + 0.003329396480694413, + 0.005796761251986027, + 0.0028617733623832464, + 0.003983360715210438, + 0.005091795232146978, + 0.0027679316699504852, + 0.003961316309869289 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.043086349964141846 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal a vast landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, simple wooden rowboat with faded blue paint visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat rocking violently up and down in turbulent gray water under torrential rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cold, monochromatic gray tone with low, thick clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007192885386757553 + ], + "opposite_logprob_tokens": [ + -7.25071907043457 + ], + "logprob_reward": -0.0007192885386757553, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007192885386757553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5937831997871399, + 0.16705968976020813, + 0.02806662768125534, + 0.06178032234311104, + 0.015225737355649471, + 0.015926722437143326, + 0.024935994297266006, + 0.028753262013196945, + 0.026463568210601807, + 0.03800487518310547 + ], + "credit_map": [ + 0.07567274570465088, + 0.06676327437162399, + 0.11180245131254196, + 0.0514875091612339, + 0.07963258028030396, + 0.0935499295592308, + 0.03572389855980873, + 0.07915081828832626, + 0.04640335962176323, + 0.01011371985077858, + 0.015331994742155075, + 0.007187606301158667, + 0.05072583630681038, + 0.016460154205560684, + 0.006284198258072138, + 0.014552813023328781, + 0.004165796097368002, + 0.0034517566673457623, + 0.004738540854305029, + 0.0023341046180576086, + 0.003927013371139765, + 0.003838760545477271, + 0.002449450781568885, + 0.0031612045131623745, + 0.0070371623151004314, + 0.008465845137834549, + 0.014707107096910477, + 0.006253970321267843, + 0.009418408386409283, + 0.005834733135998249, + 0.005164564121514559, + 0.0048985350877046585, + 0.0013507675612345338, + 0.001863015117123723, + 0.002504297997802496, + 0.001871999353170395, + 0.0017243489855900407, + 0.0022621206007897854, + 0.0017021183157339692, + 0.0019470686092972755, + 0.002862404566258192, + 0.002122014295309782, + 0.00244538439437747, + 0.0010874413419514894, + 0.0015089227817952633, + 0.0031540419440716505, + 0.0010466949315741658, + 0.0016998187638819218, + 0.0027088390197604895, + 0.0033098969142884016, + 0.003294781781733036, + 0.0023990292102098465, + 0.0042447312735021114, + 0.005152992904186249, + 0.001842051511630416, + 0.0019836709834635258, + 0.0043303039856255054, + 0.003426210954785347, + 0.004556531086564064, + 0.002549766330048442, + 0.00404145335778594, + 0.0037720329128205776, + 0.003182964865118265, + 0.0028939980547875166, + 0.0039450922049582005, + 0.004283099435269833, + 0.004323303699493408, + 0.0021378234960138798, + 0.003214728320017457, + 0.003444002941250801, + 0.002657870762050152, + 0.0024576494470238686, + 0.004615307319909334, + 0.004962559789419174, + 0.007788477465510368, + 0.0039910608902573586, + 0.004725325852632523, + 0.004462091252207756, + 0.0035211776848882437, + 0.00393887422978878 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.0007192885386757553 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_13/index_tensor([2570, 2570])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_13/index_tensor([2570, 2570])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9e4491d9235a808f48e904fbfdd8d227490b9440 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_13/index_tensor([2570, 2570])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "This low-angle, extreme close-up captures the massive drive wheel of a massive locomotive in breathtakingly high-speed slow motion, as the camera slowly tracks along it. Forged from heavy, dark steel, its surface smeared with oil and speckled with rust, the wheel slams violently against the polished, rain-soaked rail. It struggles agonizingly slowly to complete its rotation, its immense force causing it to slip and grind, sending a brilliant explosion of orange and yellow sparks flying from the point of contact. Water droplets clinging to the wheel's metal surface are atomized by the intense vibrations. In the background, the black crushed stone ballast of the track drenches and gleams in the diffuse gray light of an overcast sky. The entire scene, immersed in a cool, monochromatic palette of industrial grays and blacks, creates an atmosphere of tension and power, a raw, mechanical struggle against immense physical resistance.", + "seed": 15360, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_13/index_2570_seed_15360_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02183909760788083, + "vlm_logprob": -0.02183909760788083, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.02183909760788083, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive locomotive drive wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking slowly along the wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wheel made of dark steel with visible oil smears and rust speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there orange and yellow sparks exploding from the contact point with the rain-soaked rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of black crushed stone ballast under a diffuse gray overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01742764189839363 + ], + "opposite_logprob_tokens": [ + -4.642427444458008 + ], + "logprob_reward": -0.01742764189839363, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.01742764189839363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14961273968219757, + 0.0888049528002739, + 0.021307889372110367, + 0.07496321201324463, + 0.013151956722140312, + 0.4516475796699524, + 0.041684720665216446, + 0.0469513013958931, + 0.037673141807317734, + 0.07420258969068527 + ], + "credit_map": [ + 0.012929173186421394, + 0.01916591078042984, + 0.014883778989315033, + 0.016515687108039856, + 0.023314177989959717, + 0.025646384805440903, + 0.02167729288339615, + 0.01548031810671091, + 0.005464719608426094, + 0.01014355942606926, + 0.004790318664163351, + 0.005110907834023237, + 0.03811616823077202, + 0.008688423782587051, + 0.013578807935118675, + 0.0029120331164449453, + 0.0014236735878512263, + 0.0024169660173356533, + 0.00226543378084898, + 0.00378606328740716, + 0.003254388924688101, + 0.003221662249416113, + 0.0027810719329863787, + 0.002158625051379204, + 0.007017593365162611, + 0.009975292719900608, + 0.006442490965127945, + 0.008160294033586979, + 0.010996182449162006, + 0.01053535845130682, + 0.01318220142275095, + 0.008653786033391953, + 0.001267074141651392, + 0.0015108032384887338, + 0.0015243071829900146, + 0.002209687139838934, + 0.0016632989281788468, + 0.00132000632584095, + 0.0018680824432522058, + 0.0017886952264234424, + 0.03156495839357376, + 0.019886022433638573, + 0.0844411849975586, + 0.18422508239746094, + 0.012167371809482574, + 0.024568133056163788, + 0.06450537592172623, + 0.030289417132735252, + 0.004779726732522249, + 0.005143893416970968, + 0.006819831673055887, + 0.005418532062321901, + 0.0028188808355480433, + 0.007575300522148609, + 0.004170123487710953, + 0.004958425648510456, + 0.0037086664233356714, + 0.004995518829673529, + 0.0068280561827123165, + 0.0038340550381690264, + 0.008006976917386055, + 0.00858019944280386, + 0.005345288198441267, + 0.0056525347754359245, + 0.003047026926651597, + 0.004772609565407038, + 0.004865976981818676, + 0.005951378494501114, + 0.006671652663499117, + 0.00563681498169899, + 0.003137857187539339, + 0.00358982034958899, + 0.003130686702206731, + 0.004994201473891735, + 0.003445209702476859, + 0.005197515711188316, + 0.019294027239084244, + 0.024867577478289604, + 0.00746042188256979, + 0.005812938790768385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.01742764189839363 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive locomotive drive wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking slowly along the wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wheel made of dark steel with visible oil smears and rust speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there orange and yellow sparks exploding from the contact point with the rain-soaked rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of black crushed stone ballast under a diffuse gray overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005962679628282785 + ], + "opposite_logprob_tokens": [ + -5.13096284866333 + ], + "logprob_reward": -0.005962679628282785, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005962679628282785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1099642738699913, + 0.0956256315112114, + 0.036714304238557816, + 0.11371961981058121, + 0.02271602302789688, + 0.30553168058395386, + 0.01906660944223404, + 0.09112304449081421, + 0.06023460999131203, + 0.1453041434288025 + ], + "credit_map": [ + 0.012064683251082897, + 0.012373046018183231, + 0.012514345347881317, + 0.014533759094774723, + 0.013386091217398643, + 0.016000621020793915, + 0.01465275976806879, + 0.014438964426517487, + 0.004393606446683407, + 0.009976311586797237, + 0.005344058386981487, + 0.004824514500796795, + 0.03892362862825394, + 0.01639864780008793, + 0.01224730908870697, + 0.003517567180097103, + 0.0020425075199455023, + 0.007556098513305187, + 0.002358376979827881, + 0.004859811160713434, + 0.005802070256322622, + 0.005236449185758829, + 0.0052207414992153645, + 0.0036382502876222134, + 0.010807681828737259, + 0.012398802675306797, + 0.013163409195840359, + 0.014313267543911934, + 0.013676186092197895, + 0.01983696222305298, + 0.017148245126008987, + 0.01237507350742817, + 0.001649922109209001, + 0.0020346208475530148, + 0.0023623681627213955, + 0.005408080760389566, + 0.0037302901037037373, + 0.0022915175650268793, + 0.0027977258432656527, + 0.0024414979852735996, + 0.02197238802909851, + 0.012014327570796013, + 0.05822150036692619, + 0.12741149961948395, + 0.005043606739491224, + 0.016436032950878143, + 0.04659373313188553, + 0.017838602885603905, + 0.0016424026107415557, + 0.002166404854506254, + 0.002628876129165292, + 0.003345024073496461, + 0.0015905844047665596, + 0.0024933568201959133, + 0.002408750355243683, + 0.0027912126388400793, + 0.009037116542458534, + 0.010651920922100544, + 0.02027176506817341, + 0.007159268483519554, + 0.009495051577687263, + 0.014819251373410225, + 0.011599436402320862, + 0.008089248090982437, + 0.004727640189230442, + 0.007809677626937628, + 0.009153107181191444, + 0.009314212016761303, + 0.010115462355315685, + 0.008202790282666683, + 0.005157599225640297, + 0.005754126701503992, + 0.006760148331522942, + 0.01869182102382183, + 0.010095875710248947, + 0.017969481647014618, + 0.03799917548894882, + 0.027512740343809128, + 0.01618209108710289, + 0.010092820972204208 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 3 + ], + "selected_reward": -0.005962679628282785 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive locomotive drive wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking slowly along the wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wheel made of dark steel with visible oil smears and rust speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there orange and yellow sparks exploding from the contact point with the rain-soaked rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of black crushed stone ballast under a diffuse gray overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004643134772777557 + ], + "opposite_logprob_tokens": [ + -5.379642963409424 + ], + "logprob_reward": -0.004643134772777557, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004643134772777557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04806475341320038, + 0.04766359552741051, + 0.07182583212852478, + 0.07740253955125809, + 0.015370236709713936, + 0.6344265937805176, + 0.018555665388703346, + 0.03348683938384056, + 0.017692888155579567, + 0.0355110764503479 + ], + "credit_map": [ + 0.004737657029181719, + 0.005292156711220741, + 0.003616092959418893, + 0.0062059061601758, + 0.0077679697424173355, + 0.008532604202628136, + 0.006364649161696434, + 0.0055477190762758255, + 0.004346057772636414, + 0.006209769286215305, + 0.004992836154997349, + 0.004047235939651728, + 0.013280213810503483, + 0.005676331929862499, + 0.006046937312930822, + 0.0030642091296613216, + 0.0043974705040454865, + 0.01168461237102747, + 0.008262014016509056, + 0.017180101945996284, + 0.006966236513108015, + 0.006619816180318594, + 0.008801927790045738, + 0.007913650944828987, + 0.0068800668232142925, + 0.010031824000179768, + 0.007483407389372587, + 0.008203615434467793, + 0.011860709637403488, + 0.012292138300836086, + 0.01219999697059393, + 0.008450780063867569, + 0.001340273185633123, + 0.0016944293165579438, + 0.0011872067116200924, + 0.002814678708091378, + 0.0029304446652531624, + 0.0015816246159374714, + 0.0022578509524464607, + 0.0015637289034202695, + 0.04547811672091484, + 0.025867341086268425, + 0.12639153003692627, + 0.26199135184288025, + 0.010387757793068886, + 0.035730235278606415, + 0.09069298952817917, + 0.03788725286722183, + 0.001380602945573628, + 0.0021606346126645803, + 0.0023593869991600513, + 0.004108387511223555, + 0.0014038142981007695, + 0.00230452255345881, + 0.00268519576638937, + 0.0021531214006245136, + 0.00303224241361022, + 0.0034772029612213373, + 0.00736357131972909, + 0.0032546864822506905, + 0.0033294563181698322, + 0.004952212329953909, + 0.005327377002686262, + 0.002750088693574071, + 0.0014715703437104821, + 0.0019942375365644693, + 0.002704678801819682, + 0.0023342170752584934, + 0.0035341735929250717, + 0.002327871276065707, + 0.0014507777523249388, + 0.0018753607291728258, + 0.001754738506861031, + 0.0032267163041979074, + 0.00237283855676651, + 0.003994476515799761, + 0.008240957744419575, + 0.009077542461454868, + 0.003718292573466897, + 0.00312551180832088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 2 + ], + "selected_reward": -0.004643134772777557 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive locomotive drive wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking slowly along the wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wheel made of dark steel with visible oil smears and rust speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there orange and yellow sparks exploding from the contact point with the rain-soaked rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of black crushed stone ballast under a diffuse gray overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043073106557130814 + ], + "opposite_logprob_tokens": [ + -3.1680731773376465 + ], + "logprob_reward": -0.043073106557130814, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043073106557130814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.034358564764261246, + 0.10173114389181137, + 0.17032650113105774, + 0.06373202800750732, + 0.010014490224421024, + 0.519442617893219, + 0.012178190983831882, + 0.028971819207072258, + 0.019213363528251648, + 0.04003118351101875 + ], + "credit_map": [ + 0.0020270489621907473, + 0.005202583037316799, + 0.002161141950637102, + 0.0028947184327989817, + 0.005065815523266792, + 0.007158693391829729, + 0.006302873138338327, + 0.0035456924233585596, + 0.0059882174246013165, + 0.012364509515464306, + 0.005683160852640867, + 0.007121240254491568, + 0.03792240843176842, + 0.012677746824920177, + 0.015031138435006142, + 0.004942731931805611, + 0.010194560512900352, + 0.027870751917362213, + 0.020358489826321602, + 0.03530372679233551, + 0.016531091183423996, + 0.01989099010825157, + 0.0181332528591156, + 0.022043656557798386, + 0.00516106141731143, + 0.00790597777813673, + 0.004182534292340279, + 0.009635481052100658, + 0.008449437096714973, + 0.0069307298399508, + 0.01325405016541481, + 0.008212759159505367, + 0.0006921500316821039, + 0.0008759654592722654, + 0.0008469503954984248, + 0.001211026799865067, + 0.001609564060345292, + 0.0012545563513413072, + 0.002167758531868458, + 0.00135652010794729, + 0.03768790885806084, + 0.020836543291807175, + 0.10595965385437012, + 0.21418489515781403, + 0.009427854791283607, + 0.02888086624443531, + 0.07319161295890808, + 0.029273364692926407, + 0.0012081158347427845, + 0.0013618138618767262, + 0.0014653007965534925, + 0.002620080253109336, + 0.0009414663654752076, + 0.0013505734968930483, + 0.0015341119142249227, + 0.0016967294504866004, + 0.003760130610316992, + 0.003062693402171135, + 0.006661831866949797, + 0.002802307950332761, + 0.0029346433002501726, + 0.003932584077119827, + 0.0036295917816460133, + 0.00218803808093071, + 0.002031223615631461, + 0.0028322511352598667, + 0.002078854013234377, + 0.0020877376664429903, + 0.0033319462090730667, + 0.0025977282784879208, + 0.0016298388363793492, + 0.0026237841229885817, + 0.0023461400996893644, + 0.005504578817635775, + 0.003044168697670102, + 0.004553537350147963, + 0.008886107243597507, + 0.0072314380668103695, + 0.004966170527040958, + 0.0034990468993782997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.043073106557130814 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive locomotive drive wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking slowly along the wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wheel made of dark steel with visible oil smears and rust speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there orange and yellow sparks exploding from the contact point with the rain-soaked rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of black crushed stone ballast under a diffuse gray overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038088925182819366 + ], + "opposite_logprob_tokens": [ + -3.2880890369415283 + ], + "logprob_reward": -0.038088925182819366, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038088925182819366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09116500616073608, + 0.10214017331600189, + 0.04772314056754112, + 0.07132641971111298, + 0.0261019729077816, + 0.43902918696403503, + 0.05560064688324928, + 0.028314510360360146, + 0.0237213596701622, + 0.11487753689289093 + ], + "credit_map": [ + 0.009628305211663246, + 0.013040579855442047, + 0.008610375225543976, + 0.010954217053949833, + 0.012430681847035885, + 0.015366642735898495, + 0.011801449581980705, + 0.009332770481705666, + 0.004491060506552458, + 0.008744683116674423, + 0.008631519973278046, + 0.005332412198185921, + 0.03200770542025566, + 0.028822030872106552, + 0.007808292284607887, + 0.00630246801301837, + 0.0024702271912246943, + 0.005787346977740526, + 0.003738891799002886, + 0.009788129478693008, + 0.006072611548006535, + 0.007660717237740755, + 0.005670483689755201, + 0.006534735206514597, + 0.006833162158727646, + 0.0063978154212236404, + 0.011020559817552567, + 0.010895132087171078, + 0.006283417344093323, + 0.01490052416920662, + 0.008427722379565239, + 0.0065680877305567265, + 0.002347726607695222, + 0.0024207213427871466, + 0.0028814037796109915, + 0.004236172419041395, + 0.0035701002925634384, + 0.0024392101913690567, + 0.003693460253998637, + 0.004513179417699575, + 0.03180340677499771, + 0.017904071137309074, + 0.0939827412366867, + 0.17794209718704224, + 0.00928227324038744, + 0.025452187284827232, + 0.05855711176991463, + 0.024105338379740715, + 0.004388363100588322, + 0.006855642888695002, + 0.009337951429188251, + 0.006156057585030794, + 0.006863265298306942, + 0.008997282944619656, + 0.007966334000229836, + 0.005035756155848503, + 0.00217015715315938, + 0.0029313531704247, + 0.004006418399512768, + 0.0032666160259395838, + 0.002848974196240306, + 0.005334397777915001, + 0.004788154736161232, + 0.0029684421606361866, + 0.002270717406645417, + 0.004141420125961304, + 0.0021026036702096462, + 0.003692346392199397, + 0.0045411488972604275, + 0.0031112716533243656, + 0.0021744659170508385, + 0.0016873866552487016, + 0.005266338121145964, + 0.013617893680930138, + 0.008791303262114525, + 0.012078046798706055, + 0.03152785822749138, + 0.024190692231059074, + 0.01222652941942215, + 0.007178887724876404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 1 + ], + "selected_reward": -0.038088925182819366 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This low-angle, extreme close-up captures the massive drive wheel of a massive locomotive in breathtakingly high-speed slow motion, as the camera slowly tracks along it. Forged from heavy, dark steel, its surface smeared with oil and speckled with rust, the wheel slams violently against the polished, rain-soaked rail. It struggles agonizingly slowly to complete its rotation, its immense force causing it to slip and grind, sending a brilliant explosion of orange and yellow sparks flying from the point of contact. Water droplets clinging to the wheel's metal surface are atomized by the intense vibrations. In the background, the black crushed stone ballast of the track drenches and gleams in the diffuse gray light of an overcast sky. The entire scene, immersed in a cool, monochromatic palette of industrial grays and blacks, creates an atmosphere of tension and power, a raw, mechanical struggle against immense physical resistance.", + "seed": 15361, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_13/index_2570_seed_15361_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.01867056074552238, + "vlm_logprob": -0.01867056074552238, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.01867056074552238, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive locomotive drive wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking slowly along the wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wheel made of dark steel with visible oil smears and rust speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there orange and yellow sparks exploding from the contact point with the rain-soaked rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of black crushed stone ballast under a diffuse gray overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01571320928633213 + ], + "opposite_logprob_tokens": [ + -5.390713214874268 + ], + "logprob_reward": -0.01571320928633213, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.01571320928633213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2571815252304077, + 0.11204420030117035, + 0.05951736494898796, + 0.09236235171556473, + 0.15932220220565796, + 0.08519574999809265, + 0.044880468398332596, + 0.05623699352145195, + 0.06812282651662827, + 0.06513635069131851 + ], + "credit_map": [ + 0.05318185314536095, + 0.05753245949745178, + 0.02147538773715496, + 0.025342850014567375, + 0.028700528666377068, + 0.01926347054541111, + 0.027258502319455147, + 0.024426434189081192, + 0.01684866100549698, + 0.02181522734463215, + 0.01196421217173338, + 0.0076787094585597515, + 0.0195775143802166, + 0.007923727855086327, + 0.017368298023939133, + 0.008867837488651276, + 0.007351972162723541, + 0.009683242067694664, + 0.006128415931016207, + 0.009862441569566727, + 0.009272508323192596, + 0.006645696237683296, + 0.005108971148729324, + 0.005464114714413881, + 0.013331898488104343, + 0.013823379762470722, + 0.014259859919548035, + 0.00981234572827816, + 0.013138513080775738, + 0.0077624451369047165, + 0.00987198855727911, + 0.010361909866333008, + 0.02471461147069931, + 0.0244482159614563, + 0.02663571387529373, + 0.014858570881187916, + 0.02766212821006775, + 0.013390502892434597, + 0.011728309094905853, + 0.015884138643741608, + 0.015211530029773712, + 0.01782308705151081, + 0.008212833665311337, + 0.006788644473999739, + 0.01241740956902504, + 0.008840287104249, + 0.006119632162153721, + 0.009782318025827408, + 0.005605585407465696, + 0.004416182171553373, + 0.003964959178119898, + 0.006341318134218454, + 0.004956308286637068, + 0.007380784023553133, + 0.005153486970812082, + 0.00706184096634388, + 0.009371857158839703, + 0.007530506234616041, + 0.007789014372974634, + 0.008060535416007042, + 0.00805854331701994, + 0.006448795553296804, + 0.004141712095588446, + 0.00483601912856102, + 0.004572591278702021, + 0.014130931347608566, + 0.009579569101333618, + 0.006132552865892649, + 0.012012033723294735, + 0.009916736744344234, + 0.006440425757318735, + 0.005337984301149845, + 0.006540859583765268, + 0.013741773553192616, + 0.007848340086638927, + 0.003983850125223398, + 0.009310099296271801, + 0.007448479998856783, + 0.0083472253754735, + 0.007915708236396313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.01571320928633213 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive locomotive drive wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking slowly along the wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wheel made of dark steel with visible oil smears and rust speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there orange and yellow sparks exploding from the contact point with the rain-soaked rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of black crushed stone ballast under a diffuse gray overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00764184445142746 + ], + "opposite_logprob_tokens": [ + -4.882641792297363 + ], + "logprob_reward": -0.00764184445142746, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00764184445142746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13690629601478577, + 0.12828713655471802, + 0.1144329085946083, + 0.12732847034931183, + 0.19684135913848877, + 0.09021782875061035, + 0.028764212504029274, + 0.10847266018390656, + 0.02962193824350834, + 0.03912713751196861 + ], + "credit_map": [ + 0.022864753380417824, + 0.02380404993891716, + 0.012911010533571243, + 0.019586412236094475, + 0.017191138118505478, + 0.012063030153512955, + 0.015094170346856117, + 0.013391721993684769, + 0.01625346578657627, + 0.017496258020401, + 0.011391771025955677, + 0.008163201622664928, + 0.013704130426049232, + 0.029248762875795364, + 0.016532082110643387, + 0.015497474931180477, + 0.016178350895643234, + 0.017508337274193764, + 0.01367554534226656, + 0.014433000236749649, + 0.01918768882751465, + 0.013734506443142891, + 0.01014719344675541, + 0.009568278677761555, + 0.02063349261879921, + 0.019647832959890366, + 0.015200871974229813, + 0.01434975303709507, + 0.01732509210705757, + 0.02194369211792946, + 0.008219809271395206, + 0.010007917881011963, + 0.02701667696237564, + 0.033437520265579224, + 0.020765962079167366, + 0.013937723822891712, + 0.03979956731200218, + 0.024120748043060303, + 0.011839172802865505, + 0.02592397853732109, + 0.015347739681601524, + 0.01822303794324398, + 0.009015075862407684, + 0.0038550111930817366, + 0.01558933686465025, + 0.012009815312922001, + 0.008457088842988014, + 0.007720720488578081, + 0.0027279569767415524, + 0.0022427712101489305, + 0.004102903883904219, + 0.004780878312885761, + 0.004018939100205898, + 0.004992036614567041, + 0.0028708293102681637, + 0.003027894301339984, + 0.015941524878144264, + 0.012802639044821262, + 0.02302582934498787, + 0.010593111626803875, + 0.008082101121544838, + 0.022578058764338493, + 0.005776784382760525, + 0.009672602638602257, + 0.003470495343208313, + 0.004701214376837015, + 0.0037077604793012142, + 0.004225048702210188, + 0.00410368712618947, + 0.002500460483133793, + 0.004024906549602747, + 0.002888366812840104, + 0.0034290621988475323, + 0.005140898283571005, + 0.0029666563495993614, + 0.0025837752036750317, + 0.006167531944811344, + 0.011717665940523148, + 0.003881783690303564, + 0.003239766927435994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.00764184445142746 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive locomotive drive wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking slowly along the wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wheel made of dark steel with visible oil smears and rust speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there orange and yellow sparks exploding from the contact point with the rain-soaked rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of black crushed stone ballast under a diffuse gray overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00249182409606874 + ], + "opposite_logprob_tokens": [ + -6.0024919509887695 + ], + "logprob_reward": -0.00249182409606874, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.00249182409606874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12949422001838684, + 0.23923063278198242, + 0.15210434794425964, + 0.10808516293764114, + 0.1059829592704773, + 0.1313437968492508, + 0.019345885142683983, + 0.035701800137758255, + 0.031704485416412354, + 0.04700673371553421 + ], + "credit_map": [ + 0.017572542652487755, + 0.018950961530208588, + 0.010938846506178379, + 0.012253062799572945, + 0.01521402969956398, + 0.01131218671798706, + 0.027883876115083694, + 0.015368706546723843, + 0.02677980251610279, + 0.030860604718327522, + 0.020580507814884186, + 0.015745392069220543, + 0.025250671431422234, + 0.07067675143480301, + 0.021199814975261688, + 0.028137095272541046, + 0.021712204441428185, + 0.02381056733429432, + 0.016564367339015007, + 0.02205432392656803, + 0.023445647209882736, + 0.018466586247086525, + 0.01362068671733141, + 0.012429970316588879, + 0.016132090240716934, + 0.019252901896834373, + 0.014259657822549343, + 0.008752461522817612, + 0.016962694004178047, + 0.0113483015447855, + 0.009407049044966698, + 0.011970003135502338, + 0.014764338731765747, + 0.018525654450058937, + 0.010983224026858807, + 0.011348407715559006, + 0.022660717368125916, + 0.008735581301152706, + 0.007465262431651354, + 0.01149977371096611, + 0.02536400966346264, + 0.03364298492670059, + 0.013136192224919796, + 0.007784067653119564, + 0.015452309511601925, + 0.012769714929163456, + 0.010915684513747692, + 0.01227883342653513, + 0.0026458618231117725, + 0.0023943926207721233, + 0.0020969416946172714, + 0.0028722805436700583, + 0.002621768042445183, + 0.0024298897478729486, + 0.0021730843000113964, + 0.002111665904521942, + 0.005204705521464348, + 0.004886852111667395, + 0.005325730890035629, + 0.004114591050893068, + 0.003363303141668439, + 0.005209997296333313, + 0.002458674367517233, + 0.005137944128364325, + 0.0027882521972060204, + 0.005145696457475424, + 0.0032265507616102695, + 0.00782540813088417, + 0.003828120417892933, + 0.002761067124083638, + 0.003565691877156496, + 0.0025636961217969656, + 0.003546707099303603, + 0.0073751891031861305, + 0.004253343679010868, + 0.004286264069378376, + 0.00772658409550786, + 0.011869353242218494, + 0.004281396511942148, + 0.0036678961478173733 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 5 + ], + "selected_reward": -0.00249182409606874 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive locomotive drive wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking slowly along the wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wheel made of dark steel with visible oil smears and rust speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there orange and yellow sparks exploding from the contact point with the rain-soaked rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of black crushed stone ballast under a diffuse gray overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0549725741147995 + ], + "opposite_logprob_tokens": [ + -2.9299726486206055 + ], + "logprob_reward": -0.0549725741147995, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549725741147995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09915343672037125, + 0.23764720559120178, + 0.10226765275001526, + 0.13458308577537537, + 0.12797681987285614, + 0.13435886800289154, + 0.04394065588712692, + 0.036478832364082336, + 0.037835922092199326, + 0.045757509768009186 + ], + "credit_map": [ + 0.01803485117852688, + 0.01922965608537197, + 0.010466733947396278, + 0.01229826733469963, + 0.009221003390848637, + 0.009803084656596184, + 0.011845421977341175, + 0.008254410699009895, + 0.027194935828447342, + 0.031175866723060608, + 0.01937444508075714, + 0.01608208566904068, + 0.024923011660575867, + 0.06473016738891602, + 0.025964464992284775, + 0.028202231973409653, + 0.012314635328948498, + 0.014874392189085484, + 0.01117162685841322, + 0.01803113892674446, + 0.01408296637237072, + 0.010387610644102097, + 0.01006622239947319, + 0.01133906003087759, + 0.021272799000144005, + 0.02139304019510746, + 0.021558918058872223, + 0.015973586589097977, + 0.016589315608143806, + 0.022459769621491432, + 0.006440136581659317, + 0.00889552105218172, + 0.017340412363409996, + 0.020097875967621803, + 0.013660909608006477, + 0.014029753394424915, + 0.020142780616879463, + 0.012262808158993721, + 0.0119892917573452, + 0.01845298521220684, + 0.020449688658118248, + 0.025492243468761444, + 0.013940678909420967, + 0.0060513741336762905, + 0.023794593289494514, + 0.015062893740832806, + 0.012778243981301785, + 0.01678914576768875, + 0.005764708388596773, + 0.005637600552290678, + 0.006520949304103851, + 0.006696726195514202, + 0.005157888401299715, + 0.0052385455928742886, + 0.00422844011336565, + 0.004695798270404339, + 0.005618551280349493, + 0.0025421916507184505, + 0.011325465515255928, + 0.0038808821700513363, + 0.001883783028461039, + 0.0042450446635484695, + 0.002367166103795171, + 0.004615750163793564, + 0.002626285422593355, + 0.005975726526230574, + 0.004961085040122271, + 0.009994648396968842, + 0.003691434394568205, + 0.005234185606241226, + 0.002920505590736866, + 0.0024320483207702637, + 0.004342879168689251, + 0.007755333092063665, + 0.004622554872184992, + 0.0034852013923227787, + 0.006907881703227758, + 0.009943200275301933, + 0.004897621460258961, + 0.0038028403650969267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 5 + ], + "selected_reward": -0.0549725741147995 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive locomotive drive wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking slowly along the wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wheel made of dark steel with visible oil smears and rust speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there orange and yellow sparks exploding from the contact point with the rain-soaked rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of black crushed stone ballast under a diffuse gray overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253335177898407 + ], + "opposite_logprob_tokens": [ + -4.387533187866211 + ], + "logprob_reward": -0.01253335177898407, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253335177898407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15694454312324524, + 0.04434453323483467, + 0.16299133002758026, + 0.07912453263998032, + 0.39300864934921265, + 0.04904583469033241, + 0.016643188893795013, + 0.03579483926296234, + 0.02144615910947323, + 0.040656376630067825 + ], + "credit_map": [ + 0.028978342190384865, + 0.03319975733757019, + 0.01760181039571762, + 0.019745325669646263, + 0.016928713768720627, + 0.014107661321759224, + 0.014945722185075283, + 0.011437230743467808, + 0.006092318799346685, + 0.006640777923166752, + 0.0048962486907839775, + 0.0029857459012418985, + 0.006231746636331081, + 0.006433382630348206, + 0.006140495650470257, + 0.004923817701637745, + 0.024293391034007072, + 0.028523670509457588, + 0.014944751746952534, + 0.02324105054140091, + 0.026017893105745316, + 0.019866837188601494, + 0.012808896601200104, + 0.013294854201376438, + 0.010754432529211044, + 0.012971645221114159, + 0.011583127081394196, + 0.00728206429630518, + 0.012207798659801483, + 0.00965867843478918, + 0.006394405849277973, + 0.008272389881312847, + 0.05742752552032471, + 0.06703626364469528, + 0.04520140215754509, + 0.03099709004163742, + 0.07562222331762314, + 0.04439786821603775, + 0.024675365537405014, + 0.047650933265686035, + 0.008345279842615128, + 0.009767073206603527, + 0.004861312918365002, + 0.002633927157148719, + 0.00770668126642704, + 0.004836764652282, + 0.004720793571323156, + 0.00617400836199522, + 0.0021916888654232025, + 0.0021925608161836863, + 0.001869260799139738, + 0.0022096612956374884, + 0.002252384554594755, + 0.002632371848449111, + 0.0016498110489919782, + 0.0016454514116048813, + 0.005319529678672552, + 0.003475610166788101, + 0.009548971429467201, + 0.003849462838843465, + 0.002984305378049612, + 0.005703384522348642, + 0.0022016922011971474, + 0.0027118874713778496, + 0.0018718206556513906, + 0.00460767513141036, + 0.0019635793287307024, + 0.005599269177764654, + 0.0019732355140149593, + 0.002169840270653367, + 0.001935271779075265, + 0.0013254698133096099, + 0.002821954432874918, + 0.005026358179748058, + 0.0029782799538224936, + 0.0023248351644724607, + 0.006699968595057726, + 0.014032715000212193, + 0.003358983900398016, + 0.0034132858272641897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.01253335177898407 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This low-angle, extreme close-up captures the massive drive wheel of a massive locomotive in breathtakingly high-speed slow motion, as the camera slowly tracks along it. Forged from heavy, dark steel, its surface smeared with oil and speckled with rust, the wheel slams violently against the polished, rain-soaked rail. It struggles agonizingly slowly to complete its rotation, its immense force causing it to slip and grind, sending a brilliant explosion of orange and yellow sparks flying from the point of contact. Water droplets clinging to the wheel's metal surface are atomized by the intense vibrations. In the background, the black crushed stone ballast of the track drenches and gleams in the diffuse gray light of an overcast sky. The entire scene, immersed in a cool, monochromatic palette of industrial grays and blacks, creates an atmosphere of tension and power, a raw, mechanical struggle against immense physical resistance.", + "seed": 15362, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_13/index_2570_seed_15362_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.020752989687025547, + "vlm_logprob": -0.020752989687025547, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.020752989687025547, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive locomotive drive wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking slowly along the wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wheel made of dark steel with visible oil smears and rust speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there orange and yellow sparks exploding from the contact point with the rain-soaked rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of black crushed stone ballast under a diffuse gray overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025814387947320938 + ], + "opposite_logprob_tokens": [ + -4.025814533233643 + ], + "logprob_reward": -0.025814387947320938, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.025814387947320938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2204611599445343, + 0.25228235125541687, + 0.11134069412946701, + 0.030853793025016785, + 0.09121925383806229, + 0.048895370215177536, + 0.025303833186626434, + 0.03570609539747238, + 0.07457150518894196, + 0.10936595499515533 + ], + "credit_map": [ + 0.02421104535460472, + 0.030781298875808716, + 0.019820736721158028, + 0.06269950419664383, + 0.019921664148569107, + 0.01106227282434702, + 0.01887807995080948, + 0.03308655694127083, + 0.020476913079619408, + 0.05438905209302902, + 0.016549725085496902, + 0.02829386480152607, + 0.048022955656051636, + 0.03372935578227043, + 0.028367510065436363, + 0.022452984005212784, + 0.006960275582969189, + 0.03059256821870804, + 0.016061345115303993, + 0.010082157328724861, + 0.024078628048300743, + 0.006230642087757587, + 0.009076070040464401, + 0.008259005844593048, + 0.0033402007538825274, + 0.004674258641898632, + 0.004656325560063124, + 0.0021948127541691065, + 0.005831046029925346, + 0.002074519172310829, + 0.004862621426582336, + 0.0032200105488300323, + 0.015176884829998016, + 0.012062203139066696, + 0.006056871265172958, + 0.007244279608130455, + 0.010188641026616096, + 0.006872666534036398, + 0.024585647508502007, + 0.00903206318616867, + 0.005967106204479933, + 0.0065779173746705055, + 0.00701909651979804, + 0.00600416911765933, + 0.00783130619674921, + 0.002403847873210907, + 0.007356325630098581, + 0.005735599435865879, + 0.00308065302670002, + 0.0036019128747284412, + 0.003733324585482478, + 0.001976031344383955, + 0.00329839950427413, + 0.003748850664123893, + 0.003117120126262307, + 0.0027475403621792793, + 0.004266614560037851, + 0.00637262174859643, + 0.005051025655120611, + 0.004074550699442625, + 0.00404743617400527, + 0.004918455146253109, + 0.0034695465583354235, + 0.0035058464854955673, + 0.010591422207653522, + 0.012767323292791843, + 0.008058048784732819, + 0.007743555586785078, + 0.006504821591079235, + 0.007121387869119644, + 0.008190546184778214, + 0.01359440479427576, + 0.036289870738983154, + 0.021055471152067184, + 0.01259499229490757, + 0.011812649667263031, + 0.00786285288631916, + 0.005939350463449955, + 0.00768823828548193, + 0.006122523918747902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.025814387947320938 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive locomotive drive wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking slowly along the wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wheel made of dark steel with visible oil smears and rust speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there orange and yellow sparks exploding from the contact point with the rain-soaked rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of black crushed stone ballast under a diffuse gray overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005962679628282785 + ], + "opposite_logprob_tokens": [ + -5.13096284866333 + ], + "logprob_reward": -0.005962679628282785, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005962679628282785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28319239616394043, + 0.13773556053638458, + 0.09700506180524826, + 0.06449577212333679, + 0.08261656016111374, + 0.061252180486917496, + 0.059606049209833145, + 0.06676188111305237, + 0.06852030754089355, + 0.07881435006856918 + ], + "credit_map": [ + 0.030928997322916985, + 0.03926641866564751, + 0.027449674904346466, + 0.07459306716918945, + 0.02425670623779297, + 0.01635654829442501, + 0.02799631468951702, + 0.042344607412815094, + 0.010678100399672985, + 0.022665472701191902, + 0.015002576634287834, + 0.016471361741423607, + 0.023254966363310814, + 0.01826978661119938, + 0.020502431318163872, + 0.010890869423747063, + 0.008938550017774105, + 0.02570284530520439, + 0.011454937048256397, + 0.005463424604386091, + 0.022186189889907837, + 0.00668899342417717, + 0.009977754205465317, + 0.006592359859496355, + 0.006965136621147394, + 0.00907883234322071, + 0.011268918402493, + 0.005906402133405209, + 0.011782380752265453, + 0.0048631406389176846, + 0.009629256092011929, + 0.005001687444746494, + 0.014717805199325085, + 0.010777383111417294, + 0.007061338983476162, + 0.006333581637591124, + 0.007566261105239391, + 0.006372485775500536, + 0.019676895812153816, + 0.010110796429216862, + 0.010096090845763683, + 0.0077314069494605064, + 0.00975679699331522, + 0.006757253315299749, + 0.005858929827809334, + 0.00328469043597579, + 0.006860689725726843, + 0.010906313546001911, + 0.007633003871887922, + 0.006994995288550854, + 0.00922352634370327, + 0.006340705323964357, + 0.005631333217024803, + 0.005712257232517004, + 0.006674566771835089, + 0.011395653709769249, + 0.011189687997102737, + 0.00829888042062521, + 0.008768814615905285, + 0.008101465180516243, + 0.0065293945372104645, + 0.006262601353228092, + 0.007360957097262144, + 0.010250067338347435, + 0.010284852236509323, + 0.00977355893701315, + 0.010431043803691864, + 0.009457663632929325, + 0.005054627079516649, + 0.007364057004451752, + 0.007823743857443333, + 0.008330747485160828, + 0.024490153416991234, + 0.01419010479003191, + 0.01091061718761921, + 0.00804969947785139, + 0.0051693483255803585, + 0.0056768679060041904, + 0.005543963517993689, + 0.0047835842706263065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.005962679628282785 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive locomotive drive wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking slowly along the wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wheel made of dark steel with visible oil smears and rust speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there orange and yellow sparks exploding from the contact point with the rain-soaked rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of black crushed stone ballast under a diffuse gray overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00595379201695323 + ], + "opposite_logprob_tokens": [ + -5.130953788757324 + ], + "logprob_reward": -0.00595379201695323, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00595379201695323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24926447868347168, + 0.13706502318382263, + 0.08587881177663803, + 0.03694524988532066, + 0.043046947568655014, + 0.11201317608356476, + 0.0344838947057724, + 0.066119484603405, + 0.11204945296049118, + 0.12313344329595566 + ], + "credit_map": [ + 0.031895097345113754, + 0.030056994408369064, + 0.021395519375801086, + 0.07531735301017761, + 0.02018703706562519, + 0.010584796778857708, + 0.021000973880290985, + 0.03882668912410736, + 0.015492726117372513, + 0.023496005684137344, + 0.012481904588639736, + 0.014577394351363182, + 0.02639048919081688, + 0.017438611015677452, + 0.014779741875827312, + 0.012408149428665638, + 0.005978096276521683, + 0.015401666052639484, + 0.011208653450012207, + 0.008269875310361385, + 0.014383368194103241, + 0.011613858863711357, + 0.011220820248126984, + 0.007802469655871391, + 0.0030501834116876125, + 0.005576895084232092, + 0.0056912885047495365, + 0.003688859986141324, + 0.008081370033323765, + 0.0020540570840239525, + 0.005837731994688511, + 0.0029648642521351576, + 0.01196903083473444, + 0.005889056716114283, + 0.0037075087893754244, + 0.0028086102101951838, + 0.0030250810086727142, + 0.004020498134195805, + 0.0045577832497656345, + 0.0070693776942789555, + 0.02361239865422249, + 0.015600613318383694, + 0.017261875793337822, + 0.01282646507024765, + 0.011824567802250385, + 0.0054950835183262825, + 0.012027833610773087, + 0.01336433831602335, + 0.0049651493318378925, + 0.004814193118363619, + 0.004131603520363569, + 0.004405202344059944, + 0.003349045757204294, + 0.0043654665350914, + 0.004401850514113903, + 0.00405138498172164, + 0.008750062435865402, + 0.00991047453135252, + 0.008553396910429, + 0.006076920311897993, + 0.008346203714609146, + 0.009253671392798424, + 0.009684019722044468, + 0.00554473465308547, + 0.016973145306110382, + 0.013019324280321598, + 0.013916810043156147, + 0.02135389856994152, + 0.007555325981229544, + 0.010212033055722713, + 0.01226726546883583, + 0.016751645132899284, + 0.0360734760761261, + 0.021564055234193802, + 0.013764231465756893, + 0.013742971234023571, + 0.009252327494323254, + 0.008512700907886028, + 0.012647513300180435, + 0.007576168514788151 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.00595379201695323 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive locomotive drive wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking slowly along the wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wheel made of dark steel with visible oil smears and rust speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there orange and yellow sparks exploding from the contact point with the rain-soaked rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of black crushed stone ballast under a diffuse gray overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054967835545539856 + ], + "opposite_logprob_tokens": [ + -2.9299678802490234 + ], + "logprob_reward": -0.054967835545539856, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054967835545539856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13677681982517242, + 0.10109449923038483, + 0.09225255250930786, + 0.052538808435201645, + 0.09323275089263916, + 0.078457772731781, + 0.06629813462495804, + 0.07047083973884583, + 0.08204676955938339, + 0.22683101892471313 + ], + "credit_map": [ + 0.011412217281758785, + 0.013539746403694153, + 0.016256874427199364, + 0.028619889169931412, + 0.02562575973570347, + 0.010547513142228127, + 0.010296502150595188, + 0.020478317514061928, + 0.022048266604542732, + 0.019128328189253807, + 0.011182737536728382, + 0.010629180818796158, + 0.010712934657931328, + 0.008312897756695747, + 0.01029050350189209, + 0.008789647370576859, + 0.009502945467829704, + 0.019311605021357536, + 0.011148257181048393, + 0.008357996121048927, + 0.017800820991396904, + 0.009630905464291573, + 0.00976148433983326, + 0.006738533731549978, + 0.004953572992235422, + 0.008027427829802036, + 0.006326391827315092, + 0.004584930371493101, + 0.010087023489177227, + 0.0031016096472740173, + 0.011590436100959778, + 0.0038674098905175924, + 0.017258770763874054, + 0.013631829991936684, + 0.009249168448150158, + 0.007659100461751223, + 0.0100028021261096, + 0.007987884804606438, + 0.015766803175210953, + 0.011676378548145294, + 0.012524588964879513, + 0.010263579897582531, + 0.011589597910642624, + 0.0058900341391563416, + 0.017069201916456223, + 0.0049746534787118435, + 0.008967872709035873, + 0.007178238593041897, + 0.006286268588155508, + 0.00869689416140318, + 0.0077894446440041065, + 0.004887823481112719, + 0.0069940160028636456, + 0.012692136690020561, + 0.006154799368232489, + 0.012796750292181969, + 0.012156227603554726, + 0.007401464972645044, + 0.01138048805296421, + 0.010043750517070293, + 0.007168347015976906, + 0.006097752135246992, + 0.009346737526357174, + 0.006876066792756319, + 0.01194160245358944, + 0.009910497814416885, + 0.01136960182338953, + 0.01708199270069599, + 0.005330319982022047, + 0.005977443885058165, + 0.010361967608332634, + 0.010073332116007805, + 0.0767034962773323, + 0.04573579505085945, + 0.026361465454101562, + 0.022142674773931503, + 0.014435057528316975, + 0.010197333991527557, + 0.02074403129518032, + 0.010511157102882862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.054967835545539856 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive locomotive drive wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking slowly along the wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wheel made of dark steel with visible oil smears and rust speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there orange and yellow sparks exploding from the contact point with the rain-soaked rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of black crushed stone ballast under a diffuse gray overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011066253297030926 + ], + "opposite_logprob_tokens": [ + -4.511066436767578 + ], + "logprob_reward": -0.011066253297030926, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011066253297030926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1780901849269867, + 0.12046557664871216, + 0.11217660456895828, + 0.04173479601740837, + 0.04332530125975609, + 0.05725915729999542, + 0.048052869737148285, + 0.07042958587408066, + 0.1267027109861374, + 0.20176321268081665 + ], + "credit_map": [ + 0.01615115813910961, + 0.028633657842874527, + 0.023226140066981316, + 0.033453088253736496, + 0.019728373736143112, + 0.014293292537331581, + 0.014501192606985569, + 0.028103286400437355, + 0.021495265886187553, + 0.019686322659254074, + 0.010346324183046818, + 0.013475494459271431, + 0.01036351639777422, + 0.016647815704345703, + 0.013586721383035183, + 0.014864115044474602, + 0.008557656779885292, + 0.026385558769106865, + 0.012211517430841923, + 0.007578314747661352, + 0.025009609758853912, + 0.012694713659584522, + 0.0106122437864542, + 0.009126986376941204, + 0.004329198505729437, + 0.006032286211848259, + 0.007293076254427433, + 0.0036910814233124256, + 0.005606840830296278, + 0.003182765794917941, + 0.0072585646994411945, + 0.00434098020195961, + 0.009444649331271648, + 0.005663572810590267, + 0.00362716568633914, + 0.0033483451697975397, + 0.00435451976954937, + 0.003415449056774378, + 0.0074828932993113995, + 0.005988705437630415, + 0.006804656703025103, + 0.0059932987205684185, + 0.013140317983925343, + 0.005768448580056429, + 0.00830165296792984, + 0.0029973755590617657, + 0.010016807354986668, + 0.004236598964780569, + 0.004793554078787565, + 0.006859106011688709, + 0.005235114600509405, + 0.005363003816455603, + 0.006482629105448723, + 0.0087653286755085, + 0.004684171173721552, + 0.005869966000318527, + 0.010075477883219719, + 0.0073470682837069035, + 0.013182396069169044, + 0.009061693213880062, + 0.007687440142035484, + 0.007656661793589592, + 0.008657090365886688, + 0.006761753931641579, + 0.014201333746314049, + 0.016844887286424637, + 0.013947494328022003, + 0.015988003462553024, + 0.00714212516322732, + 0.011420746333897114, + 0.018434328958392143, + 0.028723791241645813, + 0.06766417622566223, + 0.039457809180021286, + 0.02415892295539379, + 0.016193488612771034, + 0.015959450975060463, + 0.011832742020487785, + 0.015365559607744217, + 0.011131059378385544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.011066253297030926 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This low-angle, extreme close-up captures the massive drive wheel of a massive locomotive in breathtakingly high-speed slow motion, as the camera slowly tracks along it. Forged from heavy, dark steel, its surface smeared with oil and speckled with rust, the wheel slams violently against the polished, rain-soaked rail. It struggles agonizingly slowly to complete its rotation, its immense force causing it to slip and grind, sending a brilliant explosion of orange and yellow sparks flying from the point of contact. Water droplets clinging to the wheel's metal surface are atomized by the intense vibrations. In the background, the black crushed stone ballast of the track drenches and gleams in the diffuse gray light of an overcast sky. The entire scene, immersed in a cool, monochromatic palette of industrial grays and blacks, creates an atmosphere of tension and power, a raw, mechanical struggle against immense physical resistance.", + "seed": 15363, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_13/index_2570_seed_15363_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02514718286693096, + "vlm_logprob": -0.02514718286693096, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.02514718286693096, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive locomotive drive wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking slowly along the wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wheel made of dark steel with visible oil smears and rust speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there orange and yellow sparks exploding from the contact point with the rain-soaked rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of black crushed stone ballast under a diffuse gray overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01703980192542076 + ], + "opposite_logprob_tokens": [ + -5.142039775848389 + ], + "logprob_reward": -0.01703980192542076, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.01703980192542076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8464223146438599, + 0.021976353600621223, + 0.014040245674550533, + 0.019859513267874718, + 0.019665056839585304, + 0.012659583240747452, + 0.01716005615890026, + 0.011708751320838928, + 0.010918475687503815, + 0.025589723140001297 + ], + "credit_map": [ + 0.06914087384939194, + 0.32971885800361633, + 0.03472514823079109, + 0.15032441914081573, + 0.04224223271012306, + 0.12201034277677536, + 0.061240844428539276, + 0.03701956570148468, + 0.0018347417935729027, + 0.0035238992422819138, + 0.002406535902991891, + 0.0035660136491060257, + 0.0019726785831153393, + 0.0022219575475901365, + 0.002393788890913129, + 0.004056734964251518, + 0.0025224238634109497, + 0.0021960893645882607, + 0.0014277022564783692, + 0.0017909554298967123, + 0.0017546936869621277, + 0.0021664206869900227, + 0.0010753367096185684, + 0.0011066236766055226, + 0.0015221287030726671, + 0.0020672818645834923, + 0.0019761729054152966, + 0.005607379600405693, + 0.0026142040733247995, + 0.0023340126499533653, + 0.0019333070376887918, + 0.001805023173801601, + 0.0019618519581854343, + 0.002780013717710972, + 0.0022268614266067743, + 0.0019931879360228777, + 0.0029384384397417307, + 0.002200349234044552, + 0.003073852276429534, + 0.002490500221028924, + 0.0011113823857158422, + 0.002143239136785269, + 0.0013372303219512105, + 0.0015766277210786939, + 0.0013586130226030946, + 0.0018433976219967008, + 0.0015257478225976229, + 0.0017633440438658, + 0.0026235266122967005, + 0.001780307968147099, + 0.0026680035516619682, + 0.0019203195115551353, + 0.0021553400438278913, + 0.0026344077195972204, + 0.0017130088526755571, + 0.00166514131706208, + 0.0010397748555988073, + 0.0019410731038078666, + 0.001416960614733398, + 0.0012660999782383442, + 0.0018001430435106158, + 0.001821710611693561, + 0.0010637416271492839, + 0.0013592456234619021, + 0.0010341706220060587, + 0.0016712939832359552, + 0.0013676780508831143, + 0.00124607072211802, + 0.0020114253275096416, + 0.001445347792468965, + 0.0007472913130186498, + 0.0013951961882412434, + 0.0013916280586272478, + 0.005690701771527529, + 0.002148758852854371, + 0.0027468926273286343, + 0.003809017827734351, + 0.004784353077411652, + 0.0018467600457370281, + 0.003171608317643404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.01703980192542076 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive locomotive drive wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking slowly along the wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wheel made of dark steel with visible oil smears and rust speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there orange and yellow sparks exploding from the contact point with the rain-soaked rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of black crushed stone ballast under a diffuse gray overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008652222342789173 + ], + "opposite_logprob_tokens": [ + -4.758652210235596 + ], + "logprob_reward": -0.008652222342789173, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008652222342789173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.811822772026062, + 0.02496783249080181, + 0.016131969168782234, + 0.019320162013173103, + 0.024867679923772812, + 0.013029365800321102, + 0.018669480457901955, + 0.016778673976659775, + 0.0204550102353096, + 0.03395702317357063 + ], + "credit_map": [ + 0.0652582123875618, + 0.3135380446910858, + 0.0333983488380909, + 0.1488751769065857, + 0.04121823236346245, + 0.11612289398908615, + 0.05770042538642883, + 0.035711389034986496, + 0.0021916332188993692, + 0.0021906443871557713, + 0.004810146521776915, + 0.005167062859982252, + 0.002774507738649845, + 0.00314346794039011, + 0.002482982585206628, + 0.002207387937232852, + 0.001481165993027389, + 0.0024526577908545732, + 0.0015684509417042136, + 0.0014956864761188626, + 0.002259777393192053, + 0.0024939188733696938, + 0.002556080697104335, + 0.0018242312362417579, + 0.0012613714206963778, + 0.0015065453480929136, + 0.002330567454919219, + 0.006666931789368391, + 0.002608029404655099, + 0.0015675589675083756, + 0.001996241509914398, + 0.0013829158851876855, + 0.0023007828276604414, + 0.002503035357221961, + 0.0018706533592194319, + 0.002801570575684309, + 0.003428656840696931, + 0.0031270612962543964, + 0.004565018694847822, + 0.004270900972187519, + 0.0016798950964584947, + 0.0014099973486736417, + 0.001616199966520071, + 0.0014895113417878747, + 0.0017378650372847915, + 0.0016887609381228685, + 0.0016978862695395947, + 0.0017092488706111908, + 0.002279232954606414, + 0.0019284938462078571, + 0.004286884795874357, + 0.0018110043602064252, + 0.0019739328417927027, + 0.002535599283874035, + 0.002007201313972473, + 0.0018471332732588053, + 0.0016451956471428275, + 0.0026455125771462917, + 0.002313190372660756, + 0.0018920082366093993, + 0.0023267322685569525, + 0.002800568239763379, + 0.0016293675871565938, + 0.0015260974178090692, + 0.0012289137812331319, + 0.0029563275165855885, + 0.0023067051079124212, + 0.002285473747178912, + 0.0048447418957948685, + 0.0030267133843153715, + 0.0016559470677748322, + 0.002150187734514475, + 0.0016838745214045048, + 0.0059968093410134315, + 0.0024613388814032078, + 0.0029480978846549988, + 0.005302474368363619, + 0.008251099847257137, + 0.002496221801266074, + 0.0048171053640544415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.008652222342789173 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive locomotive drive wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking slowly along the wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wheel made of dark steel with visible oil smears and rust speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there orange and yellow sparks exploding from the contact point with the rain-soaked rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of black crushed stone ballast under a diffuse gray overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006739386357367039 + ], + "opposite_logprob_tokens": [ + -5.006739616394043 + ], + "logprob_reward": -0.006739386357367039, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006739386357367039, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6426815390586853, + 0.03373733162879944, + 0.06808200478553772, + 0.0430021695792675, + 0.09445487707853317, + 0.025459716096520424, + 0.04326394945383072, + 0.01500505581498146, + 0.015672102570533752, + 0.018641196191310883 + ], + "credit_map": [ + 0.04967842996120453, + 0.24096593260765076, + 0.028411880135536194, + 0.1155015379190445, + 0.03586834669113159, + 0.09533317387104034, + 0.0463881716132164, + 0.030534135177731514, + 0.0035138654056936502, + 0.0046923463232815266, + 0.003981794696301222, + 0.004289034754037857, + 0.0040533049032092094, + 0.004011252894997597, + 0.0045888349413871765, + 0.0046069021336734295, + 0.009105008095502853, + 0.010672200471162796, + 0.005364537239074707, + 0.008327718824148178, + 0.009966908022761345, + 0.00949273444712162, + 0.0077637541107833385, + 0.0073891496285796165, + 0.006198204588145018, + 0.005282557103782892, + 0.0033349047880619764, + 0.004956306889653206, + 0.006003865040838718, + 0.00887189619243145, + 0.004414860624819994, + 0.003939579240977764, + 0.009770968928933144, + 0.00899973139166832, + 0.007933388464152813, + 0.015127993188798428, + 0.012308585457503796, + 0.010845555923879147, + 0.013648285530507565, + 0.015820376574993134, + 0.002360099693760276, + 0.0029659641440957785, + 0.003487928304821253, + 0.003159863641485572, + 0.0033409451134502888, + 0.002756342990323901, + 0.003892046632245183, + 0.003496528370305896, + 0.004023940768092871, + 0.00690604979172349, + 0.006371585186570883, + 0.00418037548661232, + 0.005939563736319542, + 0.006753280758857727, + 0.005395946558564901, + 0.0036932143848389387, + 0.0013597056968137622, + 0.0020158477127552032, + 0.001990601187571883, + 0.0014971670461818576, + 0.0024361421819776297, + 0.002487637335434556, + 0.0013619756791740656, + 0.0018559809541329741, + 0.0013715538661926985, + 0.002185169607400894, + 0.0030065930914133787, + 0.001838216558098793, + 0.002645346336066723, + 0.0018360387766733766, + 0.000982142984867096, + 0.001807043212465942, + 0.0009535475983284414, + 0.0038272978272289038, + 0.0016644971910864115, + 0.002144476165995002, + 0.0026724219787865877, + 0.003966969437897205, + 0.0013134627370163798, + 0.002098525408655405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.006739386357367039 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive locomotive drive wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking slowly along the wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wheel made of dark steel with visible oil smears and rust speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there orange and yellow sparks exploding from the contact point with the rain-soaked rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of black crushed stone ballast under a diffuse gray overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002726197242737 + ], + "opposite_logprob_tokens": [ + -2.6950273513793945 + ], + "logprob_reward": -0.07002726197242737, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002726197242737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7643232345581055, + 0.042774301022291183, + 0.0211787112057209, + 0.028686610981822014, + 0.03977685794234276, + 0.022730374708771706, + 0.020344775170087814, + 0.02672460861504078, + 0.011278009042143822, + 0.022182589396834373 + ], + "credit_map": [ + 0.05966421216726303, + 0.2904108166694641, + 0.03210333362221718, + 0.14948880672454834, + 0.03762373328208923, + 0.10711660981178284, + 0.053745172917842865, + 0.0341704785823822, + 0.004503062926232815, + 0.00499908160418272, + 0.005580648314207792, + 0.006964045576751232, + 0.003604684490710497, + 0.00581195717677474, + 0.004655918106436729, + 0.006654898636043072, + 0.0026322323828935623, + 0.0036086367908865213, + 0.0018307839054614305, + 0.0023046249989420176, + 0.0030275420285761356, + 0.00347120501101017, + 0.0024097315035760403, + 0.0018939513247460127, + 0.0029456480406224728, + 0.0024892031215131283, + 0.002483075950294733, + 0.006581543013453484, + 0.0036511817015707493, + 0.00533420592546463, + 0.0030486267060041428, + 0.0021531216334551573, + 0.0037058787420392036, + 0.003266099141910672, + 0.003045107237994671, + 0.004707242827862501, + 0.0058997515588998795, + 0.005063467659056187, + 0.006636565085500479, + 0.007452741730958223, + 0.0017440422670915723, + 0.002785405609756708, + 0.0021047189366072416, + 0.0025575528852641582, + 0.00425958214327693, + 0.0030677448958158493, + 0.003542750608175993, + 0.0026685725897550583, + 0.002672127215191722, + 0.0027550228405743837, + 0.002342815976589918, + 0.0025404486805200577, + 0.003204475389793515, + 0.0029526755679398775, + 0.002227491233497858, + 0.001649715006351471, + 0.0024953957181423903, + 0.005100942216813564, + 0.0034029013477265835, + 0.003391505451872945, + 0.0037148783449083567, + 0.003959542606025934, + 0.0024510459043085575, + 0.0022083919029682875, + 0.0010209801839664578, + 0.0016226540319621563, + 0.002072183880954981, + 0.0015195620944723487, + 0.001600629067979753, + 0.001367417280562222, + 0.0008065606816671789, + 0.001268020598217845, + 0.0014604127500206232, + 0.003796449163928628, + 0.0020639870781451464, + 0.0020790661219507456, + 0.0031564573291689157, + 0.004897939041256905, + 0.0015998244052752852, + 0.0031284503638744354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.07002726197242737 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive locomotive drive wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking slowly along the wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wheel made of dark steel with visible oil smears and rust speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there orange and yellow sparks exploding from the contact point with the rain-soaked rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of black crushed stone ballast under a diffuse gray overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023277241736650467 + ], + "opposite_logprob_tokens": [ + -3.7732772827148438 + ], + "logprob_reward": -0.023277241736650467, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023277241736650467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.9036605954170227, + 0.020784752443432808, + 0.007891077548265457, + 0.007797733880579472, + 0.008774721063673496, + 0.013546447269618511, + 0.021845223382115364, + 0.004438711795955896, + 0.006868510041385889, + 0.004392287228256464 + ], + "credit_map": [ + 0.0718909502029419, + 0.3468249440193176, + 0.03769100457429886, + 0.171251118183136, + 0.044613201171159744, + 0.12653854489326477, + 0.06472659856081009, + 0.04012414440512657, + 0.002517023356631398, + 0.0026550735346972942, + 0.0022724007721990347, + 0.0023652445524930954, + 0.0018412090139463544, + 0.0025757532566785812, + 0.0018572629196569324, + 0.0047007836401462555, + 0.000913294788915664, + 0.0010273385560140014, + 0.000946374493651092, + 0.0008360647480003536, + 0.0010573866311460733, + 0.0013728494523093104, + 0.0009864898165687919, + 0.0007512776646763086, + 0.000973768881522119, + 0.0009577735909260809, + 0.0008541826391592622, + 0.0011146112810820341, + 0.0010418437886983156, + 0.0010778671130537987, + 0.0007532701129093766, + 0.0010244158329442143, + 0.0011551580391824245, + 0.0010901793139055371, + 0.0011120568960905075, + 0.0013130332808941603, + 0.001212762319482863, + 0.0009827256435528398, + 0.001131566590629518, + 0.0007772387471050024, + 0.0014529795153066516, + 0.0018691285513341427, + 0.0023480188101530075, + 0.0017994253430515528, + 0.0015221540816128254, + 0.0013426749501377344, + 0.0014700896572321653, + 0.0017419755458831787, + 0.0036223733332008123, + 0.0021364728454500437, + 0.004113077186048031, + 0.002118033589795232, + 0.0027899816632270813, + 0.0028649859596043825, + 0.002239394001662731, + 0.0019609008450061083, + 0.000522579241078347, + 0.000734504428692162, + 0.0006187930703163147, + 0.0005702574853785336, + 0.0006394887459464371, + 0.0006638906197622418, + 0.0003062384785152972, + 0.0003829590859822929, + 0.000537281041033566, + 0.0012330058962106705, + 0.0008151950314640999, + 0.0009191319113597274, + 0.0014580817660316825, + 0.0008615651167929173, + 0.0004993219044990838, + 0.0005449267337098718, + 0.00025476343580521643, + 0.0007619521347805858, + 0.00037366169271990657, + 0.0005324985831975937, + 0.0006703396793454885, + 0.0009293612674809992, + 0.0003115949220955372, + 0.0005581150180660188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.023277241736650467 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_14/index_tensor([6789, 6789])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_14/index_tensor([6789, 6789])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..79c2dce5a3f5fcc4b8f9ca7f258407c36fa05d38 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_14/index_tensor([6789, 6789])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "This first-person point-of-view (RST) shot is captured from a camera mounted low on the bow of a vibrant red kayak, plunging headfirst into the churning rapids of a powerful river. Rendered in the style of high-speed nature photography, the entire scene captures the physics of motion and light with hyperreal clarity. The camera pitches and dips dramatically with each powerful surge of the water, angling downward into deep troughs and then slamming upward on foaming crests, conveying the immense, chaotic force of the current. Frozen in harsh, direct sunlight, each drop of water kicked up by the boat's bow appears as a crystalline, shimmering sphere, refracting light through the turbulent turquoise and white water like tiny prisms. The kayak is navigating a deep, sun-scorched canyon, flanked by towering, layered, reddish-orange sandstone walls rising vertically on either side. As the kayak surges forward, these ancient, weathered canyon walls blur past in a dazzling yet sharp spectacle of striped colors, creating a profound sense of speed and raw natural power that is both exhilarating and overwhelming.", + "seed": 17979, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_14/index_6789_seed_17979_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.20520684625953436, + "vlm_logprob": -0.20520684625953436, + "vlm_yes_no_margin": 2.35, + "vlm_token_credit": -0.20520684625953436, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person point-of-view mounted on a vibrant red kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kayak navigating through churning rapids in a deep, sun-scorched canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are towering, layered, reddish-orange sandstone walls visible rising vertically on either side of the river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do water droplets kicked up by the bow appear as crystalline, shimmering spheres refracting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pitch and dip dramatically with the water's surges, angling downward into troughs and upward on crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04099005460739136 + ], + "opposite_logprob_tokens": [ + -3.290990114212036 + ], + "logprob_reward": -0.04099005460739136, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04099005460739136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13089416921138763, + 0.09174598753452301, + 0.04829731583595276, + 0.24029740691184998, + 0.05572696402668953, + 0.056335240602493286, + 0.07265673577785492, + 0.0732373595237732, + 0.18202349543571472, + 0.04878537356853485 + ], + "credit_map": [ + 0.016497593373060226, + 0.01085418090224266, + 0.021624233573675156, + 0.019222619011998177, + 0.02417050115764141, + 0.013789953663945198, + 0.01328881736844778, + 0.011446259915828705, + 0.011430694721639156, + 0.012207014486193657, + 0.011009602807462215, + 0.013914288952946663, + 0.01474084798246622, + 0.008153858594596386, + 0.0107734939083457, + 0.009516186080873013, + 0.004298178944736719, + 0.012357063591480255, + 0.004352741874754429, + 0.004026414826512337, + 0.00824481900781393, + 0.003990771248936653, + 0.005275750532746315, + 0.005751574411988258, + 0.03579913452267647, + 0.10394813865423203, + 0.023565029725432396, + 0.006404538173228502, + 0.01218035165220499, + 0.008322804234921932, + 0.022138578817248344, + 0.027938827872276306, + 0.014148030430078506, + 0.00658849161118269, + 0.01177110057324171, + 0.003963068593293428, + 0.00457623740658164, + 0.0032554930076003075, + 0.0037673756014555693, + 0.007657165173441172, + 0.009299008175730705, + 0.012168974615633488, + 0.00636430736631155, + 0.00483111385256052, + 0.00658368319272995, + 0.004037105944007635, + 0.004292461089789867, + 0.008758584968745708, + 0.01022791862487793, + 0.009989495389163494, + 0.0120165403932333, + 0.012171683833003044, + 0.006257354747503996, + 0.007483988534659147, + 0.006574481725692749, + 0.007935266010463238, + 0.00686583062633872, + 0.00744393328204751, + 0.013174745254218578, + 0.013841398991644382, + 0.006081790663301945, + 0.0072792344726622105, + 0.01238671038299799, + 0.006163712125271559, + 0.014260496012866497, + 0.011108215898275375, + 0.024592041969299316, + 0.016461461782455444, + 0.01668570376932621, + 0.03160989657044411, + 0.029961757361888885, + 0.03734390810132027, + 0.005994240287691355, + 0.0056146434508264065, + 0.00447780080139637, + 0.00887201726436615, + 0.004655166063457727, + 0.005078385584056377, + 0.006349233444780111, + 0.007743885740637779 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.04099005460739136 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person point-of-view mounted on a vibrant red kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kayak navigating through churning rapids in a deep, sun-scorched canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are towering, layered, reddish-orange sandstone walls visible rising vertically on either side of the river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do water droplets kicked up by the bow appear as crystalline, shimmering spheres refracting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pitch and dip dramatically with the water's surges, angling downward into troughs and upward on crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933507919311523 + ], + "opposite_logprob_tokens": [ + -0.6933507919311523 + ], + "logprob_reward": -0.6933507919311523, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933507919311523, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.13942307233810425, + 0.03540915250778198, + 0.08637059479951859, + 0.1886991262435913, + 0.11806544661521912, + 0.029875895008444786, + 0.10280706733465195, + 0.04622446373105049, + 0.09071771800518036, + 0.16240742802619934 + ], + "credit_map": [ + 0.021452464163303375, + 0.013461408205330372, + 0.023507151752710342, + 0.0244390107691288, + 0.02139747142791748, + 0.012017260305583477, + 0.011439667083323002, + 0.011708651669323444, + 0.0037369688507169485, + 0.007611675653606653, + 0.0039166356436908245, + 0.005898964125663042, + 0.003982068970799446, + 0.0034689365420490503, + 0.0029902588576078415, + 0.003803645260632038, + 0.007094037719070911, + 0.007441841531544924, + 0.006717743817716837, + 0.0215100459754467, + 0.022842368111014366, + 0.005562600679695606, + 0.008180155418813229, + 0.007021816447377205, + 0.019225722178816795, + 0.07836863398551941, + 0.04264821857213974, + 0.006285227835178375, + 0.009319152683019638, + 0.005657622590661049, + 0.009010623209178448, + 0.01818394474685192, + 0.039131853729486465, + 0.008921164087951183, + 0.03119305893778801, + 0.00891116727143526, + 0.006662657484412193, + 0.006046419497579336, + 0.005474691744893789, + 0.011724435724318027, + 0.004970787558704615, + 0.004870863631367683, + 0.003462972119450569, + 0.0049810647033154964, + 0.0024142966140061617, + 0.0019584510009735823, + 0.002237093634903431, + 0.0049803657457232475, + 0.019412437453866005, + 0.021557994186878204, + 0.01251477375626564, + 0.016725506633520126, + 0.007936013862490654, + 0.007063787896186113, + 0.008724190294742584, + 0.008872367441654205, + 0.007518202997744083, + 0.009844057261943817, + 0.0067246900871396065, + 0.008979037404060364, + 0.0034116029273718596, + 0.0027933260425925255, + 0.00428732018917799, + 0.0026662286836653948, + 0.00759028410539031, + 0.006344479974359274, + 0.01199998427182436, + 0.010616721585392952, + 0.0068093640729784966, + 0.014992289245128632, + 0.015460381284356117, + 0.016904223710298538, + 0.01621220074594021, + 0.019999008625745773, + 0.020874688401818275, + 0.028224054723978043, + 0.010108070448040962, + 0.014588107354938984, + 0.01876688003540039, + 0.033634431660175323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.6933507919311523 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person point-of-view mounted on a vibrant red kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kayak navigating through churning rapids in a deep, sun-scorched canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are towering, layered, reddish-orange sandstone walls visible rising vertically on either side of the river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do water droplets kicked up by the bow appear as crystalline, shimmering spheres refracting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pitch and dip dramatically with the water's surges, angling downward into troughs and upward on crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20161330699920654 + ], + "opposite_logprob_tokens": [ + -1.7016133069992065 + ], + "logprob_reward": -0.20161330699920654, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20161330699920654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14496774971485138, + 0.060742273926734924, + 0.05367160961031914, + 0.20365440845489502, + 0.1929556429386139, + 0.02243961952626705, + 0.10373365134000778, + 0.06681837886571884, + 0.07389377057552338, + 0.07712288200855255 + ], + "credit_map": [ + 0.012456224299967289, + 0.013118688017129898, + 0.02764292247593403, + 0.02696014568209648, + 0.027006611227989197, + 0.015193002298474312, + 0.012507930397987366, + 0.010082220658659935, + 0.0074819461442530155, + 0.010946986265480518, + 0.006819508504122496, + 0.008359024301171303, + 0.009469142183661461, + 0.005514941178262234, + 0.006353380158543587, + 0.005797350779175758, + 0.004397610668092966, + 0.010360457003116608, + 0.004749973304569721, + 0.00866629183292389, + 0.005403969436883926, + 0.006826323457062244, + 0.006950653623789549, + 0.006316327024251223, + 0.02612612023949623, + 0.08528324216604233, + 0.03404765576124191, + 0.006530513055622578, + 0.010804872959852219, + 0.007085879798978567, + 0.01193277072161436, + 0.021843362599611282, + 0.06596889346837997, + 0.013763259164988995, + 0.05373302847146988, + 0.013386452570557594, + 0.010874854400753975, + 0.009963151067495346, + 0.006769233848899603, + 0.01849677413702011, + 0.00363338366150856, + 0.004064510110765696, + 0.003085001837462187, + 0.0029584872536361217, + 0.0016527745174244046, + 0.0015098146395757794, + 0.0022472890559583902, + 0.003288359148427844, + 0.015016861259937286, + 0.019431106746196747, + 0.015130696818232536, + 0.017020495608448982, + 0.009078782051801682, + 0.008240977302193642, + 0.009422867558896542, + 0.010391857475042343, + 0.0101534528657794, + 0.014631940983235836, + 0.009785134345293045, + 0.013179086148738861, + 0.005150137934833765, + 0.003725908463820815, + 0.006155005190521479, + 0.0040377103723585606, + 0.0063833873718976974, + 0.006339322775602341, + 0.01044605765491724, + 0.007611283101141453, + 0.00614321231842041, + 0.012255113571882248, + 0.009581196121871471, + 0.015134194865822792, + 0.008229807950556278, + 0.008188576437532902, + 0.013657101429998875, + 0.010349099524319172, + 0.005127418786287308, + 0.008146910928189754, + 0.007523037027567625, + 0.015900924801826477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.20161330699920654 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person point-of-view mounted on a vibrant red kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kayak navigating through churning rapids in a deep, sun-scorched canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are towering, layered, reddish-orange sandstone walls visible rising vertically on either side of the river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do water droplets kicked up by the bow appear as crystalline, shimmering spheres refracting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pitch and dip dramatically with the water's surges, angling downward into troughs and upward on crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899979501962662 + ], + "opposite_logprob_tokens": [ + -2.5789997577667236 + ], + "logprob_reward": -0.07899979501962662, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899979501962662, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22535818815231323, + 0.03164152801036835, + 0.07555171847343445, + 0.20949827134609222, + 0.09878358989953995, + 0.040429890155792236, + 0.05172301456332207, + 0.09091649949550629, + 0.11530066281557083, + 0.060796599835157394 + ], + "credit_map": [ + 0.03570660576224327, + 0.026636850088834763, + 0.03911515697836876, + 0.03845663741230965, + 0.026846734806895256, + 0.02083735354244709, + 0.01785886287689209, + 0.01989998109638691, + 0.004621070809662342, + 0.005254444200545549, + 0.004909988958388567, + 0.00474189268425107, + 0.0037272595800459385, + 0.002564516384154558, + 0.002311898861080408, + 0.003510459326207638, + 0.007636426482349634, + 0.010254871100187302, + 0.005622313357889652, + 0.015588969923555851, + 0.020229116082191467, + 0.00473744934424758, + 0.005488956812769175, + 0.00599362188950181, + 0.03597814962267876, + 0.08956817537546158, + 0.013480600900948048, + 0.006759193260222673, + 0.011774081736803055, + 0.007023695390671492, + 0.0195778738707304, + 0.025336522608995438, + 0.02996930107474327, + 0.009135248139500618, + 0.02312384359538555, + 0.006471423897892237, + 0.0066362167708575726, + 0.005310301203280687, + 0.005172132048755884, + 0.012965139001607895, + 0.00710565410554409, + 0.006477306596934795, + 0.0058077676221728325, + 0.006485715042799711, + 0.0031827494967728853, + 0.0021236848551779985, + 0.002416659379377961, + 0.006830358412116766, + 0.00863336119800806, + 0.006642364896833897, + 0.009391924366354942, + 0.0075289541855454445, + 0.004110184498131275, + 0.003839566372334957, + 0.005714490544050932, + 0.00586217176169157, + 0.012030897662043571, + 0.014070735312998295, + 0.01266203448176384, + 0.02170574851334095, + 0.00685251085087657, + 0.005784309934824705, + 0.011542017571628094, + 0.006268252152949572, + 0.012061438523232937, + 0.008283130824565887, + 0.012809056788682938, + 0.022857755422592163, + 0.00786503218114376, + 0.011135146021842957, + 0.026576021686196327, + 0.013713099993765354, + 0.00558737525716424, + 0.007948457263410091, + 0.004585850518196821, + 0.013345729559659958, + 0.006516100373119116, + 0.0073111094534397125, + 0.00865133572369814, + 0.006850648671388626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.07899979501962662 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person point-of-view mounted on a vibrant red kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kayak navigating through churning rapids in a deep, sun-scorched canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are towering, layered, reddish-orange sandstone walls visible rising vertically on either side of the river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do water droplets kicked up by the bow appear as crystalline, shimmering spheres refracting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pitch and dip dramatically with the water's surges, angling downward into troughs and upward on crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011080282740294933 + ], + "opposite_logprob_tokens": [ + -4.511080265045166 + ], + "logprob_reward": -0.011080282740294933, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080282740294933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1598934382200241, + 0.07661741226911545, + 0.11036434024572372, + 0.3175508677959442, + 0.036245569586753845, + 0.03800185024738312, + 0.05258364975452423, + 0.0507342591881752, + 0.0866113230586052, + 0.07139723002910614 + ], + "credit_map": [ + 0.019783640280365944, + 0.01244026143103838, + 0.02920239046216011, + 0.03483952209353447, + 0.023184331133961678, + 0.015961330384016037, + 0.012605370953679085, + 0.011876591481268406, + 0.011797420680522919, + 0.012872889637947083, + 0.008831949904561043, + 0.013565609231591225, + 0.00970463640987873, + 0.0056996154598891735, + 0.007438779808580875, + 0.006706514395773411, + 0.010419236496090889, + 0.014164506457746029, + 0.011412992142140865, + 0.01393281202763319, + 0.03129391744732857, + 0.009304861538112164, + 0.009335233829915524, + 0.010500794276595116, + 0.025276662781834602, + 0.12474031001329422, + 0.07638514041900635, + 0.01218213327229023, + 0.017474837601184845, + 0.01280131097882986, + 0.018413420766592026, + 0.030277099460363388, + 0.009724169038236141, + 0.003814588999375701, + 0.007986444979906082, + 0.0031329807825386524, + 0.0025382861495018005, + 0.001955056097358465, + 0.0028417243156582117, + 0.004252324812114239, + 0.0072642439045012, + 0.008860153146088123, + 0.003151694545522332, + 0.0036432950291782618, + 0.0034052510745823383, + 0.001804219326004386, + 0.0022700733970850706, + 0.007602918893098831, + 0.009767255745828152, + 0.008622256107628345, + 0.009989697486162186, + 0.00906652957201004, + 0.003466811263933778, + 0.0036454631481319666, + 0.004027939401566982, + 0.003997703082859516, + 0.0055650644935667515, + 0.005805624183267355, + 0.008148614317178726, + 0.01033499464392662, + 0.004326102789491415, + 0.004678006283938885, + 0.00794107187539339, + 0.00393478712067008, + 0.007152988575398922, + 0.006411666050553322, + 0.012782772071659565, + 0.007749331649392843, + 0.007045182399451733, + 0.014438940212130547, + 0.013327674940228462, + 0.017702778801321983, + 0.00697851600125432, + 0.005934678018093109, + 0.012132128700613976, + 0.008204174228012562, + 0.004209664184600115, + 0.009008890949189663, + 0.0069107236340641975, + 0.018018463626503944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.011080282740294933 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This first-person point-of-view (RST) shot is captured from a camera mounted low on the bow of a vibrant red kayak, plunging headfirst into the churning rapids of a powerful river. Rendered in the style of high-speed nature photography, the entire scene captures the physics of motion and light with hyperreal clarity. The camera pitches and dips dramatically with each powerful surge of the water, angling downward into deep troughs and then slamming upward on foaming crests, conveying the immense, chaotic force of the current. Frozen in harsh, direct sunlight, each drop of water kicked up by the boat's bow appears as a crystalline, shimmering sphere, refracting light through the turbulent turquoise and white water like tiny prisms. The kayak is navigating a deep, sun-scorched canyon, flanked by towering, layered, reddish-orange sandstone walls rising vertically on either side. As the kayak surges forward, these ancient, weathered canyon walls blur past in a dazzling yet sharp spectacle of striped colors, creating a profound sense of speed and raw natural power that is both exhilarating and overwhelming.", + "seed": 17980, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_14/index_6789_seed_17980_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.14772354727610945, + "vlm_logprob": -0.14772354727610945, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.14772354727610945, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person point-of-view mounted on a vibrant red kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kayak navigating through churning rapids in a deep, sun-scorched canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are towering, layered, reddish-orange sandstone walls visible rising vertically on either side of the river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do water droplets kicked up by the bow appear as crystalline, shimmering spheres refracting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pitch and dip dramatically with the water's surges, angling downward into troughs and upward on crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04974912479519844 + ], + "opposite_logprob_tokens": [ + -3.0497491359710693 + ], + "logprob_reward": -0.04974912479519844, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04974912479519844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19206537306308746, + 0.06356204301118851, + 0.06866277754306793, + 0.07437196373939514, + 0.07757575064897537, + 0.08207481354475021, + 0.04658859595656395, + 0.15035685896873474, + 0.09185005724430084, + 0.15289172530174255 + ], + "credit_map": [ + 0.022321168333292007, + 0.01885829120874405, + 0.023043127730488777, + 0.024326998740434647, + 0.02423637919127941, + 0.016237758100032806, + 0.031619682908058167, + 0.03142198175191879, + 0.010165338404476643, + 0.009621509350836277, + 0.008152462542057037, + 0.004402550403028727, + 0.007825789973139763, + 0.008399009704589844, + 0.007729519624263048, + 0.007265870925039053, + 0.01482333242893219, + 0.010541807860136032, + 0.006104110274463892, + 0.004498359747231007, + 0.0063792201690375805, + 0.00853810366243124, + 0.009602987207472324, + 0.008174862712621689, + 0.010732032358646393, + 0.009559066966176033, + 0.009841667488217354, + 0.0045311846770346165, + 0.00948326475918293, + 0.012766985222697258, + 0.008884962648153305, + 0.008572805672883987, + 0.011073878966271877, + 0.007844552397727966, + 0.011705459095537663, + 0.009494083002209663, + 0.009168107062578201, + 0.008437978103756905, + 0.013201579451560974, + 0.006650118622928858, + 0.013805139809846878, + 0.008503437042236328, + 0.016214869916439056, + 0.008408552967011929, + 0.00629540765658021, + 0.009375116787850857, + 0.010591805912554264, + 0.00888048205524683, + 0.01460681576281786, + 0.004799639340490103, + 0.005910276435315609, + 0.0031960774213075638, + 0.0026958861853927374, + 0.004751348402351141, + 0.0045685069635510445, + 0.006060048472136259, + 0.03229709714651108, + 0.011654726229608059, + 0.0211697518825531, + 0.012444275431334972, + 0.013061142526566982, + 0.016386089846491814, + 0.014264511875808239, + 0.02907927893102169, + 0.014406351372599602, + 0.007730951998382807, + 0.0060462080873548985, + 0.006243013311177492, + 0.00994779821485281, + 0.03059627115726471, + 0.004513324238359928, + 0.012366140261292458, + 0.020351015031337738, + 0.010326717980206013, + 0.006463850848376751, + 0.008669020608067513, + 0.014051485806703568, + 0.07290270179510117, + 0.005404671188443899, + 0.014722272753715515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.04974912479519844 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person point-of-view mounted on a vibrant red kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kayak navigating through churning rapids in a deep, sun-scorched canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are towering, layered, reddish-orange sandstone walls visible rising vertically on either side of the river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do water droplets kicked up by the bow appear as crystalline, shimmering spheres refracting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pitch and dip dramatically with the water's surges, angling downward into troughs and upward on crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761170983314514 + ], + "opposite_logprob_tokens": [ + -0.8261170983314514 + ], + "logprob_reward": -0.5761170983314514, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761170983314514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24117586016654968, + 0.0622258298099041, + 0.16049619019031525, + 0.11479666084051132, + 0.08552150428295135, + 0.08477721363306046, + 0.06474244594573975, + 0.03436722233891487, + 0.11882425844669342, + 0.03307276591658592 + ], + "credit_map": [ + 0.036755070090293884, + 0.027137767523527145, + 0.03438746929168701, + 0.03003913350403309, + 0.04159488156437874, + 0.018893025815486908, + 0.026079140603542328, + 0.02628939412534237, + 0.00935355480760336, + 0.012728256173431873, + 0.010099281556904316, + 0.005539272911846638, + 0.005693760700523853, + 0.0063877468928694725, + 0.004907888360321522, + 0.007516075391322374, + 0.0290101058781147, + 0.02694590575993061, + 0.017601855099201202, + 0.008740661665797234, + 0.012037422508001328, + 0.021907739341259003, + 0.017467735335230827, + 0.026784779503941536, + 0.016258759424090385, + 0.01612871140241623, + 0.016149301081895828, + 0.006585847120732069, + 0.012598350644111633, + 0.017597690224647522, + 0.01629512943327427, + 0.013182878494262695, + 0.015949247404932976, + 0.010417022742331028, + 0.019186846911907196, + 0.007886203937232494, + 0.007624960504472256, + 0.0074725341983139515, + 0.010570816695690155, + 0.006413881666958332, + 0.017989415675401688, + 0.009938783012330532, + 0.016345568001270294, + 0.00758563494309783, + 0.0068826815113425255, + 0.006088220980018377, + 0.012350877746939659, + 0.007596041541546583, + 0.011022889986634254, + 0.00794941931962967, + 0.011369035579264164, + 0.006264005322009325, + 0.00530184805393219, + 0.010182738304138184, + 0.007216557860374451, + 0.0054359580390155315, + 0.007977042347192764, + 0.004040991887450218, + 0.0035783855710178614, + 0.002237569075077772, + 0.003911239560693502, + 0.0047649783082306385, + 0.003453594632446766, + 0.004403423983603716, + 0.01577461138367653, + 0.009766589850187302, + 0.007081018295139074, + 0.008206278085708618, + 0.013187951408326626, + 0.03810396417975426, + 0.00936316978186369, + 0.017340684309601784, + 0.0034431254025548697, + 0.004806095268577337, + 0.0024986108765006065, + 0.0024796719662845135, + 0.006682354025542736, + 0.006769956089556217, + 0.0029099909588694572, + 0.0034829648211598396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.5761170983314514 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person point-of-view mounted on a vibrant red kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kayak navigating through churning rapids in a deep, sun-scorched canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are towering, layered, reddish-orange sandstone walls visible rising vertically on either side of the river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do water droplets kicked up by the bow appear as crystalline, shimmering spheres refracting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pitch and dip dramatically with the water's surges, angling downward into troughs and upward on crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205500662326813 + ], + "opposite_logprob_tokens": [ + -2.8120551109313965 + ], + "logprob_reward": -0.06205500662326813, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205500662326813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14790599048137665, + 0.06403211504220963, + 0.1821340024471283, + 0.09708478301763535, + 0.08065249025821686, + 0.11186513304710388, + 0.062061648815870285, + 0.10463715344667435, + 0.08545324206352234, + 0.0641733780503273 + ], + "credit_map": [ + 0.03140213340520859, + 0.01639564335346222, + 0.019157273694872856, + 0.020076854154467583, + 0.018497781828045845, + 0.010812301188707352, + 0.014979373663663864, + 0.01658463478088379, + 0.010380322113633156, + 0.01174204796552658, + 0.010515560396015644, + 0.00559920072555542, + 0.007055836729705334, + 0.007217037957161665, + 0.005389234982430935, + 0.006132881157100201, + 0.025716667994856834, + 0.035959526896476746, + 0.026056040078401566, + 0.010131869465112686, + 0.014201920479536057, + 0.020134201273322105, + 0.021927280351519585, + 0.028006505221128464, + 0.015389802865684032, + 0.015481228940188885, + 0.01345014851540327, + 0.005760838743299246, + 0.00950622744858265, + 0.013851014897227287, + 0.012933436781167984, + 0.010712089948356152, + 0.0160620529204607, + 0.015679745003581047, + 0.016228223219513893, + 0.005995549727231264, + 0.005458339117467403, + 0.007005395367741585, + 0.008225658908486366, + 0.0059975264593958855, + 0.022119950503110886, + 0.014886051416397095, + 0.021797670051455498, + 0.010105590336024761, + 0.00928690005093813, + 0.009965850971639156, + 0.013821038417518139, + 0.009882094338536263, + 0.009291679598391056, + 0.009375869296491146, + 0.012516663409769535, + 0.005920506548136473, + 0.004991835914552212, + 0.009372149594128132, + 0.006021933630108833, + 0.004571016877889633, + 0.03699614480137825, + 0.010533640161156654, + 0.00785033218562603, + 0.0041324119083583355, + 0.004898217041045427, + 0.010892932303249836, + 0.004724422935396433, + 0.02460905723273754, + 0.012030224315822124, + 0.00671553798019886, + 0.004974438808858395, + 0.0060560330748558044, + 0.008625226095318794, + 0.025853058323264122, + 0.007355090230703354, + 0.013843639753758907, + 0.00969642587006092, + 0.009232647716999054, + 0.004853819962590933, + 0.0035487788263708353, + 0.010045971721410751, + 0.015323346480727196, + 0.0042329216375947, + 0.0072394683957099915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.06205500662326813 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person point-of-view mounted on a vibrant red kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kayak navigating through churning rapids in a deep, sun-scorched canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are towering, layered, reddish-orange sandstone walls visible rising vertically on either side of the river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do water droplets kicked up by the bow appear as crystalline, shimmering spheres refracting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pitch and dip dramatically with the water's surges, angling downward into troughs and upward on crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306625574827194 + ], + "opposite_logprob_tokens": [ + -3.1680662631988525 + ], + "logprob_reward": -0.04306625574827194, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306625574827194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06384825706481934, + 0.022799158468842506, + 0.10236424207687378, + 0.025678355246782303, + 0.0288303941488266, + 0.05851612985134125, + 0.02655975893139839, + 0.10678251087665558, + 0.040267493575811386, + 0.5243536233901978 + ], + "credit_map": [ + 0.008830432780086994, + 0.008162702433764935, + 0.009691956453025341, + 0.007726629730314016, + 0.009783911518752575, + 0.005893065128475428, + 0.007500695995986462, + 0.006258866284042597, + 0.00286021176725626, + 0.003874168498441577, + 0.0029549559112638235, + 0.002326244255527854, + 0.002074945718050003, + 0.0028422677423805, + 0.002098769648000598, + 0.003767595626413822, + 0.021410074084997177, + 0.015670275315642357, + 0.008381822146475315, + 0.006195754744112492, + 0.005539920646697283, + 0.012761534191668034, + 0.013797486200928688, + 0.018607383593916893, + 0.0033715786412358284, + 0.0037442531902343035, + 0.00393833639100194, + 0.0017534452490508556, + 0.0019388357177376747, + 0.0045424108393490314, + 0.004300469998270273, + 0.002089026616886258, + 0.004535223823040724, + 0.003408720949664712, + 0.0033264686353504658, + 0.0023177836555987597, + 0.0025742007419466972, + 0.0029040048830211163, + 0.004707222804427147, + 0.005056771915405989, + 0.009601420722901821, + 0.0072554172948002815, + 0.008352935314178467, + 0.005684277508407831, + 0.005205324850976467, + 0.00801875814795494, + 0.008806612342596054, + 0.00559138972312212, + 0.00351309054531157, + 0.004162924829870462, + 0.005392166320234537, + 0.001937579712830484, + 0.0025093588046729565, + 0.0030431896448135376, + 0.003869655542075634, + 0.002131795044988394, + 0.036549441516399384, + 0.010377883911132812, + 0.007605021819472313, + 0.004141459707170725, + 0.003938403911888599, + 0.010467851534485817, + 0.00474883709102869, + 0.028953613713383675, + 0.006916825659573078, + 0.0046248529106378555, + 0.003241791622713208, + 0.0025889892131090164, + 0.0061904811300337315, + 0.009270435199141502, + 0.0030982347670942545, + 0.004335887264460325, + 0.07878412306308746, + 0.03567628562450409, + 0.02238178811967373, + 0.03270585834980011, + 0.05090097337961197, + 0.23867163062095642, + 0.02030782774090767, + 0.04492522031068802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 2 + ], + "selected_reward": -0.04306625574827194 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person point-of-view mounted on a vibrant red kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kayak navigating through churning rapids in a deep, sun-scorched canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are towering, layered, reddish-orange sandstone walls visible rising vertically on either side of the river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do water droplets kicked up by the bow appear as crystalline, shimmering spheres refracting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pitch and dip dramatically with the water's surges, angling downward into troughs and upward on crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630250882357359 + ], + "opposite_logprob_tokens": [ + -4.882630348205566 + ], + "logprob_reward": -0.007630250882357359, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630250882357359, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12054961174726486, + 0.09654086083173752, + 0.12877880036830902, + 0.09730435907840729, + 0.07552921026945114, + 0.06284729391336441, + 0.1148451417684555, + 0.1612342745065689, + 0.05947380140423775, + 0.08289669454097748 + ], + "credit_map": [ + 0.0177915096282959, + 0.015201413072645664, + 0.017970971763134003, + 0.014361271634697914, + 0.014259347692131996, + 0.008104639127850533, + 0.014944987371563911, + 0.017915448173880577, + 0.013355770148336887, + 0.018122149631381035, + 0.014569748193025589, + 0.010557475499808788, + 0.011861510574817657, + 0.009049060754477978, + 0.008978496305644512, + 0.010046642273664474, + 0.018139425665140152, + 0.02180768921971321, + 0.015296834520995617, + 0.009005219675600529, + 0.010056380182504654, + 0.013807864859700203, + 0.02314414642751217, + 0.017521223053336143, + 0.013400018215179443, + 0.013145839795470238, + 0.01132623478770256, + 0.008905869908630848, + 0.008825053460896015, + 0.010918298736214638, + 0.020448802039027214, + 0.010334236547350883, + 0.012264832854270935, + 0.011987458914518356, + 0.010743902064859867, + 0.006490728817880154, + 0.007028337102383375, + 0.006544296629726887, + 0.014228771440684795, + 0.006240875460207462, + 0.012704040855169296, + 0.009886614978313446, + 0.011896248906850815, + 0.0057479022070765495, + 0.005172444507479668, + 0.006396072451025248, + 0.007147124968469143, + 0.003896840615198016, + 0.019144941121339798, + 0.017350247129797935, + 0.02275409922003746, + 0.008391994051635265, + 0.009725593030452728, + 0.014205929823219776, + 0.01159460749477148, + 0.011677711270749569, + 0.047701556235551834, + 0.01652895286679268, + 0.01563999243080616, + 0.007691671140491962, + 0.01097415667027235, + 0.020242156460881233, + 0.010092907585203648, + 0.032362863421440125, + 0.008256389759480953, + 0.006471410393714905, + 0.004548071883618832, + 0.0031758558470755816, + 0.006697093602269888, + 0.017778629437088966, + 0.00288660847581923, + 0.009659736417233944, + 0.01219616923481226, + 0.009028289467096329, + 0.004289434291422367, + 0.004420371260493994, + 0.009272534400224686, + 0.030869537964463234, + 0.003908705431967974, + 0.008911646902561188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.007630250882357359 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This first-person point-of-view (RST) shot is captured from a camera mounted low on the bow of a vibrant red kayak, plunging headfirst into the churning rapids of a powerful river. Rendered in the style of high-speed nature photography, the entire scene captures the physics of motion and light with hyperreal clarity. The camera pitches and dips dramatically with each powerful surge of the water, angling downward into deep troughs and then slamming upward on foaming crests, conveying the immense, chaotic force of the current. Frozen in harsh, direct sunlight, each drop of water kicked up by the boat's bow appears as a crystalline, shimmering sphere, refracting light through the turbulent turquoise and white water like tiny prisms. The kayak is navigating a deep, sun-scorched canyon, flanked by towering, layered, reddish-orange sandstone walls rising vertically on either side. As the kayak surges forward, these ancient, weathered canyon walls blur past in a dazzling yet sharp spectacle of striped colors, creating a profound sense of speed and raw natural power that is both exhilarating and overwhelming.", + "seed": 17981, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_14/index_6789_seed_17981_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10690947202965617, + "vlm_logprob": -0.10690947202965617, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.10690947202965617, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person point-of-view mounted on a vibrant red kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kayak navigating through churning rapids in a deep, sun-scorched canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are towering, layered, reddish-orange sandstone walls visible rising vertically on either side of the river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do water droplets kicked up by the bow appear as crystalline, shimmering spheres refracting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pitch and dip dramatically with the water's surges, angling downward into troughs and upward on crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04095733165740967 + ], + "opposite_logprob_tokens": [ + -3.290957450866699 + ], + "logprob_reward": -0.04095733165740967, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04095733165740967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08889779448509216, + 0.421441912651062, + 0.05753914266824722, + 0.029168827459216118, + 0.04500231519341469, + 0.21077372133731842, + 0.024302247911691666, + 0.021669376641511917, + 0.07094481587409973, + 0.030259856954216957 + ], + "credit_map": [ + 0.010276724584400654, + 0.009321290999650955, + 0.011638796888291836, + 0.006850019097328186, + 0.011378513649106026, + 0.011263127438724041, + 0.00855732150375843, + 0.01961200125515461, + 0.03979825973510742, + 0.013498720712959766, + 0.1243300810456276, + 0.06269125640392303, + 0.028869112953543663, + 0.11992005258798599, + 0.017749473452568054, + 0.014584951102733612, + 0.008128635585308075, + 0.007671290542930365, + 0.008146063424646854, + 0.008733214810490608, + 0.005659022368490696, + 0.006769964471459389, + 0.007891501300036907, + 0.004539450164884329, + 0.003718681400641799, + 0.0031565430108457804, + 0.0048270574770867825, + 0.002524999901652336, + 0.0031580391805619, + 0.006717857904732227, + 0.0021836627274751663, + 0.0028819856233894825, + 0.006387803703546524, + 0.0074316393584012985, + 0.00610280130058527, + 0.0060726916417479515, + 0.004902202170342207, + 0.0061434535309672356, + 0.0046906801871955395, + 0.0032710435334593058, + 0.034651048481464386, + 0.02347429469227791, + 0.03935171291232109, + 0.038378577679395676, + 0.017033057287335396, + 0.016409480944275856, + 0.02777235582470894, + 0.013703183270990849, + 0.005229654256254435, + 0.004914755001664162, + 0.003531752387061715, + 0.002294115722179413, + 0.002566505456343293, + 0.0018522960599511862, + 0.0015514687402173877, + 0.0023616980761289597, + 0.001991985598579049, + 0.002472724998369813, + 0.003236706368625164, + 0.003269459819421172, + 0.0031020264141261578, + 0.0023410816211253405, + 0.002620964078232646, + 0.002634428208693862, + 0.0068956054747104645, + 0.0041940766386687756, + 0.010694462805986404, + 0.013530205935239792, + 0.00563022680580616, + 0.008845459669828415, + 0.011663192883133888, + 0.009491585195064545, + 0.0038500286173075438, + 0.0026141200214624405, + 0.002100155455991626, + 0.009148884564638138, + 0.0028260918334126472, + 0.0029749423265457153, + 0.005035253241658211, + 0.0017103840364143252 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.04095733165740967 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person point-of-view mounted on a vibrant red kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kayak navigating through churning rapids in a deep, sun-scorched canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are towering, layered, reddish-orange sandstone walls visible rising vertically on either side of the river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do water droplets kicked up by the bow appear as crystalline, shimmering spheres refracting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pitch and dip dramatically with the water's surges, angling downward into troughs and upward on crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870355188846588 + ], + "opposite_logprob_tokens": [ + -1.1370354890823364 + ], + "logprob_reward": -0.3870355188846588, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870355188846588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06142919510602951, + 0.2506181001663208, + 0.3721608817577362, + 0.014124838635325432, + 0.0831439346075058, + 0.06934411078691483, + 0.016462532803416252, + 0.007201259024441242, + 0.09854993224143982, + 0.026965202763676643 + ], + "credit_map": [ + 0.00836807768791914, + 0.007970369420945644, + 0.009908298961818218, + 0.006522837560623884, + 0.00629182206466794, + 0.00634011859074235, + 0.005036041606217623, + 0.010991633869707584, + 0.03340977057814598, + 0.024319998919963837, + 0.03281912952661514, + 0.02235945127904415, + 0.06497160345315933, + 0.045263923704624176, + 0.013601214624941349, + 0.013873028568923473, + 0.05892891809344292, + 0.057041723281145096, + 0.05125619098544121, + 0.053268179297447205, + 0.03941543027758598, + 0.04087524116039276, + 0.04476992040872574, + 0.026605274528265, + 0.0021962320897728205, + 0.0014197031268849969, + 0.0018578025046736002, + 0.0013048889813944697, + 0.0020644464530050755, + 0.0021051750518381596, + 0.0011980157578364015, + 0.001978575950488448, + 0.010603245347738266, + 0.014788046479225159, + 0.011002186685800552, + 0.009098927490413189, + 0.010598355904221535, + 0.012779251672327518, + 0.00759318470954895, + 0.006680735386908054, + 0.009294875897467136, + 0.009005358442664146, + 0.010532170534133911, + 0.007468340918421745, + 0.0069627705961465836, + 0.009808176197111607, + 0.008181462995707989, + 0.00809096172451973, + 0.003440195694565773, + 0.002523391041904688, + 0.0023112159688025713, + 0.0019255953375250101, + 0.00255132676102221, + 0.0017423242097720504, + 0.0010101775405928493, + 0.0009583067731000483, + 0.0007735653198324144, + 0.0009091200772672892, + 0.00100559217389673, + 0.0009981772163882852, + 0.0007346693309955299, + 0.0011926449369639158, + 0.0007283702725544572, + 0.000859120162203908, + 0.010302001610398293, + 0.005820356775075197, + 0.010485963895916939, + 0.023655707016587257, + 0.0073847039602696896, + 0.01430736854672432, + 0.020047692582011223, + 0.006546138785779476, + 0.003114105435088277, + 0.003031341126188636, + 0.0020170470234006643, + 0.004365785513073206, + 0.0027841064147651196, + 0.007565188687294722, + 0.002524212235584855, + 0.0015634186565876007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 8 + ], + "selected_reward": -0.3870355188846588 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person point-of-view mounted on a vibrant red kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kayak navigating through churning rapids in a deep, sun-scorched canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are towering, layered, reddish-orange sandstone walls visible rising vertically on either side of the river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do water droplets kicked up by the bow appear as crystalline, shimmering spheres refracting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pitch and dip dramatically with the water's surges, angling downward into troughs and upward on crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979913167655468 + ], + "opposite_logprob_tokens": [ + -3.529799222946167 + ], + "logprob_reward": -0.02979913167655468, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979913167655468, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039927393198013306, + 0.43033862113952637, + 0.2779243290424347, + 0.010610970668494701, + 0.04425261169672012, + 0.10221635550260544, + 0.023705165833234787, + 0.018432756885886192, + 0.039366528391838074, + 0.013225321657955647 + ], + "credit_map": [ + 0.005502352025359869, + 0.0054647973738610744, + 0.006251248996704817, + 0.004315770231187344, + 0.004175782669335604, + 0.004514401312917471, + 0.003331660060212016, + 0.00637137982994318, + 0.03702094405889511, + 0.020308902487158775, + 0.12233089655637741, + 0.060582369565963745, + 0.040296025574207306, + 0.1177254244685173, + 0.018057016655802727, + 0.014017075300216675, + 0.04340052604675293, + 0.041981812566518784, + 0.0379544273018837, + 0.04079123213887215, + 0.029573243111371994, + 0.030347105115652084, + 0.03372736647725105, + 0.02014860510826111, + 0.0016576305497437716, + 0.0015399648109450936, + 0.0012135470751672983, + 0.001070687547326088, + 0.0015442345757037401, + 0.0011031660251319408, + 0.0011327645042911172, + 0.0013489751145243645, + 0.005741596687585115, + 0.007160647306591272, + 0.006180260796099901, + 0.0049262079410254955, + 0.00398988276720047, + 0.008351455442607403, + 0.004471611697226763, + 0.003430945798754692, + 0.014825412072241306, + 0.020179348066449165, + 0.013162235729396343, + 0.009425139054656029, + 0.011278996244072914, + 0.014665251597762108, + 0.008297503925859928, + 0.010382473468780518, + 0.006462763529270887, + 0.0036390135064721107, + 0.0035567879676818848, + 0.0025717972312122583, + 0.002443925244733691, + 0.0019828584045171738, + 0.0014254520647227764, + 0.0016225689323619008, + 0.0017591525102034211, + 0.002959800185635686, + 0.0022954149171710014, + 0.0025377864949405193, + 0.0023006752599030733, + 0.002676967531442642, + 0.0019859103485941887, + 0.001917049172334373, + 0.003572873305529356, + 0.0026290123350918293, + 0.004733718000352383, + 0.009944675490260124, + 0.002638385398313403, + 0.005065913312137127, + 0.007494734134525061, + 0.0032872187439352274, + 0.0011356764007359743, + 0.0015109570231288671, + 0.0012913497630506754, + 0.002738112350925803, + 0.0016085777897387743, + 0.002488631522282958, + 0.0015406972961500287, + 0.0009113195701502264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 5 + ], + "selected_reward": -0.02979913167655468 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person point-of-view mounted on a vibrant red kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kayak navigating through churning rapids in a deep, sun-scorched canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are towering, layered, reddish-orange sandstone walls visible rising vertically on either side of the river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do water droplets kicked up by the bow appear as crystalline, shimmering spheres refracting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pitch and dip dramatically with the water's surges, angling downward into troughs and upward on crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001859694719315 + ], + "opposite_logprob_tokens": [ + -2.6950185298919678 + ], + "logprob_reward": -0.07001859694719315, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001859694719315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06425106525421143, + 0.2636776566505432, + 0.15220926702022552, + 0.06948459148406982, + 0.12518197298049927, + 0.05240637809038162, + 0.08388427644968033, + 0.08179318159818649, + 0.045031074434518814, + 0.062080517411231995 + ], + "credit_map": [ + 0.007746273186057806, + 0.009152173064649105, + 0.011784265749156475, + 0.008985558524727821, + 0.007548497524112463, + 0.007094076834619045, + 0.005613822024315596, + 0.006326398346573114, + 0.022059716284275055, + 0.01513686403632164, + 0.07164492458105087, + 0.03440907970070839, + 0.030697818845510483, + 0.0663163959980011, + 0.012683503329753876, + 0.010729369707405567, + 0.023018617182970047, + 0.01900206133723259, + 0.019683534279465675, + 0.0228551235049963, + 0.017901087179780006, + 0.020162319764494896, + 0.017102345824241638, + 0.012484177015721798, + 0.00949508510529995, + 0.010956097394227982, + 0.006692092400044203, + 0.005842171609401703, + 0.01282125897705555, + 0.0076616136357188225, + 0.00575691694393754, + 0.010259363800287247, + 0.019268697127699852, + 0.023366045206785202, + 0.015192090533673763, + 0.010991418734192848, + 0.016035115346312523, + 0.020406072959303856, + 0.00978267751634121, + 0.010139856487512589, + 0.007565208710730076, + 0.010859145782887936, + 0.005290159489959478, + 0.00681200809776783, + 0.006195388734340668, + 0.0072909449227154255, + 0.004036443307995796, + 0.004357080906629562, + 0.020880447700619698, + 0.012537897564470768, + 0.01132755633443594, + 0.015527143143117428, + 0.009475833736360073, + 0.0049339705146849155, + 0.0041747791692614555, + 0.005026640370488167, + 0.007491801865398884, + 0.012374280020594597, + 0.009693605825304985, + 0.012292871251702309, + 0.00799211859703064, + 0.012574977241456509, + 0.00911789946258068, + 0.010255629196763039, + 0.004038630984723568, + 0.00352686271071434, + 0.005898713134229183, + 0.011335385031998158, + 0.003144874470308423, + 0.0045402804389595985, + 0.00817816611379385, + 0.004368165507912636, + 0.00531263230368495, + 0.005275004077702761, + 0.004990189336240292, + 0.01726653054356575, + 0.005724173504859209, + 0.011052843183279037, + 0.00910111889243126, + 0.003358027897775173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.07001859694719315 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person point-of-view mounted on a vibrant red kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kayak navigating through churning rapids in a deep, sun-scorched canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are towering, layered, reddish-orange sandstone walls visible rising vertically on either side of the river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do water droplets kicked up by the bow appear as crystalline, shimmering spheres refracting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pitch and dip dramatically with the water's surges, angling downward into troughs and upward on crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006736780982464552 + ], + "opposite_logprob_tokens": [ + -5.006736755371094 + ], + "logprob_reward": -0.006736780982464552, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006736780982464552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10238805413246155, + 0.07645892351865768, + 0.28524795174598694, + 0.058248888701200485, + 0.09455293416976929, + 0.12374933809041977, + 0.03912672773003578, + 0.08003777265548706, + 0.04696556553244591, + 0.09322382509708405 + ], + "credit_map": [ + 0.011533609591424465, + 0.0225810669362545, + 0.013745542615652084, + 0.011270673014223576, + 0.011955727823078632, + 0.012229003943502903, + 0.007517414167523384, + 0.011555011384189129, + 0.008724577724933624, + 0.00820143148303032, + 0.013154350221157074, + 0.011224343441426754, + 0.01033945195376873, + 0.007321344222873449, + 0.006134629715234041, + 0.011358791962265968, + 0.04087812453508377, + 0.04128650203347206, + 0.03630107641220093, + 0.042670994997024536, + 0.03130088374018669, + 0.037922173738479614, + 0.031037557870149612, + 0.02385067753493786, + 0.007727575488388538, + 0.006674709264189005, + 0.008924673311412334, + 0.004602019675076008, + 0.007275737356394529, + 0.01443227007985115, + 0.003705762792378664, + 0.004906142130494118, + 0.012002449482679367, + 0.0141641516238451, + 0.012547331862151623, + 0.012456094846129417, + 0.012722472660243511, + 0.014673206023871899, + 0.009302424266934395, + 0.006684805732220411, + 0.0182991623878479, + 0.017599618062376976, + 0.015645364299416542, + 0.016334783285856247, + 0.017429141327738762, + 0.02090444602072239, + 0.009495938196778297, + 0.008040881715714931, + 0.0067145442590117455, + 0.006272498052567244, + 0.003845195984467864, + 0.005259636789560318, + 0.006858018692582846, + 0.004427340347319841, + 0.002237176988273859, + 0.0035123166162520647, + 0.00814618356525898, + 0.010371341370046139, + 0.011716135777533054, + 0.009567850269377232, + 0.006173585541546345, + 0.012572357431054115, + 0.010985083878040314, + 0.010505234822630882, + 0.004142439924180508, + 0.0030904205050319433, + 0.0055526080541312695, + 0.010804089717566967, + 0.0035260417498648167, + 0.0036111483350396156, + 0.00923370011150837, + 0.007005118299275637, + 0.006577747408300638, + 0.006700791884213686, + 0.005079534836113453, + 0.026019731536507607, + 0.005618872586637735, + 0.027346758171916008, + 0.01177548710256815, + 0.004104904364794493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.006736780982464552 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This first-person point-of-view (RST) shot is captured from a camera mounted low on the bow of a vibrant red kayak, plunging headfirst into the churning rapids of a powerful river. Rendered in the style of high-speed nature photography, the entire scene captures the physics of motion and light with hyperreal clarity. The camera pitches and dips dramatically with each powerful surge of the water, angling downward into deep troughs and then slamming upward on foaming crests, conveying the immense, chaotic force of the current. Frozen in harsh, direct sunlight, each drop of water kicked up by the boat's bow appears as a crystalline, shimmering sphere, refracting light through the turbulent turquoise and white water like tiny prisms. The kayak is navigating a deep, sun-scorched canyon, flanked by towering, layered, reddish-orange sandstone walls rising vertically on either side. As the kayak surges forward, these ancient, weathered canyon walls blur past in a dazzling yet sharp spectacle of striped colors, creating a profound sense of speed and raw natural power that is both exhilarating and overwhelming.", + "seed": 17982, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_14/index_6789_seed_17982_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.031664344295859334, + "vlm_logprob": -0.031664344295859334, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.031664344295859334, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person point-of-view mounted on a vibrant red kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kayak navigating through churning rapids in a deep, sun-scorched canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are towering, layered, reddish-orange sandstone walls visible rising vertically on either side of the river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do water droplets kicked up by the bow appear as crystalline, shimmering spheres refracting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pitch and dip dramatically with the water's surges, angling downward into troughs and upward on crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0170426145195961 + ], + "opposite_logprob_tokens": [ + -4.267042636871338 + ], + "logprob_reward": -0.0170426145195961, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.0170426145195961, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18985222280025482, + 0.046884629875421524, + 0.13452450931072235, + 0.08612053096294403, + 0.061585161834955215, + 0.22998115420341492, + 0.02055264264345169, + 0.1268937885761261, + 0.054776307195425034, + 0.04882907122373581 + ], + "credit_map": [ + 0.02006220445036888, + 0.014137123711407185, + 0.04603113606572151, + 0.01741751842200756, + 0.010008149780333042, + 0.010528342798352242, + 0.02276461385190487, + 0.04890312999486923, + 0.006307661999017, + 0.006531700491905212, + 0.010391905903816223, + 0.004900994244962931, + 0.0048407199792563915, + 0.0033523254096508026, + 0.0041907974518835545, + 0.00636852253228426, + 0.017003780230879784, + 0.022676357999444008, + 0.010953039862215519, + 0.01539846695959568, + 0.015506010502576828, + 0.022496167570352554, + 0.01226514670997858, + 0.018225522711873055, + 0.011418613605201244, + 0.016057122498750687, + 0.005902712699025869, + 0.006176259834319353, + 0.016081232577562332, + 0.01247195154428482, + 0.009638042189180851, + 0.008374583907425404, + 0.010805857367813587, + 0.005076990928500891, + 0.008377620950341225, + 0.00714655639603734, + 0.0033900002017617226, + 0.00966289360076189, + 0.00824113842099905, + 0.008884101174771786, + 0.010675815865397453, + 0.005029444117099047, + 0.01158277690410614, + 0.027707496657967567, + 0.008487474173307419, + 0.011390769854187965, + 0.09194885939359665, + 0.06315850466489792, + 0.0018270905129611492, + 0.0015794190112501383, + 0.0030557222198694944, + 0.0031168784480541945, + 0.0015611941926181316, + 0.0024198247119784355, + 0.004050831310451031, + 0.002941680606454611, + 0.017183708027005196, + 0.014765387400984764, + 0.015721479430794716, + 0.01317925751209259, + 0.017627347260713577, + 0.017530424520373344, + 0.016936257481575012, + 0.013949915766716003, + 0.00696796178817749, + 0.006245411466807127, + 0.004873208235949278, + 0.008328325115144253, + 0.00837702490389347, + 0.009115996770560741, + 0.0069311391562223434, + 0.003937236499041319, + 0.0041662598960101604, + 0.0036868792958557606, + 0.007876462303102016, + 0.0069351294077932835, + 0.005091229919344187, + 0.005885403137654066, + 0.01118530984967947, + 0.004002391826361418 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.0170426145195961 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person point-of-view mounted on a vibrant red kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kayak navigating through churning rapids in a deep, sun-scorched canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are towering, layered, reddish-orange sandstone walls visible rising vertically on either side of the river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do water droplets kicked up by the bow appear as crystalline, shimmering spheres refracting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pitch and dip dramatically with the water's surges, angling downward into troughs and upward on crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000881433486938 + ], + "opposite_logprob_tokens": [ + -2.6950087547302246 + ], + "logprob_reward": -0.07000881433486938, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000881433486938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14049264788627625, + 0.022030562162399292, + 0.07945652306079865, + 0.14764191210269928, + 0.11729029566049576, + 0.2182552069425583, + 0.018396982923150063, + 0.03234861046075821, + 0.1455167531967163, + 0.07857049256563187 + ], + "credit_map": [ + 0.02186991646885872, + 0.00665766466408968, + 0.04430360719561577, + 0.012735622003674507, + 0.004765297751873732, + 0.006372920703142881, + 0.014183573424816132, + 0.029604047536849976, + 0.0027078527491539717, + 0.004082577768713236, + 0.003461503190919757, + 0.003207694971933961, + 0.0017918250523507595, + 0.001419154112227261, + 0.0025553235318511724, + 0.002804629970341921, + 0.010942901484668255, + 0.01263980008661747, + 0.007038808427751064, + 0.00991547666490078, + 0.006945257540792227, + 0.013504453003406525, + 0.007200464606285095, + 0.011269363574683666, + 0.024813862517476082, + 0.03189816698431969, + 0.01189451850950718, + 0.013155559077858925, + 0.01495949737727642, + 0.019629210233688354, + 0.016507955268025398, + 0.014783140271902084, + 0.024180591106414795, + 0.010540933348238468, + 0.01662008836865425, + 0.010752018541097641, + 0.0055761998519301414, + 0.020377466455101967, + 0.014176013879477978, + 0.015066984109580517, + 0.010768785141408443, + 0.005839381366968155, + 0.010956624522805214, + 0.0261320061981678, + 0.00802228506654501, + 0.010578498244285583, + 0.08654884248971939, + 0.059408798813819885, + 0.0024734665639698505, + 0.0013360140146687627, + 0.001878273906186223, + 0.0019881795160472393, + 0.0010881524067372084, + 0.0025896376464515924, + 0.0033548984210938215, + 0.003688360331580043, + 0.0051216306164860725, + 0.003294402966275811, + 0.005620554089546204, + 0.00509607931599021, + 0.004138478543609381, + 0.0031681370455771685, + 0.0027935237158089876, + 0.0031158030033111572, + 0.017415223643183708, + 0.01275942288339138, + 0.014600665308535099, + 0.016463611274957657, + 0.031866904348134995, + 0.02503172494471073, + 0.01595882885158062, + 0.011420367285609245, + 0.009314932860434055, + 0.010576053522527218, + 0.015292054042220116, + 0.011603275313973427, + 0.006540240719914436, + 0.008469958789646626, + 0.009791987016797066, + 0.006981983315199614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 8 + ], + "selected_reward": -0.07000881433486938 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person point-of-view mounted on a vibrant red kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kayak navigating through churning rapids in a deep, sun-scorched canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are towering, layered, reddish-orange sandstone walls visible rising vertically on either side of the river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do water droplets kicked up by the bow appear as crystalline, shimmering spheres refracting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pitch and dip dramatically with the water's surges, angling downward into troughs and upward on crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058008871972561 + ], + "opposite_logprob_tokens": [ + -3.8955800533294678 + ], + "logprob_reward": -0.02058008871972561, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058008871972561, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13525748252868652, + 0.03474781662225723, + 0.036730002611875534, + 0.07612013816833496, + 0.10595802217721939, + 0.42025265097618103, + 0.018325109034776688, + 0.039294738322496414, + 0.08600188046693802, + 0.047312136739492416 + ], + "credit_map": [ + 0.023571213707327843, + 0.0060117500834167, + 0.04372122511267662, + 0.012646950781345367, + 0.006149377673864365, + 0.005703998263925314, + 0.01240980252623558, + 0.025043170899152756, + 0.004384167026728392, + 0.005257444456219673, + 0.00766347162425518, + 0.004811798222362995, + 0.0027863045688718557, + 0.001942622708156705, + 0.003963886294513941, + 0.003938120324164629, + 0.004991178400814533, + 0.0056154667399823666, + 0.0034079006873071194, + 0.004455350339412689, + 0.0034004475455731153, + 0.005477448459714651, + 0.0038603167049586773, + 0.005521893501281738, + 0.01340553816407919, + 0.015187835320830345, + 0.006272674072533846, + 0.007592411711812019, + 0.006413597613573074, + 0.01041074376553297, + 0.009509274736046791, + 0.007328059058636427, + 0.023095382377505302, + 0.00812378991395235, + 0.014171551913022995, + 0.010001905262470245, + 0.005520990584045649, + 0.017910536378622055, + 0.013386779464781284, + 0.013747084885835648, + 0.019957344979047775, + 0.009272635914385319, + 0.020249884575605392, + 0.050582028925418854, + 0.015015456825494766, + 0.02120407484471798, + 0.1686607301235199, + 0.11531047523021698, + 0.003052996238693595, + 0.0018531795358285308, + 0.002313022967427969, + 0.002044886350631714, + 0.0011651264503598213, + 0.0022473405115306377, + 0.0032105722930282354, + 0.0024379841051995754, + 0.005172801204025745, + 0.004648249596357346, + 0.00562314223498106, + 0.004526473116129637, + 0.005497889127582312, + 0.005042380653321743, + 0.004347043111920357, + 0.004436760675162077, + 0.010927467606961727, + 0.007534051313996315, + 0.008276493288576603, + 0.010051574558019638, + 0.018626630306243896, + 0.014776553958654404, + 0.009491493925452232, + 0.006317611318081617, + 0.006421904545277357, + 0.004390527494251728, + 0.008807126432657242, + 0.006796148139983416, + 0.004754413850605488, + 0.005782086402177811, + 0.0062501635402441025, + 0.004109764937311411 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.02058008871972561 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person point-of-view mounted on a vibrant red kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kayak navigating through churning rapids in a deep, sun-scorched canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are towering, layered, reddish-orange sandstone walls visible rising vertically on either side of the river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do water droplets kicked up by the bow appear as crystalline, shimmering spheres refracting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pitch and dip dramatically with the water's surges, angling downward into troughs and upward on crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430605448782444 + ], + "opposite_logprob_tokens": [ + -3.168060541152954 + ], + "logprob_reward": -0.0430605448782444, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430605448782444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2617660164833069, + 0.036358967423439026, + 0.20309776067733765, + 0.053519368171691895, + 0.040376290678977966, + 0.054776474833488464, + 0.008567200973629951, + 0.05785734951496124, + 0.13350430130958557, + 0.15017619729042053 + ], + "credit_map": [ + 0.046737536787986755, + 0.028234651312232018, + 0.022619465366005898, + 0.021573493257164955, + 0.025286277756094933, + 0.018856201320886612, + 0.031998731195926666, + 0.06645969301462173, + 0.0042878794483840466, + 0.004593739751726389, + 0.007847025990486145, + 0.0054504601284861565, + 0.002834423677995801, + 0.0024267167318612337, + 0.003453036304563284, + 0.005465686321258545, + 0.02400759980082512, + 0.025010446086525917, + 0.018890032544732094, + 0.031952325254678726, + 0.015600616112351418, + 0.03514157235622406, + 0.02494795247912407, + 0.027547229081392288, + 0.007131604012101889, + 0.01126087922602892, + 0.006046060472726822, + 0.007241758052259684, + 0.004367339424788952, + 0.007207886315882206, + 0.005208807066082954, + 0.005055037792772055, + 0.009747490286827087, + 0.0037484022323042154, + 0.004807617515325546, + 0.0036572986282408237, + 0.0020391233265399933, + 0.006303752772510052, + 0.0050845020450651646, + 0.004988108295947313, + 0.003335388842970133, + 0.003029331099241972, + 0.0030742455273866653, + 0.006379385478794575, + 0.00238686241209507, + 0.002780293347314, + 0.020070215687155724, + 0.013720756396651268, + 0.001208530506119132, + 0.0010897715110331774, + 0.0013964161043986678, + 0.0008061561384238303, + 0.0006565578514710069, + 0.0008846972486935556, + 0.001620119553990662, + 0.0009049529326148331, + 0.008116574957966805, + 0.007030760403722525, + 0.007949821650981903, + 0.007815234363079071, + 0.007045116275548935, + 0.006428605876863003, + 0.007298904936760664, + 0.006172333378344774, + 0.009047278203070164, + 0.008690757676959038, + 0.015296567231416702, + 0.01266922615468502, + 0.042191676795482635, + 0.0287149790674448, + 0.009471595287322998, + 0.007422225549817085, + 0.019174477085471153, + 0.017472220584750175, + 0.0300673209130764, + 0.02305551990866661, + 0.013113732449710369, + 0.015240863896906376, + 0.01788763888180256, + 0.014164441265165806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.0430605448782444 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person point-of-view mounted on a vibrant red kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kayak navigating through churning rapids in a deep, sun-scorched canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are towering, layered, reddish-orange sandstone walls visible rising vertically on either side of the river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do water droplets kicked up by the bow appear as crystalline, shimmering spheres refracting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pitch and dip dramatically with the water's surges, angling downward into troughs and upward on crests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007629659026861191 + ], + "opposite_logprob_tokens": [ + -4.882629871368408 + ], + "logprob_reward": -0.007629659026861191, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007629659026861191, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2046617865562439, + 0.05069233477115631, + 0.11232351511716843, + 0.09903083741664886, + 0.10602255165576935, + 0.23149794340133667, + 0.03401390463113785, + 0.05846456065773964, + 0.02518855594098568, + 0.07810395956039429 + ], + "credit_map": [ + 0.042768631130456924, + 0.01298580877482891, + 0.06232504919171333, + 0.017865944653749466, + 0.010306360200047493, + 0.011108716018497944, + 0.01623300462961197, + 0.031068265438079834, + 0.006839259061962366, + 0.006513859145343304, + 0.01202213391661644, + 0.006301491055637598, + 0.003522235667333007, + 0.002479435643181205, + 0.005835536401718855, + 0.007178379688411951, + 0.014201737940311432, + 0.01347988098859787, + 0.00911215040832758, + 0.019127514213323593, + 0.010323861613869667, + 0.0202640350908041, + 0.011571303941309452, + 0.014243027195334435, + 0.00947667844593525, + 0.010274708271026611, + 0.010205528698861599, + 0.008838048204779625, + 0.019479677081108093, + 0.015766337513923645, + 0.010790438391268253, + 0.014199420809745789, + 0.015379410237073898, + 0.008511058986186981, + 0.013293076306581497, + 0.014270040206611156, + 0.006975796073675156, + 0.013836591504514217, + 0.01703217811882496, + 0.016724396497011185, + 0.012917594984173775, + 0.011354435235261917, + 0.014842559583485126, + 0.027490725740790367, + 0.010017910972237587, + 0.011701398529112339, + 0.08440050482749939, + 0.058772820979356766, + 0.00296575715765357, + 0.0023091663606464863, + 0.004433848895132542, + 0.004164041019976139, + 0.002148042432963848, + 0.004712134134024382, + 0.009176668711006641, + 0.004104243591427803, + 0.009709292091429234, + 0.006114301737397909, + 0.00581973884254694, + 0.009231144562363625, + 0.006642965134233236, + 0.005257606506347656, + 0.006294602528214455, + 0.009394912049174309, + 0.004717960953712463, + 0.001833189046010375, + 0.002089647576212883, + 0.004044619854539633, + 0.00304383784532547, + 0.0027478167321532965, + 0.0036270893178880215, + 0.003084395080804825, + 0.008975613862276077, + 0.013922919519245625, + 0.012814328074455261, + 0.011429081670939922, + 0.00856001302599907, + 0.008579403162002563, + 0.008115384727716446, + 0.005707218777388334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.007629659026861191 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_15/index_tensor([6468, 6468])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_15/index_tensor([6468, 6468])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ce67a4eb9cf467cc46b532f619f25a30c3316fcf --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_15/index_tensor([6468, 6468])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, bird's-eye view, taken from directly above, captures a tranquil lake, scattered across it by countless tiny, heart-shaped water chestnut leaves like hand-painted green sprites. Each leaf embodies a saturated, soft emerald green, with rounded edges and delicate veins visible. The water is crystal clear, a deep blue-green hue. Warm afternoon sunlight filters through the ripples, creating countless dancing, shimmering golden specks that resemble highlights in hand-drawn animation. The leaves sway slowly and rhythmically with the gentle rise and fall of the water, sometimes gently bobbing and sometimes spinning leisurely in place. The entire scene is imbued with the delicate, warm quality of a Ghibli animation, with vibrant, yet not glaring, colors creating a magical moment of profound tranquility, healing, and a touch of nostalgia.", + "seed": 25656, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_15/index_6468_seed_25656_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.40230808183550837, + "vlm_logprob": -0.40230808183550837, + "vlm_yes_no_margin": 1.725, + "vlm_token_credit": -0.40230808183550837, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static bird's-eye view directly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the water chestnut leaves shaped like hearts and colored saturated emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water crystal clear with a deep blue-green hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves sway slowly and rhythmically with the water, sometimes bobbing or spinning in place?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit the visual style of a Ghibli animation with warm, vibrant, non-glaring colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1891106218099594 + ], + "opposite_logprob_tokens": [ + -1.8141106367111206 + ], + "logprob_reward": -0.1891106218099594, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1891106218099594, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06458274275064468, + 0.01949864625930786, + 0.006049532908946276, + 0.7881367802619934, + 0.014021898619830608, + 0.0068405489437282085, + 0.02296609804034233, + 0.01329877506941557, + 0.011651051230728626, + 0.052953917533159256 + ], + "credit_map": [ + 0.00638101901859045, + 0.018943289294838905, + 0.014748835004866123, + 0.004985361360013485, + 0.00591009808704257, + 0.004042880143970251, + 0.0056625488214194775, + 0.003908711485564709, + 0.0032781551126390696, + 0.00209665484726429, + 0.0029947725124657154, + 0.0021692190784960985, + 0.002150224754586816, + 0.002488114172592759, + 0.0023327006492763758, + 0.00198880466632545, + 0.0007568650180473924, + 0.000594346784055233, + 0.0007860676269046962, + 0.0014342920621857047, + 0.0005133389495313168, + 0.0005226688808761537, + 0.0007812703843228519, + 0.0006606829119846225, + 0.07367581129074097, + 0.19293056428432465, + 0.06296101212501526, + 0.03048345446586609, + 0.05125559866428375, + 0.022468766197562218, + 0.2479381561279297, + 0.10642340779304504, + 0.0015765540301799774, + 0.0013891769340261817, + 0.001496900338679552, + 0.0012722628889605403, + 0.001416343729943037, + 0.0022456739097833633, + 0.0022952482104301453, + 0.002329738112166524, + 0.0006069598020985723, + 0.00048157485434785485, + 0.0006899783620610833, + 0.0008954987861216068, + 0.0006805200246162713, + 0.0007112083840183914, + 0.0011990518542006612, + 0.0015757570508867502, + 0.003160009626299143, + 0.004592799115926027, + 0.0022515153978019953, + 0.001901094103232026, + 0.0018272693268954754, + 0.0019419417949393392, + 0.003085159230977297, + 0.0042063090950250626, + 0.0016115382313728333, + 0.002694851951673627, + 0.001185585861094296, + 0.002276437822729349, + 0.0011326003586873412, + 0.0009344185818918049, + 0.0018794086063280702, + 0.0015839340630918741, + 0.0012187216198071837, + 0.0015090632950887084, + 0.0014117872342467308, + 0.002764443401247263, + 0.0014819162897765636, + 0.0012523648329079151, + 0.00125559838488698, + 0.0007571560563519597, + 0.004622394684702158, + 0.012816575355827808, + 0.005583480000495911, + 0.004739528987556696, + 0.0043878741562366486, + 0.003125830553472042, + 0.01324375532567501, + 0.004434478003531694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.1891106218099594 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static bird's-eye view directly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the water chestnut leaves shaped like hearts and colored saturated emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water crystal clear with a deep blue-green hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves sway slowly and rhythmically with the water, sometimes bobbing or spinning in place?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit the visual style of a Ghibli animation with warm, vibrant, non-glaring colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495745688676834 + ], + "opposite_logprob_tokens": [ + -2.929957389831543 + ], + "logprob_reward": -0.05495745688676834, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495745688676834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6892213225364685, + 0.056082941591739655, + 0.04707051068544388, + 0.06142422929406166, + 0.01455614622682333, + 0.014836878515779972, + 0.013958733528852463, + 0.01604452170431614, + 0.018745584413409233, + 0.06805922091007233 + ], + "credit_map": [ + 0.04121572896838188, + 0.022738739848136902, + 0.19725297391414642, + 0.06091402843594551, + 0.02569739706814289, + 0.025062045082449913, + 0.26944294571876526, + 0.04689740762114525, + 0.007336215581744909, + 0.0050303214229643345, + 0.010128358379006386, + 0.009989113546907902, + 0.006592751946300268, + 0.005025274585932493, + 0.007476798724383116, + 0.004504107404500246, + 0.004477321170270443, + 0.0044593969359993935, + 0.0064385877922177315, + 0.007293988019227982, + 0.006206463556736708, + 0.006830491591244936, + 0.005435852333903313, + 0.00592840975150466, + 0.005375193431973457, + 0.012958013452589512, + 0.004890657961368561, + 0.002428442006930709, + 0.003919599112123251, + 0.003439952852204442, + 0.01902354136109352, + 0.009388833306729794, + 0.002040137769654393, + 0.003362467745319009, + 0.0015881512081250548, + 0.0016454999567940831, + 0.0017410271102562547, + 0.0014592362567782402, + 0.001379582448862493, + 0.001340043731033802, + 0.0019931881688535213, + 0.0024943132884800434, + 0.001488732174038887, + 0.0018372664926573634, + 0.0019401639001443982, + 0.0019309954950585961, + 0.001626128563657403, + 0.001526090782135725, + 0.0013068979606032372, + 0.0013240004191175103, + 0.00185356754809618, + 0.0015318201621994376, + 0.0012533999979496002, + 0.0018244960810989141, + 0.0017901005921885371, + 0.003074450884014368, + 0.002219931222498417, + 0.003097750246524811, + 0.0018507902277633548, + 0.0024653798900544643, + 0.0016721567371860147, + 0.0013509448617696762, + 0.0018605414079502225, + 0.0015270269941538572, + 0.0015991342952474952, + 0.002825462957844138, + 0.0017382586374878883, + 0.004196287598460913, + 0.0024029347114264965, + 0.0017543940339237452, + 0.0025175653863698244, + 0.0017115474911406636, + 0.005274660885334015, + 0.015042203478515148, + 0.007049170322716236, + 0.007371618412435055, + 0.0058305524289608, + 0.004070725757628679, + 0.017644500359892845, + 0.005775784142315388 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.05495745688676834 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static bird's-eye view directly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the water chestnut leaves shaped like hearts and colored saturated emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water crystal clear with a deep blue-green hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves sway slowly and rhythmically with the water, sometimes bobbing or spinning in place?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit the visual style of a Ghibli animation with warm, vibrant, non-glaring colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4062734842300415 + ], + "opposite_logprob_tokens": [ + -0.2812734842300415 + ], + "logprob_reward": -1.4062734842300415, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4062734842300415, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.17302845418453217, + 0.029107511043548584, + 0.009344319812953472, + 0.7088021039962769, + 0.017038611695170403, + 0.008785232901573181, + 0.012326759286224842, + 0.004485188517719507, + 0.003874644637107849, + 0.03320716693997383 + ], + "credit_map": [ + 0.011625953018665314, + 0.026006050407886505, + 0.031550221145153046, + 0.013341815210878849, + 0.008084497414529324, + 0.007723562885075808, + 0.062288038432598114, + 0.012408330105245113, + 0.0022333741653710604, + 0.001712471479550004, + 0.004473279695957899, + 0.008261012844741344, + 0.0022680782712996006, + 0.0021798191592097282, + 0.004360176157206297, + 0.003619300900027156, + 0.0008303080103360116, + 0.0008653453551232815, + 0.0008916562073864043, + 0.0024696639738976955, + 0.0013783120084553957, + 0.0008357023471035063, + 0.0011603091843426228, + 0.0009130238322541118, + 0.06641869992017746, + 0.17470578849315643, + 0.05691000819206238, + 0.027428198605775833, + 0.04635121300816536, + 0.020426403731107712, + 0.22217915952205658, + 0.09438265860080719, + 0.0031548147089779377, + 0.004304768983274698, + 0.001597340451553464, + 0.0014566252939403057, + 0.0018001708667725325, + 0.001348896767012775, + 0.001810160931199789, + 0.0015658342745155096, + 0.0013535951729863882, + 0.0013862596824765205, + 0.0010331135708838701, + 0.001072965911589563, + 0.0008860985981300473, + 0.0011648056097328663, + 0.0010153522016480565, + 0.0008730432600714266, + 0.001461624982766807, + 0.002597829094156623, + 0.0009923503967002034, + 0.0011082651326432824, + 0.0010691223433241248, + 0.0018139091553166509, + 0.0013494204031303525, + 0.0019342385930940509, + 0.0003888820647262037, + 0.0006091583054512739, + 0.0003979744215030223, + 0.0009146862430498004, + 0.000485812546685338, + 0.0005070124752819538, + 0.000630748167168349, + 0.0005509147886186838, + 0.00037720680120401084, + 0.00048324608360417187, + 0.0004240620764903724, + 0.0008384753600694239, + 0.0004949642461724579, + 0.0004357754369266331, + 0.00043906577047891915, + 0.0003818492987193167, + 0.0027113608084619045, + 0.008547227829694748, + 0.0036779202055186033, + 0.002226039534434676, + 0.0030296973418444395, + 0.0018275876063853502, + 0.00795313436537981, + 0.0032342036720365286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -1.4062734842300415 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static bird's-eye view directly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the water chestnut leaves shaped like hearts and colored saturated emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water crystal clear with a deep blue-green hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves sway slowly and rhythmically with the water, sometimes bobbing or spinning in place?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit the visual style of a Ghibli animation with warm, vibrant, non-glaring colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012550774961709976 + ], + "opposite_logprob_tokens": [ + -4.3875508308410645 + ], + "logprob_reward": -0.012550774961709976, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012550774961709976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052290648221969604, + 0.030060382559895515, + 0.01997387409210205, + 0.7583404779434204, + 0.017563777044415474, + 0.015848776325583458, + 0.019956417381763458, + 0.017017671838402748, + 0.0099747683852911, + 0.05897316709160805 + ], + "credit_map": [ + 0.006508300080895424, + 0.003051341976970434, + 0.011721688322722912, + 0.0038181543350219727, + 0.004789266735315323, + 0.004908925388008356, + 0.013475160114467144, + 0.0040178196504712105, + 0.003324182704091072, + 0.0027691528666764498, + 0.004891056567430496, + 0.005556622985750437, + 0.003425062168389559, + 0.003247383749112487, + 0.003684160066768527, + 0.003162764711305499, + 0.001595991663634777, + 0.0020863404497504234, + 0.0020568296313285828, + 0.006685929372906685, + 0.0018999241292476654, + 0.0014916121726855636, + 0.0023723740596324205, + 0.0017848752904683352, + 0.07120261341333389, + 0.18752416968345642, + 0.06071805953979492, + 0.029461467638611794, + 0.049700699746608734, + 0.02202484756708145, + 0.2370702028274536, + 0.1006385013461113, + 0.002638448029756546, + 0.0023756599985063076, + 0.0017166173784062266, + 0.0018777059158310294, + 0.0019359516445547342, + 0.0019007414812222123, + 0.0022638554219156504, + 0.0028548003174364567, + 0.0024228168185800314, + 0.0030190697871148586, + 0.0018330463208258152, + 0.001974251586943865, + 0.0014271712861955166, + 0.001828639185987413, + 0.0017880217637866735, + 0.0015557616716250777, + 0.002405563136562705, + 0.004970821551978588, + 0.0017432394670322537, + 0.001991033786907792, + 0.0023078657686710358, + 0.001941682188771665, + 0.002186903264373541, + 0.002409311244264245, + 0.001751381205394864, + 0.0018256669864058495, + 0.0020582256838679314, + 0.0027375039644539356, + 0.001602941774763167, + 0.001585371675901115, + 0.002726778155192733, + 0.002729803789407015, + 0.0009913693647831678, + 0.001577582792378962, + 0.0009358208044432104, + 0.002208324382081628, + 0.0012187015963718295, + 0.0010581080568954349, + 0.0010517188347876072, + 0.0009331440669484437, + 0.005173496436327696, + 0.0055251345038414, + 0.01208857074379921, + 0.00797264277935028, + 0.004801186267286539, + 0.007536225952208042, + 0.011095842346549034, + 0.004780077841132879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.012550774961709976 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static bird's-eye view directly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the water chestnut leaves shaped like hearts and colored saturated emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water crystal clear with a deep blue-green hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves sway slowly and rhythmically with the water, sometimes bobbing or spinning in place?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit the visual style of a Ghibli animation with warm, vibrant, non-glaring colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486480712890625 + ], + "opposite_logprob_tokens": [ + -1.2236480712890625 + ], + "logprob_reward": -0.3486480712890625, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486480712890625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03747503459453583, + 0.010954272001981735, + 0.014768070541322231, + 0.8792800903320312, + 0.014941513538360596, + 0.006748480256646872, + 0.006824350915849209, + 0.011530050076544285, + 0.0048588295467197895, + 0.012619341723620892 + ], + "credit_map": [ + 0.003162199864163995, + 0.005762265995144844, + 0.01005686167627573, + 0.002565002767369151, + 0.003352973610162735, + 0.001974145183339715, + 0.008605971932411194, + 0.001995617989450693, + 0.0011512432247400284, + 0.0008723948849365115, + 0.002013332676142454, + 0.0013425805373117328, + 0.0012044166214764118, + 0.0012050167424604297, + 0.0019328199559822679, + 0.0012324686395004392, + 0.0011563810985535383, + 0.0013070340501144528, + 0.0015758712543174624, + 0.004943482112139463, + 0.0010656524682417512, + 0.001109133823774755, + 0.0019944661762565374, + 0.0016160508384928107, + 0.0818723663687706, + 0.21569791436195374, + 0.07009033858776093, + 0.033799685537815094, + 0.05711040645837784, + 0.025075098499655724, + 0.27712973952293396, + 0.11850450932979584, + 0.0024304352700710297, + 0.002930350136011839, + 0.0016403190093114972, + 0.0013067827094346285, + 0.0014516521478071809, + 0.0016053694998845458, + 0.0019828490912914276, + 0.0015937568387016654, + 0.0009298358345404267, + 0.0011695014545693994, + 0.0007106835255399346, + 0.0007937952759675682, + 0.0007416479638777673, + 0.0008330331183969975, + 0.0008667538058944046, + 0.0007032291032373905, + 0.0009783718269318342, + 0.00178253382910043, + 0.0005393147002905607, + 0.0007302846061065793, + 0.0007457667961716652, + 0.0005149419885128736, + 0.0006800637347623706, + 0.0008530737832188606, + 0.001156927552074194, + 0.0019095387542620301, + 0.0010267370380461216, + 0.0021105571649968624, + 0.0012033384991809726, + 0.0008496603113599122, + 0.0016445519868284464, + 0.0016287392936646938, + 0.000477110588690266, + 0.0009222002699971199, + 0.0005372044397518039, + 0.001099382876418531, + 0.0005584642640314996, + 0.00045286418753676116, + 0.0005043878336437047, + 0.00030721581424586475, + 0.0009662335505709052, + 0.003026226069778204, + 0.0012717832578346133, + 0.001210549846291542, + 0.0012083284091204405, + 0.0007375694694928825, + 0.003381845774129033, + 0.0008168058702722192 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.3486480712890625 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, bird's-eye view, taken from directly above, captures a tranquil lake, scattered across it by countless tiny, heart-shaped water chestnut leaves like hand-painted green sprites. Each leaf embodies a saturated, soft emerald green, with rounded edges and delicate veins visible. The water is crystal clear, a deep blue-green hue. Warm afternoon sunlight filters through the ripples, creating countless dancing, shimmering golden specks that resemble highlights in hand-drawn animation. The leaves sway slowly and rhythmically with the gentle rise and fall of the water, sometimes gently bobbing and sometimes spinning leisurely in place. The entire scene is imbued with the delicate, warm quality of a Ghibli animation, with vibrant, yet not glaring, colors creating a magical moment of profound tranquility, healing, and a touch of nostalgia.", + "seed": 25657, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_15/index_6468_seed_25657_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.30667589195072653, + "vlm_logprob": -0.30667589195072653, + "vlm_yes_no_margin": 1.95, + "vlm_token_credit": -0.30667589195072653, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static bird's-eye view directly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the water chestnut leaves shaped like hearts and colored saturated emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water crystal clear with a deep blue-green hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves sway slowly and rhythmically with the water, sometimes bobbing or spinning in place?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit the visual style of a Ghibli animation with warm, vibrant, non-glaring colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.26499465107917786 + ], + "opposite_logprob_tokens": [ + -1.5149946212768555 + ], + "logprob_reward": -0.26499465107917786, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.26499465107917786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15717598795890808, + 0.0647183209657669, + 0.07507804781198502, + 0.06909902393817902, + 0.12914052605628967, + 0.08678683638572693, + 0.07682915776968002, + 0.2485324740409851, + 0.03389187902212143, + 0.058747656643390656 + ], + "credit_map": [ + 0.021654821932315826, + 0.013816313818097115, + 0.011264868080615997, + 0.016668129712343216, + 0.015526204369962215, + 0.02613813616335392, + 0.016931230202317238, + 0.03517629951238632, + 0.005559782963246107, + 0.005452232901006937, + 0.007791444659233093, + 0.008400826714932919, + 0.0068861995823681355, + 0.009718998335301876, + 0.00955003872513771, + 0.011358806863427162, + 0.011250128038227558, + 0.0077882613986730576, + 0.010848446749150753, + 0.008627845905721188, + 0.00746410246938467, + 0.009283237159252167, + 0.00966220535337925, + 0.010153829120099545, + 0.011826763860881329, + 0.005909303203225136, + 0.006347015034407377, + 0.010212782770395279, + 0.007454042322933674, + 0.009407347068190575, + 0.009629893116652966, + 0.00831188540905714, + 0.017757302150130272, + 0.01626117154955864, + 0.01355357002466917, + 0.015241176821291447, + 0.0171233918517828, + 0.01654592901468277, + 0.015832727774977684, + 0.016825266182422638, + 0.013050304725766182, + 0.010679475031793118, + 0.008946728892624378, + 0.008796828798949718, + 0.012281008996069431, + 0.01029540691524744, + 0.010406034998595715, + 0.012331063859164715, + 0.00990336574614048, + 0.00829258095473051, + 0.007865768857300282, + 0.005762823391705751, + 0.010251727886497974, + 0.009082340635359287, + 0.010959699749946594, + 0.014710856601595879, + 0.015300371684134007, + 0.009123926050961018, + 0.009429204277694225, + 0.009702644310891628, + 0.013129905797541142, + 0.08670391142368317, + 0.014616884291172028, + 0.09052567183971405, + 0.00285426271148026, + 0.00417301757261157, + 0.002729570958763361, + 0.0027628703974187374, + 0.004392686299979687, + 0.006050747353583574, + 0.0034631695598363876, + 0.007465556263923645, + 0.008665683679282665, + 0.00627625547349453, + 0.006015298422425985, + 0.006836066488176584, + 0.006865187548100948, + 0.006848357152193785, + 0.008835344575345516, + 0.008405473083257675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.26499465107917786 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static bird's-eye view directly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the water chestnut leaves shaped like hearts and colored saturated emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water crystal clear with a deep blue-green hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves sway slowly and rhythmically with the water, sometimes bobbing or spinning in place?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit the visual style of a Ghibli animation with warm, vibrant, non-glaring colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038109004497528076 + ], + "opposite_logprob_tokens": [ + -3.288109064102173 + ], + "logprob_reward": -0.038109004497528076, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038109004497528076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25170621275901794, + 0.0563005767762661, + 0.05413274094462395, + 0.07755127549171448, + 0.13368748128414154, + 0.0717187449336052, + 0.03717176243662834, + 0.23937363922595978, + 0.04356604814529419, + 0.03479146584868431 + ], + "credit_map": [ + 0.016929306089878082, + 0.02043430134654045, + 0.03996508568525314, + 0.03271183744072914, + 0.03098008781671524, + 0.042892664670944214, + 0.02937418222427368, + 0.03841874375939369, + 0.00663165794685483, + 0.004418946336954832, + 0.009332499466836452, + 0.0051854983903467655, + 0.005691819824278355, + 0.009257766418159008, + 0.008140448480844498, + 0.007641942240297794, + 0.009626738727092743, + 0.006062828004360199, + 0.006620573811233044, + 0.005766705144196749, + 0.006872914265841246, + 0.00567703228443861, + 0.007428567390888929, + 0.0060773808509111404, + 0.011648355983197689, + 0.007235904689878225, + 0.008094203658401966, + 0.010135319083929062, + 0.007440568413585424, + 0.011609070003032684, + 0.010285084135830402, + 0.011102762073278427, + 0.011782901361584663, + 0.010774227790534496, + 0.021218204870820045, + 0.016494760289788246, + 0.015834452584385872, + 0.018201680853962898, + 0.01753195933997631, + 0.021849289536476135, + 0.008971964940428734, + 0.004837638232856989, + 0.0054005589336156845, + 0.007174840196967125, + 0.007278971374034882, + 0.010121663101017475, + 0.012073788791894913, + 0.015859320759773254, + 0.006269319448620081, + 0.0049694860354065895, + 0.003539067693054676, + 0.0028726626187562943, + 0.00584851810708642, + 0.0045224460773169994, + 0.005498907994478941, + 0.003651355393230915, + 0.016322869807481766, + 0.008588957600295544, + 0.015470610931515694, + 0.01593708246946335, + 0.022819798439741135, + 0.07561299204826355, + 0.020678440108895302, + 0.06394290924072266, + 0.004819895606487989, + 0.005867248401045799, + 0.0036450326442718506, + 0.004613232798874378, + 0.005619564093649387, + 0.009269346483051777, + 0.003834380768239498, + 0.005897349212318659, + 0.0050621782429516315, + 0.0036736687179654837, + 0.0047998856753110886, + 0.0033882653806358576, + 0.0056370943784713745, + 0.004179326817393303, + 0.0036056246608495712, + 0.004445421975106001 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.038109004497528076 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static bird's-eye view directly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the water chestnut leaves shaped like hearts and colored saturated emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water crystal clear with a deep blue-green hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves sway slowly and rhythmically with the water, sometimes bobbing or spinning in place?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit the visual style of a Ghibli animation with warm, vibrant, non-glaring colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0538387298583984 + ], + "opposite_logprob_tokens": [ + -0.42883870005607605 + ], + "logprob_reward": -1.0538387298583984, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0538387298583984, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.15134398639202118, + 0.1456015706062317, + 0.04187833517789841, + 0.09249062091112137, + 0.10800404846668243, + 0.0763501524925232, + 0.06515048444271088, + 0.2076735496520996, + 0.04806942120194435, + 0.0634378045797348 + ], + "credit_map": [ + 0.016802961006760597, + 0.01250887755304575, + 0.017859360203146935, + 0.020399514585733414, + 0.014862007461488247, + 0.02782888524234295, + 0.0212031751871109, + 0.01987919956445694, + 0.01807747222483158, + 0.01243418175727129, + 0.011153343133628368, + 0.020120516419410706, + 0.015297133475542068, + 0.018061010167002678, + 0.025230849161744118, + 0.02522706799209118, + 0.004494518972933292, + 0.004471654072403908, + 0.004380549769848585, + 0.004594792611896992, + 0.0034800979774445295, + 0.007599500007927418, + 0.0062922523356974125, + 0.006564966402947903, + 0.016529612243175507, + 0.00867738388478756, + 0.014219069853425026, + 0.011009152047336102, + 0.006867096293717623, + 0.014606640674173832, + 0.009372700937092304, + 0.011208968237042427, + 0.013500557281076908, + 0.014289683662354946, + 0.015760159119963646, + 0.011348721571266651, + 0.012710782699286938, + 0.013611557893455029, + 0.011500079184770584, + 0.015282503329217434, + 0.009107244201004505, + 0.006712873466312885, + 0.008991344831883907, + 0.010807173326611519, + 0.01172648649662733, + 0.008616654202342033, + 0.010484866797924042, + 0.009903507307171822, + 0.010122651234269142, + 0.007238322403281927, + 0.0053035602904856205, + 0.003584533231332898, + 0.007816560566425323, + 0.009026043117046356, + 0.009368005208671093, + 0.012690803036093712, + 0.019076082855463028, + 0.0158771313726902, + 0.020901653915643692, + 0.02456398867070675, + 0.03705592826008797, + 0.03272848576307297, + 0.022219674661755562, + 0.03525061160326004, + 0.0032647494226694107, + 0.004616432823240757, + 0.003349629696458578, + 0.0039217667654156685, + 0.006720105651766062, + 0.01097929198294878, + 0.006565264891833067, + 0.00865217950195074, + 0.009681777097284794, + 0.00690813921391964, + 0.005890585947781801, + 0.0050476957112550735, + 0.007771536707878113, + 0.009012369439005852, + 0.0070608751848340034, + 0.012064824812114239 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -1.0538387298583984 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static bird's-eye view directly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the water chestnut leaves shaped like hearts and colored saturated emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water crystal clear with a deep blue-green hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves sway slowly and rhythmically with the water, sometimes bobbing or spinning in place?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit the visual style of a Ghibli animation with warm, vibrant, non-glaring colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016079628840088844 + ], + "opposite_logprob_tokens": [ + -4.141079425811768 + ], + "logprob_reward": -0.016079628840088844, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016079628840088844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12716257572174072, + 0.049414873123168945, + 0.10754919052124023, + 0.14045222103595734, + 0.043325025588274, + 0.029289621859788895, + 0.057894762605428696, + 0.23344025015830994, + 0.15551136434078217, + 0.055960167199373245 + ], + "credit_map": [ + 0.01472054049372673, + 0.011258776299655437, + 0.01474690716713667, + 0.018273351714015007, + 0.013752427883446217, + 0.01673591509461403, + 0.018069379031658173, + 0.01960526779294014, + 0.0043059103190898895, + 0.0034248584415763617, + 0.006886780261993408, + 0.00885824952274561, + 0.0033259198535233736, + 0.011059078387916088, + 0.006660498678684235, + 0.004893575794994831, + 0.006550068501383066, + 0.010231499560177326, + 0.014909404329955578, + 0.011417227797210217, + 0.014014977030456066, + 0.016969013959169388, + 0.019021524116396904, + 0.014435473829507828, + 0.015948845073580742, + 0.012711388058960438, + 0.01672547124326229, + 0.013792498968541622, + 0.01908981241285801, + 0.017981654033064842, + 0.01745028980076313, + 0.02675226703286171, + 0.0048735779710114, + 0.004360751248896122, + 0.004888494033366442, + 0.004669944290071726, + 0.0053598093800246716, + 0.006908938754349947, + 0.005587238352745771, + 0.0066762748174369335, + 0.004584353882819414, + 0.0029592541977763176, + 0.002871415577828884, + 0.003200025297701359, + 0.004646356217563152, + 0.0035895435139536858, + 0.002971311565488577, + 0.004467361606657505, + 0.00874978955835104, + 0.006103108171373606, + 0.0046386998146772385, + 0.004342259373515844, + 0.007277824450284243, + 0.008754628710448742, + 0.008393579162657261, + 0.00963487010449171, + 0.015531199984252453, + 0.016509676352143288, + 0.018813401460647583, + 0.01651643216609955, + 0.023111999034881592, + 0.05334435775876045, + 0.029722947627305984, + 0.05989024043083191, + 0.013492756523191929, + 0.013258353807032108, + 0.01298163179308176, + 0.014028637669980526, + 0.018912430852651596, + 0.031117761507630348, + 0.024128152057528496, + 0.027591636404395103, + 0.007386285346001387, + 0.006633054465055466, + 0.007842166349291801, + 0.0049728029407560825, + 0.005718478467315435, + 0.007320627570152283, + 0.008666451089084148, + 0.007420303765684366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 3 + ], + "selected_reward": -0.016079628840088844 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static bird's-eye view directly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the water chestnut leaves shaped like hearts and colored saturated emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water crystal clear with a deep blue-green hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves sway slowly and rhythmically with the water, sometimes bobbing or spinning in place?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit the visual style of a Ghibli animation with warm, vibrant, non-glaring colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16035744547843933 + ], + "opposite_logprob_tokens": [ + -1.9103574752807617 + ], + "logprob_reward": -0.16035744547843933, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16035744547843933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11727665364742279, + 0.1482328176498413, + 0.1048143059015274, + 0.07018382847309113, + 0.03842991590499878, + 0.07112365961074829, + 0.035510845482349396, + 0.29614830017089844, + 0.03673712536692619, + 0.08154257386922836 + ], + "credit_map": [ + 0.01447733212262392, + 0.013232695870101452, + 0.012299804016947746, + 0.023502910509705544, + 0.01155894435942173, + 0.014861064963042736, + 0.013704788871109486, + 0.013639112934470177, + 0.019185194745659828, + 0.014711770229041576, + 0.012705094181001186, + 0.021894514560699463, + 0.024834908545017242, + 0.017300421372056007, + 0.01683533750474453, + 0.020765557885169983, + 0.010037551634013653, + 0.012890077196061611, + 0.009729200974106789, + 0.009940270334482193, + 0.013943993486464024, + 0.016638940200209618, + 0.015500102192163467, + 0.01613415777683258, + 0.014437372796237469, + 0.00881543755531311, + 0.00627743499353528, + 0.007464216556400061, + 0.00723546976223588, + 0.011172342114150524, + 0.0062356144189834595, + 0.008545936085283756, + 0.0070697832852602005, + 0.005404679570347071, + 0.003949031233787537, + 0.003946198616176844, + 0.004146409220993519, + 0.006049869116395712, + 0.003159918589517474, + 0.0047040218487381935, + 0.010606263764202595, + 0.008782649412751198, + 0.006315149832516909, + 0.005342206452041864, + 0.006358536891639233, + 0.011043648235499859, + 0.010475951246917248, + 0.01219924446195364, + 0.006615988910198212, + 0.005162614863365889, + 0.004725580103695393, + 0.0030391751788556576, + 0.0040023718029260635, + 0.004472438711673021, + 0.0031501473858952522, + 0.004342527594417334, + 0.024220643565058708, + 0.021861841902136803, + 0.03145832195878029, + 0.02589074894785881, + 0.030967917293310165, + 0.06665952503681183, + 0.04057244211435318, + 0.05451685190200806, + 0.0028867695946246386, + 0.0032994442153722048, + 0.0038406315725296736, + 0.0033417195081710815, + 0.00378841208294034, + 0.00724086444824934, + 0.003948420751839876, + 0.008390864357352257, + 0.014796413481235504, + 0.00927188154309988, + 0.010824868455529213, + 0.008081817999482155, + 0.009114129468798637, + 0.009650874882936478, + 0.007829933427274227, + 0.011972649022936821 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.16035744547843933 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, bird's-eye view, taken from directly above, captures a tranquil lake, scattered across it by countless tiny, heart-shaped water chestnut leaves like hand-painted green sprites. Each leaf embodies a saturated, soft emerald green, with rounded edges and delicate veins visible. The water is crystal clear, a deep blue-green hue. Warm afternoon sunlight filters through the ripples, creating countless dancing, shimmering golden specks that resemble highlights in hand-drawn animation. The leaves sway slowly and rhythmically with the gentle rise and fall of the water, sometimes gently bobbing and sometimes spinning leisurely in place. The entire scene is imbued with the delicate, warm quality of a Ghibli animation, with vibrant, yet not glaring, colors creating a magical moment of profound tranquility, healing, and a touch of nostalgia.", + "seed": 25658, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_15/index_6468_seed_25658_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08574530258774757, + "vlm_logprob": -0.08574530258774757, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.08574530258774757, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static bird's-eye view directly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the water chestnut leaves shaped like hearts and colored saturated emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water crystal clear with a deep blue-green hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves sway slowly and rhythmically with the water, sometimes bobbing or spinning in place?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit the visual style of a Ghibli animation with warm, vibrant, non-glaring colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2357267439365387 + ], + "opposite_logprob_tokens": [ + -1.6107267141342163 + ], + "logprob_reward": -0.2357267439365387, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2357267439365387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11604311317205429, + 0.06856696307659149, + 0.1408083736896515, + 0.03003372810781002, + 0.0204875897616148, + 0.15154513716697693, + 0.1773354560136795, + 0.21585513651371002, + 0.05013202503323555, + 0.029192408546805382 + ], + "credit_map": [ + 0.01159279141575098, + 0.011026022955775261, + 0.013378070667386055, + 0.02308937907218933, + 0.012918109074234962, + 0.011989269405603409, + 0.013561089523136616, + 0.01848839409649372, + 0.010038208216428757, + 0.009113376960158348, + 0.0055749439634382725, + 0.006850214209407568, + 0.009862358681857586, + 0.008532718755304813, + 0.006967852357774973, + 0.011627285741269588, + 0.011465301737189293, + 0.015149418264627457, + 0.011801555752754211, + 0.026406271383166313, + 0.02199973724782467, + 0.017238037660717964, + 0.016120675951242447, + 0.020627375692129135, + 0.002667760942131281, + 0.0022648745216429234, + 0.0038334678392857313, + 0.00709124282002449, + 0.0022478438913822174, + 0.003279702039435506, + 0.004464665427803993, + 0.004184171557426453, + 0.0020465352572500706, + 0.0023112173657864332, + 0.0025781027507036924, + 0.0030322198290377855, + 0.0020964499562978745, + 0.002161373384296894, + 0.0030742294620722532, + 0.0031874626874923706, + 0.01599998027086258, + 0.01647363230586052, + 0.023465435951948166, + 0.024585280567407608, + 0.020661620423197746, + 0.018153265118598938, + 0.019198985770344734, + 0.013006933033466339, + 0.026196900755167007, + 0.020006705075502396, + 0.01943831704556942, + 0.019036980345845222, + 0.025651345029473305, + 0.020729223266243935, + 0.025212178006768227, + 0.02106381766498089, + 0.017063520848751068, + 0.016774626448750496, + 0.0295576062053442, + 0.02796282060444355, + 0.031708184629678726, + 0.02710692025721073, + 0.03795349597930908, + 0.027727970853447914, + 0.00395056651905179, + 0.008215358480811119, + 0.004689020570367575, + 0.00783387478441, + 0.008121776394546032, + 0.009647675789892673, + 0.004384530242532492, + 0.0032892264425754547, + 0.0014398330822587013, + 0.001990517135709524, + 0.0024406835436820984, + 0.002896428108215332, + 0.007640528492629528, + 0.005988216493278742, + 0.0026337127201259136, + 0.004162488039582968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.2357267439365387 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static bird's-eye view directly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the water chestnut leaves shaped like hearts and colored saturated emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water crystal clear with a deep blue-green hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves sway slowly and rhythmically with the water, sometimes bobbing or spinning in place?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit the visual style of a Ghibli animation with warm, vibrant, non-glaring colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029814405366778374 + ], + "opposite_logprob_tokens": [ + -3.5298144817352295 + ], + "logprob_reward": -0.029814405366778374, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029814405366778374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.160955011844635, + 0.1648939549922943, + 0.05639122053980827, + 0.04128003120422363, + 0.03598680719733238, + 0.0687347799539566, + 0.14966291189193726, + 0.15418171882629395, + 0.07647818326950073, + 0.09143539518117905 + ], + "credit_map": [ + 0.017904022708535194, + 0.021751554682850838, + 0.019912781193852425, + 0.019378526136279106, + 0.01914827711880207, + 0.017066508531570435, + 0.021716196089982986, + 0.024077147245407104, + 0.02206442877650261, + 0.01929580606520176, + 0.022578567266464233, + 0.01780335046350956, + 0.025530483573675156, + 0.0159715935587883, + 0.019977247342467308, + 0.021672483533620834, + 0.006742960773408413, + 0.008000539615750313, + 0.006078304257243872, + 0.009187287651002407, + 0.006645047105848789, + 0.006520222872495651, + 0.005629397928714752, + 0.007587461266666651, + 0.003269354347139597, + 0.0038595867808908224, + 0.004603675566613674, + 0.009830071590840816, + 0.005288127344101667, + 0.004337682388722897, + 0.004859714303165674, + 0.005231819115579128, + 0.0036278781481087208, + 0.004219505470246077, + 0.004222044721245766, + 0.005540850106626749, + 0.0049906205385923386, + 0.004731767810881138, + 0.003917946480214596, + 0.004736193455755711, + 0.007424080744385719, + 0.005826004780828953, + 0.009309748187661171, + 0.01175815798342228, + 0.006869735196232796, + 0.007134289480745792, + 0.01067996397614479, + 0.009732795879244804, + 0.024522075429558754, + 0.017957432195544243, + 0.017766349017620087, + 0.015319190919399261, + 0.022664565593004227, + 0.01740334928035736, + 0.01939345709979534, + 0.014636488631367683, + 0.010819974355399609, + 0.01071784645318985, + 0.019812023267149925, + 0.01829933561384678, + 0.025647062808275223, + 0.01931190863251686, + 0.027505921199917793, + 0.022067643702030182, + 0.006105922628194094, + 0.016410402953624725, + 0.006605023518204689, + 0.007963458076119423, + 0.016712725162506104, + 0.009948437102138996, + 0.006582627072930336, + 0.006149583961814642, + 0.007137809880077839, + 0.003080949652940035, + 0.0055185407400131226, + 0.02429146133363247, + 0.007965631783008575, + 0.010448293760418892, + 0.0077074612490832806, + 0.02528524212539196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.029814405366778374 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static bird's-eye view directly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the water chestnut leaves shaped like hearts and colored saturated emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water crystal clear with a deep blue-green hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves sway slowly and rhythmically with the water, sometimes bobbing or spinning in place?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit the visual style of a Ghibli animation with warm, vibrant, non-glaring colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789836049079895 + ], + "opposite_logprob_tokens": [ + -2.5789835453033447 + ], + "logprob_reward": -0.0789836049079895, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789836049079895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10739884525537491, + 0.08812795579433441, + 0.0319351889193058, + 0.05037129297852516, + 0.12068653851747513, + 0.11887765675783157, + 0.044638969004154205, + 0.20160436630249023, + 0.179365873336792, + 0.0569932758808136 + ], + "credit_map": [ + 0.012988158501684666, + 0.015964914113283157, + 0.011143486015498638, + 0.012829439714550972, + 0.013645106926560402, + 0.013308263383805752, + 0.011472763493657112, + 0.016046715900301933, + 0.011460939422249794, + 0.009373679757118225, + 0.012510308064520359, + 0.010421219281852245, + 0.012602579779922962, + 0.009388183243572712, + 0.011289961636066437, + 0.011081092990934849, + 0.0027353146579116583, + 0.0045670149847865105, + 0.002953849732875824, + 0.00862062443047762, + 0.0026319867465645075, + 0.0037261489778757095, + 0.003880181349813938, + 0.0028200673405081034, + 0.0035292766988277435, + 0.004238552413880825, + 0.0072240084409713745, + 0.011008070781826973, + 0.0046081249602139, + 0.005527810659259558, + 0.009450413286685944, + 0.004785038530826569, + 0.013004908338189125, + 0.01563890464603901, + 0.012494551949203014, + 0.017076127231121063, + 0.012927896343171597, + 0.013020899146795273, + 0.014474714174866676, + 0.022048531100153923, + 0.015189271420240402, + 0.013939562253654003, + 0.020333416759967804, + 0.013448496349155903, + 0.01277992408722639, + 0.01192411221563816, + 0.01567651890218258, + 0.01558635663241148, + 0.005028105340898037, + 0.0034234432969242334, + 0.0046609691344201565, + 0.005213348660618067, + 0.007010295055806637, + 0.005205876659601927, + 0.0069507756270468235, + 0.007146153599023819, + 0.016084037721157074, + 0.01587466150522232, + 0.02966584824025631, + 0.02800792083144188, + 0.027151891961693764, + 0.024317800998687744, + 0.033781085163354874, + 0.026721114292740822, + 0.01365147065371275, + 0.016990426927804947, + 0.016762230545282364, + 0.030223194509744644, + 0.030934741720557213, + 0.035660937428474426, + 0.021137669682502747, + 0.014005208387970924, + 0.0035295304842293262, + 0.0038024845998734236, + 0.0042189061641693115, + 0.009298205375671387, + 0.010771366767585278, + 0.007874280214309692, + 0.003858308307826519, + 0.013640196062624454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 4 + ], + "selected_reward": -0.0789836049079895 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static bird's-eye view directly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the water chestnut leaves shaped like hearts and colored saturated emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water crystal clear with a deep blue-green hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves sway slowly and rhythmically with the water, sometimes bobbing or spinning in place?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit the visual style of a Ghibli animation with warm, vibrant, non-glaring colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014199061319231987 + ], + "opposite_logprob_tokens": [ + -4.264199256896973 + ], + "logprob_reward": -0.014199061319231987, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014199061319231987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1915685385465622, + 0.09113402664661407, + 0.12924546003341675, + 0.057979755103588104, + 0.13413110375404358, + 0.15647274255752563, + 0.09459862858057022, + 0.05487833917140961, + 0.036686308681964874, + 0.05330507457256317 + ], + "credit_map": [ + 0.020080475136637688, + 0.021980799734592438, + 0.016104351729154587, + 0.025952713564038277, + 0.024497903883457184, + 0.021408751606941223, + 0.02664569579064846, + 0.03489784896373749, + 0.010532969608902931, + 0.009253616444766521, + 0.01431975793093443, + 0.011374388821423054, + 0.012427832931280136, + 0.011791171506047249, + 0.008666770532727242, + 0.01276751235127449, + 0.009970653802156448, + 0.014956984668970108, + 0.010660270228981972, + 0.029411455616354942, + 0.014006993733346462, + 0.017354676499962807, + 0.015152016654610634, + 0.0177324041724205, + 0.004483907017856836, + 0.005542008671909571, + 0.008957864716649055, + 0.015953997150063515, + 0.005264129023998976, + 0.006597993429750204, + 0.005862127989530563, + 0.005317730363458395, + 0.013155425898730755, + 0.02162492275238037, + 0.01133791171014309, + 0.0166724044829607, + 0.015247146598994732, + 0.01525446493178606, + 0.01747661642730236, + 0.023362213745713234, + 0.015437846072018147, + 0.016165364533662796, + 0.024018481373786926, + 0.023039616644382477, + 0.02252836525440216, + 0.019189471378922462, + 0.02157970890402794, + 0.014513901434838772, + 0.016423897817730904, + 0.011157538741827011, + 0.010703257285058498, + 0.009436608292162418, + 0.01351337879896164, + 0.010629547759890556, + 0.012861885130405426, + 0.009872508235275745, + 0.004477354697883129, + 0.004462576936930418, + 0.00536270160228014, + 0.00544273154810071, + 0.008236879482865334, + 0.007087565027177334, + 0.011039355769753456, + 0.008769175037741661, + 0.0031336909160017967, + 0.0057340022176504135, + 0.0036423224955797195, + 0.005871957633644342, + 0.0060244398191571236, + 0.005148912779986858, + 0.0035848289262503386, + 0.0035461559891700745, + 0.005571894813328981, + 0.0034004307817667723, + 0.0038994571659713984, + 0.009932314045727253, + 0.006384273525327444, + 0.0058501060120761395, + 0.00438575679436326, + 0.013880842365324497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.014199061319231987 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static bird's-eye view directly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the water chestnut leaves shaped like hearts and colored saturated emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water crystal clear with a deep blue-green hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves sway slowly and rhythmically with the water, sometimes bobbing or spinning in place?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit the visual style of a Ghibli animation with warm, vibrant, non-glaring colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000269740819931 + ], + "opposite_logprob_tokens": [ + -2.695002794265747 + ], + "logprob_reward": -0.07000269740819931, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000269740819931, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14420001208782196, + 0.06516465544700623, + 0.04333392158150673, + 0.03525999188423157, + 0.09579101949930191, + 0.05686085298657417, + 0.17788216471672058, + 0.21064193546772003, + 0.13268040120601654, + 0.038185037672519684 + ], + "credit_map": [ + 0.015716789290308952, + 0.019617320969700813, + 0.013417224399745464, + 0.022559164091944695, + 0.017443416640162468, + 0.01832428202033043, + 0.014002816751599312, + 0.023118993267416954, + 0.006332388147711754, + 0.008053661324083805, + 0.007176091894507408, + 0.00794744398444891, + 0.009419282898306847, + 0.006959188729524612, + 0.008827601559460163, + 0.010448998771607876, + 0.003860027529299259, + 0.006812619511038065, + 0.003973386716097593, + 0.006295030005276203, + 0.006580411922186613, + 0.004709299653768539, + 0.004928047768771648, + 0.006175100337713957, + 0.003327617421746254, + 0.003494192147627473, + 0.0064451927319169044, + 0.005040482152253389, + 0.004378139041364193, + 0.0036972372326999903, + 0.004427154548466206, + 0.004449975676834583, + 0.010351092554628849, + 0.01254303939640522, + 0.010142345912754536, + 0.012715098448097706, + 0.01226012036204338, + 0.010331515222787857, + 0.01154213584959507, + 0.01590566709637642, + 0.00824948400259018, + 0.006463341414928436, + 0.008359809406101704, + 0.007325352169573307, + 0.00640935217961669, + 0.006489960942417383, + 0.00676274998113513, + 0.006800800561904907, + 0.027071243152022362, + 0.02015070617198944, + 0.018816927447915077, + 0.016318008303642273, + 0.025848664343357086, + 0.020776789635419846, + 0.025684522464871407, + 0.023215321823954582, + 0.01679745502769947, + 0.017742054536938667, + 0.02811901830136776, + 0.027906347066164017, + 0.031189342960715294, + 0.02512526884675026, + 0.03545747697353363, + 0.028304971754550934, + 0.008813215419650078, + 0.02263529971241951, + 0.011528206057846546, + 0.01784363202750683, + 0.029214289039373398, + 0.02030288241803646, + 0.012577082961797714, + 0.00976579263806343, + 0.00477387523278594, + 0.002712007611989975, + 0.004928163718432188, + 0.0043798452243208885, + 0.004773986060172319, + 0.004895803518593311, + 0.004800766706466675, + 0.006920591928064823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.07000269740819931 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, bird's-eye view, taken from directly above, captures a tranquil lake, scattered across it by countless tiny, heart-shaped water chestnut leaves like hand-painted green sprites. Each leaf embodies a saturated, soft emerald green, with rounded edges and delicate veins visible. The water is crystal clear, a deep blue-green hue. Warm afternoon sunlight filters through the ripples, creating countless dancing, shimmering golden specks that resemble highlights in hand-drawn animation. The leaves sway slowly and rhythmically with the gentle rise and fall of the water, sometimes gently bobbing and sometimes spinning leisurely in place. The entire scene is imbued with the delicate, warm quality of a Ghibli animation, with vibrant, yet not glaring, colors creating a magical moment of profound tranquility, healing, and a touch of nostalgia.", + "seed": 25659, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_15/index_6468_seed_25659_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09073655940592289, + "vlm_logprob": -0.09073655940592289, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.09073655940592289, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static bird's-eye view directly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the water chestnut leaves shaped like hearts and colored saturated emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water crystal clear with a deep blue-green hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves sway slowly and rhythmically with the water, sometimes bobbing or spinning in place?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit the visual style of a Ghibli animation with warm, vibrant, non-glaring colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2444932758808136 + ], + "opposite_logprob_tokens": [ + -1.6194932460784912 + ], + "logprob_reward": -0.2444932758808136, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2444932758808136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16895103454589844, + 0.04328017309308052, + 0.23490847647190094, + 0.01775306463241577, + 0.04023543372750282, + 0.13510848581790924, + 0.03491343557834625, + 0.1313256174325943, + 0.15561705827713013, + 0.03790721669793129 + ], + "credit_map": [ + 0.018597813323140144, + 0.033226095139980316, + 0.013026181608438492, + 0.022165941074490547, + 0.030300555750727654, + 0.016658499836921692, + 0.017731133848428726, + 0.017244800925254822, + 0.005962891038507223, + 0.006161997094750404, + 0.004992353729903698, + 0.0060476139187812805, + 0.004667883738875389, + 0.004389779642224312, + 0.004569781478494406, + 0.006487877108156681, + 0.012966261245310307, + 0.014564922079443932, + 0.009952410124242306, + 0.006617947947233915, + 0.08595356345176697, + 0.08025344461202621, + 0.015935100615024567, + 0.008664822205901146, + 0.002594074234366417, + 0.0025862918701022863, + 0.0017453469336032867, + 0.002039415528997779, + 0.002467383397743106, + 0.002102089812979102, + 0.001820883946493268, + 0.0023975796066224575, + 0.005305904429405928, + 0.005760198924690485, + 0.0029167698230594397, + 0.0037040517199784517, + 0.005772917997092009, + 0.00395475747063756, + 0.004777395632117987, + 0.008043436333537102, + 0.020480381324887276, + 0.011235188692808151, + 0.010858750902116299, + 0.00807611271739006, + 0.04331163316965103, + 0.021145261824131012, + 0.01093448419123888, + 0.009066682308912277, + 0.005735841114073992, + 0.005276357755064964, + 0.002755036810413003, + 0.0037374356761574745, + 0.0030296246986836195, + 0.005985678173601627, + 0.004635411314666271, + 0.0037580495700240135, + 0.034638844430446625, + 0.011046869680285454, + 0.04317177087068558, + 0.00704818032681942, + 0.0044365511275827885, + 0.01368444412946701, + 0.009611050598323345, + 0.007687909062951803, + 0.036944277584552765, + 0.0096594812348485, + 0.055504970252513885, + 0.010625200346112251, + 0.009757786989212036, + 0.00726420059800148, + 0.015479369089007378, + 0.010381769388914108, + 0.0033663439098745584, + 0.004015012644231319, + 0.008340712636709213, + 0.003203543834388256, + 0.0036275621969252825, + 0.006896280683577061, + 0.006099735852330923, + 0.002358024474233389 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.2444932758808136 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static bird's-eye view directly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the water chestnut leaves shaped like hearts and colored saturated emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water crystal clear with a deep blue-green hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves sway slowly and rhythmically with the water, sometimes bobbing or spinning in place?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit the visual style of a Ghibli animation with warm, vibrant, non-glaring colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026358038187026978 + ], + "opposite_logprob_tokens": [ + -3.651358127593994 + ], + "logprob_reward": -0.026358038187026978, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026358038187026978, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06804551929235458, + 0.060698628425598145, + 0.3998253643512726, + 0.013164444826543331, + 0.030855555087327957, + 0.10149561613798141, + 0.0828358381986618, + 0.12411138415336609, + 0.01689809560775757, + 0.10206957161426544 + ], + "credit_map": [ + 0.00848329160362482, + 0.014705716632306576, + 0.005629121325910091, + 0.007107125129550695, + 0.01110236719250679, + 0.008024409413337708, + 0.006408660672605038, + 0.006584823131561279, + 0.007055155001580715, + 0.010170127265155315, + 0.005280034616589546, + 0.006042253691703081, + 0.012009531259536743, + 0.006856192369014025, + 0.006179395597428083, + 0.007105938158929348, + 0.027614887803792953, + 0.02693112939596176, + 0.03287884593009949, + 0.017400283366441727, + 0.07876969128847122, + 0.13298703730106354, + 0.06684327125549316, + 0.01640024036169052, + 0.0013548501301556826, + 0.0016745368484407663, + 0.0014432420721277595, + 0.0014971747295930982, + 0.0014110675547271967, + 0.001169484807178378, + 0.0016588125145062804, + 0.0029552755877375603, + 0.003592723747715354, + 0.0035418099723756313, + 0.0027432036586105824, + 0.002902855398133397, + 0.004507268313318491, + 0.0030502404551953077, + 0.00408715195953846, + 0.006430302746593952, + 0.014387565664947033, + 0.008919481188058853, + 0.010353242978453636, + 0.01226583868265152, + 0.02301209233701229, + 0.01175803691148758, + 0.013934765011072159, + 0.0068645901046693325, + 0.01172663364559412, + 0.006106297951191664, + 0.028407851234078407, + 0.006034145597368479, + 0.004166481085121632, + 0.012445969507098198, + 0.008516095578670502, + 0.005432359874248505, + 0.027357565239071846, + 0.011897769756615162, + 0.03339057415723801, + 0.008418252691626549, + 0.007891868241131306, + 0.014433677308261395, + 0.011366428807377815, + 0.009355246089398861, + 0.0014445078559219837, + 0.0026686592027544975, + 0.0018712752498686314, + 0.00234371330589056, + 0.0016610868042334914, + 0.0022419882006943226, + 0.0021458768751472235, + 0.00252098822966218, + 0.012581353075802326, + 0.008406241424381733, + 0.026776215061545372, + 0.009817292913794518, + 0.009850668720901012, + 0.00919888075441122, + 0.015685973688960075, + 0.00975294690579176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 9 + ], + "selected_reward": -0.026358038187026978 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static bird's-eye view directly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the water chestnut leaves shaped like hearts and colored saturated emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water crystal clear with a deep blue-green hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves sway slowly and rhythmically with the water, sometimes bobbing or spinning in place?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit the visual style of a Ghibli animation with warm, vibrant, non-glaring colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001025974750519 + ], + "opposite_logprob_tokens": [ + -2.695010185241699 + ], + "logprob_reward": -0.07001025974750519, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001025974750519, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08475011587142944, + 0.06382937729358673, + 0.3670850694179535, + 0.0241006538271904, + 0.10223961621522903, + 0.10774548351764679, + 0.034990813583135605, + 0.1569906771183014, + 0.03482591733336449, + 0.02344224974513054 + ], + "credit_map": [ + 0.011284581385552883, + 0.019032930955290794, + 0.0062154545448720455, + 0.008605636656284332, + 0.017107544466853142, + 0.007651667110621929, + 0.008415734395384789, + 0.00643656961619854, + 0.011905097402632236, + 0.009254515171051025, + 0.007272951304912567, + 0.0060400883667171, + 0.007221425883471966, + 0.007036218419671059, + 0.006155917886644602, + 0.0089431656524539, + 0.027083653956651688, + 0.022812852635979652, + 0.039508841931819916, + 0.02240273542702198, + 0.05615741387009621, + 0.06260799616575241, + 0.11688674241304398, + 0.019624851644039154, + 0.0033283703960478306, + 0.003477039746940136, + 0.002190382918342948, + 0.0035436118487268686, + 0.001953889848664403, + 0.0023111256305128336, + 0.0023515422362834215, + 0.004944691900163889, + 0.01139420922845602, + 0.010222931392490864, + 0.009986935183405876, + 0.008472389541566372, + 0.01115079689770937, + 0.01110804546624422, + 0.009144573472440243, + 0.03075972944498062, + 0.019058050587773323, + 0.00734727643430233, + 0.0039876410737633705, + 0.003049027407541871, + 0.04140602797269821, + 0.019997136667370796, + 0.009257026948034763, + 0.0036432889755815268, + 0.004357959143817425, + 0.003486224450170994, + 0.010983806103467941, + 0.002541717840358615, + 0.0020376930478960276, + 0.005731887184083462, + 0.003560737706720829, + 0.002290787873789668, + 0.04158182069659233, + 0.013565942645072937, + 0.053173840045928955, + 0.008877688087522984, + 0.00506279431283474, + 0.014990570023655891, + 0.011621460318565369, + 0.008116556331515312, + 0.008856679312884808, + 0.0023244500625878572, + 0.010823972523212433, + 0.0022703257855027914, + 0.0020227176137268543, + 0.0021200724877417088, + 0.0039485907182097435, + 0.0024591090623289347, + 0.00231923651881516, + 0.002362894592806697, + 0.005145559087395668, + 0.0015422161668539047, + 0.001949335914105177, + 0.005262818653136492, + 0.0030600321479141712, + 0.0018001566641032696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 5 + ], + "selected_reward": -0.07001025974750519 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static bird's-eye view directly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the water chestnut leaves shaped like hearts and colored saturated emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water crystal clear with a deep blue-green hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves sway slowly and rhythmically with the water, sometimes bobbing or spinning in place?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit the visual style of a Ghibli animation with warm, vibrant, non-glaring colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012542063370347023 + ], + "opposite_logprob_tokens": [ + -4.387542247772217 + ], + "logprob_reward": -0.012542063370347023, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012542063370347023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0925387442111969, + 0.03239668905735016, + 0.18983647227287292, + 0.05197892338037491, + 0.02536696009337902, + 0.08294442296028137, + 0.05823834240436554, + 0.3080902099609375, + 0.12057897448539734, + 0.038030244410037994 + ], + "credit_map": [ + 0.009160034358501434, + 0.012904955074191093, + 0.009835273958742619, + 0.01749301701784134, + 0.015898684039711952, + 0.00856509618461132, + 0.010386794805526733, + 0.008294891566038132, + 0.0035620671696960926, + 0.005104240961372852, + 0.0031077503226697445, + 0.0035883751697838306, + 0.0058044809848070145, + 0.003618097398430109, + 0.003218219615519047, + 0.004393456969410181, + 0.020416246727108955, + 0.013027451001107693, + 0.01735675521194935, + 0.010541480034589767, + 0.019153926521539688, + 0.059069789946079254, + 0.040339358150959015, + 0.009931458160281181, + 0.006471785716712475, + 0.0075617278926074505, + 0.00477822870016098, + 0.005690364167094231, + 0.007005279418081045, + 0.005655656103044748, + 0.007010097615420818, + 0.0078057837672531605, + 0.0032988961320370436, + 0.0033648363314568996, + 0.0020676711574196815, + 0.002194164087995887, + 0.0030642610508948565, + 0.0023297297302633524, + 0.0031125242821872234, + 0.005934877786785364, + 0.015734875574707985, + 0.005770498421043158, + 0.0044991434551775455, + 0.004903995897620916, + 0.027989661321043968, + 0.013152669183909893, + 0.007367010693997145, + 0.0035265653859823942, + 0.008502626791596413, + 0.004822171293199062, + 0.019980764016509056, + 0.004286450333893299, + 0.0034779352135956287, + 0.007230737712234259, + 0.005547458305954933, + 0.004390198737382889, + 0.0816272720694542, + 0.02611154317855835, + 0.1047946885228157, + 0.017526237294077873, + 0.01055934838950634, + 0.028048867359757423, + 0.0230672936886549, + 0.016354968771338463, + 0.02823135256767273, + 0.007552380207926035, + 0.04311738908290863, + 0.008005495183169842, + 0.0068015833385288715, + 0.005169559270143509, + 0.01197111140936613, + 0.00973009318113327, + 0.005991656333208084, + 0.002208957215771079, + 0.014376065693795681, + 0.00243094889447093, + 0.0017338868929073215, + 0.0033479805570095778, + 0.0053691198118031025, + 0.002571630757302046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 8 + ], + "selected_reward": -0.012542063370347023 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static bird's-eye view directly above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the water chestnut leaves shaped like hearts and colored saturated emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water crystal clear with a deep blue-green hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves sway slowly and rhythmically with the water, sometimes bobbing or spinning in place?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit the visual style of a Ghibli animation with warm, vibrant, non-glaring colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10027915984392166 + ], + "opposite_logprob_tokens": [ + -2.3502790927886963 + ], + "logprob_reward": -0.10027915984392166, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10027915984392166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10873058438301086, + 0.027263395488262177, + 0.26126253604888916, + 0.009913847781717777, + 0.016955699771642685, + 0.09316437691450119, + 0.03828183561563492, + 0.2280120849609375, + 0.17935490608215332, + 0.03706063702702522 + ], + "credit_map": [ + 0.01355777122080326, + 0.015933560207486153, + 0.013856579549610615, + 0.017045438289642334, + 0.012617191299796104, + 0.014501472003757954, + 0.010812428779900074, + 0.010406154207885265, + 0.0043671648018062115, + 0.00692033814266324, + 0.0025575184263288975, + 0.003412933088839054, + 0.0030686240643262863, + 0.0025723897852003574, + 0.001845862832851708, + 0.002518566558137536, + 0.01653018221259117, + 0.016193801537156105, + 0.028779704123735428, + 0.01670714281499386, + 0.03680351376533508, + 0.04611744359135628, + 0.08617512136697769, + 0.013955656439065933, + 0.0011096546659246087, + 0.0014497707597911358, + 0.0009205174865201116, + 0.0010792040266096592, + 0.001281381701119244, + 0.001120484434068203, + 0.0015900774160400033, + 0.0013627585722133517, + 0.002147303195670247, + 0.0017441489035263658, + 0.0021326388232409954, + 0.0015214644372463226, + 0.00299050216563046, + 0.0022208476439118385, + 0.0018946505151689053, + 0.0023041467647999525, + 0.014235838316380978, + 0.0066214390099048615, + 0.004910398740321398, + 0.004285318311303854, + 0.03395393490791321, + 0.01679757609963417, + 0.008471892215311527, + 0.0038879818748682737, + 0.005279141943901777, + 0.0033158434089273214, + 0.013208964839577675, + 0.0027111168019473553, + 0.0020270722452551126, + 0.0059500811621546745, + 0.0034428220242261887, + 0.002346797613427043, + 0.05914562568068504, + 0.019612999632954597, + 0.08122118562459946, + 0.013877607882022858, + 0.007732488680630922, + 0.016802968457341194, + 0.017936700955033302, + 0.01168252807110548, + 0.031229006126523018, + 0.011389477178454399, + 0.06381580978631973, + 0.01180080696940422, + 0.011639317497611046, + 0.016000058501958847, + 0.020292935892939568, + 0.013187505304813385, + 0.007539144717156887, + 0.0023099409881979227, + 0.014319009147584438, + 0.002253095619380474, + 0.0019693050999194384, + 0.001944062882103026, + 0.0049160742200911045, + 0.0018100073793902993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 8 + ], + "selected_reward": -0.10027915984392166 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_2/index_tensor([6484, 6484])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_2/index_tensor([6484, 6484])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..beec1f7d2abebe4c37cfa7d14ddd08658766e437 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_2/index_tensor([6484, 6484])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-angle tracking shot follows an ancient four-wheeled carriage as it gallops along a muddy, rain-soaked dirt road. Two sturdy, dark brown horses, their muscles bulging, white, hot air emanating from their nostrils, strain to pull the carriage. The driver, wearing a worn black felt hat and a coarse cloth cape, grips the frayed leather reins tightly, his body swaying back and forth with the carriage's violent jolts, his face a look of focused determination. The carriage's body is constructed of dark, weathered wood, covered in scratches and caked mud. Its massive wooden wheels are rimmed with crude iron hoops. As the carriage races forward, each revolution of the wheels violently crushes the muddy, puddled road, sending large splatters of dark brown mud mixed with grass scraps flying high to the rear and sides. Some of the mud even splashes onto the camera, momentarily blurring the image. Under the diffuse light of the gloomy evening, the entire scene is dominated by dark brown, gray and earthy ochre tones, creating a sense of tension, urgency, primitive movement and a wild atmosphere of breaking free from constraints.", + "seed": 31645, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_2/index_6484_seed_31645_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.025941463466733695, + "vlm_logprob": -0.025941463466733695, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.025941463466733695, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a four-wheeled carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dark brown horses pulling the carriage with visible bulging muscles and steam emanating from their nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a worn black felt hat and a coarse cloth cape while gripping frayed leather reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carriage constructed of dark, weathered wood with massive wooden wheels rimmed with crude iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience moments of blurring caused by mud splashing onto it as the carriage races forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03226509317755699 + ], + "opposite_logprob_tokens": [ + -5.1572651863098145 + ], + "logprob_reward": -0.03226509317755699, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.03226509317755699, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10177356749773026, + 0.025083351880311966, + 0.013128655031323433, + 0.04534114524722099, + 0.021310238167643547, + 0.04924120008945465, + 0.05638464540243149, + 0.5388520956039429, + 0.04243664816021919, + 0.1064484640955925 + ], + "credit_map": [ + 0.016918474808335304, + 0.010032105259597301, + 0.021302947774529457, + 0.006894303485751152, + 0.007063958793878555, + 0.01786762848496437, + 0.006004620343446732, + 0.01568952202796936, + 0.0033226681407541037, + 0.002777159446850419, + 0.004969432950019836, + 0.0018558463780209422, + 0.0031781538855284452, + 0.003706792602315545, + 0.0027868421748280525, + 0.0024864545557647943, + 0.0013154229382053018, + 0.0015659565106034279, + 0.002023026579990983, + 0.001705600880086422, + 0.0014908307930454612, + 0.001469587441533804, + 0.001580347423441708, + 0.001977881882339716, + 0.005105193704366684, + 0.006110879592597485, + 0.004648812115192413, + 0.00499989977106452, + 0.007396019995212555, + 0.0050522200763225555, + 0.003912213258445263, + 0.008115904405713081, + 0.002467030892148614, + 0.0034183249808847904, + 0.0035422227811068296, + 0.002216615481302142, + 0.003312804503366351, + 0.0030823973938822746, + 0.0020601905416697264, + 0.0012106510112062097, + 0.004526864737272263, + 0.004793320316821337, + 0.0072980355471372604, + 0.004545878618955612, + 0.005569924600422382, + 0.00818203017115593, + 0.005852924659848213, + 0.008472221903502941, + 0.0055909538641572, + 0.005170935764908791, + 0.006391486153006554, + 0.005867364816367626, + 0.009339838288724422, + 0.014953574165701866, + 0.00461948849260807, + 0.004451001062989235, + 0.0685308650135994, + 0.019641663879156113, + 0.21043837070465088, + 0.025698866695165634, + 0.027829986065626144, + 0.1543130874633789, + 0.011717982590198517, + 0.02068132907152176, + 0.008781593292951584, + 0.0036478734109550714, + 0.0082553094252944, + 0.004640608560293913, + 0.003274439601227641, + 0.0060423593968153, + 0.004312286619096994, + 0.0034821766894310713, + 0.021702665835618973, + 0.030368326231837273, + 0.00727608660236001, + 0.010483644902706146, + 0.009315607137978077, + 0.010958706960082054, + 0.008775145746767521, + 0.007568282075226307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.03226509317755699 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a four-wheeled carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dark brown horses pulling the carriage with visible bulging muscles and steam emanating from their nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a worn black felt hat and a coarse cloth cape while gripping frayed leather reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carriage constructed of dark, weathered wood with massive wooden wheels rimmed with crude iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience moments of blurring caused by mud splashing onto it as the carriage races forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371446579694748 + ], + "opposite_logprob_tokens": [ + -3.408714532852173 + ], + "logprob_reward": -0.03371446579694748, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371446579694748, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10773096233606339, + 0.06168549507856369, + 0.03731323778629303, + 0.061952248215675354, + 0.053672995418310165, + 0.07305102050304413, + 0.10495886206626892, + 0.32765576243400574, + 0.051610201597213745, + 0.12036919593811035 + ], + "credit_map": [ + 0.017964275553822517, + 0.00753902830183506, + 0.022563880309462547, + 0.006944329012185335, + 0.008183843456208706, + 0.020601581782102585, + 0.009183387272059917, + 0.01475063432008028, + 0.008749240078032017, + 0.008229349739849567, + 0.0080023854970932, + 0.0056447661481797695, + 0.008283385075628757, + 0.01024140790104866, + 0.005962940398603678, + 0.006572014652192593, + 0.0031625975389033556, + 0.005180551670491695, + 0.0038620687555521727, + 0.004631711635738611, + 0.006379352882504463, + 0.005765270907431841, + 0.00402036914601922, + 0.004311314783990383, + 0.00571074103936553, + 0.009178305976092815, + 0.007623028941452503, + 0.006227875594049692, + 0.006706652697175741, + 0.008763473480939865, + 0.005795695353299379, + 0.011946474201977253, + 0.005952378734946251, + 0.0069547388702631, + 0.00441333744674921, + 0.0033026316668838263, + 0.01330387033522129, + 0.011063877493143082, + 0.005191609263420105, + 0.003490550210699439, + 0.006839863955974579, + 0.004850446712225676, + 0.008075108751654625, + 0.010028271935880184, + 0.008336999453604221, + 0.01264934055507183, + 0.00833013653755188, + 0.013940853998064995, + 0.01116656232625246, + 0.011757542379200459, + 0.012935161590576172, + 0.01067379955202341, + 0.015391497872769833, + 0.025052545592188835, + 0.008381081745028496, + 0.009600672870874405, + 0.04650232195854187, + 0.014622359536588192, + 0.12917309999465942, + 0.0164166409522295, + 0.019516758620738983, + 0.07786191999912262, + 0.010513950139284134, + 0.013048711232841015, + 0.008912872523069382, + 0.006006781477481127, + 0.007745824288576841, + 0.006981150712817907, + 0.00367341679520905, + 0.010030112229287624, + 0.0036930155474692583, + 0.004567030351608992, + 0.018772244453430176, + 0.03388162702322006, + 0.007953041233122349, + 0.016172923147678375, + 0.009123963303864002, + 0.012829751707613468, + 0.01266592275351286, + 0.008969726040959358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.03371446579694748 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a four-wheeled carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dark brown horses pulling the carriage with visible bulging muscles and steam emanating from their nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a worn black felt hat and a coarse cloth cape while gripping frayed leather reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carriage constructed of dark, weathered wood with massive wooden wheels rimmed with crude iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience moments of blurring caused by mud splashing onto it as the carriage races forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018196435645222664 + ], + "opposite_logprob_tokens": [ + -4.0181965827941895 + ], + "logprob_reward": -0.018196435645222664, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018196435645222664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10884445905685425, + 0.0330238901078701, + 0.08478346467018127, + 0.11570142209529877, + 0.04603668674826622, + 0.12289407104253769, + 0.09202354401350021, + 0.1397935450077057, + 0.05945231765508652, + 0.19744658470153809 + ], + "credit_map": [ + 0.015488945879042149, + 0.008897664956748486, + 0.023793581873178482, + 0.010125566273927689, + 0.009362975135445595, + 0.016541002318263054, + 0.00888038333505392, + 0.01575433276593685, + 0.00409025838598609, + 0.003076057881116867, + 0.007255499716848135, + 0.0033278779592365026, + 0.0036473064683377743, + 0.004509154241532087, + 0.003631765954196453, + 0.003485969500616193, + 0.00792505033314228, + 0.00957163330167532, + 0.015421008691191673, + 0.010079558938741684, + 0.017591269686818123, + 0.010677719488739967, + 0.007720002438873053, + 0.005797222722321749, + 0.010033654049038887, + 0.017786439508199692, + 0.01058714184910059, + 0.014442996121942997, + 0.017364416271448135, + 0.013716562651097775, + 0.011183557100594044, + 0.020586654543876648, + 0.005155443679541349, + 0.005338950548321009, + 0.006553612649440765, + 0.0037598395720124245, + 0.00767930643633008, + 0.005780868697911501, + 0.0070869498886168, + 0.004681718535721302, + 0.010443526320159435, + 0.01230582781136036, + 0.014887012541294098, + 0.026141781359910965, + 0.014371596276760101, + 0.015344849787652493, + 0.012246464379131794, + 0.01715300790965557, + 0.01296860259026289, + 0.011463209055364132, + 0.014140988700091839, + 0.009361899457871914, + 0.012838786467909813, + 0.01138843223452568, + 0.0062767621129751205, + 0.013584858737885952, + 0.01728280819952488, + 0.012524191290140152, + 0.04181286692619324, + 0.010825816541910172, + 0.015234608203172684, + 0.022556232288479805, + 0.010571309365332127, + 0.008985710330307484, + 0.009640242904424667, + 0.006538189481943846, + 0.007097129710018635, + 0.007012230344116688, + 0.0054440624080598354, + 0.013904276303946972, + 0.003970715217292309, + 0.005845475941896439, + 0.046404942870140076, + 0.03037356212735176, + 0.01987553760409355, + 0.0171017087996006, + 0.01727505959570408, + 0.029411163181066513, + 0.020144591107964516, + 0.01686001755297184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.018196435645222664 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a four-wheeled carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dark brown horses pulling the carriage with visible bulging muscles and steam emanating from their nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a worn black felt hat and a coarse cloth cape while gripping frayed leather reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carriage constructed of dark, weathered wood with massive wooden wheels rimmed with crude iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience moments of blurring caused by mud splashing onto it as the carriage races forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024851649068295956 + ], + "opposite_logprob_tokens": [ + -6.002485275268555 + ], + "logprob_reward": -0.0024851649068295956, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024851649068295956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03900963068008423, + 0.031865403056144714, + 0.019540973007678986, + 0.08788144588470459, + 0.01682467572391033, + 0.21162782609462738, + 0.12982423603534698, + 0.26675790548324585, + 0.12383060157299042, + 0.07283734530210495 + ], + "credit_map": [ + 0.005799906328320503, + 0.004045011941343546, + 0.0045025963336229324, + 0.00393984280526638, + 0.004872222431004047, + 0.0085215512663126, + 0.0032475765328854322, + 0.004080924671143293, + 0.004419241566210985, + 0.0038699747528880835, + 0.005896039307117462, + 0.0034468546509742737, + 0.0035433294251561165, + 0.0049870493821799755, + 0.0032588320318609476, + 0.002444083336740732, + 0.001941226189956069, + 0.0028281749691814184, + 0.0027377833612263203, + 0.0020421009976416826, + 0.0038191776257008314, + 0.002619264181703329, + 0.0014674229314550757, + 0.0020858217030763626, + 0.00860470812767744, + 0.01389323826879263, + 0.009069615043699741, + 0.008734663017094135, + 0.014030996710062027, + 0.011882196180522442, + 0.008874758146703243, + 0.012791270390152931, + 0.00140478799585253, + 0.002154679736122489, + 0.0019387915963307023, + 0.001123698428273201, + 0.0032055696938186884, + 0.002412031404674053, + 0.001722863526083529, + 0.0028622529935091734, + 0.019303234294056892, + 0.02024077996611595, + 0.02666354365646839, + 0.0455479733645916, + 0.026880547404289246, + 0.017980938777327538, + 0.02320440672338009, + 0.031806398183107376, + 0.014264910481870174, + 0.01223845686763525, + 0.017534710466861725, + 0.012224908918142319, + 0.019721031188964844, + 0.03240053355693817, + 0.012885617092251778, + 0.008554067462682724, + 0.03227422758936882, + 0.011063257232308388, + 0.09372417628765106, + 0.013028968125581741, + 0.015943074598908424, + 0.08087977766990662, + 0.009062981233000755, + 0.010781449265778065, + 0.02041333168745041, + 0.01253021601587534, + 0.01729906164109707, + 0.01640205830335617, + 0.00952241662889719, + 0.023652473464608192, + 0.011627579107880592, + 0.012383472174406052, + 0.009494022466242313, + 0.010677590034902096, + 0.007934609428048134, + 0.008225012570619583, + 0.0072014848701655865, + 0.010265211574733257, + 0.010637813247740269, + 0.008401603437960148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.0024851649068295956 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a four-wheeled carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dark brown horses pulling the carriage with visible bulging muscles and steam emanating from their nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a worn black felt hat and a coarse cloth cape while gripping frayed leather reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carriage constructed of dark, weathered wood with massive wooden wheels rimmed with crude iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience moments of blurring caused by mud splashing onto it as the carriage races forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304615780711174 + ], + "opposite_logprob_tokens": [ + -3.168046236038208 + ], + "logprob_reward": -0.04304615780711174, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304615780711174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07138539105653763, + 0.025171879678964615, + 0.026616428047418594, + 0.033472996205091476, + 0.06902395933866501, + 0.08865994960069656, + 0.08440523594617844, + 0.3262491524219513, + 0.06682280451059341, + 0.20819224417209625 + ], + "credit_map": [ + 0.01298497337847948, + 0.005639428272843361, + 0.012648224830627441, + 0.004746105056256056, + 0.006657383870333433, + 0.014415128156542778, + 0.005894969217479229, + 0.008399181067943573, + 0.003699572291225195, + 0.003002479439601302, + 0.00446556368842721, + 0.0030143745243549347, + 0.0025714028161019087, + 0.002874986967071891, + 0.002684131497517228, + 0.0028593679890036583, + 0.002568131545558572, + 0.004156208597123623, + 0.004000148270279169, + 0.002854570047929883, + 0.005084387492388487, + 0.003421361558139324, + 0.0020018871873617172, + 0.0025297333486378193, + 0.005194730591028929, + 0.004482303746044636, + 0.004122599493712187, + 0.0037575566675513983, + 0.0052543203346431255, + 0.003969855140894651, + 0.003490249626338482, + 0.0032013801392167807, + 0.00988755188882351, + 0.00913218129426241, + 0.007538273930549622, + 0.004548316355794668, + 0.01582110859453678, + 0.009397542104125023, + 0.007112982217222452, + 0.005586002953350544, + 0.008782248012721539, + 0.008751689456403255, + 0.011176401749253273, + 0.010049947537481785, + 0.01235914696007967, + 0.014285493642091751, + 0.009598221629858017, + 0.013656800612807274, + 0.010378722101449966, + 0.007328111212700605, + 0.011013742536306381, + 0.009340709075331688, + 0.01056975219398737, + 0.01814912259578705, + 0.006724147591739893, + 0.010900932364165783, + 0.049049679189920425, + 0.013430607505142689, + 0.12961506843566895, + 0.016590414568781853, + 0.01971697248518467, + 0.07528618723154068, + 0.009693576022982597, + 0.012866659089922905, + 0.014332184568047523, + 0.00671116728335619, + 0.012031742371618748, + 0.007245072163641453, + 0.004697374999523163, + 0.011477303691208363, + 0.004858392756432295, + 0.005469561088830233, + 0.04597482830286026, + 0.054710861295461655, + 0.024926360696554184, + 0.023093819618225098, + 0.011899491772055626, + 0.018636353313922882, + 0.017061075195670128, + 0.01188945397734642 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 5 + ], + "selected_reward": -0.04304615780711174 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows an ancient four-wheeled carriage as it gallops along a muddy, rain-soaked dirt road. Two sturdy, dark brown horses, their muscles bulging, white, hot air emanating from their nostrils, strain to pull the carriage. The driver, wearing a worn black felt hat and a coarse cloth cape, grips the frayed leather reins tightly, his body swaying back and forth with the carriage's violent jolts, his face a look of focused determination. The carriage's body is constructed of dark, weathered wood, covered in scratches and caked mud. Its massive wooden wheels are rimmed with crude iron hoops. As the carriage races forward, each revolution of the wheels violently crushes the muddy, puddled road, sending large splatters of dark brown mud mixed with grass scraps flying high to the rear and sides. Some of the mud even splashes onto the camera, momentarily blurring the image. Under the diffuse light of the gloomy evening, the entire scene is dominated by dark brown, gray and earthy ochre tones, creating a sense of tension, urgency, primitive movement and a wild atmosphere of breaking free from constraints.", + "seed": 31646, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_2/index_6484_seed_31646_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.019583229999989273, + "vlm_logprob": -0.019583229999989273, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.019583229999989273, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a four-wheeled carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dark brown horses pulling the carriage with visible bulging muscles and steam emanating from their nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a worn black felt hat and a coarse cloth cape while gripping frayed leather reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carriage constructed of dark, weathered wood with massive wooden wheels rimmed with crude iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience moments of blurring caused by mud splashing onto it as the carriage races forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04954550787806511 + ], + "opposite_logprob_tokens": [ + -3.6745455265045166 + ], + "logprob_reward": -0.04954550787806511, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.04954550787806511, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1970900446176529, + 0.1737372875213623, + 0.06676987558603287, + 0.018497943878173828, + 0.0810040608048439, + 0.06468886137008667, + 0.03193357214331627, + 0.21723200380802155, + 0.0580088347196579, + 0.09103750437498093 + ], + "credit_map": [ + 0.024151792749762535, + 0.02189159020781517, + 0.015183113515377045, + 0.017594723030924797, + 0.03283204510807991, + 0.03634846955537796, + 0.025133023038506508, + 0.023955300450325012, + 0.011928333900868893, + 0.07202132046222687, + 0.005681532435119152, + 0.04228278249502182, + 0.017841262742877007, + 0.008175699040293694, + 0.009697469882667065, + 0.006108877249062061, + 0.005765647627413273, + 0.008089509792625904, + 0.006149997003376484, + 0.0055303433910012245, + 0.006361307576298714, + 0.009517529979348183, + 0.011730247177183628, + 0.013625293970108032, + 0.0018599479226395488, + 0.0017136860406026244, + 0.0012802876299247146, + 0.0014140581479296088, + 0.002228970406576991, + 0.0022191149182617664, + 0.0028720912523567677, + 0.00490978779271245, + 0.006441229954361916, + 0.011742017231881618, + 0.006605822127312422, + 0.007300578523427248, + 0.014205203391611576, + 0.01376507431268692, + 0.01239852886646986, + 0.008545608259737492, + 0.006211106665432453, + 0.0059810299426317215, + 0.008319640532135963, + 0.014198163524270058, + 0.004089749418199062, + 0.009014187380671501, + 0.009767111390829086, + 0.00710787670686841, + 0.00497336033731699, + 0.005282977130264044, + 0.006214423105120659, + 0.0040528136305511, + 0.0037692468613386154, + 0.0027822288684546947, + 0.002429028507322073, + 0.002429493935778737, + 0.06793322414159775, + 0.02031908556818962, + 0.006743522826582193, + 0.00622784486040473, + 0.0219422560185194, + 0.07501549273729324, + 0.011037942953407764, + 0.008012617006897926, + 0.005758573301136494, + 0.014232346788048744, + 0.004451450891792774, + 0.005666698329150677, + 0.00666614668443799, + 0.007736163679510355, + 0.006922557018697262, + 0.006574899423867464, + 0.009805425070226192, + 0.02513442374765873, + 0.009806294925510883, + 0.009268077090382576, + 0.009053748100996017, + 0.015803677961230278, + 0.006054284051060677, + 0.006111572030931711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.04954550787806511 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a four-wheeled carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dark brown horses pulling the carriage with visible bulging muscles and steam emanating from their nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a worn black felt hat and a coarse cloth cape while gripping frayed leather reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carriage constructed of dark, weathered wood with massive wooden wheels rimmed with crude iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience moments of blurring caused by mud splashing onto it as the carriage races forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014221038669347763 + ], + "opposite_logprob_tokens": [ + -4.26422119140625 + ], + "logprob_reward": -0.014221038669347763, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014221038669347763, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11029611527919769, + 0.24066497385501862, + 0.023354992270469666, + 0.021560246124863625, + 0.12170044332742691, + 0.058118101209402084, + 0.03475990146398544, + 0.2146417647600174, + 0.08195662498474121, + 0.0929468497633934 + ], + "credit_map": [ + 0.011994860135018826, + 0.015223155729472637, + 0.010715489275753498, + 0.009058313444256783, + 0.016702579334378242, + 0.01825140230357647, + 0.01302147377282381, + 0.0153288459405303, + 0.01713673211634159, + 0.09817397594451904, + 0.007824213244020939, + 0.06204809993505478, + 0.022293873131275177, + 0.010784556157886982, + 0.013248132541775703, + 0.00915538053959608, + 0.0030046687461435795, + 0.004080915357917547, + 0.0020736511796712875, + 0.003073509084060788, + 0.002812133403494954, + 0.0031226975843310356, + 0.002992645837366581, + 0.0021947729401290417, + 0.00209523132070899, + 0.002257022773846984, + 0.0022859666496515274, + 0.0024537097197026014, + 0.003426564624533057, + 0.002959252567961812, + 0.002641041297465563, + 0.0034414571709930897, + 0.008084030821919441, + 0.015450915321707726, + 0.008955352008342743, + 0.012350411154329777, + 0.017659615725278854, + 0.024084800854325294, + 0.018143625929951668, + 0.016971692442893982, + 0.0064964499324560165, + 0.005068354308605194, + 0.004575453232973814, + 0.008079715073108673, + 0.012263186275959015, + 0.008868809789419174, + 0.007536560297012329, + 0.005229571368545294, + 0.005403686780482531, + 0.0046394746750593185, + 0.006797765847295523, + 0.005113320890814066, + 0.002824608003720641, + 0.0035027035046368837, + 0.0038319516461342573, + 0.002646390115842223, + 0.04645823687314987, + 0.029126938432455063, + 0.008881216868758202, + 0.012138282880187035, + 0.04012982174754143, + 0.04391110688447952, + 0.022994335740804672, + 0.011001833714544773, + 0.009929804131388664, + 0.0186154805123806, + 0.00728302588686347, + 0.0066971248015761375, + 0.01009109802544117, + 0.011494692414999008, + 0.008399467915296555, + 0.009445932693779469, + 0.009651479311287403, + 0.021904122084379196, + 0.012718018144369125, + 0.007380951661616564, + 0.010391908697783947, + 0.017459578812122345, + 0.006919987965375185, + 0.006520809140056372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 4 + ], + "selected_reward": -0.014221038669347763 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a four-wheeled carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dark brown horses pulling the carriage with visible bulging muscles and steam emanating from their nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a worn black felt hat and a coarse cloth cape while gripping frayed leather reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carriage constructed of dark, weathered wood with massive wooden wheels rimmed with crude iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience moments of blurring caused by mud splashing onto it as the carriage races forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011085470207035542 + ], + "opposite_logprob_tokens": [ + -4.511085510253906 + ], + "logprob_reward": -0.011085470207035542, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011085470207035542, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10070433467626572, + 0.0835181400179863, + 0.036929577589035034, + 0.03868832439184189, + 0.19385874271392822, + 0.06878211349248886, + 0.10021566599607468, + 0.13265126943588257, + 0.09534645080566406, + 0.14930540323257446 + ], + "credit_map": [ + 0.014727551490068436, + 0.012317687273025513, + 0.011545144952833652, + 0.011280802078545094, + 0.01665428839623928, + 0.010576109401881695, + 0.010841996408998966, + 0.01276075467467308, + 0.007625051308423281, + 0.021421346813440323, + 0.006183835212141275, + 0.01612086221575737, + 0.010260745882987976, + 0.006061003543436527, + 0.008356199599802494, + 0.007489101495593786, + 0.0031190761364996433, + 0.008484198711812496, + 0.003091110149398446, + 0.005385148338973522, + 0.0036278842017054558, + 0.0045604342594742775, + 0.0047315857373178005, + 0.003930141218006611, + 0.0041091241873800755, + 0.004481094889342785, + 0.0035391857381910086, + 0.0025575733743608, + 0.0041518318466842175, + 0.004611562471836805, + 0.004576796665787697, + 0.01066115777939558, + 0.01458557415753603, + 0.02090735174715519, + 0.014859277755022049, + 0.024129988625645638, + 0.026867106556892395, + 0.03872154653072357, + 0.02912740968167782, + 0.024660484865307808, + 0.007222490385174751, + 0.0050243837758898735, + 0.005466553382575512, + 0.008193407207727432, + 0.007917443290352821, + 0.01713515818119049, + 0.0088755302131176, + 0.008947145193815231, + 0.014965453185141087, + 0.01147508155554533, + 0.010931932367384434, + 0.01302075944840908, + 0.01162684801965952, + 0.00863858126103878, + 0.017116764560341835, + 0.012440243735909462, + 0.018626153469085693, + 0.022766748443245888, + 0.00810418464243412, + 0.010383771732449532, + 0.031196394935250282, + 0.015378031879663467, + 0.017234981060028076, + 0.00896099116653204, + 0.009789341129362583, + 0.01903589442372322, + 0.009184658527374268, + 0.009427111595869064, + 0.01125215645879507, + 0.013598233461380005, + 0.011227956973016262, + 0.01183109637349844, + 0.01392005942761898, + 0.033919159322977066, + 0.01921258307993412, + 0.01352080050855875, + 0.015627557411789894, + 0.02994074672460556, + 0.010978453792631626, + 0.012186050415039062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 7 + ], + "selected_reward": -0.011085470207035542 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a four-wheeled carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dark brown horses pulling the carriage with visible bulging muscles and steam emanating from their nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a worn black felt hat and a coarse cloth cape while gripping frayed leather reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carriage constructed of dark, weathered wood with massive wooden wheels rimmed with crude iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience moments of blurring caused by mud splashing onto it as the carriage races forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024877809919416904 + ], + "opposite_logprob_tokens": [ + -6.002487659454346 + ], + "logprob_reward": -0.0024877809919416904, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024877809919416904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039666712284088135, + 0.047027215361595154, + 0.019219957292079926, + 0.011749227531254292, + 0.02565157413482666, + 0.04425327479839325, + 0.08606014400720596, + 0.39808541536331177, + 0.24724891781806946, + 0.08103758096694946 + ], + "credit_map": [ + 0.0050324806943535805, + 0.003932728432118893, + 0.0034021451137959957, + 0.0052867187187075615, + 0.005756168160587549, + 0.004318026825785637, + 0.004988552536815405, + 0.006949889473617077, + 0.0037639399524778128, + 0.01728953793644905, + 0.002524539129808545, + 0.004677088465541601, + 0.009488512761890888, + 0.003006458282470703, + 0.0032200089190155268, + 0.0030571273528039455, + 0.0019313845550641418, + 0.0018285431433469057, + 0.0019609553273767233, + 0.0015332590555772185, + 0.0025585400871932507, + 0.0027389556635171175, + 0.003158065490424633, + 0.003510252805426717, + 0.001165520865470171, + 0.0015755662461742759, + 0.0010501391952857375, + 0.0013962299562990665, + 0.0009902809979394078, + 0.0011151072103530169, + 0.0015449331840500236, + 0.0029114503413438797, + 0.0026339564938098192, + 0.0028464144561439753, + 0.0025473337154835463, + 0.0026767689269036055, + 0.0019481955096125603, + 0.004178247880190611, + 0.005550832487642765, + 0.0032698246650397778, + 0.005403093062341213, + 0.004306298680603504, + 0.006359564606100321, + 0.00846222136169672, + 0.0034990524873137474, + 0.005575796123594046, + 0.0058598448522388935, + 0.0047874064184725285, + 0.013048099353909492, + 0.009114811196923256, + 0.01284065842628479, + 0.011753096245229244, + 0.008477774448692799, + 0.006813796237111092, + 0.013697703368961811, + 0.010314205661416054, + 0.11599796265363693, + 0.03754345327615738, + 0.012266883626580238, + 0.011471536941826344, + 0.0447935089468956, + 0.1378927081823349, + 0.024051442742347717, + 0.01406791154295206, + 0.020162351429462433, + 0.02963516116142273, + 0.025115877389907837, + 0.027999546378850937, + 0.026624545454978943, + 0.0335765965282917, + 0.046720366925001144, + 0.037414491176605225, + 0.007647986523807049, + 0.013837547041475773, + 0.01620495691895485, + 0.009408396668732166, + 0.009282308630645275, + 0.011826476082205772, + 0.0076213134452700615, + 0.005208594724535942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.0024877809919416904 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a four-wheeled carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dark brown horses pulling the carriage with visible bulging muscles and steam emanating from their nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a worn black felt hat and a coarse cloth cape while gripping frayed leather reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carriage constructed of dark, weathered wood with massive wooden wheels rimmed with crude iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience moments of blurring caused by mud splashing onto it as the carriage races forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057635225355625 + ], + "opposite_logprob_tokens": [ + -3.895576238632202 + ], + "logprob_reward": -0.02057635225355625, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057635225355625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051712047308683395, + 0.06543219834566116, + 0.0171709805727005, + 0.03498296067118645, + 0.14264823496341705, + 0.05069566145539284, + 0.06986761093139648, + 0.048152100294828415, + 0.25242576003074646, + 0.26691246032714844 + ], + "credit_map": [ + 0.006544508505612612, + 0.005574382375925779, + 0.004158023279160261, + 0.003982840105891228, + 0.009478979744017124, + 0.007828564383089542, + 0.00460710097104311, + 0.009537652134895325, + 0.005890603177249432, + 0.009483556263148785, + 0.005933959968388081, + 0.0087197907269001, + 0.013146592304110527, + 0.005823250859975815, + 0.009565417654812336, + 0.006869026459753513, + 0.0023276126012206078, + 0.002736651571467519, + 0.0018741022795438766, + 0.0019768907222896814, + 0.0023362028878182173, + 0.0019864097703248262, + 0.0020196852274239063, + 0.0019134256290271878, + 0.0033881578128784895, + 0.0037490534596145153, + 0.002902558771893382, + 0.002861824817955494, + 0.0034563359804451466, + 0.003694049781188369, + 0.004205052740871906, + 0.010725927539169788, + 0.009463983587920666, + 0.017327765002846718, + 0.009314061142504215, + 0.011846247129142284, + 0.02377607300877571, + 0.027300193905830383, + 0.022472485899925232, + 0.02114742249250412, + 0.005441684741526842, + 0.007515248842537403, + 0.005603752098977566, + 0.008358309045433998, + 0.006651601754128933, + 0.006870081182569265, + 0.00625609653070569, + 0.003998885862529278, + 0.010564655996859074, + 0.009309832938015461, + 0.009589689783751965, + 0.009392189793288708, + 0.006557921413332224, + 0.007352486718446016, + 0.009335975162684917, + 0.0077648586593568325, + 0.00695423036813736, + 0.008296741172671318, + 0.0037288214080035686, + 0.0032542869448661804, + 0.00742545397952199, + 0.009670337662100792, + 0.004971339367330074, + 0.00385088799521327, + 0.026994744315743446, + 0.07278136909008026, + 0.016934361308813095, + 0.017642851918935776, + 0.03550251945853233, + 0.040850378572940826, + 0.01851038821041584, + 0.02320915088057518, + 0.031980354338884354, + 0.06622203439474106, + 0.04076400026679039, + 0.03506556898355484, + 0.023150121793150902, + 0.034367673099040985, + 0.020036451518535614, + 0.01532625500112772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.02057635225355625 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows an ancient four-wheeled carriage as it gallops along a muddy, rain-soaked dirt road. Two sturdy, dark brown horses, their muscles bulging, white, hot air emanating from their nostrils, strain to pull the carriage. The driver, wearing a worn black felt hat and a coarse cloth cape, grips the frayed leather reins tightly, his body swaying back and forth with the carriage's violent jolts, his face a look of focused determination. The carriage's body is constructed of dark, weathered wood, covered in scratches and caked mud. Its massive wooden wheels are rimmed with crude iron hoops. As the carriage races forward, each revolution of the wheels violently crushes the muddy, puddled road, sending large splatters of dark brown mud mixed with grass scraps flying high to the rear and sides. Some of the mud even splashes onto the camera, momentarily blurring the image. Under the diffuse light of the gloomy evening, the entire scene is dominated by dark brown, gray and earthy ochre tones, creating a sense of tension, urgency, primitive movement and a wild atmosphere of breaking free from constraints.", + "seed": 31647, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_2/index_6484_seed_31647_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.023384066065773368, + "vlm_logprob": -0.023384066065773368, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.023384066065773368, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a four-wheeled carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dark brown horses pulling the carriage with visible bulging muscles and steam emanating from their nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a worn black felt hat and a coarse cloth cape while gripping frayed leather reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carriage constructed of dark, weathered wood with massive wooden wheels rimmed with crude iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience moments of blurring caused by mud splashing onto it as the carriage races forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03576267883181572 + ], + "opposite_logprob_tokens": [ + -4.410762786865234 + ], + "logprob_reward": -0.03576267883181572, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.03576267883181572, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7162880301475525, + 0.096373051404953, + 0.012159937992691994, + 0.009416879154741764, + 0.017085984349250793, + 0.01772911846637726, + 0.019019315019249916, + 0.027863143011927605, + 0.060264937579631805, + 0.02379954420030117 + ], + "credit_map": [ + 0.09044494479894638, + 0.011428801342844963, + 0.31362995505332947, + 0.032633908092975616, + 0.039509933441877365, + 0.028584185987710953, + 0.028313586488366127, + 0.17174269258975983, + 0.011551975272595882, + 0.008738520555198193, + 0.008464206010103226, + 0.009503506124019623, + 0.009823378175497055, + 0.02599499560892582, + 0.009627499617636204, + 0.012668972834944725, + 0.0010424306383356452, + 0.0008147534681484103, + 0.0011347790714353323, + 0.0009092213003896177, + 0.0022872372064739466, + 0.002736488590016961, + 0.0014782402431592345, + 0.0017567884642630816, + 0.0010576614877209067, + 0.0007163869449868798, + 0.0007851127884350717, + 0.0005141449510119855, + 0.0020108381286263466, + 0.002625466091558337, + 0.0007162714609876275, + 0.0009909974178299308, + 0.002119134645909071, + 0.0013703893637284636, + 0.0017258322332054377, + 0.001450669253244996, + 0.0035633454099297523, + 0.002604912733659148, + 0.0014900038950145245, + 0.002761696930974722, + 0.0017057915683835745, + 0.003083430463448167, + 0.001821638084948063, + 0.0022029562387615442, + 0.0021711126901209354, + 0.001976383849978447, + 0.001955286832526326, + 0.0028125180397182703, + 0.0020585369784384966, + 0.0024110758677124977, + 0.0030084813479334116, + 0.002717725932598114, + 0.0017022138927131891, + 0.002884059213101864, + 0.00242311111651361, + 0.0018141110194846988, + 0.0037356961984187365, + 0.002698647091165185, + 0.004101232159882784, + 0.003176687052473426, + 0.0029500839300453663, + 0.004503079690039158, + 0.0033106429036706686, + 0.003387073054909706, + 0.00607100548222661, + 0.0058211530558764935, + 0.006666815374046564, + 0.005275530740618706, + 0.010259476490318775, + 0.008556998334825039, + 0.00711579155176878, + 0.0104981679469347, + 0.004611389245837927, + 0.002817639848217368, + 0.0035176046658307314, + 0.003226514672860503, + 0.002908907365053892, + 0.002553545171394944, + 0.0018859560368582606, + 0.0022779887076467276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.03576267883181572 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a four-wheeled carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dark brown horses pulling the carriage with visible bulging muscles and steam emanating from their nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a worn black felt hat and a coarse cloth cape while gripping frayed leather reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carriage constructed of dark, weathered wood with massive wooden wheels rimmed with crude iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience moments of blurring caused by mud splashing onto it as the carriage races forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026370925828814507 + ], + "opposite_logprob_tokens": [ + -3.6513710021972656 + ], + "logprob_reward": -0.026370925828814507, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026370925828814507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34385618567466736, + 0.14949645102024078, + 0.07474349439144135, + 0.06192360445857048, + 0.0433327741920948, + 0.04693009704351425, + 0.028446199372410774, + 0.06035551428794861, + 0.14427658915519714, + 0.04663911461830139 + ], + "credit_map": [ + 0.10023730248212814, + 0.01231731940060854, + 0.057670995593070984, + 0.017369946464896202, + 0.015833811834454536, + 0.02355172298848629, + 0.015574101358652115, + 0.10130096971988678, + 0.019916966557502747, + 0.011787084862589836, + 0.020273307338356972, + 0.025425860658288002, + 0.012612088583409786, + 0.02483983151614666, + 0.019354715943336487, + 0.015286584384739399, + 0.007116149179637432, + 0.008201824501156807, + 0.00756592582911253, + 0.005135813727974892, + 0.008519052527844906, + 0.01670246757566929, + 0.008200251497328281, + 0.013302013278007507, + 0.005290699191391468, + 0.005290968809276819, + 0.0075659179128706455, + 0.0029691862873733044, + 0.019309336319565773, + 0.007726969197392464, + 0.005716622807085514, + 0.008053908124566078, + 0.004351010080426931, + 0.004727629013359547, + 0.006337485276162624, + 0.004145967774093151, + 0.008425911888480186, + 0.004903367254883051, + 0.004537719301879406, + 0.005903684534132481, + 0.0038401884958148003, + 0.006893078796565533, + 0.006455675698816776, + 0.005339027848094702, + 0.008230662904679775, + 0.005298500414937735, + 0.004485079552978277, + 0.00638788053765893, + 0.003317401045933366, + 0.004105074796825647, + 0.003464405657723546, + 0.005191814620047808, + 0.0030149242375046015, + 0.0030256137251853943, + 0.003861438250169158, + 0.0024655265733599663, + 0.008536303415894508, + 0.007188798859715462, + 0.007581894285976887, + 0.007714258041232824, + 0.007015350740402937, + 0.008922980166971684, + 0.007555752992630005, + 0.0058401767164468765, + 0.018761413171887398, + 0.021568890661001205, + 0.022997410967946053, + 0.012655061669647694, + 0.016582049429416656, + 0.017072729766368866, + 0.014453170821070671, + 0.020185865461826324, + 0.004298410378396511, + 0.005440276116132736, + 0.006186007522046566, + 0.006021062843501568, + 0.004713786765933037, + 0.007861804217100143, + 0.005067404359579086, + 0.007050362415611744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.026370925828814507 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a four-wheeled carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dark brown horses pulling the carriage with visible bulging muscles and steam emanating from their nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a worn black felt hat and a coarse cloth cape while gripping frayed leather reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carriage constructed of dark, weathered wood with massive wooden wheels rimmed with crude iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience moments of blurring caused by mud splashing onto it as the carriage races forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014212576672434807 + ], + "opposite_logprob_tokens": [ + -4.264212608337402 + ], + "logprob_reward": -0.014212576672434807, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014212576672434807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3563419580459595, + 0.05613352358341217, + 0.05257643386721611, + 0.025888776406645775, + 0.06426899880170822, + 0.05773046985268593, + 0.08974599093198776, + 0.04985342174768448, + 0.14502881467342377, + 0.10243154317140579 + ], + "credit_map": [ + 0.03365209326148033, + 0.007046508602797985, + 0.12341447174549103, + 0.013264129869639874, + 0.014303234405815601, + 0.019740844145417213, + 0.013016396202147007, + 0.13190431892871857, + 0.00757446326315403, + 0.007292082067579031, + 0.005583316087722778, + 0.007332200184464455, + 0.006331705488264561, + 0.009118134155869484, + 0.005990905221551657, + 0.006910725496709347, + 0.005805667024105787, + 0.005426504649221897, + 0.0064608692191541195, + 0.0034529748372733593, + 0.005189469084143639, + 0.011724681593477726, + 0.005914685782045126, + 0.008601585403084755, + 0.002279704436659813, + 0.0012873399537056684, + 0.001854635658673942, + 0.001319927629083395, + 0.005483763292431831, + 0.008673377335071564, + 0.0016341162845492363, + 0.0033559163566678762, + 0.006065574008971453, + 0.004073409363627434, + 0.009263276122510433, + 0.005417705047875643, + 0.013222741894423962, + 0.007552755996584892, + 0.0068700737319886684, + 0.011803469620645046, + 0.0066105942241847515, + 0.008046865463256836, + 0.004650862421840429, + 0.005230143666267395, + 0.008791966363787651, + 0.007640285417437553, + 0.00566995469853282, + 0.011089801788330078, + 0.008638129569590092, + 0.011371527798473835, + 0.013885063119232655, + 0.011951913125813007, + 0.01194128766655922, + 0.011755550280213356, + 0.01189989410340786, + 0.008302642032504082, + 0.007145150564610958, + 0.005574156064540148, + 0.005847245920449495, + 0.005110220052301884, + 0.006342095788568258, + 0.007429785560816526, + 0.007360885385423899, + 0.00504388427361846, + 0.0194228608161211, + 0.019466450437903404, + 0.022292815148830414, + 0.014271700754761696, + 0.015899043530225754, + 0.016756299883127213, + 0.016344832256436348, + 0.02057482674717903, + 0.025943314656615257, + 0.013509396463632584, + 0.013178057037293911, + 0.009359549731016159, + 0.01067271362990141, + 0.01095531415194273, + 0.00876834150403738, + 0.010044860653579235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.014212576672434807 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a four-wheeled carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dark brown horses pulling the carriage with visible bulging muscles and steam emanating from their nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a worn black felt hat and a coarse cloth cape while gripping frayed leather reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carriage constructed of dark, weathered wood with massive wooden wheels rimmed with crude iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience moments of blurring caused by mud splashing onto it as the carriage races forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002486948622390628 + ], + "opposite_logprob_tokens": [ + -6.0024871826171875 + ], + "logprob_reward": -0.002486948622390628, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002486948622390628, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2878866195678711, + 0.10355950146913528, + 0.05065786838531494, + 0.03357454016804695, + 0.03115510568022728, + 0.02790174074470997, + 0.08246804773807526, + 0.11222659051418304, + 0.1805732101202011, + 0.08999678492546082 + ], + "credit_map": [ + 0.021811218932271004, + 0.0047491323202848434, + 0.11772476136684418, + 0.014307450503110886, + 0.013236073777079582, + 0.014170605689287186, + 0.011788851581513882, + 0.090098537504673, + 0.012860714457929134, + 0.011460786685347557, + 0.007982881739735603, + 0.011797814629971981, + 0.010701227933168411, + 0.02447556145489216, + 0.00973481684923172, + 0.014545688405632973, + 0.00553167937323451, + 0.005973763298243284, + 0.0060224393382668495, + 0.004343770444393158, + 0.004980279132723808, + 0.0074989208951592445, + 0.005766937974840403, + 0.010540079325437546, + 0.002704763785004616, + 0.002676171250641346, + 0.0027850705664604902, + 0.0017186857294291258, + 0.005031001754105091, + 0.009968390688300133, + 0.0039058513939380646, + 0.004784604534506798, + 0.003475619712844491, + 0.002606697380542755, + 0.0035047680139541626, + 0.0025967289693653584, + 0.0067181154154241085, + 0.003554978407919407, + 0.003757889149710536, + 0.004940307233482599, + 0.0032095839269459248, + 0.0039703757502138615, + 0.0028508605901151896, + 0.002739593619480729, + 0.0035089123994112015, + 0.0034087966196238995, + 0.0028293016366660595, + 0.005384315736591816, + 0.007423238828778267, + 0.01021790411323309, + 0.011708428151905537, + 0.012377322651445866, + 0.009725227020680904, + 0.010250951163470745, + 0.013028453104197979, + 0.007736523635685444, + 0.016522599384188652, + 0.012873072177171707, + 0.015734070912003517, + 0.012040984816849232, + 0.010016193613409996, + 0.012650590389966965, + 0.01948522962629795, + 0.012903857044875622, + 0.023645849898457527, + 0.02463000826537609, + 0.026576995849609375, + 0.016539840027689934, + 0.02302544005215168, + 0.01938769780099392, + 0.018447531387209892, + 0.028319843113422394, + 0.019233308732509613, + 0.011544962413609028, + 0.011300014331936836, + 0.009125109761953354, + 0.008432330563664436, + 0.008969311602413654, + 0.009064587764441967, + 0.012327160686254501 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.002486948622390628 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a four-wheeled carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dark brown horses pulling the carriage with visible bulging muscles and steam emanating from their nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a worn black felt hat and a coarse cloth cape while gripping frayed leather reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carriage constructed of dark, weathered wood with massive wooden wheels rimmed with crude iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience moments of blurring caused by mud splashing onto it as the carriage races forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808720037341118 + ], + "opposite_logprob_tokens": [ + -3.2880871295928955 + ], + "logprob_reward": -0.03808720037341118, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808720037341118, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4848144054412842, + 0.06635700911283493, + 0.06720814853906631, + 0.016956135630607605, + 0.021526537835597992, + 0.07217634469270706, + 0.0682789757847786, + 0.07178567349910736, + 0.06767313182353973, + 0.06322366744279861 + ], + "credit_map": [ + 0.1138330027461052, + 0.01246984675526619, + 0.18907901644706726, + 0.027105772867798805, + 0.02847502939403057, + 0.021555915474891663, + 0.023136092349886894, + 0.06915969401597977, + 0.0062708319164812565, + 0.004482432268559933, + 0.005810856353491545, + 0.008694591000676155, + 0.006084619089961052, + 0.016552116721868515, + 0.00852695107460022, + 0.009934600442647934, + 0.006883890368044376, + 0.0070575326681137085, + 0.008461222052574158, + 0.006504977587610483, + 0.00742374500259757, + 0.015981486067175865, + 0.005995098035782576, + 0.008900202810764313, + 0.001997691113501787, + 0.0015402913559228182, + 0.0024096891283988953, + 0.001070822705514729, + 0.003481522900983691, + 0.0036236303858458996, + 0.0014505707658827305, + 0.0013819163432344794, + 0.0022208828013390303, + 0.002327526919543743, + 0.0030353933107107878, + 0.0020431815646588802, + 0.004212609026581049, + 0.002389061963185668, + 0.0025578250642865896, + 0.002740056486800313, + 0.007682148367166519, + 0.011029861867427826, + 0.006560939364135265, + 0.006530903745442629, + 0.01011376641690731, + 0.009488753974437714, + 0.0059747821651399136, + 0.014795190654695034, + 0.00779755087569356, + 0.00726809399202466, + 0.010284966789186, + 0.010713575407862663, + 0.007335742469877005, + 0.008584504947066307, + 0.007863188162446022, + 0.00843135453760624, + 0.011554940603673458, + 0.0061674825847148895, + 0.008222841657698154, + 0.007872785441577435, + 0.009312962181866169, + 0.009428960271179676, + 0.011595705524086952, + 0.007629991974681616, + 0.006374059710651636, + 0.009825961664319038, + 0.009969087317585945, + 0.006960973609238863, + 0.009532547555863857, + 0.007764165755361319, + 0.00893581472337246, + 0.008310521952807903, + 0.008629631251096725, + 0.006564201787114143, + 0.008163857273757458, + 0.007642258889973164, + 0.005683090537786484, + 0.0067572989501059055, + 0.007150223944336176, + 0.01263310294598341 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.03808720037341118 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows an ancient four-wheeled carriage as it gallops along a muddy, rain-soaked dirt road. Two sturdy, dark brown horses, their muscles bulging, white, hot air emanating from their nostrils, strain to pull the carriage. The driver, wearing a worn black felt hat and a coarse cloth cape, grips the frayed leather reins tightly, his body swaying back and forth with the carriage's violent jolts, his face a look of focused determination. The carriage's body is constructed of dark, weathered wood, covered in scratches and caked mud. Its massive wooden wheels are rimmed with crude iron hoops. As the carriage races forward, each revolution of the wheels violently crushes the muddy, puddled road, sending large splatters of dark brown mud mixed with grass scraps flying high to the rear and sides. Some of the mud even splashes onto the camera, momentarily blurring the image. Under the diffuse light of the gloomy evening, the entire scene is dominated by dark brown, gray and earthy ochre tones, creating a sense of tension, urgency, primitive movement and a wild atmosphere of breaking free from constraints.", + "seed": 31648, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_2/index_6484_seed_31648_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.018657980812713505, + "vlm_logprob": -0.018657980812713505, + "vlm_yes_no_margin": 4.624999952316284, + "vlm_token_credit": -0.018657980812713505, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a four-wheeled carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dark brown horses pulling the carriage with visible bulging muscles and steam emanating from their nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a worn black felt hat and a coarse cloth cape while gripping frayed leather reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carriage constructed of dark, weathered wood with massive wooden wheels rimmed with crude iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience moments of blurring caused by mud splashing onto it as the carriage races forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038365673273801804 + ], + "opposite_logprob_tokens": [ + -4.788365840911865 + ], + "logprob_reward": -0.038365673273801804, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.038365673273801804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17572394013404846, + 0.20870810747146606, + 0.06814558804035187, + 0.08496289700269699, + 0.08388292044401169, + 0.0711265429854393, + 0.049138281494379044, + 0.10465829074382782, + 0.08531694114208221, + 0.06833653897047043 + ], + "credit_map": [ + 0.01687472313642502, + 0.018232135102152824, + 0.02463465742766857, + 0.02242974191904068, + 0.027400905266404152, + 0.03315892815589905, + 0.014100811444222927, + 0.018892038613557816, + 0.030373718589544296, + 0.021213769912719727, + 0.030487826094031334, + 0.01975284516811371, + 0.03558178246021271, + 0.041553281247615814, + 0.015658093616366386, + 0.014086795970797539, + 0.007797266822308302, + 0.008140072226524353, + 0.011318834498524666, + 0.004635106772184372, + 0.013244252651929855, + 0.007922791875898838, + 0.00897213350981474, + 0.0061151268891990185, + 0.012063964270055294, + 0.010984725318849087, + 0.008875531144440174, + 0.006558738648891449, + 0.010050555691123009, + 0.024543097242712975, + 0.0063428026624023914, + 0.005543478764593601, + 0.01056598499417305, + 0.01230272650718689, + 0.009554966352880001, + 0.009057569317519665, + 0.00931358989328146, + 0.023809313774108887, + 0.004621414467692375, + 0.004657345823943615, + 0.008771078661084175, + 0.008980165235698223, + 0.007640765979886055, + 0.008509302511811256, + 0.009116463363170624, + 0.01273611094802618, + 0.005960995331406593, + 0.009411661885678768, + 0.003475002245977521, + 0.005580225959420204, + 0.006467740051448345, + 0.011972717940807343, + 0.004557895008474588, + 0.005536479409784079, + 0.004807421006262302, + 0.00674080103635788, + 0.013554454781115055, + 0.011126823723316193, + 0.01082853227853775, + 0.015593483112752438, + 0.013113215565681458, + 0.01133564580231905, + 0.015810370445251465, + 0.013295762240886688, + 0.012797115370631218, + 0.014439848251640797, + 0.007963915355503559, + 0.016476474702358246, + 0.007558824494481087, + 0.005911888089030981, + 0.011678679846227169, + 0.008490188047289848, + 0.009920970536768436, + 0.0070773884654045105, + 0.0077657210640609264, + 0.011757179163396358, + 0.006357879843562841, + 0.007516268640756607, + 0.007991554215550423, + 0.009949573315680027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.038365673273801804 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a four-wheeled carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dark brown horses pulling the carriage with visible bulging muscles and steam emanating from their nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a worn black felt hat and a coarse cloth cape while gripping frayed leather reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carriage constructed of dark, weathered wood with massive wooden wheels rimmed with crude iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience moments of blurring caused by mud splashing onto it as the carriage races forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020599940791726112 + ], + "opposite_logprob_tokens": [ + -3.895599842071533 + ], + "logprob_reward": -0.020599940791726112, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020599940791726112, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13157841563224792, + 0.15726931393146515, + 0.10134897381067276, + 0.17433154582977295, + 0.221969872713089, + 0.032992102205753326, + 0.09318161755800247, + 0.04112282395362854, + 0.02160058170557022, + 0.024604810401797295 + ], + "credit_map": [ + 0.01761174388229847, + 0.025695666670799255, + 0.02012861706316471, + 0.017027849331498146, + 0.013748418539762497, + 0.014534554444253445, + 0.011519046500325203, + 0.011312502436339855, + 0.022590938955545425, + 0.011035433039069176, + 0.01799294538795948, + 0.015693802386522293, + 0.018403444439172745, + 0.038697171956300735, + 0.012505006045103073, + 0.020350564271211624, + 0.01023037452250719, + 0.014320909045636654, + 0.009606546722352505, + 0.0068868715316057205, + 0.010555367916822433, + 0.022769330069422722, + 0.015701690688729286, + 0.0112778814509511, + 0.029289821162819862, + 0.011406268924474716, + 0.014511250890791416, + 0.011342225596308708, + 0.01787399873137474, + 0.06598219275474548, + 0.014500880613923073, + 0.00942488294094801, + 0.02448926866054535, + 0.044075433164834976, + 0.029214782640337944, + 0.022202476859092712, + 0.022487537935376167, + 0.0467214398086071, + 0.014524743892252445, + 0.018254192546010017, + 0.002976101590320468, + 0.003655422246083617, + 0.0029244464822113514, + 0.004138119053095579, + 0.00625646160915494, + 0.006465858779847622, + 0.0023928622249513865, + 0.00418282812461257, + 0.005673144478350878, + 0.011006169952452183, + 0.010218897834420204, + 0.019263656809926033, + 0.010814184322953224, + 0.012027374468743801, + 0.008020540699362755, + 0.01615765132009983, + 0.004913081414997578, + 0.004271632991731167, + 0.004010763019323349, + 0.005210681818425655, + 0.004186625126749277, + 0.006796238012611866, + 0.006681310012936592, + 0.005052484571933746, + 0.0029568711761385202, + 0.0025766997132450342, + 0.00188830541446805, + 0.0035932334139943123, + 0.00252355239354074, + 0.0013736146502196789, + 0.0039057680405676365, + 0.002782535972073674, + 0.003937308210879564, + 0.0030281622894108295, + 0.003316597081720829, + 0.002882762812077999, + 0.0024476652033627033, + 0.003148901741951704, + 0.003190332790836692, + 0.002653075847774744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 1 + ], + "selected_reward": -0.020599940791726112 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a four-wheeled carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dark brown horses pulling the carriage with visible bulging muscles and steam emanating from their nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a worn black felt hat and a coarse cloth cape while gripping frayed leather reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carriage constructed of dark, weathered wood with massive wooden wheels rimmed with crude iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience moments of blurring caused by mud splashing onto it as the carriage races forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014206935651600361 + ], + "opposite_logprob_tokens": [ + -4.264206886291504 + ], + "logprob_reward": -0.014206935651600361, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014206935651600361, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2567766010761261, + 0.16220444440841675, + 0.13642865419387817, + 0.06304807960987091, + 0.1077585220336914, + 0.06402476131916046, + 0.07239187508821487, + 0.05934646353125572, + 0.03678763285279274, + 0.04123295471072197 + ], + "credit_map": [ + 0.035534609109163284, + 0.041720665991306305, + 0.040861550718545914, + 0.02911386266350746, + 0.02870519831776619, + 0.036235224455595016, + 0.02003851905465126, + 0.024566980078816414, + 0.02337760478258133, + 0.01231161318719387, + 0.024613099172711372, + 0.014374829828739166, + 0.02439716085791588, + 0.03515097498893738, + 0.01151017565280199, + 0.016468985006213188, + 0.013565778732299805, + 0.033197298645973206, + 0.014996204525232315, + 0.01640777662396431, + 0.014103199355304241, + 0.021294914186000824, + 0.014754118397831917, + 0.008109369315207005, + 0.008139986544847488, + 0.010418874211609364, + 0.010401982814073563, + 0.005049987696111202, + 0.012061232700943947, + 0.008468802087008953, + 0.00453900545835495, + 0.0039682090282440186, + 0.017766563221812248, + 0.013357113115489483, + 0.008886873722076416, + 0.009310131892561913, + 0.009965457953512669, + 0.0368487574160099, + 0.005704615730792284, + 0.005919012241065502, + 0.00628709327429533, + 0.0051958817057311535, + 0.005347029305994511, + 0.007824650965631008, + 0.013410019688308239, + 0.01242789439857006, + 0.005490141920745373, + 0.008042056113481522, + 0.003549522953107953, + 0.005272521637380123, + 0.006355482153594494, + 0.025214312598109245, + 0.006131215952336788, + 0.0031465415377169847, + 0.004533499013632536, + 0.018188782036304474, + 0.006642645224928856, + 0.0072571588680148125, + 0.0075971875339746475, + 0.008259566500782967, + 0.0070998454466462135, + 0.00624621519818902, + 0.008017593063414097, + 0.008226254023611546, + 0.005619282368570566, + 0.004754712805151939, + 0.003310274565592408, + 0.007411609403789043, + 0.0030214618891477585, + 0.003089379286393523, + 0.005073530599474907, + 0.004507381469011307, + 0.005666172131896019, + 0.003977777436375618, + 0.005437582731246948, + 0.004955356940627098, + 0.0034866665955632925, + 0.005414952524006367, + 0.006510688457638025, + 0.005783755332231522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.014206935651600361 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a four-wheeled carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dark brown horses pulling the carriage with visible bulging muscles and steam emanating from their nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a worn black felt hat and a coarse cloth cape while gripping frayed leather reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carriage constructed of dark, weathered wood with massive wooden wheels rimmed with crude iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience moments of blurring caused by mud splashing onto it as the carriage races forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019377756398171186 + ], + "opposite_logprob_tokens": [ + -6.2519378662109375 + ], + "logprob_reward": -0.0019377756398171186, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019377756398171186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2651651203632355, + 0.08253468573093414, + 0.04286462068557739, + 0.07961609214544296, + 0.09925886988639832, + 0.048382535576820374, + 0.1215939149260521, + 0.11717259883880615, + 0.0685124546289444, + 0.07489916682243347 + ], + "credit_map": [ + 0.03631221503019333, + 0.04050308093428612, + 0.040078055113554, + 0.03155070170760155, + 0.03138083219528198, + 0.036111436784267426, + 0.021227719262242317, + 0.028001099824905396, + 0.008297068066895008, + 0.007209894713014364, + 0.009974566288292408, + 0.009640185162425041, + 0.010016273707151413, + 0.01717105135321617, + 0.011225458234548569, + 0.009000184945762157, + 0.005552218295633793, + 0.0041335346177220345, + 0.004670553375035524, + 0.004243630915880203, + 0.011627363972365856, + 0.004778937436640263, + 0.004320204723626375, + 0.0035381768830120564, + 0.01052478514611721, + 0.008408538065850735, + 0.010361157357692719, + 0.007054079324007034, + 0.011880436912178993, + 0.017526334151625633, + 0.006864689756184816, + 0.006996070966124535, + 0.013190807774662971, + 0.019620463252067566, + 0.01243943627923727, + 0.011974241584539413, + 0.009303377009928226, + 0.016416186466813087, + 0.008774700574576855, + 0.007539653684943914, + 0.0050283027812838554, + 0.003734801895916462, + 0.0053261262364685535, + 0.004799703136086464, + 0.008960512466728687, + 0.008040225133299828, + 0.004986421670764685, + 0.007506439927965403, + 0.006799512542784214, + 0.012168248184025288, + 0.011516542173922062, + 0.028531672433018684, + 0.011785654351115227, + 0.01583961769938469, + 0.01041713822633028, + 0.0245355274528265, + 0.01001505646854639, + 0.014573431573808193, + 0.01639684848487377, + 0.01537500973790884, + 0.014451246708631516, + 0.014791772700846195, + 0.021164659410715103, + 0.010404575616121292, + 0.00976274162530899, + 0.010491025634109974, + 0.008231358602643013, + 0.00821224320679903, + 0.007569874171167612, + 0.006727668456733227, + 0.010703817009925842, + 0.006813723128288984, + 0.009650848805904388, + 0.007622108329087496, + 0.008409354835748672, + 0.009073356166481972, + 0.006263534538447857, + 0.011951825581490993, + 0.01158431638032198, + 0.010343823581933975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.0019377756398171186 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a four-wheeled carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dark brown horses pulling the carriage with visible bulging muscles and steam emanating from their nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a worn black felt hat and a coarse cloth cape while gripping frayed leather reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carriage constructed of dark, weathered wood with massive wooden wheels rimmed with crude iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera experience moments of blurring caused by mud splashing onto it as the carriage races forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018179578706622124 + ], + "opposite_logprob_tokens": [ + -4.018179416656494 + ], + "logprob_reward": -0.018179578706622124, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018179578706622124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18758292496204376, + 0.16240382194519043, + 0.08755708485841751, + 0.096861831843853, + 0.07948102056980133, + 0.03913313150405884, + 0.07505369931459427, + 0.05367988348007202, + 0.11776505410671234, + 0.10048159211874008 + ], + "credit_map": [ + 0.02571513131260872, + 0.027672072872519493, + 0.03767189010977745, + 0.02287900820374489, + 0.020241791382431984, + 0.0217173732817173, + 0.014199923723936081, + 0.017485741525888443, + 0.018156085163354874, + 0.017255648970603943, + 0.024792032316327095, + 0.01526518352329731, + 0.026517661288380623, + 0.03091059811413288, + 0.010199138894677162, + 0.019307466223835945, + 0.0075944880954921246, + 0.015630004927515984, + 0.009581453166902065, + 0.01112580019980669, + 0.014187042601406574, + 0.009520391933619976, + 0.013277627527713776, + 0.006640270352363586, + 0.014241044409573078, + 0.010597228072583675, + 0.008497255854308605, + 0.006608597002923489, + 0.008905668742954731, + 0.033395953476428986, + 0.00728214904665947, + 0.0073339375667274, + 0.01407563779503107, + 0.007444154936820269, + 0.0073133124969899654, + 0.005671321414411068, + 0.00798287708312273, + 0.02743094228208065, + 0.004855969920754433, + 0.004706800449639559, + 0.0036644621286541224, + 0.0032776235602796078, + 0.0031257709488272667, + 0.0036748622078448534, + 0.00707298144698143, + 0.00815809890627861, + 0.003642834024503827, + 0.006516495253890753, + 0.0046280971728265285, + 0.007412836421281099, + 0.007912810891866684, + 0.023267818614840508, + 0.011723416857421398, + 0.007000817451626062, + 0.006960034836083651, + 0.0061478684656322, + 0.0054929424077272415, + 0.005727711133658886, + 0.006673149764537811, + 0.006800530478358269, + 0.007023544982075691, + 0.007243762258440256, + 0.007160183973610401, + 0.007558058947324753, + 0.0201211366802454, + 0.014147903770208359, + 0.01414444949477911, + 0.01555979810655117, + 0.011048068292438984, + 0.009562523104250431, + 0.02081330679357052, + 0.01236786786466837, + 0.016219455748796463, + 0.011586060747504234, + 0.01577644608914852, + 0.00984505657106638, + 0.006890235934406519, + 0.012209908105432987, + 0.016688397154211998, + 0.011266033165156841 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.018179578706622124 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_3/index_tensor([3667, 3667])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_3/index_tensor([3667, 3667])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f7d5a1961879c76b45bec8cc50aac1c51ad8665a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_3/index_tensor([3667, 3667])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a stunningly close-up, static shot, under Roger Deakins' signature strong sidelight, a pair of weathered hands delicately twist a few flat, emerald-green leaves of West Lake Longjing tea. This precise, focused light, cast from the left side of the frame, sculpts the contours of the knuckles and the fine wrinkles of the skin, while sharply illuminating the spiral veins of the fingertips. As the tea leaves are gently twisted between the fingers, light penetrates the thin leaves, revealing their veins and creating an almost translucent jade-like texture. The extremely shallow depth of field completely blurs the background, softening it into a deep, serene dark palette, focusing the entire visual focus on these experienced hands and the fresh leaves. The intense contrast of light and dark creates an atmosphere of concentration, tranquility, and ritual. Every subtle movement appears steady and deliberate, imbued with a reverence for nature.", + "seed": 25633, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_3/index_3667_seed_25633_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1668504357803613, + "vlm_logprob": -0.1668504357803613, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.1668504357803613, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of weathered hands with visible knuckles and fine wrinkles visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are flat, emerald-green leaves resembling West Lake Longjing tea being twisted between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting strong sidelight coming from the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background extremely blurred with a deep, serene dark palette due to shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static without any movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004439142532646656 + ], + "opposite_logprob_tokens": [ + -6.254439353942871 + ], + "logprob_reward": -0.004439142532646656, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.004439142532646656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15679693222045898, + 0.068741075694561, + 0.07493943721055984, + 0.07115168869495392, + 0.08620604872703552, + 0.139008566737175, + 0.06135972961783409, + 0.08033960312604904, + 0.15749908983707428, + 0.10395779460668564 + ], + "credit_map": [ + 0.012785033322870731, + 0.02461147867143154, + 0.028942951932549477, + 0.03242776542901993, + 0.014222525991499424, + 0.01084323599934578, + 0.017066165804862976, + 0.015897780656814575, + 0.012159516103565693, + 0.00896807573735714, + 0.01105694379657507, + 0.010277526453137398, + 0.008169380016624928, + 0.0037769952323287725, + 0.005656099412590265, + 0.008676540106534958, + 0.007874337024986744, + 0.010229536332190037, + 0.009258902631700039, + 0.010832749307155609, + 0.010684320703148842, + 0.005125816445797682, + 0.011287989094853401, + 0.009645781479775906, + 0.00950058363378048, + 0.011247980408370495, + 0.010754005052149296, + 0.013407121412456036, + 0.006474291905760765, + 0.0032196701504290104, + 0.008555371314287186, + 0.007992664352059364, + 0.009990010410547256, + 0.008625966496765614, + 0.012827586382627487, + 0.018741026520729065, + 0.009087990038096905, + 0.005667875055223703, + 0.010747038759291172, + 0.010518559254705906, + 0.020551061257719994, + 0.015628425404429436, + 0.02142396569252014, + 0.02922048419713974, + 0.01222479809075594, + 0.006450498942285776, + 0.019310420379042625, + 0.014198912307620049, + 0.006844285875558853, + 0.005881043151021004, + 0.008859978057444096, + 0.007343212608247995, + 0.010364748537540436, + 0.00456087663769722, + 0.011620580218732357, + 0.005885002668946981, + 0.011982351541519165, + 0.006182564422488213, + 0.014152558520436287, + 0.011586776934564114, + 0.010628469288349152, + 0.0077426438219845295, + 0.008492186665534973, + 0.00957205519080162, + 0.02300010435283184, + 0.017140045762062073, + 0.023037482053041458, + 0.013651651330292225, + 0.02133948542177677, + 0.01402097474783659, + 0.027237597852945328, + 0.018071752041578293, + 0.012402351014316082, + 0.00907975248992443, + 0.013780156150460243, + 0.010336312465369701, + 0.015147814527153969, + 0.00902474019676447, + 0.011585619300603867, + 0.022601043805480003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.004439142532646656 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of weathered hands with visible knuckles and fine wrinkles visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are flat, emerald-green leaves resembling West Lake Longjing tea being twisted between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting strong sidelight coming from the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background extremely blurred with a deep, serene dark palette due to shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static without any movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007643027231097221 + ], + "opposite_logprob_tokens": [ + -4.882643222808838 + ], + "logprob_reward": -0.007643027231097221, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007643027231097221, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1130295842885971, + 0.08319050073623657, + 0.2276284545660019, + 0.0756707563996315, + 0.14258453249931335, + 0.057571571320295334, + 0.08602625131607056, + 0.08901875466108322, + 0.06307125091552734, + 0.06220833584666252 + ], + "credit_map": [ + 0.010583851486444473, + 0.014120884239673615, + 0.01755172200500965, + 0.017677366733551025, + 0.01735193468630314, + 0.00973476655781269, + 0.01121074240654707, + 0.014798316173255444, + 0.014127751812338829, + 0.010336235165596008, + 0.011906804516911507, + 0.011616009287536144, + 0.007749687414616346, + 0.0044889324344694614, + 0.008722509257495403, + 0.014242569915950298, + 0.02460470236837864, + 0.03706313297152519, + 0.027963483706116676, + 0.0296307522803545, + 0.0458906888961792, + 0.00969279557466507, + 0.03342636302113533, + 0.019356539472937584, + 0.00870057474821806, + 0.0079500125721097, + 0.01170390099287033, + 0.016456924378871918, + 0.010667072609066963, + 0.004259586334228516, + 0.00891465786844492, + 0.007018017116934061, + 0.018117688596248627, + 0.01435681339353323, + 0.022244445979595184, + 0.02114262804389, + 0.02033527009189129, + 0.010990463197231293, + 0.020011648535728455, + 0.015385587699711323, + 0.007241566199809313, + 0.005322424694895744, + 0.008997074328362942, + 0.0080441664904356, + 0.012319252826273441, + 0.0022533482406288385, + 0.008594071492552757, + 0.004799664486199617, + 0.011705437675118446, + 0.006874986924231052, + 0.014663086272776127, + 0.008844892494380474, + 0.014210198074579239, + 0.008256264962255955, + 0.011927974410355091, + 0.009543412365019321, + 0.011729050427675247, + 0.011790147051215172, + 0.014571917243301868, + 0.010966812260448933, + 0.013869380578398705, + 0.007280559279024601, + 0.007753881625831127, + 0.011057003401219845, + 0.007828389294445515, + 0.009175061248242855, + 0.008333085104823112, + 0.0067389653995633125, + 0.011279921978712082, + 0.004994745831936598, + 0.007989240810275078, + 0.006731838453561068, + 0.008742492645978928, + 0.007826754823327065, + 0.008582096546888351, + 0.006353138945996761, + 0.01024636160582304, + 0.004020641557872295, + 0.009038124233484268, + 0.007398726884275675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.007643027231097221 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of weathered hands with visible knuckles and fine wrinkles visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are flat, emerald-green leaves resembling West Lake Longjing tea being twisted between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting strong sidelight coming from the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background extremely blurred with a deep, serene dark palette due to shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static without any movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12705299258232117 + ], + "opposite_logprob_tokens": [ + -2.1270530223846436 + ], + "logprob_reward": -0.12705299258232117, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12705299258232117, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08081363141536713, + 0.148000568151474, + 0.08631555736064911, + 0.07316192984580994, + 0.06333765387535095, + 0.08859917521476746, + 0.028835203498601913, + 0.19105570018291473, + 0.1798105537891388, + 0.060070037841796875 + ], + "credit_map": [ + 0.01196361891925335, + 0.011383265256881714, + 0.01088864915072918, + 0.014535085298120975, + 0.010136585682630539, + 0.00576922157779336, + 0.007782589644193649, + 0.0083546107634902, + 0.01826300658285618, + 0.017306270077824593, + 0.02085539512336254, + 0.02381264790892601, + 0.03276238590478897, + 0.0061768521554768085, + 0.016450466588139534, + 0.012373537756502628, + 0.011204119771718979, + 0.010364511981606483, + 0.014681932516396046, + 0.017039705067873, + 0.01155335083603859, + 0.0036293731536716223, + 0.011417679488658905, + 0.006424893159419298, + 0.013585339300334454, + 0.009440591558814049, + 0.010775933973491192, + 0.01312361378222704, + 0.0060568321496248245, + 0.0038316736463457346, + 0.007781561929732561, + 0.008566383272409439, + 0.008768392726778984, + 0.007028589025139809, + 0.010911673307418823, + 0.010490740649402142, + 0.008246442303061485, + 0.0039642988704144955, + 0.008732951246201992, + 0.00519456947222352, + 0.013267772272229195, + 0.0075224619358778, + 0.014921411871910095, + 0.010316969826817513, + 0.022074826061725616, + 0.004019374493509531, + 0.009543441236019135, + 0.006932917051017284, + 0.00481925206258893, + 0.0028245323337614536, + 0.004641965497285128, + 0.003536979667842388, + 0.003939141985028982, + 0.0020432444289326668, + 0.004175476264208555, + 0.002854613121598959, + 0.026104753836989403, + 0.018994346261024475, + 0.03284720703959465, + 0.028116358444094658, + 0.026326805353164673, + 0.012374764308333397, + 0.021548844873905182, + 0.02474263310432434, + 0.03148303180932999, + 0.022284047678112984, + 0.02552415244281292, + 0.019579509273171425, + 0.019383162260055542, + 0.016727443784475327, + 0.02176038920879364, + 0.023068809881806374, + 0.007789310533553362, + 0.0045168776996433735, + 0.008191501721739769, + 0.0060304333455860615, + 0.005847369786351919, + 0.0056770797818899155, + 0.011099200695753098, + 0.0109182670712471 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 1 + ], + "selected_reward": -0.12705299258232117 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of weathered hands with visible knuckles and fine wrinkles visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are flat, emerald-green leaves resembling West Lake Longjing tea being twisted between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting strong sidelight coming from the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background extremely blurred with a deep, serene dark palette due to shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static without any movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017143089789897203 + ], + "opposite_logprob_tokens": [ + -6.37671422958374 + ], + "logprob_reward": -0.0017143089789897203, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017143089789897203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11642663925886154, + 0.215948686003685, + 0.058389727026224136, + 0.05026301369071007, + 0.04623614251613617, + 0.15398845076560974, + 0.03809459134936333, + 0.09988842159509659, + 0.1434159129858017, + 0.07734842598438263 + ], + "credit_map": [ + 0.015100860968232155, + 0.02137853391468525, + 0.017413431778550148, + 0.020688684657216072, + 0.012942180968821049, + 0.0075108748860657215, + 0.009837464429438114, + 0.011554604396224022, + 0.03412850201129913, + 0.030937038362026215, + 0.026949530467391014, + 0.027510564774274826, + 0.03236245736479759, + 0.012955360114574432, + 0.021091772243380547, + 0.03001346066594124, + 0.007520459592342377, + 0.006518456153571606, + 0.010689891874790192, + 0.0108175752684474, + 0.006120353005826473, + 0.004199748393148184, + 0.0066588521003723145, + 0.005864391569048166, + 0.010915949940681458, + 0.004740261938422918, + 0.0074888127855956554, + 0.008277680724859238, + 0.004724063444882631, + 0.0022471630945801735, + 0.006261566653847694, + 0.005607519764453173, + 0.005911642219871283, + 0.005084417760372162, + 0.006717340089380741, + 0.008337088860571384, + 0.004282907582819462, + 0.003229941939935088, + 0.00641678087413311, + 0.0062560285441577435, + 0.021688729524612427, + 0.015149634331464767, + 0.021767040714621544, + 0.031032126396894455, + 0.01691148616373539, + 0.010102374479174614, + 0.021966945379972458, + 0.015370121225714684, + 0.006189877167344093, + 0.0037079574540257454, + 0.006553737912327051, + 0.005495522636920214, + 0.005150137934833765, + 0.0025062880013138056, + 0.004904787056148052, + 0.003586284816265106, + 0.025242788717150688, + 0.00743102328851819, + 0.01867019012570381, + 0.01344564650207758, + 0.009149116463959217, + 0.00915736798197031, + 0.008216111920773983, + 0.008576170541346073, + 0.025787897408008575, + 0.013176171109080315, + 0.021048491820693016, + 0.013897896744310856, + 0.017632540315389633, + 0.015067577362060547, + 0.01710391230881214, + 0.01970142126083374, + 0.009210591204464436, + 0.006094896234571934, + 0.010679776780307293, + 0.005956483073532581, + 0.012218429706990719, + 0.005555990617722273, + 0.013528745621442795, + 0.014103513211011887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 8 + ], + "selected_reward": -0.0017143089789897203 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of weathered hands with visible knuckles and fine wrinkles visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are flat, emerald-green leaves resembling West Lake Longjing tea being twisted between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting strong sidelight coming from the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background extremely blurred with a deep, serene dark palette due to shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static without any movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6934027075767517 + ], + "opposite_logprob_tokens": [ + -0.6934027075767517 + ], + "logprob_reward": -0.6934027075767517, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6934027075767517, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10898276418447495, + 0.110415518283844, + 0.03465396165847778, + 0.07506369799375534, + 0.05811465159058571, + 0.34624403715133667, + 0.024105224758386612, + 0.06225143000483513, + 0.11010409146547318, + 0.07006463408470154 + ], + "credit_map": [ + 0.011258329264819622, + 0.014325922355055809, + 0.02020350843667984, + 0.023364538326859474, + 0.01052380632609129, + 0.006945684086531401, + 0.008917533792555332, + 0.013443425297737122, + 0.013569706119596958, + 0.016832219436764717, + 0.015905899927020073, + 0.019683225080370903, + 0.011492956429719925, + 0.008245960809290409, + 0.010640779510140419, + 0.01404476072639227, + 0.003851396031677723, + 0.004246785771101713, + 0.005337637849152088, + 0.005772264674305916, + 0.0033976687118411064, + 0.0029504650738090277, + 0.004548022523522377, + 0.004549721721559763, + 0.013738233596086502, + 0.007534034550189972, + 0.012461735866963863, + 0.013839778490364552, + 0.007821645587682724, + 0.0030425225850194693, + 0.008350680582225323, + 0.008275052532553673, + 0.009186461567878723, + 0.004700209479779005, + 0.009666398167610168, + 0.01174075249582529, + 0.005333836656063795, + 0.0045744143426418304, + 0.005621813237667084, + 0.0072907619178295135, + 0.04323321953415871, + 0.04381665587425232, + 0.051710281521081924, + 0.0793340876698494, + 0.031878724694252014, + 0.015458514913916588, + 0.047436341643333435, + 0.03337620571255684, + 0.003878844901919365, + 0.002725726692005992, + 0.003518257290124893, + 0.0025686719454824924, + 0.005555156618356705, + 0.0014496714575216174, + 0.0028142365626990795, + 0.001594657078385353, + 0.011276176199316978, + 0.005454380996525288, + 0.011366108432412148, + 0.008418683893978596, + 0.005825995001941919, + 0.005507026333361864, + 0.0059147062711417675, + 0.008488344959914684, + 0.013027038425207138, + 0.011727877892553806, + 0.013295521959662437, + 0.010038142092525959, + 0.014024031348526478, + 0.011614724062383175, + 0.023707151412963867, + 0.012669590301811695, + 0.010110114701092243, + 0.007384380791336298, + 0.010153266601264477, + 0.006715457420796156, + 0.01178863737732172, + 0.0045121097937226295, + 0.010170399211347103, + 0.009230260737240314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 8 + ], + "selected_reward": -0.6934027075767517 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "In a stunningly close-up, static shot, under Roger Deakins' signature strong sidelight, a pair of weathered hands delicately twist a few flat, emerald-green leaves of West Lake Longjing tea. This precise, focused light, cast from the left side of the frame, sculpts the contours of the knuckles and the fine wrinkles of the skin, while sharply illuminating the spiral veins of the fingertips. As the tea leaves are gently twisted between the fingers, light penetrates the thin leaves, revealing their veins and creating an almost translucent jade-like texture. The extremely shallow depth of field completely blurs the background, softening it into a deep, serene dark palette, focusing the entire visual focus on these experienced hands and the fresh leaves. The intense contrast of light and dark creates an atmosphere of concentration, tranquility, and ritual. Every subtle movement appears steady and deliberate, imbued with a reverence for nature.", + "seed": 25634, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_3/index_3667_seed_25634_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1648661142680794, + "vlm_logprob": -0.1648661142680794, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.1648661142680794, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of weathered hands with visible knuckles and fine wrinkles visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are flat, emerald-green leaves resembling West Lake Longjing tea being twisted between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting strong sidelight coming from the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background extremely blurred with a deep, serene dark palette due to shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static without any movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006978305988013744 + ], + "opposite_logprob_tokens": [ + -5.256978511810303 + ], + "logprob_reward": -0.006978305988013744, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.006978305988013744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1159321591258049, + 0.04414411261677742, + 0.06366228312253952, + 0.058450404554605484, + 0.07449520379304886, + 0.2636356055736542, + 0.10740001499652863, + 0.06228901073336601, + 0.15563073754310608, + 0.05436047539114952 + ], + "credit_map": [ + 0.013936678878962994, + 0.01142465602606535, + 0.01912480965256691, + 0.018420523032546043, + 0.011828077957034111, + 0.011219884268939495, + 0.014334571547806263, + 0.015642955899238586, + 0.005749018397182226, + 0.003443122608587146, + 0.007967231795191765, + 0.008469619788229465, + 0.0037686428986489773, + 0.0029293654952198267, + 0.005739626009017229, + 0.006077482830733061, + 0.00991379376500845, + 0.006659248378127813, + 0.0080147385597229, + 0.011139915324747562, + 0.003520217491313815, + 0.005006713327020407, + 0.007296377792954445, + 0.012111272662878036, + 0.008470895700156689, + 0.007269450929015875, + 0.008713352493941784, + 0.010237135924398899, + 0.0037500152830034494, + 0.005282653961330652, + 0.0069427089765667915, + 0.007784191053360701, + 0.015351064503192902, + 0.007780547719448805, + 0.014881697483360767, + 0.013018355704843998, + 0.0034141961950808764, + 0.005635598208755255, + 0.005869969725608826, + 0.008543774485588074, + 0.01331098098307848, + 0.009855633601546288, + 0.10181570053100586, + 0.10400430113077164, + 0.007090622093528509, + 0.006888261530548334, + 0.008409172296524048, + 0.01226092129945755, + 0.013940935023128986, + 0.011108859442174435, + 0.030136970803141594, + 0.027878817170858383, + 0.00557295186445117, + 0.005673758219927549, + 0.005574776325374842, + 0.0075129494071006775, + 0.007354579400271177, + 0.005546010565012693, + 0.011884479783475399, + 0.014925827272236347, + 0.0036816755309700966, + 0.004417198710143566, + 0.006758806761354208, + 0.0077204350382089615, + 0.0170550886541605, + 0.009724571369588375, + 0.04218670353293419, + 0.04500386863946915, + 0.007880598306655884, + 0.007143336348235607, + 0.012011365965008736, + 0.014625204727053642, + 0.007349048275500536, + 0.006119901780039072, + 0.00701545923948288, + 0.009116260334849358, + 0.005300379358232021, + 0.004419988952577114, + 0.007086950354278088, + 0.00795248243957758 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.006978305988013744 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of weathered hands with visible knuckles and fine wrinkles visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are flat, emerald-green leaves resembling West Lake Longjing tea being twisted between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting strong sidelight coming from the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background extremely blurred with a deep, serene dark palette due to shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static without any movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008647376671433449 + ], + "opposite_logprob_tokens": [ + -4.758647441864014 + ], + "logprob_reward": -0.008647376671433449, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008647376671433449, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08341825008392334, + 0.0587974451482296, + 0.0895327627658844, + 0.07338734716176987, + 0.060184624046087265, + 0.06656108051538467, + 0.10551458597183228, + 0.24440209567546844, + 0.1745324432849884, + 0.04366935044527054 + ], + "credit_map": [ + 0.010861225426197052, + 0.006260902155190706, + 0.013659351505339146, + 0.01905977725982666, + 0.006809794344007969, + 0.007668083533644676, + 0.012415336444973946, + 0.006683771498501301, + 0.0069931247271597385, + 0.0072177438996732235, + 0.008824223652482033, + 0.009194458834826946, + 0.006449677050113678, + 0.005294088274240494, + 0.008462353609502316, + 0.006361772771924734, + 0.011016493663191795, + 0.007558700628578663, + 0.012008768506348133, + 0.016825556755065918, + 0.006514828186482191, + 0.006889174226671457, + 0.013790080323815346, + 0.014929153956472874, + 0.004937948193401098, + 0.003851822344586253, + 0.0096218166872859, + 0.01363598182797432, + 0.008950614370405674, + 0.004478526767343283, + 0.018155081197619438, + 0.009755549021065235, + 0.00821537896990776, + 0.005872167181223631, + 0.009382334537804127, + 0.009736224077641964, + 0.005821768194437027, + 0.0057242619805037975, + 0.006342343054711819, + 0.009090145118534565, + 0.005690297111868858, + 0.0060338410548865795, + 0.013077862560749054, + 0.014480739831924438, + 0.006502676755189896, + 0.0030799605883657932, + 0.007975003682076931, + 0.0097206961363554, + 0.009877746924757957, + 0.01001057866960764, + 0.02647564187645912, + 0.027621982619166374, + 0.007774795405566692, + 0.004883293062448502, + 0.008976073935627937, + 0.009894464164972305, + 0.0218913983553648, + 0.029722152277827263, + 0.03672192618250847, + 0.03829577937722206, + 0.02630394697189331, + 0.015487313270568848, + 0.044508788734674454, + 0.03147076070308685, + 0.017787348479032516, + 0.013908586464822292, + 0.03815458342432976, + 0.04644107073545456, + 0.012447429820895195, + 0.011952990666031837, + 0.01700880192220211, + 0.016831612214446068, + 0.0049950918182730675, + 0.004545765928924084, + 0.005740531720221043, + 0.008558491244912148, + 0.003915468696504831, + 0.003624435281381011, + 0.0065229530446231365, + 0.005766611546278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.008647376671433449 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of weathered hands with visible knuckles and fine wrinkles visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are flat, emerald-green leaves resembling West Lake Longjing tea being twisted between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting strong sidelight coming from the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background extremely blurred with a deep, serene dark palette due to shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static without any movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866386204957962 + ], + "opposite_logprob_tokens": [ + -3.048663854598999 + ], + "logprob_reward": -0.04866386204957962, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866386204957962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12043802440166473, + 0.07917975634336472, + 0.03800097852945328, + 0.0621219165623188, + 0.12695010006427765, + 0.3631032407283783, + 0.026956729590892792, + 0.0410691574215889, + 0.06774712353944778, + 0.07443302869796753 + ], + "credit_map": [ + 0.0127332778647542, + 0.011597516015172005, + 0.017447752878069878, + 0.019885677844285965, + 0.013335363939404488, + 0.012203829362988472, + 0.016707396134734154, + 0.016527196392416954, + 0.013266499154269695, + 0.006678876932710409, + 0.015622245147824287, + 0.011587433516979218, + 0.005847916007041931, + 0.006144908722490072, + 0.008407351560890675, + 0.01162451971322298, + 0.004498528782278299, + 0.004267600830644369, + 0.004225472453981638, + 0.007233864162117243, + 0.002758907387033105, + 0.001947057549841702, + 0.006853059865534306, + 0.00621648458763957, + 0.008611957542598248, + 0.0066517735831439495, + 0.00875843781977892, + 0.010658040642738342, + 0.00534931430593133, + 0.0036517020780593157, + 0.008836035616695881, + 0.00960464682430029, + 0.0186329148709774, + 0.011804182082414627, + 0.028372082859277725, + 0.02712200954556465, + 0.007044343277812004, + 0.008713250048458576, + 0.010491418652236462, + 0.01476989034563303, + 0.042270317673683167, + 0.030756903812289238, + 0.10335930436849594, + 0.10894741863012314, + 0.017197750508785248, + 0.014283481054008007, + 0.019218478351831436, + 0.02706953138113022, + 0.003340873634442687, + 0.004222668241709471, + 0.004929122049361467, + 0.005156700033694506, + 0.0032575202640146017, + 0.0014093139907345176, + 0.0022920093033462763, + 0.0023485207930207253, + 0.003478534519672394, + 0.0036261952482163906, + 0.007167249917984009, + 0.009824689477682114, + 0.0028417243156582117, + 0.0032520380336791277, + 0.005620989948511124, + 0.005257732700556517, + 0.006760040298104286, + 0.005102754104882479, + 0.014531461521983147, + 0.016687985509634018, + 0.005781205836683512, + 0.004817965906113386, + 0.006799019873142242, + 0.007266687694936991, + 0.010079490020871162, + 0.00875889789313078, + 0.0104021355509758, + 0.011733954772353172, + 0.007336197886615992, + 0.006929924711585045, + 0.009052651934325695, + 0.010139770805835724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.04866386204957962 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of weathered hands with visible knuckles and fine wrinkles visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are flat, emerald-green leaves resembling West Lake Longjing tea being twisted between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting strong sidelight coming from the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background extremely blurred with a deep, serene dark palette due to shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static without any movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021972341928631067 + ], + "opposite_logprob_tokens": [ + -6.127197265625 + ], + "logprob_reward": -0.0021972341928631067, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021972341928631067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10887964814901352, + 0.10083209723234177, + 0.021489353850483894, + 0.03539981693029404, + 0.07484711706638336, + 0.4435894787311554, + 0.0355503112077713, + 0.04905560612678528, + 0.047800883650779724, + 0.08255564421415329 + ], + "credit_map": [ + 0.012031789869070053, + 0.011774714104831219, + 0.01883520558476448, + 0.01809375174343586, + 0.008847451768815517, + 0.014535537920892239, + 0.010940774343907833, + 0.013820432126522064, + 0.011956602334976196, + 0.011277877725660801, + 0.015888091176748276, + 0.017003776505589485, + 0.008655441924929619, + 0.006857306230813265, + 0.012682951986789703, + 0.01651005633175373, + 0.002688829554244876, + 0.0021657587494701147, + 0.002949923975393176, + 0.0038416623137891293, + 0.0017577674007043242, + 0.0016802719328552485, + 0.003120322711765766, + 0.003284817561507225, + 0.005803659558296204, + 0.0038083104882389307, + 0.007342809345573187, + 0.007189195603132248, + 0.002141204196959734, + 0.0017395602772012353, + 0.0033152818214148283, + 0.004059795290231705, + 0.013866685330867767, + 0.008304652757942677, + 0.016649434342980385, + 0.015507377684116364, + 0.004631349816918373, + 0.005289198365062475, + 0.004824679810553789, + 0.005773743614554405, + 0.016058262437582016, + 0.011820008978247643, + 0.18166591227054596, + 0.18888024985790253, + 0.007484500762075186, + 0.01101558655500412, + 0.01092176791280508, + 0.01574322208762169, + 0.003966824617236853, + 0.004288621712476015, + 0.00835781916975975, + 0.008144001476466656, + 0.0029521530959755182, + 0.001721089705824852, + 0.0028479110915213823, + 0.003271890804171562, + 0.004306081682443619, + 0.004038482904434204, + 0.01170788798481226, + 0.01171116903424263, + 0.0029627038165926933, + 0.003428298979997635, + 0.005888219922780991, + 0.005012760404497385, + 0.0037742832209914923, + 0.003252235474064946, + 0.008448978886008263, + 0.010170212015509605, + 0.00549545232206583, + 0.0038778267335146666, + 0.0067503368481993675, + 0.00603155791759491, + 0.010296492837369442, + 0.0077350749634206295, + 0.013383550569415092, + 0.015457426197826862, + 0.00561814708635211, + 0.005632707383483648, + 0.010406537912786007, + 0.01402570866048336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.0021972341928631067 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of weathered hands with visible knuckles and fine wrinkles visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are flat, emerald-green leaves resembling West Lake Longjing tea being twisted between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting strong sidelight coming from the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background extremely blurred with a deep, serene dark palette due to shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static without any movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578437924385071 + ], + "opposite_logprob_tokens": [ + -0.6328437924385071 + ], + "logprob_reward": -0.7578437924385071, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578437924385071, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.13098472356796265, + 0.05608544498682022, + 0.12680643796920776, + 0.07646716386079788, + 0.09611741453409195, + 0.12915630638599396, + 0.05019821971654892, + 0.08378296345472336, + 0.0763629600405693, + 0.17403832077980042 + ], + "credit_map": [ + 0.016389504075050354, + 0.012312026694417, + 0.02000279352068901, + 0.019937356933951378, + 0.009380425326526165, + 0.017517169937491417, + 0.014924445189535618, + 0.020520994439721107, + 0.0054098800756037235, + 0.00405909214168787, + 0.007242112420499325, + 0.01081124972552061, + 0.004601685330271721, + 0.00456927390769124, + 0.007211268879473209, + 0.012180881574749947, + 0.021099135279655457, + 0.014125792309641838, + 0.017407210543751717, + 0.023279884830117226, + 0.006738108582794666, + 0.007796324323862791, + 0.014136080630123615, + 0.022223908454179764, + 0.013190766796469688, + 0.008120729587972164, + 0.011788859032094479, + 0.013487574644386768, + 0.0042175534181296825, + 0.0062131742015480995, + 0.008210781961679459, + 0.011237715370953083, + 0.015799565240740776, + 0.009738544933497906, + 0.02198462001979351, + 0.018691444769501686, + 0.0056122480891644955, + 0.00834096409380436, + 0.006458192598074675, + 0.009491835720837116, + 0.009028544649481773, + 0.009273475036025047, + 0.04243769496679306, + 0.04461856558918953, + 0.005643413867801428, + 0.004107826389372349, + 0.006535042077302933, + 0.0075117312371730804, + 0.005774552933871746, + 0.0059729404747486115, + 0.0098716476932168, + 0.009550528600811958, + 0.0034573981538414955, + 0.003651874838396907, + 0.006121146958321333, + 0.005798132624477148, + 0.008090602234005928, + 0.00897427648305893, + 0.01912030018866062, + 0.021405398845672607, + 0.0048559801653027534, + 0.00600239634513855, + 0.007442942354828119, + 0.007891063578426838, + 0.007142858114093542, + 0.005986912176012993, + 0.013859232887625694, + 0.01384017150849104, + 0.008189731277525425, + 0.006040287669748068, + 0.010886825621128082, + 0.010416935198009014, + 0.02354609966278076, + 0.021999265998601913, + 0.019061662256717682, + 0.03050367720425129, + 0.015158277936279774, + 0.014678706414997578, + 0.024138513952493668, + 0.02495211735367775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.7578437924385071 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "In a stunningly close-up, static shot, under Roger Deakins' signature strong sidelight, a pair of weathered hands delicately twist a few flat, emerald-green leaves of West Lake Longjing tea. This precise, focused light, cast from the left side of the frame, sculpts the contours of the knuckles and the fine wrinkles of the skin, while sharply illuminating the spiral veins of the fingertips. As the tea leaves are gently twisted between the fingers, light penetrates the thin leaves, revealing their veins and creating an almost translucent jade-like texture. The extremely shallow depth of field completely blurs the background, softening it into a deep, serene dark palette, focusing the entire visual focus on these experienced hands and the fresh leaves. The intense contrast of light and dark creates an atmosphere of concentration, tranquility, and ritual. Every subtle movement appears steady and deliberate, imbued with a reverence for nature.", + "seed": 25635, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_3/index_3667_seed_25635_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06368065471760928, + "vlm_logprob": -0.06368065471760928, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.06368065471760928, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of weathered hands with visible knuckles and fine wrinkles visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are flat, emerald-green leaves resembling West Lake Longjing tea being twisted between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting strong sidelight coming from the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background extremely blurred with a deep, serene dark palette due to shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static without any movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003922389354556799 + ], + "opposite_logprob_tokens": [ + -6.378922462463379 + ], + "logprob_reward": -0.003922389354556799, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.003922389354556799, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17780232429504395, + 0.14127978682518005, + 0.0715174674987793, + 0.046024151146411896, + 0.0982355996966362, + 0.1963571459054947, + 0.06864599883556366, + 0.05485057830810547, + 0.07685519009828568, + 0.0684317797422409 + ], + "credit_map": [ + 0.03981217369437218, + 0.01914484240114689, + 0.023749416694045067, + 0.015102403238415718, + 0.01506096962839365, + 0.016029661521315575, + 0.0259380079805851, + 0.02296483889222145, + 0.02222108095884323, + 0.016283534467220306, + 0.018136946484446526, + 0.013334953226149082, + 0.013387211598455906, + 0.013851161114871502, + 0.024216046556830406, + 0.019848845899105072, + 0.014999901875853539, + 0.008527690544724464, + 0.008878855034708977, + 0.010729989968240261, + 0.0034671733155846596, + 0.0053408388048410416, + 0.008558560162782669, + 0.011014454066753387, + 0.00438700569793582, + 0.007422884926199913, + 0.0058905500918626785, + 0.004021720495074987, + 0.0035722502507269382, + 0.00338911940343678, + 0.009192523546516895, + 0.008148100227117538, + 0.017636416479945183, + 0.01550283096730709, + 0.010226543992757797, + 0.010067264549434185, + 0.009473462589085102, + 0.005990002769976854, + 0.016524482518434525, + 0.012814600951969624, + 0.022020619362592697, + 0.00949837639927864, + 0.03426916524767876, + 0.029772531241178513, + 0.011504766531288624, + 0.020391643047332764, + 0.03213992714881897, + 0.0367601215839386, + 0.007643285673111677, + 0.006321406923234463, + 0.012268106453120708, + 0.008203443139791489, + 0.006552902050316334, + 0.007549166213721037, + 0.009861732833087444, + 0.010245952755212784, + 0.004816207569092512, + 0.004560008179396391, + 0.014012080617249012, + 0.005375630687922239, + 0.005785176530480385, + 0.007404555566608906, + 0.005756511352956295, + 0.007140412461012602, + 0.00986821111291647, + 0.008012078702449799, + 0.00967198982834816, + 0.007263867184519768, + 0.008607595227658749, + 0.012048653326928616, + 0.01001407578587532, + 0.011368714272975922, + 0.00892207957804203, + 0.006355348974466324, + 0.014507030136883259, + 0.007419457659125328, + 0.005149914883077145, + 0.008214143104851246, + 0.008771986700594425, + 0.009091818705201149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.003922389354556799 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of weathered hands with visible knuckles and fine wrinkles visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are flat, emerald-green leaves resembling West Lake Longjing tea being twisted between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting strong sidelight coming from the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background extremely blurred with a deep, serene dark palette due to shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static without any movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009792386554181576 + ], + "opposite_logprob_tokens": [ + -4.634792327880859 + ], + "logprob_reward": -0.009792386554181576, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009792386554181576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2672080993652344, + 0.14904288947582245, + 0.1007101759314537, + 0.061203282326459885, + 0.08571942895650864, + 0.06835728138685226, + 0.0938238725066185, + 0.05735424906015396, + 0.07337221503257751, + 0.0432085320353508 + ], + "credit_map": [ + 0.03056010976433754, + 0.022189276292920113, + 0.03633706271648407, + 0.04402502626180649, + 0.023095330223441124, + 0.040284253656864166, + 0.02934381738305092, + 0.04137323051691055, + 0.014021058566868305, + 0.014229439198970795, + 0.013792821206152439, + 0.012086402624845505, + 0.01469107624143362, + 0.010993888601660728, + 0.041640911251306534, + 0.027587296441197395, + 0.016971683129668236, + 0.012250110507011414, + 0.009681463241577148, + 0.010962273925542831, + 0.006523557007312775, + 0.006038537248969078, + 0.02208866737782955, + 0.016193877905607224, + 0.005054329056292772, + 0.0053063007071614265, + 0.006440493278205395, + 0.0074650621972978115, + 0.004232749342918396, + 0.005499440710991621, + 0.01610654406249523, + 0.011098363436758518, + 0.006882813293486834, + 0.005730388220399618, + 0.00899315346032381, + 0.010815026238560677, + 0.01075859647244215, + 0.005578064825385809, + 0.022379986941814423, + 0.01458139531314373, + 0.005980781279504299, + 0.004494854714721441, + 0.010849982500076294, + 0.01036049798130989, + 0.007541131228208542, + 0.00714309886097908, + 0.010670499876141548, + 0.011316430754959583, + 0.008144119754433632, + 0.006356855388730764, + 0.015040500089526176, + 0.014232045039534569, + 0.009527194313704967, + 0.013730294071137905, + 0.012459702789783478, + 0.014333167113363743, + 0.0069615659303963184, + 0.00399126997217536, + 0.006305644288659096, + 0.004763786680996418, + 0.008903260342776775, + 0.008371399715542793, + 0.01104920357465744, + 0.007008117623627186, + 0.005783021450042725, + 0.006709694396704435, + 0.006378017365932465, + 0.005837799049913883, + 0.01670190505683422, + 0.01238112710416317, + 0.011134442873299122, + 0.00844621006399393, + 0.0038467461708933115, + 0.004173745401203632, + 0.004365865606814623, + 0.004349996335804462, + 0.004722468554973602, + 0.00725982803851366, + 0.008325591683387756, + 0.006164290010929108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.009792386554181576 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of weathered hands with visible knuckles and fine wrinkles visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are flat, emerald-green leaves resembling West Lake Longjing tea being twisted between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting strong sidelight coming from the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background extremely blurred with a deep, serene dark palette due to shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static without any movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031528025865555 + ], + "opposite_logprob_tokens": [ + -2.3503153324127197 + ], + "logprob_reward": -0.10031528025865555, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031528025865555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10399939119815826, + 0.03157149627804756, + 0.06202782690525055, + 0.037927694618701935, + 0.09959065914154053, + 0.188870370388031, + 0.06020238623023033, + 0.09358026832342148, + 0.2516247034072876, + 0.07060522586107254 + ], + "credit_map": [ + 0.01336574088782072, + 0.008450285531580448, + 0.014804317615926266, + 0.015918975695967674, + 0.009376781061291695, + 0.014435128308832645, + 0.012700995430350304, + 0.014947156421840191, + 0.004108275752514601, + 0.0030611816328018904, + 0.003745222697034478, + 0.004792253952473402, + 0.003521162550896406, + 0.002938390476629138, + 0.005048620514571667, + 0.004356385208666325, + 0.008411766961216927, + 0.009839914739131927, + 0.008878643624484539, + 0.007864140905439854, + 0.004579755011945963, + 0.004882406909018755, + 0.007858901284635067, + 0.00971229188144207, + 0.004006375093013048, + 0.0031464030034840107, + 0.00491017010062933, + 0.005587636958807707, + 0.0035848466213792562, + 0.003001217031851411, + 0.006865326780825853, + 0.006825712509453297, + 0.019191358238458633, + 0.016977699473500252, + 0.015138210728764534, + 0.010729648172855377, + 0.00654340535402298, + 0.008192507550120354, + 0.010468428954482079, + 0.012349397875368595, + 0.021459262818098068, + 0.011458082124590874, + 0.032952554523944855, + 0.027521837502717972, + 0.011627055704593658, + 0.02280391938984394, + 0.02902885340154171, + 0.03201879933476448, + 0.006277002859860659, + 0.00570704136043787, + 0.009082349948585033, + 0.006399437785148621, + 0.005864928476512432, + 0.009194144047796726, + 0.00899762287735939, + 0.008679850958287716, + 0.008895516395568848, + 0.0071813310496509075, + 0.012268292717635632, + 0.013320721685886383, + 0.007472177967429161, + 0.020390460267663002, + 0.011375623755156994, + 0.012676136568188667, + 0.03835389018058777, + 0.030372565612196922, + 0.02463231422007084, + 0.020859304815530777, + 0.02620253525674343, + 0.04181138798594475, + 0.031838223338127136, + 0.03755446523427963, + 0.007231582887470722, + 0.006309906952083111, + 0.010731196030974388, + 0.008936732076108456, + 0.004382885992527008, + 0.011125004850327969, + 0.01219943631440401, + 0.009688465856015682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.10031528025865555 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of weathered hands with visible knuckles and fine wrinkles visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are flat, emerald-green leaves resembling West Lake Longjing tea being twisted between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting strong sidelight coming from the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background extremely blurred with a deep, serene dark palette due to shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static without any movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002817115979269147 + ], + "opposite_logprob_tokens": [ + -5.877817153930664 + ], + "logprob_reward": -0.002817115979269147, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002817115979269147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09842918813228607, + 0.07183515280485153, + 0.07703062891960144, + 0.05821482464671135, + 0.2890002727508545, + 0.05150391906499863, + 0.10136842727661133, + 0.08104068785905838, + 0.06466841697692871, + 0.10690850019454956 + ], + "credit_map": [ + 0.011143445037305355, + 0.011824451386928558, + 0.012605182826519012, + 0.010595332831144333, + 0.0066579654812812805, + 0.013435058295726776, + 0.014903065748512745, + 0.017264675348997116, + 0.010240038856863976, + 0.0072777122259140015, + 0.007949091494083405, + 0.006691449321806431, + 0.0053191413171589375, + 0.008349938318133354, + 0.015287903137505054, + 0.010719872079789639, + 0.013192174956202507, + 0.009880702011287212, + 0.009976662695407867, + 0.011547043919563293, + 0.005439423955976963, + 0.00525628449395299, + 0.010172955691814423, + 0.011565376073122025, + 0.0061067803762853146, + 0.007375585846602917, + 0.007985942997038364, + 0.008280003443360329, + 0.005187527742236853, + 0.003987353295087814, + 0.010161920450627804, + 0.009129705838859081, + 0.054763272404670715, + 0.058218732476234436, + 0.040381480008363724, + 0.03345021605491638, + 0.015369447879493237, + 0.02317395806312561, + 0.02744205854833126, + 0.03620109334588051, + 0.005106807220727205, + 0.0043284837156534195, + 0.008045547641813755, + 0.0067838383838534355, + 0.00758561072871089, + 0.005075161345303059, + 0.007506008725613356, + 0.007072455249726772, + 0.013544989749789238, + 0.009327265433967113, + 0.019174661487340927, + 0.012417776510119438, + 0.0059230513870716095, + 0.012373160570859909, + 0.015261377207934856, + 0.013346138410270214, + 0.011811112985014915, + 0.0077069890685379505, + 0.012424669228494167, + 0.008679633028805256, + 0.004981303587555885, + 0.014386487193405628, + 0.00961795262992382, + 0.011432540602982044, + 0.00779576925560832, + 0.007331604138016701, + 0.006275990046560764, + 0.004666178487241268, + 0.008009983226656914, + 0.009621729142963886, + 0.011678407900035381, + 0.009288749657571316, + 0.012141712009906769, + 0.012202220037579536, + 0.019275251775979996, + 0.011301346123218536, + 0.010055705904960632, + 0.015889372676610947, + 0.012045939452946186, + 0.013996950350701809 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 6 + ], + "selected_reward": -0.002817115979269147 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of weathered hands with visible knuckles and fine wrinkles visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are flat, emerald-green leaves resembling West Lake Longjing tea being twisted between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting strong sidelight coming from the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background extremely blurred with a deep, serene dark palette due to shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static without any movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20155610144138336 + ], + "opposite_logprob_tokens": [ + -1.7015560865402222 + ], + "logprob_reward": -0.20155610144138336, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20155610144138336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23356960713863373, + 0.08194465935230255, + 0.10742270946502686, + 0.061969056725502014, + 0.09781060367822647, + 0.10626063495874405, + 0.0449402779340744, + 0.09894774854183197, + 0.06742368638515472, + 0.09971103817224503 + ], + "credit_map": [ + 0.03140261024236679, + 0.020408207550644875, + 0.03571626916527748, + 0.03466397896409035, + 0.012300591915845871, + 0.025188153609633446, + 0.03670729324221611, + 0.03718249499797821, + 0.012534772977232933, + 0.009234672412276268, + 0.009510747157037258, + 0.0134279178455472, + 0.005503015127032995, + 0.008922222070395947, + 0.010174756869673729, + 0.012636552564799786, + 0.025964079424738884, + 0.017242318019270897, + 0.012509306892752647, + 0.014039617031812668, + 0.005617525894194841, + 0.005981359630823135, + 0.009883816353976727, + 0.016184691339731216, + 0.005622713826596737, + 0.005581223871558905, + 0.008997684344649315, + 0.007993705570697784, + 0.003746818983927369, + 0.004446730483323336, + 0.014084908179938793, + 0.011495273560285568, + 0.012373506091535091, + 0.01569710299372673, + 0.009459775872528553, + 0.012822610326111317, + 0.007863848470151424, + 0.008851950988173485, + 0.018591761589050293, + 0.012150043621659279, + 0.009716402739286423, + 0.009864709340035915, + 0.014986556023359299, + 0.010775645263493061, + 0.016137849539518356, + 0.014950569719076157, + 0.01432441920042038, + 0.015504486858844757, + 0.003705673385411501, + 0.0036489395424723625, + 0.0076014017686247826, + 0.0059637222439050674, + 0.005235180724412203, + 0.005975157022476196, + 0.005747065879404545, + 0.0070631359703838825, + 0.011691904626786709, + 0.008419048972427845, + 0.012902278453111649, + 0.010550431907176971, + 0.009914249181747437, + 0.020223861560225487, + 0.012985783629119396, + 0.012260186485946178, + 0.007467036135494709, + 0.006841019261628389, + 0.007614552974700928, + 0.006860243156552315, + 0.013893421739339828, + 0.009673775173723698, + 0.007347630336880684, + 0.007726006209850311, + 0.013146081939339638, + 0.011213406920433044, + 0.01167220901697874, + 0.009547887369990349, + 0.01294486504048109, + 0.013881606049835682, + 0.014337197877466679, + 0.012967784889042377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.20155610144138336 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunningly close-up, static shot, under Roger Deakins' signature strong sidelight, a pair of weathered hands delicately twist a few flat, emerald-green leaves of West Lake Longjing tea. This precise, focused light, cast from the left side of the frame, sculpts the contours of the knuckles and the fine wrinkles of the skin, while sharply illuminating the spiral veins of the fingertips. As the tea leaves are gently twisted between the fingers, light penetrates the thin leaves, revealing their veins and creating an almost translucent jade-like texture. The extremely shallow depth of field completely blurs the background, softening it into a deep, serene dark palette, focusing the entire visual focus on these experienced hands and the fresh leaves. The intense contrast of light and dark creates an atmosphere of concentration, tranquility, and ritual. Every subtle movement appears steady and deliberate, imbued with a reverence for nature.", + "seed": 25636, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_3/index_3667_seed_25636_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07048015878535807, + "vlm_logprob": -0.07048015878535807, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.07048015878535807, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of weathered hands with visible knuckles and fine wrinkles visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are flat, emerald-green leaves resembling West Lake Longjing tea being twisted between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting strong sidelight coming from the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background extremely blurred with a deep, serene dark palette due to shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static without any movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003128636162728071 + ], + "opposite_logprob_tokens": [ + -6.253128528594971 + ], + "logprob_reward": -0.003128636162728071, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.003128636162728071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13556551933288574, + 0.049410734325647354, + 0.03053983300924301, + 0.09258277714252472, + 0.08898896723985672, + 0.060001250356435776, + 0.029183821752667427, + 0.23790216445922852, + 0.24893788993358612, + 0.026887087151408195 + ], + "credit_map": [ + 0.022894177585840225, + 0.007524346467107534, + 0.01457310002297163, + 0.01955263502895832, + 0.01598622091114521, + 0.02191087044775486, + 0.016704170033335686, + 0.016419997438788414, + 0.004982431884855032, + 0.003324724268168211, + 0.006392243783921003, + 0.006701869890093803, + 0.006139812525361776, + 0.0065333545207977295, + 0.007374705281108618, + 0.007961592637002468, + 0.0019631634932011366, + 0.002727075247094035, + 0.0027116804849356413, + 0.0032860205974429846, + 0.005363999865949154, + 0.005418153014034033, + 0.005199878942221403, + 0.0038698615971952677, + 0.0063529363833367825, + 0.008511517196893692, + 0.00860908254981041, + 0.013960772193968296, + 0.014572720043361187, + 0.016642654314637184, + 0.01253298670053482, + 0.011400110088288784, + 0.006166914012283087, + 0.0093459477648139, + 0.007588833570480347, + 0.01255160290747881, + 0.014021298848092556, + 0.01564197428524494, + 0.01261597778648138, + 0.011056415736675262, + 0.003934355918318033, + 0.004080329090356827, + 0.007417510263621807, + 0.00862465426325798, + 0.009817223064601421, + 0.0075357574969530106, + 0.007862092927098274, + 0.01072932593524456, + 0.0019448912935331464, + 0.002020547864958644, + 0.006136848125606775, + 0.003486944828182459, + 0.002706740517169237, + 0.002586914459243417, + 0.003219813806936145, + 0.007081119809299707, + 0.013979139737784863, + 0.008870288729667664, + 0.07698354125022888, + 0.019875600934028625, + 0.012074775993824005, + 0.014946133829653263, + 0.014903642237186432, + 0.07626906037330627, + 0.016603806987404823, + 0.011999514885246754, + 0.08135613054037094, + 0.020415347069501877, + 0.009866866283118725, + 0.017095979303121567, + 0.017438629642128944, + 0.07416161894798279, + 0.002476262394338846, + 0.0028387662023305893, + 0.003206387860700488, + 0.003471849486231804, + 0.0034175270702689886, + 0.003900935873389244, + 0.003463462460786104, + 0.0041118948720395565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.003128636162728071 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of weathered hands with visible knuckles and fine wrinkles visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are flat, emerald-green leaves resembling West Lake Longjing tea being twisted between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting strong sidelight coming from the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background extremely blurred with a deep, serene dark palette due to shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static without any movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636430785059929 + ], + "opposite_logprob_tokens": [ + -3.651364326477051 + ], + "logprob_reward": -0.02636430785059929, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636430785059929, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09028123319149017, + 0.0904458686709404, + 0.05298112705349922, + 0.03016815334558487, + 0.04310552030801773, + 0.0347021147608757, + 0.037292007356882095, + 0.5242688655853271, + 0.07434830069541931, + 0.02240675315260887 + ], + "credit_map": [ + 0.0070162583142519, + 0.008533947169780731, + 0.011908894404768944, + 0.013543065637350082, + 0.010962056927382946, + 0.013677820563316345, + 0.011540348641574383, + 0.01309884898364544, + 0.008099726401269436, + 0.0062705883756279945, + 0.01299035269767046, + 0.02057579532265663, + 0.008907046169042587, + 0.008606869727373123, + 0.010493339039385319, + 0.014502151869237423, + 0.003988351672887802, + 0.0065514263696968555, + 0.005308265797793865, + 0.008032667450606823, + 0.007771143224090338, + 0.006593936122953892, + 0.00879366509616375, + 0.005941673647612333, + 0.0017977498937398195, + 0.0027703384403139353, + 0.0032672772649675608, + 0.004216872621327639, + 0.005874486640095711, + 0.004330421332269907, + 0.004112531431019306, + 0.003798478050157428, + 0.0030026272870600224, + 0.004349885042756796, + 0.0038974029012024403, + 0.006562096998095512, + 0.006454875692725182, + 0.0077866422943770885, + 0.005536191631108522, + 0.005515797529369593, + 0.0031650615856051445, + 0.00510929711163044, + 0.004053178709000349, + 0.005456720478832722, + 0.004469919949769974, + 0.0037138357292860746, + 0.004000500310212374, + 0.004733597859740257, + 0.0031858403235673904, + 0.00421930430456996, + 0.005409586243331432, + 0.005182382185012102, + 0.0034500746987760067, + 0.005094059277325869, + 0.004100539721548557, + 0.006650220137089491, + 0.030464498326182365, + 0.01652909256517887, + 0.16540755331516266, + 0.0435287281870842, + 0.03252573311328888, + 0.03661959245800972, + 0.03381307050585747, + 0.16538064181804657, + 0.004243355244398117, + 0.003357366193085909, + 0.024468975141644478, + 0.005802254192531109, + 0.0033094878308475018, + 0.004582101479172707, + 0.0044476063922047615, + 0.02413715235888958, + 0.0026537454687058926, + 0.003040337935090065, + 0.003694940824061632, + 0.002656603930518031, + 0.0019268745090812445, + 0.0025136510375887156, + 0.002606387482956052, + 0.003314211731776595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.02636430785059929 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of weathered hands with visible knuckles and fine wrinkles visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are flat, emerald-green leaves resembling West Lake Longjing tea being twisted between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting strong sidelight coming from the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background extremely blurred with a deep, serene dark palette due to shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static without any movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16036871075630188 + ], + "opposite_logprob_tokens": [ + -1.9103686809539795 + ], + "logprob_reward": -0.16036871075630188, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16036871075630188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06865223497152328, + 0.028264351189136505, + 0.03962292894721031, + 0.1847568303346634, + 0.26628875732421875, + 0.11371693760156631, + 0.053626466542482376, + 0.15759579837322235, + 0.0733071118593216, + 0.014168648049235344 + ], + "credit_map": [ + 0.008659280836582184, + 0.005543589126318693, + 0.008222714997828007, + 0.010743623599410057, + 0.007100860588252544, + 0.009798035956919193, + 0.008066048845648766, + 0.010518073104321957, + 0.0029151493217796087, + 0.0029662656597793102, + 0.0038103063125163317, + 0.00441812863573432, + 0.0028195285703986883, + 0.0029049385339021683, + 0.003506365465000272, + 0.004923665430396795, + 0.003135163802653551, + 0.003564652055501938, + 0.003938679583370686, + 0.005362049210816622, + 0.006320539396256208, + 0.005589394364506006, + 0.0064537664875388145, + 0.005258684512227774, + 0.01100952923297882, + 0.015048383735120296, + 0.018982958048582077, + 0.043209198862314224, + 0.021505726501345634, + 0.01760590262711048, + 0.03219400718808174, + 0.025201091542840004, + 0.01979302056133747, + 0.02354092337191105, + 0.025445682927966118, + 0.048910122364759445, + 0.036415908485651016, + 0.039716463536024094, + 0.036527104675769806, + 0.03593950346112251, + 0.007649907376617193, + 0.009983864612877369, + 0.010374114848673344, + 0.018789375200867653, + 0.020414667204022408, + 0.01501383911818266, + 0.016225384548306465, + 0.015265780501067638, + 0.003927320241928101, + 0.008171502500772476, + 0.005558487493544817, + 0.006764199584722519, + 0.00848415307700634, + 0.008226299658417702, + 0.007022470701485872, + 0.005472029559314251, + 0.010104937478899956, + 0.0041769868694245815, + 0.05741611123085022, + 0.012574239633977413, + 0.003972452599555254, + 0.007360407616943121, + 0.006868691183626652, + 0.055121976882219315, + 0.004500442184507847, + 0.0030784220434725285, + 0.025302201509475708, + 0.00574140902608633, + 0.002164724748581648, + 0.004127204418182373, + 0.004578418098390102, + 0.023814283311367035, + 0.0011332719586789608, + 0.0010600424138829112, + 0.0021782240364700556, + 0.0013902001082897186, + 0.001442166161723435, + 0.0017741179326549172, + 0.0016855206340551376, + 0.0035051037557423115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 7 + ], + "selected_reward": -0.16036871075630188 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of weathered hands with visible knuckles and fine wrinkles visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are flat, emerald-green leaves resembling West Lake Longjing tea being twisted between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting strong sidelight coming from the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background extremely blurred with a deep, serene dark palette due to shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static without any movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002196520334109664 + ], + "opposite_logprob_tokens": [ + -6.127196311950684 + ], + "logprob_reward": -0.002196520334109664, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002196520334109664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07878715544939041, + 0.05884355306625366, + 0.03391147404909134, + 0.1405133754014969, + 0.14570550620555878, + 0.06337139010429382, + 0.07278332114219666, + 0.2692841589450836, + 0.1149011105298996, + 0.021898949518799782 + ], + "credit_map": [ + 0.012516983784735203, + 0.006870793178677559, + 0.010057840496301651, + 0.013033025898039341, + 0.007897094823420048, + 0.008754769340157509, + 0.010600844398140907, + 0.00905579887330532, + 0.0045775556936860085, + 0.0054833292961120605, + 0.006299324333667755, + 0.007255914621055126, + 0.009034279733896255, + 0.008671019226312637, + 0.008631988428533077, + 0.008890138007700443, + 0.0021134349517524242, + 0.0035802556667476892, + 0.003430234966799617, + 0.005707563366740942, + 0.004236389417201281, + 0.004824010655283928, + 0.005700097419321537, + 0.004319484811276197, + 0.007134784944355488, + 0.0131128691136837, + 0.014946413226425648, + 0.028272753581404686, + 0.021626392379403114, + 0.01493606623262167, + 0.022447070106863976, + 0.018037011846899986, + 0.011254427954554558, + 0.017219366505742073, + 0.011809012852609158, + 0.022731153294444084, + 0.022705966606736183, + 0.02240210399031639, + 0.019721277058124542, + 0.017862191423773766, + 0.005635984241962433, + 0.004168510437011719, + 0.006107551045715809, + 0.009649377316236496, + 0.010926431976258755, + 0.00968858040869236, + 0.008510920219123363, + 0.00868403259664774, + 0.0064420378766953945, + 0.011916290037333965, + 0.008929425850510597, + 0.00996134988963604, + 0.0072023035027086735, + 0.011287632398307323, + 0.008054681122303009, + 0.008989593014121056, + 0.018334731459617615, + 0.007142563350498676, + 0.09790493547916412, + 0.02118506096303463, + 0.005219803657382727, + 0.013091671280562878, + 0.014087498188018799, + 0.09231790155172348, + 0.007169673684984446, + 0.008185633458197117, + 0.03319249674677849, + 0.009506943635642529, + 0.006512196268886328, + 0.008112614043056965, + 0.009079416282474995, + 0.033142127096652985, + 0.0019656678196042776, + 0.0021032991353422403, + 0.002413846319541335, + 0.0026258297730237246, + 0.002675802679732442, + 0.0035020492505282164, + 0.0035573297645896673, + 0.0030551222153007984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 3 + ], + "selected_reward": -0.002196520334109664 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of weathered hands with visible knuckles and fine wrinkles visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are flat, emerald-green leaves resembling West Lake Longjing tea being twisted between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting strong sidelight coming from the left side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background extremely blurred with a deep, serene dark palette due to shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static without any movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16034261882305145 + ], + "opposite_logprob_tokens": [ + -1.9103425741195679 + ], + "logprob_reward": -0.16034261882305145, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16034261882305145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12329037487506866, + 0.036852192133665085, + 0.03474665433168411, + 0.061220455914735794, + 0.11455776542425156, + 0.042315322905778885, + 0.08816586434841156, + 0.20521675050258636, + 0.24373042583465576, + 0.04990415275096893 + ], + "credit_map": [ + 0.007514937315136194, + 0.008350257761776447, + 0.013778544031083584, + 0.01663539558649063, + 0.01662874035537243, + 0.02041853405535221, + 0.01733534410595894, + 0.022628620266914368, + 0.004133664537221193, + 0.0034269248135387897, + 0.003520892933011055, + 0.005111974664032459, + 0.0049096145667135715, + 0.005175565369427204, + 0.006191703025251627, + 0.0043818531557917595, + 0.0023254812695086002, + 0.0037322931457310915, + 0.004258609376847744, + 0.004608967807143927, + 0.004160315729677677, + 0.005269284825772047, + 0.004935919772833586, + 0.005455780774354935, + 0.004430942703038454, + 0.00432509696111083, + 0.006242101080715656, + 0.00793750025331974, + 0.00995097029954195, + 0.010324308648705482, + 0.009619300253689289, + 0.008390238508582115, + 0.007427281700074673, + 0.014290177263319492, + 0.010631664656102657, + 0.016738593578338623, + 0.017463453114032745, + 0.018361771479249, + 0.01434409525245428, + 0.015300736762583256, + 0.004542158450931311, + 0.0036803483963012695, + 0.004876192193478346, + 0.008124299347400665, + 0.005933176260441542, + 0.0043010651133954525, + 0.005160147789865732, + 0.005697938613593578, + 0.006885773502290249, + 0.01186034083366394, + 0.009740767069160938, + 0.012559721246361732, + 0.01084539107978344, + 0.015002771280705929, + 0.012174037285149097, + 0.009097054600715637, + 0.02423769421875477, + 0.01307977456599474, + 0.020195091143250465, + 0.02680078148841858, + 0.0198327898979187, + 0.03446221351623535, + 0.03768717125058174, + 0.02892124466598034, + 0.014911429956555367, + 0.007916987873613834, + 0.08074963092803955, + 0.01786518469452858, + 0.010663151741027832, + 0.0156781654804945, + 0.013735027052462101, + 0.08221085369586945, + 0.004367177374660969, + 0.007620716001838446, + 0.006492147222161293, + 0.005320376716554165, + 0.004642019513994455, + 0.00585537776350975, + 0.0056342314928770065, + 0.009972105734050274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.16034261882305145 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_4/index_tensor([3469, 3469])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_4/index_tensor([3469, 3469])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a8c77a75893e51aa12f4c2c96e88837cf659c75e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_4/index_tensor([3469, 3469])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning close-up, captured from a level angle, the camera held still, focuses on a vibrant red spider lily (Banana lily) swaying gently in the breeze. Its velvety scarlet petals curl gracefully back like ribbons, revealing dozens of slender, curving stamens radiating from the center. At the tip of each stamen hangs a shimmering, translucent dewdrop, shimmering like tiny diamonds in the soft, diffuse morning light. An imperceptible draft of air causes the entire flower to vibrate in a slow, rhythmic tremor, gently swaying the dewdrops on the stamens, capturing and refracting fleeting light. The incredibly shallow depth of field completely blurs the background, transforming it into a dreamy, soft patch of light composed of varying shades of green and earth tones, placing all visual focus on the flower's vitality. The entire picture is rich and saturated in color, creating a tranquil, focused and poetic atmosphere, as if time has frozen at this moment, just to show this brief and extreme beauty.", + "seed": 12729, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_4/index_3469_seed_12729_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03776740189641714, + "vlm_logprob": -0.03776740189641714, + "vlm_yes_no_margin": 3.774999976158142, + "vlm_token_credit": -0.03776740189641714, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red spider lily (Banana lily)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flower's petals velvety, scarlet, and curling back like ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there dozens of slender, curving stamens with shimmering translucent dewdrops at their tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held still at a level angle with a shallow depth of field that blurs the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the flower swaying gently in the breeze while the dewdrops vibrate in a slow, rhythmic tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022293316200375557 + ], + "opposite_logprob_tokens": [ + -3.8972933292388916 + ], + "logprob_reward": -0.022293316200375557, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.022293316200375557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2122744619846344, + 0.1308167278766632, + 0.2035800814628601, + 0.0910482257604599, + 0.024425098672509193, + 0.07216627895832062, + 0.038402389734983444, + 0.0916908010840416, + 0.06435557454824448, + 0.07124034315347672 + ], + "credit_map": [ + 0.0413484089076519, + 0.020314624533057213, + 0.03918365761637688, + 0.021429937332868576, + 0.022752530872821808, + 0.023978153243660927, + 0.016948264092206955, + 0.026318883523344994, + 0.013736879453063011, + 0.008289691992104053, + 0.03357735648751259, + 0.008894385769963264, + 0.013368578627705574, + 0.01257073599845171, + 0.008249957114458084, + 0.03212914243340492, + 0.031604327261447906, + 0.006243049167096615, + 0.07646116614341736, + 0.008405386470258236, + 0.006712664384394884, + 0.010943513363599777, + 0.006644033826887608, + 0.05656593292951584, + 0.01549938227981329, + 0.002089231973513961, + 0.03611592575907707, + 0.0035063624382019043, + 0.0026239214930683374, + 0.002871184842661023, + 0.0025199935771524906, + 0.025822225958108902, + 0.0036233454011380672, + 0.0020022382959723473, + 0.005549508612602949, + 0.002060655038803816, + 0.0027289672289043665, + 0.0023921846877783537, + 0.0018830950139090419, + 0.004185104742646217, + 0.005914899520576, + 0.00584468012675643, + 0.008967495523393154, + 0.0073007214814424515, + 0.012538059614598751, + 0.01075182855129242, + 0.012009691447019577, + 0.008838905952870846, + 0.004544392693787813, + 0.005747491028159857, + 0.006327440496534109, + 0.0038115675561130047, + 0.004476370755583048, + 0.003272679401561618, + 0.003909946419298649, + 0.006312501151114702, + 0.012607808224856853, + 0.010302241891622543, + 0.013940554112195969, + 0.01005422044545412, + 0.014465943910181522, + 0.009465723298490047, + 0.00831435713917017, + 0.012539951130747795, + 0.005506434012204409, + 0.008769609034061432, + 0.010011672973632812, + 0.010070282965898514, + 0.007780325599014759, + 0.005130574572831392, + 0.007323637139052153, + 0.009763037785887718, + 0.009653151966631413, + 0.011091378517448902, + 0.011688007973134518, + 0.005768804345279932, + 0.009066489525139332, + 0.007981478236615658, + 0.007167236879467964, + 0.00882379524409771 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.022293316200375557 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red spider lily (Banana lily)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flower's petals velvety, scarlet, and curling back like ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there dozens of slender, curving stamens with shimmering translucent dewdrops at their tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held still at a level angle with a shallow depth of field that blurs the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the flower swaying gently in the breeze while the dewdrops vibrate in a slow, rhythmic tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011099735274910927 + ], + "opposite_logprob_tokens": [ + -4.511099815368652 + ], + "logprob_reward": -0.011099735274910927, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011099735274910927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10837434977293015, + 0.12438887357711792, + 0.31052225828170776, + 0.1210581511259079, + 0.05375325307250023, + 0.02426021918654442, + 0.021613480523228645, + 0.04210193455219269, + 0.12092486768960953, + 0.07300268113613129 + ], + "credit_map": [ + 0.02663384936749935, + 0.009745276533067226, + 0.019753126427531242, + 0.010456872172653675, + 0.008508330211043358, + 0.010516571812331676, + 0.007859417237341404, + 0.01490091159939766, + 0.019530627876520157, + 0.004597390070557594, + 0.04475279897451401, + 0.005874299444258213, + 0.004134111572057009, + 0.00667343195527792, + 0.004740780219435692, + 0.03408542647957802, + 0.05827532708644867, + 0.00930796004831791, + 0.12147828191518784, + 0.011582547798752785, + 0.010109882801771164, + 0.012152685783803463, + 0.00974434893578291, + 0.07787121087312698, + 0.022177735343575478, + 0.0025141851510852575, + 0.04751495271921158, + 0.004589889198541641, + 0.002905031433328986, + 0.003938570618629456, + 0.003266793442890048, + 0.03415099158883095, + 0.004797898232936859, + 0.003677989123389125, + 0.013161065988242626, + 0.003907349891960621, + 0.006048485636711121, + 0.005442753434181213, + 0.006869907025247812, + 0.00984780490398407, + 0.003031464060768485, + 0.001959142042323947, + 0.004515268839895725, + 0.002855248050764203, + 0.0034803994931280613, + 0.0023941064719110727, + 0.002492726081982255, + 0.0035318643786013126, + 0.002295388840138912, + 0.0028456482104957104, + 0.0035801734775304794, + 0.0026060142554342747, + 0.0025821325834840536, + 0.0027763957623392344, + 0.0021226219832897186, + 0.0028051051776856184, + 0.003484587650746107, + 0.00814114697277546, + 0.004778342787176371, + 0.003622027114033699, + 0.007513054180890322, + 0.005090905353426933, + 0.00482795387506485, + 0.004643916618078947, + 0.011934111826121807, + 0.009011575020849705, + 0.023918962106108665, + 0.021513892337679863, + 0.008637151680886745, + 0.008794127032160759, + 0.012636933475732803, + 0.02447810024023056, + 0.009262443520128727, + 0.011467507109045982, + 0.011103203520178795, + 0.006437345873564482, + 0.009130305610597134, + 0.009726564399898052, + 0.00736404862254858, + 0.008511263877153397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.011099735274910927 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red spider lily (Banana lily)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flower's petals velvety, scarlet, and curling back like ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there dozens of slender, curving stamens with shimmering translucent dewdrops at their tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held still at a level angle with a shallow depth of field that blurs the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the flower swaying gently in the breeze while the dewdrops vibrate in a slow, rhythmic tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014207170344889164 + ], + "opposite_logprob_tokens": [ + -4.264207363128662 + ], + "logprob_reward": -0.014207170344889164, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014207170344889164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09168680012226105, + 0.1505792737007141, + 0.07558626681566238, + 0.14629516005516052, + 0.025382686406373978, + 0.018530307337641716, + 0.03084344044327736, + 0.054060157388448715, + 0.35940006375312805, + 0.04763582721352577 + ], + "credit_map": [ + 0.016566146165132523, + 0.008800985291600227, + 0.01688324846327305, + 0.007484361995011568, + 0.011125409044325352, + 0.010795452632009983, + 0.009103608317673206, + 0.010927592404186726, + 0.02144555188715458, + 0.009272700175642967, + 0.04730129614472389, + 0.008684013970196247, + 0.009883149527013302, + 0.01122693158686161, + 0.006771962624043226, + 0.03599366918206215, + 0.01196476723998785, + 0.006446409039199352, + 0.01684444583952427, + 0.005075428169220686, + 0.005559978540986776, + 0.009168097749352455, + 0.004238103982061148, + 0.016289031133055687, + 0.023403365164995193, + 0.002790414495393634, + 0.05946291610598564, + 0.005622150842100382, + 0.0033326363191008568, + 0.004573110956698656, + 0.003716390812769532, + 0.04339417442679405, + 0.0024219185579568148, + 0.002810549223795533, + 0.0034905392676591873, + 0.0026110021863132715, + 0.004868954420089722, + 0.0026868884451687336, + 0.00353707792237401, + 0.0029557575471699238, + 0.002889009891077876, + 0.0017775010783225298, + 0.0030101733282208443, + 0.0014894303167238832, + 0.0019442398333922029, + 0.0026880500372499228, + 0.0024128067307174206, + 0.002319097053259611, + 0.003234350588172674, + 0.002538479631766677, + 0.004636537283658981, + 0.0049086627550423145, + 0.002756798639893532, + 0.004574512131512165, + 0.0043715620413422585, + 0.0038225387688726187, + 0.007050772197544575, + 0.006491443142294884, + 0.007861189544200897, + 0.006644515320658684, + 0.006041103973984718, + 0.004609048832207918, + 0.008283928968012333, + 0.0070781526155769825, + 0.041954003274440765, + 0.02099352329969406, + 0.07859437167644501, + 0.05955824628472328, + 0.020366588607430458, + 0.027506377547979355, + 0.0297390203922987, + 0.08068794012069702, + 0.0067090122029185295, + 0.00657792529091239, + 0.007628221530467272, + 0.004720550496131182, + 0.005234831012785435, + 0.006605925969779491, + 0.004554058890789747, + 0.005605300422757864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 3 + ], + "selected_reward": -0.014207170344889164 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red spider lily (Banana lily)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flower's petals velvety, scarlet, and curling back like ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there dozens of slender, curving stamens with shimmering translucent dewdrops at their tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held still at a level angle with a shallow depth of field that blurs the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the flower swaying gently in the breeze while the dewdrops vibrate in a slow, rhythmic tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014200588688254356 + ], + "opposite_logprob_tokens": [ + -4.264200687408447 + ], + "logprob_reward": -0.014200588688254356, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014200588688254356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09475827217102051, + 0.05637675151228905, + 0.16354571282863617, + 0.24086135625839233, + 0.12247687578201294, + 0.0452563501894474, + 0.07695522904396057, + 0.09090627729892731, + 0.04978357255458832, + 0.05907958373427391 + ], + "credit_map": [ + 0.02458515763282776, + 0.004836035892367363, + 0.02784695103764534, + 0.00555409025400877, + 0.007683681324124336, + 0.005361215211451054, + 0.006719054188579321, + 0.01217208057641983, + 0.014987500384449959, + 0.0033515929244458675, + 0.015280828811228275, + 0.003367280587553978, + 0.004416418727487326, + 0.00459499703720212, + 0.0026472611352801323, + 0.007730872370302677, + 0.02556784078478813, + 0.004215022549033165, + 0.06120496243238449, + 0.006350736599415541, + 0.004620148800313473, + 0.006661136168986559, + 0.005112484097480774, + 0.04981338977813721, + 0.0371546745300293, + 0.004453021101653576, + 0.0996677353978157, + 0.008415634743869305, + 0.004681561607867479, + 0.00609868299216032, + 0.005572938825935125, + 0.07481712102890015, + 0.01742306910455227, + 0.006000800523906946, + 0.03562923148274422, + 0.006245646625757217, + 0.00887619610875845, + 0.005856836214661598, + 0.009285780601203442, + 0.03315931558609009, + 0.004666105844080448, + 0.0028934406582266092, + 0.00768744433298707, + 0.005630565341562033, + 0.004512612707912922, + 0.0055947876535356045, + 0.008000003173947334, + 0.006271385587751865, + 0.008107447996735573, + 0.0072663999162614346, + 0.012408609502017498, + 0.00711254496127367, + 0.008159554563462734, + 0.012958750128746033, + 0.010818340815603733, + 0.010123582556843758, + 0.008864442817866802, + 0.009979807771742344, + 0.01695915311574936, + 0.00693547772243619, + 0.01112381462007761, + 0.00770958187058568, + 0.011129318736493587, + 0.01820467971265316, + 0.006676735822111368, + 0.0076760281808674335, + 0.008279402740299702, + 0.004352056421339512, + 0.007455947808921337, + 0.00380146992392838, + 0.004233705345541239, + 0.007308228872716427, + 0.007121144328266382, + 0.0077060917392373085, + 0.008501851931214333, + 0.00670631742104888, + 0.0074673970229923725, + 0.006144919898360968, + 0.007789746858179569, + 0.007642112672328949 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 4 + ], + "selected_reward": -0.014200588688254356 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red spider lily (Banana lily)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flower's petals velvety, scarlet, and curling back like ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there dozens of slender, curving stamens with shimmering translucent dewdrops at their tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held still at a level angle with a shallow depth of field that blurs the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the flower swaying gently in the breeze while the dewdrops vibrate in a slow, rhythmic tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270361989736557 + ], + "opposite_logprob_tokens": [ + -2.1270360946655273 + ], + "logprob_reward": -0.1270361989736557, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1270361989736557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03466015309095383, + 0.0654132142663002, + 0.07941491901874542, + 0.47151312232017517, + 0.06880775839090347, + 0.029678842052817345, + 0.05021032691001892, + 0.06243930384516716, + 0.08394034206867218, + 0.05392208322882652 + ], + "credit_map": [ + 0.008162040263414383, + 0.0029959857929497957, + 0.006320328917354345, + 0.002854946767911315, + 0.003334740409627557, + 0.0034656114876270294, + 0.0029751521069556475, + 0.004551345016807318, + 0.00939954724162817, + 0.0034346911124885082, + 0.02033006027340889, + 0.0036465327721089125, + 0.003166758455336094, + 0.004921669606119394, + 0.0026495754718780518, + 0.017864374443888664, + 0.011197264306247234, + 0.0023960229009389877, + 0.030612923204898834, + 0.003127068979665637, + 0.002782427240163088, + 0.0033492292277514935, + 0.002731695305556059, + 0.02321827970445156, + 0.07888466119766235, + 0.008333439938724041, + 0.19539187848567963, + 0.015535131096839905, + 0.009576814249157906, + 0.011455077677965164, + 0.0095905726775527, + 0.1427454799413681, + 0.006232555489987135, + 0.007560525555163622, + 0.008041059598326683, + 0.006802618969231844, + 0.012524589896202087, + 0.00854256097227335, + 0.011057062074542046, + 0.008046776987612247, + 0.004565829411149025, + 0.0020639458671212196, + 0.006085759960114956, + 0.002985689789056778, + 0.0022835489362478256, + 0.002806826727464795, + 0.0026392515283077955, + 0.006247987970709801, + 0.005565503612160683, + 0.006609209347516298, + 0.007925364188849926, + 0.005659349262714386, + 0.006597249768674374, + 0.006857278756797314, + 0.004684717860072851, + 0.006311648525297642, + 0.006938681472092867, + 0.007346724160015583, + 0.009322692640125751, + 0.006544116418808699, + 0.007367059122771025, + 0.006026225630193949, + 0.009264631196856499, + 0.00962916947901249, + 0.00802364107221365, + 0.007327703759074211, + 0.015468482859432697, + 0.015226666815578938, + 0.0065641324035823345, + 0.006071779411286116, + 0.009369798004627228, + 0.01588812842965126, + 0.007745198905467987, + 0.007416771724820137, + 0.008537597954273224, + 0.005331642925739288, + 0.006251407787203789, + 0.006745888851583004, + 0.0054964544251561165, + 0.0063971164636313915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 2 + ], + "selected_reward": -0.1270361989736557 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up, captured from a level angle, the camera held still, focuses on a vibrant red spider lily (Banana lily) swaying gently in the breeze. Its velvety scarlet petals curl gracefully back like ribbons, revealing dozens of slender, curving stamens radiating from the center. At the tip of each stamen hangs a shimmering, translucent dewdrop, shimmering like tiny diamonds in the soft, diffuse morning light. An imperceptible draft of air causes the entire flower to vibrate in a slow, rhythmic tremor, gently swaying the dewdrops on the stamens, capturing and refracting fleeting light. The incredibly shallow depth of field completely blurs the background, transforming it into a dreamy, soft patch of light composed of varying shades of green and earth tones, placing all visual focus on the flower's vitality. The entire picture is rich and saturated in color, creating a tranquil, focused and poetic atmosphere, as if time has frozen at this moment, just to show this brief and extreme beauty.", + "seed": 12730, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_4/index_3469_seed_12730_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05205406993627548, + "vlm_logprob": -0.05205406993627548, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.05205406993627548, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red spider lily (Banana lily)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flower's petals velvety, scarlet, and curling back like ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there dozens of slender, curving stamens with shimmering translucent dewdrops at their tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held still at a level angle with a shallow depth of field that blurs the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the flower swaying gently in the breeze while the dewdrops vibrate in a slow, rhythmic tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008354594931006432 + ], + "opposite_logprob_tokens": [ + -4.883354663848877 + ], + "logprob_reward": -0.008354594931006432, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.008354594931006432, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.023523278534412384, + 0.012047883123159409, + 0.013809246942400932, + 0.006761716678738594, + 0.01360953226685524, + 0.02694578468799591, + 0.021252358332276344, + 0.016958925873041153, + 0.6669260859489441, + 0.19816510379314423 + ], + "credit_map": [ + 0.004763834178447723, + 0.0016170883318409324, + 0.003991232253611088, + 0.0024152984842658043, + 0.0013584209373220801, + 0.0017792372964322567, + 0.0018847581231966615, + 0.005713413469493389, + 0.001541414763778448, + 0.0010676240781322122, + 0.0024055466055870056, + 0.0010009091347455978, + 0.0019556942861527205, + 0.0011042223777621984, + 0.0008693572599440813, + 0.002103117061778903, + 0.0014782976359128952, + 0.0013175393687561154, + 0.002812757855281234, + 0.0009917650604620576, + 0.002312423661351204, + 0.0014563992153853178, + 0.0010938229970633984, + 0.0023462430108338594, + 0.0009223590604960918, + 0.0007616037037223577, + 0.0011275691213086247, + 0.0008487243903800845, + 0.000660354329738766, + 0.0007178015657700598, + 0.0007493281154893339, + 0.0009739778470247984, + 0.0016901448834687471, + 0.0011420603841543198, + 0.0019681628327816725, + 0.0015336006181314588, + 0.002453983761370182, + 0.0015584933571517467, + 0.001788993482477963, + 0.0014740951592102647, + 0.0032559363171458244, + 0.002058945829048753, + 0.0037888633087277412, + 0.003213834948837757, + 0.003426373703405261, + 0.0029266600031405687, + 0.0032821400091052055, + 0.0049930354580283165, + 0.002408236963674426, + 0.001978805521503091, + 0.003711816854774952, + 0.0027424388099461794, + 0.0027225881349295378, + 0.0031425082124769688, + 0.002567536197602749, + 0.001978430198505521, + 0.002565657952800393, + 0.0014692292315885425, + 0.0032995399087667465, + 0.0017440185183659196, + 0.002195315668359399, + 0.0023057276848703623, + 0.0015930836088955402, + 0.0017863558605313301, + 0.08567710220813751, + 0.03012010082602501, + 0.20518869161605835, + 0.05985374003648758, + 0.052067190408706665, + 0.03687911480665207, + 0.06294587254524231, + 0.13419437408447266, + 0.03206751123070717, + 0.004298975225538015, + 0.07656478881835938, + 0.011193935759365559, + 0.005267745815217495, + 0.0069559551775455475, + 0.009463378228247166, + 0.052352845668792725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.008354594931006432 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red spider lily (Banana lily)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flower's petals velvety, scarlet, and curling back like ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there dozens of slender, curving stamens with shimmering translucent dewdrops at their tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held still at a level angle with a shallow depth of field that blurs the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the flower swaying gently in the breeze while the dewdrops vibrate in a slow, rhythmic tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018211886286735535 + ], + "opposite_logprob_tokens": [ + -4.018211841583252 + ], + "logprob_reward": -0.018211886286735535, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018211886286735535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046970173716545105, + 0.015487466007471085, + 0.01925523951649666, + 0.010686192661523819, + 0.022195560857653618, + 0.033673547208309174, + 0.06626570969820023, + 0.07439567148685455, + 0.5034369230270386, + 0.20763354003429413 + ], + "credit_map": [ + 0.008278082124888897, + 0.003096587723121047, + 0.00901508703827858, + 0.004299377091228962, + 0.002907219808548689, + 0.0029346912633627653, + 0.0036432037595659494, + 0.012795927003026009, + 0.0018079153960570693, + 0.0008622343884781003, + 0.004027700517326593, + 0.0009924252517521381, + 0.0010730919893831015, + 0.0012628730619326234, + 0.0013316817348822951, + 0.004129542503505945, + 0.0019910172559320927, + 0.0017580067506060004, + 0.0034146648831665516, + 0.0016130841104313731, + 0.003189742797985673, + 0.0024847760796546936, + 0.0018087952630594373, + 0.002995152957737446, + 0.0016679257387295365, + 0.0012528708903118968, + 0.0016854575369507074, + 0.0012322403490543365, + 0.0007411282858811319, + 0.001126595539972186, + 0.0012221044162288308, + 0.0017578703118488193, + 0.0026000961661338806, + 0.002062245039269328, + 0.0033533296082168818, + 0.0032511595636606216, + 0.0027638438623398542, + 0.0022355651017278433, + 0.002510802587494254, + 0.003418518230319023, + 0.00394764170050621, + 0.0020659922156482935, + 0.006237702444195747, + 0.00338377826847136, + 0.004181503783911467, + 0.0034937106538563967, + 0.0053339675068855286, + 0.005029249005019665, + 0.008801179938018322, + 0.0039574154652655125, + 0.017494898289442062, + 0.006914076860994101, + 0.005072041880339384, + 0.007044410798698664, + 0.007135321851819754, + 0.009846359491348267, + 0.009288511238992214, + 0.00490031810477376, + 0.01966889388859272, + 0.00523785687983036, + 0.011178886517882347, + 0.007005991879850626, + 0.0065767159685492516, + 0.010538498871028423, + 0.0675620585680008, + 0.01994405686855316, + 0.16213467717170715, + 0.04414907097816467, + 0.03343113884329796, + 0.026105836033821106, + 0.04392954334616661, + 0.10618049651384354, + 0.033322274684906006, + 0.004914990626275539, + 0.07886844128370285, + 0.011649329215288162, + 0.006985082291066647, + 0.007662671618163586, + 0.010730491019785404, + 0.053500253707170486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.018211886286735535 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red spider lily (Banana lily)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flower's petals velvety, scarlet, and curling back like ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there dozens of slender, curving stamens with shimmering translucent dewdrops at their tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held still at a level angle with a shallow depth of field that blurs the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the flower swaying gently in the breeze while the dewdrops vibrate in a slow, rhythmic tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016078105196356773 + ], + "opposite_logprob_tokens": [ + -4.141077995300293 + ], + "logprob_reward": -0.016078105196356773, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016078105196356773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04403901472687721, + 0.018789930269122124, + 0.013414029963314533, + 0.009133442305028439, + 0.028996482491493225, + 0.045668162405490875, + 0.056508153676986694, + 0.035925596952438354, + 0.34380683302879333, + 0.4037182629108429 + ], + "credit_map": [ + 0.008905932307243347, + 0.0027307006530463696, + 0.008663415908813477, + 0.0039419629611074924, + 0.0020714541897177696, + 0.0026719600427895784, + 0.0031357179395854473, + 0.01191787887364626, + 0.0018239771015942097, + 0.0019405308412387967, + 0.0032858888152986765, + 0.0016139073995873332, + 0.0034109812695533037, + 0.001962914364412427, + 0.001512866117991507, + 0.0032388654071837664, + 0.0015156720764935017, + 0.001297449111007154, + 0.0023753775749355555, + 0.0010612660553306341, + 0.002122224308550358, + 0.001764166634529829, + 0.0011759265325963497, + 0.0021019489504396915, + 0.001483964268118143, + 0.0007511303410865366, + 0.001624536351300776, + 0.0011215502163395286, + 0.0008023206028155982, + 0.0013382673496380448, + 0.0008179043652489781, + 0.0011937698582187295, + 0.002896466525271535, + 0.00281531666405499, + 0.003917977213859558, + 0.004318774677813053, + 0.0036376803182065487, + 0.0038994343485683203, + 0.004420737270265818, + 0.003090099198743701, + 0.004594655707478523, + 0.0032048390712589025, + 0.007520233280956745, + 0.004501475021243095, + 0.004898244049400091, + 0.007158289197832346, + 0.00737739959731698, + 0.0064130318351089954, + 0.005275423638522625, + 0.005490689538419247, + 0.0077632819302380085, + 0.007681591436266899, + 0.006338538136333227, + 0.007125587202608585, + 0.007412693928927183, + 0.009420354850590229, + 0.004922454711049795, + 0.0020713028497993946, + 0.009697147645056248, + 0.0033854690846055746, + 0.0043961843475699425, + 0.002993559930473566, + 0.003086841432377696, + 0.005372640211135149, + 0.046571776270866394, + 0.01299748569726944, + 0.11248788982629776, + 0.02930755913257599, + 0.02226894162595272, + 0.017621930688619614, + 0.028786130249500275, + 0.07376515865325928, + 0.06307021528482437, + 0.007998360320925713, + 0.156661257147789, + 0.02324640192091465, + 0.00810202956199646, + 0.01431246753782034, + 0.018472857773303986, + 0.11185470223426819 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.016078105196356773 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red spider lily (Banana lily)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flower's petals velvety, scarlet, and curling back like ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there dozens of slender, curving stamens with shimmering translucent dewdrops at their tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held still at a level angle with a shallow depth of field that blurs the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the flower swaying gently in the breeze while the dewdrops vibrate in a slow, rhythmic tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016068484634160995 + ], + "opposite_logprob_tokens": [ + -4.141068458557129 + ], + "logprob_reward": -0.016068484634160995, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016068484634160995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05488152056932449, + 0.017639780417084694, + 0.016710378229618073, + 0.014452355913817883, + 0.021193254739046097, + 0.03319977968931198, + 0.050509124994277954, + 0.06820500642061234, + 0.24291367828845978, + 0.4802950918674469 + ], + "credit_map": [ + 0.007323201280087233, + 0.004202715586870909, + 0.01378975436091423, + 0.0066502392292022705, + 0.004425015300512314, + 0.0057950266636908054, + 0.004269011784344912, + 0.00842655822634697, + 0.0019958983175456524, + 0.001450236770324409, + 0.0038860891945660114, + 0.0014582184376195073, + 0.001763858599588275, + 0.0015671547735109925, + 0.0015186461387202144, + 0.00399967934936285, + 0.001977456035092473, + 0.0016274641966447234, + 0.0024166482035070658, + 0.002137857722118497, + 0.0029280888848006725, + 0.0016686632297933102, + 0.0017366762040182948, + 0.002217525150626898, + 0.002600543200969696, + 0.0014305805088952184, + 0.0023592703510075808, + 0.0018866303144022822, + 0.00124041095841676, + 0.0018604984506964684, + 0.0014844657853245735, + 0.001589957275427878, + 0.003539217635989189, + 0.0020106974989175797, + 0.002848041709512472, + 0.003329469356685877, + 0.002333316719159484, + 0.002369175897911191, + 0.001922953873872757, + 0.002840383443981409, + 0.004149128682911396, + 0.0027491094078868628, + 0.005945183802396059, + 0.0037223363760858774, + 0.004466033540666103, + 0.004299029242247343, + 0.00320247164927423, + 0.004666488151997328, + 0.004923663567751646, + 0.004677044227719307, + 0.005688531789928675, + 0.006375657394528389, + 0.006967138033360243, + 0.009234337136149406, + 0.006546188611537218, + 0.006096563767641783, + 0.005691258702427149, + 0.007365013938397169, + 0.011453704908490181, + 0.0058443862944841385, + 0.014749381691217422, + 0.009199732914566994, + 0.006600071210414171, + 0.007301465608179569, + 0.035696759819984436, + 0.005884016864001751, + 0.08832420408725739, + 0.014646120369434357, + 0.005606717895716429, + 0.009145903401076794, + 0.017041100189089775, + 0.06656887382268906, + 0.07666195183992386, + 0.009434499777853489, + 0.1871606856584549, + 0.02813103422522545, + 0.010494348593056202, + 0.017757659777998924, + 0.021563110873103142, + 0.1290917992591858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.016068484634160995 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red spider lily (Banana lily)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flower's petals velvety, scarlet, and curling back like ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there dozens of slender, curving stamens with shimmering translucent dewdrops at their tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held still at a level angle with a shallow depth of field that blurs the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the flower swaying gently in the breeze while the dewdrops vibrate in a slow, rhythmic tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20155727863311768 + ], + "opposite_logprob_tokens": [ + -1.7015572786331177 + ], + "logprob_reward": -0.20155727863311768, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20155727863311768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08470579981803894, + 0.017012372612953186, + 0.01239821407943964, + 0.018671732395887375, + 0.028077956289052963, + 0.044721513986587524, + 0.041567444801330566, + 0.038499001413583755, + 0.4198721647262573, + 0.29447370767593384 + ], + "credit_map": [ + 0.021651845425367355, + 0.005094871856272221, + 0.018152335658669472, + 0.008374850265681744, + 0.0044820429757237434, + 0.00603090226650238, + 0.00589261157438159, + 0.015026346780359745, + 0.0019219544483348727, + 0.0014269916573539376, + 0.003426539245992899, + 0.0013384028570726514, + 0.0021776168141514063, + 0.0012549902312457561, + 0.001605394994840026, + 0.0038604827132076025, + 0.0017459244700148702, + 0.0010700112907215953, + 0.0022463181521743536, + 0.0014412994496524334, + 0.0010641571134328842, + 0.0010764586040750146, + 0.0014444541884586215, + 0.002309591742232442, + 0.0022685059811919928, + 0.0014849005965515971, + 0.003395659849047661, + 0.0021584052592515945, + 0.00258759967982769, + 0.002930226270109415, + 0.00161945098079741, + 0.00222698412835598, + 0.0031134539749473333, + 0.0021825425792485476, + 0.004435369744896889, + 0.0033917787950485945, + 0.003878757357597351, + 0.002859528874978423, + 0.003136150073260069, + 0.005080378148704767, + 0.0075171273201704025, + 0.0022845130879431963, + 0.011677243746817112, + 0.004096283111721277, + 0.004509781487286091, + 0.003181702923029661, + 0.004979460500180721, + 0.006475407164543867, + 0.0066087921150028706, + 0.0024076481349766254, + 0.01159727480262518, + 0.004377732984721661, + 0.0048050303012132645, + 0.0033123581670224667, + 0.003654711414128542, + 0.004803902003914118, + 0.003168924245983362, + 0.004526520147919655, + 0.00477248802781105, + 0.003380361944437027, + 0.008525094948709011, + 0.005535611882805824, + 0.005107569973915815, + 0.00348243722692132, + 0.05757816880941391, + 0.015247218310832977, + 0.1400715708732605, + 0.03472944721579552, + 0.02642485871911049, + 0.020937031134963036, + 0.03162909299135208, + 0.09325482696294785, + 0.04351992905139923, + 0.005958480294793844, + 0.11215909570455551, + 0.017033414915204048, + 0.007028612773865461, + 0.011215286329388618, + 0.014114558696746826, + 0.08344437927007675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.20155727863311768 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up, captured from a level angle, the camera held still, focuses on a vibrant red spider lily (Banana lily) swaying gently in the breeze. Its velvety scarlet petals curl gracefully back like ribbons, revealing dozens of slender, curving stamens radiating from the center. At the tip of each stamen hangs a shimmering, translucent dewdrop, shimmering like tiny diamonds in the soft, diffuse morning light. An imperceptible draft of air causes the entire flower to vibrate in a slow, rhythmic tremor, gently swaying the dewdrops on the stamens, capturing and refracting fleeting light. The incredibly shallow depth of field completely blurs the background, transforming it into a dreamy, soft patch of light composed of varying shades of green and earth tones, placing all visual focus on the flower's vitality. The entire picture is rich and saturated in color, creating a tranquil, focused and poetic atmosphere, as if time has frozen at this moment, just to show this brief and extreme beauty.", + "seed": 12731, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_4/index_3469_seed_12731_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06447945442050695, + "vlm_logprob": -0.06447945442050695, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.06447945442050695, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red spider lily (Banana lily)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flower's petals velvety, scarlet, and curling back like ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there dozens of slender, curving stamens with shimmering translucent dewdrops at their tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held still at a level angle with a shallow depth of field that blurs the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the flower swaying gently in the breeze while the dewdrops vibrate in a slow, rhythmic tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019701508805155754 + ], + "opposite_logprob_tokens": [ + -4.0197014808654785 + ], + "logprob_reward": -0.019701508805155754, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.019701508805155754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3525461256504059, + 0.09860904514789581, + 0.06896735727787018, + 0.08705156296491623, + 0.04080775007605553, + 0.046664223074913025, + 0.07217858731746674, + 0.037928298115730286, + 0.17525553703308105, + 0.019991522654891014 + ], + "credit_map": [ + 0.0639614462852478, + 0.04459359496831894, + 0.050156909972429276, + 0.04349681735038757, + 0.03654681518673897, + 0.034771475940942764, + 0.046193961054086685, + 0.03282508999109268, + 0.015267780050635338, + 0.010401390492916107, + 0.013629510067403316, + 0.017322178930044174, + 0.0109736118465662, + 0.006040636915713549, + 0.01535044051706791, + 0.00962350144982338, + 0.012168572284281254, + 0.00933146383613348, + 0.009603452868759632, + 0.0077364263124763966, + 0.008494159206748009, + 0.007300521247088909, + 0.007817219011485577, + 0.0065155355259776115, + 0.013383091427385807, + 0.0118453623726964, + 0.013610685244202614, + 0.008753187023103237, + 0.011966788209974766, + 0.010122653096914291, + 0.008362086489796638, + 0.009007710963487625, + 0.005551802925765514, + 0.00556101743131876, + 0.006472247652709484, + 0.005949719808995724, + 0.004689037799835205, + 0.0030032263603061438, + 0.006154671777039766, + 0.003426029346883297, + 0.007382700219750404, + 0.006665672641247511, + 0.007779042236506939, + 0.0038221071008592844, + 0.006138953380286694, + 0.004200796131044626, + 0.00490595493465662, + 0.005768993869423866, + 0.014064685441553593, + 0.008179065771400928, + 0.012722711078822613, + 0.007552502676844597, + 0.008786157704889774, + 0.007037945091724396, + 0.006145397666841745, + 0.007690117694437504, + 0.007148828823119402, + 0.004315723665058613, + 0.00702625373378396, + 0.0030085404869168997, + 0.004521836061030626, + 0.004069369286298752, + 0.004183403681963682, + 0.0036543412134051323, + 0.020415155217051506, + 0.02109527587890625, + 0.028205424547195435, + 0.025927746668457985, + 0.02067819982767105, + 0.012359309010207653, + 0.029826907441020012, + 0.016747521236538887, + 0.0034930012188851833, + 0.002621231134980917, + 0.004204663913697004, + 0.0023523021955043077, + 0.0024563856422901154, + 0.001554748509079218, + 0.0014913583872839808, + 0.0018178316531702876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.019701508805155754 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red spider lily (Banana lily)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flower's petals velvety, scarlet, and curling back like ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there dozens of slender, curving stamens with shimmering translucent dewdrops at their tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held still at a level angle with a shallow depth of field that blurs the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the flower swaying gently in the breeze while the dewdrops vibrate in a slow, rhythmic tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014225387014448643 + ], + "opposite_logprob_tokens": [ + -4.264225482940674 + ], + "logprob_reward": -0.014225387014448643, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014225387014448643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19863031804561615, + 0.12030748277902603, + 0.15754199028015137, + 0.08392069488763809, + 0.030546877533197403, + 0.06280677020549774, + 0.1525713950395584, + 0.07636097073554993, + 0.070775605738163, + 0.046537820249795914 + ], + "credit_map": [ + 0.04040937125682831, + 0.019814446568489075, + 0.03739994391798973, + 0.028662927448749542, + 0.018008988350629807, + 0.014337892644107342, + 0.020568830892443657, + 0.019427942112088203, + 0.020656868815422058, + 0.014220035634934902, + 0.016236064955592155, + 0.018131788820028305, + 0.014200495555996895, + 0.01003706082701683, + 0.01623246632516384, + 0.01059271302074194, + 0.02985454723238945, + 0.022628016769886017, + 0.01975984126329422, + 0.016596511006355286, + 0.02421080879867077, + 0.014388130977749825, + 0.01745697110891342, + 0.012647171504795551, + 0.010678218677639961, + 0.01124475710093975, + 0.012031740508973598, + 0.010054283775389194, + 0.011716379784047604, + 0.009087435901165009, + 0.0084427734836936, + 0.010665114037692547, + 0.004320177715271711, + 0.0034746513701975346, + 0.004895067308098078, + 0.004530713893473148, + 0.003413253929466009, + 0.0018209863919764757, + 0.0050152563489973545, + 0.0030767754651606083, + 0.010172663256525993, + 0.007848892360925674, + 0.009948495775461197, + 0.006928218994289637, + 0.006747039034962654, + 0.006835068576037884, + 0.008903823792934418, + 0.005422569811344147, + 0.027299582958221436, + 0.018256770446896553, + 0.026810698211193085, + 0.01360015757381916, + 0.013804546557366848, + 0.017174851149320602, + 0.01345143560320139, + 0.022173363715410233, + 0.01038570236414671, + 0.010385025292634964, + 0.009151026606559753, + 0.0093778595328331, + 0.0120063666254282, + 0.006076191086322069, + 0.011719313450157642, + 0.007259492762386799, + 0.01125013455748558, + 0.009430252946913242, + 0.010058121755719185, + 0.008558773435652256, + 0.0073849717155098915, + 0.008189589716494083, + 0.009903861209750175, + 0.005999910179525614, + 0.010211816057562828, + 0.00426058005541563, + 0.00804243702441454, + 0.005745836067944765, + 0.003581926692277193, + 0.0040271650068461895, + 0.004823397379368544, + 0.005844665225595236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.014225387014448643 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red spider lily (Banana lily)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flower's petals velvety, scarlet, and curling back like ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there dozens of slender, curving stamens with shimmering translucent dewdrops at their tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held still at a level angle with a shallow depth of field that blurs the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the flower swaying gently in the breeze while the dewdrops vibrate in a slow, rhythmic tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018197020515799522 + ], + "opposite_logprob_tokens": [ + -4.018197059631348 + ], + "logprob_reward": -0.018197020515799522, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018197020515799522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20953862369060516, + 0.12136394530534744, + 0.11559345573186874, + 0.050719935446977615, + 0.06501160562038422, + 0.08724159002304077, + 0.12536658346652985, + 0.09190621972084045, + 0.0680151879787445, + 0.06524278223514557 + ], + "credit_map": [ + 0.031176751479506493, + 0.020506879314780235, + 0.029820434749126434, + 0.039792221039533615, + 0.01981183886528015, + 0.018900800496339798, + 0.028249403461813927, + 0.021280312910676003, + 0.019754156470298767, + 0.009974855929613113, + 0.021659675985574722, + 0.014339612796902657, + 0.009359787218272686, + 0.00949107576161623, + 0.01161215640604496, + 0.025172624737024307, + 0.014440543949604034, + 0.016271580010652542, + 0.012498946860432625, + 0.014523347839713097, + 0.017607254907488823, + 0.010884404182434082, + 0.01635223999619484, + 0.013015147298574448, + 0.0070907860063016415, + 0.006250884849578142, + 0.008746160194277763, + 0.004727286286652088, + 0.005844848230481148, + 0.005240265745669603, + 0.005440372508019209, + 0.007379337213933468, + 0.009534923359751701, + 0.00607656966894865, + 0.012922962196171284, + 0.010498616844415665, + 0.005844923667609692, + 0.004869766533374786, + 0.007896955125033855, + 0.007366892881691456, + 0.01166969072073698, + 0.009472586214542389, + 0.014442513696849346, + 0.011349315755069256, + 0.010159114375710487, + 0.007913283072412014, + 0.013700714334845543, + 0.008534383960068226, + 0.0237758569419384, + 0.015052673406898975, + 0.022707847878336906, + 0.011150328442454338, + 0.012118478305637836, + 0.012393365614116192, + 0.011166253127157688, + 0.017001790925860405, + 0.016371846199035645, + 0.010030520148575306, + 0.014978642575442791, + 0.009825248271226883, + 0.010766223073005676, + 0.008564045652747154, + 0.013728996738791466, + 0.007640708237886429, + 0.010521631687879562, + 0.008973874151706696, + 0.010051104240119457, + 0.00669785076752305, + 0.008562443777918816, + 0.008017322048544884, + 0.007443474605679512, + 0.007747496012598276, + 0.013935072347521782, + 0.00549249118193984, + 0.011939230374991894, + 0.009355578571557999, + 0.004351829178631306, + 0.005319445393979549, + 0.00688799936324358, + 0.007961136288940907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.018197020515799522 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red spider lily (Banana lily)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flower's petals velvety, scarlet, and curling back like ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there dozens of slender, curving stamens with shimmering translucent dewdrops at their tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held still at a level angle with a shallow depth of field that blurs the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the flower swaying gently in the breeze while the dewdrops vibrate in a slow, rhythmic tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818847469985485 + ], + "opposite_logprob_tokens": [ + -4.0181884765625 + ], + "logprob_reward": -0.01818847469985485, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818847469985485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16647695004940033, + 0.08726678788661957, + 0.13283634185791016, + 0.08083877712488174, + 0.10617855191230774, + 0.07518598437309265, + 0.09118212759494781, + 0.06684721261262894, + 0.12341897189617157, + 0.06976829469203949 + ], + "credit_map": [ + 0.034788694232702255, + 0.015152549371123314, + 0.03600253537297249, + 0.015434283763170242, + 0.013990064151585102, + 0.01877288520336151, + 0.016290463507175446, + 0.0160454660654068, + 0.016911836341023445, + 0.010789302177727222, + 0.01204872690141201, + 0.010810685344040394, + 0.008882268331944942, + 0.008714878000319004, + 0.00931190513074398, + 0.009797181002795696, + 0.024527890607714653, + 0.012323045171797276, + 0.023024767637252808, + 0.019320325925946236, + 0.008640076033771038, + 0.012140901759266853, + 0.015853943303227425, + 0.017005398869514465, + 0.014535214751958847, + 0.00912237074226141, + 0.013270185329020023, + 0.007111782673746347, + 0.007156193722039461, + 0.011670812964439392, + 0.008775738999247551, + 0.009196477010846138, + 0.018444914370775223, + 0.011479823850095272, + 0.019307509064674377, + 0.019911540672183037, + 0.01052080187946558, + 0.007448038551956415, + 0.009610684588551521, + 0.009455237537622452, + 0.01269067544490099, + 0.008406643755733967, + 0.014607531018555164, + 0.011607077904045582, + 0.006877145729959011, + 0.0062373122200369835, + 0.007970144972205162, + 0.006789451465010643, + 0.015234572812914848, + 0.01155550591647625, + 0.011510083451867104, + 0.010407829657196999, + 0.012706421315670013, + 0.009506444446742535, + 0.009676833637058735, + 0.01058442983776331, + 0.0125992801040411, + 0.00613947119563818, + 0.009186345152556896, + 0.006804176140576601, + 0.006478806491941214, + 0.006919734179973602, + 0.008552496321499348, + 0.010166901163756847, + 0.026058409363031387, + 0.010289187543094158, + 0.027743089944124222, + 0.009326215833425522, + 0.009110025130212307, + 0.014457791112363338, + 0.01113691832870245, + 0.01529733557254076, + 0.016369570046663284, + 0.006409451365470886, + 0.014583505690097809, + 0.006436443421989679, + 0.004399207420647144, + 0.008231277577579021, + 0.00573546951636672, + 0.007603374309837818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.01818847469985485 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red spider lily (Banana lily)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flower's petals velvety, scarlet, and curling back like ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there dozens of slender, curving stamens with shimmering translucent dewdrops at their tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held still at a level angle with a shallow depth of field that blurs the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the flower swaying gently in the breeze while the dewdrops vibrate in a slow, rhythmic tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.252084881067276 + ], + "opposite_logprob_tokens": [ + -1.5020848512649536 + ], + "logprob_reward": -0.252084881067276, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.252084881067276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27256476879119873, + 0.16274553537368774, + 0.05220753327012062, + 0.078192338347435, + 0.05396956577897072, + 0.06204836070537567, + 0.1373479962348938, + 0.04922226443886757, + 0.09139727056026459, + 0.04030441492795944 + ], + "credit_map": [ + 0.05150146782398224, + 0.02237858809530735, + 0.045904893428087234, + 0.04292877018451691, + 0.02070559561252594, + 0.02395705133676529, + 0.03502458706498146, + 0.030163800343871117, + 0.03445494920015335, + 0.019361283630132675, + 0.028342651203274727, + 0.013867674395442009, + 0.0169744361191988, + 0.016751592978835106, + 0.015223134309053421, + 0.017769811674952507, + 0.006291790399700403, + 0.006382419727742672, + 0.007151654921472073, + 0.009114213287830353, + 0.00712288822978735, + 0.0050225271843373775, + 0.005738307721912861, + 0.005383732262998819, + 0.013880806043744087, + 0.009187697432935238, + 0.011603892780840397, + 0.010188153013586998, + 0.007354513742029667, + 0.007851227186620235, + 0.007785444147884846, + 0.010340606793761253, + 0.009361782111227512, + 0.006264011841267347, + 0.011190044693648815, + 0.00628247857093811, + 0.005245329812169075, + 0.004656832665205002, + 0.0057683358900249004, + 0.005200752057135105, + 0.0059546600095927715, + 0.010107455775141716, + 0.007416028995066881, + 0.008503026328980923, + 0.008884808048605919, + 0.005765055771917105, + 0.00855272077023983, + 0.006864605937153101, + 0.02496405877172947, + 0.013411933556199074, + 0.02742774784564972, + 0.01182345487177372, + 0.010664486326277256, + 0.014722169376909733, + 0.01341258268803358, + 0.020921556279063225, + 0.007231808267533779, + 0.005692692007869482, + 0.007508740294724703, + 0.00791450310498476, + 0.006108168046921492, + 0.0036850266624242067, + 0.005978171247988939, + 0.005103151313960552, + 0.016505610197782516, + 0.007961692288517952, + 0.014397735707461834, + 0.0064001609571278095, + 0.009199867956340313, + 0.011952174827456474, + 0.008639674633741379, + 0.016340354457497597, + 0.009580097161233425, + 0.002297775587067008, + 0.010144582018256187, + 0.004065812565386295, + 0.0023019167128950357, + 0.0032821630593389273, + 0.003660531947389245, + 0.004971534013748169 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.252084881067276 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up, captured from a level angle, the camera held still, focuses on a vibrant red spider lily (Banana lily) swaying gently in the breeze. Its velvety scarlet petals curl gracefully back like ribbons, revealing dozens of slender, curving stamens radiating from the center. At the tip of each stamen hangs a shimmering, translucent dewdrop, shimmering like tiny diamonds in the soft, diffuse morning light. An imperceptible draft of air causes the entire flower to vibrate in a slow, rhythmic tremor, gently swaying the dewdrops on the stamens, capturing and refracting fleeting light. The incredibly shallow depth of field completely blurs the background, transforming it into a dreamy, soft patch of light composed of varying shades of green and earth tones, placing all visual focus on the flower's vitality. The entire picture is rich and saturated in color, creating a tranquil, focused and poetic atmosphere, as if time has frozen at this moment, just to show this brief and extreme beauty.", + "seed": 12732, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_4/index_3469_seed_12732_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.058814899437129496, + "vlm_logprob": -0.058814899437129496, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.058814899437129496, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red spider lily (Banana lily)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flower's petals velvety, scarlet, and curling back like ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there dozens of slender, curving stamens with shimmering translucent dewdrops at their tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held still at a level angle with a shallow depth of field that blurs the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the flower swaying gently in the breeze while the dewdrops vibrate in a slow, rhythmic tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025217320770025253 + ], + "opposite_logprob_tokens": [ + -3.775217294692993 + ], + "logprob_reward": -0.025217320770025253, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.025217320770025253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34578144550323486, + 0.10641530901193619, + 0.05538863316178322, + 0.07144461572170258, + 0.13360044360160828, + 0.05145168676972389, + 0.05032208189368248, + 0.02168809063732624, + 0.02037651464343071, + 0.1435312181711197 + ], + "credit_map": [ + 0.04869857057929039, + 0.030853992328047752, + 0.05272171273827553, + 0.04609201103448868, + 0.046051524579524994, + 0.03925911337137222, + 0.03253951296210289, + 0.04956500604748726, + 0.015726670622825623, + 0.011687939055263996, + 0.01745535247027874, + 0.01193181611597538, + 0.013160246424376965, + 0.009343416430056095, + 0.008918337523937225, + 0.018191533163189888, + 0.00817036535590887, + 0.004641377367079258, + 0.00907604768872261, + 0.006978415418416262, + 0.00612471392378211, + 0.004788579884916544, + 0.007221125066280365, + 0.008388010784983635, + 0.01100365910679102, + 0.004305724520236254, + 0.012911117635667324, + 0.008616885170340538, + 0.0071839033626019955, + 0.006924419663846493, + 0.008379071019589901, + 0.012119829654693604, + 0.01616291143000126, + 0.01489713042974472, + 0.0202551931142807, + 0.016644742339849472, + 0.019450029358267784, + 0.011447781696915627, + 0.017544154077768326, + 0.017198503017425537, + 0.006857626140117645, + 0.005506813991814852, + 0.008685624226927757, + 0.006340928841382265, + 0.005440647713840008, + 0.004197753965854645, + 0.006462864112108946, + 0.007959424518048763, + 0.007042274810373783, + 0.004597165156155825, + 0.010367737151682377, + 0.005595763213932514, + 0.005485466681420803, + 0.004487567115575075, + 0.005245962645858526, + 0.007500146981328726, + 0.0032698167487978935, + 0.0019452456617727876, + 0.003904758719727397, + 0.0026033956091850996, + 0.0021010402124375105, + 0.0017262527253478765, + 0.002582397311925888, + 0.0035551826003938913, + 0.0030973267275840044, + 0.0015056764241307974, + 0.003594514913856983, + 0.0024959007278084755, + 0.001793706207536161, + 0.0018995552090927958, + 0.003076908178627491, + 0.002912926720455289, + 0.010832511819899082, + 0.010412429459393024, + 0.037727031856775284, + 0.010304979048669338, + 0.014328721910715103, + 0.017499752342700958, + 0.011565128341317177, + 0.030860677361488342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.025217320770025253 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red spider lily (Banana lily)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flower's petals velvety, scarlet, and curling back like ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there dozens of slender, curving stamens with shimmering translucent dewdrops at their tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held still at a level angle with a shallow depth of field that blurs the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the flower swaying gently in the breeze while the dewdrops vibrate in a slow, rhythmic tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012567726895213127 + ], + "opposite_logprob_tokens": [ + -4.387567520141602 + ], + "logprob_reward": -0.012567726895213127, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012567726895213127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16059637069702148, + 0.11990813165903091, + 0.10261911898851395, + 0.03336659446358681, + 0.0421363040804863, + 0.18321669101715088, + 0.05838080868124962, + 0.06021075323224068, + 0.07440871000289917, + 0.1651565134525299 + ], + "credit_map": [ + 0.03375520557165146, + 0.013872954063117504, + 0.02854403667151928, + 0.022721443325281143, + 0.014553169719874859, + 0.01636326126754284, + 0.011106360703706741, + 0.01967993564903736, + 0.016551034525036812, + 0.012897324748337269, + 0.017890147864818573, + 0.01458771526813507, + 0.015718912705779076, + 0.012070076540112495, + 0.013553264550864697, + 0.016639649868011475, + 0.012881968170404434, + 0.007850371301174164, + 0.016138572245836258, + 0.013530842028558254, + 0.011565799824893475, + 0.011125447228550911, + 0.013595441356301308, + 0.01593068055808544, + 0.00528543209657073, + 0.0023474080953747034, + 0.0049630431458354, + 0.005389035679399967, + 0.0037158923223614693, + 0.0030946002807468176, + 0.003884645877406001, + 0.004686536733061075, + 0.0050927926786243916, + 0.004837791435420513, + 0.0064085982739925385, + 0.005890773143619299, + 0.005655480083078146, + 0.0038355442229658365, + 0.004636580590158701, + 0.005778740160167217, + 0.02201339602470398, + 0.024542205035686493, + 0.02946558967232704, + 0.021387429907917976, + 0.021415486931800842, + 0.01863652653992176, + 0.019848547875881195, + 0.025907497853040695, + 0.00936119258403778, + 0.005660022143274546, + 0.010302240028977394, + 0.006495886016637087, + 0.004038824234157801, + 0.0050729913637042046, + 0.006249383091926575, + 0.011200268752872944, + 0.010115152224898338, + 0.005475109443068504, + 0.010802153497934341, + 0.007398085668683052, + 0.004253637511283159, + 0.005655896849930286, + 0.006410755682736635, + 0.010099961422383785, + 0.008883662521839142, + 0.008673837408423424, + 0.008336829021573067, + 0.01051403395831585, + 0.010518566705286503, + 0.009265794418752193, + 0.00921719241887331, + 0.008998794481158257, + 0.015689700841903687, + 0.014961916022002697, + 0.02906722202897072, + 0.016402825713157654, + 0.016567833721637726, + 0.029135478660464287, + 0.01843634992837906, + 0.024895191192626953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.012567726895213127 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red spider lily (Banana lily)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flower's petals velvety, scarlet, and curling back like ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there dozens of slender, curving stamens with shimmering translucent dewdrops at their tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held still at a level angle with a shallow depth of field that blurs the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the flower swaying gently in the breeze while the dewdrops vibrate in a slow, rhythmic tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0181954987347126 + ], + "opposite_logprob_tokens": [ + -4.018195629119873 + ], + "logprob_reward": -0.0181954987347126, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.0181954987347126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1812603920698166, + 0.11614004522562027, + 0.10628265142440796, + 0.028236711397767067, + 0.08076971769332886, + 0.20127980411052704, + 0.07149209082126617, + 0.061309024691581726, + 0.04793388396501541, + 0.1052955836057663 + ], + "credit_map": [ + 0.03967741131782532, + 0.014559607021510601, + 0.034064050763845444, + 0.023861706256866455, + 0.01537313498556614, + 0.01510481908917427, + 0.014812422916293144, + 0.02380727231502533, + 0.017867255955934525, + 0.01269135344773531, + 0.0144210709258914, + 0.015205337665975094, + 0.016461294144392014, + 0.010632230900228024, + 0.011163825169205666, + 0.01769767329096794, + 0.013012322597205639, + 0.0082010542973876, + 0.015980714932084084, + 0.014426447451114655, + 0.013533709570765495, + 0.011986865662038326, + 0.013764235191047192, + 0.015377317555248737, + 0.00477548548951745, + 0.0024897833354771137, + 0.004380958620458841, + 0.0038741196040064096, + 0.0028851586394011974, + 0.0024833776988089085, + 0.0029076486825942993, + 0.004440183285623789, + 0.01062056515365839, + 0.008573503233492374, + 0.01187724620103836, + 0.01135225035250187, + 0.009177039377391338, + 0.007414885796606541, + 0.011777359992265701, + 0.009976880624890327, + 0.024576425552368164, + 0.02588430978357792, + 0.03276730701327324, + 0.023873345926404, + 0.02356422133743763, + 0.01957145892083645, + 0.022085951641201973, + 0.028956808149814606, + 0.01139044389128685, + 0.006054286379367113, + 0.012504471465945244, + 0.008972760289907455, + 0.004980920813977718, + 0.007219580467790365, + 0.0075914496555924416, + 0.012778181582689285, + 0.009256561286747456, + 0.005934663582593203, + 0.010808445513248444, + 0.007752075791358948, + 0.005740653723478317, + 0.005304854828864336, + 0.0073255496099591255, + 0.009186227805912495, + 0.006812675856053829, + 0.004367440473288298, + 0.007528987713158131, + 0.006837392225861549, + 0.004288677591830492, + 0.005230354145169258, + 0.004738172981888056, + 0.008130191825330257, + 0.012343916110694408, + 0.012241440825164318, + 0.014015126042068005, + 0.010195511393249035, + 0.008370707742869854, + 0.018631571903824806, + 0.01373509131371975, + 0.01576222851872444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.0181954987347126 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red spider lily (Banana lily)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flower's petals velvety, scarlet, and curling back like ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there dozens of slender, curving stamens with shimmering translucent dewdrops at their tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held still at a level angle with a shallow depth of field that blurs the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the flower swaying gently in the breeze while the dewdrops vibrate in a slow, rhythmic tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543358840048313 + ], + "opposite_logprob_tokens": [ + -4.387543201446533 + ], + "logprob_reward": -0.012543358840048313, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543358840048313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10726713389158249, + 0.14667560160160065, + 0.06978441029787064, + 0.04919328913092613, + 0.0883820652961731, + 0.139700785279274, + 0.13809029757976532, + 0.09799662977457047, + 0.06546485424041748, + 0.09744486957788467 + ], + "credit_map": [ + 0.015101214870810509, + 0.009132996201515198, + 0.016032302752137184, + 0.0192415788769722, + 0.010848312638700008, + 0.00910580437630415, + 0.01109354104846716, + 0.016711387783288956, + 0.023295927792787552, + 0.011002939194440842, + 0.025634869933128357, + 0.023894818499684334, + 0.012427682988345623, + 0.009365753270685673, + 0.011824565008282661, + 0.029229063540697098, + 0.009976444765925407, + 0.006713749375194311, + 0.009843233041465282, + 0.011251899413764477, + 0.007778806611895561, + 0.0068848407827317715, + 0.007428454235196114, + 0.009906991384923458, + 0.011952591128647327, + 0.003989018499851227, + 0.005815485026687384, + 0.0048707411624491215, + 0.005285844672471285, + 0.004883912391960621, + 0.005012498237192631, + 0.0073832059279084206, + 0.012562760151922703, + 0.008002148009836674, + 0.014483828097581863, + 0.011360256001353264, + 0.013063773512840271, + 0.009385901503264904, + 0.009269656613469124, + 0.010253749787807465, + 0.01465186569839716, + 0.018834605813026428, + 0.023723069578409195, + 0.01756618171930313, + 0.018581995740532875, + 0.015345864929258823, + 0.015644259750843048, + 0.01535297092050314, + 0.021931041032075882, + 0.009650864638388157, + 0.027453536167740822, + 0.013029800727963448, + 0.009136815555393696, + 0.01251896470785141, + 0.01608213596045971, + 0.028287163004279137, + 0.016526376828551292, + 0.009055132046341896, + 0.015578380785882473, + 0.011324837803840637, + 0.009534906595945358, + 0.009053638204932213, + 0.010187924839556217, + 0.016735447570681572, + 0.008014610968530178, + 0.008066248148679733, + 0.007840326987206936, + 0.009024476632475853, + 0.009370788931846619, + 0.008125522173941135, + 0.007377911359071732, + 0.007644974160939455, + 0.012842909432947636, + 0.009236063808202744, + 0.018619216978549957, + 0.008999234065413475, + 0.01077269297093153, + 0.008411494083702564, + 0.011668999679386616, + 0.016894256696105003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 6 + ], + "selected_reward": -0.012543358840048313 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red spider lily (Banana lily)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flower's petals velvety, scarlet, and curling back like ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there dozens of slender, curving stamens with shimmering translucent dewdrops at their tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held still at a level angle with a shallow depth of field that blurs the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the flower swaying gently in the breeze while the dewdrops vibrate in a slow, rhythmic tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2255505919456482 + ], + "opposite_logprob_tokens": [ + -1.600550651550293 + ], + "logprob_reward": -0.2255505919456482, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255505919456482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1534566432237625, + 0.0484393872320652, + 0.06155438721179962, + 0.04631667584180832, + 0.055335063487291336, + 0.10132088512182236, + 0.16792283952236176, + 0.11595752090215683, + 0.11431165784597397, + 0.13538497686386108 + ], + "credit_map": [ + 0.022307682782411575, + 0.017540965229272842, + 0.023120513185858727, + 0.025355935096740723, + 0.014342999085783958, + 0.01312322448939085, + 0.0170489139854908, + 0.020616402849555016, + 0.009600000455975533, + 0.004823144059628248, + 0.007607733365148306, + 0.007610308472067118, + 0.004228169564157724, + 0.0032358516473323107, + 0.0039708903059363365, + 0.00736328586935997, + 0.00992717407643795, + 0.004867730662226677, + 0.009407490491867065, + 0.009528735652565956, + 0.005899578798562288, + 0.005336876027286053, + 0.006980959326028824, + 0.009605837985873222, + 0.009131433442234993, + 0.0033411949407309294, + 0.006067417096346617, + 0.00783737376332283, + 0.002875681035220623, + 0.0031738816760480404, + 0.004843798000365496, + 0.009045899845659733, + 0.008084306493401527, + 0.00565878301858902, + 0.00842567440122366, + 0.006738088093698025, + 0.006418128032237291, + 0.00470283767208457, + 0.007751330733299255, + 0.007555915974080563, + 0.011420086026191711, + 0.012578791938722134, + 0.016449516639113426, + 0.011640080250799656, + 0.010825995355844498, + 0.009090875275433064, + 0.013283468782901764, + 0.01603206992149353, + 0.025516871362924576, + 0.009544217959046364, + 0.02923830784857273, + 0.017347389832139015, + 0.0064260391518473625, + 0.01535689178854227, + 0.023084374144673347, + 0.04140874370932579, + 0.021273192018270493, + 0.010226917453110218, + 0.0197885874658823, + 0.013992064632475376, + 0.008179213851690292, + 0.010506602004170418, + 0.01280477736145258, + 0.019186167046427727, + 0.013887699693441391, + 0.012611824087798595, + 0.014934798702597618, + 0.015281054191291332, + 0.0147446533665061, + 0.013252702541649342, + 0.014790759421885014, + 0.01480817049741745, + 0.015447311103343964, + 0.016385596245527267, + 0.01942918822169304, + 0.016930628567934036, + 0.013212593272328377, + 0.015703406184911728, + 0.017900776118040085, + 0.020375482738018036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.2255505919456482 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_5/index_tensor([2061, 2061])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_5/index_tensor([2061, 2061])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0cb11afbfd2732fc568cb801b54a494f55f57335 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_5/index_tensor([2061, 2061])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping aerial tracking shot smoothly tracks a massive dragon as it soars above a boundless, roiling sea of clouds. The golden afterglow of the setting sun paints the cloudtops a warm gold and soft pink, while the deep valleys between them sink into a serene indigo shade. The dragon's massive body is covered in countless cyan scales, each one like a finely crafted gem, shimmering in the light with iridescent hues ranging from deep turquoise to brilliant aquamarine. Its long, sinuous neck curves gracefully, its head crowned by towering obsidian horns. A pair of intelligent eyes, burning with amber light, calmly survey the world below. Its massive, resilient wing membranes are a translucent light blue, revealing their complex vein structure as the sunlight penetrates. With each powerful yet deliberate flap of its wings, its massive body glides effortlessly forward, its long, bony tail piercing it with a spur traces a graceful arc behind it, stirring the clouds in its wake. The entire scene is imbued with a sense of epic grandeur and tranquil power, creating an awe-inspiring atmosphere of ancient mythology.", + "seed": 3761, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_5/index_2061_seed_3761_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.010064213769510388, + "vlm_logprob": -0.010064213769510388, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.010064213769510388, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a sweeping aerial tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon visible soaring above a boundless sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cloudtops painted in warm gold and soft pink with deep indigo valleys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's body covered in cyan scales with iridescent turquoise and aquamarine hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dragon flap its translucent light blue wings while gliding forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02233400195837021 + ], + "opposite_logprob_tokens": [ + -4.397334098815918 + ], + "logprob_reward": -0.02233400195837021, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.02233400195837021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13280965387821198, + 0.08428747206926346, + 0.09450308233499527, + 0.06659775972366333, + 0.10419292002916336, + 0.06907113641500473, + 0.05775289610028267, + 0.20921772718429565, + 0.13846011459827423, + 0.04310726374387741 + ], + "credit_map": [ + 0.014429900795221329, + 0.011761954985558987, + 0.007747999392449856, + 0.011043654754757881, + 0.020806720480322838, + 0.03096204437315464, + 0.012797510251402855, + 0.02325987070798874, + 0.011983273550868034, + 0.01060712244361639, + 0.011925064958631992, + 0.008892376907169819, + 0.013678564690053463, + 0.012036968022584915, + 0.00658611673861742, + 0.008577988483011723, + 0.008680347353219986, + 0.006433775182813406, + 0.007686752825975418, + 0.017552414909005165, + 0.008134753443300724, + 0.007999259047210217, + 0.0260448195040226, + 0.011970962397754192, + 0.005899820942431688, + 0.004525374621152878, + 0.0077280206605792046, + 0.006822451017796993, + 0.014064833521842957, + 0.010198396630585194, + 0.007801446132361889, + 0.009557412937283516, + 0.01443040743470192, + 0.011772998608648777, + 0.016507398337125778, + 0.015035337768495083, + 0.010096022859215736, + 0.013429904356598854, + 0.01304374448955059, + 0.009877102449536324, + 0.013536229729652405, + 0.0052581229247152805, + 0.0074789030477404594, + 0.007347874343395233, + 0.005370395258069038, + 0.016003193333745003, + 0.006657485384494066, + 0.007418932858854532, + 0.007908240891993046, + 0.006005748175084591, + 0.006857675965875387, + 0.00728470366448164, + 0.006483836565166712, + 0.01356071699410677, + 0.004966161213815212, + 0.0046858093701303005, + 0.019294923171401024, + 0.021804992109537125, + 0.026342114433646202, + 0.022562064230442047, + 0.02838289923965931, + 0.03210468590259552, + 0.027241451665759087, + 0.031484607607126236, + 0.009731704369187355, + 0.008694133721292019, + 0.032522160559892654, + 0.012569422833621502, + 0.012535516172647476, + 0.017570748925209045, + 0.01710512861609459, + 0.02773129567503929, + 0.004538574256002903, + 0.00612612022086978, + 0.00684657646343112, + 0.004464569967240095, + 0.0035201553255319595, + 0.008460483513772488, + 0.004929913207888603, + 0.004220867063850164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.02233400195837021 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a sweeping aerial tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon visible soaring above a boundless sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cloudtops painted in warm gold and soft pink with deep indigo valleys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's body covered in cyan scales with iridescent turquoise and aquamarine hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dragon flap its translucent light blue wings while gliding forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004652508534491062 + ], + "opposite_logprob_tokens": [ + -5.379652500152588 + ], + "logprob_reward": -0.004652508534491062, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004652508534491062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0657968670129776, + 0.05949065089225769, + 0.19766488671302795, + 0.08996204286813736, + 0.3106288015842438, + 0.03113453835248947, + 0.049453023821115494, + 0.10215181857347488, + 0.06084781885147095, + 0.03286953642964363 + ], + "credit_map": [ + 0.007610865402966738, + 0.006913816090673208, + 0.005058180075138807, + 0.003977533429861069, + 0.008034205995500088, + 0.01636597141623497, + 0.00773725938051939, + 0.01009904034435749, + 0.008494549430906773, + 0.005633864551782608, + 0.0067993211559951305, + 0.005654403939843178, + 0.010668031871318817, + 0.00803791917860508, + 0.006234371103346348, + 0.007968193851411343, + 0.011683511547744274, + 0.013465345837175846, + 0.01915932074189186, + 0.044765837490558624, + 0.010741294361650944, + 0.01684960536658764, + 0.031205331906676292, + 0.04979465901851654, + 0.007071879226714373, + 0.006526405923068523, + 0.00923014897853136, + 0.010129842907190323, + 0.016314588487148285, + 0.020059626549482346, + 0.01227753795683384, + 0.008352016098797321, + 0.03811163455247879, + 0.018924176692962646, + 0.05489489808678627, + 0.03597130998969078, + 0.02187265083193779, + 0.0383930578827858, + 0.04204835370182991, + 0.060412727296352386, + 0.0041013481095433235, + 0.002527889097109437, + 0.0031387454364448786, + 0.00453698355704546, + 0.0030078738927841187, + 0.004259859677404165, + 0.0035182691644877195, + 0.006043571978807449, + 0.006391060072928667, + 0.005671249236911535, + 0.007058417424559593, + 0.005399757996201515, + 0.00657551409676671, + 0.008072552271187305, + 0.005133497063070536, + 0.005150980316102505, + 0.010194920003414154, + 0.004663459025323391, + 0.02388414368033409, + 0.006624286528676748, + 0.006179762538522482, + 0.014475078321993351, + 0.021127628162503242, + 0.015002538450062275, + 0.004951933864504099, + 0.004944301210343838, + 0.010380786843597889, + 0.00763991242274642, + 0.005434514954686165, + 0.006992015056312084, + 0.012880093418061733, + 0.007624265272170305, + 0.004901014268398285, + 0.004240028560161591, + 0.005124278366565704, + 0.002738790586590767, + 0.003143906593322754, + 0.0057677156291902065, + 0.0038000070489943027, + 0.0031537963077425957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 7 + ], + "selected_reward": -0.004652508534491062 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a sweeping aerial tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon visible soaring above a boundless sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cloudtops painted in warm gold and soft pink with deep indigo valleys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's body covered in cyan scales with iridescent turquoise and aquamarine hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dragon flap its translucent light blue wings while gliding forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008643004111945629 + ], + "opposite_logprob_tokens": [ + -4.75864315032959 + ], + "logprob_reward": -0.008643004111945629, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008643004111945629, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11129390448331833, + 0.07526379078626633, + 0.10922160744667053, + 0.10650429129600525, + 0.2221621721982956, + 0.019174952059984207, + 0.08077585697174072, + 0.1385900378227234, + 0.1056765466928482, + 0.031336840242147446 + ], + "credit_map": [ + 0.010540202260017395, + 0.009275976568460464, + 0.0077188569121062756, + 0.0105751883238554, + 0.015032732859253883, + 0.02232755720615387, + 0.014820916578173637, + 0.021002469584345818, + 0.008013625629246235, + 0.00772954011335969, + 0.009651081636548042, + 0.008146993815898895, + 0.012397979386150837, + 0.016673194244503975, + 0.005385732278227806, + 0.007265640888363123, + 0.012225096113979816, + 0.00897114910185337, + 0.013018555007874966, + 0.015882909297943115, + 0.006610853131860495, + 0.007265015970915556, + 0.018931012600660324, + 0.026317017152905464, + 0.00878220982849598, + 0.009741386398673058, + 0.013570382259786129, + 0.012693820521235466, + 0.015122265554964542, + 0.01794460602104664, + 0.014004042372107506, + 0.014645572751760483, + 0.025467563420534134, + 0.022411683574318886, + 0.03852347284555435, + 0.03198511525988579, + 0.02254854515194893, + 0.026253147050738335, + 0.028827356174588203, + 0.026145288720726967, + 0.0029522296972572803, + 0.0013696851674467325, + 0.002449115738272667, + 0.002438983414322138, + 0.001666351337917149, + 0.0037966056261211634, + 0.002101657446473837, + 0.002400323748588562, + 0.011282805353403091, + 0.00693603465333581, + 0.009608648717403412, + 0.010148163884878159, + 0.00954697746783495, + 0.01931341551244259, + 0.007225904148072004, + 0.006713906768709421, + 0.014561222866177559, + 0.013294047676026821, + 0.0229512806981802, + 0.013685750775039196, + 0.01742803491652012, + 0.01659349352121353, + 0.0216667577624321, + 0.018409444019198418, + 0.008066672831773758, + 0.00914832204580307, + 0.014126766473054886, + 0.013894952833652496, + 0.01161202508956194, + 0.015147591009736061, + 0.022938527166843414, + 0.01074168086051941, + 0.004352238494902849, + 0.0031425405759364367, + 0.0048523470759391785, + 0.002893856493756175, + 0.0034269087482243776, + 0.0052060228772461414, + 0.004536630120128393, + 0.002926293760538101 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.008643004111945629 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a sweeping aerial tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon visible soaring above a boundless sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cloudtops painted in warm gold and soft pink with deep indigo valleys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's body covered in cyan scales with iridescent turquoise and aquamarine hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dragon flap its translucent light blue wings while gliding forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003616105066612363 + ], + "opposite_logprob_tokens": [ + -5.6286163330078125 + ], + "logprob_reward": -0.003616105066612363, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003616105066612363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09254338592290878, + 0.1647867113351822, + 0.11170358210802078, + 0.0637710690498352, + 0.05293010175228119, + 0.037547603249549866, + 0.1611349880695343, + 0.13797014951705933, + 0.12685681879520416, + 0.0507555790245533 + ], + "credit_map": [ + 0.010223755612969398, + 0.008691981434822083, + 0.008409208618104458, + 0.007239392958581448, + 0.014971605502068996, + 0.015858037397265434, + 0.010335522703826427, + 0.01681388169527054, + 0.019895486533641815, + 0.018340574577450752, + 0.023833664134144783, + 0.01865769736468792, + 0.025560151785612106, + 0.02797800861299038, + 0.014160587452352047, + 0.016360526904463768, + 0.006373313721269369, + 0.006562224589288235, + 0.00965211447328329, + 0.02109602279961109, + 0.009569592773914337, + 0.009430993348360062, + 0.02244727313518524, + 0.026572050526738167, + 0.005344842094928026, + 0.006162693724036217, + 0.006764231715351343, + 0.007688084151595831, + 0.010362952947616577, + 0.010747124440968037, + 0.00834923516958952, + 0.008351911790668964, + 0.006794029846787453, + 0.006228878628462553, + 0.005881974473595619, + 0.00764957582578063, + 0.006192747037857771, + 0.006986563093960285, + 0.007368783000856638, + 0.0058275507763028145, + 0.004981903824955225, + 0.002986157312989235, + 0.00531049445271492, + 0.003958855289965868, + 0.0032350500114262104, + 0.007526594214141369, + 0.004421591758728027, + 0.0051269568502902985, + 0.020642109215259552, + 0.013264311477541924, + 0.01773233525454998, + 0.02162964455783367, + 0.01586739718914032, + 0.03506605699658394, + 0.02117300033569336, + 0.015760133042931557, + 0.015536301769316196, + 0.013585957698523998, + 0.024717705324292183, + 0.014500819146633148, + 0.01585904322564602, + 0.013563293032348156, + 0.020243335515260696, + 0.019963707774877548, + 0.010688716545701027, + 0.009608607739210129, + 0.022038152441382408, + 0.015122849494218826, + 0.012320652604103088, + 0.014845565892755985, + 0.025334084406495094, + 0.01689818874001503, + 0.00734739750623703, + 0.006007826421409845, + 0.007291705813258886, + 0.006558550521731377, + 0.0037301506381481886, + 0.010435429401695728, + 0.005078078247606754, + 0.004306440707296133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 7 + ], + "selected_reward": -0.003616105066612363 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a sweeping aerial tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon visible soaring above a boundless sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cloudtops painted in warm gold and soft pink with deep indigo valleys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's body covered in cyan scales with iridescent turquoise and aquamarine hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dragon flap its translucent light blue wings while gliding forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075449176132679 + ], + "opposite_logprob_tokens": [ + -4.511075496673584 + ], + "logprob_reward": -0.011075449176132679, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075449176132679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13574542105197906, + 0.06342437863349915, + 0.14189046621322632, + 0.13519415259361267, + 0.13999846577644348, + 0.029468806460499763, + 0.031415846198797226, + 0.15137170255184174, + 0.11885862052440643, + 0.05263219401240349 + ], + "credit_map": [ + 0.01347566768527031, + 0.010456925258040428, + 0.007964028976857662, + 0.010338167659938335, + 0.020880931988358498, + 0.03247954323887825, + 0.01330041978508234, + 0.02684973180294037, + 0.008780430071055889, + 0.006213836837559938, + 0.009765644557774067, + 0.005855603609234095, + 0.00890795886516571, + 0.0105411596596241, + 0.006773569155484438, + 0.006586168427020311, + 0.007208327762782574, + 0.010008199140429497, + 0.013901367783546448, + 0.042948897927999496, + 0.009327687323093414, + 0.00923033058643341, + 0.015779590234160423, + 0.03348606452345848, + 0.015426374040544033, + 0.010822582989931107, + 0.014932048507034779, + 0.018274405971169472, + 0.021200478076934814, + 0.013668273575603962, + 0.018456721678376198, + 0.02241324819624424, + 0.017011819407343864, + 0.015417025424540043, + 0.023402882739901543, + 0.022124961018562317, + 0.013453151099383831, + 0.016429752111434937, + 0.019489653408527374, + 0.012669212184846401, + 0.004111330024898052, + 0.0024788053706288338, + 0.005634008906781673, + 0.0033013869542628527, + 0.002474419306963682, + 0.004748188424855471, + 0.0033742072992026806, + 0.003346457611769438, + 0.004623692482709885, + 0.004016796592622995, + 0.005272626876831055, + 0.0035630634520202875, + 0.004001220688223839, + 0.003750736592337489, + 0.0031606650445610285, + 0.0030270398128777742, + 0.01768535189330578, + 0.015277724713087082, + 0.02560415118932724, + 0.01599436067044735, + 0.017447592690587044, + 0.013439293019473553, + 0.027441762387752533, + 0.018481450155377388, + 0.010134520940482616, + 0.010338942520320415, + 0.01737549714744091, + 0.015444867312908173, + 0.012174555100500584, + 0.015710480511188507, + 0.024669310078024864, + 0.013010439462959766, + 0.008105862885713577, + 0.0065575288608670235, + 0.007139736786484718, + 0.0048090252093970776, + 0.004237275570631027, + 0.007079703733325005, + 0.00840336736291647, + 0.006299689877778292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 4 + ], + "selected_reward": -0.011075449176132679 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping aerial tracking shot smoothly tracks a massive dragon as it soars above a boundless, roiling sea of clouds. The golden afterglow of the setting sun paints the cloudtops a warm gold and soft pink, while the deep valleys between them sink into a serene indigo shade. The dragon's massive body is covered in countless cyan scales, each one like a finely crafted gem, shimmering in the light with iridescent hues ranging from deep turquoise to brilliant aquamarine. Its long, sinuous neck curves gracefully, its head crowned by towering obsidian horns. A pair of intelligent eyes, burning with amber light, calmly survey the world below. Its massive, resilient wing membranes are a translucent light blue, revealing their complex vein structure as the sunlight penetrates. With each powerful yet deliberate flap of its wings, its massive body glides effortlessly forward, its long, bony tail piercing it with a spur traces a graceful arc behind it, stirring the clouds in its wake. The entire scene is imbued with a sense of epic grandeur and tranquil power, creating an awe-inspiring atmosphere of ancient mythology.", + "seed": 3762, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_5/index_2061_seed_3762_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.009564650361426175, + "vlm_logprob": -0.009564650361426175, + "vlm_yes_no_margin": 5.025, + "vlm_token_credit": -0.009564650361426175, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a sweeping aerial tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon visible soaring above a boundless sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cloudtops painted in warm gold and soft pink with deep indigo valleys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's body covered in cyan scales with iridescent turquoise and aquamarine hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dragon flap its translucent light blue wings while gliding forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011995912529528141 + ], + "opposite_logprob_tokens": [ + -5.261995792388916 + ], + "logprob_reward": -0.011995912529528141, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.011995912529528141, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07677381485700607, + 0.12209996581077576, + 0.06239788606762886, + 0.04503946006298065, + 0.05951603502035141, + 0.045932888984680176, + 0.06082875281572342, + 0.1688649207353592, + 0.30878913402557373, + 0.049757156521081924 + ], + "credit_map": [ + 0.008268291130661964, + 0.0096916938200593, + 0.009846016764640808, + 0.007117386907339096, + 0.013442832976579666, + 0.008458529599010944, + 0.010269434191286564, + 0.009679632261395454, + 0.008185776881873608, + 0.008959869854152203, + 0.016648387536406517, + 0.01400166004896164, + 0.014096301048994064, + 0.014830428175628185, + 0.025106731802225113, + 0.020270809531211853, + 0.004758703988045454, + 0.0067755295895040035, + 0.01103162206709385, + 0.0065032439306378365, + 0.006839183624833822, + 0.010512800887227058, + 0.00993630662560463, + 0.006040493026375771, + 0.003834093688055873, + 0.004624285735189915, + 0.006138524506241083, + 0.0048296027816832066, + 0.005271516274660826, + 0.00806527677923441, + 0.006574267987161875, + 0.005701894406229258, + 0.006083824206143618, + 0.005022857338190079, + 0.008176233619451523, + 0.0051042973063886166, + 0.005281392950564623, + 0.014255586080253124, + 0.008475064300000668, + 0.007116785738617182, + 0.005944305099546909, + 0.0069930776953697205, + 0.005069959908723831, + 0.005204432178288698, + 0.005141108762472868, + 0.006856185384094715, + 0.004348252899944782, + 0.006375569384545088, + 0.007344451732933521, + 0.00766740133985877, + 0.011775701306760311, + 0.005003257188946009, + 0.005919375456869602, + 0.007006065454334021, + 0.007322308607399464, + 0.008790189400315285, + 0.017926866188645363, + 0.02662714570760727, + 0.018885444849729538, + 0.02443479560315609, + 0.01515942718833685, + 0.021554360166192055, + 0.025147832930088043, + 0.019129043444991112, + 0.05063496530056, + 0.02738925814628601, + 0.028174808248877525, + 0.0387302041053772, + 0.026428580284118652, + 0.06079358607530594, + 0.046523138880729675, + 0.03011459670960903, + 0.010799355804920197, + 0.0060346974059939384, + 0.005295656621456146, + 0.004732815083116293, + 0.005643452517688274, + 0.0056683095172047615, + 0.005492410622537136, + 0.006090455688536167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 1 + ], + "selected_reward": -0.011995912529528141 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a sweeping aerial tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon visible soaring above a boundless sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cloudtops painted in warm gold and soft pink with deep indigo valleys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's body covered in cyan scales with iridescent turquoise and aquamarine hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dragon flap its translucent light blue wings while gliding forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019528857665136456 + ], + "opposite_logprob_tokens": [ + -6.251952648162842 + ], + "logprob_reward": -0.0019528857665136456, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019528857665136456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10423837602138519, + 0.1439385563135147, + 0.09834782034158707, + 0.13517405092716217, + 0.085685595870018, + 0.025509364902973175, + 0.04386277124285698, + 0.1796582043170929, + 0.10012322664260864, + 0.08346199989318848 + ], + "credit_map": [ + 0.010698196478188038, + 0.013604845851659775, + 0.012961294502019882, + 0.009734752587974072, + 0.020938027650117874, + 0.011728496290743351, + 0.013402494601905346, + 0.01117027085274458, + 0.009934727102518082, + 0.016597246751189232, + 0.0242594126611948, + 0.01364654116332531, + 0.013498685322701931, + 0.015689710155129433, + 0.02849080041050911, + 0.02182144485414028, + 0.006375123746693134, + 0.007530273403972387, + 0.013649237342178822, + 0.007513845339417458, + 0.008746492676436901, + 0.012791061773896217, + 0.02636277861893177, + 0.015379020944237709, + 0.006923947483301163, + 0.009720283560454845, + 0.021691182628273964, + 0.011105027049779892, + 0.012868283316493034, + 0.03379569947719574, + 0.025342021137475967, + 0.013727623037993908, + 0.007184167858213186, + 0.013458747416734695, + 0.013688873499631882, + 0.009861821308732033, + 0.007661769632250071, + 0.011884684674441814, + 0.014924497343599796, + 0.0070210411213338375, + 0.002781163901090622, + 0.00259978580288589, + 0.0035028879065066576, + 0.0026992049533873796, + 0.0035649139899760485, + 0.004621502477675676, + 0.0027806442230939865, + 0.0029592635110020638, + 0.005058708600699902, + 0.004616978112608194, + 0.007837295532226562, + 0.003831700189039111, + 0.005488211289048195, + 0.005416742525994778, + 0.006127285771071911, + 0.005485852248966694, + 0.019887195900082588, + 0.021757822483778, + 0.01811445690691471, + 0.025394165888428688, + 0.01738390326499939, + 0.017638444900512695, + 0.035988688468933105, + 0.023493541404604912, + 0.018039993941783905, + 0.00980164110660553, + 0.0078220684081316, + 0.00971030630171299, + 0.008558656089007854, + 0.021502230316400528, + 0.01568477600812912, + 0.00900355726480484, + 0.016204379498958588, + 0.015688162297010422, + 0.007460369728505611, + 0.006906033027917147, + 0.006121529266238213, + 0.009324639104306698, + 0.008998574689030647, + 0.012758315540850163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 3 + ], + "selected_reward": -0.0019528857665136456 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a sweeping aerial tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon visible soaring above a boundless sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cloudtops painted in warm gold and soft pink with deep indigo valleys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's body covered in cyan scales with iridescent turquoise and aquamarine hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dragon flap its translucent light blue wings while gliding forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0160687193274498 + ], + "opposite_logprob_tokens": [ + -4.141068935394287 + ], + "logprob_reward": -0.0160687193274498, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0160687193274498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08760058879852295, + 0.21149280667304993, + 0.08155510574579239, + 0.04770432412624359, + 0.04663146287202835, + 0.03523540124297142, + 0.024431932717561722, + 0.13917329907417297, + 0.21652056276798248, + 0.1096544861793518 + ], + "credit_map": [ + 0.008735260926187038, + 0.007353165186941624, + 0.019038524478673935, + 0.008211643435060978, + 0.014011019840836525, + 0.010986042208969593, + 0.01128319650888443, + 0.007981734350323677, + 0.017432259395718575, + 0.01750895380973816, + 0.02613966539502144, + 0.022618163377046585, + 0.030937187373638153, + 0.02995024062693119, + 0.03850528597831726, + 0.02840103954076767, + 0.005483850371092558, + 0.005321995820850134, + 0.013163045048713684, + 0.004387132823467255, + 0.008236751891672611, + 0.020988425239920616, + 0.011950166895985603, + 0.012023745104670525, + 0.003770894603803754, + 0.0046906424686312675, + 0.007807777728885412, + 0.004232687409967184, + 0.005089078098535538, + 0.007940702140331268, + 0.008846037089824677, + 0.005326503422111273, + 0.005225024186074734, + 0.005750380922108889, + 0.007518493104726076, + 0.006288873963057995, + 0.004154319409281015, + 0.007352936547249556, + 0.005154499318450689, + 0.005186932627111673, + 0.0037896072026342154, + 0.00418616458773613, + 0.004847843665629625, + 0.0033600260503590107, + 0.004760488867759705, + 0.0067573185078799725, + 0.00371453445404768, + 0.003819417906925082, + 0.003113353392109275, + 0.002478028880432248, + 0.0028661105316132307, + 0.002986927516758442, + 0.0035352495033293962, + 0.0037815559189766645, + 0.0030122511088848114, + 0.002658453304320574, + 0.016699491068720818, + 0.02456938847899437, + 0.015691116452217102, + 0.015453464351594448, + 0.01738668605685234, + 0.017631666734814644, + 0.015437609516084194, + 0.01630386710166931, + 0.038677968084812164, + 0.018180329352617264, + 0.017976481467485428, + 0.023840533569455147, + 0.018262162804603577, + 0.04689979925751686, + 0.031856197863817215, + 0.020827079191803932, + 0.017338590696454048, + 0.026128217577934265, + 0.0052671586163342, + 0.006489449646323919, + 0.007421793416142464, + 0.007310611195862293, + 0.009055131115019321, + 0.0306435264647007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 7 + ], + "selected_reward": -0.0160687193274498 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a sweeping aerial tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon visible soaring above a boundless sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cloudtops painted in warm gold and soft pink with deep indigo valleys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's body covered in cyan scales with iridescent turquoise and aquamarine hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dragon flap its translucent light blue wings while gliding forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734057795256376 + ], + "opposite_logprob_tokens": [ + -5.0067338943481445 + ], + "logprob_reward": -0.006734057795256376, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734057795256376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1846475452184677, + 0.04723271355032921, + 0.01915440522134304, + 0.10166726261377335, + 0.1292937844991684, + 0.042985789477825165, + 0.023268159478902817, + 0.02988259680569172, + 0.1519012600183487, + 0.2699664831161499 + ], + "credit_map": [ + 0.015531585551798344, + 0.021821279078722, + 0.02202264405786991, + 0.01687590777873993, + 0.04333319887518883, + 0.021602125838398933, + 0.023463720455765724, + 0.019997093826532364, + 0.003014581510797143, + 0.0029711457900702953, + 0.004987036809325218, + 0.004334606230258942, + 0.004650626797229052, + 0.004315507132560015, + 0.012107823975384235, + 0.010851385071873665, + 0.0024952595122158527, + 0.002661981387063861, + 0.0023904223926365376, + 0.0015092657413333654, + 0.0026510071475058794, + 0.00291361752897501, + 0.002527512377128005, + 0.0020053391344845295, + 0.0054529812186956406, + 0.007334625814110041, + 0.016132812947034836, + 0.008078095503151417, + 0.010038192383944988, + 0.02457105554640293, + 0.019887486472725868, + 0.010172017849981785, + 0.010412589646875858, + 0.01888751983642578, + 0.02014206349849701, + 0.011592797935009003, + 0.012378516606986523, + 0.01848423480987549, + 0.023706045001745224, + 0.013690004125237465, + 0.005330762360244989, + 0.005972030106931925, + 0.0064461310394108295, + 0.005639028735458851, + 0.0040225861594080925, + 0.0069849491119384766, + 0.003762394655495882, + 0.004827904514968395, + 0.003124580718576908, + 0.002451136242598295, + 0.0034213934559375048, + 0.0024374607019126415, + 0.0027718532364815474, + 0.0029346782248467207, + 0.003565338207408786, + 0.0025617193896323442, + 0.003528333269059658, + 0.005238588433712721, + 0.0024822501000016928, + 0.003965346142649651, + 0.0035850955173373222, + 0.0037818499840795994, + 0.0032336630392819643, + 0.004067470785230398, + 0.02679121494293213, + 0.013177608139812946, + 0.013032340444624424, + 0.01651054248213768, + 0.01293690875172615, + 0.03362962603569031, + 0.021117161959409714, + 0.014705849811434746, + 0.035365357995033264, + 0.06350292265415192, + 0.013209213502705097, + 0.010503358207643032, + 0.022855451330542564, + 0.023562906309962273, + 0.017609549686312675, + 0.08335772901773453 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.006734057795256376 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a sweeping aerial tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon visible soaring above a boundless sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cloudtops painted in warm gold and soft pink with deep indigo valleys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's body covered in cyan scales with iridescent turquoise and aquamarine hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dragon flap its translucent light blue wings while gliding forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011071676388382912 + ], + "opposite_logprob_tokens": [ + -4.511071681976318 + ], + "logprob_reward": -0.011071676388382912, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011071676388382912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0753050446510315, + 0.06520723551511765, + 0.11943315714597702, + 0.12525558471679688, + 0.12824977934360504, + 0.04358670860528946, + 0.0514121912419796, + 0.09451568126678467, + 0.23683670163154602, + 0.06019780412316322 + ], + "credit_map": [ + 0.008117909543216228, + 0.007813490927219391, + 0.009465810842812061, + 0.008505723439157009, + 0.015586864203214645, + 0.00843307375907898, + 0.009942437522113323, + 0.0074397423304617405, + 0.004299560096114874, + 0.006740165408700705, + 0.008281459100544453, + 0.006862211506813765, + 0.006046791095286608, + 0.006993670016527176, + 0.014134231023490429, + 0.011849154718220234, + 0.014716651290655136, + 0.011481069028377533, + 0.018817773088812828, + 0.010484561324119568, + 0.01611565239727497, + 0.0165341068059206, + 0.019822433590888977, + 0.011460927315056324, + 0.008423703722655773, + 0.009962665848433971, + 0.01995447278022766, + 0.014208202250301838, + 0.012959650717675686, + 0.027715258300304413, + 0.020097065716981888, + 0.011934584937989712, + 0.011531379073858261, + 0.01896056905388832, + 0.020522022619843483, + 0.011962920427322388, + 0.009574972093105316, + 0.02161974459886551, + 0.02265382744371891, + 0.011424364522099495, + 0.004967757035046816, + 0.005504370667040348, + 0.006200952921062708, + 0.005817533936351538, + 0.005349067971110344, + 0.006550754886120558, + 0.005480162799358368, + 0.0037161146756261587, + 0.005962451919913292, + 0.00546286441385746, + 0.011165818199515343, + 0.004534176550805569, + 0.0056221638806164265, + 0.006599240005016327, + 0.006188902538269758, + 0.005876580718904734, + 0.010692381300032139, + 0.014202970080077648, + 0.009994265623390675, + 0.010881289839744568, + 0.010051501914858818, + 0.01318549644201994, + 0.012210641987621784, + 0.013297141529619694, + 0.033636223524808884, + 0.02073894627392292, + 0.023737939074635506, + 0.03349724039435387, + 0.022655345499515533, + 0.0366181880235672, + 0.04143737256526947, + 0.024515481665730476, + 0.011541174724698067, + 0.013555902987718582, + 0.005482925102114677, + 0.0043417164124548435, + 0.006051329430192709, + 0.006715308874845505, + 0.005947224795818329, + 0.006562230177223682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 3 + ], + "selected_reward": -0.011071676388382912 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping aerial tracking shot smoothly tracks a massive dragon as it soars above a boundless, roiling sea of clouds. The golden afterglow of the setting sun paints the cloudtops a warm gold and soft pink, while the deep valleys between them sink into a serene indigo shade. The dragon's massive body is covered in countless cyan scales, each one like a finely crafted gem, shimmering in the light with iridescent hues ranging from deep turquoise to brilliant aquamarine. Its long, sinuous neck curves gracefully, its head crowned by towering obsidian horns. A pair of intelligent eyes, burning with amber light, calmly survey the world below. Its massive, resilient wing membranes are a translucent light blue, revealing their complex vein structure as the sunlight penetrates. With each powerful yet deliberate flap of its wings, its massive body glides effortlessly forward, its long, bony tail piercing it with a spur traces a graceful arc behind it, stirring the clouds in its wake. The entire scene is imbued with a sense of epic grandeur and tranquil power, creating an awe-inspiring atmosphere of ancient mythology.", + "seed": 3763, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_5/index_2061_seed_3763_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.010251575801521539, + "vlm_logprob": -0.010251575801521539, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.010251575801521539, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a sweeping aerial tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon visible soaring above a boundless sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cloudtops painted in warm gold and soft pink with deep indigo valleys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's body covered in cyan scales with iridescent turquoise and aquamarine hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dragon flap its translucent light blue wings while gliding forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02095009945333004 + ], + "opposite_logprob_tokens": [ + -4.2709503173828125 + ], + "logprob_reward": -0.02095009945333004, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.02095009945333004, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07324577122926712, + 0.028715044260025024, + 0.04327016323804855, + 0.04784176126122475, + 0.2954016625881195, + 0.04821091145277023, + 0.15101835131645203, + 0.07372963428497314, + 0.14214283227920532, + 0.09642383456230164 + ], + "credit_map": [ + 0.01400180533528328, + 0.006647863890975714, + 0.012791336514055729, + 0.006151873152703047, + 0.006187522783875465, + 0.010655565187335014, + 0.007379383314400911, + 0.009430419653654099, + 0.005093664862215519, + 0.0022479586768895388, + 0.002858934225514531, + 0.0034781156573444605, + 0.002814345294609666, + 0.004148043226450682, + 0.0036170496605336666, + 0.004456934053450823, + 0.003460019826889038, + 0.0030043190345168114, + 0.004084975924342871, + 0.007411150727421045, + 0.003309610765427351, + 0.004006090108305216, + 0.009167599491775036, + 0.008826399222016335, + 0.002693233313038945, + 0.007759032305330038, + 0.008341553620994091, + 0.00734383799135685, + 0.0037795568350702524, + 0.0059692226350307465, + 0.006080572959035635, + 0.005874753929674625, + 0.006845132913440466, + 0.009190797805786133, + 0.030824361369013786, + 0.013256918638944626, + 0.011297622695565224, + 0.018817629665136337, + 0.11296204477548599, + 0.0922071784734726, + 0.00579831562936306, + 0.004016389604657888, + 0.004547412041574717, + 0.005472312215715647, + 0.00938939768821001, + 0.008517561480402946, + 0.00443927152082324, + 0.0060302503407001495, + 0.017817534506320953, + 0.019516875967383385, + 0.02098015882074833, + 0.012769294902682304, + 0.015592988580465317, + 0.01801077090203762, + 0.020538324490189552, + 0.025792403146624565, + 0.011741346679627895, + 0.012841112911701202, + 0.008481702767312527, + 0.006502612493932247, + 0.007434403523802757, + 0.00854630209505558, + 0.007534740027040243, + 0.010647411458194256, + 0.01663697138428688, + 0.011113754473626614, + 0.014402643777430058, + 0.015477809123694897, + 0.014856548048555851, + 0.02010306715965271, + 0.026570765301585197, + 0.022981278598308563, + 0.009965349920094013, + 0.020431578159332275, + 0.009579461999237537, + 0.012523261830210686, + 0.007739757653325796, + 0.009715155698359013, + 0.014599348418414593, + 0.01186993159353733 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 8 + ], + "selected_reward": -0.02095009945333004 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a sweeping aerial tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon visible soaring above a boundless sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cloudtops painted in warm gold and soft pink with deep indigo valleys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's body covered in cyan scales with iridescent turquoise and aquamarine hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dragon flap its translucent light blue wings while gliding forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00865151360630989 + ], + "opposite_logprob_tokens": [ + -4.7586517333984375 + ], + "logprob_reward": -0.00865151360630989, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00865151360630989, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14253157377243042, + 0.11357740312814713, + 0.06839005649089813, + 0.08117900043725967, + 0.19974736869335175, + 0.04598499462008476, + 0.15023362636566162, + 0.07226992398500443, + 0.07711458206176758, + 0.0489715039730072 + ], + "credit_map": [ + 0.01706136204302311, + 0.011645506136119366, + 0.02337576076388359, + 0.013579057529568672, + 0.012624554336071014, + 0.026606477797031403, + 0.01902776025235653, + 0.01861107535660267, + 0.024828067049384117, + 0.010419105179607868, + 0.01019735261797905, + 0.01320570893585682, + 0.00944880023598671, + 0.017023874446749687, + 0.012804110534489155, + 0.015650371089577675, + 0.010547990910708904, + 0.004289278760552406, + 0.008825472556054592, + 0.011533846147358418, + 0.006685349158942699, + 0.007230360526591539, + 0.010276135057210922, + 0.009001625701785088, + 0.004571782425045967, + 0.01163893099874258, + 0.011037647724151611, + 0.007000120356678963, + 0.005977773107588291, + 0.017179690301418304, + 0.010916411876678467, + 0.012856640852987766, + 0.0072743697091937065, + 0.011372575536370277, + 0.02105380967259407, + 0.013895454816520214, + 0.009394202381372452, + 0.015200295485556126, + 0.06677522510290146, + 0.05478142574429512, + 0.005292695015668869, + 0.005024215206503868, + 0.004573649261146784, + 0.005971498787403107, + 0.008057406172156334, + 0.007037782110273838, + 0.004433588124811649, + 0.00559415714815259, + 0.018197830766439438, + 0.019751567393541336, + 0.015321663580834866, + 0.01763065904378891, + 0.02012021094560623, + 0.02218014933168888, + 0.01736750267446041, + 0.019664039835333824, + 0.014648514799773693, + 0.008526667952537537, + 0.006925014313310385, + 0.005455699749290943, + 0.006394114810973406, + 0.004805956035852432, + 0.011179335415363312, + 0.014334617182612419, + 0.0075457338243722916, + 0.007269824855029583, + 0.006699095480144024, + 0.008945460431277752, + 0.00801294855773449, + 0.010696222074329853, + 0.015448715537786484, + 0.012496582232415676, + 0.005497310310602188, + 0.008224552497267723, + 0.006731375120580196, + 0.006709143053740263, + 0.003946325276046991, + 0.0043328735046088696, + 0.0072618029080331326, + 0.006268121302127838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.00865151360630989 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a sweeping aerial tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon visible soaring above a boundless sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cloudtops painted in warm gold and soft pink with deep indigo valleys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's body covered in cyan scales with iridescent turquoise and aquamarine hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dragon flap its translucent light blue wings while gliding forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005948222242295742 + ], + "opposite_logprob_tokens": [ + -5.130948066711426 + ], + "logprob_reward": -0.005948222242295742, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005948222242295742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15479695796966553, + 0.050521064549684525, + 0.05652133375406265, + 0.13279347121715546, + 0.16053740680217743, + 0.12498125433921814, + 0.10633356124162674, + 0.07032723724842072, + 0.07828125357627869, + 0.06490644812583923 + ], + "credit_map": [ + 0.01808558590710163, + 0.013086601160466671, + 0.019510623067617416, + 0.01414581760764122, + 0.016297727823257446, + 0.028093168511986732, + 0.025067927315831184, + 0.0205095037817955, + 0.006730466615408659, + 0.007468685507774353, + 0.005133866332471371, + 0.006431330926716328, + 0.005696343258023262, + 0.005953934974968433, + 0.0063311392441391945, + 0.00677530188113451, + 0.005329860839992762, + 0.004992992151528597, + 0.005690779536962509, + 0.010331692174077034, + 0.0047486890107393265, + 0.00565373944118619, + 0.009749757125973701, + 0.01002382580190897, + 0.0062777078710496426, + 0.013789641670882702, + 0.01903364062309265, + 0.011053387075662613, + 0.006014156620949507, + 0.02047736756503582, + 0.024445097893476486, + 0.03170246258378029, + 0.005429170560091734, + 0.012186937965452671, + 0.014489946886897087, + 0.010087314993143082, + 0.007107231765985489, + 0.014904960989952087, + 0.050249308347702026, + 0.04608255252242088, + 0.01286089513450861, + 0.014966452494263649, + 0.01196411531418562, + 0.014619351364672184, + 0.021469293162226677, + 0.019149914383888245, + 0.013962263241410255, + 0.015988977625966072, + 0.014498603530228138, + 0.011421537958085537, + 0.015207812190055847, + 0.008662444539368153, + 0.010565443895757198, + 0.01396375522017479, + 0.014261928386986256, + 0.017752036452293396, + 0.01369076780974865, + 0.00855658482760191, + 0.006485347170382738, + 0.004651362542062998, + 0.006369428709149361, + 0.005970089230686426, + 0.010001088492572308, + 0.014602569863200188, + 0.010080909356474876, + 0.007281745318323374, + 0.008754542097449303, + 0.010696149431169033, + 0.006365809589624405, + 0.008951638825237751, + 0.013738187961280346, + 0.01241227425634861, + 0.007069042883813381, + 0.008601776324212551, + 0.006502868607640266, + 0.007869139313697815, + 0.006354551296681166, + 0.009996997192502022, + 0.008678070269525051, + 0.00983400084078312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.005948222242295742 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a sweeping aerial tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon visible soaring above a boundless sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cloudtops painted in warm gold and soft pink with deep indigo valleys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's body covered in cyan scales with iridescent turquoise and aquamarine hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dragon flap its translucent light blue wings while gliding forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004635777790099382 + ], + "opposite_logprob_tokens": [ + -5.379635810852051 + ], + "logprob_reward": -0.004635777790099382, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004635777790099382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05130862817168236, + 0.07632521539926529, + 0.050030212849378586, + 0.03544149175286293, + 0.10630979388952255, + 0.06254752725362778, + 0.10812339931726456, + 0.2811242341995239, + 0.1202920600771904, + 0.10849740356206894 + ], + "credit_map": [ + 0.005628480110317469, + 0.004991321824491024, + 0.006574307568371296, + 0.006225281860679388, + 0.005420570727437735, + 0.007182687986642122, + 0.008512888103723526, + 0.006773088593035936, + 0.008617079816758633, + 0.007720125373452902, + 0.010124760679900646, + 0.010909688659012318, + 0.01075271237641573, + 0.011783120222389698, + 0.008517027832567692, + 0.007900700904428959, + 0.0064903381280601025, + 0.003505920758470893, + 0.004891255870461464, + 0.011270671151578426, + 0.0038555075880140066, + 0.004051735624670982, + 0.008061118423938751, + 0.00790366716682911, + 0.002074613468721509, + 0.003972529899328947, + 0.004720158409327269, + 0.004397702869027853, + 0.001984288450330496, + 0.005038162227720022, + 0.005176435224711895, + 0.008077600039541721, + 0.004568484611809254, + 0.005406945012509823, + 0.011361434124410152, + 0.0073964158073067665, + 0.005106001626700163, + 0.008147741667926311, + 0.03445304185152054, + 0.02986973337829113, + 0.007136577740311623, + 0.006238757632672787, + 0.006462670862674713, + 0.007303198799490929, + 0.00816370826214552, + 0.012319379486143589, + 0.007007545325905085, + 0.007915688678622246, + 0.015975283458828926, + 0.016365066170692444, + 0.011294800788164139, + 0.011605551466345787, + 0.012090393342077732, + 0.014408931136131287, + 0.012516967952251434, + 0.013866405934095383, + 0.04108813405036926, + 0.035060424357652664, + 0.01889982260763645, + 0.021878542378544807, + 0.027800776064395905, + 0.025645378977060318, + 0.0634971410036087, + 0.04725400358438492, + 0.010281512513756752, + 0.009509867988526821, + 0.013132471591234207, + 0.017494460567831993, + 0.01141664944589138, + 0.023625191301107407, + 0.01700553111732006, + 0.017826378345489502, + 0.015017244964838028, + 0.01199161447584629, + 0.013218927197158337, + 0.015575726516544819, + 0.012160570360720158, + 0.01175275631248951, + 0.016246965155005455, + 0.012533598579466343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.004635777790099382 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a sweeping aerial tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon visible soaring above a boundless sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cloudtops painted in warm gold and soft pink with deep indigo valleys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's body covered in cyan scales with iridescent turquoise and aquamarine hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dragon flap its translucent light blue wings while gliding forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072265915572643 + ], + "opposite_logprob_tokens": [ + -4.511072158813477 + ], + "logprob_reward": -0.011072265915572643, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072265915572643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16440525650978088, + 0.054631758481264114, + 0.061479486525058746, + 0.07711409032344818, + 0.12821203470230103, + 0.05712196230888367, + 0.14330273866653442, + 0.08066481351852417, + 0.13248883187770844, + 0.10057902336120605 + ], + "credit_map": [ + 0.03687504678964615, + 0.01799398474395275, + 0.02222132496535778, + 0.016786837950348854, + 0.015250778757035732, + 0.023223256692290306, + 0.014397663995623589, + 0.017656359821558, + 0.007799671962857246, + 0.005483033135533333, + 0.0053999037481844425, + 0.0074022733606398106, + 0.00719849206507206, + 0.007178068161010742, + 0.006141228601336479, + 0.008029085583984852, + 0.005548873450607061, + 0.004875244572758675, + 0.006006191484630108, + 0.013777748681604862, + 0.0070104836486279964, + 0.006010301411151886, + 0.010840009897947311, + 0.00741063104942441, + 0.005056251771748066, + 0.006846201606094837, + 0.008273177780210972, + 0.010421364568173885, + 0.006701069418340921, + 0.01224415935575962, + 0.00885001104325056, + 0.018721850588917732, + 0.0035734844859689474, + 0.005305744707584381, + 0.012283974327147007, + 0.006136362906545401, + 0.006240581627935171, + 0.008432322181761265, + 0.04685412719845772, + 0.039385441690683365, + 0.0058217537589371204, + 0.005485770758241415, + 0.006023669149726629, + 0.007901862263679504, + 0.008284817449748516, + 0.009044351056218147, + 0.007325663696974516, + 0.007234079763293266, + 0.017322326079010963, + 0.01649184711277485, + 0.01845584437251091, + 0.011978554539382458, + 0.015919381752610207, + 0.016419826075434685, + 0.021778713911771774, + 0.024936242029070854, + 0.014644790440797806, + 0.009893073700368404, + 0.00784453097730875, + 0.006854720879346132, + 0.007669794373214245, + 0.007571708410978317, + 0.013746351934969425, + 0.012439834885299206, + 0.013105605728924274, + 0.010745498351752758, + 0.014138303697109222, + 0.019091732800006866, + 0.01231037825345993, + 0.019358081743121147, + 0.023110616952180862, + 0.020628614351153374, + 0.00986399780958891, + 0.01789429783821106, + 0.013663218356668949, + 0.01417368184775114, + 0.008431497029960155, + 0.008210677653551102, + 0.017043283209204674, + 0.011298366822302341 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.011072265915572643 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping aerial tracking shot smoothly tracks a massive dragon as it soars above a boundless, roiling sea of clouds. The golden afterglow of the setting sun paints the cloudtops a warm gold and soft pink, while the deep valleys between them sink into a serene indigo shade. The dragon's massive body is covered in countless cyan scales, each one like a finely crafted gem, shimmering in the light with iridescent hues ranging from deep turquoise to brilliant aquamarine. Its long, sinuous neck curves gracefully, its head crowned by towering obsidian horns. A pair of intelligent eyes, burning with amber light, calmly survey the world below. Its massive, resilient wing membranes are a translucent light blue, revealing their complex vein structure as the sunlight penetrates. With each powerful yet deliberate flap of its wings, its massive body glides effortlessly forward, its long, bony tail piercing it with a spur traces a graceful arc behind it, stirring the clouds in its wake. The entire scene is imbued with a sense of epic grandeur and tranquil power, creating an awe-inspiring atmosphere of ancient mythology.", + "seed": 3764, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_5/index_2061_seed_3764_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.013369798753410577, + "vlm_logprob": -0.013369798753410577, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.013369798753410577, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a sweeping aerial tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon visible soaring above a boundless sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cloudtops painted in warm gold and soft pink with deep indigo valleys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's body covered in cyan scales with iridescent turquoise and aquamarine hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dragon flap its translucent light blue wings while gliding forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02584006078541279 + ], + "opposite_logprob_tokens": [ + -3.9008400440216064 + ], + "logprob_reward": -0.02584006078541279, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02584006078541279, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09452294558286667, + 0.025085950270295143, + 0.04711022228002548, + 0.02658068761229515, + 0.6026584506034851, + 0.07731153815984726, + 0.03572353720664978, + 0.01580015942454338, + 0.05823608115315437, + 0.016970379278063774 + ], + "credit_map": [ + 0.0069580962881445885, + 0.009818016551434994, + 0.018070559948682785, + 0.011740412563085556, + 0.008270949125289917, + 0.014611035585403442, + 0.011820591054856777, + 0.013233286328613758, + 0.0030663914512842894, + 0.0036135781556367874, + 0.005085306242108345, + 0.0024874822702258825, + 0.002169026294723153, + 0.0035654499661177397, + 0.0029497428331524134, + 0.0021489732898771763, + 0.005082776304334402, + 0.002572617493569851, + 0.012025888077914715, + 0.005332481116056442, + 0.0030870388727635145, + 0.014125009067356586, + 0.002149621956050396, + 0.002734788693487644, + 0.004617845173925161, + 0.003059182781726122, + 0.0036619279999285936, + 0.0033283145166933537, + 0.003128127893432975, + 0.003963434603065252, + 0.0019458738388493657, + 0.002875980921089649, + 0.04942142590880394, + 0.05555909126996994, + 0.19360370934009552, + 0.029885662719607353, + 0.013233101926743984, + 0.06371767818927765, + 0.04182484745979309, + 0.1554129421710968, + 0.008119476959109306, + 0.006942604668438435, + 0.014350270852446556, + 0.0068063498474657536, + 0.004583160392940044, + 0.01042400673031807, + 0.014780028723180294, + 0.011305633932352066, + 0.008233041502535343, + 0.0025161011144518852, + 0.005895978771150112, + 0.001955475891008973, + 0.0029116463847458363, + 0.0027573262341320515, + 0.005747214891016483, + 0.005706753581762314, + 0.0024387717712670565, + 0.0014259304152801633, + 0.0022781873121857643, + 0.00176798936445266, + 0.0009491522796452045, + 0.0022034163121134043, + 0.0023026885464787483, + 0.002434022491797805, + 0.006409723777323961, + 0.00566783407703042, + 0.004562250338494778, + 0.0039529153145849705, + 0.005550691857933998, + 0.01262142788618803, + 0.007301449775695801, + 0.012169786728918552, + 0.0018232434522360563, + 0.0019497304456308484, + 0.0023432429879903793, + 0.0015303679974749684, + 0.0018857698887586594, + 0.00234712241217494, + 0.0022515705786645412, + 0.0028393305838108063 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.02584006078541279 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a sweeping aerial tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon visible soaring above a boundless sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cloudtops painted in warm gold and soft pink with deep indigo valleys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's body covered in cyan scales with iridescent turquoise and aquamarine hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dragon flap its translucent light blue wings while gliding forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005961731541901827 + ], + "opposite_logprob_tokens": [ + -5.130961894989014 + ], + "logprob_reward": -0.005961731541901827, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005961731541901827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1831187605857849, + 0.07430814206600189, + 0.10903158783912659, + 0.044945210218429565, + 0.3710096478462219, + 0.06919781118631363, + 0.025573058053851128, + 0.022706275805830956, + 0.08524969220161438, + 0.014859856106340885 + ], + "credit_map": [ + 0.013055557385087013, + 0.0141759617254138, + 0.027310557663440704, + 0.02406051754951477, + 0.015722231939435005, + 0.04202248901128769, + 0.0232378002256155, + 0.023533638566732407, + 0.007397798355668783, + 0.008118359372019768, + 0.009036186151206493, + 0.009920335374772549, + 0.005508516449481249, + 0.01852640137076378, + 0.006449534557759762, + 0.009351005777716637, + 0.014607474207878113, + 0.007085603196173906, + 0.015227866359055042, + 0.019540589302778244, + 0.007997422479093075, + 0.02690771408379078, + 0.007361415773630142, + 0.010303499177098274, + 0.004382031969726086, + 0.004954845178872347, + 0.008081088773906231, + 0.00463164271786809, + 0.004017155151814222, + 0.0069229742512106895, + 0.005729206372052431, + 0.0062262630090117455, + 0.03007366694509983, + 0.0334833562374115, + 0.11786423623561859, + 0.018654469400644302, + 0.00928869191557169, + 0.0392807200551033, + 0.026126088574528694, + 0.09623844921588898, + 0.007888471707701683, + 0.00568485539406538, + 0.00668672239407897, + 0.005781731568276882, + 0.0045445580035448074, + 0.005312385968863964, + 0.020584462210536003, + 0.012714625336229801, + 0.004910819698125124, + 0.0018586795777082443, + 0.004300880245864391, + 0.0017801851499825716, + 0.0016999350627884269, + 0.0019045162480324507, + 0.005077983718365431, + 0.004040057770907879, + 0.0023300182074308395, + 0.0012954435078427196, + 0.0037653418257832527, + 0.0023906559217721224, + 0.0014171568909659982, + 0.002667828695848584, + 0.00541525287553668, + 0.003424578346312046, + 0.009666066616773605, + 0.004279470536857843, + 0.007754963357001543, + 0.005330677144229412, + 0.00951287429779768, + 0.021813087165355682, + 0.010542945004999638, + 0.016349604353308678, + 0.0014052360784262419, + 0.0018239583587273955, + 0.0017215387197211385, + 0.0012275711633265018, + 0.0013242835411801934, + 0.002187746576964855, + 0.0023516991641372442, + 0.0028178223874419928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.005961731541901827 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a sweeping aerial tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon visible soaring above a boundless sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cloudtops painted in warm gold and soft pink with deep indigo valleys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's body covered in cyan scales with iridescent turquoise and aquamarine hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dragon flap its translucent light blue wings while gliding forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195182360708714 + ], + "opposite_logprob_tokens": [ + -4.264194965362549 + ], + "logprob_reward": -0.014195182360708714, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195182360708714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07256495952606201, + 0.030320649966597557, + 0.03219952806830406, + 0.02396034263074398, + 0.6316856741905212, + 0.015619230456650257, + 0.07406485080718994, + 0.00824018381536007, + 0.10642215609550476, + 0.004922396503388882 + ], + "credit_map": [ + 0.005222149193286896, + 0.006153253372758627, + 0.014277324080467224, + 0.0094422223046422, + 0.005376823246479034, + 0.014728139154613018, + 0.00828452780842781, + 0.009080528281629086, + 0.00380010181106627, + 0.0037259929813444614, + 0.005568533204495907, + 0.004063589032739401, + 0.002124740509316325, + 0.005409783683717251, + 0.0020200242288410664, + 0.003607887076213956, + 0.003297412069514394, + 0.001975074876099825, + 0.008640637621283531, + 0.002555662067607045, + 0.002026286907494068, + 0.0100401034578681, + 0.0015401728451251984, + 0.0021241807844489813, + 0.0031256009824573994, + 0.002492241794243455, + 0.0026822304353117943, + 0.003004721598699689, + 0.0028292827773839235, + 0.003588806837797165, + 0.003124932060018182, + 0.003112529870122671, + 0.05226865038275719, + 0.057913247495889664, + 0.20288234949111938, + 0.03039398416876793, + 0.013331501744687557, + 0.0666244775056839, + 0.044039517641067505, + 0.16423198580741882, + 0.0015584988286718726, + 0.0012385636800900102, + 0.0034258163068443537, + 0.0013562452513724566, + 0.0009131580009125173, + 0.0021180433686822653, + 0.002751612104475498, + 0.0022572935558855534, + 0.012710710987448692, + 0.005711250472813845, + 0.012299853377044201, + 0.00535165797919035, + 0.008119101636111736, + 0.007644275669008493, + 0.011952049098908901, + 0.010275961831212044, + 0.0007171272300183773, + 0.001030638930387795, + 0.0009752946207299829, + 0.0008814805187284946, + 0.0007276867399923503, + 0.0013319678837433457, + 0.0013849990209564567, + 0.0011909888125956059, + 0.011813712306320667, + 0.006247757934033871, + 0.010673447512090206, + 0.007755001075565815, + 0.009684030897915363, + 0.0215053241699934, + 0.013607235625386238, + 0.025135651230812073, + 0.0005257516750134528, + 0.000567536277230829, + 0.000635625037830323, + 0.000385404477128759, + 0.00037885212805122137, + 0.0005254194838926196, + 0.0007482381770387292, + 0.0011555696837604046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 6 + ], + "selected_reward": -0.014195182360708714 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a sweeping aerial tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon visible soaring above a boundless sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cloudtops painted in warm gold and soft pink with deep indigo valleys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's body covered in cyan scales with iridescent turquoise and aquamarine hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dragon flap its translucent light blue wings while gliding forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009778574109077454 + ], + "opposite_logprob_tokens": [ + -4.6347784996032715 + ], + "logprob_reward": -0.009778574109077454, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009778574109077454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03328738361597061, + 0.07388772815465927, + 0.03103967197239399, + 0.022074410691857338, + 0.7347748279571533, + 0.05792997404932976, + 0.013019046746194363, + 0.014579576440155506, + 0.012556721456348896, + 0.006850742269307375 + ], + "credit_map": [ + 0.0027424839790910482, + 0.003632141277194023, + 0.006978416349738836, + 0.004385985434055328, + 0.0027333060279488564, + 0.003449140815064311, + 0.004051653202623129, + 0.0053142523393034935, + 0.008356922306120396, + 0.008878922089934349, + 0.013414015993475914, + 0.010441887192428112, + 0.005095443222671747, + 0.016673635691404343, + 0.004715339280664921, + 0.006311559118330479, + 0.0029884339310228825, + 0.0014709234237670898, + 0.010108960792422295, + 0.001733994809910655, + 0.00169148831628263, + 0.009703107178211212, + 0.0015120008029043674, + 0.0018307616701349616, + 0.003484155284240842, + 0.00445460295304656, + 0.0028873928822577, + 0.0025865717325359583, + 0.0024037088733166456, + 0.0026950391475111246, + 0.0017127480823546648, + 0.0018501897575333714, + 0.060536522418260574, + 0.06710173189640045, + 0.23500974476337433, + 0.03559041768312454, + 0.016332440078258514, + 0.07704991102218628, + 0.051319267600774765, + 0.19183474779129028, + 0.006034738849848509, + 0.0076026879251003265, + 0.010229303501546383, + 0.004720728844404221, + 0.004733613692224026, + 0.006839371286332607, + 0.01059916615486145, + 0.007170361001044512, + 0.0020206456538289785, + 0.0015138897579163313, + 0.001893086708150804, + 0.0009560135658830404, + 0.001317511429078877, + 0.0013927002437412739, + 0.001916062319651246, + 0.0020091356709599495, + 0.0019945725798606873, + 0.0017770384438335896, + 0.0022603990510106087, + 0.0018860041163861752, + 0.0011704216012731194, + 0.0013029592810198665, + 0.001899703755043447, + 0.002288475865498185, + 0.0011971357744187117, + 0.000654056086204946, + 0.0011331512359902263, + 0.0009361577685922384, + 0.001358075300231576, + 0.0029583817813545465, + 0.0018279077485203743, + 0.002491855062544346, + 0.0008149015484377742, + 0.0008170509245246649, + 0.0008063487475737929, + 0.0006185937090776861, + 0.0006018647109158337, + 0.0011021023383364081, + 0.0008375377510674298, + 0.0012523423647508025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 5 + ], + "selected_reward": -0.009778574109077454 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a sweeping aerial tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon visible soaring above a boundless sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cloudtops painted in warm gold and soft pink with deep indigo valleys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's body covered in cyan scales with iridescent turquoise and aquamarine hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dragon flap its translucent light blue wings while gliding forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011073444969952106 + ], + "opposite_logprob_tokens": [ + -4.511073589324951 + ], + "logprob_reward": -0.011073444969952106, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011073444969952106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0908866599202156, + 0.06157328188419342, + 0.0481484979391098, + 0.03134353458881378, + 0.6119059324264526, + 0.04072748124599457, + 0.07293560355901718, + 0.0070458874106407166, + 0.01623694971203804, + 0.019196217879652977 + ], + "credit_map": [ + 0.007770895026624203, + 0.007420504465699196, + 0.02113095298409462, + 0.012009277008473873, + 0.006706547923386097, + 0.013046683743596077, + 0.010669496841728687, + 0.012132292613387108, + 0.006412319839000702, + 0.007268862333148718, + 0.010828424245119095, + 0.008132976479828358, + 0.004233416635543108, + 0.013016498647630215, + 0.005805612076073885, + 0.00587516650557518, + 0.0044995518401265144, + 0.0025983841624110937, + 0.013097490184009075, + 0.004622648470103741, + 0.0032809446565806866, + 0.015072545036673546, + 0.0022994454484432936, + 0.0026774846483021975, + 0.006106512621045113, + 0.004722085781395435, + 0.0037779652047902346, + 0.003310444066300988, + 0.0030480248387902975, + 0.003326786682009697, + 0.0032028283458203077, + 0.0038488840218633413, + 0.05062144622206688, + 0.056414537131786346, + 0.19642500579357147, + 0.02972559630870819, + 0.013998324982821941, + 0.06441842019557953, + 0.042076896876096725, + 0.15822568535804749, + 0.003591743530705571, + 0.006175106856971979, + 0.00646673608571291, + 0.003041138406842947, + 0.0036136324051767588, + 0.004767062608152628, + 0.007439496926963329, + 0.005632559768855572, + 0.014073067344725132, + 0.007409564685076475, + 0.012350940145552158, + 0.005524890031665564, + 0.006315693724900484, + 0.004386198706924915, + 0.011863362975418568, + 0.011011878959834576, + 0.0010590660385787487, + 0.0006161964265629649, + 0.0009711217717267573, + 0.0006103584310039878, + 0.0005465043941512704, + 0.0007699750130996108, + 0.0012474051909521222, + 0.0012252592714503407, + 0.0016847489168867469, + 0.0029584853909909725, + 0.001585545134730637, + 0.0014896694337949157, + 0.0015366984298452735, + 0.002502105664461851, + 0.0017326729139313102, + 0.002747023245319724, + 0.0018424985464662313, + 0.0017546103335916996, + 0.002488011959940195, + 0.0018097746651619673, + 0.0020880696829408407, + 0.0021012243814766407, + 0.0035458928905427456, + 0.003566135186702013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.011073444969952106 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_6/index_tensor([816, 816])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_6/index_tensor([816, 816])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8be954c6e9a7bb4be2d98233bbdfefd322fb989b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_6/index_tensor([816, 816])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, deliberately doll-like shot at eye level captures a female engineer in her 30s, her dark brown hair tied in a practical ponytail, standing in a clean, modern workshop. She wears a navy short-sleeved polo shirt and dark work pants, her expression focused. Her attention is divided between a sleek black tablet held steady in her left hand and a 3D printer humming on the workbench before her. The printer, a precision component constructed of brushed aluminum and matte black plastic, operates within a transparent acrylic casing. Inside, the extruder head moves with robotic precision, meticulously depositing thin, molten layers of vibrant electric blue filament onto a partially formed, complex geometric prototype. As she watches the object build layer by layer, her right thumb makes a decisive, tiny swipe across the tablet's glowing screen, adjusting parameters. Simultaneously, a stylized, translucent 3D wireframe animation glowing in the same electric blue hue overlays the physical object, visually deconstructing its complex internal structure and emphasizing the digital-to-physical creation process in real time. The scene is lit primarily by cool, bright overhead studio lights, which contrast softly with the warm, focused glow from the printer’s internal LEDs and the tablet screen, casting soft reflections on her face and creating a palpable atmosphere of focused innovation and cutting-edge creation.", + "seed": 636, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_6/index_816_seed_636_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.012883196072652935, + "vlm_logprob": -0.012883196072652935, + "vlm_yes_no_margin": 4.724999952316284, + "vlm_token_credit": -0.012883196072652935, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female engineer in her 30s with dark brown hair tied in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a navy short-sleeved polo shirt and dark work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a 3D printer with a transparent acrylic casing visible on the workbench, extruding electric blue filament?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stylized, translucent 3D wireframe animation glowing in electric blue overlaying the physical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool overhead studio lights contrasting with the warm glow from the printer's internal LEDs and tablet screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020395908504724503 + ], + "opposite_logprob_tokens": [ + -4.020395755767822 + ], + "logprob_reward": -0.020395908504724503, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.020395908504724503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5238991379737854, + 0.0893276259303093, + 0.05267021059989929, + 0.09262175112962723, + 0.06337906420230865, + 0.0382741279900074, + 0.032645661383867264, + 0.043288253247737885, + 0.0355420745909214, + 0.028352128341794014 + ], + "credit_map": [ + 0.09318435192108154, + 0.043399687856435776, + 0.047168172895908356, + 0.02685941755771637, + 0.046627484261989594, + 0.1311751753091812, + 0.06846443563699722, + 0.06702040135860443, + 0.011758765205740929, + 0.007120807189494371, + 0.009732488542795181, + 0.008007262833416462, + 0.008994163013994694, + 0.00869604293256998, + 0.017619797959923744, + 0.017398295924067497, + 0.007730306126177311, + 0.0032090304885059595, + 0.006598394364118576, + 0.003961001522839069, + 0.004791193176060915, + 0.008274617604911327, + 0.0073230331763625145, + 0.010782632045447826, + 0.011624738574028015, + 0.005984462331980467, + 0.00647871196269989, + 0.004951878916472197, + 0.008781281299889088, + 0.01092169713228941, + 0.022118525579571724, + 0.021760443225502968, + 0.00779744004830718, + 0.008182544261217117, + 0.00807698629796505, + 0.004961828235536814, + 0.004988228436559439, + 0.011958743445575237, + 0.0064375512301921844, + 0.01097574271261692, + 0.005350256804376841, + 0.0038487594574689865, + 0.006829298101365566, + 0.0027151007670909166, + 0.004033947829157114, + 0.006144552957266569, + 0.004785432014614344, + 0.004566777031868696, + 0.00409275246784091, + 0.0033360091038048267, + 0.0037133176811039448, + 0.002538548782467842, + 0.003000489203259349, + 0.0030503966845571995, + 0.005520978011190891, + 0.00739316875115037, + 0.005603228695690632, + 0.004774249158799648, + 0.005124243441969156, + 0.004855274688452482, + 0.004877683240920305, + 0.007087027654051781, + 0.0040090251713991165, + 0.00695751653984189, + 0.005136375781148672, + 0.003771894844248891, + 0.004183564800769091, + 0.004244384355843067, + 0.0035922639071941376, + 0.0038503846153616905, + 0.004963040817528963, + 0.005800164770334959, + 0.0038358387537300587, + 0.003476895857602358, + 0.004759449046105146, + 0.0025551742874085903, + 0.0025998905766755342, + 0.0030370617751032114, + 0.003124737413600087, + 0.004963076673448086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.020395908504724503 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female engineer in her 30s with dark brown hair tied in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a navy short-sleeved polo shirt and dark work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a 3D printer with a transparent acrylic casing visible on the workbench, extruding electric blue filament?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stylized, translucent 3D wireframe animation glowing in electric blue overlaying the physical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool overhead studio lights contrasting with the warm glow from the printer's internal LEDs and tablet screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003625251119956374 + ], + "opposite_logprob_tokens": [ + -5.628625392913818 + ], + "logprob_reward": -0.003625251119956374, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003625251119956374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4305914342403412, + 0.09183397889137268, + 0.0767672210931778, + 0.08049725741147995, + 0.06450972706079483, + 0.0769711285829544, + 0.024107426404953003, + 0.04377632588148117, + 0.058287132531404495, + 0.052658334374427795 + ], + "credit_map": [ + 0.08576565980911255, + 0.04275581240653992, + 0.04245579615235329, + 0.039534278213977814, + 0.039018414914608, + 0.08035927265882492, + 0.041373301297426224, + 0.05932891368865967, + 0.016416039317846298, + 0.014224880374968052, + 0.015886174514889717, + 0.010442644357681274, + 0.006376172881573439, + 0.008783749304711819, + 0.009314000606536865, + 0.0103903254494071, + 0.014199147932231426, + 0.012746589258313179, + 0.012353350408375263, + 0.010623179376125336, + 0.0057611726224422455, + 0.006225655786693096, + 0.006813577841967344, + 0.008044550195336342, + 0.010856689885258675, + 0.004313585348427296, + 0.006099253427237272, + 0.005107737611979246, + 0.0069718859158456326, + 0.013836078345775604, + 0.017446309328079224, + 0.015865718945860863, + 0.00754605233669281, + 0.007546712644398212, + 0.0074747707694768906, + 0.009894970804452896, + 0.005516838748008013, + 0.009475532919168472, + 0.007734858430922031, + 0.009320000186562538, + 0.009286588057875633, + 0.009261413477361202, + 0.013599403202533722, + 0.00986512191593647, + 0.005788419861346483, + 0.005251006688922644, + 0.009878283366560936, + 0.014040890149772167, + 0.003434872953221202, + 0.0020274044945836067, + 0.003939544782042503, + 0.0018383240094408393, + 0.0020418891217559576, + 0.0033644558861851692, + 0.003291400382295251, + 0.004169534891843796, + 0.0051429131999611855, + 0.005899353884160519, + 0.007167330011725426, + 0.007844693027436733, + 0.005045177880674601, + 0.0038967691361904144, + 0.0028956064488738775, + 0.005884483456611633, + 0.007807866670191288, + 0.006732241250574589, + 0.0075601753778755665, + 0.006915051490068436, + 0.005657679866999388, + 0.005921534728258848, + 0.005702055525034666, + 0.011990527622401714, + 0.007618822157382965, + 0.005602260585874319, + 0.006818457506597042, + 0.0045775449834764, + 0.004285357892513275, + 0.0077269067987799644, + 0.006404519081115723, + 0.00962446816265583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.003625251119956374 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female engineer in her 30s with dark brown hair tied in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a navy short-sleeved polo shirt and dark work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a 3D printer with a transparent acrylic casing visible on the workbench, extruding electric blue filament?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stylized, translucent 3D wireframe animation glowing in electric blue overlaying the physical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool overhead studio lights contrasting with the warm glow from the printer's internal LEDs and tablet screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046432530507445335 + ], + "opposite_logprob_tokens": [ + -5.379643440246582 + ], + "logprob_reward": -0.0046432530507445335, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046432530507445335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3976627588272095, + 0.091722272336483, + 0.06678812205791473, + 0.11194680631160736, + 0.16189922392368317, + 0.034049756824970245, + 0.026386503130197525, + 0.028228450566530228, + 0.03255684673786163, + 0.04875919595360756 + ], + "credit_map": [ + 0.060319382697343826, + 0.02496877871453762, + 0.034081365913152695, + 0.02591169998049736, + 0.0394282341003418, + 0.09530682116746902, + 0.060287248343229294, + 0.0573592446744442, + 0.013776086270809174, + 0.0076859816908836365, + 0.0071244011633098125, + 0.00895744003355503, + 0.008247350342571735, + 0.010603819042444229, + 0.017109636217355728, + 0.018217554315924644, + 0.009686078876256943, + 0.004155260045081377, + 0.005173790734261274, + 0.0041850474663078785, + 0.006694518029689789, + 0.011913478374481201, + 0.013084940612316132, + 0.011895005591213703, + 0.013202724047005177, + 0.006341037806123495, + 0.009381166659295559, + 0.00630063284188509, + 0.009382883086800575, + 0.011388803832232952, + 0.028016820549964905, + 0.02793274074792862, + 0.02322131209075451, + 0.010334385558962822, + 0.011822037398815155, + 0.010545101016759872, + 0.014540786854922771, + 0.03233291581273079, + 0.02446903847157955, + 0.03463364392518997, + 0.006299850530922413, + 0.0030570202507078648, + 0.0036040081176906824, + 0.002251529833301902, + 0.0030779519584029913, + 0.005994075443595648, + 0.0036271344870328903, + 0.00613818597048521, + 0.003808040404692292, + 0.0026025392580777407, + 0.0029283096082508564, + 0.0016972701996564865, + 0.0023582137655466795, + 0.003737234277650714, + 0.003933049272745848, + 0.00532184774056077, + 0.004120158962905407, + 0.0018142769113183022, + 0.002432806883007288, + 0.0019334476673975587, + 0.0024143224582076073, + 0.007545898202806711, + 0.0029760419856756926, + 0.004991499707102776, + 0.004545261617749929, + 0.0025783218443393707, + 0.0030902535654604435, + 0.0029788785614073277, + 0.0032492061145603657, + 0.0053430162370204926, + 0.0031897383742034435, + 0.007582169491797686, + 0.007102800998836756, + 0.00416713347658515, + 0.004724441561847925, + 0.0032020763028413057, + 0.00617991155013442, + 0.008446748368442059, + 0.006871404126286507, + 0.008064678870141506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.0046432530507445335 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female engineer in her 30s with dark brown hair tied in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a navy short-sleeved polo shirt and dark work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a 3D printer with a transparent acrylic casing visible on the workbench, extruding electric blue filament?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stylized, translucent 3D wireframe animation glowing in electric blue overlaying the physical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool overhead studio lights contrasting with the warm glow from the printer's internal LEDs and tablet screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005945378448814154 + ], + "opposite_logprob_tokens": [ + -5.130945205688477 + ], + "logprob_reward": -0.005945378448814154, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005945378448814154, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38462740182876587, + 0.15035955607891083, + 0.06164034083485603, + 0.09188970178365707, + 0.09158942103385925, + 0.05501420423388481, + 0.042060282081365585, + 0.05967250093817711, + 0.01970347948372364, + 0.04344315454363823 + ], + "credit_map": [ + 0.052347131073474884, + 0.03611862286925316, + 0.04720747470855713, + 0.02519945800304413, + 0.054482314735651016, + 0.054033517837524414, + 0.06110293045639992, + 0.05413592606782913, + 0.015882564708590508, + 0.010007300414144993, + 0.010905938223004341, + 0.013833886943757534, + 0.015363974496722221, + 0.013307323679327965, + 0.03305386006832123, + 0.038004692643880844, + 0.01159394159913063, + 0.004623141605407, + 0.006789331324398518, + 0.00327988900244236, + 0.008727632462978363, + 0.009099430404603481, + 0.008662504144012928, + 0.008864475414156914, + 0.013056251220405102, + 0.004986703861504793, + 0.0052576856687664986, + 0.005220848601311445, + 0.009810217656195164, + 0.014520751312375069, + 0.0183233842253685, + 0.02071385458111763, + 0.013405470177531242, + 0.00769750727340579, + 0.0077317385002970695, + 0.006406418047845364, + 0.009757609106600285, + 0.010225650854408741, + 0.01590942218899727, + 0.020455598831176758, + 0.008075101301074028, + 0.0040161097422242165, + 0.006370339542627335, + 0.0036243454087525606, + 0.00490502268075943, + 0.009111111983656883, + 0.007194776553660631, + 0.011717396788299084, + 0.0052295890636742115, + 0.003423630492761731, + 0.0040153940208256245, + 0.0023436378687620163, + 0.005193125456571579, + 0.006828625220805407, + 0.005529904272407293, + 0.009496377781033516, + 0.008075878024101257, + 0.004772762767970562, + 0.005693953949958086, + 0.004839404486119747, + 0.004471100866794586, + 0.012095498852431774, + 0.005567216780036688, + 0.014156688004732132, + 0.0032694616820663214, + 0.001763396430760622, + 0.002296739025041461, + 0.0016761055449023843, + 0.002233091974630952, + 0.0026536365039646626, + 0.00281434110365808, + 0.0029967068694531918, + 0.0065116602927446365, + 0.0031117117032408714, + 0.005806020461022854, + 0.0030782336834818125, + 0.0057897381484508514, + 0.00639372831210494, + 0.00669996440410614, + 0.006052099633961916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.005945378448814154 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female engineer in her 30s with dark brown hair tied in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a navy short-sleeved polo shirt and dark work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a 3D printer with a transparent acrylic casing visible on the workbench, extruding electric blue filament?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stylized, translucent 3D wireframe animation glowing in electric blue overlaying the physical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool overhead studio lights contrasting with the warm glow from the printer's internal LEDs and tablet screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029806189239025116 + ], + "opposite_logprob_tokens": [ + -3.529806137084961 + ], + "logprob_reward": -0.029806189239025116, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029806189239025116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33284085988998413, + 0.1554039865732193, + 0.08385113626718521, + 0.07945692539215088, + 0.04293394461274147, + 0.10655886679887772, + 0.03682831674814224, + 0.04678880423307419, + 0.052038587629795074, + 0.06329867243766785 + ], + "credit_map": [ + 0.0446709506213665, + 0.03247639536857605, + 0.029071001335978508, + 0.02647010236978531, + 0.03413647040724754, + 0.06838102638721466, + 0.05337852984666824, + 0.04425634443759918, + 0.016358831897377968, + 0.010773180983960629, + 0.011520215310156345, + 0.014015787281095982, + 0.015366032719612122, + 0.015926314517855644, + 0.033581286668777466, + 0.03786233440041542, + 0.010454287752509117, + 0.005936305969953537, + 0.006873714737594128, + 0.00634832913056016, + 0.008301145397126675, + 0.013873329386115074, + 0.013400102034211159, + 0.018663916736841202, + 0.011759656481444836, + 0.006141815800219774, + 0.005428264383226633, + 0.004541824571788311, + 0.008791493251919746, + 0.014045815914869308, + 0.009191184304654598, + 0.019556861370801926, + 0.006354116834700108, + 0.0047128344886004925, + 0.00421722698956728, + 0.0031980813946574926, + 0.005310032516717911, + 0.005821238271892071, + 0.007572328206151724, + 0.005748082417994738, + 0.014131929725408554, + 0.0065189627930521965, + 0.009482043795287609, + 0.007151008117944002, + 0.007594285998493433, + 0.011821219697594643, + 0.018340948969125748, + 0.03151845932006836, + 0.004930085968226194, + 0.004007076378911734, + 0.004396979231387377, + 0.002532442333176732, + 0.00427640974521637, + 0.006862656679004431, + 0.00407086405903101, + 0.005751798860728741, + 0.006077378522604704, + 0.005629673134535551, + 0.005264615174382925, + 0.004683034494519234, + 0.00283613009378314, + 0.008109652437269688, + 0.004826056305319071, + 0.009362258017063141, + 0.00876062735915184, + 0.004236043896526098, + 0.004535631742328405, + 0.004062369465827942, + 0.005567761138081551, + 0.0064018056727945805, + 0.00777176721021533, + 0.010702577419579029, + 0.009860317222774029, + 0.006811533588916063, + 0.005812872666865587, + 0.004270375706255436, + 0.008171159774065018, + 0.008643724955618382, + 0.00976816937327385, + 0.00996051449328661 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.029806189239025116 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, deliberately doll-like shot at eye level captures a female engineer in her 30s, her dark brown hair tied in a practical ponytail, standing in a clean, modern workshop. She wears a navy short-sleeved polo shirt and dark work pants, her expression focused. Her attention is divided between a sleek black tablet held steady in her left hand and a 3D printer humming on the workbench before her. The printer, a precision component constructed of brushed aluminum and matte black plastic, operates within a transparent acrylic casing. Inside, the extruder head moves with robotic precision, meticulously depositing thin, molten layers of vibrant electric blue filament onto a partially formed, complex geometric prototype. As she watches the object build layer by layer, her right thumb makes a decisive, tiny swipe across the tablet's glowing screen, adjusting parameters. Simultaneously, a stylized, translucent 3D wireframe animation glowing in the same electric blue hue overlays the physical object, visually deconstructing its complex internal structure and emphasizing the digital-to-physical creation process in real time. The scene is lit primarily by cool, bright overhead studio lights, which contrast softly with the warm, focused glow from the printer’s internal LEDs and the tablet screen, casting soft reflections on her face and creating a palpable atmosphere of focused innovation and cutting-edge creation.", + "seed": 637, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_6/index_816_seed_637_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.019985012151300906, + "vlm_logprob": -0.019985012151300906, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.019985012151300906, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female engineer in her 30s with dark brown hair tied in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a navy short-sleeved polo shirt and dark work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a 3D printer with a transparent acrylic casing visible on the workbench, extruding electric blue filament?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stylized, translucent 3D wireframe animation glowing in electric blue overlaying the physical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool overhead studio lights contrasting with the warm glow from the printer's internal LEDs and tablet screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05647055059671402 + ], + "opposite_logprob_tokens": [ + -2.9314706325531006 + ], + "logprob_reward": -0.05647055059671402, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05647055059671402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39431455731391907, + 0.10886938124895096, + 0.07136460393667221, + 0.11944118142127991, + 0.05013680458068848, + 0.04984789714217186, + 0.057431094348430634, + 0.019264889881014824, + 0.0883430615067482, + 0.040986523032188416 + ], + "credit_map": [ + 0.07337089627981186, + 0.05720328167080879, + 0.06578905880451202, + 0.05051819607615471, + 0.04032284393906593, + 0.05017080530524254, + 0.028961312025785446, + 0.027978166937828064, + 0.014318326488137245, + 0.01852080039680004, + 0.018515970557928085, + 0.017020178958773613, + 0.009210774675011635, + 0.014488931745290756, + 0.008389844559133053, + 0.008404560387134552, + 0.010450911708176136, + 0.01213523093611002, + 0.009558068588376045, + 0.009273996576666832, + 0.0067984904162585735, + 0.008943700231611729, + 0.006315796636044979, + 0.007888410240411758, + 0.013137619942426682, + 0.011066747829318047, + 0.017393754795193672, + 0.022126322612166405, + 0.01064155250787735, + 0.014517580159008503, + 0.012650422751903534, + 0.01790717802941799, + 0.0068707517348229885, + 0.011025913991034031, + 0.007098004221916199, + 0.007352801971137524, + 0.005218678154051304, + 0.004840705078095198, + 0.0044573103077709675, + 0.003272639587521553, + 0.00850466825067997, + 0.011038513854146004, + 0.005833270028233528, + 0.006653090938925743, + 0.004459820222109556, + 0.004200816620141268, + 0.0041736094281077385, + 0.004984105937182903, + 0.0070815361104905605, + 0.008983252570033073, + 0.008159367367625237, + 0.007978248409926891, + 0.005344129633158445, + 0.009154556319117546, + 0.005431364290416241, + 0.005298637319356203, + 0.0029612681828439236, + 0.003353095380589366, + 0.0030484595336019993, + 0.002773476066067815, + 0.0014558205148205161, + 0.0022425197530537844, + 0.0011934522772207856, + 0.002236798172816634, + 0.0071484739892184734, + 0.01062981877475977, + 0.010113531723618507, + 0.012808511033654213, + 0.009575325064361095, + 0.011299054138362408, + 0.009176707826554775, + 0.017591632902622223, + 0.006043833680450916, + 0.00696677016094327, + 0.005924417171627283, + 0.006556002423167229, + 0.004268087446689606, + 0.003965781070291996, + 0.004078743048012257, + 0.0031828870996832848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.05647055059671402 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female engineer in her 30s with dark brown hair tied in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a navy short-sleeved polo shirt and dark work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a 3D printer with a transparent acrylic casing visible on the workbench, extruding electric blue filament?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stylized, translucent 3D wireframe animation glowing in electric blue overlaying the physical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool overhead studio lights contrasting with the warm glow from the printer's internal LEDs and tablet screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052617397159338 + ], + "opposite_logprob_tokens": [ + -5.2552618980407715 + ], + "logprob_reward": -0.0052617397159338, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052617397159338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36609122157096863, + 0.11604730784893036, + 0.11110489815473557, + 0.0729813501238823, + 0.06247708573937416, + 0.04632051661610603, + 0.03800887241959572, + 0.03160422295331955, + 0.06441482901573181, + 0.0909496396780014 + ], + "credit_map": [ + 0.055236782878637314, + 0.06832961738109589, + 0.05522125959396362, + 0.047296032309532166, + 0.03967801108956337, + 0.042793527245521545, + 0.028722383081912994, + 0.028813639655709267, + 0.01525914017111063, + 0.022549716755747795, + 0.019610615447163582, + 0.021451463922858238, + 0.011263235472142696, + 0.009439615532755852, + 0.008722643367946148, + 0.007750887423753738, + 0.014063225127756596, + 0.019215097650885582, + 0.01617416739463806, + 0.014640654437243938, + 0.01036092359572649, + 0.013531479984521866, + 0.01055037323385477, + 0.012568972073495388, + 0.007348408456891775, + 0.009788844734430313, + 0.00890460703521967, + 0.015435637906193733, + 0.006397997494786978, + 0.008992708288133144, + 0.007250702939927578, + 0.008862446062266827, + 0.009936238639056683, + 0.012650308199226856, + 0.007303992751985788, + 0.009538998827338219, + 0.0062104701064527035, + 0.004433806985616684, + 0.004760848358273506, + 0.007642414886504412, + 0.0055878241546452045, + 0.009152748622000217, + 0.005293779540807009, + 0.006429353263229132, + 0.004524564370512962, + 0.0062815602868795395, + 0.005087667144834995, + 0.003963018301874399, + 0.005295213777571917, + 0.00808743853121996, + 0.005591911729425192, + 0.00557124987244606, + 0.0036928346380591393, + 0.0035784251522272825, + 0.003041790332645178, + 0.003150004195049405, + 0.005009881686419249, + 0.005164439789950848, + 0.005105230491608381, + 0.004978693090379238, + 0.0026527298614382744, + 0.0029690912924706936, + 0.002513503422960639, + 0.0032106542494148016, + 0.005658548790961504, + 0.006018306128680706, + 0.006510299630463123, + 0.01047710981220007, + 0.007444876711815596, + 0.011281115002930164, + 0.006994443479925394, + 0.010030128993093967, + 0.014588594436645508, + 0.02107030712068081, + 0.012225252576172352, + 0.014255504123866558, + 0.0071369982324540615, + 0.007332243025302887, + 0.006747291889041662, + 0.007593441754579544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0052617397159338 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female engineer in her 30s with dark brown hair tied in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a navy short-sleeved polo shirt and dark work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a 3D printer with a transparent acrylic casing visible on the workbench, extruding electric blue filament?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stylized, translucent 3D wireframe animation glowing in electric blue overlaying the physical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool overhead studio lights contrasting with the warm glow from the printer's internal LEDs and tablet screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008642531931400299 + ], + "opposite_logprob_tokens": [ + -4.758642673492432 + ], + "logprob_reward": -0.008642531931400299, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008642531931400299, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3980155289173126, + 0.12229068577289581, + 0.07896088808774948, + 0.05582122877240181, + 0.030108021572232246, + 0.0911622866988182, + 0.030150093138217926, + 0.03166238218545914, + 0.07442954182624817, + 0.08739927411079407 + ], + "credit_map": [ + 0.04269558563828468, + 0.036499347537755966, + 0.0437818318605423, + 0.08193259686231613, + 0.052050843834877014, + 0.049456220120191574, + 0.04146740213036537, + 0.050131745636463165, + 0.00949443131685257, + 0.010865668766200542, + 0.014865322969853878, + 0.025632472708821297, + 0.012071963399648666, + 0.02021142654120922, + 0.012711556628346443, + 0.01643786020576954, + 0.006355158984661102, + 0.008822825737297535, + 0.010744472034275532, + 0.009836952202022076, + 0.008610490709543228, + 0.012323918752372265, + 0.009642541408538818, + 0.012624532915651798, + 0.004984399303793907, + 0.0048944102600216866, + 0.00651493901386857, + 0.009542716667056084, + 0.00522034615278244, + 0.0074180662631988525, + 0.0058098407462239265, + 0.011436515487730503, + 0.0032590227201581, + 0.0048355963081121445, + 0.00315962010063231, + 0.004284235183149576, + 0.004386637359857559, + 0.004864347632974386, + 0.002705694641917944, + 0.002612870652228594, + 0.00764116458594799, + 0.00767529895529151, + 0.009050430729985237, + 0.016499565914273262, + 0.0101108243688941, + 0.0178613793104887, + 0.012009479105472565, + 0.010314146056771278, + 0.0025740826968103647, + 0.0037230090238153934, + 0.0049290903843939304, + 0.004697263706475496, + 0.003730410011485219, + 0.003915088716894388, + 0.0025065226946026087, + 0.0040746284648776054, + 0.0038763030897825956, + 0.0046099950559437275, + 0.005658084526658058, + 0.0053361509926617146, + 0.0021874711383134127, + 0.003008428728207946, + 0.002797634806483984, + 0.004188318736851215, + 0.006551295518875122, + 0.007657816633582115, + 0.00638565281406045, + 0.010751020163297653, + 0.009882716462016106, + 0.011639595963060856, + 0.007756724022328854, + 0.013804730027914047, + 0.007776262238621712, + 0.009764550253748894, + 0.00935498345643282, + 0.014382156543433666, + 0.009087697602808475, + 0.010621950961649418, + 0.010844580829143524, + 0.015567108057439327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.008642531931400299 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female engineer in her 30s with dark brown hair tied in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a navy short-sleeved polo shirt and dark work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a 3D printer with a transparent acrylic casing visible on the workbench, extruding electric blue filament?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stylized, translucent 3D wireframe animation glowing in electric blue overlaying the physical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool overhead studio lights contrasting with the warm glow from the printer's internal LEDs and tablet screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003191736526787281 + ], + "opposite_logprob_tokens": [ + -5.753191947937012 + ], + "logprob_reward": -0.003191736526787281, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003191736526787281, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4098740816116333, + 0.09309419244527817, + 0.05823875963687897, + 0.06004726141691208, + 0.048726558685302734, + 0.07751791924238205, + 0.04402737691998482, + 0.04879651963710785, + 0.057996924966573715, + 0.10168041288852692 + ], + "credit_map": [ + 0.05181790888309479, + 0.0550973117351532, + 0.04839526116847992, + 0.06299372017383575, + 0.042470354586839676, + 0.06355646997690201, + 0.03746557980775833, + 0.04807746410369873, + 0.0073081678710877895, + 0.010446470230817795, + 0.019419806078076363, + 0.014422939158976078, + 0.008334279991686344, + 0.01152802724391222, + 0.00853474996984005, + 0.013099754229187965, + 0.005310785956680775, + 0.00798999983817339, + 0.008488578721880913, + 0.010160723701119423, + 0.004816785454750061, + 0.005762206390500069, + 0.006628185510635376, + 0.009081498719751835, + 0.006333065684884787, + 0.005584170576184988, + 0.006376068107783794, + 0.008701971732079983, + 0.006207114085555077, + 0.010068453848361969, + 0.0061568571254611015, + 0.010619555599987507, + 0.004865893162786961, + 0.005244948901236057, + 0.006128258537501097, + 0.0066073560155928135, + 0.0060640377923846245, + 0.006710198242217302, + 0.004964564461261034, + 0.008141303434967995, + 0.006909152027219534, + 0.0070602670311927795, + 0.008731364272534847, + 0.011330629698932171, + 0.0069734943099319935, + 0.009183717891573906, + 0.008047708310186863, + 0.01928158849477768, + 0.003700424451380968, + 0.00526761170476675, + 0.004695572424679995, + 0.005862709134817123, + 0.005353295244276524, + 0.00682434206828475, + 0.005676475819200277, + 0.006646942812949419, + 0.0036194478161633015, + 0.003493044525384903, + 0.004393919371068478, + 0.007853806018829346, + 0.0059042684733867645, + 0.01145066600292921, + 0.004941189661622047, + 0.007140174508094788, + 0.005149975884705782, + 0.0050195991061627865, + 0.005163541994988918, + 0.008796203881502151, + 0.009130004793405533, + 0.009516282938420773, + 0.006567154545336962, + 0.008654164150357246, + 0.007873945869505405, + 0.010819637216627598, + 0.012155995704233646, + 0.014895614236593246, + 0.011130016297101974, + 0.012512859888374805, + 0.012380312196910381, + 0.019912036135792732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.003191736526787281 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female engineer in her 30s with dark brown hair tied in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a navy short-sleeved polo shirt and dark work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a 3D printer with a transparent acrylic casing visible on the workbench, extruding electric blue filament?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stylized, translucent 3D wireframe animation glowing in electric blue overlaying the physical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool overhead studio lights contrasting with the warm glow from the printer's internal LEDs and tablet screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026358501985669136 + ], + "opposite_logprob_tokens": [ + -3.6513586044311523 + ], + "logprob_reward": -0.026358501985669136, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026358501985669136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41848042607307434, + 0.08339398354291916, + 0.05490417033433914, + 0.0396086610853672, + 0.08433474600315094, + 0.04770580679178238, + 0.07546989619731903, + 0.06310693919658661, + 0.05687694251537323, + 0.07611840963363647 + ], + "credit_map": [ + 0.038914743810892105, + 0.03829646110534668, + 0.0562417097389698, + 0.0683979019522667, + 0.0472613200545311, + 0.07521867007017136, + 0.04603137820959091, + 0.04811827465891838, + 0.012427237816154957, + 0.016889438033103943, + 0.011746019124984741, + 0.012488690204918385, + 0.0065847174264490604, + 0.008369149640202522, + 0.006498834118247032, + 0.008389904163777828, + 0.00505019398406148, + 0.007961130701005459, + 0.008584808558225632, + 0.007563533261418343, + 0.0045597609132528305, + 0.0059278132393956184, + 0.005864082369953394, + 0.00939285010099411, + 0.0047978004440665245, + 0.006604902446269989, + 0.005583768244832754, + 0.007611694745719433, + 0.0029760580509901047, + 0.0039556329138576984, + 0.0035134158097207546, + 0.004565389361232519, + 0.006985860411077738, + 0.006564869545400143, + 0.010116984136402607, + 0.011988731101155281, + 0.009803812019526958, + 0.011704460717737675, + 0.011278285644948483, + 0.01589173637330532, + 0.00528289470821619, + 0.004206802695989609, + 0.007440511602908373, + 0.007432938553392887, + 0.0041112094186246395, + 0.006029825191944838, + 0.005388614255934954, + 0.007813013158738613, + 0.00851465668529272, + 0.007319179829210043, + 0.010707966051995754, + 0.012995781376957893, + 0.007043962366878986, + 0.012259327806532383, + 0.008068989962339401, + 0.00856003351509571, + 0.005885110702365637, + 0.007310543674975634, + 0.009151914156973362, + 0.01189010962843895, + 0.005469826515763998, + 0.009459187276661396, + 0.00579787464812398, + 0.008142373524606228, + 0.003962638322263956, + 0.006112227216362953, + 0.005403082352131605, + 0.009890131652355194, + 0.006235858425498009, + 0.008205949328839779, + 0.006566268391907215, + 0.010500784032046795, + 0.008739029057323933, + 0.007443651556968689, + 0.010763612575829029, + 0.011193379759788513, + 0.007954244501888752, + 0.008788920938968658, + 0.00853207241743803, + 0.012703501619398594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.026358501985669136 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, deliberately doll-like shot at eye level captures a female engineer in her 30s, her dark brown hair tied in a practical ponytail, standing in a clean, modern workshop. She wears a navy short-sleeved polo shirt and dark work pants, her expression focused. Her attention is divided between a sleek black tablet held steady in her left hand and a 3D printer humming on the workbench before her. The printer, a precision component constructed of brushed aluminum and matte black plastic, operates within a transparent acrylic casing. Inside, the extruder head moves with robotic precision, meticulously depositing thin, molten layers of vibrant electric blue filament onto a partially formed, complex geometric prototype. As she watches the object build layer by layer, her right thumb makes a decisive, tiny swipe across the tablet's glowing screen, adjusting parameters. Simultaneously, a stylized, translucent 3D wireframe animation glowing in the same electric blue hue overlays the physical object, visually deconstructing its complex internal structure and emphasizing the digital-to-physical creation process in real time. The scene is lit primarily by cool, bright overhead studio lights, which contrast softly with the warm, focused glow from the printer’s internal LEDs and the tablet screen, casting soft reflections on her face and creating a palpable atmosphere of focused innovation and cutting-edge creation.", + "seed": 638, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_6/index_816_seed_638_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.020496103912591934, + "vlm_logprob": -0.020496103912591934, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.020496103912591934, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female engineer in her 30s with dark brown hair tied in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a navy short-sleeved polo shirt and dark work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a 3D printer with a transparent acrylic casing visible on the workbench, extruding electric blue filament?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stylized, translucent 3D wireframe animation glowing in electric blue overlaying the physical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool overhead studio lights contrasting with the warm glow from the printer's internal LEDs and tablet screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06337269395589828 + ], + "opposite_logprob_tokens": [ + -2.8133726119995117 + ], + "logprob_reward": -0.06337269395589828, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06337269395589828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18401195108890533, + 0.056657496839761734, + 0.08040782809257507, + 0.06981013715267181, + 0.27731308341026306, + 0.05165106803178787, + 0.13914048671722412, + 0.034390319138765335, + 0.058950915932655334, + 0.0476667545735836 + ], + "credit_map": [ + 0.0295705646276474, + 0.05441901832818985, + 0.01817343197762966, + 0.021978070959448814, + 0.018318941816687584, + 0.012357719242572784, + 0.01218472234904766, + 0.017009470611810684, + 0.007221198175102472, + 0.009729196317493916, + 0.012539686635136604, + 0.007417607121169567, + 0.007204944733530283, + 0.004087380599230528, + 0.005050864536315203, + 0.0034066166263073683, + 0.011308715678751469, + 0.014296171255409718, + 0.019605882465839386, + 0.009377269074320793, + 0.008902733214199543, + 0.0065578892827034, + 0.004599335137754679, + 0.0057598259299993515, + 0.00613089744001627, + 0.005890212021768093, + 0.025246785953640938, + 0.004247857723385096, + 0.016327938064932823, + 0.0047388155944645405, + 0.004329517018049955, + 0.0028981133364140987, + 0.02175378054380417, + 0.010331372730433941, + 0.1207684725522995, + 0.058913704007864, + 0.03217291459441185, + 0.012432248331606388, + 0.01235556323081255, + 0.008584998548030853, + 0.008003701455891132, + 0.0070059653371572495, + 0.007614823989570141, + 0.007626407314091921, + 0.007365882396697998, + 0.004594412166625261, + 0.005103278439491987, + 0.00433659041300416, + 0.03773678094148636, + 0.03178510069847107, + 0.015568599104881287, + 0.010876612737774849, + 0.017026469111442566, + 0.00921358447521925, + 0.00875047966837883, + 0.008182845078408718, + 0.0039616357535123825, + 0.003100794507190585, + 0.006214780732989311, + 0.004451664164662361, + 0.005628087092190981, + 0.0025024968199431896, + 0.002751101739704609, + 0.005779758095741272, + 0.00766871077939868, + 0.004790940787643194, + 0.016795242205262184, + 0.0033523396123200655, + 0.004399022087454796, + 0.00513879582285881, + 0.002963294740766287, + 0.013842565007507801, + 0.006253581494092941, + 0.006232958287000656, + 0.01002756878733635, + 0.004191932268440723, + 0.0041475906036794186, + 0.004422456491738558, + 0.0029408892150968313, + 0.009449772536754608 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.06337269395589828 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female engineer in her 30s with dark brown hair tied in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a navy short-sleeved polo shirt and dark work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a 3D printer with a transparent acrylic casing visible on the workbench, extruding electric blue filament?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stylized, translucent 3D wireframe animation glowing in electric blue overlaying the physical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool overhead studio lights contrasting with the warm glow from the printer's internal LEDs and tablet screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036210936959832907 + ], + "opposite_logprob_tokens": [ + -5.6286211013793945 + ], + "logprob_reward": -0.0036210936959832907, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036210936959832907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23292836546897888, + 0.0956152081489563, + 0.14568284153938293, + 0.06551205366849899, + 0.09499362111091614, + 0.03184084966778755, + 0.10278008133172989, + 0.06933068484067917, + 0.062395308166742325, + 0.09892097115516663 + ], + "credit_map": [ + 0.03419531509280205, + 0.03803061693906784, + 0.025542916730046272, + 0.07656751573085785, + 0.016339406371116638, + 0.012856139801442623, + 0.015401272103190422, + 0.01399516686797142, + 0.012552167288959026, + 0.01216841209679842, + 0.016238441690802574, + 0.024272175505757332, + 0.012889115139842033, + 0.005920587573200464, + 0.005762504879385233, + 0.0058118002489209175, + 0.021513663232326508, + 0.018318932503461838, + 0.03763587400317192, + 0.021839486435055733, + 0.024074966087937355, + 0.008482500910758972, + 0.006682738196104765, + 0.007134684361517429, + 0.005446953698992729, + 0.004860410466790199, + 0.023209895938634872, + 0.004402982071042061, + 0.015910496935248375, + 0.004638828802853823, + 0.0033768031280487776, + 0.003665682626888156, + 0.0073922257870435715, + 0.006103545427322388, + 0.03069785051047802, + 0.03219446539878845, + 0.006239715497940779, + 0.0037554693408310413, + 0.004870634991675615, + 0.0037397139240056276, + 0.004062377382069826, + 0.003322875825688243, + 0.0032650900539010763, + 0.002792635001242161, + 0.0052739218808710575, + 0.003982719965279102, + 0.0035309249069541693, + 0.005610303953289986, + 0.020784806460142136, + 0.0352172814309597, + 0.009354651905596256, + 0.009376254864037037, + 0.01192709431052208, + 0.005579506047070026, + 0.005810143891721964, + 0.004730342887341976, + 0.00798453576862812, + 0.0048280986957252026, + 0.024254918098449707, + 0.003547106636688113, + 0.0057423654943704605, + 0.005431684665381908, + 0.002838235581293702, + 0.014703740365803242, + 0.009150641970336437, + 0.005870044231414795, + 0.013782051391899586, + 0.009225327521562576, + 0.004981940146535635, + 0.004846624564379454, + 0.0032010534778237343, + 0.011337620206177235, + 0.012630508281290531, + 0.007505315355956554, + 0.02319546602666378, + 0.0073123713955283165, + 0.007615400943905115, + 0.00852569192647934, + 0.006101383827626705, + 0.026034832000732422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.0036210936959832907 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female engineer in her 30s with dark brown hair tied in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a navy short-sleeved polo shirt and dark work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a 3D printer with a transparent acrylic casing visible on the workbench, extruding electric blue filament?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stylized, translucent 3D wireframe animation glowing in electric blue overlaying the physical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool overhead studio lights contrasting with the warm glow from the printer's internal LEDs and tablet screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003194113029167056 + ], + "opposite_logprob_tokens": [ + -5.753194332122803 + ], + "logprob_reward": -0.003194113029167056, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003194113029167056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2728375196456909, + 0.1393531858921051, + 0.06308267265558243, + 0.020130671560764313, + 0.1813843846321106, + 0.029830895364284515, + 0.09391411393880844, + 0.07181945443153381, + 0.043359581381082535, + 0.08428752422332764 + ], + "credit_map": [ + 0.03152862936258316, + 0.027919495478272438, + 0.04522208869457245, + 0.024454638361930847, + 0.04212096706032753, + 0.028414350003004074, + 0.035225000232458115, + 0.03795233741402626, + 0.014900236390531063, + 0.011263895779848099, + 0.01971874013543129, + 0.012114736251533031, + 0.02673596329987049, + 0.015007960610091686, + 0.020070267841219902, + 0.019541386514902115, + 0.005392969120293856, + 0.004392276983708143, + 0.01410934329032898, + 0.006472966633737087, + 0.017117025330662727, + 0.004626981448382139, + 0.005084504839032888, + 0.005886605009436607, + 0.0023993204813450575, + 0.003185684559866786, + 0.003295070258900523, + 0.0017745285294950008, + 0.0020693771075457335, + 0.002678380813449621, + 0.002185453660786152, + 0.002542856615036726, + 0.015360632911324501, + 0.00848589837551117, + 0.07714106142520905, + 0.024259507656097412, + 0.030849028378725052, + 0.009949143044650555, + 0.008124569430947304, + 0.0072145406156778336, + 0.00480780890211463, + 0.0034616410266608, + 0.0031641998793929815, + 0.004561235196888447, + 0.003239213488996029, + 0.00358153460547328, + 0.0036145506892353296, + 0.0034007099457085133, + 0.013523365370929241, + 0.025358818471431732, + 0.010409471578896046, + 0.008231759071350098, + 0.010836074128746986, + 0.006957359611988068, + 0.010484742932021618, + 0.008112521842122078, + 0.007641914766281843, + 0.009246952831745148, + 0.02361021749675274, + 0.010366738773882389, + 0.008914966136217117, + 0.003668735967949033, + 0.0039698779582977295, + 0.004400050733238459, + 0.004623593762516975, + 0.006269839592278004, + 0.0038393239956349134, + 0.005317577626556158, + 0.005933776963502169, + 0.00533548416569829, + 0.00457704346626997, + 0.0074629406444728374, + 0.007958252914249897, + 0.012820056639611721, + 0.017242006957530975, + 0.005358577240258455, + 0.011371377855539322, + 0.007409755140542984, + 0.009443647228181362, + 0.012683854438364506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.003194113029167056 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female engineer in her 30s with dark brown hair tied in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a navy short-sleeved polo shirt and dark work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a 3D printer with a transparent acrylic casing visible on the workbench, extruding electric blue filament?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stylized, translucent 3D wireframe animation glowing in electric blue overlaying the physical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool overhead studio lights contrasting with the warm glow from the printer's internal LEDs and tablet screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005945378448814154 + ], + "opposite_logprob_tokens": [ + -5.130945205688477 + ], + "logprob_reward": -0.005945378448814154, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005945378448814154, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20551376044750214, + 0.09154156595468521, + 0.055911432951688766, + 0.0932730883359909, + 0.2413143664598465, + 0.06638069450855255, + 0.10196226835250854, + 0.06571383029222488, + 0.025894122198224068, + 0.05249491333961487 + ], + "credit_map": [ + 0.022220876067876816, + 0.01827850006520748, + 0.027882667258381844, + 0.012905977666378021, + 0.03314719349145889, + 0.03080250695347786, + 0.027445243671536446, + 0.032830800861120224, + 0.008248948492109776, + 0.005675861611962318, + 0.024726949632167816, + 0.00693485839292407, + 0.015127675607800484, + 0.01021959912031889, + 0.00976499728858471, + 0.01084267906844616, + 0.004973461385816336, + 0.004839986562728882, + 0.006495312321931124, + 0.004180216230452061, + 0.008766747079789639, + 0.007556011434644461, + 0.008526570163667202, + 0.010573129169642925, + 0.00730561651289463, + 0.005870799999684095, + 0.02786402590572834, + 0.004336315207183361, + 0.024505315348505974, + 0.008298837579786777, + 0.008587121032178402, + 0.006505064200609922, + 0.019145183265209198, + 0.01163827907294035, + 0.10077842324972153, + 0.038765352219343185, + 0.03769823536276817, + 0.012651988305151463, + 0.011276270262897015, + 0.00936062354594469, + 0.006246752105653286, + 0.0044832853600382805, + 0.007456755265593529, + 0.004522829782217741, + 0.013957391493022442, + 0.00846132542937994, + 0.010781574994325638, + 0.010470783337950706, + 0.015853457152843475, + 0.021429577842354774, + 0.007808982860296965, + 0.006903656292706728, + 0.021004104986786842, + 0.00906210858374834, + 0.013221628963947296, + 0.006678745150566101, + 0.004860013723373413, + 0.0053373221307992935, + 0.0084384148940444, + 0.008255955763161182, + 0.011405923403799534, + 0.0051954155787825584, + 0.005147706717252731, + 0.017073076218366623, + 0.002339124446734786, + 0.003342093899846077, + 0.0020516214426606894, + 0.002928979927673936, + 0.004666441585868597, + 0.004176432266831398, + 0.003505389904603362, + 0.0028840394224971533, + 0.004626361653208733, + 0.00289541226811707, + 0.015386977232992649, + 0.0024505318142473698, + 0.005981976632028818, + 0.004574261140078306, + 0.0038858677726238966, + 0.012693526223301888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.005945378448814154 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female engineer in her 30s with dark brown hair tied in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a navy short-sleeved polo shirt and dark work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a 3D printer with a transparent acrylic casing visible on the workbench, extruding electric blue filament?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stylized, translucent 3D wireframe animation glowing in electric blue overlaying the physical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool overhead studio lights contrasting with the warm glow from the printer's internal LEDs and tablet screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026347240433096886 + ], + "opposite_logprob_tokens": [ + -3.6513471603393555 + ], + "logprob_reward": -0.026347240433096886, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026347240433096886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21765215694904327, + 0.07057331502437592, + 0.07432568073272705, + 0.07488803565502167, + 0.23760873079299927, + 0.03115701489150524, + 0.06275700777769089, + 0.11204598844051361, + 0.033067960292100906, + 0.08592408895492554 + ], + "credit_map": [ + 0.027097603306174278, + 0.024599596858024597, + 0.03403163328766823, + 0.023646052926778793, + 0.035976897925138474, + 0.023595888167619705, + 0.020754149183630943, + 0.0279503483325243, + 0.006710132118314505, + 0.005955190863460302, + 0.01705668307840824, + 0.006155665032565594, + 0.013850192539393902, + 0.006374649237841368, + 0.007516058161854744, + 0.00695475609973073, + 0.006592074874788523, + 0.006724887527525425, + 0.019782450050115585, + 0.006464507430791855, + 0.020461969077587128, + 0.005165599286556244, + 0.00477018766105175, + 0.004364004358649254, + 0.006103296298533678, + 0.0038251488003879786, + 0.025094790384173393, + 0.0032489399891346693, + 0.019622372463345528, + 0.007083794567734003, + 0.004978719167411327, + 0.004930985160171986, + 0.016309021040797234, + 0.012802639044821262, + 0.08416708558797836, + 0.07951869815587997, + 0.008946685120463371, + 0.010830282233655453, + 0.011884018778800964, + 0.013150307349860668, + 0.0034498069435358047, + 0.0025980535428971052, + 0.00429792795330286, + 0.0028680877294391394, + 0.0046008275821805, + 0.004785756580531597, + 0.004338269587606192, + 0.004218286834657192, + 0.009421567432582378, + 0.016225937753915787, + 0.007351959589868784, + 0.006947407033294439, + 0.008227648213505745, + 0.00517096696421504, + 0.0046607814729213715, + 0.004750747233629227, + 0.010852810926735401, + 0.005966157652437687, + 0.036302000284194946, + 0.006960676051676273, + 0.010326131246984005, + 0.009288646280765533, + 0.005196291953325272, + 0.027153274044394493, + 0.0037770031485706568, + 0.003525572130456567, + 0.004917806945741177, + 0.003193553304299712, + 0.005048262886703014, + 0.0032161453273147345, + 0.0034064408391714096, + 0.005983178503811359, + 0.007761518470942974, + 0.007014704402536154, + 0.017500413581728935, + 0.004418382421135902, + 0.01098864059895277, + 0.009083724580705166, + 0.008636970072984695, + 0.020519733428955078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.026347240433096886 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, deliberately doll-like shot at eye level captures a female engineer in her 30s, her dark brown hair tied in a practical ponytail, standing in a clean, modern workshop. She wears a navy short-sleeved polo shirt and dark work pants, her expression focused. Her attention is divided between a sleek black tablet held steady in her left hand and a 3D printer humming on the workbench before her. The printer, a precision component constructed of brushed aluminum and matte black plastic, operates within a transparent acrylic casing. Inside, the extruder head moves with robotic precision, meticulously depositing thin, molten layers of vibrant electric blue filament onto a partially formed, complex geometric prototype. As she watches the object build layer by layer, her right thumb makes a decisive, tiny swipe across the tablet's glowing screen, adjusting parameters. Simultaneously, a stylized, translucent 3D wireframe animation glowing in the same electric blue hue overlays the physical object, visually deconstructing its complex internal structure and emphasizing the digital-to-physical creation process in real time. The scene is lit primarily by cool, bright overhead studio lights, which contrast softly with the warm, focused glow from the printer’s internal LEDs and the tablet screen, casting soft reflections on her face and creating a palpable atmosphere of focused innovation and cutting-edge creation.", + "seed": 639, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_6/index_816_seed_639_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.016513488721102476, + "vlm_logprob": -0.016513488721102476, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.016513488721102476, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female engineer in her 30s with dark brown hair tied in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a navy short-sleeved polo shirt and dark work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a 3D printer with a transparent acrylic casing visible on the workbench, extruding electric blue filament?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stylized, translucent 3D wireframe animation glowing in electric blue overlaying the physical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool overhead studio lights contrasting with the warm glow from the printer's internal LEDs and tablet screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.050852127373218536 + ], + "opposite_logprob_tokens": [ + -3.050852060317993 + ], + "logprob_reward": -0.050852127373218536, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.050852127373218536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3535054922103882, + 0.07735549658536911, + 0.17350926995277405, + 0.08802185952663422, + 0.04479677230119705, + 0.039704013615846634, + 0.06690014898777008, + 0.04387442395091057, + 0.06688486784696579, + 0.04544771462678909 + ], + "credit_map": [ + 0.055125314742326736, + 0.07819274812936783, + 0.02670469507575035, + 0.06275743991136551, + 0.04374208301305771, + 0.028573796153068542, + 0.03309283033013344, + 0.025316568091511726, + 0.010774885304272175, + 0.012619283981621265, + 0.006291759200394154, + 0.01015359628945589, + 0.009356698021292686, + 0.012647961266338825, + 0.008440541103482246, + 0.00707075884565711, + 0.03588247299194336, + 0.03799884393811226, + 0.025201674550771713, + 0.02480194717645645, + 0.015233419835567474, + 0.013437006622552872, + 0.012867535464465618, + 0.00808636099100113, + 0.013659651391208172, + 0.00725406501442194, + 0.01030889991670847, + 0.009045894257724285, + 0.012787343934178352, + 0.011293296702206135, + 0.01127587165683508, + 0.012396827340126038, + 0.007157703395932913, + 0.004936168901622295, + 0.005830390844494104, + 0.005033578258007765, + 0.005581882316619158, + 0.005287324544042349, + 0.005017783027142286, + 0.005951939150691032, + 0.005166543181985617, + 0.00605645589530468, + 0.003807679284363985, + 0.005927907302975655, + 0.004768934566527605, + 0.004875536076724529, + 0.0043565197847783566, + 0.004744436126202345, + 0.0068340846337378025, + 0.011507516726851463, + 0.004387205466628075, + 0.009575321339070797, + 0.014200056903064251, + 0.008548680692911148, + 0.007715620566159487, + 0.004131658934056759, + 0.007437111344188452, + 0.011551751755177975, + 0.005230186507105827, + 0.007114844396710396, + 0.004094291944056749, + 0.0030184732750058174, + 0.003173123812302947, + 0.002254634164273739, + 0.006090610753744841, + 0.01003530528396368, + 0.010452318005263805, + 0.006923284847289324, + 0.010843810625374317, + 0.00712633179500699, + 0.009290006943047047, + 0.006123194471001625, + 0.005598684772849083, + 0.008933357894420624, + 0.007621967699378729, + 0.006845278199762106, + 0.0055182743817567825, + 0.004087912384420633, + 0.0035256980918347836, + 0.0033165381755679846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.050852127373218536 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female engineer in her 30s with dark brown hair tied in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a navy short-sleeved polo shirt and dark work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a 3D printer with a transparent acrylic casing visible on the workbench, extruding electric blue filament?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stylized, translucent 3D wireframe animation glowing in electric blue overlaying the physical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool overhead studio lights contrasting with the warm glow from the printer's internal LEDs and tablet screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00362489465624094 + ], + "opposite_logprob_tokens": [ + -5.62862491607666 + ], + "logprob_reward": -0.00362489465624094, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00362489465624094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1880207359790802, + 0.05517756566405296, + 0.19016338884830475, + 0.08670718967914581, + 0.06421280652284622, + 0.026438197121024132, + 0.053136181086301804, + 0.11386001110076904, + 0.1938745230436325, + 0.028409402817487717 + ], + "credit_map": [ + 0.023387964814901352, + 0.03538832440972328, + 0.016526656225323677, + 0.045511625707149506, + 0.02231895737349987, + 0.013270189985632896, + 0.018239067867398262, + 0.013377944938838482, + 0.009684408083558083, + 0.007759728468954563, + 0.005949368234723806, + 0.006310649216175079, + 0.006531973835080862, + 0.008733443915843964, + 0.005146552808582783, + 0.0050614443607628345, + 0.02978263795375824, + 0.021860074251890182, + 0.05562012642621994, + 0.03375215083360672, + 0.0156721081584692, + 0.012773144990205765, + 0.012747160159051418, + 0.007955994457006454, + 0.009068170562386513, + 0.010599151253700256, + 0.015995783731341362, + 0.013606616295874119, + 0.010311517864465714, + 0.0089879110455513, + 0.008694475516676903, + 0.009443563409149647, + 0.00815325416624546, + 0.008353336714208126, + 0.010979305021464825, + 0.015557333827018738, + 0.005916162859648466, + 0.005211610812693834, + 0.006433082278817892, + 0.0036087259650230408, + 0.0035341023467481136, + 0.003805212676525116, + 0.0028433669358491898, + 0.006257454864680767, + 0.0032855949830263853, + 0.002687731757760048, + 0.0023272372782230377, + 0.0016974974423646927, + 0.004640389699488878, + 0.012720691040158272, + 0.0026805740781128407, + 0.012469841167330742, + 0.0068694972433149815, + 0.0050406488589942455, + 0.004277843050658703, + 0.004436696879565716, + 0.017733246088027954, + 0.013729761354625225, + 0.030315162613987923, + 0.014472789131104946, + 0.010363972745835781, + 0.008641264401376247, + 0.00865984708070755, + 0.009943963959813118, + 0.017867768183350563, + 0.06243707239627838, + 0.050542399287223816, + 0.02464486099779606, + 0.013780530542135239, + 0.010209753178060055, + 0.008338368497788906, + 0.006053749937564135, + 0.0032591461203992367, + 0.0043504564091563225, + 0.004357099998742342, + 0.00372655363753438, + 0.004128222353756428, + 0.0031395317055284977, + 0.002785034943372011, + 0.002663358813151717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.00362489465624094 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female engineer in her 30s with dark brown hair tied in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a navy short-sleeved polo shirt and dark work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a 3D printer with a transparent acrylic casing visible on the workbench, extruding electric blue filament?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stylized, translucent 3D wireframe animation glowing in electric blue overlaying the physical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool overhead studio lights contrasting with the warm glow from the printer's internal LEDs and tablet screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005258300807327032 + ], + "opposite_logprob_tokens": [ + -5.255258083343506 + ], + "logprob_reward": -0.005258300807327032, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005258300807327032, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16537192463874817, + 0.03470388427376747, + 0.08116525411605835, + 0.07238119840621948, + 0.02297649160027504, + 0.028125161305069923, + 0.029821297153830528, + 0.05538167059421539, + 0.4850444197654724, + 0.02502877078950405 + ], + "credit_map": [ + 0.017075523734092712, + 0.03802139684557915, + 0.022802961990237236, + 0.021678069606423378, + 0.022692037746310234, + 0.014302344992756844, + 0.015765774995088577, + 0.013033812865614891, + 0.0036747243721038103, + 0.006199977360665798, + 0.0027825471479445696, + 0.004603327251970768, + 0.0049117375165224075, + 0.00517298374325037, + 0.003595295362174511, + 0.0037632870953530073, + 0.0076137688010931015, + 0.008383463136851788, + 0.008525296114385128, + 0.012326735071837902, + 0.01966412551701069, + 0.010259773582220078, + 0.00877804309129715, + 0.005614041350781918, + 0.00610607722774148, + 0.00628244411200285, + 0.004972349386662245, + 0.006681125145405531, + 0.012068316340446472, + 0.011864449828863144, + 0.010756329633295536, + 0.013650094158947468, + 0.001967907650396228, + 0.0019931765273213387, + 0.0018069486832246184, + 0.002977753756567836, + 0.0032701899763196707, + 0.003551200497895479, + 0.003775084624066949, + 0.0036342297680675983, + 0.0021831723861396313, + 0.003721784334629774, + 0.0016917340690270066, + 0.003724223468452692, + 0.004012290854007006, + 0.005174453370273113, + 0.004038625862449408, + 0.00357887358404696, + 0.0026401663199067116, + 0.0047652763314545155, + 0.0017714306013658643, + 0.00481127854436636, + 0.005385049153119326, + 0.00344209186732769, + 0.0027974783442914486, + 0.0042085228487849236, + 0.00569799542427063, + 0.007997674867510796, + 0.004479691386222839, + 0.008662105537950993, + 0.010012000799179077, + 0.007314031943678856, + 0.005472721066325903, + 0.0057454463094472885, + 0.045644160360097885, + 0.168000265955925, + 0.13009333610534668, + 0.06304360181093216, + 0.02439854107797146, + 0.02454567700624466, + 0.018918368965387344, + 0.010400427505373955, + 0.002211214043200016, + 0.005134709179401398, + 0.006246073637157679, + 0.003623007796704769, + 0.002071741968393326, + 0.002233802806586027, + 0.0016824391204863787, + 0.0018257815390825272 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.005258300807327032 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female engineer in her 30s with dark brown hair tied in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a navy short-sleeved polo shirt and dark work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a 3D printer with a transparent acrylic casing visible on the workbench, extruding electric blue filament?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stylized, translucent 3D wireframe animation glowing in electric blue overlaying the physical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool overhead studio lights contrasting with the warm glow from the printer's internal LEDs and tablet screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004637676291167736 + ], + "opposite_logprob_tokens": [ + -5.379637718200684 + ], + "logprob_reward": -0.004637676291167736, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004637676291167736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09696271270513535, + 0.043531790375709534, + 0.04866021126508713, + 0.056040603667497635, + 0.02054266631603241, + 0.025982610881328583, + 0.01969548873603344, + 0.034430261701345444, + 0.6299960613250732, + 0.024157557636499405 + ], + "credit_map": [ + 0.013743641786277294, + 0.016243964433670044, + 0.010377632454037666, + 0.023156343027949333, + 0.011065546423196793, + 0.007937475107610226, + 0.008143438026309013, + 0.006294670049101114, + 0.005736622493714094, + 0.006618674844503403, + 0.004712364636361599, + 0.005490358453243971, + 0.005546980537474155, + 0.0058920979499816895, + 0.005223388317972422, + 0.004311301279813051, + 0.004527070093899965, + 0.009168049320578575, + 0.004930531606078148, + 0.01068209856748581, + 0.007850362919270992, + 0.004580265376716852, + 0.004060207400470972, + 0.0028616287745535374, + 0.0050757871940732, + 0.00782310776412487, + 0.00467002484947443, + 0.006410242989659309, + 0.009003832936286926, + 0.007770476397126913, + 0.006286944262683392, + 0.009000184945762157, + 0.002051146235316992, + 0.0018937530694529414, + 0.002001577988266945, + 0.003237286116927862, + 0.003423356683924794, + 0.002860703505575657, + 0.0024750572629272938, + 0.0025997855700552464, + 0.0021547190845012665, + 0.003266415325924754, + 0.0016892594285309315, + 0.003154987236484885, + 0.0037255280185490847, + 0.005653987172991037, + 0.002946831751614809, + 0.0033908828627318144, + 0.0022380887530744076, + 0.002480789553374052, + 0.0016772921662777662, + 0.003248205641284585, + 0.0033846155274659395, + 0.002728178631514311, + 0.001915005617775023, + 0.0020233129616826773, + 0.0029542390257120132, + 0.007949325256049633, + 0.0033142329193651676, + 0.006188369356095791, + 0.003215908072888851, + 0.0037286956794559956, + 0.003108938690274954, + 0.0039705513045191765, + 0.059277597814798355, + 0.21545149385929108, + 0.1681481897830963, + 0.08204495906829834, + 0.03376798331737518, + 0.03188871592283249, + 0.02491559274494648, + 0.014501512050628662, + 0.002837135922163725, + 0.003715100232511759, + 0.003966222051531076, + 0.0033943962771445513, + 0.0031572282314300537, + 0.002894449047744274, + 0.0023476197384297848, + 0.001845405320636928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.004637676291167736 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female engineer in her 30s with dark brown hair tied in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a navy short-sleeved polo shirt and dark work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a 3D printer with a transparent acrylic casing visible on the workbench, extruding electric blue filament?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stylized, translucent 3D wireframe animation glowing in electric blue overlaying the physical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool overhead studio lights contrasting with the warm glow from the printer's internal LEDs and tablet screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018194444477558136 + ], + "opposite_logprob_tokens": [ + -4.018194675445557 + ], + "logprob_reward": -0.018194444477558136, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018194444477558136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10982537269592285, + 0.04271203279495239, + 0.07436740398406982, + 0.0942956730723381, + 0.022682370617985725, + 0.022323060780763626, + 0.037733014672994614, + 0.11534997820854187, + 0.46168211102485657, + 0.019028980284929276 + ], + "credit_map": [ + 0.01959117129445076, + 0.023165885359048843, + 0.009671354666352272, + 0.016653023660182953, + 0.012406171299517155, + 0.010285434313118458, + 0.010141631588339806, + 0.007910692133009434, + 0.006026045884937048, + 0.007693686988204718, + 0.0051610032096505165, + 0.004578925669193268, + 0.004558355547487736, + 0.006338680163025856, + 0.0042793480679392815, + 0.004075984936207533, + 0.005884110927581787, + 0.01679070107638836, + 0.005942557938396931, + 0.021743396297097206, + 0.008951838128268719, + 0.004582044668495655, + 0.005990910343825817, + 0.004481848794966936, + 0.008180060423910618, + 0.00881048385053873, + 0.007363243028521538, + 0.01028031948953867, + 0.016938891261816025, + 0.012514523230493069, + 0.014548433013260365, + 0.015659712255001068, + 0.0029472080059349537, + 0.002582450630143285, + 0.002240173751488328, + 0.0036209099926054478, + 0.002859033877030015, + 0.0027509164065122604, + 0.002527434378862381, + 0.0031542438082396984, + 0.0017971661873161793, + 0.0028570150025188923, + 0.0014780666679143906, + 0.0025875025894492865, + 0.0032743047922849655, + 0.0037391213700175285, + 0.002639477839693427, + 0.003950405865907669, + 0.003858483163639903, + 0.0037744936998933554, + 0.002755764639005065, + 0.008273721672594547, + 0.00485495338216424, + 0.005719735287129879, + 0.004270509816706181, + 0.004225351382046938, + 0.012722131796181202, + 0.0169575996696949, + 0.011504877358675003, + 0.019691409543156624, + 0.014964810572564602, + 0.013929628767073154, + 0.012372659519314766, + 0.013206865638494492, + 0.043422944843769073, + 0.15970420837402344, + 0.12647010385990143, + 0.06018822267651558, + 0.020822061225771904, + 0.02284599095582962, + 0.018067779019474983, + 0.010160769335925579, + 0.0015692648012191057, + 0.0024032597430050373, + 0.0029371092095971107, + 0.0021175688598304987, + 0.0031888813246041536, + 0.002359263366088271, + 0.002281441120430827, + 0.002172190696001053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.018194444477558136 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_7/index_tensor([9380, 9380])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_7/index_tensor([9380, 9380])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..85d8b1e04b5b4df55c7930b0b6bc80ec4ec49050 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_7/index_tensor([9380, 9380])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "long, high-angle drone shot, slowly and steadily advancing, follows a quaint, handcrafted wooden sampan as it leisurely circles a mysterious island shrouded in thick morning mist. The boat's hull is constructed from deep brown aged timber, its grain visible. Its sides, covered in moss and water stains from years of wear and tear, appear weathered. As the camera zooms in, the water's surface takes center stage. The early morning sun, just breaking through the clouds, casts a golden glow across the deep, calm lake, creating a shimmering, flickering streak of light that complements the gentle ripples created by the boat's cut. The island itself is lush and verdant, but its contours and peaks are largely hidden within a shifting, dreamy white mist, revealing only glimpses of verdant green. The scene's palette, dominated by the dusky brown of the boat, the dark green of the lake, and the warm golden hues of the sunlight, creates a mysterious atmosphere of profound tranquility, poetry, and a sense of discovery.", + "seed": 19653, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_7/index_9380_seed_19653_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08089084941893816, + "vlm_logprob": -0.08089084941893816, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.08089084941893816, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handcrafted wooden sampan (boat)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep brown aged timber with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat's side covered in moss and water stains appearing weathered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle drone shot that slowly advances and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a calm lake with a mysterious island shrouded in thick morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028607888147234917 + ], + "opposite_logprob_tokens": [ + -4.1536078453063965 + ], + "logprob_reward": -0.028607888147234917, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.028607888147234917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14508524537086487, + 0.12187979370355606, + 0.0686069056391716, + 0.051777590066194534, + 0.28969550132751465, + 0.10767364501953125, + 0.020724862813949585, + 0.047499991953372955, + 0.04149818420410156, + 0.10555830597877502 + ], + "credit_map": [ + 0.02451876550912857, + 0.02040313184261322, + 0.02637648396193981, + 0.02449869178235531, + 0.017416955903172493, + 0.011465713381767273, + 0.009166031144559383, + 0.011239469982683659, + 0.013857070356607437, + 0.010098346509039402, + 0.032773666083812714, + 0.03874697908759117, + 0.0069573563523590565, + 0.00692331837490201, + 0.006394932512193918, + 0.006128119770437479, + 0.011741175316274166, + 0.009749024175107479, + 0.007596940267831087, + 0.010973276570439339, + 0.007455180399119854, + 0.003891010768711567, + 0.00920585822314024, + 0.007994435727596283, + 0.005161661189049482, + 0.005524090491235256, + 0.00555451912805438, + 0.013597920536994934, + 0.005642835982143879, + 0.0029789202380925417, + 0.009708459489047527, + 0.003609187202528119, + 0.013671360909938812, + 0.020202675834298134, + 0.02644176036119461, + 0.11424917727708817, + 0.010960971936583519, + 0.008115522563457489, + 0.052259039133787155, + 0.04379497468471527, + 0.016685334965586662, + 0.025089412927627563, + 0.0071271006017923355, + 0.020536506548523903, + 0.006188465282320976, + 0.0030592738185077906, + 0.021525003015995026, + 0.007462545298039913, + 0.00319882039912045, + 0.00509822741150856, + 0.0020720793399959803, + 0.0025538322515785694, + 0.0018409918993711472, + 0.0012277981732040644, + 0.0027250556740909815, + 0.0020080574322491884, + 0.005938495509326458, + 0.009192620404064655, + 0.005550234112888575, + 0.0067869918420910835, + 0.00478044617921114, + 0.0038162486162036657, + 0.007734119892120361, + 0.0037008377257734537, + 0.004717655014246702, + 0.007037491537630558, + 0.005293305031955242, + 0.005689795594662428, + 0.004217468202114105, + 0.00373309338465333, + 0.006747937295585871, + 0.0040614367462694645, + 0.012325834482908249, + 0.025208842009305954, + 0.013666070066392422, + 0.011447339318692684, + 0.013041624799370766, + 0.009751442819833755, + 0.01122971810400486, + 0.00888743158429861 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.028607888147234917 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handcrafted wooden sampan (boat)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep brown aged timber with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat's side covered in moss and water stains appearing weathered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle drone shot that slowly advances and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a calm lake with a mysterious island shrouded in thick morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018239863216876984 + ], + "opposite_logprob_tokens": [ + -4.018239974975586 + ], + "logprob_reward": -0.018239863216876984, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018239863216876984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040664296597242355, + 0.1475227028131485, + 0.03670215606689453, + 0.03064507432281971, + 0.28640982508659363, + 0.27385175228118896, + 0.015330685302615166, + 0.06362804770469666, + 0.03751935437321663, + 0.06772615760564804 + ], + "credit_map": [ + 0.005820399150252342, + 0.003903575474396348, + 0.006003630813211203, + 0.007811891846358776, + 0.0035010536666959524, + 0.0029109639581292868, + 0.005132004152983427, + 0.005580779165029526, + 0.01567084714770317, + 0.0081283925101161, + 0.044620249420404434, + 0.04507935047149658, + 0.01070295087993145, + 0.007004104554653168, + 0.008452493697404861, + 0.007864315994083881, + 0.005358730908483267, + 0.005054261535406113, + 0.005273872055113316, + 0.006231103092432022, + 0.006124828942120075, + 0.0021211826242506504, + 0.003467014990746975, + 0.003071160987019539, + 0.003059608396142721, + 0.0033379080705344677, + 0.0032915438059717417, + 0.006574408616870642, + 0.0038271790836006403, + 0.00177347706630826, + 0.004758833907544613, + 0.004022113513201475, + 0.013174344785511494, + 0.022949280217289925, + 0.024801114574074745, + 0.10918378084897995, + 0.012632993049919605, + 0.008591491729021072, + 0.05287047475576401, + 0.04220633581280708, + 0.029577700421214104, + 0.04913496971130371, + 0.019033391028642654, + 0.07493481785058975, + 0.011735277250409126, + 0.006785064935684204, + 0.07157790660858154, + 0.011072608642280102, + 0.0018463589949533343, + 0.003387963166460395, + 0.0017644745530560613, + 0.002199193462729454, + 0.0012280623195692897, + 0.0009219818166457117, + 0.0022597014904022217, + 0.0017229500226676464, + 0.009289526380598545, + 0.010532403364777565, + 0.006180930882692337, + 0.008631527423858643, + 0.0048804315738379955, + 0.0028694572392851114, + 0.017205482348799706, + 0.0040382929146289825, + 0.005422743037343025, + 0.007747225929051638, + 0.004771954379975796, + 0.0036612250842154026, + 0.004610504023730755, + 0.0030831575859338045, + 0.004912864416837692, + 0.0033096815459430218, + 0.007247147615998983, + 0.01980993151664734, + 0.00906996987760067, + 0.005381268914788961, + 0.008047939278185368, + 0.005338954273611307, + 0.006066998932510614, + 0.006763951387256384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 1 + ], + "selected_reward": -0.018239863216876984 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handcrafted wooden sampan (boat)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep brown aged timber with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat's side covered in moss and water stains appearing weathered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle drone shot that slowly advances and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a calm lake with a mysterious island shrouded in thick morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014215397648513317 + ], + "opposite_logprob_tokens": [ + -4.264215469360352 + ], + "logprob_reward": -0.014215397648513317, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014215397648513317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07756214588880539, + 0.1021767184138298, + 0.05236531421542168, + 0.025211406871676445, + 0.36459898948669434, + 0.17164357006549835, + 0.031002389267086983, + 0.05133816599845886, + 0.03672698140144348, + 0.08737428486347198 + ], + "credit_map": [ + 0.009710204787552357, + 0.0067125242203474045, + 0.014676489867269993, + 0.013743915595114231, + 0.00607784790918231, + 0.0046835048124194145, + 0.0069814217276871204, + 0.014976241625845432, + 0.011174514889717102, + 0.006549500860273838, + 0.027142085134983063, + 0.026907529681921005, + 0.0080935163423419, + 0.007959977723658085, + 0.006847592536360025, + 0.007502018474042416, + 0.007227674592286348, + 0.00660034641623497, + 0.007038375828415155, + 0.009635336697101593, + 0.007959969341754913, + 0.0026865259278565645, + 0.005440365988761187, + 0.0057767233811318874, + 0.002712070709094405, + 0.0028725641313940287, + 0.002648369176313281, + 0.0042341467924416065, + 0.0037848784122616053, + 0.0017929356545209885, + 0.003925548400729895, + 0.003240896388888359, + 0.016801098361611366, + 0.028213052079081535, + 0.03417261317372322, + 0.14161953330039978, + 0.014165619388222694, + 0.01011332031339407, + 0.06519056111574173, + 0.0543232187628746, + 0.018354985862970352, + 0.02734525501728058, + 0.010201897472143173, + 0.04786435887217522, + 0.006078146863728762, + 0.003983173985034227, + 0.05011525750160217, + 0.007700521964579821, + 0.003489091992378235, + 0.0061510903760790825, + 0.0042183115147054195, + 0.005239240825176239, + 0.0027368110604584217, + 0.0017471702303737402, + 0.004755223635584116, + 0.0026654545217752457, + 0.006331845186650753, + 0.009829810820519924, + 0.00589324627071619, + 0.007649602368474007, + 0.004183802288025618, + 0.0034041423350572586, + 0.01023394986987114, + 0.0038117696531116962, + 0.004238293506205082, + 0.006661310791969299, + 0.005536112003028393, + 0.0050343736074864864, + 0.0034309097100049257, + 0.002451728330925107, + 0.006054426543414593, + 0.0033198287710547447, + 0.010371319949626923, + 0.026380568742752075, + 0.009486946277320385, + 0.007880926132202148, + 0.010920294560492039, + 0.004658161196857691, + 0.008658979088068008, + 0.009017095901072025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 1 + ], + "selected_reward": -0.014215397648513317 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handcrafted wooden sampan (boat)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep brown aged timber with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat's side covered in moss and water stains appearing weathered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle drone shot that slowly advances and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a calm lake with a mysterious island shrouded in thick morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28132757544517517 + ], + "opposite_logprob_tokens": [ + -1.4063276052474976 + ], + "logprob_reward": -0.28132757544517517, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28132757544517517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12071507424116135, + 0.1996844857931137, + 0.03705720230937004, + 0.01911969855427742, + 0.24235716462135315, + 0.17766457796096802, + 0.01993963122367859, + 0.1197051852941513, + 0.01560214627534151, + 0.048154789954423904 + ], + "credit_map": [ + 0.020504282787442207, + 0.016604239121079445, + 0.02104194462299347, + 0.01923041231930256, + 0.014298105612397194, + 0.00893519725650549, + 0.00842344481498003, + 0.01167745515704155, + 0.02073809877038002, + 0.014661876484751701, + 0.06077319383621216, + 0.06524882465600967, + 0.009075772017240524, + 0.007971256971359253, + 0.008799516595900059, + 0.012415952049195766, + 0.006745611317455769, + 0.00558999739587307, + 0.0037048959638923407, + 0.005955717526376247, + 0.005386447999626398, + 0.002090862952172756, + 0.00462903268635273, + 0.002954634139314294, + 0.002591297961771488, + 0.0024910669308155775, + 0.0022292197681963444, + 0.0036904572043567896, + 0.0022574227768927813, + 0.001225789193995297, + 0.002743396209552884, + 0.0018910500220954418, + 0.009303764440119267, + 0.011758138425648212, + 0.02433580532670021, + 0.10602812469005585, + 0.008215918205678463, + 0.006148156244307756, + 0.051915235817432404, + 0.024652015417814255, + 0.036974452435970306, + 0.06366186589002609, + 0.01223459281027317, + 0.01737259142100811, + 0.012384368106722832, + 0.005207172129303217, + 0.017455564811825752, + 0.012373965233564377, + 0.003358029294759035, + 0.0026762965135276318, + 0.00278547708876431, + 0.0026719397865235806, + 0.0016344566829502583, + 0.0010973279131576419, + 0.00410919776186347, + 0.001606905134394765, + 0.024316200986504555, + 0.010007137432694435, + 0.006649872753769159, + 0.015539243817329407, + 0.007822617888450623, + 0.004670656751841307, + 0.043470289558172226, + 0.007229152601212263, + 0.002118340227752924, + 0.00361964525654912, + 0.0023399621713906527, + 0.0013245107838883996, + 0.0021678260527551174, + 0.0009642041404731572, + 0.0015437847469002008, + 0.0015238720225170255, + 0.0058732102625072, + 0.01328691840171814, + 0.005673238541930914, + 0.006342113483697176, + 0.00558910146355629, + 0.003895207541063428, + 0.00407585920765996, + 0.0034191401209682226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 5 + ], + "selected_reward": -0.28132757544517517 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handcrafted wooden sampan (boat)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep brown aged timber with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat's side covered in moss and water stains appearing weathered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle drone shot that slowly advances and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a calm lake with a mysterious island shrouded in thick morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206352263689041 + ], + "opposite_logprob_tokens": [ + -2.812063455581665 + ], + "logprob_reward": -0.06206352263689041, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206352263689041, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06091088801622391, + 0.07037960737943649, + 0.09662503004074097, + 0.014204056933522224, + 0.232484370470047, + 0.16331139206886292, + 0.01803443394601345, + 0.1918029636144638, + 0.023664923384785652, + 0.12858225405216217 + ], + "credit_map": [ + 0.009022237733006477, + 0.007758158724755049, + 0.010328306816518307, + 0.009786824695765972, + 0.00738830491900444, + 0.0048045567236840725, + 0.004105035215616226, + 0.007717473898082972, + 0.009680571965873241, + 0.007288875989615917, + 0.01682054065167904, + 0.019072219729423523, + 0.004309259355068207, + 0.00442527886480093, + 0.0035637172404676676, + 0.005219149403274059, + 0.018641358241438866, + 0.014122279360890388, + 0.010179207660257816, + 0.016164613887667656, + 0.012471995316445827, + 0.004646500572562218, + 0.010523159056901932, + 0.00987592525780201, + 0.001553499256260693, + 0.0015899762511253357, + 0.0012644573580473661, + 0.0022223698906600475, + 0.0012592831626534462, + 0.0010587871074676514, + 0.003004671074450016, + 0.002251015044748783, + 0.009679924696683884, + 0.011164267547428608, + 0.021892592310905457, + 0.09763646125793457, + 0.008472474291920662, + 0.0063505047000944614, + 0.047390926629304886, + 0.029897252097725868, + 0.02151259407401085, + 0.03502005711197853, + 0.009679931215941906, + 0.03977141156792641, + 0.006954681593924761, + 0.0036538238637149334, + 0.039682451635599136, + 0.007036454044282436, + 0.0038456586189568043, + 0.0029317475855350494, + 0.0020309167448431253, + 0.001772083225660026, + 0.001452228520065546, + 0.0009255463955923915, + 0.003778408979997039, + 0.0012978456215932965, + 0.03306451439857483, + 0.019828232005238533, + 0.014401928521692753, + 0.025930961593985558, + 0.01220962405204773, + 0.006450981367379427, + 0.06864520162343979, + 0.011271525174379349, + 0.0024447315372526646, + 0.005809478461742401, + 0.0031367356423288584, + 0.0027994338888674974, + 0.002060927450656891, + 0.0013595782220363617, + 0.003567826934158802, + 0.0024862142745405436, + 0.015369577333331108, + 0.036928389221429825, + 0.015248519368469715, + 0.012176156975328922, + 0.014983303844928741, + 0.010687238536775112, + 0.011806252412497997, + 0.011382830329239368 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 5 + ], + "selected_reward": -0.06206352263689041 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, high-angle drone shot, slowly and steadily advancing, follows a quaint, handcrafted wooden sampan as it leisurely circles a mysterious island shrouded in thick morning mist. The boat's hull is constructed from deep brown aged timber, its grain visible. Its sides, covered in moss and water stains from years of wear and tear, appear weathered. As the camera zooms in, the water's surface takes center stage. The early morning sun, just breaking through the clouds, casts a golden glow across the deep, calm lake, creating a shimmering, flickering streak of light that complements the gentle ripples created by the boat's cut. The island itself is lush and verdant, but its contours and peaks are largely hidden within a shifting, dreamy white mist, revealing only glimpses of verdant green. The scene's palette, dominated by the dusky brown of the boat, the dark green of the lake, and the warm golden hues of the sunlight, creates a mysterious atmosphere of profound tranquility, poetry, and a sense of discovery.", + "seed": 19654, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_7/index_9380_seed_19654_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.034876182489097116, + "vlm_logprob": -0.034876182489097116, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.034876182489097116, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handcrafted wooden sampan (boat)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep brown aged timber with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat's side covered in moss and water stains appearing weathered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle drone shot that slowly advances and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a calm lake with a mysterious island shrouded in thick morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05619189515709877 + ], + "opposite_logprob_tokens": [ + -3.056191921234131 + ], + "logprob_reward": -0.05619189515709877, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05619189515709877, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1492975503206253, + 0.0513046458363533, + 0.03935345262289047, + 0.051765501499176025, + 0.018571609631180763, + 0.2858799695968628, + 0.17446322739124298, + 0.08089478313922882, + 0.11757231503725052, + 0.030896976590156555 + ], + "credit_map": [ + 0.013515655882656574, + 0.050165966153144836, + 0.010293981991708279, + 0.013397079892456532, + 0.021512627601623535, + 0.010990767739713192, + 0.01419956423342228, + 0.015221918933093548, + 0.005353717133402824, + 0.006980359088629484, + 0.005218859296292067, + 0.006394441705197096, + 0.00568437110632658, + 0.004150540102273226, + 0.010631204582750797, + 0.006891153287142515, + 0.0044970037415623665, + 0.005583113059401512, + 0.0029514096677303314, + 0.003940952010452747, + 0.005723868031054735, + 0.004266363102942705, + 0.006799979601055384, + 0.005590766668319702, + 0.00519068818539381, + 0.010192178189754486, + 0.004177391994744539, + 0.0044220685958862305, + 0.007978714071214199, + 0.005811044946312904, + 0.007315848022699356, + 0.006677569355815649, + 0.002251722849905491, + 0.0025148424319922924, + 0.0016670494806021452, + 0.0022816865239292383, + 0.002915085293352604, + 0.0021868476178497076, + 0.002943004947155714, + 0.001811372465454042, + 0.008616712875664234, + 0.11405888199806213, + 0.003394814906641841, + 0.006549106445163488, + 0.1298651546239853, + 0.007359171286225319, + 0.009214308112859726, + 0.00682185310870409, + 0.020484743639826775, + 0.026921894401311874, + 0.010860009118914604, + 0.01629672572016716, + 0.04423265904188156, + 0.022511014714837074, + 0.020153839141130447, + 0.013002355583012104, + 0.012666871771216393, + 0.012812037020921707, + 0.00485247140750289, + 0.008801302872598171, + 0.0077443430200219154, + 0.00788937695324421, + 0.008654549717903137, + 0.017473839223384857, + 0.011402875185012817, + 0.01922176033258438, + 0.011331216432154179, + 0.013215791434049606, + 0.01898135058581829, + 0.017590904608368874, + 0.01467494573444128, + 0.011153476312756538, + 0.004617397673428059, + 0.005652747116982937, + 0.0030863515567034483, + 0.004351269919425249, + 0.003792043775320053, + 0.0041517638601362705, + 0.003388690296560526, + 0.0018567148363217711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.05619189515709877 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handcrafted wooden sampan (boat)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep brown aged timber with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat's side covered in moss and water stains appearing weathered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle drone shot that slowly advances and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a calm lake with a mysterious island shrouded in thick morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048679303377866745 + ], + "opposite_logprob_tokens": [ + -3.0486793518066406 + ], + "logprob_reward": -0.048679303377866745, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048679303377866745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045199859887361526, + 0.06276882439851761, + 0.03545108065009117, + 0.026977138593792915, + 0.022065510973334312, + 0.46114981174468994, + 0.027836604043841362, + 0.24794383347034454, + 0.0423850454390049, + 0.028222233057022095 + ], + "credit_map": [ + 0.005088265053927898, + 0.0070106713101267815, + 0.00427608098834753, + 0.006431988440454006, + 0.005861278623342514, + 0.004410778637975454, + 0.006515716668218374, + 0.005605085287243128, + 0.006311081349849701, + 0.012755308300256729, + 0.005749195348471403, + 0.007959187030792236, + 0.007092030253261328, + 0.006400962360203266, + 0.008857136592268944, + 0.007643927354365587, + 0.004325365647673607, + 0.004253062419593334, + 0.0032136987429112196, + 0.005336374510079622, + 0.00522936275228858, + 0.0039743236266076565, + 0.005035439506173134, + 0.004083454143255949, + 0.0030427423771470785, + 0.005098867695778608, + 0.0019586216658353806, + 0.0025261160917580128, + 0.00434342585504055, + 0.003611889434978366, + 0.003291736589744687, + 0.0031037405133247375, + 0.002089988673105836, + 0.0035744148772209883, + 0.0017524441936984658, + 0.0019086445681750774, + 0.005283792037516832, + 0.003448250237852335, + 0.0019508304540067911, + 0.002057147677987814, + 0.012630521319806576, + 0.18647563457489014, + 0.00468592019751668, + 0.009731216356158257, + 0.2099381983280182, + 0.012417620047926903, + 0.0143801374360919, + 0.01089059840887785, + 0.004833537619560957, + 0.004936519544571638, + 0.0018847477622330189, + 0.0022543766535818577, + 0.004573045298457146, + 0.003938146401196718, + 0.0026100038085132837, + 0.002806230215355754, + 0.026646653190255165, + 0.06131761893630028, + 0.01186422910541296, + 0.024096081033349037, + 0.04382137954235077, + 0.019863778725266457, + 0.022303251549601555, + 0.038030870258808136, + 0.003997777123004198, + 0.007638200651854277, + 0.004165022633969784, + 0.004450206644833088, + 0.006471924018114805, + 0.006690281443297863, + 0.0048957751132547855, + 0.004075861535966396, + 0.004411908332258463, + 0.005674743093550205, + 0.002211397746577859, + 0.0027047451585531235, + 0.003871479071676731, + 0.003646666882559657, + 0.003399597480893135, + 0.002301699249073863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 1 + ], + "selected_reward": -0.048679303377866745 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handcrafted wooden sampan (boat)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep brown aged timber with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat's side covered in moss and water stains appearing weathered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle drone shot that slowly advances and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a calm lake with a mysterious island shrouded in thick morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0110840555280447 + ], + "opposite_logprob_tokens": [ + -4.511084079742432 + ], + "logprob_reward": -0.0110840555280447, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.0110840555280447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1256181299686432, + 0.06801809370517731, + 0.03399629890918732, + 0.041592225432395935, + 0.05478864908218384, + 0.46963149309158325, + 0.04624930024147034, + 0.09409677982330322, + 0.04031584784388542, + 0.025693155825138092 + ], + "credit_map": [ + 0.010977689176797867, + 0.04110665246844292, + 0.008315845392644405, + 0.012321625836193562, + 0.021563945338129997, + 0.008537599816918373, + 0.012648040428757668, + 0.010146725922822952, + 0.0062399972230196, + 0.016155682504177094, + 0.006058845203369856, + 0.009256109595298767, + 0.0074285767041146755, + 0.005703185219317675, + 0.0094204256311059, + 0.007755273021757603, + 0.003966114949434996, + 0.005406386684626341, + 0.002303259214386344, + 0.0036704165395349264, + 0.005758071783930063, + 0.0036218189634382725, + 0.005276412237435579, + 0.003993819933384657, + 0.004829904530197382, + 0.006834743078798056, + 0.0033494376111775637, + 0.0033574237022548914, + 0.006408333778381348, + 0.004930380266159773, + 0.006115952506661415, + 0.005766049958765507, + 0.005343344993889332, + 0.011198606342077255, + 0.002604067325592041, + 0.003966640681028366, + 0.014750623144209385, + 0.005950033199042082, + 0.006957839243113995, + 0.0040174927562475204, + 0.012839989736676216, + 0.18936537206172943, + 0.004806446377187967, + 0.009943979792296886, + 0.2139548361301422, + 0.013249988667666912, + 0.014787508174777031, + 0.010683352127671242, + 0.007738535292446613, + 0.007644192315638065, + 0.0035035188775509596, + 0.004699103534221649, + 0.005835669580847025, + 0.006030346266925335, + 0.0057437364012002945, + 0.005054198205471039, + 0.007564419414848089, + 0.01681501977145672, + 0.004457506816834211, + 0.008780648931860924, + 0.02599332109093666, + 0.009660959243774414, + 0.008046247996389866, + 0.01277865283191204, + 0.004576067440211773, + 0.004462237935513258, + 0.003867947496473789, + 0.0029669543728232384, + 0.008953879587352276, + 0.009878719225525856, + 0.002794404048472643, + 0.002815637271851301, + 0.0032865628600120544, + 0.005481193773448467, + 0.00246951705776155, + 0.0028972697909921408, + 0.002907923422753811, + 0.004426333121955395, + 0.0026676387060433626, + 0.0015567177906632423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.0110840555280447 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handcrafted wooden sampan (boat)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep brown aged timber with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat's side covered in moss and water stains appearing weathered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle drone shot that slowly advances and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a calm lake with a mysterious island shrouded in thick morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864637553691864 + ], + "opposite_logprob_tokens": [ + -3.0486464500427246 + ], + "logprob_reward": -0.04864637553691864, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864637553691864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05185064673423767, + 0.04699527472257614, + 0.031867753714323044, + 0.061327967792749405, + 0.04411397501826286, + 0.17072077095508575, + 0.10682277381420135, + 0.4409884512424469, + 0.02974560856819153, + 0.015566765330731869 + ], + "credit_map": [ + 0.004333146382123232, + 0.010455382987856865, + 0.004350228235125542, + 0.005956846754997969, + 0.0072298734448850155, + 0.005957831162959337, + 0.0070815072394907475, + 0.006485829595476389, + 0.004364301450550556, + 0.007241297047585249, + 0.004913492128252983, + 0.007583774626255035, + 0.005734311882406473, + 0.003864323254674673, + 0.006540042348206043, + 0.0067537338472902775, + 0.002707257168367505, + 0.00389655283652246, + 0.0026792052667587996, + 0.005663275718688965, + 0.005535233300179243, + 0.003231795970350504, + 0.0046936580911278725, + 0.0034607734996825457, + 0.006591515615582466, + 0.009224851615726948, + 0.005532646551728249, + 0.004833375569432974, + 0.011858458630740643, + 0.008483730256557465, + 0.0074429321102797985, + 0.007360456977039576, + 0.004306422546505928, + 0.008230138570070267, + 0.0030104878824204206, + 0.005064168479293585, + 0.009917729534208775, + 0.004490375518798828, + 0.005785741843283176, + 0.0033089115750044584, + 0.008102278225123882, + 0.06059366837143898, + 0.0041038356721401215, + 0.006695624440908432, + 0.07067938894033432, + 0.006453295238316059, + 0.007607275620102882, + 0.006485410965979099, + 0.012356670573353767, + 0.017099885269999504, + 0.005430394317954779, + 0.00923985242843628, + 0.03156987950205803, + 0.011436655186116695, + 0.01213886123150587, + 0.007550571579486132, + 0.04807373136281967, + 0.10631956160068512, + 0.02185857482254505, + 0.04651542007923126, + 0.0785902589559555, + 0.03539609909057617, + 0.0399838425219059, + 0.06425096094608307, + 0.0036282213404774666, + 0.004710135515779257, + 0.0024504116736352444, + 0.002531383652240038, + 0.007071647327393293, + 0.00612635537981987, + 0.0018173075513914227, + 0.0014101462438702583, + 0.0016953309532254934, + 0.003138461848720908, + 0.0015118863666430116, + 0.0014929035678505898, + 0.0019112923182547092, + 0.002353927120566368, + 0.0016645214054733515, + 0.0017984414007514715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.04864637553691864 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handcrafted wooden sampan (boat)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep brown aged timber with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat's side covered in moss and water stains appearing weathered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle drone shot that slowly advances and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a calm lake with a mysterious island shrouded in thick morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009779282845556736 + ], + "opposite_logprob_tokens": [ + -4.634779453277588 + ], + "logprob_reward": -0.009779282845556736, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009779282845556736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13992860913276672, + 0.030311504378914833, + 0.03588435798883438, + 0.036723632365465164, + 0.0868256464600563, + 0.44995397329330444, + 0.06354203820228577, + 0.07693229615688324, + 0.05760883167386055, + 0.02228902466595173 + ], + "credit_map": [ + 0.011502415873110294, + 0.044515155255794525, + 0.008908662013709545, + 0.011023043654859066, + 0.018489476293325424, + 0.012326192110776901, + 0.013208027929067612, + 0.019955649971961975, + 0.00325564737431705, + 0.006794488988816738, + 0.0026420187205076218, + 0.003576883114874363, + 0.0033272793516516685, + 0.0038120748940855265, + 0.003563665086403489, + 0.003339446382597089, + 0.003607802093029022, + 0.006571280304342508, + 0.0027589837554842234, + 0.003978924825787544, + 0.0053657665848731995, + 0.004937076475471258, + 0.0043159895576536655, + 0.004348535556346178, + 0.0037832981906831264, + 0.008399495854973793, + 0.00396614009514451, + 0.004083447623997927, + 0.00532658863812685, + 0.0028711222112178802, + 0.0041835433803498745, + 0.004109998699277639, + 0.007658626884222031, + 0.01929575204849243, + 0.004916850011795759, + 0.00788336805999279, + 0.023734165355563164, + 0.008810790255665779, + 0.009053564630448818, + 0.0054725282825529575, + 0.013078399002552032, + 0.18186478316783905, + 0.004845003131777048, + 0.009421168826520443, + 0.2033407986164093, + 0.012272447347640991, + 0.01406104862689972, + 0.011070344597101212, + 0.006797716952860355, + 0.010510596446692944, + 0.003842828329652548, + 0.005631649866700172, + 0.01685597561299801, + 0.007651342544704676, + 0.007205863017588854, + 0.005046065431088209, + 0.007004922721534967, + 0.019133955240249634, + 0.00314392545260489, + 0.005481099709868431, + 0.022457899525761604, + 0.006443285848945379, + 0.005138803739100695, + 0.008128401823341846, + 0.009459166787564754, + 0.014021835289895535, + 0.0035234391689300537, + 0.003012642730027437, + 0.011784862726926804, + 0.010271159932017326, + 0.0030709397979080677, + 0.0024647836107760668, + 0.0031922804191708565, + 0.0049558961763978004, + 0.002311664866283536, + 0.002200379269197583, + 0.002877431455999613, + 0.0035250552464276552, + 0.0019493126310408115, + 0.0012770048342645168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.009779282845556736 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, high-angle drone shot, slowly and steadily advancing, follows a quaint, handcrafted wooden sampan as it leisurely circles a mysterious island shrouded in thick morning mist. The boat's hull is constructed from deep brown aged timber, its grain visible. Its sides, covered in moss and water stains from years of wear and tear, appear weathered. As the camera zooms in, the water's surface takes center stage. The early morning sun, just breaking through the clouds, casts a golden glow across the deep, calm lake, creating a shimmering, flickering streak of light that complements the gentle ripples created by the boat's cut. The island itself is lush and verdant, but its contours and peaks are largely hidden within a shifting, dreamy white mist, revealing only glimpses of verdant green. The scene's palette, dominated by the dusky brown of the boat, the dark green of the lake, and the warm golden hues of the sunlight, creates a mysterious atmosphere of profound tranquility, poetry, and a sense of discovery.", + "seed": 19655, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_7/index_9380_seed_19655_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05870750928297639, + "vlm_logprob": -0.05870750928297639, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.05870750928297639, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handcrafted wooden sampan (boat)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep brown aged timber with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat's side covered in moss and water stains appearing weathered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle drone shot that slowly advances and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a calm lake with a mysterious island shrouded in thick morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017206674441695213 + ], + "opposite_logprob_tokens": [ + -4.39220666885376 + ], + "logprob_reward": -0.017206674441695213, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.017206674441695213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11722993850708008, + 0.025225263088941574, + 0.06676529347896576, + 0.04836461693048477, + 0.028540905565023422, + 0.11966472119092941, + 0.08532442897558212, + 0.4502606689929962, + 0.044230129569768906, + 0.014394048601388931 + ], + "credit_map": [ + 0.01442186813801527, + 0.01001474354416132, + 0.015710683539509773, + 0.021773163229227066, + 0.014734504744410515, + 0.0113000413402915, + 0.01488275546580553, + 0.014392177574336529, + 0.0025547430850565434, + 0.0013067355612292886, + 0.004898471757769585, + 0.0048249633982777596, + 0.0022663965355604887, + 0.0022981157526373863, + 0.003806658321991563, + 0.0032691799569875, + 0.010572219267487526, + 0.003781628794968128, + 0.008794555440545082, + 0.011695876717567444, + 0.00774743827059865, + 0.00650734081864357, + 0.008528083562850952, + 0.009138158522546291, + 0.003725605783984065, + 0.003391593461856246, + 0.0073571219108998775, + 0.007862633094191551, + 0.0057801916263997555, + 0.004152543842792511, + 0.008552012033760548, + 0.007542915176600218, + 0.005990302190184593, + 0.0014152195071801543, + 0.004103172570466995, + 0.0040777381509542465, + 0.0022475167643278837, + 0.0024701508227735758, + 0.003932089079171419, + 0.004304717294871807, + 0.010599409230053425, + 0.0037663220427930355, + 0.030017316341400146, + 0.01093381829559803, + 0.006380167324095964, + 0.021868666633963585, + 0.021922452375292778, + 0.014176561497151852, + 0.010157127864658833, + 0.005937198176980019, + 0.012662054039537907, + 0.017032939940690994, + 0.006021365523338318, + 0.00813803169876337, + 0.013708021491765976, + 0.011667692102491856, + 0.05194877088069916, + 0.042627327144145966, + 0.1191958412528038, + 0.07598254084587097, + 0.03562839701771736, + 0.04335098713636398, + 0.03641588240861893, + 0.04511092230677605, + 0.005461814347654581, + 0.005150308832526207, + 0.00642732810229063, + 0.004339943639934063, + 0.004278246313333511, + 0.009058893658220768, + 0.003136794315651059, + 0.006376800127327442, + 0.0014340976485982537, + 0.0010614763014018536, + 0.0026982908602803946, + 0.002095387550070882, + 0.0011567053152248263, + 0.0015773907070979476, + 0.0023816993925720453, + 0.001989000244066119 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.017206674441695213 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handcrafted wooden sampan (boat)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep brown aged timber with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat's side covered in moss and water stains appearing weathered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle drone shot that slowly advances and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a calm lake with a mysterious island shrouded in thick morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033735211938619614 + ], + "opposite_logprob_tokens": [ + -3.4087352752685547 + ], + "logprob_reward": -0.033735211938619614, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033735211938619614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11461471766233444, + 0.049251098185777664, + 0.06369731575250626, + 0.030100157484412193, + 0.030243096873164177, + 0.21182462573051453, + 0.1045222356915474, + 0.3393917381763458, + 0.018529681488871574, + 0.03782530128955841 + ], + "credit_map": [ + 0.01448778249323368, + 0.0064499257132411, + 0.016247067600488663, + 0.016505107283592224, + 0.012403018772602081, + 0.013324212282896042, + 0.015105036087334156, + 0.020092559978365898, + 0.005103124305605888, + 0.0033351487945765257, + 0.00591205433011055, + 0.009904060512781143, + 0.007081589195877314, + 0.0049779475666582584, + 0.007414595223963261, + 0.005522575229406357, + 0.01071470882743597, + 0.004783764015883207, + 0.006625869311392307, + 0.010006595402956009, + 0.007707072421908379, + 0.005201654974371195, + 0.008297733031213284, + 0.01035991869866848, + 0.003389745019376278, + 0.002073488663882017, + 0.0035225055180490017, + 0.00407274579629302, + 0.003315843641757965, + 0.0022327364422380924, + 0.006196789443492889, + 0.005296303890645504, + 0.007138227578252554, + 0.0019126014085486531, + 0.0039864820428192616, + 0.003364765318110585, + 0.003010382177308202, + 0.003369241254404187, + 0.002753378124907613, + 0.0047080181539058685, + 0.016840752214193344, + 0.005893896333873272, + 0.05593251809477806, + 0.017822137102484703, + 0.010797519236803055, + 0.05301860719919205, + 0.03180772438645363, + 0.01971147581934929, + 0.011545573361217976, + 0.009239698760211468, + 0.014339729212224483, + 0.018019655719399452, + 0.007118785288184881, + 0.016105694696307182, + 0.013564348220825195, + 0.01458874810487032, + 0.03937271982431412, + 0.03502398356795311, + 0.08922506123781204, + 0.05699429661035538, + 0.027778882533311844, + 0.0294199101626873, + 0.028419584035873413, + 0.03315729647874832, + 0.0024955878034234047, + 0.0028321496210992336, + 0.001464132103137672, + 0.0012706124689429998, + 0.0017311840783804655, + 0.004078639671206474, + 0.001836314215324819, + 0.0028210601303726435, + 0.003461692715063691, + 0.0020588769111782312, + 0.004021854139864445, + 0.005124269984662533, + 0.003180101281031966, + 0.004868474323302507, + 0.0075480458326637745, + 0.007561987265944481 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.033735211938619614 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handcrafted wooden sampan (boat)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep brown aged timber with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat's side covered in moss and water stains appearing weathered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle drone shot that slowly advances and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a calm lake with a mysterious island shrouded in thick morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011090421117842197 + ], + "opposite_logprob_tokens": [ + -4.511090278625488 + ], + "logprob_reward": -0.011090421117842197, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011090421117842197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1036379411816597, + 0.03096799924969673, + 0.05102858692407608, + 0.08483098447322845, + 0.028012624010443687, + 0.2288980782032013, + 0.07856862246990204, + 0.26145875453948975, + 0.05137753114104271, + 0.08121887594461441 + ], + "credit_map": [ + 0.011028208769857883, + 0.009103468619287014, + 0.012693358585238457, + 0.01940680854022503, + 0.01345868781208992, + 0.008208191022276878, + 0.01473365817219019, + 0.015005561523139477, + 0.0037306046579033136, + 0.0020218905992805958, + 0.003978346474468708, + 0.0048463731072843075, + 0.003177970414981246, + 0.0032074712216854095, + 0.005183686967939138, + 0.004821657203137875, + 0.006346633657813072, + 0.0033396901562809944, + 0.006457222159951925, + 0.008103488013148308, + 0.0061164842918515205, + 0.0049413531087338924, + 0.007615603506565094, + 0.00810810923576355, + 0.006397868040949106, + 0.0056837256997823715, + 0.010320574045181274, + 0.012269395403563976, + 0.007446574047207832, + 0.0067610591650009155, + 0.020213905721902847, + 0.015737885609269142, + 0.006160383578389883, + 0.0018921622540801764, + 0.0033477272372692823, + 0.0030536153353750706, + 0.003726257709786296, + 0.002918246667832136, + 0.0031372513622045517, + 0.0037769796326756477, + 0.016315868124365807, + 0.005862693767994642, + 0.06415856629610062, + 0.018277689814567566, + 0.010788515210151672, + 0.06470976769924164, + 0.029897112399339676, + 0.01888786070048809, + 0.010031859390437603, + 0.007268787361681461, + 0.0094681978225708, + 0.011806788854300976, + 0.005019159056246281, + 0.013319543562829494, + 0.00972256250679493, + 0.011931722983717918, + 0.029934730380773544, + 0.02657260186970234, + 0.06792251765727997, + 0.043288279324769974, + 0.022472338750958443, + 0.02403128892183304, + 0.021824916824698448, + 0.02541208639740944, + 0.005554962903261185, + 0.007295479066669941, + 0.009985285811126232, + 0.006181345321238041, + 0.004896616097539663, + 0.008045670576393604, + 0.0039694164879620075, + 0.005448752548545599, + 0.007609634194523096, + 0.005377460271120071, + 0.00783875398337841, + 0.010740759782493114, + 0.008392177522182465, + 0.011093498207628727, + 0.015491222962737083, + 0.01467536948621273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.011090421117842197 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handcrafted wooden sampan (boat)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep brown aged timber with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat's side covered in moss and water stains appearing weathered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle drone shot that slowly advances and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a calm lake with a mysterious island shrouded in thick morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22555239498615265 + ], + "opposite_logprob_tokens": [ + -1.6005524396896362 + ], + "logprob_reward": -0.22555239498615265, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22555239498615265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34140655398368835, + 0.05921167880296707, + 0.026178790256381035, + 0.1208360493183136, + 0.05330963432788849, + 0.22057315707206726, + 0.046184610575437546, + 0.03185272589325905, + 0.016883837059140205, + 0.08356296271085739 + ], + "credit_map": [ + 0.037179943174123764, + 0.028557393699884415, + 0.049853477627038956, + 0.062011733651161194, + 0.037878476083278656, + 0.03258679434657097, + 0.04219060763716698, + 0.05114812031388283, + 0.0073259142227470875, + 0.003522144863381982, + 0.008561419323086739, + 0.010363788343966007, + 0.006627450697124004, + 0.004913642071187496, + 0.00867888517677784, + 0.009218435734510422, + 0.0034927241504192352, + 0.0018657597247511148, + 0.0023391067516058683, + 0.0035132537595927715, + 0.003316788235679269, + 0.003365668235346675, + 0.004558277782052755, + 0.003727211384102702, + 0.008382325991988182, + 0.007533695083111525, + 0.015889983624219894, + 0.01843046024441719, + 0.010294992476701736, + 0.008823845535516739, + 0.030672820284962654, + 0.02080792561173439, + 0.007486832793802023, + 0.003471223171800375, + 0.006642874330282211, + 0.007011109031736851, + 0.007813060656189919, + 0.005082021467387676, + 0.008588653057813644, + 0.00721386494114995, + 0.017632264643907547, + 0.0056951018050313, + 0.06339602917432785, + 0.018075130879878998, + 0.009992125444114208, + 0.0594649538397789, + 0.027379579842090607, + 0.018937982618808746, + 0.004959038458764553, + 0.004449593368917704, + 0.00617248797789216, + 0.006448470521718264, + 0.004043016582727432, + 0.004711336921900511, + 0.008601102977991104, + 0.006799561902880669, + 0.0032661326695233583, + 0.0027292717713862658, + 0.008388477377593517, + 0.005537483375519514, + 0.002760018687695265, + 0.0030289406422525644, + 0.0028969664126634598, + 0.0032454333268105984, + 0.001387556898407638, + 0.0017455309862270951, + 0.0034083949867635965, + 0.0023975137155503035, + 0.0016471152193844318, + 0.0018936611013486981, + 0.0022167351562529802, + 0.0021873293444514275, + 0.008275315165519714, + 0.0068703000433743, + 0.009803050197660923, + 0.00998734775930643, + 0.008773265406489372, + 0.012643259018659592, + 0.013543583452701569, + 0.013666841201484203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.22555239498615265 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handcrafted wooden sampan (boat)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep brown aged timber with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat's side covered in moss and water stains appearing weathered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle drone shot that slowly advances and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a calm lake with a mysterious island shrouded in thick morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005952843930572271 + ], + "opposite_logprob_tokens": [ + -5.130952835083008 + ], + "logprob_reward": -0.005952843930572271, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005952843930572271, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11593309789896011, + 0.01706426776945591, + 0.029963254928588867, + 0.045986440032720566, + 0.028363816440105438, + 0.12023523449897766, + 0.08274780213832855, + 0.08676102757453918, + 0.34395483136177063, + 0.12899023294448853 + ], + "credit_map": [ + 0.013024037703871727, + 0.010181189514696598, + 0.016954952850937843, + 0.022310130298137665, + 0.013474816456437111, + 0.00975459348410368, + 0.014145322144031525, + 0.01608806662261486, + 0.0020053423941135406, + 0.0013792180689051747, + 0.002392664784565568, + 0.0025835493579506874, + 0.0017869172152131796, + 0.0013012824347242713, + 0.0023254130501300097, + 0.003289880696684122, + 0.003767717396840453, + 0.0019929981790483, + 0.0036701872013509274, + 0.00327905616723001, + 0.0044927336275577545, + 0.0034414154943078756, + 0.004242394585162401, + 0.005076753441244364, + 0.0029867186676710844, + 0.0031524882651865482, + 0.005849895533174276, + 0.0057573639787733555, + 0.00521713076159358, + 0.003844634396955371, + 0.01028071716427803, + 0.008897489868104458, + 0.002799283480271697, + 0.001860702526755631, + 0.003858808893710375, + 0.0035731059033423662, + 0.003664680989459157, + 0.0028128933627158403, + 0.005534874275326729, + 0.004259468521922827, + 0.013416387140750885, + 0.003247662214562297, + 0.028368642553687096, + 0.010235201567411423, + 0.005872555077075958, + 0.027528082951903343, + 0.020136376842856407, + 0.011430319398641586, + 0.00874217040836811, + 0.008900993503630161, + 0.0099464301019907, + 0.012184402905404568, + 0.00620662746950984, + 0.013085760176181793, + 0.011326080188155174, + 0.012355337850749493, + 0.008187392726540565, + 0.005060748662799597, + 0.023661203682422638, + 0.01611185446381569, + 0.006471662782132626, + 0.006702380254864693, + 0.009394198656082153, + 0.01117158867418766, + 0.04682614281773567, + 0.062005218118429184, + 0.029643407091498375, + 0.02218254655599594, + 0.04384872689843178, + 0.0679616704583168, + 0.025576334446668625, + 0.045910775661468506, + 0.012975811026990414, + 0.011755120009183884, + 0.01423090323805809, + 0.01398764830082655, + 0.011626237072050571, + 0.018294157460331917, + 0.017100000753998756, + 0.029020346701145172 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.005952843930572271 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, high-angle drone shot, slowly and steadily advancing, follows a quaint, handcrafted wooden sampan as it leisurely circles a mysterious island shrouded in thick morning mist. The boat's hull is constructed from deep brown aged timber, its grain visible. Its sides, covered in moss and water stains from years of wear and tear, appear weathered. As the camera zooms in, the water's surface takes center stage. The early morning sun, just breaking through the clouds, casts a golden glow across the deep, calm lake, creating a shimmering, flickering streak of light that complements the gentle ripples created by the boat's cut. The island itself is lush and verdant, but its contours and peaks are largely hidden within a shifting, dreamy white mist, revealing only glimpses of verdant green. The scene's palette, dominated by the dusky brown of the boat, the dark green of the lake, and the warm golden hues of the sunlight, creates a mysterious atmosphere of profound tranquility, poetry, and a sense of discovery.", + "seed": 19656, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_7/index_9380_seed_19656_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.023105545528233053, + "vlm_logprob": -0.023105545528233053, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.023105545528233053, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handcrafted wooden sampan (boat)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep brown aged timber with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat's side covered in moss and water stains appearing weathered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle drone shot that slowly advances and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a calm lake with a mysterious island shrouded in thick morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020947763696312904 + ], + "opposite_logprob_tokens": [ + -4.2709479331970215 + ], + "logprob_reward": -0.020947763696312904, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.020947763696312904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.139567032456398, + 0.06803761422634125, + 0.035172704607248306, + 0.07857388257980347, + 0.053661953657865524, + 0.12200729548931122, + 0.12579211592674255, + 0.15363334119319916, + 0.057040948420763016, + 0.1665131151676178 + ], + "credit_map": [ + 0.014792900532484055, + 0.011990347877144814, + 0.016725003719329834, + 0.019619043916463852, + 0.02837788313627243, + 0.014814984984695911, + 0.021449008956551552, + 0.011797853745520115, + 0.006477195769548416, + 0.007133624516427517, + 0.004338075406849384, + 0.011520573869347572, + 0.008659464307129383, + 0.015287274494767189, + 0.006181422155350447, + 0.008439983241260052, + 0.005139652639627457, + 0.004675659351050854, + 0.0017541386187076569, + 0.004910971969366074, + 0.011571841314435005, + 0.0033985828049480915, + 0.001981128007173538, + 0.0017407301347702742, + 0.008938116021454334, + 0.018366502597928047, + 0.005233959760516882, + 0.006334768608212471, + 0.014675146900117397, + 0.013644322752952576, + 0.005432284437119961, + 0.005948781035840511, + 0.0050917258486151695, + 0.0072884769178926945, + 0.005097461864352226, + 0.009609692730009556, + 0.007759867236018181, + 0.007625860627740622, + 0.004890745505690575, + 0.0062981233932077885, + 0.014639567583799362, + 0.012336641550064087, + 0.018641717731952667, + 0.011691334657371044, + 0.027158129960298538, + 0.01743725873529911, + 0.00852226559072733, + 0.011580376885831356, + 0.016058221459388733, + 0.021681474521756172, + 0.021852945908904076, + 0.012306246906518936, + 0.017504200339317322, + 0.013524436391890049, + 0.013271131552755833, + 0.00959346629679203, + 0.013667655177414417, + 0.015233982354402542, + 0.022719856351614, + 0.016755329445004463, + 0.028705419972538948, + 0.026020627468824387, + 0.01413082703948021, + 0.016399646177887917, + 0.006990592926740646, + 0.00555863743647933, + 0.009582476690411568, + 0.004658723250031471, + 0.014400412328541279, + 0.008281472139060497, + 0.0038220835849642754, + 0.0037465475033968687, + 0.03756938874721527, + 0.012129068374633789, + 0.027110818773508072, + 0.013453071936964989, + 0.017140474170446396, + 0.03015209175646305, + 0.01333754975348711, + 0.015620659105479717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.020947763696312904 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handcrafted wooden sampan (boat)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep brown aged timber with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat's side covered in moss and water stains appearing weathered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle drone shot that slowly advances and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a calm lake with a mysterious island shrouded in thick morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372944891452789 + ], + "opposite_logprob_tokens": [ + -3.4087295532226562 + ], + "logprob_reward": -0.03372944891452789, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372944891452789, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3509010970592499, + 0.058091070502996445, + 0.10473905503749847, + 0.04357187822461128, + 0.02035478875041008, + 0.11476573348045349, + 0.03534109890460968, + 0.037893105298280716, + 0.07244181632995605, + 0.1619003266096115 + ], + "credit_map": [ + 0.05392349883913994, + 0.022007713094353676, + 0.0237555131316185, + 0.037039946764707565, + 0.11293770372867584, + 0.045774828642606735, + 0.037319548428058624, + 0.01814231276512146, + 0.0063441358506679535, + 0.006690292619168758, + 0.005119542591273785, + 0.0058517176657915115, + 0.00829656794667244, + 0.01033694576472044, + 0.005804758053272963, + 0.009647110477089882, + 0.006293787620961666, + 0.009336160495877266, + 0.0030061991419643164, + 0.00698512839153409, + 0.04019538313150406, + 0.03180443122982979, + 0.0031463189516216516, + 0.003971643280237913, + 0.005977068096399307, + 0.00832778587937355, + 0.003191061085090041, + 0.003922407049685717, + 0.009859748184680939, + 0.00685493927448988, + 0.002306252019479871, + 0.0031326194293797016, + 0.00212981179356575, + 0.0027567725628614426, + 0.002214704407379031, + 0.004296867176890373, + 0.0021060965955257416, + 0.002880786545574665, + 0.0018186357337981462, + 0.0021511155646294355, + 0.01326882466673851, + 0.01586824841797352, + 0.017061268910765648, + 0.011534064076840878, + 0.022850167006254196, + 0.01604115031659603, + 0.007737046107649803, + 0.010404967702925205, + 0.004118027165532112, + 0.004805486183613539, + 0.007236323319375515, + 0.0020703976042568684, + 0.008372055366635323, + 0.0041685933247208595, + 0.0028928876854479313, + 0.0016773248789831996, + 0.00362870329990983, + 0.004392734728753567, + 0.005076326429843903, + 0.0038155773654580116, + 0.0057437787763774395, + 0.007432234939187765, + 0.0035153517965227365, + 0.004288397263735533, + 0.008507742546498775, + 0.006245715543627739, + 0.012987788766622543, + 0.004912670236080885, + 0.022383661940693855, + 0.008542919531464577, + 0.004222682677209377, + 0.004638635087758303, + 0.03501347452402115, + 0.011975780129432678, + 0.02583877183496952, + 0.014343907125294209, + 0.017801964655518532, + 0.027653370052576065, + 0.013727131299674511, + 0.015545940957963467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.03372944891452789 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handcrafted wooden sampan (boat)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep brown aged timber with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat's side covered in moss and water stains appearing weathered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle drone shot that slowly advances and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a calm lake with a mysterious island shrouded in thick morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608150638639927 + ], + "opposite_logprob_tokens": [ + -4.1410813331604 + ], + "logprob_reward": -0.01608150638639927, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608150638639927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09140840172767639, + 0.09328863769769669, + 0.12833112478256226, + 0.09964200854301453, + 0.04778129607439041, + 0.14852268993854523, + 0.06139324605464935, + 0.06975328922271729, + 0.07767248153686523, + 0.18220683932304382 + ], + "credit_map": [ + 0.01091007050126791, + 0.010545837692916393, + 0.013798083178699017, + 0.010832874104380608, + 0.015936363488435745, + 0.010069737210869789, + 0.012012713588774204, + 0.007302721496671438, + 0.010653512552380562, + 0.010954885743558407, + 0.009220222011208534, + 0.01568111777305603, + 0.012164822779595852, + 0.013536600396037102, + 0.008111415430903435, + 0.012966062873601913, + 0.009805632755160332, + 0.019760683178901672, + 0.00609419122338295, + 0.008928384631872177, + 0.040806036442518234, + 0.03162769600749016, + 0.004961722996085882, + 0.006346777081489563, + 0.010616152547299862, + 0.024556554853916168, + 0.0075166537426412106, + 0.009642551653087139, + 0.015097555704414845, + 0.01577703282237053, + 0.007361475843936205, + 0.009074036963284016, + 0.006646868772804737, + 0.006638894788920879, + 0.006380589213222265, + 0.004050562623888254, + 0.006437067873775959, + 0.007937244139611721, + 0.0043450696393847466, + 0.005345000885426998, + 0.016203658655285835, + 0.022078432142734528, + 0.02103554457426071, + 0.014340689405798912, + 0.0313456729054451, + 0.021812355145812035, + 0.009423978626728058, + 0.012282363139092922, + 0.006285557523369789, + 0.00746167404577136, + 0.013285303488373756, + 0.005264333449304104, + 0.011824602261185646, + 0.007012803573161364, + 0.00553390895947814, + 0.004725067876279354, + 0.005638479720801115, + 0.00787093210965395, + 0.01143371220678091, + 0.006086294073611498, + 0.011665700934827328, + 0.014476926997303963, + 0.006426994688808918, + 0.006154245231300592, + 0.008966922760009766, + 0.0060213422402739525, + 0.014640294946730137, + 0.0036228427197784185, + 0.02463860623538494, + 0.011164164170622826, + 0.004323115572333336, + 0.004295187536627054, + 0.03139355033636093, + 0.017673812806606293, + 0.026359038427472115, + 0.018033530563116074, + 0.02137216553092003, + 0.03551044315099716, + 0.018567107617855072, + 0.013297189027071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 2 + ], + "selected_reward": -0.01608150638639927 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handcrafted wooden sampan (boat)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep brown aged timber with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat's side covered in moss and water stains appearing weathered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle drone shot that slowly advances and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a calm lake with a mysterious island shrouded in thick morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033695563673973083 + ], + "opposite_logprob_tokens": [ + -3.4086954593658447 + ], + "logprob_reward": -0.033695563673973083, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033695563673973083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17180131375789642, + 0.09120739996433258, + 0.10250348597764969, + 0.11038396507501602, + 0.14602208137512207, + 0.021497061476111412, + 0.03196946531534195, + 0.075429268181324, + 0.16900213062763214, + 0.08018383383750916 + ], + "credit_map": [ + 0.02407207153737545, + 0.01536573376506567, + 0.014802335761487484, + 0.022391073405742645, + 0.04469888284802437, + 0.020422974601387978, + 0.019621919840574265, + 0.010426328517496586, + 0.015039867721498013, + 0.011896598152816296, + 0.007843738421797752, + 0.011729951947927475, + 0.015919020399451256, + 0.01344307977706194, + 0.00728077907115221, + 0.008054360747337341, + 0.010597724467515945, + 0.012415110133588314, + 0.007068428676575422, + 0.01670328713953495, + 0.01211159210652113, + 0.02310548722743988, + 0.008558094501495361, + 0.011943768709897995, + 0.011482195928692818, + 0.03051409125328064, + 0.007047286257147789, + 0.011789040639996529, + 0.017279982566833496, + 0.016574209555983543, + 0.006913906894624233, + 0.008783247321844101, + 0.016240553930401802, + 0.024228530004620552, + 0.015311836265027523, + 0.02070005238056183, + 0.017239123582839966, + 0.017527204006910324, + 0.015755783766508102, + 0.0190189927816391, + 0.00259387562982738, + 0.0019619406666606665, + 0.0034385209437459707, + 0.0026351292617619038, + 0.003978675696998835, + 0.0033002470154315233, + 0.0017899761442095041, + 0.0017986954189836979, + 0.0036093578673899174, + 0.003737351391464472, + 0.0052301413379609585, + 0.0028668094892054796, + 0.006550679448992014, + 0.003955305553972721, + 0.0030547422356903553, + 0.0029650821816176176, + 0.0068510426208376884, + 0.007066971622407436, + 0.010811347514390945, + 0.007017653435468674, + 0.011887683533132076, + 0.01714337058365345, + 0.006865169387310743, + 0.0077860294841229916, + 0.01895655132830143, + 0.014788358472287655, + 0.030675554648041725, + 0.011823245324194431, + 0.046904828399419785, + 0.020434807986021042, + 0.014674379490315914, + 0.010744396597146988, + 0.01880323700606823, + 0.005972936283797026, + 0.013535701669752598, + 0.005463223904371262, + 0.007377850357443094, + 0.015597174875438213, + 0.006451264023780823, + 0.0069824447855353355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.033695563673973083 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handcrafted wooden sampan (boat)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep brown aged timber with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat's side covered in moss and water stains appearing weathered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle drone shot that slowly advances and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a calm lake with a mysterious island shrouded in thick morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011073444969952106 + ], + "opposite_logprob_tokens": [ + -4.511073589324951 + ], + "logprob_reward": -0.011073444969952106, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011073444969952106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2987746596336365, + 0.14338476955890656, + 0.12441886216402054, + 0.041447993367910385, + 0.05017166584730148, + 0.056616004556417465, + 0.03798748180270195, + 0.044617392122745514, + 0.0690990760922432, + 0.13348208367824554 + ], + "credit_map": [ + 0.04602688178420067, + 0.02883591502904892, + 0.021550102159380913, + 0.04163966327905655, + 0.06723421812057495, + 0.03160322085022926, + 0.03785967826843262, + 0.024025004357099533, + 0.014570847153663635, + 0.014699467457830906, + 0.011860853061079979, + 0.02233031578361988, + 0.01887531392276287, + 0.03686804696917534, + 0.012096407823264599, + 0.012083514593541622, + 0.011661049909889698, + 0.017077714204788208, + 0.004857235588133335, + 0.008165335282683372, + 0.04094446823000908, + 0.030005276203155518, + 0.005290788598358631, + 0.006416989024728537, + 0.00510387122631073, + 0.006407815497368574, + 0.0034536486491560936, + 0.0034161629155278206, + 0.010876130312681198, + 0.00672910688444972, + 0.0026152911596000195, + 0.002845965325832367, + 0.004045930691063404, + 0.006913486402481794, + 0.005082345567643642, + 0.009510227479040623, + 0.006540497299283743, + 0.007336883340030909, + 0.003984563518315554, + 0.006757731083780527, + 0.006249940488487482, + 0.0071511040441691875, + 0.00794622115790844, + 0.005263026803731918, + 0.013154656626284122, + 0.00835531298071146, + 0.0037569240666925907, + 0.004738818388432264, + 0.0038637297693639994, + 0.004864455666393042, + 0.007832301780581474, + 0.003211510367691517, + 0.007183399982750416, + 0.00403253547847271, + 0.00365861551836133, + 0.003340934170410037, + 0.0038102592807263136, + 0.005321112461388111, + 0.005725935101509094, + 0.0028613856993615627, + 0.007942252792418003, + 0.011217890307307243, + 0.003690254408866167, + 0.00404830276966095, + 0.008979964070022106, + 0.004601913969963789, + 0.013261072337627411, + 0.0037859156727790833, + 0.020683055743575096, + 0.00884318072348833, + 0.004660087171941996, + 0.004283890128135681, + 0.02877890132367611, + 0.010465848259627819, + 0.02204279601573944, + 0.010943790897727013, + 0.012725143693387508, + 0.02575131319463253, + 0.011670557782053947, + 0.011103728786110878 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.011073444969952106 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_8/index_tensor([8917, 8917])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_8/index_tensor([8917, 8917])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d8ee2437031fb653bb61e9f5c1c70b241892b876 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_8/index_tensor([8917, 8917])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-angle, wide-angle lens slowly pulls up and back, overlooking a vast green plain bathed in the golden afterglow of sunset, dotted with a few weathered boulders. From all sides of the frame, groups of people dressed in ancient attire approach with a determined and deliberate pace. Their attire varies, from rough earth-toned linen robes to flowing tunics in deep blue and forest green, embellished with simple embroidery. Some even wear worn leather wrist guards and simple breastplates. Long shadows stretch out behind them. They exchange no words, yet their movements are in silent sync. Their paths flow like countless streams, steadily and inexorably converging toward the center of the plain. As the crowd converges, they seamlessly merge into an ever-expanding, densely packed circle. Standing shoulder to shoulder, facing the center, they ultimately form a vast, united, and powerful entity. The scene, bathed in warm, dramatic light, exudes a solemn, awe-inspiring sense of historical purpose, heralding the arrival of a momentous occasion.", + "seed": 30460, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_8/index_8917_seed_30460_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02604283131659031, + "vlm_logprob": -0.02604283131659031, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.02604283131659031, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide-angle view that slowly pulls up and back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a vast green plain bathed in the golden afterglow of sunset?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are groups of people wearing ancient attire (linen robes, tunics, leather guards) approaching from all sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the people move with a determined pace, forming a densely packed circle that converges toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows visible stretching out behind the people?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06336396932601929 + ], + "opposite_logprob_tokens": [ + -3.938364028930664 + ], + "logprob_reward": -0.06336396932601929, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.06336396932601929, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22039565443992615, + 0.107608862221241, + 0.11556452512741089, + 0.06814373284578323, + 0.024672379717230797, + 0.12868990004062653, + 0.15261821448802948, + 0.04793630167841911, + 0.09473003447055817, + 0.039640363305807114 + ], + "credit_map": [ + 0.020283224061131477, + 0.026461781933903694, + 0.024352144449949265, + 0.02810220792889595, + 0.02915521338582039, + 0.03369816765189171, + 0.03509721904993057, + 0.023245690390467644, + 0.03155793994665146, + 0.008685793727636337, + 0.006929423194378614, + 0.0059867361560463905, + 0.009509812109172344, + 0.035458002239465714, + 0.005354087334126234, + 0.004127060528844595, + 0.02716093324124813, + 0.00948187243193388, + 0.00845396053045988, + 0.011662447825074196, + 0.017012767493724823, + 0.010951321572065353, + 0.013609104789793491, + 0.017232121899724007, + 0.006396916229277849, + 0.009162877686321735, + 0.009890359826385975, + 0.010150788351893425, + 0.006194887682795525, + 0.008273890241980553, + 0.007762203924357891, + 0.010311812162399292, + 0.0022281582932919264, + 0.0049832542426884174, + 0.0027133433613926172, + 0.0028421804308891296, + 0.0036662276834249496, + 0.0024075484834611416, + 0.002621239749714732, + 0.003210428636521101, + 0.009470521472394466, + 0.00982162170112133, + 0.009466052986681461, + 0.018682746216654778, + 0.007398378569632769, + 0.006320614833384752, + 0.047323841601610184, + 0.02020612731575966, + 0.019488530233502388, + 0.02012057416141033, + 0.01840885542333126, + 0.01758117973804474, + 0.01651761867105961, + 0.02153271622955799, + 0.017661264166235924, + 0.021307487040758133, + 0.005613510496914387, + 0.005829737987369299, + 0.007409024517983198, + 0.008905651979148388, + 0.004730852786451578, + 0.004924883134663105, + 0.005369655322283506, + 0.0051529849879443645, + 0.009535856544971466, + 0.008778930641710758, + 0.01709792949259281, + 0.013575354591012001, + 0.012476122938096523, + 0.007850143127143383, + 0.013238895684480667, + 0.012176798656582832, + 0.0037876255810260773, + 0.0035499820951372385, + 0.005633927881717682, + 0.004779580980539322, + 0.0037231692112982273, + 0.0037860951852053404, + 0.007965441793203354, + 0.006414542440325022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.06336396932601929 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide-angle view that slowly pulls up and back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a vast green plain bathed in the golden afterglow of sunset?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are groups of people wearing ancient attire (linen robes, tunics, leather guards) approaching from all sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the people move with a determined pace, forming a densely packed circle that converges toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows visible stretching out behind the people?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008660495281219482 + ], + "opposite_logprob_tokens": [ + -4.758660316467285 + ], + "logprob_reward": -0.008660495281219482, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008660495281219482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25054043531417847, + 0.4244156777858734, + 0.035772353410720825, + 0.04026613384485245, + 0.08236405998468399, + 0.06015224754810333, + 0.03224855288863182, + 0.018564535304903984, + 0.028692934662103653, + 0.02698313258588314 + ], + "credit_map": [ + 0.03827592357993126, + 0.03636746108531952, + 0.02571643702685833, + 0.024777786806225777, + 0.03294664993882179, + 0.032869853079319, + 0.03686460480093956, + 0.022721704095602036, + 0.08318356424570084, + 0.039362628012895584, + 0.034419648349285126, + 0.05526461452245712, + 0.027425315231084824, + 0.13553780317306519, + 0.026438171043992043, + 0.02278389036655426, + 0.007172801997512579, + 0.003958751913160086, + 0.003145484020933509, + 0.0034140064381062984, + 0.0050284541212022305, + 0.003484538523480296, + 0.004121012985706329, + 0.005447304341942072, + 0.004081476014107466, + 0.0063507938757538795, + 0.004870241042226553, + 0.005153610836714506, + 0.005249810870736837, + 0.005364707671105862, + 0.004700703080743551, + 0.004494784399867058, + 0.005569338332861662, + 0.0037086112424731255, + 0.028050901368260384, + 0.004103419370949268, + 0.003861786797642708, + 0.003344013122841716, + 0.005285867024213076, + 0.028440114110708237, + 0.004206607583910227, + 0.00540740555152297, + 0.004342118743807077, + 0.0201729703694582, + 0.0041595883667469025, + 0.002444764832034707, + 0.015340014360845089, + 0.0040787761099636555, + 0.0031466512009501457, + 0.00420233141630888, + 0.004547709133476019, + 0.004812699276953936, + 0.0030477182008326054, + 0.004557259846478701, + 0.0036510417703539133, + 0.0042831385508179665, + 0.0018801658879965544, + 0.002478280570358038, + 0.003256190102547407, + 0.0021438179537653923, + 0.0025456289295107126, + 0.001736806589178741, + 0.0016464277869090438, + 0.0028772172518074512, + 0.002383976010605693, + 0.002620761515572667, + 0.004189955070614815, + 0.0025487153325229883, + 0.0036000553518533707, + 0.003995366394519806, + 0.004649695940315723, + 0.004704406950622797, + 0.0036269905976951122, + 0.0029738617595285177, + 0.005190887022763491, + 0.002863189671188593, + 0.0030270270071923733, + 0.001946555101312697, + 0.0045425742864608765, + 0.0028120428323745728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.008660495281219482 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide-angle view that slowly pulls up and back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a vast green plain bathed in the golden afterglow of sunset?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are groups of people wearing ancient attire (linen robes, tunics, leather guards) approaching from all sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the people move with a determined pace, forming a densely packed circle that converges toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows visible stretching out behind the people?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01420082338154316 + ], + "opposite_logprob_tokens": [ + -4.264200687408447 + ], + "logprob_reward": -0.01420082338154316, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01420082338154316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12319207936525345, + 0.24302548170089722, + 0.04768675938248634, + 0.049329839646816254, + 0.25178664922714233, + 0.10765646398067474, + 0.05186726525425911, + 0.04263553023338318, + 0.03426261246204376, + 0.04855729639530182 + ], + "credit_map": [ + 0.022792115807533264, + 0.024586515501141548, + 0.008386081084609032, + 0.010457252152264118, + 0.01745992712676525, + 0.018685437738895416, + 0.012909641489386559, + 0.007915111258625984, + 0.021960068494081497, + 0.025382405146956444, + 0.0166776143014431, + 0.07222829014062881, + 0.01601872406899929, + 0.0600491464138031, + 0.015420685522258282, + 0.015288527123630047, + 0.01290645357221365, + 0.004136316478252411, + 0.005200065206736326, + 0.004481008276343346, + 0.0043882206082344055, + 0.003586983773857355, + 0.005002406891435385, + 0.007985302247107029, + 0.004994760267436504, + 0.006528613623231649, + 0.005667486228048801, + 0.0052336351945996284, + 0.005738896317780018, + 0.008254148997366428, + 0.006441428326070309, + 0.006470870226621628, + 0.017838668078184128, + 0.007422892842441797, + 0.09166586399078369, + 0.011202163062989712, + 0.006280852016061544, + 0.007701748050749302, + 0.01621283032000065, + 0.09346165508031845, + 0.007182386238127947, + 0.006332792341709137, + 0.040871452540159225, + 0.011313220486044884, + 0.004945557098835707, + 0.004870333708822727, + 0.017239229753613472, + 0.014901488088071346, + 0.005916539579629898, + 0.005937593523412943, + 0.008532091975212097, + 0.010337378829717636, + 0.005434836260974407, + 0.003891380038112402, + 0.0055495984852313995, + 0.006267849821597338, + 0.005177155137062073, + 0.006757842842489481, + 0.005030058324337006, + 0.005589967127889395, + 0.006568793207406998, + 0.004415140952914953, + 0.004757679998874664, + 0.004338892176747322, + 0.003125122981145978, + 0.003458182094618678, + 0.005007745698094368, + 0.00329068792052567, + 0.006897751707583666, + 0.004464771598577499, + 0.004282252863049507, + 0.003736096201464534, + 0.003926741424947977, + 0.0049994527362287045, + 0.0038648920599371195, + 0.006325691007077694, + 0.0030809089075773954, + 0.004393473267555237, + 0.011461259797215462, + 0.010504874400794506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.01420082338154316 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide-angle view that slowly pulls up and back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a vast green plain bathed in the golden afterglow of sunset?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are groups of people wearing ancient attire (linen robes, tunics, leather guards) approaching from all sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the people move with a determined pace, forming a densely packed circle that converges toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows visible stretching out behind the people?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014194712042808533 + ], + "opposite_logprob_tokens": [ + -4.264194488525391 + ], + "logprob_reward": -0.014194712042808533, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014194712042808533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16966703534126282, + 0.13771630823612213, + 0.06341195851564407, + 0.03662986308336258, + 0.11849924176931381, + 0.048623695969581604, + 0.08355214446783066, + 0.07689767330884933, + 0.22849184274673462, + 0.036510225385427475 + ], + "credit_map": [ + 0.019061435014009476, + 0.020568430423736572, + 0.020724525675177574, + 0.023783911019563675, + 0.017948083579540253, + 0.019677603617310524, + 0.030798612162470818, + 0.017104431986808777, + 0.010414063930511475, + 0.012325584888458252, + 0.013076812960207462, + 0.037738457322120667, + 0.01090988889336586, + 0.03588787093758583, + 0.009007671847939491, + 0.008355948142707348, + 0.013042046688497066, + 0.005807744339108467, + 0.005443945061415434, + 0.0064813848584890366, + 0.00686233676970005, + 0.007868043147027493, + 0.006827234756201506, + 0.01107922662049532, + 0.0032162887509912252, + 0.00450195511803031, + 0.0035306941717863083, + 0.006013528443872929, + 0.0036159909795969725, + 0.005146062467247248, + 0.004622755106538534, + 0.005982587579637766, + 0.008340560831129551, + 0.007290261797606945, + 0.03766534477472305, + 0.00634749373421073, + 0.005360928364098072, + 0.005631093867123127, + 0.008238300681114197, + 0.03962525725364685, + 0.003840695833787322, + 0.006561117246747017, + 0.009283879771828651, + 0.014301798306405544, + 0.0038005488459020853, + 0.0034916503354907036, + 0.003509885398671031, + 0.0038341195322573185, + 0.007545081432908773, + 0.010644963942468166, + 0.013132806867361069, + 0.014084369875490665, + 0.010679205879569054, + 0.009379912167787552, + 0.007865694351494312, + 0.010220112279057503, + 0.007608190178871155, + 0.007037112955003977, + 0.01394655555486679, + 0.01395484246313572, + 0.00602104002609849, + 0.007923144847154617, + 0.008782302029430866, + 0.011624488048255444, + 0.018260587006807327, + 0.02452532760798931, + 0.025696678087115288, + 0.02193445712327957, + 0.054254449903964996, + 0.0332043394446373, + 0.027242818847298622, + 0.023373162373900414, + 0.003205613000318408, + 0.004280016291886568, + 0.004181739408522844, + 0.004536793567240238, + 0.0032837488688528538, + 0.003180610714480281, + 0.0075279888696968555, + 0.006313713267445564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.014194712042808533 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide-angle view that slowly pulls up and back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a vast green plain bathed in the golden afterglow of sunset?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are groups of people wearing ancient attire (linen robes, tunics, leather guards) approaching from all sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the people move with a determined pace, forming a densely packed circle that converges toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows visible stretching out behind the people?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029794156551361084 + ], + "opposite_logprob_tokens": [ + -3.529794216156006 + ], + "logprob_reward": -0.029794156551361084, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029794156551361084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18719546496868134, + 0.34395724534988403, + 0.04836833477020264, + 0.030343545600771904, + 0.13850422203540802, + 0.04983752593398094, + 0.04129532352089882, + 0.03818874806165695, + 0.07792496681213379, + 0.04438460245728493 + ], + "credit_map": [ + 0.029379690065979958, + 0.024549638852477074, + 0.020093606784939766, + 0.020602615550160408, + 0.021445848047733307, + 0.029302896931767464, + 0.02559535950422287, + 0.01622581109404564, + 0.052854038774967194, + 0.02700321190059185, + 0.022864965721964836, + 0.10601600259542465, + 0.0275163222104311, + 0.06346212327480316, + 0.02150397002696991, + 0.022736622020602226, + 0.014459164813160896, + 0.003333844244480133, + 0.004567399621009827, + 0.005462991073727608, + 0.003743743058294058, + 0.00456920824944973, + 0.004464739002287388, + 0.0077672433108091354, + 0.002764277858659625, + 0.0055282521061599255, + 0.0033289496786892414, + 0.0045195696875452995, + 0.004185920115560293, + 0.0030144918709993362, + 0.0035350772086530924, + 0.003467004979029298, + 0.00990383606404066, + 0.007008988875895739, + 0.046939101070165634, + 0.006942338310182095, + 0.005907902028411627, + 0.005927901715040207, + 0.009427214041352272, + 0.046446945518255234, + 0.0032617750111967325, + 0.00309996516443789, + 0.016346722841262817, + 0.003855150891467929, + 0.003268774366006255, + 0.0025116223841905594, + 0.01366998627781868, + 0.003823524573817849, + 0.003319286974146962, + 0.004347066394984722, + 0.006630868650972843, + 0.007492764387279749, + 0.004640381783246994, + 0.005130595527589321, + 0.004612252116203308, + 0.005122109781950712, + 0.0031668455339968204, + 0.004337864462286234, + 0.010133028030395508, + 0.004067585337907076, + 0.0030533282551914454, + 0.0036477919202297926, + 0.002759892027825117, + 0.007022413890808821, + 0.006474586203694344, + 0.008305396884679794, + 0.009360818192362785, + 0.005892318673431873, + 0.018389999866485596, + 0.012672025710344315, + 0.009125816635787487, + 0.0077039944007992744, + 0.004872522316873074, + 0.0030270384158939123, + 0.008530806750059128, + 0.0041553252376616, + 0.003170184325426817, + 0.003650657832622528, + 0.008408153429627419, + 0.008569915778934956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.029794156551361084 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, wide-angle lens slowly pulls up and back, overlooking a vast green plain bathed in the golden afterglow of sunset, dotted with a few weathered boulders. From all sides of the frame, groups of people dressed in ancient attire approach with a determined and deliberate pace. Their attire varies, from rough earth-toned linen robes to flowing tunics in deep blue and forest green, embellished with simple embroidery. Some even wear worn leather wrist guards and simple breastplates. Long shadows stretch out behind them. They exchange no words, yet their movements are in silent sync. Their paths flow like countless streams, steadily and inexorably converging toward the center of the plain. As the crowd converges, they seamlessly merge into an ever-expanding, densely packed circle. Standing shoulder to shoulder, facing the center, they ultimately form a vast, united, and powerful entity. The scene, bathed in warm, dramatic light, exudes a solemn, awe-inspiring sense of historical purpose, heralding the arrival of a momentous occasion.", + "seed": 30461, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_8/index_8917_seed_30461_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.023393259011209012, + "vlm_logprob": -0.023393259011209012, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.023393259011209012, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide-angle view that slowly pulls up and back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a vast green plain bathed in the golden afterglow of sunset?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are groups of people wearing ancient attire (linen robes, tunics, leather guards) approaching from all sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the people move with a determined pace, forming a densely packed circle that converges toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows visible stretching out behind the people?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022011850029230118 + ], + "opposite_logprob_tokens": [ + -5.147011756896973 + ], + "logprob_reward": -0.022011850029230118, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.022011850029230118, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45332884788513184, + 0.033671360462903976, + 0.023154795169830322, + 0.02503090351819992, + 0.03472992405295372, + 0.022298315539956093, + 0.07646673917770386, + 0.09405907243490219, + 0.21748727560043335, + 0.019772810861468315 + ], + "credit_map": [ + 0.027642779052257538, + 0.16450516879558563, + 0.1214742362499237, + 0.03182823583483696, + 0.03235626220703125, + 0.024348460137844086, + 0.016425739973783493, + 0.03474794700741768, + 0.0040235742926597595, + 0.005241198465228081, + 0.003928361460566521, + 0.0050527979619801044, + 0.00413254089653492, + 0.00531752360984683, + 0.0031346241012215614, + 0.002840736648067832, + 0.0028061014600098133, + 0.0034960408229380846, + 0.002775683766230941, + 0.003226375672966242, + 0.002918618032708764, + 0.002907498273998499, + 0.0029785791411995888, + 0.0020458968356251717, + 0.003157049184665084, + 0.002866795752197504, + 0.0034747417084872723, + 0.004354236181825399, + 0.0027953977696597576, + 0.0041167251765728, + 0.002144401893019676, + 0.0021215565502643585, + 0.003353849519044161, + 0.003735905047506094, + 0.0038743074983358383, + 0.006124884355813265, + 0.0031515450682491064, + 0.0036441225092858076, + 0.006583858747035265, + 0.004261448513716459, + 0.003642517374828458, + 0.0031895821448415518, + 0.002969504799693823, + 0.002555037848651409, + 0.0026178660336881876, + 0.0018915360560640693, + 0.002982356818392873, + 0.0024499131832271814, + 0.00836306530982256, + 0.010650711134076118, + 0.012395902536809444, + 0.007532231044024229, + 0.010659794323146343, + 0.014379788190126419, + 0.006076738703995943, + 0.00640849769115448, + 0.012359200976788998, + 0.008920717984437943, + 0.02215602621436119, + 0.011285219341516495, + 0.008546433411538601, + 0.0076307072304189205, + 0.013333608396351337, + 0.009827149100601673, + 0.03015667200088501, + 0.026344183832406998, + 0.02421080693602562, + 0.031417928636074066, + 0.019135253503918648, + 0.025189630687236786, + 0.025138404220342636, + 0.03589437156915665, + 0.0031624515540897846, + 0.001660255016759038, + 0.0037373313680291176, + 0.003013884648680687, + 0.0018748330185189843, + 0.002250482328236103, + 0.002000578911975026, + 0.0020729920361191034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.022011850029230118 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide-angle view that slowly pulls up and back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a vast green plain bathed in the golden afterglow of sunset?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are groups of people wearing ancient attire (linen robes, tunics, leather guards) approaching from all sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the people move with a determined pace, forming a densely packed circle that converges toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows visible stretching out behind the people?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01420564204454422 + ], + "opposite_logprob_tokens": [ + -4.264205455780029 + ], + "logprob_reward": -0.01420564204454422, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01420564204454422, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5680766105651855, + 0.05124688521027565, + 0.01541661936789751, + 0.038599468767642975, + 0.030942969024181366, + 0.05810423940420151, + 0.061789292842149734, + 0.061005521565675735, + 0.045418523252010345, + 0.06939990073442459 + ], + "credit_map": [ + 0.03405402973294258, + 0.21587584912776947, + 0.10038425773382187, + 0.03152134269475937, + 0.045505862683057785, + 0.03497844561934471, + 0.0252077616751194, + 0.08054905384778976, + 0.005886740051209927, + 0.008724240586161613, + 0.006803913973271847, + 0.008807672187685966, + 0.005020684562623501, + 0.006670008879154921, + 0.005130271892994642, + 0.004203349351882935, + 0.001941939233802259, + 0.002123894402757287, + 0.0017985397716984153, + 0.0016082341317087412, + 0.002574475249275565, + 0.002127038547769189, + 0.0018186290981248021, + 0.0014238691655918956, + 0.004261675290763378, + 0.0028701932169497013, + 0.007048493716865778, + 0.006244571879506111, + 0.003895396599546075, + 0.005564495455473661, + 0.004340686835348606, + 0.004373956471681595, + 0.004223733674734831, + 0.0026733921840786934, + 0.004511069040745497, + 0.005349166225641966, + 0.002559819957241416, + 0.0035883078817278147, + 0.004968047142028809, + 0.0030694345477968454, + 0.0086197005584836, + 0.008776305243372917, + 0.006694404873996973, + 0.006431441754102707, + 0.006688385736197233, + 0.00781633984297514, + 0.006840804126113653, + 0.0062368582002818584, + 0.012005123309791088, + 0.006757555063813925, + 0.018597416579723358, + 0.00640445901080966, + 0.0033141314052045345, + 0.0048294588923454285, + 0.005443934351205826, + 0.004437214229255915, + 0.006463347934186459, + 0.00698644807562232, + 0.008536217734217644, + 0.007783882319927216, + 0.00685749901458621, + 0.010427341796457767, + 0.00862676091492176, + 0.0053240228444337845, + 0.005246725399047136, + 0.006049609277397394, + 0.005378998350352049, + 0.009842620231211185, + 0.004253357648849487, + 0.004614813718944788, + 0.004470916464924812, + 0.005561482161283493, + 0.011961604468524456, + 0.0060876780189573765, + 0.007876711897552013, + 0.014377545565366745, + 0.006628657691180706, + 0.006811984349042177, + 0.005948095582425594, + 0.009707621298730373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.01420564204454422 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide-angle view that slowly pulls up and back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a vast green plain bathed in the golden afterglow of sunset?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are groups of people wearing ancient attire (linen robes, tunics, leather guards) approaching from all sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the people move with a determined pace, forming a densely packed circle that converges toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows visible stretching out behind the people?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011084291152656078 + ], + "opposite_logprob_tokens": [ + -4.511084079742432 + ], + "logprob_reward": -0.011084291152656078, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011084291152656078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14499792456626892, + 0.03823292255401611, + 0.047478996217250824, + 0.23067606985569, + 0.0291848573833704, + 0.03437787666916847, + 0.1515945941209793, + 0.059626106172800064, + 0.12075335532426834, + 0.14307722449302673 + ], + "credit_map": [ + 0.014833834953606129, + 0.014620482921600342, + 0.03281513974070549, + 0.02094370871782303, + 0.013248427771031857, + 0.011968252249062061, + 0.012600748799741268, + 0.023967353627085686, + 0.0043065305799245834, + 0.005599442403763533, + 0.004375055432319641, + 0.005682644434273243, + 0.004850077908486128, + 0.004796327091753483, + 0.004328220151364803, + 0.004294627346098423, + 0.005679930094629526, + 0.005471598822623491, + 0.005760534200817347, + 0.006118119694292545, + 0.005925023928284645, + 0.005643562879413366, + 0.006654094438999891, + 0.006226137280464172, + 0.025607876479625702, + 0.023933110758662224, + 0.03588540852069855, + 0.03720841184258461, + 0.02587019093334675, + 0.03445976972579956, + 0.02737702801823616, + 0.020334288477897644, + 0.0033075667452067137, + 0.0037027786020189524, + 0.003892803331837058, + 0.004067148081958294, + 0.002856988226994872, + 0.004092401824891567, + 0.00427702022716403, + 0.0029881554655730724, + 0.004652294330298901, + 0.003315136069431901, + 0.005138115957379341, + 0.003824294311925769, + 0.005409825127571821, + 0.0038721247110515833, + 0.004988559987396002, + 0.003177531296387315, + 0.03234952688217163, + 0.02000710368156433, + 0.03626421093940735, + 0.011603654362261295, + 0.00935282465070486, + 0.01666594110429287, + 0.01425909623503685, + 0.01109224557876587, + 0.007449505850672722, + 0.005610535852611065, + 0.007298419717699289, + 0.008494390174746513, + 0.007250859402120113, + 0.009590496309101582, + 0.008003032766282558, + 0.005928868427872658, + 0.015469114296138287, + 0.01774720475077629, + 0.010657617822289467, + 0.015050778165459633, + 0.009344995953142643, + 0.0127218933776021, + 0.013708328828215599, + 0.026053430512547493, + 0.018559681251645088, + 0.015239300206303596, + 0.020183997228741646, + 0.03177192807197571, + 0.012201743200421333, + 0.012017452158033848, + 0.014369698241353035, + 0.018733439967036247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.011084291152656078 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide-angle view that slowly pulls up and back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a vast green plain bathed in the golden afterglow of sunset?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are groups of people wearing ancient attire (linen robes, tunics, leather guards) approaching from all sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the people move with a determined pace, forming a densely packed circle that converges toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows visible stretching out behind the people?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007627766579389572 + ], + "opposite_logprob_tokens": [ + -4.882627964019775 + ], + "logprob_reward": -0.007627766579389572, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007627766579389572, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24002763628959656, + 0.04525119438767433, + 0.07143202424049377, + 0.08647169917821884, + 0.05645923316478729, + 0.050079088658094406, + 0.060924749821424484, + 0.1055450364947319, + 0.11487416923046112, + 0.1689351499080658 + ], + "credit_map": [ + 0.022698866203427315, + 0.0362570695579052, + 0.04945798218250275, + 0.024336116388440132, + 0.02182733081281185, + 0.01651187054812908, + 0.017114030197262764, + 0.05182437226176262, + 0.004232169594615698, + 0.005786421708762646, + 0.005428158212453127, + 0.00800251867622137, + 0.004492396488785744, + 0.005229986272752285, + 0.00603075884282589, + 0.00604878319427371, + 0.007684189826250076, + 0.007969615049660206, + 0.008532941341400146, + 0.00933894794434309, + 0.008929403498768806, + 0.009674111381173134, + 0.011543854139745235, + 0.007758963853120804, + 0.006672340910881758, + 0.007066060788929462, + 0.014879122376441956, + 0.012289387173950672, + 0.010540232993662357, + 0.008801226504147053, + 0.017687847837805748, + 0.008535482920706272, + 0.006922692991793156, + 0.005190418101847172, + 0.008260926231741905, + 0.01004404854029417, + 0.005488819908350706, + 0.006473764777183533, + 0.008710515685379505, + 0.005368046462535858, + 0.008549568243324757, + 0.006823296193033457, + 0.00880060251802206, + 0.005945856217294931, + 0.0033934200182557106, + 0.005258683580905199, + 0.005572025664150715, + 0.005735631566494703, + 0.005798887461423874, + 0.010158277116715908, + 0.010838935151696205, + 0.008320107124745846, + 0.006856406107544899, + 0.006144852377474308, + 0.007713282015174627, + 0.005093999207019806, + 0.01942412368953228, + 0.010545170865952969, + 0.013391682878136635, + 0.014523951336741447, + 0.009125311858952045, + 0.01108249370008707, + 0.011963436380028725, + 0.015488864853978157, + 0.017038071528077126, + 0.012083464302122593, + 0.010659688152372837, + 0.022950679063796997, + 0.010390874929726124, + 0.009097776375710964, + 0.010273185558617115, + 0.02238042838871479, + 0.026870040223002434, + 0.01937144622206688, + 0.025658125057816505, + 0.025387929752469063, + 0.01434769481420517, + 0.015983454883098602, + 0.017680836841464043, + 0.023635627701878548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.007627766579389572 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide-angle view that slowly pulls up and back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a vast green plain bathed in the golden afterglow of sunset?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are groups of people wearing ancient attire (linen robes, tunics, leather guards) approaching from all sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the people move with a determined pace, forming a densely packed circle that converges toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows visible stretching out behind the people?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203674525022507 + ], + "opposite_logprob_tokens": [ + -2.8120367527008057 + ], + "logprob_reward": -0.06203674525022507, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203674525022507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19845259189605713, + 0.06844880431890488, + 0.015735317021608353, + 0.09548764675855637, + 0.04999389871954918, + 0.1549815684556961, + 0.13177087903022766, + 0.1295829564332962, + 0.0645885095000267, + 0.09095782786607742 + ], + "credit_map": [ + 0.013548143208026886, + 0.06791054457426071, + 0.01666346937417984, + 0.014594084583222866, + 0.01805642992258072, + 0.01424643024802208, + 0.011245264671742916, + 0.042188216000795364, + 0.007835772819817066, + 0.013452792540192604, + 0.008245660923421383, + 0.009170176461338997, + 0.006591806188225746, + 0.011561576277017593, + 0.0059678563848137856, + 0.005623162258416414, + 0.0017351724673062563, + 0.001941961352713406, + 0.00198373687453568, + 0.0019463127246126533, + 0.0018695559119805694, + 0.002388168126344681, + 0.0020684730261564255, + 0.001801936188712716, + 0.007593146525323391, + 0.006747884675860405, + 0.013713635504245758, + 0.010838107205927372, + 0.011977681890130043, + 0.014736742712557316, + 0.0183501485735178, + 0.011530296877026558, + 0.00559286680072546, + 0.006773402448743582, + 0.004923376254737377, + 0.006683411076664925, + 0.004899929277598858, + 0.006423096638172865, + 0.009215271100401878, + 0.005482547450810671, + 0.0331372432410717, + 0.023327087983489037, + 0.016491862013936043, + 0.018390057608485222, + 0.016842693090438843, + 0.017238011583685875, + 0.0148925194516778, + 0.014662093482911587, + 0.02479640394449234, + 0.010014215484261513, + 0.05031115934252739, + 0.010684706270694733, + 0.006466473452746868, + 0.009904411621391773, + 0.011411464773118496, + 0.008182048797607422, + 0.01531376875936985, + 0.011426636017858982, + 0.03768383339047432, + 0.017162777483463287, + 0.009626779705286026, + 0.013576135970652103, + 0.01694760099053383, + 0.007845413871109486, + 0.008003980852663517, + 0.0094487015157938, + 0.006463554687798023, + 0.012773925438523293, + 0.007303004153072834, + 0.0073293414898216724, + 0.00778571143746376, + 0.005480287596583366, + 0.019446376711130142, + 0.007890922017395496, + 0.012441745959222317, + 0.013329753652215004, + 0.009247304871678352, + 0.009861590340733528, + 0.006335320416837931, + 0.012404813431203365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.06203674525022507 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, wide-angle lens slowly pulls up and back, overlooking a vast green plain bathed in the golden afterglow of sunset, dotted with a few weathered boulders. From all sides of the frame, groups of people dressed in ancient attire approach with a determined and deliberate pace. Their attire varies, from rough earth-toned linen robes to flowing tunics in deep blue and forest green, embellished with simple embroidery. Some even wear worn leather wrist guards and simple breastplates. Long shadows stretch out behind them. They exchange no words, yet their movements are in silent sync. Their paths flow like countless streams, steadily and inexorably converging toward the center of the plain. As the crowd converges, they seamlessly merge into an ever-expanding, densely packed circle. Standing shoulder to shoulder, facing the center, they ultimately form a vast, united, and powerful entity. The scene, bathed in warm, dramatic light, exudes a solemn, awe-inspiring sense of historical purpose, heralding the arrival of a momentous occasion.", + "seed": 30462, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_8/index_8917_seed_30462_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01246357662603259, + "vlm_logprob": -0.01246357662603259, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.01246357662603259, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide-angle view that slowly pulls up and back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a vast green plain bathed in the golden afterglow of sunset?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are groups of people wearing ancient attire (linen robes, tunics, leather guards) approaching from all sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the people move with a determined pace, forming a densely packed circle that converges toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows visible stretching out behind the people?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02797122113406658 + ], + "opposite_logprob_tokens": [ + -4.652971267700195 + ], + "logprob_reward": -0.02797122113406658, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.02797122113406658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11484741419553757, + 0.11939186602830887, + 0.03387649357318878, + 0.10349807143211365, + 0.04113210365176201, + 0.12085024267435074, + 0.06915180385112762, + 0.03570593148469925, + 0.2779071033000946, + 0.08363891392946243 + ], + "credit_map": [ + 0.010740782134234905, + 0.017434056848287582, + 0.00993700698018074, + 0.011809447780251503, + 0.020919477567076683, + 0.016018090769648552, + 0.011216169223189354, + 0.01677238754928112, + 0.023670701310038567, + 0.013616284355521202, + 0.011737392283976078, + 0.012743917293846607, + 0.01814226061105728, + 0.01453869417309761, + 0.012164749205112457, + 0.012777871452271938, + 0.004076769575476646, + 0.005392922088503838, + 0.004331807605922222, + 0.004494977183640003, + 0.0033143784385174513, + 0.003995316568762064, + 0.003424382768571377, + 0.0048459433019161224, + 0.007358430419117212, + 0.007848584093153477, + 0.01827056333422661, + 0.01783878728747368, + 0.012428255751729012, + 0.008017879910767078, + 0.01920810341835022, + 0.012527468614280224, + 0.0033052547369152308, + 0.002919204765930772, + 0.00642310781404376, + 0.006828050594776869, + 0.003694831859320402, + 0.00423826277256012, + 0.009057317860424519, + 0.0046660760417580605, + 0.010050889104604721, + 0.011081687174737453, + 0.018428541719913483, + 0.024663100019097328, + 0.00963257160037756, + 0.015700096264481544, + 0.01872287131845951, + 0.012570490129292011, + 0.006276379805058241, + 0.005389086902141571, + 0.005464334972202778, + 0.014852643013000488, + 0.008451459929347038, + 0.006394986063241959, + 0.016511784866452217, + 0.005811136215925217, + 0.006122673861682415, + 0.003005382139235735, + 0.005486952606588602, + 0.004334789235144854, + 0.001967270392924547, + 0.0026836139149963856, + 0.004325480200350285, + 0.007779770530760288, + 0.033747438341379166, + 0.016038648784160614, + 0.08534017950296402, + 0.016294052824378014, + 0.014064610935747623, + 0.07887428253889084, + 0.01232532225549221, + 0.021222572773694992, + 0.010269192978739738, + 0.008490054868161678, + 0.010080048814415932, + 0.016691293567419052, + 0.009813846088945866, + 0.010844451375305653, + 0.007333710789680481, + 0.010116321966052055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 1 + ], + "selected_reward": -0.02797122113406658 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide-angle view that slowly pulls up and back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a vast green plain bathed in the golden afterglow of sunset?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are groups of people wearing ancient attire (linen robes, tunics, leather guards) approaching from all sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the people move with a determined pace, forming a densely packed circle that converges toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows visible stretching out behind the people?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006752410903573036 + ], + "opposite_logprob_tokens": [ + -5.0067524909973145 + ], + "logprob_reward": -0.006752410903573036, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006752410903573036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10162804275751114, + 0.1099708154797554, + 0.05694116652011871, + 0.2806278467178345, + 0.0593532919883728, + 0.103236123919487, + 0.06658127158880234, + 0.018009133636951447, + 0.13006964325904846, + 0.07358267158269882 + ], + "credit_map": [ + 0.010565340518951416, + 0.014005460776388645, + 0.016546791419386864, + 0.011069699190557003, + 0.010141639038920403, + 0.013950079679489136, + 0.010958944447338581, + 0.014390078373253345, + 0.02282511442899704, + 0.011665145866572857, + 0.012063286267220974, + 0.014623112045228481, + 0.015372177585959435, + 0.011434154585003853, + 0.00980179663747549, + 0.01218600757420063, + 0.007705208845436573, + 0.006080859340727329, + 0.007515888661146164, + 0.007587578147649765, + 0.004698492586612701, + 0.011036555282771587, + 0.006158716045320034, + 0.006157862488180399, + 0.02002195455133915, + 0.01993492990732193, + 0.04362950101494789, + 0.04005953669548035, + 0.05408144369721413, + 0.025437552481889725, + 0.05039457604289055, + 0.027068322524428368, + 0.0038545597344636917, + 0.0038685062900185585, + 0.014402558095753193, + 0.0076283677481114864, + 0.0032622008584439754, + 0.0071457764133811, + 0.013356921263039112, + 0.005834395997226238, + 0.006672586314380169, + 0.006118715275079012, + 0.01524890586733818, + 0.030593393370509148, + 0.006674989126622677, + 0.008075673133134842, + 0.02275937609374523, + 0.007092474494129419, + 0.012014059349894524, + 0.004332725889980793, + 0.005401314236223698, + 0.01123160868883133, + 0.00721268355846405, + 0.0074231927283108234, + 0.01382035855203867, + 0.005145317874848843, + 0.002375754527747631, + 0.0017614462412893772, + 0.003675696672871709, + 0.0015053603565320373, + 0.002660116646438837, + 0.002215965185314417, + 0.001794608193449676, + 0.0020201830193400383, + 0.013506641611456871, + 0.008611153811216354, + 0.031595710664987564, + 0.007323532830923796, + 0.0056015667505562305, + 0.016780680045485497, + 0.0076904636807739735, + 0.03895988687872887, + 0.009439072571694851, + 0.007962985895574093, + 0.007444786373525858, + 0.008439485915005207, + 0.010634603910148144, + 0.010950833559036255, + 0.006342184264212847, + 0.012368710711598396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 1 + ], + "selected_reward": -0.006752410903573036 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide-angle view that slowly pulls up and back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a vast green plain bathed in the golden afterglow of sunset?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are groups of people wearing ancient attire (linen robes, tunics, leather guards) approaching from all sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the people move with a determined pace, forming a densely packed circle that converges toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows visible stretching out behind the people?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006743411999195814 + ], + "opposite_logprob_tokens": [ + -5.006743431091309 + ], + "logprob_reward": -0.006743411999195814, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006743411999195814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13794530928134918, + 0.20893116295337677, + 0.05000009015202522, + 0.11150379478931427, + 0.02444295771420002, + 0.05199043080210686, + 0.22183911502361298, + 0.020759565755724907, + 0.10236561298370361, + 0.070221908390522 + ], + "credit_map": [ + 0.010197438299655914, + 0.019566381350159645, + 0.01128990389406681, + 0.014091283082962036, + 0.02008204348385334, + 0.023183509707450867, + 0.015713009983301163, + 0.023821745067834854, + 0.04216784983873367, + 0.017937863245606422, + 0.018580008298158646, + 0.022711142897605896, + 0.025477154180407524, + 0.02496459148824215, + 0.03399210795760155, + 0.02310045249760151, + 0.006185806822031736, + 0.003783526597544551, + 0.006495585199445486, + 0.005427433643490076, + 0.004389117006212473, + 0.010654263198375702, + 0.006529095117002726, + 0.006535264663398266, + 0.00835536140948534, + 0.008993452414870262, + 0.017151324078440666, + 0.015530562028288841, + 0.01668001152575016, + 0.012751958332955837, + 0.020081715658307076, + 0.011959423311054707, + 0.0015654375310987234, + 0.0012499503791332245, + 0.007671318482607603, + 0.002939187455922365, + 0.0014520190889015794, + 0.0018022883450612426, + 0.004144963808357716, + 0.0036177923902869225, + 0.0035064169205725193, + 0.0025228927843272686, + 0.003715453902259469, + 0.016548557206988335, + 0.003306391416117549, + 0.0029537593945860863, + 0.01576823927462101, + 0.0036687299143522978, + 0.011782549321651459, + 0.009692039340734482, + 0.008363195694983006, + 0.08243928849697113, + 0.008160045370459557, + 0.008096964098513126, + 0.08214409649372101, + 0.01116094272583723, + 0.002731884131208062, + 0.002543503185734153, + 0.0035740958992391825, + 0.0021461965516209602, + 0.0024293658789247274, + 0.0019584381952881813, + 0.0029173896182328463, + 0.0024586948566138744, + 0.012788236141204834, + 0.009621416218578815, + 0.029049130156636238, + 0.006786927580833435, + 0.004809231963008642, + 0.01184904295951128, + 0.005329494830220938, + 0.022132141515612602, + 0.006683496292680502, + 0.007392019033432007, + 0.007699036039412022, + 0.011639205738902092, + 0.010095678269863129, + 0.009999101981520653, + 0.006749898660928011, + 0.009963476099073887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.006743411999195814 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide-angle view that slowly pulls up and back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a vast green plain bathed in the golden afterglow of sunset?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are groups of people wearing ancient attire (linen robes, tunics, leather guards) approaching from all sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the people move with a determined pace, forming a densely packed circle that converges toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows visible stretching out behind the people?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009777866303920746 + ], + "opposite_logprob_tokens": [ + -4.634778022766113 + ], + "logprob_reward": -0.009777866303920746, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009777866303920746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06234489753842354, + 0.10873173177242279, + 0.05009722709655762, + 0.25011423230171204, + 0.03856494277715683, + 0.08500728011131287, + 0.1858983337879181, + 0.03272045776247978, + 0.1133292019367218, + 0.07319174706935883 + ], + "credit_map": [ + 0.005934107583016157, + 0.010491302236914635, + 0.005011282861232758, + 0.007362807169556618, + 0.010318925604224205, + 0.008824224583804607, + 0.007511793170124292, + 0.006890449207276106, + 0.027887240052223206, + 0.011210744269192219, + 0.012865293771028519, + 0.013247105292975903, + 0.012366839684545994, + 0.008825089782476425, + 0.010591733269393444, + 0.011737688444554806, + 0.007109040394425392, + 0.0062520927749574184, + 0.006443385034799576, + 0.005609503481537104, + 0.0051683085039258, + 0.008801513351500034, + 0.0060055493377149105, + 0.004707835614681244, + 0.017826998606324196, + 0.017612747848033905, + 0.03832966834306717, + 0.03534458950161934, + 0.04923267289996147, + 0.021513959392905235, + 0.04597935080528259, + 0.024274231866002083, + 0.0028055724687874317, + 0.0021247544791549444, + 0.01095419842749834, + 0.005135708954185247, + 0.0027974669355899096, + 0.004320418927818537, + 0.0045199329033494, + 0.0059068892151117325, + 0.005785020999610424, + 0.005122571252286434, + 0.009917587973177433, + 0.02273436449468136, + 0.008128400892019272, + 0.007584249600768089, + 0.019940678030252457, + 0.005794404074549675, + 0.012678543105721474, + 0.011206437833607197, + 0.011964808218181133, + 0.06068839877843857, + 0.009902087971568108, + 0.007639301009476185, + 0.0598653145134449, + 0.011953439563512802, + 0.003740175161510706, + 0.0041027064435184, + 0.004769092425704002, + 0.0038475932087749243, + 0.004717497155070305, + 0.0032940322998911142, + 0.004313261713832617, + 0.003936099819839001, + 0.011052864603698254, + 0.0061707631684839725, + 0.02314160205423832, + 0.007343688048422337, + 0.005351099651306868, + 0.01733163371682167, + 0.0071491957642138, + 0.03578835353255272, + 0.007531026378273964, + 0.008035961538553238, + 0.008218204602599144, + 0.014615041203796864, + 0.007799237500876188, + 0.009987680241465569, + 0.007270487491041422, + 0.009734098799526691 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 8 + ], + "selected_reward": -0.009777866303920746 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide-angle view that slowly pulls up and back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a vast green plain bathed in the golden afterglow of sunset?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are groups of people wearing ancient attire (linen robes, tunics, leather guards) approaching from all sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the people move with a determined pace, forming a densely packed circle that converges toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows visible stretching out behind the people?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072972789406776 + ], + "opposite_logprob_tokens": [ + -4.511073112487793 + ], + "logprob_reward": -0.011072972789406776, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072972789406776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06227457523345947, + 0.0339738167822361, + 0.0236620232462883, + 0.3251885771751404, + 0.22824008762836456, + 0.09828653931617737, + 0.11816783994436264, + 0.012197701260447502, + 0.0680471733212471, + 0.029961654916405678 + ], + "credit_map": [ + 0.00438044173642993, + 0.009056110866367817, + 0.007691322825849056, + 0.009632383473217487, + 0.008437557145953178, + 0.00896618515253067, + 0.006311869248747826, + 0.007798711769282818, + 0.004089674446731806, + 0.00413919473066926, + 0.003755918238312006, + 0.005410986021161079, + 0.0048482404090464115, + 0.004778796341270208, + 0.0029579049441963434, + 0.003993100021034479, + 0.0028974118176847696, + 0.0031896079890429974, + 0.0037659399677067995, + 0.0028439718298614025, + 0.002260226756334305, + 0.0023642159067094326, + 0.002810599049553275, + 0.0035300536546856165, + 0.02112843096256256, + 0.023270169273018837, + 0.050880078226327896, + 0.047733329236507416, + 0.05701679363846779, + 0.027474088594317436, + 0.06231454759836197, + 0.03537115454673767, + 0.011213650926947594, + 0.010293865576386452, + 0.05655227601528168, + 0.03100777044892311, + 0.009679879061877728, + 0.014905675314366817, + 0.06629735976457596, + 0.028289619833230972, + 0.0062615154311060905, + 0.0047914632596075535, + 0.007577369920909405, + 0.032783281058073044, + 0.006636342033743858, + 0.005883991252630949, + 0.028392352163791656, + 0.00596023490652442, + 0.00703020952641964, + 0.005372627172619104, + 0.004589971620589495, + 0.04139331355690956, + 0.0054655070416629314, + 0.00713521521538496, + 0.04130367189645767, + 0.005877327639609575, + 0.0014807652914896607, + 0.001416621496900916, + 0.0021753094624727964, + 0.0015954674454405904, + 0.0016699216794222593, + 0.0011617576237767935, + 0.001148968585766852, + 0.0015488903736695647, + 0.006525834556668997, + 0.0037417318671941757, + 0.005300663877278566, + 0.004017476923763752, + 0.005439425352960825, + 0.019841885194182396, + 0.003922893200069666, + 0.019257256761193275, + 0.0037870975211262703, + 0.0026592607609927654, + 0.003194465534761548, + 0.004882724490016699, + 0.00315132481046021, + 0.004426934290677309, + 0.002432519569993019, + 0.005427329335361719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 6 + ], + "selected_reward": -0.011072972789406776 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, wide-angle lens slowly pulls up and back, overlooking a vast green plain bathed in the golden afterglow of sunset, dotted with a few weathered boulders. From all sides of the frame, groups of people dressed in ancient attire approach with a determined and deliberate pace. Their attire varies, from rough earth-toned linen robes to flowing tunics in deep blue and forest green, embellished with simple embroidery. Some even wear worn leather wrist guards and simple breastplates. Long shadows stretch out behind them. They exchange no words, yet their movements are in silent sync. Their paths flow like countless streams, steadily and inexorably converging toward the center of the plain. As the crowd converges, they seamlessly merge into an ever-expanding, densely packed circle. Standing shoulder to shoulder, facing the center, they ultimately form a vast, united, and powerful entity. The scene, bathed in warm, dramatic light, exudes a solemn, awe-inspiring sense of historical purpose, heralding the arrival of a momentous occasion.", + "seed": 30463, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_8/index_8917_seed_30463_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01756026390939951, + "vlm_logprob": -0.01756026390939951, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.01756026390939951, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide-angle view that slowly pulls up and back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a vast green plain bathed in the golden afterglow of sunset?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are groups of people wearing ancient attire (linen robes, tunics, leather guards) approaching from all sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the people move with a determined pace, forming a densely packed circle that converges toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows visible stretching out behind the people?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.031923841685056686 + ], + "opposite_logprob_tokens": [ + -4.781923770904541 + ], + "logprob_reward": -0.031923841685056686, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.031923841685056686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1415976583957672, + 0.14158424735069275, + 0.07002489268779755, + 0.11213746666908264, + 0.13565945625305176, + 0.060564056038856506, + 0.16794265806674957, + 0.03775167837738991, + 0.053007178008556366, + 0.07973068952560425 + ], + "credit_map": [ + 0.017036370933055878, + 0.01634608767926693, + 0.014681746251881123, + 0.013755337335169315, + 0.018166719004511833, + 0.013072398491203785, + 0.030088908970355988, + 0.01845008134841919, + 0.016152629628777504, + 0.01140294037759304, + 0.020324595272541046, + 0.016865435987710953, + 0.013718044385313988, + 0.023916494101285934, + 0.021191397681832314, + 0.01801271364092827, + 0.010986048728227615, + 0.010024487040936947, + 0.009050038643181324, + 0.009049064479768276, + 0.005056648049503565, + 0.008471754379570484, + 0.01061563566327095, + 0.006771221291273832, + 0.01053221058100462, + 0.013809391297399998, + 0.010352559387683868, + 0.014245830476284027, + 0.011706177145242691, + 0.019567005336284637, + 0.014266615733504295, + 0.017657680436968803, + 0.02498898282647133, + 0.01381507609039545, + 0.016481226310133934, + 0.019564518705010414, + 0.014340314082801342, + 0.013588493689894676, + 0.019911030307412148, + 0.012969809584319592, + 0.008571689948439598, + 0.006047871895134449, + 0.007863830775022507, + 0.007811382412910461, + 0.007947090081870556, + 0.007976319640874863, + 0.007213988807052374, + 0.007131884340196848, + 0.021514881402254105, + 0.01627792790532112, + 0.023087013512849808, + 0.024047881364822388, + 0.014867736026644707, + 0.015139694325625896, + 0.024158544838428497, + 0.028848983347415924, + 0.0037093819119036198, + 0.0060773794539272785, + 0.00427625048905611, + 0.0030013578943908215, + 0.006952384021133184, + 0.003666271222755313, + 0.005608221981674433, + 0.004460431635379791, + 0.005432572215795517, + 0.005961109884083271, + 0.00825117900967598, + 0.00634233932942152, + 0.007566560059785843, + 0.004595682490617037, + 0.007384768221527338, + 0.007472964935004711, + 0.010588252916932106, + 0.007761449087411165, + 0.009753357619047165, + 0.010382168926298618, + 0.011537845246493816, + 0.008084011264145374, + 0.010173836722970009, + 0.011449763551354408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.031923841685056686 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide-angle view that slowly pulls up and back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a vast green plain bathed in the golden afterglow of sunset?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are groups of people wearing ancient attire (linen robes, tunics, leather guards) approaching from all sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the people move with a determined pace, forming a densely packed circle that converges toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows visible stretching out behind the people?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008655649609863758 + ], + "opposite_logprob_tokens": [ + -4.758655548095703 + ], + "logprob_reward": -0.008655649609863758, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008655649609863758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10989323258399963, + 0.1362321525812149, + 0.17475956678390503, + 0.06638844311237335, + 0.08860170841217041, + 0.023700060322880745, + 0.11782429367303848, + 0.07101175934076309, + 0.05374623090028763, + 0.15784257650375366 + ], + "credit_map": [ + 0.013718994334340096, + 0.01258140429854393, + 0.014795240014791489, + 0.015623657964169979, + 0.011231856420636177, + 0.013198445551097393, + 0.014911738224327564, + 0.013831887394189835, + 0.011212294921278954, + 0.009474091231822968, + 0.014622685499489307, + 0.01957051269710064, + 0.014460359700024128, + 0.024059763178229332, + 0.01649041287600994, + 0.026342036202549934, + 0.015435278415679932, + 0.015505034476518631, + 0.012388544157147408, + 0.04071281477808952, + 0.010910720564424992, + 0.02161329798400402, + 0.03716948255896568, + 0.021024391055107117, + 0.005068243481218815, + 0.007315352093428373, + 0.0068589020520448685, + 0.006942220963537693, + 0.008097467944025993, + 0.014779383316636086, + 0.008250384591519833, + 0.009076487272977829, + 0.010710532777011395, + 0.00941645447164774, + 0.00699809892103076, + 0.010591906495392323, + 0.012577020563185215, + 0.015803391113877296, + 0.013559340499341488, + 0.008944966830313206, + 0.003179956926032901, + 0.002178381197154522, + 0.0026170078199356794, + 0.0026691195089370012, + 0.0032721443567425013, + 0.0036643529310822487, + 0.003334964392706752, + 0.002784132957458496, + 0.011582998558878899, + 0.01017751358449459, + 0.014893908053636551, + 0.01575355976819992, + 0.013168619014322758, + 0.010959913022816181, + 0.024911895394325256, + 0.016375891864299774, + 0.00561040872707963, + 0.012486476451158524, + 0.006453641224652529, + 0.00511977169662714, + 0.015698356553912163, + 0.006532854400575161, + 0.009998729452490807, + 0.009111521765589714, + 0.006071425508707762, + 0.00552908331155777, + 0.008771080523729324, + 0.006330493837594986, + 0.006675715558230877, + 0.0065854210406541824, + 0.0057068136520683765, + 0.008076198399066925, + 0.0349748320877552, + 0.010762442834675312, + 0.03477299585938454, + 0.01358137745410204, + 0.011236059479415417, + 0.019285891205072403, + 0.013794193044304848, + 0.019434774294495583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 1 + ], + "selected_reward": -0.008655649609863758 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide-angle view that slowly pulls up and back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a vast green plain bathed in the golden afterglow of sunset?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are groups of people wearing ancient attire (linen robes, tunics, leather guards) approaching from all sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the people move with a determined pace, forming a densely packed circle that converges toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows visible stretching out behind the people?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007641962729394436 + ], + "opposite_logprob_tokens": [ + -4.882641792297363 + ], + "logprob_reward": -0.007641962729394436, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007641962729394436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19570070505142212, + 0.12955527007579803, + 0.06299994140863419, + 0.04350702837109566, + 0.10073667764663696, + 0.049919430166482925, + 0.06461426615715027, + 0.09978377819061279, + 0.07677163183689117, + 0.17641125619411469 + ], + "credit_map": [ + 0.016176579520106316, + 0.014074111357331276, + 0.021520135924220085, + 0.029817620292305946, + 0.022288119420409203, + 0.024359211325645447, + 0.028234481811523438, + 0.039230458438396454, + 0.01566215232014656, + 0.01332843117415905, + 0.01753123290836811, + 0.016129296272993088, + 0.010586030781269073, + 0.02356734499335289, + 0.014176794327795506, + 0.018573984503746033, + 0.005851584952324629, + 0.0058959489688277245, + 0.005837132688611746, + 0.009573997929692268, + 0.0074411677196621895, + 0.007601271383464336, + 0.00879072118550539, + 0.012008121237158775, + 0.00425333296880126, + 0.0034146527759730816, + 0.004508177749812603, + 0.007213090546429157, + 0.004073005635291338, + 0.006981142796576023, + 0.006423726212233305, + 0.006639903876930475, + 0.01658416911959648, + 0.009995990432798862, + 0.010742632672190666, + 0.012201475910842419, + 0.0117798438295722, + 0.011809351854026318, + 0.013776836916804314, + 0.013846375979483128, + 0.00965785700827837, + 0.005112756043672562, + 0.00542434211820364, + 0.0057557690888643265, + 0.006409919820725918, + 0.005006407853215933, + 0.006187706254422665, + 0.006364679429680109, + 0.005938040092587471, + 0.006174544803798199, + 0.008975505828857422, + 0.005365215241909027, + 0.008930554613471031, + 0.006329139228910208, + 0.011423869989812374, + 0.011477404274046421, + 0.008992386981844902, + 0.017346300184726715, + 0.008290869183838367, + 0.007159554865211248, + 0.026911504566669464, + 0.011410754173994064, + 0.010084880515933037, + 0.009587538428604603, + 0.009209898300468922, + 0.008526602759957314, + 0.01389623899012804, + 0.008083242923021317, + 0.00895309541374445, + 0.00781982485204935, + 0.011116599664092064, + 0.009166132658720016, + 0.015694700181484222, + 0.017687464132905006, + 0.02591368556022644, + 0.024518391117453575, + 0.023514583706855774, + 0.013832599855959415, + 0.03153761476278305, + 0.023712219670414925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.007641962729394436 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide-angle view that slowly pulls up and back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a vast green plain bathed in the golden afterglow of sunset?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are groups of people wearing ancient attire (linen robes, tunics, leather guards) approaching from all sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the people move with a determined pace, forming a densely packed circle that converges toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows visible stretching out behind the people?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978246983140707 + ], + "opposite_logprob_tokens": [ + -4.634782314300537 + ], + "logprob_reward": -0.00978246983140707, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978246983140707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16266609728336334, + 0.1284858137369156, + 0.050676412880420685, + 0.06926346570253372, + 0.08607463538646698, + 0.038496192544698715, + 0.11379819363355637, + 0.07676547765731812, + 0.10802965611219406, + 0.16574397683143616 + ], + "credit_map": [ + 0.018872113898396492, + 0.014394585974514484, + 0.014521663077175617, + 0.01744741015136242, + 0.014600010588765144, + 0.017018737271428108, + 0.043513670563697815, + 0.022297918796539307, + 0.015424617566168308, + 0.009732958860695362, + 0.018425824120640755, + 0.014172702096402645, + 0.012134869582951069, + 0.022983476519584656, + 0.018651144579052925, + 0.016960225999355316, + 0.004748286679387093, + 0.005563509184867144, + 0.004535614978522062, + 0.009471271187067032, + 0.0054651168175041676, + 0.00594751350581646, + 0.00874506589025259, + 0.006200034637004137, + 0.00719478027895093, + 0.010985271073877811, + 0.006299236789345741, + 0.010317186824977398, + 0.005977061577141285, + 0.009278052486479282, + 0.011979835107922554, + 0.007232039701193571, + 0.01195916160941124, + 0.006921485997736454, + 0.007901014760136604, + 0.010962124913930893, + 0.010203477926552296, + 0.00904669426381588, + 0.01648651994764805, + 0.012594158761203289, + 0.007637761998921633, + 0.003007961204275489, + 0.004829635843634605, + 0.004105080850422382, + 0.003822724102064967, + 0.00565040810033679, + 0.00405614310875535, + 0.005386476870626211, + 0.012784370221197605, + 0.011660481803119183, + 0.012641419656574726, + 0.014095496386289597, + 0.020383140072226524, + 0.011104666627943516, + 0.01809713989496231, + 0.013031474314630032, + 0.007507483474910259, + 0.006796194706112146, + 0.008437932468950748, + 0.0045041656121611595, + 0.009288479574024677, + 0.00489468639716506, + 0.021524064242839813, + 0.013812472112476826, + 0.00983158964663744, + 0.01381757389754057, + 0.017254110425710678, + 0.016574230045080185, + 0.015525241382420063, + 0.010053514502942562, + 0.013884512707591057, + 0.011088892817497253, + 0.015639901161193848, + 0.016634630039334297, + 0.02669375203549862, + 0.021626878529787064, + 0.023319194093346596, + 0.011520364321768284, + 0.028276121243834496, + 0.022033128887414932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.00978246983140707 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide-angle view that slowly pulls up and back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a vast green plain bathed in the golden afterglow of sunset?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are groups of people wearing ancient attire (linen robes, tunics, leather guards) approaching from all sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the people move with a determined pace, forming a densely packed circle that converges toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows visible stretching out behind the people?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029797395691275597 + ], + "opposite_logprob_tokens": [ + -3.529797315597534 + ], + "logprob_reward": -0.029797395691275597, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029797395691275597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17180851101875305, + 0.13666166365146637, + 0.17215260863304138, + 0.06656091660261154, + 0.08943045139312744, + 0.048369813710451126, + 0.0904807448387146, + 0.06540998816490173, + 0.07220317423343658, + 0.08692212402820587 + ], + "credit_map": [ + 0.02106454223394394, + 0.017138708382844925, + 0.022038208320736885, + 0.024091651663184166, + 0.01822461001574993, + 0.021729428321123123, + 0.02236083336174488, + 0.02516053058207035, + 0.012363037094473839, + 0.00993132870644331, + 0.019450772553682327, + 0.025356678292155266, + 0.008213533088564873, + 0.014452018775045872, + 0.016723360866308212, + 0.030170947313308716, + 0.014930589124560356, + 0.01740814931690693, + 0.014536233618855476, + 0.03086341731250286, + 0.01406081672757864, + 0.020566808059811592, + 0.033879078924655914, + 0.025907527655363083, + 0.005084303207695484, + 0.006976898293942213, + 0.006237104535102844, + 0.006910983007401228, + 0.009284782223403454, + 0.012427243404090405, + 0.009182307869195938, + 0.010457292199134827, + 0.010360386222600937, + 0.011206804774701595, + 0.011507554911077023, + 0.007990519516170025, + 0.010037881322205067, + 0.009243657812476158, + 0.016868477687239647, + 0.012215172871947289, + 0.006028069648891687, + 0.004542297217994928, + 0.005542926024645567, + 0.006908086594194174, + 0.00423424132168293, + 0.006889307405799627, + 0.007819297723472118, + 0.006405593361705542, + 0.010362333618104458, + 0.010524524375796318, + 0.011952796019613743, + 0.010764646343886852, + 0.00927064660936594, + 0.008584328927099705, + 0.016035206615924835, + 0.012986265122890472, + 0.00514220492914319, + 0.009499859996140003, + 0.006715249735862017, + 0.005847351159900427, + 0.009867715649306774, + 0.00710379658266902, + 0.012302812188863754, + 0.008930997923016548, + 0.00794248003512621, + 0.007809425704181194, + 0.011739326640963554, + 0.007627933286130428, + 0.01040150597691536, + 0.008013250306248665, + 0.009414413012564182, + 0.009254836477339268, + 0.011902721598744392, + 0.0061079878360033035, + 0.01742812991142273, + 0.01259244978427887, + 0.008916324004530907, + 0.005445521324872971, + 0.014194735325872898, + 0.010334254242479801 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.029797395691275597 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_9/index_tensor([5302, 5302])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_9/index_tensor([5302, 5302])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9bd94c33f87cf0209a85d603f1321012471b83ac --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_9/index_tensor([5302, 5302])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot, shot from directly above the desk, then orbits in an incredibly slow, smooth arc around the center of the writing action, immersing the viewer in the meticulous detail of the composition. A sheet of slightly yellowed rice paper lies flat on a solid wooden desk, its dark grain and subtle signs of age. A slender, powerful hand, defined by distinct joints, grips a dark brown bamboo-handled brush with expert grace. The suspended wrist demonstrates precise control of pressure. A soft, warm light slants down from above and to the side, like afternoon sunlight filtering through a window lattice, gilding the pristine inkstone and jade paperweight with a gentle patina. As the camera pans, we can clearly see the glossy black brush tip, saturated with rich ink, moving effortlessly across the paper. With each stroke, the thick black ink slowly seeps through the fibers of the rice paper, creating a unique texture with a slightly rough edge and a rich, lustrous center. The dominant color palette of the entire scene, composed of deep wood tones, warm beige, and pure black, creates an atmosphere of tranquility, concentration, and a sense of classic charm and Zen.", + "seed": 5245, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_9/index_5302_seed_5245_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03574282424524426, + "vlm_logprob": -0.03574282424524426, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.03574282424524426, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from directly above the desk and does it orbit in a slow, smooth arc around the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender hand with distinct joints gripping a dark brown bamboo-handled brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sheet of slightly yellowed rice paper lying flat on a solid wooden desk with visible dark grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glossy black brush tip saturated with rich ink moving across the paper, creating thick black ink strokes that seep into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm light slanting from above and to the side, gilding a pristine inkstone and jade paperweight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12898045778274536 + ], + "opposite_logprob_tokens": [ + -2.3789803981781006 + ], + "logprob_reward": -0.12898045778274536, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.12898045778274536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41921353340148926, + 0.04352512210607529, + 0.0658349096775055, + 0.04942018538713455, + 0.04310554265975952, + 0.09734819084405899, + 0.057657383382320404, + 0.1080828607082367, + 0.0755341425538063, + 0.04027824103832245 + ], + "credit_map": [ + 0.04445647820830345, + 0.05672115832567215, + 0.05545608326792717, + 0.06382299214601517, + 0.03667315095663071, + 0.034032005816698074, + 0.06203727051615715, + 0.06601439416408539, + 0.003948050085455179, + 0.004937584511935711, + 0.007964055985212326, + 0.006384815089404583, + 0.004662493243813515, + 0.0032792221754789352, + 0.006945391651242971, + 0.00540350703522563, + 0.01021116878837347, + 0.004458561539649963, + 0.008205372840166092, + 0.012406410649418831, + 0.006237409543246031, + 0.007159141823649406, + 0.006470839027315378, + 0.010686001740396023, + 0.010272786021232605, + 0.006264661438763142, + 0.005151836201548576, + 0.005432340782135725, + 0.004346493165940046, + 0.00513136200606823, + 0.004148241598159075, + 0.008672465570271015, + 0.005540239624679089, + 0.005843132734298706, + 0.0053258417174220085, + 0.006097155157476664, + 0.005634089931845665, + 0.004262091126292944, + 0.0035767315421253443, + 0.006826257798820734, + 0.01607748679816723, + 0.0053679561242461205, + 0.009214990772306919, + 0.015140349045395851, + 0.00637897290289402, + 0.009085746482014656, + 0.01185646466910839, + 0.024226222187280655, + 0.0053890664130449295, + 0.01035365927964449, + 0.004924508277326822, + 0.00803288072347641, + 0.007547149434685707, + 0.007507746107876301, + 0.00677402364090085, + 0.007128347642719746, + 0.016400909051299095, + 0.00928251352161169, + 0.012684316374361515, + 0.008615135215222836, + 0.008292646147310734, + 0.019582083448767662, + 0.008260668255388737, + 0.0249645859003067, + 0.00773450406268239, + 0.008528681471943855, + 0.010964102111756802, + 0.013266662135720253, + 0.00885793473571539, + 0.008372919633984566, + 0.009380838833749294, + 0.008428502827882767, + 0.0034937288146466017, + 0.003921908792108297, + 0.0033818893134593964, + 0.0041566528379917145, + 0.005851410329341888, + 0.0057620401494205, + 0.005848313681781292, + 0.007862293161451817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.12898045778274536 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from directly above the desk and does it orbit in a slow, smooth arc around the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender hand with distinct joints gripping a dark brown bamboo-handled brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sheet of slightly yellowed rice paper lying flat on a solid wooden desk with visible dark grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glossy black brush tip saturated with rich ink moving across the paper, creating thick black ink strokes that seep into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm light slanting from above and to the side, gilding a pristine inkstone and jade paperweight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636883594095707 + ], + "opposite_logprob_tokens": [ + -3.6513688564300537 + ], + "logprob_reward": -0.02636883594095707, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636883594095707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3418917655944824, + 0.15573590993881226, + 0.10199666023254395, + 0.03239183872938156, + 0.014909645542502403, + 0.07641521096229553, + 0.08499188721179962, + 0.03767833113670349, + 0.08714867383241653, + 0.06684011220932007 + ], + "credit_map": [ + 0.035958338528871536, + 0.03366222232580185, + 0.028346683830022812, + 0.04448631405830383, + 0.047768719494342804, + 0.0487939789891243, + 0.05215371400117874, + 0.050721753388643265, + 0.011573589406907558, + 0.011991996318101883, + 0.013522573746740818, + 0.013958376832306385, + 0.032408829778432846, + 0.01893460750579834, + 0.032505374401807785, + 0.020840533077716827, + 0.006995904725044966, + 0.006636919919401407, + 0.008016188628971577, + 0.024464964866638184, + 0.006439106073230505, + 0.015438119880855083, + 0.008482401259243488, + 0.025523051619529724, + 0.005879328586161137, + 0.006097991019487381, + 0.002955774310976267, + 0.003984407987445593, + 0.0031571118161082268, + 0.0032108970917761326, + 0.0023409684654325247, + 0.004765355959534645, + 0.0017339910846203566, + 0.0018805203726515174, + 0.0012922825990244746, + 0.002375449985265732, + 0.0020271725952625275, + 0.0015452340012416244, + 0.0016614767955616117, + 0.002393517177551985, + 0.010026941075921059, + 0.0060704974457621574, + 0.011681975796818733, + 0.00976559054106474, + 0.006984082981944084, + 0.008016417734324932, + 0.01485453825443983, + 0.0090151596814394, + 0.01000604871660471, + 0.011150682345032692, + 0.006750890053808689, + 0.007984304800629616, + 0.006512971129268408, + 0.013469655066728592, + 0.0057429540902376175, + 0.02337437868118286, + 0.004049644339829683, + 0.0017477330984547734, + 0.005912733264267445, + 0.004758286289870739, + 0.004028345923870802, + 0.004017074126750231, + 0.005512596573680639, + 0.007651911582797766, + 0.0072922008112072945, + 0.00802346970885992, + 0.0161945428699255, + 0.015965444967150688, + 0.013916869647800922, + 0.007991816848516464, + 0.008452799171209335, + 0.009311513975262642, + 0.004064876586198807, + 0.0032827239483594894, + 0.005681290756911039, + 0.006290548946708441, + 0.008739393204450607, + 0.013466141186654568, + 0.01016622968018055, + 0.01514890231192112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.02636883594095707 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from directly above the desk and does it orbit in a slow, smooth arc around the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender hand with distinct joints gripping a dark brown bamboo-handled brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sheet of slightly yellowed rice paper lying flat on a solid wooden desk with visible dark grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glossy black brush tip saturated with rich ink moving across the paper, creating thick black ink strokes that seep into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm light slanting from above and to the side, gilding a pristine inkstone and jade paperweight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004098350182175636 + ], + "opposite_logprob_tokens": [ + -5.504098415374756 + ], + "logprob_reward": -0.004098350182175636, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004098350182175636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23949022591114044, + 0.18154844641685486, + 0.07876143604516983, + 0.05256374925374985, + 0.11431713402271271, + 0.10395347326993942, + 0.029444724321365356, + 0.06940586119890213, + 0.056623294949531555, + 0.07389167696237564 + ], + "credit_map": [ + 0.033893901854753494, + 0.031230509281158447, + 0.0283953994512558, + 0.03535467013716698, + 0.022935276851058006, + 0.022137627005577087, + 0.03351948410272598, + 0.03202337399125099, + 0.02224983461201191, + 0.018403954803943634, + 0.024666229262948036, + 0.01879475638270378, + 0.02678005024790764, + 0.0164374727755785, + 0.029006414115428925, + 0.025209732353687286, + 0.010563233867287636, + 0.0059795081615448, + 0.008200788870453835, + 0.011998433619737625, + 0.008753033354878426, + 0.012832270003855228, + 0.007022847421467304, + 0.013411320745944977, + 0.007657908834517002, + 0.006631847005337477, + 0.009379391558468342, + 0.007787626702338457, + 0.004332595504820347, + 0.004473336506634951, + 0.0046876887790858746, + 0.007613353431224823, + 0.012727142311632633, + 0.008408956229686737, + 0.022246092557907104, + 0.015957225114107132, + 0.012883470393717289, + 0.009629970416426659, + 0.013445680029690266, + 0.01901860162615776, + 0.014592169784009457, + 0.0062795537523925304, + 0.008794105611741543, + 0.017899353057146072, + 0.005816355813294649, + 0.008912621065974236, + 0.014043333940207958, + 0.027615981176495552, + 0.006793085020035505, + 0.0040992326103150845, + 0.004185783211141825, + 0.004107670392841101, + 0.002107740845531225, + 0.002435852773487568, + 0.002649996429681778, + 0.003065364435315132, + 0.0069440328516066074, + 0.006262112874537706, + 0.00752669433131814, + 0.00754888541996479, + 0.004811582621186972, + 0.013261182233691216, + 0.005754139740020037, + 0.017297230660915375, + 0.004818646237254143, + 0.0057384660467505455, + 0.007849344052374363, + 0.005545378662645817, + 0.009475159458816051, + 0.008469967171549797, + 0.007233798503875732, + 0.007492538075894117, + 0.005965464282780886, + 0.008652460761368275, + 0.007007674314081669, + 0.006857668049633503, + 0.01662803255021572, + 0.009045413695275784, + 0.01121307909488678, + 0.008521880954504013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.004098350182175636 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from directly above the desk and does it orbit in a slow, smooth arc around the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender hand with distinct joints gripping a dark brown bamboo-handled brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sheet of slightly yellowed rice paper lying flat on a solid wooden desk with visible dark grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glossy black brush tip saturated with rich ink moving across the paper, creating thick black ink strokes that seep into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm light slanting from above and to the side, gilding a pristine inkstone and jade paperweight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735597271472216 + ], + "opposite_logprob_tokens": [ + -5.006735801696777 + ], + "logprob_reward": -0.006735597271472216, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735597271472216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30438628792762756, + 0.22389909625053406, + 0.0609956830739975, + 0.031937647610902786, + 0.03616383299231529, + 0.05834759771823883, + 0.0940595269203186, + 0.024363607168197632, + 0.12271657586097717, + 0.043130144476890564 + ], + "credit_map": [ + 0.03277989476919174, + 0.03987618535757065, + 0.041287776082754135, + 0.05288506671786308, + 0.030317967757582664, + 0.025727644562721252, + 0.03833606094121933, + 0.04317568242549896, + 0.023712847381830215, + 0.025146590545773506, + 0.028407009318470955, + 0.022567782551050186, + 0.034644175320863724, + 0.02130805142223835, + 0.03570869565010071, + 0.03240394592285156, + 0.004387877881526947, + 0.003622121410444379, + 0.004237846005707979, + 0.015593655407428741, + 0.004846295341849327, + 0.008073797449469566, + 0.00411398196592927, + 0.01612011343240738, + 0.005684415344148874, + 0.0055177826434373856, + 0.0043044788762927055, + 0.0033021538984030485, + 0.002591657219454646, + 0.0034868125803768635, + 0.0033106314949691296, + 0.003739712992683053, + 0.0033977816347032785, + 0.004752843640744686, + 0.004169551655650139, + 0.0050127385184168816, + 0.0036570406518876553, + 0.0051957243122160435, + 0.005289190448820591, + 0.004688962362706661, + 0.0057708569802343845, + 0.004133746959269047, + 0.005236517172306776, + 0.009769723750650883, + 0.005237384233623743, + 0.004907328635454178, + 0.013074830174446106, + 0.010217208415269852, + 0.006304677110165358, + 0.015794826671481133, + 0.009331436827778816, + 0.010081141255795956, + 0.008447090163826942, + 0.01456652395427227, + 0.00729265995323658, + 0.022241180762648582, + 0.0027361069805920124, + 0.0018061869777739048, + 0.0034522057976573706, + 0.004301996901631355, + 0.002976116258651018, + 0.0025665436405688524, + 0.0023162548895925283, + 0.004208195488899946, + 0.008595767430961132, + 0.009832894429564476, + 0.02189512923359871, + 0.02533729560673237, + 0.012695079669356346, + 0.01568310707807541, + 0.012872625142335892, + 0.015804659575223923, + 0.004358886741101742, + 0.0038208425976336002, + 0.005494312383234501, + 0.005172736011445522, + 0.006345578469336033, + 0.0055019366554915905, + 0.006534965708851814, + 0.005900886375457048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.006735597271472216 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from directly above the desk and does it orbit in a slow, smooth arc around the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender hand with distinct joints gripping a dark brown bamboo-handled brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sheet of slightly yellowed rice paper lying flat on a solid wooden desk with visible dark grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glossy black brush tip saturated with rich ink moving across the paper, creating thick black ink strokes that seep into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm light slanting from above and to the side, gilding a pristine inkstone and jade paperweight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253088004887104 + ], + "opposite_logprob_tokens": [ + -4.38753080368042 + ], + "logprob_reward": -0.01253088004887104, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253088004887104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2570367455482483, + 0.05038589984178543, + 0.0591292604804039, + 0.03095952421426773, + 0.04930684715509415, + 0.11979838460683823, + 0.08989396691322327, + 0.13798297941684723, + 0.12855303287506104, + 0.07695338129997253 + ], + "credit_map": [ + 0.02721302956342697, + 0.029819821938872337, + 0.01989288069307804, + 0.029424430802464485, + 0.02292863465845585, + 0.021281475201249123, + 0.05579466000199318, + 0.05068179592490196, + 0.004022693727165461, + 0.004962363746017218, + 0.007661149371415377, + 0.00957665964961052, + 0.006261924281716347, + 0.004276982042938471, + 0.006910789757966995, + 0.0067133367992937565, + 0.005034576170146465, + 0.005960067268460989, + 0.00452560605481267, + 0.0076761809177696705, + 0.006699622143059969, + 0.010822799056768417, + 0.007398937828838825, + 0.011011465452611446, + 0.005421267822384834, + 0.0036953496746718884, + 0.0035443799570202827, + 0.0026660633739084005, + 0.0025154498871415854, + 0.0046021114103496075, + 0.0036237938329577446, + 0.0048911054618656635, + 0.0066688875667750835, + 0.005621703807264566, + 0.004347637761384249, + 0.009654129855334759, + 0.005393024068325758, + 0.004529303405433893, + 0.004354109987616539, + 0.008738049305975437, + 0.013662855140864849, + 0.008483527228236198, + 0.010863701812922955, + 0.021522171795368195, + 0.008612791076302528, + 0.010897431522607803, + 0.014194408431649208, + 0.03156149014830589, + 0.016758687794208527, + 0.007748041767627001, + 0.011978574097156525, + 0.010034536942839622, + 0.0066040679812431335, + 0.011505141854286194, + 0.0072365147061645985, + 0.01802840083837509, + 0.020017532631754875, + 0.013908653520047665, + 0.013972301036119461, + 0.009416898712515831, + 0.010951591655611992, + 0.027026880532503128, + 0.00951762218028307, + 0.033171478658914566, + 0.009398444555699825, + 0.011303276754915714, + 0.013399407267570496, + 0.013808428309857845, + 0.02242252230644226, + 0.01900361478328705, + 0.015297945588827133, + 0.02391938678920269, + 0.005513854790478945, + 0.00659363204613328, + 0.011593177914619446, + 0.009148159995675087, + 0.013229358941316605, + 0.007715278305113316, + 0.013025025837123394, + 0.010134892538189888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.01253088004887104 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from directly above the desk, then orbits in an incredibly slow, smooth arc around the center of the writing action, immersing the viewer in the meticulous detail of the composition. A sheet of slightly yellowed rice paper lies flat on a solid wooden desk, its dark grain and subtle signs of age. A slender, powerful hand, defined by distinct joints, grips a dark brown bamboo-handled brush with expert grace. The suspended wrist demonstrates precise control of pressure. A soft, warm light slants down from above and to the side, like afternoon sunlight filtering through a window lattice, gilding the pristine inkstone and jade paperweight with a gentle patina. As the camera pans, we can clearly see the glossy black brush tip, saturated with rich ink, moving effortlessly across the paper. With each stroke, the thick black ink slowly seeps through the fibers of the rice paper, creating a unique texture with a slightly rough edge and a rich, lustrous center. The dominant color palette of the entire scene, composed of deep wood tones, warm beige, and pure black, creates an atmosphere of tranquility, concentration, and a sense of classic charm and Zen.", + "seed": 5246, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_9/index_5302_seed_5246_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.10495947981253266, + "vlm_logprob": -0.10495947981253266, + "vlm_yes_no_margin": 3.549999988079071, + "vlm_token_credit": -0.10495947981253266, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from directly above the desk and does it orbit in a slow, smooth arc around the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender hand with distinct joints gripping a dark brown bamboo-handled brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sheet of slightly yellowed rice paper lying flat on a solid wooden desk with visible dark grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glossy black brush tip saturated with rich ink moving across the paper, creating thick black ink strokes that seep into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm light slanting from above and to the side, gilding a pristine inkstone and jade paperweight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4571576714515686 + ], + "opposite_logprob_tokens": [ + -1.0821576118469238 + ], + "logprob_reward": -0.4571576714515686, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.4571576714515686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3461180329322815, + 0.08033852279186249, + 0.060478657484054565, + 0.05323967710137367, + 0.06194606050848961, + 0.09139744937419891, + 0.1204705685377121, + 0.04477810859680176, + 0.057154905050992966, + 0.08407802134752274 + ], + "credit_map": [ + 0.09228762239217758, + 0.06412665545940399, + 0.02171264961361885, + 0.015612625516951084, + 0.054053667932748795, + 0.021384522318840027, + 0.014088847674429417, + 0.0628514364361763, + 0.010416733101010323, + 0.010294090956449509, + 0.012735781259834766, + 0.011162818409502506, + 0.013001090846955776, + 0.008050479926168919, + 0.0047664949670434, + 0.009911036118865013, + 0.006310383323580027, + 0.006380843929946423, + 0.007670472376048565, + 0.006478995084762573, + 0.010843666270375252, + 0.011326892301440239, + 0.004677671007812023, + 0.006789734587073326, + 0.006578362546861172, + 0.00899265892803669, + 0.008943439461290836, + 0.005539103876799345, + 0.0065960343927145, + 0.0067926859483122826, + 0.003583621233701706, + 0.006213770713657141, + 0.0068299793638288975, + 0.010239757597446442, + 0.008188816718757153, + 0.0055697918869555, + 0.0068689752370119095, + 0.0076231579296290874, + 0.00883693527430296, + 0.007788646966218948, + 0.007755076978355646, + 0.028296202421188354, + 0.007345377467572689, + 0.004711993038654327, + 0.008181216195225716, + 0.0063517275266349316, + 0.0048944964073598385, + 0.023861365392804146, + 0.014512605965137482, + 0.01677655428647995, + 0.01495283842086792, + 0.010137560777366161, + 0.015015983954071999, + 0.013683411292731762, + 0.009010420180857182, + 0.026381203904747963, + 0.006059487350285053, + 0.007842274382710457, + 0.004216084722429514, + 0.004008314572274685, + 0.005727154668420553, + 0.004931014962494373, + 0.004320541396737099, + 0.0076732393354177475, + 0.005097341723740101, + 0.006823030300438404, + 0.007111175451427698, + 0.00912287924438715, + 0.007029881235212088, + 0.0067160590551793575, + 0.006364505738019943, + 0.008890033699572086, + 0.009146258234977722, + 0.007942542433738708, + 0.020195985212922096, + 0.009011348709464073, + 0.009336614981293678, + 0.011139736510813236, + 0.007713199593126774, + 0.009592335671186447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.4571576714515686 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from directly above the desk and does it orbit in a slow, smooth arc around the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender hand with distinct joints gripping a dark brown bamboo-handled brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sheet of slightly yellowed rice paper lying flat on a solid wooden desk with visible dark grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glossy black brush tip saturated with rich ink moving across the paper, creating thick black ink strokes that seep into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm light slanting from above and to the side, gilding a pristine inkstone and jade paperweight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373037278652191 + ], + "opposite_logprob_tokens": [ + -3.4087302684783936 + ], + "logprob_reward": -0.03373037278652191, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373037278652191, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10527196526527405, + 0.13540111482143402, + 0.2519369125366211, + 0.06441842019557953, + 0.07400412857532501, + 0.05749301612377167, + 0.0262710340321064, + 0.02388577349483967, + 0.07446667551994324, + 0.18685093522071838 + ], + "credit_map": [ + 0.01632782444357872, + 0.015350199304521084, + 0.016827236860990524, + 0.010708496905863285, + 0.017761090770363808, + 0.008443434722721577, + 0.0073211947456002235, + 0.012532487511634827, + 0.024514203891158104, + 0.020998019725084305, + 0.01381734013557434, + 0.011274593882262707, + 0.023092765361070633, + 0.01668533682823181, + 0.007131280843168497, + 0.017887569963932037, + 0.02795455791056156, + 0.025538058951497078, + 0.02991262823343277, + 0.02441515401005745, + 0.04670577868819237, + 0.05539335682988167, + 0.020540334284305573, + 0.021477026864886284, + 0.007267416454851627, + 0.006988398265093565, + 0.007617331575602293, + 0.005259225144982338, + 0.012555258348584175, + 0.014587403275072575, + 0.004932540934532881, + 0.005210845731198788, + 0.006382315885275602, + 0.007556835189461708, + 0.010949895717203617, + 0.006681003607809544, + 0.009502091445028782, + 0.015884259715676308, + 0.009004156105220318, + 0.008043569512665272, + 0.006090985145419836, + 0.01363092940300703, + 0.004266111645847559, + 0.004656887147575617, + 0.0045435428619384766, + 0.008304999209940434, + 0.006432566791772842, + 0.009566990658640862, + 0.0028509139083325863, + 0.003310068743303418, + 0.003126955358311534, + 0.0026367823593318462, + 0.003941941540688276, + 0.0029077024664729834, + 0.003048485377803445, + 0.004448182415217161, + 0.002809553174301982, + 0.0032648106571286917, + 0.0027293870225548744, + 0.002265948336571455, + 0.0033987846691161394, + 0.003435808001086116, + 0.0027765845879912376, + 0.0032048970460891724, + 0.010152202099561691, + 0.00853637047111988, + 0.010623161680996418, + 0.0058766682632267475, + 0.008598117157816887, + 0.012241757474839687, + 0.008390954695641994, + 0.010047446005046368, + 0.031819116324186325, + 0.019546737894415855, + 0.04240889102220535, + 0.016279812902212143, + 0.014721546322107315, + 0.019603393971920013, + 0.019040657207369804, + 0.02343078702688217 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 1 + ], + "selected_reward": -0.03373037278652191 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from directly above the desk and does it orbit in a slow, smooth arc around the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender hand with distinct joints gripping a dark brown bamboo-handled brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sheet of slightly yellowed rice paper lying flat on a solid wooden desk with visible dark grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glossy black brush tip saturated with rich ink moving across the paper, creating thick black ink strokes that seep into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm light slanting from above and to the side, gilding a pristine inkstone and jade paperweight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059514218010008335 + ], + "opposite_logprob_tokens": [ + -5.130951404571533 + ], + "logprob_reward": -0.0059514218010008335, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059514218010008335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19678723812103271, + 0.05993668735027313, + 0.14152617752552032, + 0.061283741146326065, + 0.04325460270047188, + 0.1082279235124588, + 0.07341331988573074, + 0.10115250200033188, + 0.058141596615314484, + 0.15627612173557281 + ], + "credit_map": [ + 0.052081815898418427, + 0.013344904407858849, + 0.01653537154197693, + 0.010853450745344162, + 0.013433492742478848, + 0.009647099301218987, + 0.008198822848498821, + 0.07269230484962463, + 0.00808851420879364, + 0.007012457121163607, + 0.009258645586669445, + 0.008594870567321777, + 0.009569469839334488, + 0.005167581140995026, + 0.003467242233455181, + 0.00877790991216898, + 0.01894466020166874, + 0.014315064065158367, + 0.01647220179438591, + 0.013952703215181828, + 0.025861039757728577, + 0.02295348234474659, + 0.011117711663246155, + 0.017909333109855652, + 0.007222508080303669, + 0.00731901079416275, + 0.010323607362806797, + 0.008012724108994007, + 0.00763370655477047, + 0.0076022157445549965, + 0.005978369154036045, + 0.007191607262939215, + 0.005045745987445116, + 0.0052473824471235275, + 0.004909506533294916, + 0.003318896982818842, + 0.005480920895934105, + 0.011652677319943905, + 0.0037227552384138107, + 0.0038767210207879543, + 0.011666090227663517, + 0.03110460564494133, + 0.009892940521240234, + 0.007128665689378977, + 0.009469467215240002, + 0.008444705978035927, + 0.008764533326029778, + 0.02175692655146122, + 0.007565788924694061, + 0.013755393214523792, + 0.010562884621322155, + 0.0041379486210644245, + 0.008434795774519444, + 0.009583148173987865, + 0.0065087624825537205, + 0.01286460179835558, + 0.010216902010142803, + 0.017572471871972084, + 0.012993128038942814, + 0.009468539617955685, + 0.0132887102663517, + 0.010053865611553192, + 0.01137614343315363, + 0.016182750463485718, + 0.007442324422299862, + 0.007599366828799248, + 0.011746558360755444, + 0.006798001937568188, + 0.005719500128179789, + 0.0064176334999501705, + 0.00382379163056612, + 0.008594424463808537, + 0.010809144005179405, + 0.027326101437211037, + 0.0423409566283226, + 0.019517719745635986, + 0.015191592276096344, + 0.013881252147257328, + 0.013848969712853432, + 0.0133604034781456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.0059514218010008335 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from directly above the desk and does it orbit in a slow, smooth arc around the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender hand with distinct joints gripping a dark brown bamboo-handled brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sheet of slightly yellowed rice paper lying flat on a solid wooden desk with visible dark grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glossy black brush tip saturated with rich ink moving across the paper, creating thick black ink strokes that seep into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm light slanting from above and to the side, gilding a pristine inkstone and jade paperweight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780345484614372 + ], + "opposite_logprob_tokens": [ + -4.634780406951904 + ], + "logprob_reward": -0.009780345484614372, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780345484614372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4047674536705017, + 0.051096104085445404, + 0.14797773957252502, + 0.03880013898015022, + 0.14001141488552094, + 0.03914820775389671, + 0.022090258076786995, + 0.034549783915281296, + 0.053402598947286606, + 0.06815627217292786 + ], + "credit_map": [ + 0.09533575177192688, + 0.07771848142147064, + 0.023205287754535675, + 0.024605602025985718, + 0.07758154720067978, + 0.02939458005130291, + 0.021843165159225464, + 0.0550830252468586, + 0.006956997327506542, + 0.009652730077505112, + 0.0064865476451814175, + 0.0056737447157502174, + 0.01025980245321989, + 0.004605178255587816, + 0.0028387822676450014, + 0.004622315987944603, + 0.016897602006793022, + 0.01500152237713337, + 0.017745373770594597, + 0.01432725042104721, + 0.028857525438070297, + 0.02850130945444107, + 0.012512237764894962, + 0.014134909957647324, + 0.0045402152463793755, + 0.00439969589933753, + 0.004884553141891956, + 0.004183229058980942, + 0.006031479220837355, + 0.00617528660222888, + 0.004237789660692215, + 0.004347888752818108, + 0.009277251549065113, + 0.016464928165078163, + 0.013126542791724205, + 0.010729864239692688, + 0.021874863654375076, + 0.03527267277240753, + 0.016364168375730515, + 0.01690111868083477, + 0.003861712059006095, + 0.00532756419852376, + 0.006207786034792662, + 0.004480284638702869, + 0.0037207785062491894, + 0.004806448705494404, + 0.005623537581413984, + 0.005120094865560532, + 0.0026246614288538694, + 0.003988527227193117, + 0.0031782237347215414, + 0.002253966173157096, + 0.002558568026870489, + 0.0022111013531684875, + 0.0017383611993864179, + 0.0035368476528674364, + 0.004096388816833496, + 0.005757065024226904, + 0.0038455426692962646, + 0.003766243578866124, + 0.004625925328582525, + 0.003767719492316246, + 0.0031037386506795883, + 0.00558715732768178, + 0.0058996123261749744, + 0.004745526239275932, + 0.008371857926249504, + 0.009243624284863472, + 0.0044765048660337925, + 0.007623225450515747, + 0.006723688915371895, + 0.00631855521351099, + 0.011872190982103348, + 0.005591670982539654, + 0.021556630730628967, + 0.00469645531848073, + 0.006006176583468914, + 0.00688900426030159, + 0.004317521583288908, + 0.007226622197777033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.009780345484614372 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from directly above the desk and does it orbit in a slow, smooth arc around the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender hand with distinct joints gripping a dark brown bamboo-handled brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sheet of slightly yellowed rice paper lying flat on a solid wooden desk with visible dark grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glossy black brush tip saturated with rich ink moving across the paper, creating thick black ink strokes that seep into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm light slanting from above and to the side, gilding a pristine inkstone and jade paperweight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018177587538957596 + ], + "opposite_logprob_tokens": [ + -4.018177509307861 + ], + "logprob_reward": -0.018177587538957596, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018177587538957596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6346840858459473, + 0.02405600994825363, + 0.049466364085674286, + 0.024487122893333435, + 0.027430346235632896, + 0.029866797849535942, + 0.014903909526765347, + 0.03523289039731026, + 0.023912463337183, + 0.13595999777317047 + ], + "credit_map": [ + 0.12279997020959854, + 0.12993977963924408, + 0.04332491382956505, + 0.043144531548023224, + 0.1420767903327942, + 0.05349652096629143, + 0.04006967693567276, + 0.05983198434114456, + 0.0027842079289257526, + 0.0032731404062360525, + 0.00379489129409194, + 0.002718627452850342, + 0.003764849156141281, + 0.0032188754994422197, + 0.0014004932017996907, + 0.003100926987826824, + 0.005863275844603777, + 0.004906386137008667, + 0.00561437476426363, + 0.00445922277867794, + 0.00915608275681734, + 0.010021000169217587, + 0.004319567698985338, + 0.005126456264406443, + 0.0025546206161379814, + 0.003277192357927561, + 0.003577392315492034, + 0.002971117850393057, + 0.003553971415385604, + 0.003544360399246216, + 0.0019189578015357256, + 0.0030895117670297623, + 0.0024655454326421022, + 0.003236612305045128, + 0.0032421601936221123, + 0.0017368324333801866, + 0.005160450004041195, + 0.005743206944316626, + 0.0029493863694369793, + 0.0028961552307009697, + 0.0037766604218631983, + 0.0027985302731394768, + 0.005199279170483351, + 0.0029315948486328125, + 0.005036606919020414, + 0.0037138196639716625, + 0.002773058135062456, + 0.003637252375483513, + 0.0013466301606968045, + 0.0018218194600194693, + 0.0024186756927520037, + 0.0014535881346091628, + 0.0017081296537071466, + 0.0019153343746438622, + 0.0013842870248481631, + 0.0028554454911500216, + 0.00476487772539258, + 0.005495354533195496, + 0.003940224647521973, + 0.0040304274298250675, + 0.004593017976731062, + 0.0039572948589921, + 0.0030977213755249977, + 0.005353972781449556, + 0.002209954196587205, + 0.003863455494865775, + 0.0025389117654412985, + 0.004093302879482508, + 0.0028265914879739285, + 0.002231918042525649, + 0.0025690775364637375, + 0.003579252865165472, + 0.023186497390270233, + 0.009846538305282593, + 0.04387350007891655, + 0.010239186696708202, + 0.011169501580297947, + 0.013709522783756256, + 0.010244839824736118, + 0.013690410181879997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.018177587538957596 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from directly above the desk, then orbits in an incredibly slow, smooth arc around the center of the writing action, immersing the viewer in the meticulous detail of the composition. A sheet of slightly yellowed rice paper lies flat on a solid wooden desk, its dark grain and subtle signs of age. A slender, powerful hand, defined by distinct joints, grips a dark brown bamboo-handled brush with expert grace. The suspended wrist demonstrates precise control of pressure. A soft, warm light slants down from above and to the side, like afternoon sunlight filtering through a window lattice, gilding the pristine inkstone and jade paperweight with a gentle patina. As the camera pans, we can clearly see the glossy black brush tip, saturated with rich ink, moving effortlessly across the paper. With each stroke, the thick black ink slowly seeps through the fibers of the rice paper, creating a unique texture with a slightly rough edge and a rich, lustrous center. The dominant color palette of the entire scene, composed of deep wood tones, warm beige, and pure black, creates an atmosphere of tranquility, concentration, and a sense of classic charm and Zen.", + "seed": 5247, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_9/index_5302_seed_5247_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.12488238643854857, + "vlm_logprob": -0.12488238643854857, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.12488238643854857, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from directly above the desk and does it orbit in a slow, smooth arc around the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender hand with distinct joints gripping a dark brown bamboo-handled brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sheet of slightly yellowed rice paper lying flat on a solid wooden desk with visible dark grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glossy black brush tip saturated with rich ink moving across the paper, creating thick black ink strokes that seep into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm light slanting from above and to the side, gilding a pristine inkstone and jade paperweight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5516428351402283 + ], + "opposite_logprob_tokens": [ + -0.9266428351402283 + ], + "logprob_reward": -0.5516428351402283, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5516428351402283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2244933843612671, + 0.0731121301651001, + 0.10566184669733047, + 0.1079503521323204, + 0.0962885245680809, + 0.04582194238901138, + 0.08980032056570053, + 0.08055268228054047, + 0.10168351233005524, + 0.07463530451059341 + ], + "credit_map": [ + 0.02865256927907467, + 0.034568797796964645, + 0.02758234739303589, + 0.0368807427585125, + 0.02042134664952755, + 0.03519942983984947, + 0.023272542282938957, + 0.017915591597557068, + 0.01012471504509449, + 0.012135898694396019, + 0.00757271284237504, + 0.005431715399026871, + 0.011323108337819576, + 0.010688993148505688, + 0.007015305105596781, + 0.008819680660963058, + 0.011404148302972317, + 0.014430554583668709, + 0.008323236368596554, + 0.019635872915387154, + 0.0111628957092762, + 0.009262220934033394, + 0.01771536096930504, + 0.01372755691409111, + 0.010268619284033775, + 0.016318997368216515, + 0.012154488824307919, + 0.017053376883268356, + 0.010322648100554943, + 0.012240207754075527, + 0.013449518010020256, + 0.016142498701810837, + 0.007756133563816547, + 0.00926356390118599, + 0.008136237971484661, + 0.009695974178612232, + 0.015469660982489586, + 0.019666561856865883, + 0.010995278134942055, + 0.015305111184716225, + 0.007709713652729988, + 0.005064037162810564, + 0.003381414106115699, + 0.004119728226214647, + 0.007109868340194225, + 0.007236385252326727, + 0.005547515582293272, + 0.005653284024447203, + 0.008676665835082531, + 0.010166733525693417, + 0.0165542084723711, + 0.008692980743944645, + 0.006660868879407644, + 0.00936937890946865, + 0.016648635268211365, + 0.013030845671892166, + 0.010872986167669296, + 0.011306801810860634, + 0.006780445575714111, + 0.007553231902420521, + 0.009544056840240955, + 0.016806337982416153, + 0.0075001223012804985, + 0.010188697837293148, + 0.012791694141924381, + 0.015600073151290417, + 0.011697378940880299, + 0.011579607613384724, + 0.018082600086927414, + 0.009543711319565773, + 0.01261462364345789, + 0.009773822501301765, + 0.00740091735497117, + 0.00845368579030037, + 0.010034937411546707, + 0.014042678289115429, + 0.007218555547297001, + 0.008432422764599323, + 0.00806338433176279, + 0.010988724417984486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.5516428351402283 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from directly above the desk and does it orbit in a slow, smooth arc around the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender hand with distinct joints gripping a dark brown bamboo-handled brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sheet of slightly yellowed rice paper lying flat on a solid wooden desk with visible dark grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glossy black brush tip saturated with rich ink moving across the paper, creating thick black ink strokes that seep into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm light slanting from above and to the side, gilding a pristine inkstone and jade paperweight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018218792974948883 + ], + "opposite_logprob_tokens": [ + -4.018218994140625 + ], + "logprob_reward": -0.018218792974948883, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018218792974948883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16666623950004578, + 0.09498675912618637, + 0.19389966130256653, + 0.056720852851867676, + 0.14508555829524994, + 0.015417058020830154, + 0.09839224815368652, + 0.03490222990512848, + 0.16102202236652374, + 0.0329074002802372 + ], + "credit_map": [ + 0.023824574425816536, + 0.029088271781802177, + 0.03358810022473335, + 0.016205204650759697, + 0.016224026679992676, + 0.016616497188806534, + 0.020465265959501266, + 0.010654287412762642, + 0.010825654491782188, + 0.01716175675392151, + 0.007898291572928429, + 0.00867527723312378, + 0.020498719066381454, + 0.011821003630757332, + 0.00850203912705183, + 0.009604009799659252, + 0.019214527681469917, + 0.024908095598220825, + 0.015296337194740772, + 0.017467601224780083, + 0.04999329522252083, + 0.025820007547736168, + 0.024359747767448425, + 0.016840051859617233, + 0.004134196322411299, + 0.006420964375138283, + 0.00431770458817482, + 0.006800237577408552, + 0.01290725078433752, + 0.007239602506160736, + 0.008830663748085499, + 0.006070232484489679, + 0.020623696967959404, + 0.02895505353808403, + 0.010745841078460217, + 0.014668959192931652, + 0.020654823631048203, + 0.014772322028875351, + 0.013690657913684845, + 0.020974205806851387, + 0.0020033351611346006, + 0.00282181054353714, + 0.0015387868043035269, + 0.0013016669545322657, + 0.002695874311029911, + 0.0018271754961460829, + 0.0013360335724428296, + 0.0018923748284578323, + 0.011493164114654064, + 0.011139451526105404, + 0.018222078680992126, + 0.010568779893219471, + 0.008617441169917583, + 0.007433960679918528, + 0.018664361909031868, + 0.012253009714186192, + 0.0034596396144479513, + 0.005766060668975115, + 0.003250933950766921, + 0.00481119379401207, + 0.004475216381251812, + 0.0051421308889985085, + 0.004435984883457422, + 0.003561068559065461, + 0.013043275102972984, + 0.013745076023042202, + 0.022364458069205284, + 0.03255350887775421, + 0.009565162472426891, + 0.021602706983685493, + 0.02742701582610607, + 0.020720819011330605, + 0.0037563859950751066, + 0.005204459186643362, + 0.003944180440157652, + 0.0045707919634878635, + 0.003650630358606577, + 0.0036968293134123087, + 0.004292286932468414, + 0.0037918377202004194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.018218792974948883 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from directly above the desk and does it orbit in a slow, smooth arc around the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender hand with distinct joints gripping a dark brown bamboo-handled brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sheet of slightly yellowed rice paper lying flat on a solid wooden desk with visible dark grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glossy black brush tip saturated with rich ink moving across the paper, creating thick black ink strokes that seep into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm light slanting from above and to the side, gilding a pristine inkstone and jade paperweight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009784476831555367 + ], + "opposite_logprob_tokens": [ + -4.634784698486328 + ], + "logprob_reward": -0.009784476831555367, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784476831555367, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06928073614835739, + 0.05576283484697342, + 0.029164521023631096, + 0.0859692320227623, + 0.22204266488552094, + 0.061547525227069855, + 0.08749295771121979, + 0.062083251774311066, + 0.28723928332328796, + 0.03941696137189865 + ], + "credit_map": [ + 0.008714298717677593, + 0.010760319419205189, + 0.007297213654965162, + 0.010672520846128464, + 0.0074976407922804356, + 0.012240187264978886, + 0.00648981099948287, + 0.005608751904219389, + 0.005986600648611784, + 0.010830928571522236, + 0.004830614197999239, + 0.004776833578944206, + 0.011347728781402111, + 0.006977209355682135, + 0.005039041396230459, + 0.005973879247903824, + 0.002709391061216593, + 0.002617534715682268, + 0.0028242759872227907, + 0.0035630096681416035, + 0.004421970807015896, + 0.002976710442453623, + 0.005585217848420143, + 0.0044664121232926846, + 0.007739506661891937, + 0.011814162135124207, + 0.006760904099792242, + 0.008787332102656364, + 0.01614142395555973, + 0.015332575887441635, + 0.008617576211690903, + 0.010775760747492313, + 0.031259581446647644, + 0.04510346055030823, + 0.016666198149323463, + 0.021583186462521553, + 0.031491562724113464, + 0.023448292165994644, + 0.021627534180879593, + 0.030862867832183838, + 0.009658333845436573, + 0.010062781162559986, + 0.007374617271125317, + 0.006322729866951704, + 0.008731350302696228, + 0.007161247543990612, + 0.004880934953689575, + 0.007355531211942434, + 0.009261497296392918, + 0.009601124562323093, + 0.018454313278198242, + 0.007969013415277004, + 0.008970078080892563, + 0.0075765978544950485, + 0.017270751297473907, + 0.008389587514102459, + 0.007818188518285751, + 0.007820137776434422, + 0.006131183356046677, + 0.007691938430070877, + 0.009172220714390278, + 0.009792516939342022, + 0.005888633895665407, + 0.007768435403704643, + 0.019170816987752914, + 0.01922614499926567, + 0.04296382889151573, + 0.05694327503442764, + 0.022721875458955765, + 0.04040375351905823, + 0.050739213824272156, + 0.03507037088274956, + 0.004432275425642729, + 0.003471117001026869, + 0.005744477733969688, + 0.007452310528606176, + 0.0030591492541134357, + 0.00482607027515769, + 0.0047141388058662415, + 0.00571742607280612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 6 + ], + "selected_reward": -0.009784476831555367 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from directly above the desk and does it orbit in a slow, smooth arc around the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender hand with distinct joints gripping a dark brown bamboo-handled brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sheet of slightly yellowed rice paper lying flat on a solid wooden desk with visible dark grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glossy black brush tip saturated with rich ink moving across the paper, creating thick black ink strokes that seep into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm light slanting from above and to the side, gilding a pristine inkstone and jade paperweight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075449176132679 + ], + "opposite_logprob_tokens": [ + -4.511075496673584 + ], + "logprob_reward": -0.011075449176132679, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075449176132679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.285074383020401, + 0.10461478680372238, + 0.18245749175548553, + 0.08995091915130615, + 0.033759359270334244, + 0.014150169678032398, + 0.049211110919713974, + 0.04503536969423294, + 0.14623358845710754, + 0.049512796103954315 + ], + "credit_map": [ + 0.03781621903181076, + 0.046435464173555374, + 0.04606820270419121, + 0.04741961881518364, + 0.021827545017004013, + 0.03160472586750984, + 0.03232702985405922, + 0.021575573831796646, + 0.011821132153272629, + 0.018224598839879036, + 0.010894372127950191, + 0.012683609500527382, + 0.01354348473250866, + 0.013677314855158329, + 0.011506186798214912, + 0.012264086864888668, + 0.013955389149487019, + 0.023111874237656593, + 0.01438150368630886, + 0.025730421766638756, + 0.036968763917684555, + 0.021046524867415428, + 0.026586035266518593, + 0.020676972344517708, + 0.007965289056301117, + 0.012652251869440079, + 0.008966472931206226, + 0.011097249574959278, + 0.014805890619754791, + 0.011356833390891552, + 0.011564993299543858, + 0.011541933752596378, + 0.0032954902853816748, + 0.0036461171694099903, + 0.0032409310806542635, + 0.0035961687099188566, + 0.005130083300173283, + 0.005351986736059189, + 0.0043389759957790375, + 0.005159605760127306, + 0.001653126091696322, + 0.0020793171133846045, + 0.0011557340621948242, + 0.00151786336209625, + 0.0022446613293141127, + 0.0017240388551726937, + 0.0015883812448009849, + 0.0021870483178645372, + 0.006510434206575155, + 0.006926459260284901, + 0.005028495565056801, + 0.0029876043554395437, + 0.007119444664567709, + 0.00430632010102272, + 0.00916258990764618, + 0.007169762626290321, + 0.005361515562981367, + 0.005393718369305134, + 0.004400528036057949, + 0.006512005347758532, + 0.004419486504048109, + 0.008163389749825, + 0.005815352313220501, + 0.004969376139342785, + 0.01761164329946041, + 0.018603188917040825, + 0.015789713710546494, + 0.019359122961759567, + 0.023037148639559746, + 0.016667520627379417, + 0.011519983410835266, + 0.023645270615816116, + 0.006547937169671059, + 0.008663815446197987, + 0.006840490270406008, + 0.0056325397454202175, + 0.0038444248493760824, + 0.005566552747040987, + 0.005686329212039709, + 0.006730708293616772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.011075449176132679 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from directly above the desk and does it orbit in a slow, smooth arc around the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender hand with distinct joints gripping a dark brown bamboo-handled brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sheet of slightly yellowed rice paper lying flat on a solid wooden desk with visible dark grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glossy black brush tip saturated with rich ink moving across the paper, creating thick black ink strokes that seep into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm light slanting from above and to the side, gilding a pristine inkstone and jade paperweight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033690378069877625 + ], + "opposite_logprob_tokens": [ + -3.4086904525756836 + ], + "logprob_reward": -0.033690378069877625, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033690378069877625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13967736065387726, + 0.12696442008018494, + 0.06837255507707596, + 0.05577091872692108, + 0.06209868937730789, + 0.05093105509877205, + 0.02411079965531826, + 0.07507342845201492, + 0.3348977565765381, + 0.062103044241666794 + ], + "credit_map": [ + 0.020885609090328217, + 0.013026941567659378, + 0.013082746416330338, + 0.014425980858504772, + 0.014249343425035477, + 0.027369318529963493, + 0.013217571191489697, + 0.023419849574565887, + 0.01602870598435402, + 0.026612458750605583, + 0.01196221448481083, + 0.011672819964587688, + 0.019186651334166527, + 0.018649592995643616, + 0.011616981588304043, + 0.011234995909035206, + 0.006379945203661919, + 0.008653608150780201, + 0.007331783417612314, + 0.006654136348515749, + 0.011317881755530834, + 0.007847562432289124, + 0.011357747949659824, + 0.008829895406961441, + 0.006537007633596659, + 0.008452075533568859, + 0.00519088888540864, + 0.009144051000475883, + 0.005755378399044275, + 0.005365961231291294, + 0.007501532323658466, + 0.007824025116860867, + 0.007955743931233883, + 0.0065920716151595116, + 0.004844125360250473, + 0.005685438401997089, + 0.007808581925928593, + 0.01004822924733162, + 0.00862613320350647, + 0.010538366623222828, + 0.007350907661020756, + 0.006666402332484722, + 0.004181073047220707, + 0.005768654402345419, + 0.009545420296490192, + 0.006069457624107599, + 0.005797929596155882, + 0.0055512068793177605, + 0.0033360961824655533, + 0.002842724323272705, + 0.0038400779012590647, + 0.0024718246422708035, + 0.0033296768087893724, + 0.0027358033694326878, + 0.002292254939675331, + 0.0032623405568301678, + 0.00873891543596983, + 0.008052121847867966, + 0.009044353850185871, + 0.01076466403901577, + 0.008535747416317463, + 0.013022695668041706, + 0.006708371452987194, + 0.010206561535596848, + 0.03820906952023506, + 0.035769253969192505, + 0.038976460695266724, + 0.051832523196935654, + 0.044154092669487, + 0.03022361546754837, + 0.04006607457995415, + 0.05566668510437012, + 0.005808799061924219, + 0.008574984967708588, + 0.007650431711226702, + 0.010701000690460205, + 0.005353153217583895, + 0.00852455385029316, + 0.007256010081619024, + 0.008234106935560703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.033690378069877625 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from directly above the desk, then orbits in an incredibly slow, smooth arc around the center of the writing action, immersing the viewer in the meticulous detail of the composition. A sheet of slightly yellowed rice paper lies flat on a solid wooden desk, its dark grain and subtle signs of age. A slender, powerful hand, defined by distinct joints, grips a dark brown bamboo-handled brush with expert grace. The suspended wrist demonstrates precise control of pressure. A soft, warm light slants down from above and to the side, like afternoon sunlight filtering through a window lattice, gilding the pristine inkstone and jade paperweight with a gentle patina. As the camera pans, we can clearly see the glossy black brush tip, saturated with rich ink, moving effortlessly across the paper. With each stroke, the thick black ink slowly seeps through the fibers of the rice paper, creating a unique texture with a slightly rough edge and a rich, lustrous center. The dominant color palette of the entire scene, composed of deep wood tones, warm beige, and pure black, creates an atmosphere of tranquility, concentration, and a sense of classic charm and Zen.", + "seed": 5248, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_9/index_5302_seed_5248_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04987282063812017, + "vlm_logprob": -0.04987282063812017, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.04987282063812017, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from directly above the desk and does it orbit in a slow, smooth arc around the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender hand with distinct joints gripping a dark brown bamboo-handled brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sheet of slightly yellowed rice paper lying flat on a solid wooden desk with visible dark grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glossy black brush tip saturated with rich ink moving across the paper, creating thick black ink strokes that seep into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm light slanting from above and to the side, gilding a pristine inkstone and jade paperweight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1921190470457077 + ], + "opposite_logprob_tokens": [ + -1.9421190023422241 + ], + "logprob_reward": -0.1921190470457077, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1921190470457077, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05381108820438385, + 0.05686647444963455, + 0.03451428934931755, + 0.5557501316070557, + 0.08189180493354797, + 0.06550829112529755, + 0.08377616107463837, + 0.015060771256685257, + 0.03306838497519493, + 0.01975257322192192 + ], + "credit_map": [ + 0.006341302767395973, + 0.005581938195973635, + 0.010593297891318798, + 0.004868705756962299, + 0.005790626164525747, + 0.006695014424622059, + 0.005802806466817856, + 0.008137394674122334, + 0.004417116288095713, + 0.00665251025930047, + 0.006558687426149845, + 0.009465587325394154, + 0.003620669012889266, + 0.006743593607097864, + 0.0046692960895597935, + 0.014739015139639378, + 0.0028803322929888964, + 0.0069256494753062725, + 0.004194530658423901, + 0.005627464037388563, + 0.0037035311106592417, + 0.003332793014124036, + 0.004349883180111647, + 0.0035001030191779137, + 0.0506153479218483, + 0.10275378823280334, + 0.15555769205093384, + 0.18880432844161987, + 0.01223961915820837, + 0.012383077293634415, + 0.015897022560238838, + 0.017499269917607307, + 0.01049723383039236, + 0.011235097423195839, + 0.008779201656579971, + 0.012380328960716724, + 0.009308699518442154, + 0.011676413007080555, + 0.008582163602113724, + 0.009432668797671795, + 0.0111250514164567, + 0.014527379535138607, + 0.006507826037704945, + 0.008562601171433926, + 0.006360841449350119, + 0.007743167690932751, + 0.005224422551691532, + 0.005457001272588968, + 0.00972865428775549, + 0.009796318598091602, + 0.012230482883751392, + 0.014991255477070808, + 0.007336893118917942, + 0.007437070365995169, + 0.011530936695635319, + 0.010724540799856186, + 0.0017192427767440677, + 0.0017243897309526801, + 0.0017544254660606384, + 0.002493414096534252, + 0.0014209632063284516, + 0.0013157124631106853, + 0.0028305347077548504, + 0.0018020884599536657, + 0.0032474445179104805, + 0.004241567570716143, + 0.005395115353167057, + 0.008226663805544376, + 0.00270175002515316, + 0.002810212317854166, + 0.002858792431652546, + 0.0035868368577212095, + 0.001986769726499915, + 0.0023630158975720406, + 0.0020515986252576113, + 0.0028535244055092335, + 0.002230794634670019, + 0.003276636591181159, + 0.0024795483332127333, + 0.0025106852408498526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 4 + ], + "selected_reward": -0.1921190470457077 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from directly above the desk and does it orbit in a slow, smooth arc around the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender hand with distinct joints gripping a dark brown bamboo-handled brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sheet of slightly yellowed rice paper lying flat on a solid wooden desk with visible dark grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glossy black brush tip saturated with rich ink moving across the paper, creating thick black ink strokes that seep into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm light slanting from above and to the side, gilding a pristine inkstone and jade paperweight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018219847232103348 + ], + "opposite_logprob_tokens": [ + -4.018219947814941 + ], + "logprob_reward": -0.018219847232103348, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018219847232103348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06607420742511749, + 0.04173863306641579, + 0.07626703381538391, + 0.4683733880519867, + 0.04558485746383667, + 0.12350475043058395, + 0.04701501876115799, + 0.054608024656772614, + 0.03332829475402832, + 0.04350581020116806 + ], + "credit_map": [ + 0.01079555694013834, + 0.009303389117121696, + 0.009234488010406494, + 0.005750447977334261, + 0.006791118532419205, + 0.011171781457960606, + 0.006326664704829454, + 0.0067007592879235744, + 0.0038630098570138216, + 0.008024360984563828, + 0.004894191399216652, + 0.006224621552973986, + 0.0039549581706523895, + 0.003653076710179448, + 0.00466116750612855, + 0.006463245488703251, + 0.006960755214095116, + 0.012924506328999996, + 0.010314269922673702, + 0.017830269411206245, + 0.0063904388807713985, + 0.006374897435307503, + 0.006643884815275669, + 0.00882801879197359, + 0.039539873600006104, + 0.09518695622682571, + 0.12069208174943924, + 0.16726118326187134, + 0.009366865269839764, + 0.009474999271333218, + 0.013212078250944614, + 0.013639342971146107, + 0.0056993355974555016, + 0.004820556845515966, + 0.006135218311101198, + 0.004996258299797773, + 0.006587013602256775, + 0.006507356651127338, + 0.003681591246277094, + 0.007157525978982449, + 0.017260132357478142, + 0.02212182804942131, + 0.01635444350540638, + 0.017750591039657593, + 0.016287967562675476, + 0.014129686169326305, + 0.010726451873779297, + 0.008873655460774899, + 0.0061523811891674995, + 0.005299166310578585, + 0.0055305203422904015, + 0.010088277980685234, + 0.005949190817773342, + 0.0049910699017345905, + 0.0046282499097287655, + 0.004376162774860859, + 0.00595663208514452, + 0.004866113420575857, + 0.010349171236157417, + 0.012052527628839016, + 0.00474921939894557, + 0.004409563262015581, + 0.00804049614816904, + 0.00418430520221591, + 0.002636933233588934, + 0.003580923890694976, + 0.00458910595625639, + 0.008565586060285568, + 0.0037008204963058233, + 0.0032721818424761295, + 0.0037821312434971333, + 0.0032006113324314356, + 0.0048270937986671925, + 0.004255364648997784, + 0.006947062909603119, + 0.005724324844777584, + 0.004916675388813019, + 0.008725205436348915, + 0.0035921819508075714, + 0.004517901223152876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.018219847232103348 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from directly above the desk and does it orbit in a slow, smooth arc around the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender hand with distinct joints gripping a dark brown bamboo-handled brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sheet of slightly yellowed rice paper lying flat on a solid wooden desk with visible dark grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glossy black brush tip saturated with rich ink moving across the paper, creating thick black ink strokes that seep into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm light slanting from above and to the side, gilding a pristine inkstone and jade paperweight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006739149335771799 + ], + "opposite_logprob_tokens": [ + -5.006739139556885 + ], + "logprob_reward": -0.006739149335771799, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006739149335771799, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03133383393287659, + 0.03408931940793991, + 0.04967553913593292, + 0.4834635555744171, + 0.06019026041030884, + 0.16401506960391998, + 0.03122749738395214, + 0.07585263252258301, + 0.04249195754528046, + 0.02766038291156292 + ], + "credit_map": [ + 0.0030059022828936577, + 0.0032805060036480427, + 0.007238220889121294, + 0.0027234007138758898, + 0.002929980168119073, + 0.003962030168622732, + 0.004308661911636591, + 0.0038851271383464336, + 0.0031891618855297565, + 0.0037975888699293137, + 0.0037199053913354874, + 0.005504630971699953, + 0.0026704519987106323, + 0.003516836790367961, + 0.0033576558344066143, + 0.008333088830113411, + 0.004908259958028793, + 0.010574204847216606, + 0.0065058874897658825, + 0.009379515424370766, + 0.003946559503674507, + 0.004262564238160849, + 0.005553398746997118, + 0.004545144271105528, + 0.0412132665514946, + 0.09465832263231277, + 0.12771916389465332, + 0.17023533582687378, + 0.010502346791327, + 0.0103650176897645, + 0.014110204763710499, + 0.014659837819635868, + 0.010539003647863865, + 0.00823740754276514, + 0.00892502348870039, + 0.009534833021461964, + 0.005542299710214138, + 0.004382757004350424, + 0.0071956501342356205, + 0.005833280738443136, + 0.03411535173654556, + 0.031942833214998245, + 0.026352079585194588, + 0.026226099580526352, + 0.00905594788491726, + 0.01087898574769497, + 0.013796025887131691, + 0.01164773479104042, + 0.004691394977271557, + 0.004435388371348381, + 0.0042806812562048435, + 0.004648915957659483, + 0.002365343039855361, + 0.0036608148366212845, + 0.0038141643162816763, + 0.00333079369738698, + 0.00978021789342165, + 0.005292132031172514, + 0.022176360711455345, + 0.019735995680093765, + 0.004280559252947569, + 0.004913944751024246, + 0.005330600310117006, + 0.004342811647802591, + 0.003820146666839719, + 0.0059274728409945965, + 0.005303651560097933, + 0.011965659447014332, + 0.0031239662785083055, + 0.003147315699607134, + 0.005587584339082241, + 0.0036161579191684723, + 0.002990873297676444, + 0.0038215036038309336, + 0.003427963238209486, + 0.003946328070014715, + 0.002572998171672225, + 0.0035560333635658026, + 0.0033862018026411533, + 0.003958477638661861 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 7 + ], + "selected_reward": -0.006739149335771799 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from directly above the desk and does it orbit in a slow, smooth arc around the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender hand with distinct joints gripping a dark brown bamboo-handled brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sheet of slightly yellowed rice paper lying flat on a solid wooden desk with visible dark grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glossy black brush tip saturated with rich ink moving across the paper, creating thick black ink strokes that seep into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm light slanting from above and to the side, gilding a pristine inkstone and jade paperweight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059483409859240055 + ], + "opposite_logprob_tokens": [ + -5.130948543548584 + ], + "logprob_reward": -0.0059483409859240055, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059483409859240055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09199468046426773, + 0.06236207112669945, + 0.038166746497154236, + 0.5027970671653748, + 0.05983654782176018, + 0.04299226403236389, + 0.05162132531404495, + 0.0691218227148056, + 0.05565513297915459, + 0.02545238845050335 + ], + "credit_map": [ + 0.009098119102418423, + 0.009489948861300945, + 0.012934553436934948, + 0.009437838569283485, + 0.011821974068880081, + 0.01509871892631054, + 0.013196294195950031, + 0.010917222127318382, + 0.004587357863783836, + 0.004472267348319292, + 0.006260953843593597, + 0.010420639999210835, + 0.0027926077600568533, + 0.006486214697360992, + 0.0045164404436945915, + 0.022825591266155243, + 0.004258195403963327, + 0.004608643241226673, + 0.006560675799846649, + 0.005148164462298155, + 0.0054654888808727264, + 0.0038149948231875896, + 0.004349548369646072, + 0.0039610364474356174, + 0.04402025789022446, + 0.10012120008468628, + 0.13161100447177887, + 0.17689737677574158, + 0.010612670332193375, + 0.010824747383594513, + 0.01451585441827774, + 0.014193934388458729, + 0.008286782540380955, + 0.0077814445830881596, + 0.008038939908146858, + 0.008698948659002781, + 0.0064666359685361385, + 0.007932884618639946, + 0.005552578717470169, + 0.007078331895172596, + 0.007587948348373175, + 0.0065800342708826065, + 0.005586141254752874, + 0.007444276940077543, + 0.003529337467625737, + 0.003327971324324608, + 0.004807556513696909, + 0.004128999076783657, + 0.006812814623117447, + 0.006434678100049496, + 0.006528433412313461, + 0.009459963999688625, + 0.004183120559900999, + 0.004782617557793856, + 0.007441618479788303, + 0.005978078581392765, + 0.008898510597646236, + 0.005239069927483797, + 0.01898282580077648, + 0.015754621475934982, + 0.00459074042737484, + 0.0033121653832495213, + 0.00803419854491949, + 0.004309685900807381, + 0.0057875895872712135, + 0.008608822710812092, + 0.006947910413146019, + 0.011712122708559036, + 0.004594946745783091, + 0.003577935043722391, + 0.009710959158837795, + 0.004714842885732651, + 0.002595119643956423, + 0.003174195298925042, + 0.0025318206753581762, + 0.0034545371308922768, + 0.002306283451616764, + 0.0028384209144860506, + 0.004204674623906612, + 0.004347336944192648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.0059483409859240055 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from directly above the desk and does it orbit in a slow, smooth arc around the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender hand with distinct joints gripping a dark brown bamboo-handled brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sheet of slightly yellowed rice paper lying flat on a solid wooden desk with visible dark grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glossy black brush tip saturated with rich ink moving across the paper, creating thick black ink strokes that seep into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm light slanting from above and to the side, gilding a pristine inkstone and jade paperweight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026337718591094017 + ], + "opposite_logprob_tokens": [ + -3.6513376235961914 + ], + "logprob_reward": -0.026337718591094017, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026337718591094017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08483466506004333, + 0.16135653853416443, + 0.06273748725652695, + 0.08711686730384827, + 0.14415337145328522, + 0.08486390858888626, + 0.09853056818246841, + 0.0648035779595375, + 0.11726245284080505, + 0.09434060752391815 + ], + "credit_map": [ + 0.009035557508468628, + 0.0096035897731781, + 0.018269583582878113, + 0.007907451130449772, + 0.008982758969068527, + 0.009773961268365383, + 0.010095951147377491, + 0.011165807954967022, + 0.014375888742506504, + 0.021056268364191055, + 0.018883250653743744, + 0.035027239471673965, + 0.013473153114318848, + 0.013499373570084572, + 0.01391852367669344, + 0.031122827902436256, + 0.006935264449566603, + 0.006903975736349821, + 0.008695229887962341, + 0.010483154095709324, + 0.007242591120302677, + 0.007349319290369749, + 0.007057541515678167, + 0.008070413023233414, + 0.005517079960554838, + 0.025513457134366035, + 0.007429832126945257, + 0.030927885323762894, + 0.0048595815896987915, + 0.0033304558601230383, + 0.004333105869591236, + 0.00520546967163682, + 0.019899077713489532, + 0.021033307537436485, + 0.01882297545671463, + 0.023728670552372932, + 0.01595689170062542, + 0.014900057576596737, + 0.017218656837940216, + 0.012593727558851242, + 0.009528744965791702, + 0.014590620063245296, + 0.012095934711396694, + 0.012914984486997128, + 0.0076325638219714165, + 0.008369357325136662, + 0.010508304461836815, + 0.009223402477800846, + 0.009306204505264759, + 0.010760498233139515, + 0.015639718621969223, + 0.021140187978744507, + 0.007191600278019905, + 0.009511982090771198, + 0.013770747929811478, + 0.011209633201360703, + 0.006554818712174892, + 0.0053112138994038105, + 0.013181757181882858, + 0.014697386883199215, + 0.006728691048920155, + 0.005252575036138296, + 0.006857735104858875, + 0.006219401489943266, + 0.011770748533308506, + 0.012856743298470974, + 0.013540521264076233, + 0.02047208696603775, + 0.01309466827660799, + 0.01061631366610527, + 0.022700268775224686, + 0.012211100198328495, + 0.0081316027790308, + 0.010299893096089363, + 0.010894560255110264, + 0.014088732190430164, + 0.010120217688381672, + 0.016925664618611336, + 0.011306393891572952, + 0.0125735467299819 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 8 + ], + "selected_reward": -0.026337718591094017 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_0/index_tensor([6887, 6887])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_0/index_tensor([6887, 6887])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fb74b185eaba6d4192e9dd1cccab9bae3e7dbd91 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_0/index_tensor([6887, 6887])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-angle, panoramic shot slowly opens, capturing a quiet classroom. The soft, cool light of an overcast day filters through the tall glass windows, creating distinct, slanting beams of light. Countless tiny particles of dust drift slowly within these beams like suspended stars. The camera then slowly descends and pans forward, finally focusing on a close-up shot of a sign-in table: a slender, fair wrist belonging to a girl with a neatly styled high ponytail. She wears a classic blue and white school uniform, neatly cuffed. Her fingers, steady and composed, grip a dark blue fountain pen. The gleaming metallic tip hovers hesitantly for a moment above a printed class list on the sign-in table. In this frozen moment, a drop of deep black ink drips silently from the pen tip, spreading rapidly and silently across the rough paper, forming an irregular dot. Then, the camera zooms in slowly and precisely, passing over the girl's hand and concentrating entirely on the spot on the list that was stained by the ink spot, clearly revealing the column title \"Review Class 2\" printed in Song typeface. Below the column, apart from the abrupt ink spot, there is still an unsettling emptiness. The whole scene is filled with a silent atmosphere of pressure and heavy choice in the extreme silence.", + "seed": 19194, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_0/index_6887_seed_19194_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1599153894931078, + "vlm_logprob": -0.1599153894931078, + "vlm_yes_no_margin": 2.475, + "vlm_token_credit": -0.1599153894931078, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle panoramic shot of a quiet classroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a slender fair wrist with a high ponytail wearing a blue and white school uniform visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of deep black ink drip from a fountain pen and spread across a printed class list?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in to reveal the column title 'Review Class 2' printed in Song typeface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include slanting beams of light from overcast windows with visible dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31499120593070984 + ], + "opposite_logprob_tokens": [ + -1.3149912357330322 + ], + "logprob_reward": -0.31499120593070984, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31499120593070984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29134470224380493, + 0.06836963444948196, + 0.05604274198412895, + 0.08248390257358551, + 0.09495577961206436, + 0.2912515103816986, + 0.02484966441988945, + 0.040544383227825165, + 0.031424347311258316, + 0.018733292818069458 + ], + "credit_map": [ + 0.02293393202126026, + 0.028157087042927742, + 0.051892660558223724, + 0.054566480219364166, + 0.028983252122998238, + 0.04780840501189232, + 0.0339469313621521, + 0.023055965080857277, + 0.00611333642154932, + 0.007996459491550922, + 0.009807378984987736, + 0.011551865376532078, + 0.005627023056149483, + 0.011913550086319447, + 0.01006818376481533, + 0.005291840527206659, + 0.006365455221384764, + 0.005489791743457317, + 0.009699131362140179, + 0.012691174633800983, + 0.00439234497025609, + 0.005299902055412531, + 0.005797183606773615, + 0.0063077607192099094, + 0.013144399039447308, + 0.01611410826444626, + 0.01058164145797491, + 0.011486111208796501, + 0.011424954049289227, + 0.006665424909442663, + 0.007310515735298395, + 0.005756745580583811, + 0.016008367761969566, + 0.013514665886759758, + 0.009584746323525906, + 0.011927199549973011, + 0.01299777440726757, + 0.010545329190790653, + 0.011589769273996353, + 0.008787926286458969, + 0.06700509041547775, + 0.07723373919725418, + 0.06383361667394638, + 0.05335480347275734, + 0.009803704917430878, + 0.008414695970714092, + 0.006602367386221886, + 0.005003505386412144, + 0.0019413132686167955, + 0.0025906802620738745, + 0.003914922475814819, + 0.0057376716285943985, + 0.002756314817816019, + 0.0029890630394220352, + 0.003263063495978713, + 0.001656635315157473, + 0.0028702609706670046, + 0.0032779518514871597, + 0.00656553078442812, + 0.010339682921767235, + 0.003325561061501503, + 0.005145838484168053, + 0.004863719921559095, + 0.004155836068093777, + 0.0025237982627004385, + 0.002980185439810157, + 0.004043448716402054, + 0.003860976779833436, + 0.00394080113619566, + 0.0033522280864417553, + 0.005280559882521629, + 0.005442348774522543, + 0.0025413488037884235, + 0.002312514465302229, + 0.0026115416549146175, + 0.001941230148077011, + 0.002844255417585373, + 0.0023766730446368456, + 0.0026373493019491434, + 0.001468380563892424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.31499120593070984 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle panoramic shot of a quiet classroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a slender fair wrist with a high ponytail wearing a blue and white school uniform visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of deep black ink drip from a fountain pen and spread across a printed class list?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in to reveal the column title 'Review Class 2' printed in Song typeface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include slanting beams of light from overcast windows with visible dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01609194651246071 + ], + "opposite_logprob_tokens": [ + -4.141091823577881 + ], + "logprob_reward": -0.01609194651246071, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01609194651246071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3325309753417969, + 0.07878054678440094, + 0.04984236881136894, + 0.07457096874713898, + 0.0742448940873146, + 0.20443224906921387, + 0.030050402507185936, + 0.07726962119340897, + 0.041303884238004684, + 0.03697410598397255 + ], + "credit_map": [ + 0.02086472511291504, + 0.03561859205365181, + 0.05191981419920921, + 0.06401985883712769, + 0.023994997143745422, + 0.048609230667352676, + 0.03641725331544876, + 0.05108647793531418, + 0.006980551406741142, + 0.009731259196996689, + 0.011926739476621151, + 0.013531486503779888, + 0.006764143239706755, + 0.009693646803498268, + 0.01014626957476139, + 0.010006442666053772, + 0.0047555947676301, + 0.007102505769580603, + 0.00694945128634572, + 0.013014895841479301, + 0.004175799898803234, + 0.004381726961582899, + 0.005674838088452816, + 0.0037875520065426826, + 0.012336025945842266, + 0.012982425279915333, + 0.010045803152024746, + 0.011643992736935616, + 0.010050738230347633, + 0.005306201055645943, + 0.007107413373887539, + 0.005098361521959305, + 0.012175419367849827, + 0.01033003255724907, + 0.00783485732972622, + 0.00986556988209486, + 0.009586479514837265, + 0.007937896996736526, + 0.009936525486409664, + 0.006578109227120876, + 0.054849158972501755, + 0.05677198991179466, + 0.029651805758476257, + 0.02639998309314251, + 0.010735353454947472, + 0.010926148854196072, + 0.008888445794582367, + 0.006209355313330889, + 0.0028588443528860807, + 0.003248009365051985, + 0.0031782847363501787, + 0.004323457833379507, + 0.00504400534555316, + 0.004511880688369274, + 0.004728709347546101, + 0.002157209673896432, + 0.005798689089715481, + 0.006843660492449999, + 0.0129384221509099, + 0.021165508776903152, + 0.006096166558563709, + 0.006546534597873688, + 0.007073146756738424, + 0.01080748625099659, + 0.002874125959351659, + 0.005238746292889118, + 0.005696588661521673, + 0.005172501318156719, + 0.0049899849109351635, + 0.003342145122587681, + 0.005165447015315294, + 0.008824344724416733, + 0.0029245149344205856, + 0.008443818427622318, + 0.00535598723217845, + 0.0035158416721969843, + 0.004917076323181391, + 0.0028304627630859613, + 0.0032547565642744303, + 0.005731645505875349 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.01609194651246071 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle panoramic shot of a quiet classroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a slender fair wrist with a high ponytail wearing a blue and white school uniform visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of deep black ink drip from a fountain pen and spread across a printed class list?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in to reveal the column title 'Review Class 2' printed in Song typeface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include slanting beams of light from overcast windows with visible dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026365352794528008 + ], + "opposite_logprob_tokens": [ + -3.651365280151367 + ], + "logprob_reward": -0.026365352794528008, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026365352794528008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07283474504947662, + 0.024475563317537308, + 0.014276321977376938, + 0.04150848090648651, + 0.03324553370475769, + 0.4795304834842682, + 0.06600956618785858, + 0.12412828952074051, + 0.07636263221502304, + 0.0676284208893776 + ], + "credit_map": [ + 0.004869695287197828, + 0.007246899418532848, + 0.01265344861894846, + 0.010670882649719715, + 0.0069961813278496265, + 0.016245337203145027, + 0.008033103309571743, + 0.006119198631495237, + 0.002333485521376133, + 0.003543962026014924, + 0.0032007317058742046, + 0.004170915577560663, + 0.0025162806268781424, + 0.0023259322624653578, + 0.003714523045346141, + 0.002669730456545949, + 0.0015884642489254475, + 0.0013313776580616832, + 0.0018245424143970013, + 0.0017447351710870862, + 0.0011444048723205924, + 0.002520746784284711, + 0.001689653261564672, + 0.002432397333905101, + 0.004705445840954781, + 0.004719903226941824, + 0.005141311790794134, + 0.0052336594089865685, + 0.007525100372731686, + 0.004534204490482807, + 0.006239689886569977, + 0.00340916495770216, + 0.005580950528383255, + 0.004611458163708448, + 0.004642216023057699, + 0.004371028859168291, + 0.003065407043322921, + 0.003253307892009616, + 0.004368790425360203, + 0.0033523733727633953, + 0.14036500453948975, + 0.1502348631620407, + 0.08141694217920303, + 0.06842810660600662, + 0.015472612343728542, + 0.009676029905676842, + 0.007763694506138563, + 0.006173212546855211, + 0.006670485716313124, + 0.0069383541122078896, + 0.006816907320171595, + 0.009800195693969727, + 0.009926830418407917, + 0.010355117730796337, + 0.010426666587591171, + 0.00507500721141696, + 0.009298929013311863, + 0.013080203905701637, + 0.015762099996209145, + 0.023381922394037247, + 0.011148081161081791, + 0.016854675486683846, + 0.01963706873357296, + 0.014965308830142021, + 0.005448840092867613, + 0.008997590281069279, + 0.010456616058945656, + 0.009897769428789616, + 0.009937578812241554, + 0.006504268851131201, + 0.010117903351783752, + 0.015002069063484669, + 0.005662782117724419, + 0.013089301064610481, + 0.008287007920444012, + 0.006899422965943813, + 0.013137506321072578, + 0.005636656191200018, + 0.008973193354904652, + 0.005942550487816334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 8 + ], + "selected_reward": -0.026365352794528008 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle panoramic shot of a quiet classroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a slender fair wrist with a high ponytail wearing a blue and white school uniform visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of deep black ink drip from a fountain pen and spread across a printed class list?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in to reveal the column title 'Review Class 2' printed in Song typeface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include slanting beams of light from overcast windows with visible dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38716545701026917 + ], + "opposite_logprob_tokens": [ + -1.1371654272079468 + ], + "logprob_reward": -0.38716545701026917, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38716545701026917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.232478067278862, + 0.05687258392572403, + 0.03051554225385189, + 0.145196333527565, + 0.08045592904090881, + 0.12995100021362305, + 0.03968270123004913, + 0.17235919833183289, + 0.03708551451563835, + 0.07540316134691238 + ], + "credit_map": [ + 0.01380909513682127, + 0.028338346630334854, + 0.04566804692149162, + 0.054417338222265244, + 0.02103388123214245, + 0.023181654512882233, + 0.02975097857415676, + 0.016278721392154694, + 0.005464774090796709, + 0.007444345857948065, + 0.008144247345626354, + 0.01169112604111433, + 0.0056919073686003685, + 0.0063970331102609634, + 0.007256332319229841, + 0.004782817326486111, + 0.003465171903371811, + 0.003553884569555521, + 0.0043896958231925964, + 0.0075729042291641235, + 0.0024320995435118675, + 0.0031199839431792498, + 0.003420452354475856, + 0.0025613480247557163, + 0.018743867054581642, + 0.020987261086702347, + 0.03942568227648735, + 0.01941903494298458, + 0.01724954880774021, + 0.009814871475100517, + 0.011814874596893787, + 0.007741190493106842, + 0.01528919581323862, + 0.012552923522889614, + 0.00820175465196371, + 0.009645760990679264, + 0.009638071060180664, + 0.007467877119779587, + 0.010169127024710178, + 0.007491217460483313, + 0.04229385033249855, + 0.04043211787939072, + 0.01288702804595232, + 0.011334235779941082, + 0.007504353765398264, + 0.007044827099889517, + 0.004329628311097622, + 0.004124951083213091, + 0.004135556519031525, + 0.004240847658365965, + 0.0049818516708910465, + 0.006243548355996609, + 0.003988285548985004, + 0.0053466190584003925, + 0.006518750451505184, + 0.004227243363857269, + 0.011910398490726948, + 0.015173815190792084, + 0.03586515784263611, + 0.059555672109127045, + 0.009572038426995277, + 0.015170389786362648, + 0.01802949793636799, + 0.0070822155103087425, + 0.0028019428718835115, + 0.0045223901979625225, + 0.004732961766421795, + 0.005557460244745016, + 0.0049283504486083984, + 0.003535561030730605, + 0.00649668276309967, + 0.004510165192186832, + 0.008295174688100815, + 0.016301117837429047, + 0.010145245119929314, + 0.0070299431681632996, + 0.013032408431172371, + 0.007757296320050955, + 0.007689939811825752, + 0.005152030382305384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.38716545701026917 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle panoramic shot of a quiet classroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a slender fair wrist with a high ponytail wearing a blue and white school uniform visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of deep black ink drip from a fountain pen and spread across a printed class list?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in to reveal the column title 'Review Class 2' printed in Song typeface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include slanting beams of light from overcast windows with visible dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496298521757126 + ], + "opposite_logprob_tokens": [ + -2.9299628734588623 + ], + "logprob_reward": -0.05496298521757126, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496298521757126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7009889483451843, + 0.03647896274924278, + 0.030969509854912758, + 0.03159094974398613, + 0.02886643260717392, + 0.06119898334145546, + 0.02573307603597641, + 0.059492796659469604, + 0.009472225792706013, + 0.015208090655505657 + ], + "credit_map": [ + 0.056558411568403244, + 0.12952624261379242, + 0.14556793868541718, + 0.16132263839244843, + 0.05887051671743393, + 0.05088028684258461, + 0.06113453581929207, + 0.03712838888168335, + 0.0037380047142505646, + 0.0035998807288706303, + 0.005824585445225239, + 0.007345462217926979, + 0.003241288708522916, + 0.004927273374050856, + 0.004900914151221514, + 0.0029015569016337395, + 0.004316615406423807, + 0.004798996262252331, + 0.004836359992623329, + 0.005306400824338198, + 0.0031367987394332886, + 0.003122401423752308, + 0.0022986906114965677, + 0.003153246594592929, + 0.004357034806162119, + 0.003844277001917362, + 0.006343130022287369, + 0.0034648573491722345, + 0.005057681351900101, + 0.0026181938592344522, + 0.003576946211978793, + 0.0023288296069949865, + 0.006221049465239048, + 0.004679994657635689, + 0.0027822302654385567, + 0.0037559624761343002, + 0.00394328311085701, + 0.002690190216526389, + 0.0025694428477436304, + 0.002224279334768653, + 0.009405503049492836, + 0.00580855505540967, + 0.019464926794171333, + 0.01624886319041252, + 0.0023866635747253895, + 0.0026505799032747746, + 0.0034657849464565516, + 0.0017681051976978779, + 0.0024498954880982637, + 0.0021732323803007603, + 0.0031455736607313156, + 0.004594234749674797, + 0.0034405572805553675, + 0.00379625940695405, + 0.002558514941483736, + 0.0035748097579926252, + 0.004300867207348347, + 0.005893476773053408, + 0.012253480032086372, + 0.020566575229167938, + 0.00315140001475811, + 0.005151793826371431, + 0.005653009284287691, + 0.0025221924297511578, + 0.0009138929308392107, + 0.0009685821132734418, + 0.0013436466688290238, + 0.0011807308765128255, + 0.001033475622534752, + 0.000706942577380687, + 0.0011609134962782264, + 0.002164040692150593, + 0.0012998151360079646, + 0.0032909223809838295, + 0.001979487482458353, + 0.0015807717572897673, + 0.0023606161121279, + 0.001400215202011168, + 0.0020973302889615297, + 0.001198932295665145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.05496298521757126 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, panoramic shot slowly opens, capturing a quiet classroom. The soft, cool light of an overcast day filters through the tall glass windows, creating distinct, slanting beams of light. Countless tiny particles of dust drift slowly within these beams like suspended stars. The camera then slowly descends and pans forward, finally focusing on a close-up shot of a sign-in table: a slender, fair wrist belonging to a girl with a neatly styled high ponytail. She wears a classic blue and white school uniform, neatly cuffed. Her fingers, steady and composed, grip a dark blue fountain pen. The gleaming metallic tip hovers hesitantly for a moment above a printed class list on the sign-in table. In this frozen moment, a drop of deep black ink drips silently from the pen tip, spreading rapidly and silently across the rough paper, forming an irregular dot. Then, the camera zooms in slowly and precisely, passing over the girl's hand and concentrating entirely on the spot on the list that was stained by the ink spot, clearly revealing the column title \"Review Class 2\" printed in Song typeface. Below the column, apart from the abrupt ink spot, there is still an unsettling emptiness. The whole scene is filled with a silent atmosphere of pressure and heavy choice in the extreme silence.", + "seed": 19195, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_0/index_6887_seed_19195_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.3208147704601288, + "vlm_logprob": -0.3208147704601288, + "vlm_yes_no_margin": 1.625, + "vlm_token_credit": -0.3208147704601288, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle panoramic shot of a quiet classroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a slender fair wrist with a high ponytail wearing a blue and white school uniform visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of deep black ink drip from a fountain pen and spread across a printed class list?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in to reveal the column title 'Review Class 2' printed in Song typeface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include slanting beams of light from overcast windows with visible dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9761958122253418 + ], + "opposite_logprob_tokens": [ + -0.4761958122253418 + ], + "logprob_reward": -0.9761958122253418, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9761958122253418, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.158620685338974, + 0.09972331672906876, + 0.06582380831241608, + 0.22551171481609344, + 0.0953851267695427, + 0.05321123078465462, + 0.03726547956466675, + 0.13304974138736725, + 0.03326160088181496, + 0.09814721345901489 + ], + "credit_map": [ + 0.019880948588252068, + 0.01584080420434475, + 0.03261636942625046, + 0.02872253768146038, + 0.013791859149932861, + 0.01265364047139883, + 0.015998631715774536, + 0.01911589875817299, + 0.012668628245592117, + 0.009614861570298672, + 0.01688249595463276, + 0.01954045332968235, + 0.012607705779373646, + 0.008114079013466835, + 0.009188653901219368, + 0.011106443591415882, + 0.006292935460805893, + 0.006532963365316391, + 0.010239161550998688, + 0.008736697025597095, + 0.008634516969323158, + 0.00756294559687376, + 0.009663831442594528, + 0.008160760626196861, + 0.05380656570196152, + 0.023232121020555496, + 0.030667660757899284, + 0.02877696417272091, + 0.018745020031929016, + 0.021701831370592117, + 0.02176091820001602, + 0.026820655912160873, + 0.019109364598989487, + 0.01479890663176775, + 0.013587694615125656, + 0.004977821838110685, + 0.015300099737942219, + 0.014974771998822689, + 0.007575189229100943, + 0.005061275791376829, + 0.006430329289287329, + 0.007141015492379665, + 0.006807464640587568, + 0.004983449354767799, + 0.0069604068994522095, + 0.008700895123183727, + 0.005295034497976303, + 0.006892633158713579, + 0.0048420666716992855, + 0.003233988070860505, + 0.0043149227276444435, + 0.0026597511023283005, + 0.009504730813205242, + 0.0029676214326173067, + 0.00485647190362215, + 0.0048859273083508015, + 0.01175630372017622, + 0.022327709943056107, + 0.007323901634663343, + 0.011931562796235085, + 0.03015534020960331, + 0.013493937440216541, + 0.020481089130043983, + 0.01557991560548544, + 0.002674604533240199, + 0.0031928783282637596, + 0.001965190516784787, + 0.00493635656312108, + 0.005064024124294519, + 0.0036717697512358427, + 0.005684089846909046, + 0.006072687450796366, + 0.005855950526893139, + 0.00375748029910028, + 0.018581349402666092, + 0.015279153361916542, + 0.004966018255800009, + 0.007861660793423653, + 0.004500482231378555, + 0.037345126271247864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.9761958122253418 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle panoramic shot of a quiet classroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a slender fair wrist with a high ponytail wearing a blue and white school uniform visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of deep black ink drip from a fountain pen and spread across a printed class list?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in to reveal the column title 'Review Class 2' printed in Song typeface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include slanting beams of light from overcast windows with visible dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11300133913755417 + ], + "opposite_logprob_tokens": [ + -2.2380013465881348 + ], + "logprob_reward": -0.11300133913755417, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11300133913755417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21028490364551544, + 0.040276847779750824, + 0.06299402564764023, + 0.2229270190000534, + 0.064087875187397, + 0.07341939210891724, + 0.03780565410852432, + 0.09158448129892349, + 0.056826189160346985, + 0.13979360461235046 + ], + "credit_map": [ + 0.02127676270902157, + 0.025358334183692932, + 0.023990977555513382, + 0.03838137164711952, + 0.02546530030667782, + 0.021776119247078896, + 0.023995205760002136, + 0.030040845274925232, + 0.004549652338027954, + 0.004330398049205542, + 0.007792838383466005, + 0.005021105520427227, + 0.00540953641757369, + 0.005111215636134148, + 0.004048849456012249, + 0.004013251978904009, + 0.006376035511493683, + 0.005881665740162134, + 0.008190823718905449, + 0.008772139437496662, + 0.006915130186825991, + 0.00849231332540512, + 0.009243220090866089, + 0.009122696705162525, + 0.06474129855632782, + 0.050908416509628296, + 0.017232079058885574, + 0.01628098636865616, + 0.017606915906071663, + 0.01951996050775051, + 0.017353083938360214, + 0.019284259527921677, + 0.008562617935240269, + 0.007686194963753223, + 0.005097977351397276, + 0.006777327507734299, + 0.011809809133410454, + 0.008245822042226791, + 0.0105246277526021, + 0.005383501295000315, + 0.006994935683906078, + 0.007745883893221617, + 0.015703028067946434, + 0.00405189860612154, + 0.011972184292972088, + 0.01501117181032896, + 0.00679734256118536, + 0.005142952315509319, + 0.00653090700507164, + 0.0032472622115164995, + 0.003880404168739915, + 0.0032923005055636168, + 0.005583154037594795, + 0.0031987447291612625, + 0.0049481941387057304, + 0.007124686613678932, + 0.007522073108702898, + 0.014206086285412312, + 0.004501544404774904, + 0.008966103196144104, + 0.018246082589030266, + 0.009328038431704044, + 0.014331983402371407, + 0.014482574537396431, + 0.0035958783701062202, + 0.006024856586009264, + 0.005289202090352774, + 0.011272676289081573, + 0.006853724364191294, + 0.006060122512280941, + 0.007419375702738762, + 0.01031035091727972, + 0.008429709821939468, + 0.005779442377388477, + 0.0240679569542408, + 0.02030888758599758, + 0.007591963279992342, + 0.011670544743537903, + 0.011885834857821465, + 0.05005926266312599 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.11300133913755417 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle panoramic shot of a quiet classroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a slender fair wrist with a high ponytail wearing a blue and white school uniform visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of deep black ink drip from a fountain pen and spread across a printed class list?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in to reveal the column title 'Review Class 2' printed in Song typeface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include slanting beams of light from overcast windows with visible dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023306824266910553 + ], + "opposite_logprob_tokens": [ + -3.7733068466186523 + ], + "logprob_reward": -0.023306824266910553, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023306824266910553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05733437463641167, + 0.021108146756887436, + 0.020978061482310295, + 0.027644695714116096, + 0.12148357182741165, + 0.02824457362294197, + 0.04809197783470154, + 0.09511450678110123, + 0.10470199584960938, + 0.4752980172634125 + ], + "credit_map": [ + 0.007944226264953613, + 0.006739538162946701, + 0.008729057386517525, + 0.01196027547121048, + 0.005917161237448454, + 0.005143268965184689, + 0.005356650799512863, + 0.00554420193657279, + 0.0025112046860158443, + 0.0025188373401761055, + 0.0016654288629069924, + 0.0029275876004248857, + 0.0040117958560585976, + 0.0030287890695035458, + 0.0024470852222293615, + 0.0019974196329712868, + 0.003439277410507202, + 0.0017714352579787374, + 0.0034596517216414213, + 0.003522532293573022, + 0.0018959969747811556, + 0.0018524135230109096, + 0.002707883482798934, + 0.0023288747761398554, + 0.00469928327947855, + 0.004759496543556452, + 0.0035564564168453217, + 0.002843237482011318, + 0.002914274577051401, + 0.0032819865737110376, + 0.0025156077463179827, + 0.003074354724958539, + 0.01420978270471096, + 0.016660844907164574, + 0.015571868978440762, + 0.009013555012643337, + 0.019693389534950256, + 0.024458525702357292, + 0.010701192542910576, + 0.011174418963491917, + 0.003934144042432308, + 0.005220692604780197, + 0.003282826626673341, + 0.002402318874374032, + 0.004124358296394348, + 0.004007084760814905, + 0.00239379471167922, + 0.0028793560341000557, + 0.008009232580661774, + 0.004589651245623827, + 0.003940807189792395, + 0.0033668926917016506, + 0.0084611801430583, + 0.003925117664039135, + 0.006214034743607044, + 0.009585065767168999, + 0.008181517012417316, + 0.015858149155974388, + 0.005069756880402565, + 0.008627771399915218, + 0.02233061008155346, + 0.009556264616549015, + 0.014724585227668285, + 0.010765857994556427, + 0.0071280961856245995, + 0.011321173049509525, + 0.008158744312822819, + 0.01163153164088726, + 0.019162127748131752, + 0.014209049753844738, + 0.020857438445091248, + 0.012233845889568329, + 0.028246628120541573, + 0.020069550722837448, + 0.07785291969776154, + 0.06747589260339737, + 0.02950974926352501, + 0.03489324450492859, + 0.05321059376001358, + 0.16403944790363312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 8 + ], + "selected_reward": -0.023306824266910553 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle panoramic shot of a quiet classroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a slender fair wrist with a high ponytail wearing a blue and white school uniform visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of deep black ink drip from a fountain pen and spread across a printed class list?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in to reveal the column title 'Review Class 2' printed in Song typeface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include slanting beams of light from overcast windows with visible dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34872353076934814 + ], + "opposite_logprob_tokens": [ + -1.2237235307693481 + ], + "logprob_reward": -0.34872353076934814, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34872353076934814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.148664653301239, + 0.04668745398521423, + 0.044784240424633026, + 0.06918297708034515, + 0.22221964597702026, + 0.06307563185691833, + 0.052517425268888474, + 0.06801386922597885, + 0.09236693382263184, + 0.1924872100353241 + ], + "credit_map": [ + 0.018088918179273605, + 0.018577177077531815, + 0.02202017232775688, + 0.03451266139745712, + 0.01583993248641491, + 0.013022550381720066, + 0.01269183773547411, + 0.013911392539739609, + 0.00637521268799901, + 0.0058380309492349625, + 0.00835490133613348, + 0.0067176357842981815, + 0.006641983054578304, + 0.004284687340259552, + 0.003941740375012159, + 0.00453325966373086, + 0.00716613931581378, + 0.0062764366157352924, + 0.006188707891851664, + 0.007598758675158024, + 0.003987000323832035, + 0.004153309855610132, + 0.005099495407193899, + 0.004314390942454338, + 0.013696519657969475, + 0.01048139575868845, + 0.006600396241992712, + 0.007157076150178909, + 0.006151890382170677, + 0.006575918290764093, + 0.007499296218156815, + 0.011020486243069172, + 0.039566945284605026, + 0.022748121991753578, + 0.03420379385352135, + 0.014779118821024895, + 0.031191837042570114, + 0.05114267021417618, + 0.015070983208715916, + 0.013516169972717762, + 0.006157404743134975, + 0.009352914057672024, + 0.008823327720165253, + 0.004089671187102795, + 0.010327771306037903, + 0.009078406728804111, + 0.007420733571052551, + 0.007825403474271297, + 0.006178803276270628, + 0.005635398905724287, + 0.005337363108992577, + 0.004485550336539745, + 0.010794982314109802, + 0.004889226984232664, + 0.009427056647837162, + 0.005769045557826757, + 0.00727119343355298, + 0.009799763560295105, + 0.005712639540433884, + 0.007142978720366955, + 0.01212113257497549, + 0.008908145129680634, + 0.009563872590661049, + 0.007494141347706318, + 0.005203460808843374, + 0.009491599164903164, + 0.006185336969792843, + 0.015488481149077415, + 0.015280538238584995, + 0.010496766306459904, + 0.014308122918009758, + 0.015912627801299095, + 0.011436574161052704, + 0.007521080318838358, + 0.03587840497493744, + 0.030313195660710335, + 0.011134088970720768, + 0.015233548358082771, + 0.008799569681286812, + 0.07217074185609818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.34872353076934814 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle panoramic shot of a quiet classroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a slender fair wrist with a high ponytail wearing a blue and white school uniform visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of deep black ink drip from a fountain pen and spread across a printed class list?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in to reveal the column title 'Review Class 2' printed in Song typeface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include slanting beams of light from overcast windows with visible dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14284634590148926 + ], + "opposite_logprob_tokens": [ + -2.0178463459014893 + ], + "logprob_reward": -0.14284634590148926, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14284634590148926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17648842930793762, + 0.21807238459587097, + 0.03836679831147194, + 0.05537126213312149, + 0.10491664707660675, + 0.022855164483189583, + 0.025280214846134186, + 0.05475194379687309, + 0.05587291345000267, + 0.24802426993846893 + ], + "credit_map": [ + 0.03302864357829094, + 0.020877070724964142, + 0.026727313175797462, + 0.026827087625861168, + 0.018034763634204865, + 0.01665407419204712, + 0.01576259732246399, + 0.018576862290501595, + 0.03015340305864811, + 0.02220102958381176, + 0.030484717339277267, + 0.042375702410936356, + 0.02670229785144329, + 0.02403920516371727, + 0.018875835463404655, + 0.023240160197019577, + 0.006330391857773066, + 0.0032582057174295187, + 0.005516943987458944, + 0.00626119552180171, + 0.004251725506037474, + 0.0031441834289580584, + 0.0056892214342951775, + 0.003914925269782543, + 0.013504758477210999, + 0.008896450512111187, + 0.009002226404845715, + 0.005820350255817175, + 0.004322501365095377, + 0.004971414804458618, + 0.0036213593557476997, + 0.0052321963012218475, + 0.02066119946539402, + 0.011004704050719738, + 0.016617240384221077, + 0.00840422697365284, + 0.01384790986776352, + 0.020902413874864578, + 0.0077870930545032024, + 0.005691849626600742, + 0.0022044903598725796, + 0.003076068125665188, + 0.0027020389679819345, + 0.0013638560194522142, + 0.005035476293414831, + 0.0036804864648729563, + 0.0027151925023645163, + 0.0020775531884282827, + 0.004623807035386562, + 0.0022858099546283484, + 0.0022176436614245176, + 0.0021468691993504763, + 0.0033719376660883427, + 0.0024424109142273664, + 0.0028742244467139244, + 0.005317509174346924, + 0.004056470934301615, + 0.0075198994018137455, + 0.0029606432653963566, + 0.005367555655539036, + 0.011801991611719131, + 0.0052701993845403194, + 0.011362314224243164, + 0.006412859074771404, + 0.004893335979431868, + 0.008132120594382286, + 0.0052473521791398525, + 0.007501873187720776, + 0.010548206977546215, + 0.0060318103060126305, + 0.006884288042783737, + 0.006633922457695007, + 0.01449801679700613, + 0.009500528685748577, + 0.044075656682252884, + 0.03794601932168007, + 0.014616203494369984, + 0.019903263077139854, + 0.014965745620429516, + 0.09251881390810013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.14284634590148926 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, panoramic shot slowly opens, capturing a quiet classroom. The soft, cool light of an overcast day filters through the tall glass windows, creating distinct, slanting beams of light. Countless tiny particles of dust drift slowly within these beams like suspended stars. The camera then slowly descends and pans forward, finally focusing on a close-up shot of a sign-in table: a slender, fair wrist belonging to a girl with a neatly styled high ponytail. She wears a classic blue and white school uniform, neatly cuffed. Her fingers, steady and composed, grip a dark blue fountain pen. The gleaming metallic tip hovers hesitantly for a moment above a printed class list on the sign-in table. In this frozen moment, a drop of deep black ink drips silently from the pen tip, spreading rapidly and silently across the rough paper, forming an irregular dot. Then, the camera zooms in slowly and precisely, passing over the girl's hand and concentrating entirely on the spot on the list that was stained by the ink spot, clearly revealing the column title \"Review Class 2\" printed in Song typeface. Below the column, apart from the abrupt ink spot, there is still an unsettling emptiness. The whole scene is filled with a silent atmosphere of pressure and heavy choice in the extreme silence.", + "seed": 19196, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_0/index_6887_seed_19196_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.17185299694538117, + "vlm_logprob": -0.17185299694538117, + "vlm_yes_no_margin": 2.225, + "vlm_token_credit": -0.17185299694538117, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle panoramic shot of a quiet classroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a slender fair wrist with a high ponytail wearing a blue and white school uniform visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of deep black ink drip from a fountain pen and spread across a printed class list?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in to reveal the column title 'Review Class 2' printed in Song typeface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include slanting beams of light from overcast windows with visible dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5773918032646179 + ], + "opposite_logprob_tokens": [ + -0.8273918032646179 + ], + "logprob_reward": -0.5773918032646179, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5773918032646179, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5692176818847656, + 0.08124593645334244, + 0.04571215435862541, + 0.1067923903465271, + 0.05816228687763214, + 0.02347901277244091, + 0.05025208741426468, + 0.010074359364807606, + 0.02233692817389965, + 0.032727159559726715 + ], + "credit_map": [ + 0.032058607786893845, + 0.07569941878318787, + 0.027692649513483047, + 0.03376736491918564, + 0.23563212156295776, + 0.03272819519042969, + 0.06987898051738739, + 0.06176033243536949, + 0.00520090339705348, + 0.007481797598302364, + 0.007068898528814316, + 0.013704877346754074, + 0.010580065660178661, + 0.01735813356935978, + 0.010381262749433517, + 0.009469999000430107, + 0.0038214100059121847, + 0.007250898517668247, + 0.0030979029834270477, + 0.004012150689959526, + 0.008851278573274612, + 0.00695960083976388, + 0.004496544133871794, + 0.007222369313240051, + 0.005751943681389093, + 0.03402307257056236, + 0.002906314330175519, + 0.004223143681883812, + 0.01211322657763958, + 0.03764292597770691, + 0.004395776893943548, + 0.005735974758863449, + 0.003269005101174116, + 0.005007558967918158, + 0.007986401207745075, + 0.0054123979061841965, + 0.007676915731281042, + 0.01481285784393549, + 0.004580676555633545, + 0.009416472166776657, + 0.0018118623411282897, + 0.0029871901497244835, + 0.0032587440218776464, + 0.0032007377594709396, + 0.0023289290256798267, + 0.0031703098211437464, + 0.00355025427415967, + 0.003170985961332917, + 0.004013374447822571, + 0.007701132446527481, + 0.0046936627477407455, + 0.006540636532008648, + 0.009788155555725098, + 0.006464255973696709, + 0.005004913080483675, + 0.006045953370630741, + 0.0013323521707206964, + 0.001246752799488604, + 0.0011109593324363232, + 0.0010445454390719533, + 0.0014334310544654727, + 0.0012581305345520377, + 0.0011005917331203818, + 0.0015475961845368147, + 0.002801441354677081, + 0.0032034770119935274, + 0.00291522778570652, + 0.004534283187240362, + 0.002798705827444792, + 0.002112325746566057, + 0.0022124152164906263, + 0.0017590520437806845, + 0.0052435253746807575, + 0.007038231939077377, + 0.005102627910673618, + 0.004716154653578997, + 0.002365284599363804, + 0.003420197870582342, + 0.0018207523971796036, + 0.0030203834176063538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.5773918032646179 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle panoramic shot of a quiet classroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a slender fair wrist with a high ponytail wearing a blue and white school uniform visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of deep black ink drip from a fountain pen and spread across a printed class list?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in to reveal the column title 'Review Class 2' printed in Song typeface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include slanting beams of light from overcast windows with visible dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813677653670311 + ], + "opposite_logprob_tokens": [ + -3.2881367206573486 + ], + "logprob_reward": -0.03813677653670311, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813677653670311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3784828782081604, + 0.0855434238910675, + 0.1139422282576561, + 0.051029376685619354, + 0.1260034739971161, + 0.0744987204670906, + 0.026171430945396423, + 0.04760598763823509, + 0.05494141951203346, + 0.041781023144721985 + ], + "credit_map": [ + 0.02510172501206398, + 0.04689589515328407, + 0.021649353206157684, + 0.025833336636424065, + 0.15218991041183472, + 0.021248726174235344, + 0.046637099236249924, + 0.03892682492733002, + 0.008295899257063866, + 0.0073142098262906075, + 0.010918110609054565, + 0.008916682563722134, + 0.012254623696208, + 0.014124628156423569, + 0.011746100150048733, + 0.011973175220191479, + 0.010712310671806335, + 0.015866879373788834, + 0.013149603269994259, + 0.019146345555782318, + 0.014897530898451805, + 0.015866050496697426, + 0.011272719129920006, + 0.013030791655182838, + 0.006309135816991329, + 0.005883736535906792, + 0.005112969316542149, + 0.00669592060148716, + 0.0070621236227452755, + 0.007528551854193211, + 0.005116958171129227, + 0.007319983560591936, + 0.005945668090134859, + 0.014484046958386898, + 0.018431229516863823, + 0.011449085548520088, + 0.010977299883961678, + 0.03532731533050537, + 0.011270085349678993, + 0.018118740990757942, + 0.0069600907154381275, + 0.008611660450696945, + 0.010782129131257534, + 0.010522134602069855, + 0.006336690858006477, + 0.011136072687804699, + 0.007801748346537352, + 0.012348191812634468, + 0.0025187593419104815, + 0.003982796333730221, + 0.0029302809853106737, + 0.0038094765041023493, + 0.003798936726525426, + 0.003122111549600959, + 0.0024461366701871157, + 0.0035629314370453358, + 0.006897058803588152, + 0.008131314069032669, + 0.00466917222365737, + 0.005906892009079456, + 0.007261110004037619, + 0.005277405027300119, + 0.003995642531663179, + 0.005467393901199102, + 0.009959742426872253, + 0.009371108375489712, + 0.006620980799198151, + 0.010748407803475857, + 0.0052723451517522335, + 0.00485460739582777, + 0.004040649626404047, + 0.0040735783986747265, + 0.005642504431307316, + 0.005732425022870302, + 0.0049695163033902645, + 0.004326080437749624, + 0.006001398432999849, + 0.005509310867637396, + 0.005215140525251627, + 0.004384648986160755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.03813677653670311 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle panoramic shot of a quiet classroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a slender fair wrist with a high ponytail wearing a blue and white school uniform visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of deep black ink drip from a fountain pen and spread across a printed class list?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in to reveal the column title 'Review Class 2' printed in Song typeface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include slanting beams of light from overcast windows with visible dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043081898242235184 + ], + "opposite_logprob_tokens": [ + -3.1680819988250732 + ], + "logprob_reward": -0.043081898242235184, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043081898242235184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06605707854032516, + 0.04382477328181267, + 0.048611950129270554, + 0.32914018630981445, + 0.14596185088157654, + 0.03874998912215233, + 0.05408138036727905, + 0.033343181014060974, + 0.15353135764598846, + 0.08669827878475189 + ], + "credit_map": [ + 0.005627099424600601, + 0.006718432996422052, + 0.007339717820286751, + 0.007708067074418068, + 0.009871107526123524, + 0.00912786927074194, + 0.011204992420971394, + 0.008459790609776974, + 0.004834350198507309, + 0.004827381577342749, + 0.004053815733641386, + 0.004512689542025328, + 0.00737622007727623, + 0.006573826540261507, + 0.006801082752645016, + 0.004845402203500271, + 0.0040085772052407265, + 0.00802112277597189, + 0.005785355810075998, + 0.004862002097070217, + 0.00883172731846571, + 0.006901306100189686, + 0.003703840309754014, + 0.0064980085007846355, + 0.016145069152116776, + 0.10733380168676376, + 0.008309361524879932, + 0.009956826455891132, + 0.027154861018061638, + 0.12771642208099365, + 0.007391368970274925, + 0.025132445618510246, + 0.0061185322701931, + 0.01041923463344574, + 0.00901861023157835, + 0.008860000409185886, + 0.011983800679445267, + 0.049891263246536255, + 0.009081496857106686, + 0.040588900446891785, + 0.003975896630436182, + 0.004701916128396988, + 0.005184556823223829, + 0.006759945768862963, + 0.0038684061728417873, + 0.003630579449236393, + 0.0047508710995316505, + 0.005877816118299961, + 0.003766907611861825, + 0.00674762949347496, + 0.005375135224312544, + 0.005790010094642639, + 0.01326790638267994, + 0.007677704561501741, + 0.005299291107803583, + 0.006156791467219591, + 0.003991106059402227, + 0.0037020156159996986, + 0.004328448791056871, + 0.003448177594691515, + 0.003689743811264634, + 0.004490191116929054, + 0.0034368583001196384, + 0.0062566353008151054, + 0.019870303571224213, + 0.018577365204691887, + 0.015019060112535954, + 0.02660137414932251, + 0.02186724543571472, + 0.01836692914366722, + 0.014195173978805542, + 0.019033895805478096, + 0.006905450951308012, + 0.011463936418294907, + 0.0071273865178227425, + 0.010342011228203773, + 0.012071896344423294, + 0.013452254235744476, + 0.010611984878778458, + 0.014723352156579494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 4 + ], + "selected_reward": -0.043081898242235184 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle panoramic shot of a quiet classroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a slender fair wrist with a high ponytail wearing a blue and white school uniform visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of deep black ink drip from a fountain pen and spread across a printed class list?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in to reveal the column title 'Review Class 2' printed in Song typeface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include slanting beams of light from overcast windows with visible dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033178329467773 + ], + "opposite_logprob_tokens": [ + -2.3503317832946777 + ], + "logprob_reward": -0.10033178329467773, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033178329467773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.266419917345047, + 0.03343454375863075, + 0.053879864513874054, + 0.11886300146579742, + 0.3178093731403351, + 0.018677977845072746, + 0.08962000161409378, + 0.011777552776038647, + 0.06424502283334732, + 0.025272725149989128 + ], + "credit_map": [ + 0.014972200617194176, + 0.03525881841778755, + 0.013385961763560772, + 0.016448751091957092, + 0.10940489917993546, + 0.018146937713027, + 0.030736954882740974, + 0.028065402060747147, + 0.0030886500608175993, + 0.0029640330467373133, + 0.003645295975729823, + 0.002922804094851017, + 0.0061321379616856575, + 0.0061536142602562904, + 0.004533483646810055, + 0.0039945244789123535, + 0.004450376145541668, + 0.008921688422560692, + 0.004342980217188597, + 0.005731048993766308, + 0.0096944784745574, + 0.008096041157841682, + 0.00547247426584363, + 0.007170774042606354, + 0.00612545944750309, + 0.028634805232286453, + 0.004775205161422491, + 0.0061407568864524364, + 0.009067758917808533, + 0.041504234075546265, + 0.0056517343036830425, + 0.0169630516320467, + 0.012554062530398369, + 0.02628461644053459, + 0.02120337449014187, + 0.019246002659201622, + 0.03848608210682869, + 0.10117840766906738, + 0.021885978057980537, + 0.07697084546089172, + 0.001684841699898243, + 0.0028831581585109234, + 0.0023575432132929564, + 0.002196777379140258, + 0.0020290713291615248, + 0.0017887965077534318, + 0.002374029951170087, + 0.003363759024068713, + 0.00722470274195075, + 0.013484152033925056, + 0.008351537398993969, + 0.011086352169513702, + 0.01985270529985428, + 0.012168023735284805, + 0.008657249622046947, + 0.008795278146862984, + 0.0016079990891739726, + 0.0015576892765238881, + 0.0012855518143624067, + 0.0017450505401939154, + 0.001598790637217462, + 0.0014271604595705867, + 0.0012908830540254712, + 0.0012644286034628749, + 0.008547105826437473, + 0.010125004686415195, + 0.008453352376818657, + 0.012043368071317673, + 0.00690843490883708, + 0.0062477909959852695, + 0.00800187699496746, + 0.003918088972568512, + 0.0028779844287782907, + 0.004704597871750593, + 0.003673387225717306, + 0.004383687395602465, + 0.0022457572631537914, + 0.003028683830052614, + 0.0021108039654791355, + 0.0022478222381323576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.10033178329467773 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle panoramic shot of a quiet classroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a slender fair wrist with a high ponytail wearing a blue and white school uniform visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of deep black ink drip from a fountain pen and spread across a printed class list?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in to reveal the column title 'Review Class 2' printed in Song typeface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include slanting beams of light from overcast windows with visible dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032272338867188 + ], + "opposite_logprob_tokens": [ + -2.350322723388672 + ], + "logprob_reward": -0.10032272338867188, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032272338867188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4752115309238434, + 0.08880620449781418, + 0.04223496839404106, + 0.07781634479761124, + 0.056225668638944626, + 0.06257135421037674, + 0.06509054452180862, + 0.029173322021961212, + 0.04715609550476074, + 0.05571394041180611 + ], + "credit_map": [ + 0.025338314473628998, + 0.03096378594636917, + 0.036731887608766556, + 0.05405073240399361, + 0.061931803822517395, + 0.04337603598833084, + 0.12403906136751175, + 0.09877991676330566, + 0.010433285497128963, + 0.011097567155957222, + 0.00901616271585226, + 0.008689073845744133, + 0.014481648802757263, + 0.014088808558881283, + 0.010273476131260395, + 0.010726177133619785, + 0.0037125328090041876, + 0.006722151301801205, + 0.004388086032122374, + 0.0043471879325807095, + 0.008611940778791904, + 0.005948908627033234, + 0.003481765277683735, + 0.005022397264838219, + 0.004965981002897024, + 0.02177571877837181, + 0.004276780411601067, + 0.0026570777408778667, + 0.010655161924660206, + 0.02526896446943283, + 0.0034488884266465902, + 0.0047677746042609215, + 0.0038862924557179213, + 0.006005510687828064, + 0.007217029575258493, + 0.004808986093848944, + 0.010784609243273735, + 0.007409859914332628, + 0.006312283221632242, + 0.009801098145544529, + 0.004843831993639469, + 0.0062775686383247375, + 0.007733880076557398, + 0.0074286386370658875, + 0.006597307510674, + 0.010284394957125187, + 0.0074277776293456554, + 0.01197795383632183, + 0.005003003869205713, + 0.008789421059191227, + 0.006816930137574673, + 0.008712268434464931, + 0.014097025617957115, + 0.00856756791472435, + 0.006557782180607319, + 0.006546540651470423, + 0.003241952508687973, + 0.004911560565233231, + 0.003492902033030987, + 0.003497192170470953, + 0.002973034977912903, + 0.0032650751527398825, + 0.003294197376817465, + 0.004497406538575888, + 0.005459196399897337, + 0.0071829636581242085, + 0.005055645946413279, + 0.008168830536305904, + 0.006302528083324432, + 0.0049823736771941185, + 0.006225671619176865, + 0.003778887679800391, + 0.008962946012616158, + 0.00956693198531866, + 0.007871272042393684, + 0.00710403872653842, + 0.006335854064673185, + 0.0067331730388104916, + 0.004537829197943211, + 0.004601895809173584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.10032272338867188 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, panoramic shot slowly opens, capturing a quiet classroom. The soft, cool light of an overcast day filters through the tall glass windows, creating distinct, slanting beams of light. Countless tiny particles of dust drift slowly within these beams like suspended stars. The camera then slowly descends and pans forward, finally focusing on a close-up shot of a sign-in table: a slender, fair wrist belonging to a girl with a neatly styled high ponytail. She wears a classic blue and white school uniform, neatly cuffed. Her fingers, steady and composed, grip a dark blue fountain pen. The gleaming metallic tip hovers hesitantly for a moment above a printed class list on the sign-in table. In this frozen moment, a drop of deep black ink drips silently from the pen tip, spreading rapidly and silently across the rough paper, forming an irregular dot. Then, the camera zooms in slowly and precisely, passing over the girl's hand and concentrating entirely on the spot on the list that was stained by the ink spot, clearly revealing the column title \"Review Class 2\" printed in Song typeface. Below the column, apart from the abrupt ink spot, there is still an unsettling emptiness. The whole scene is filled with a silent atmosphere of pressure and heavy choice in the extreme silence.", + "seed": 19197, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_0/index_6887_seed_19197_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.26064865551888944, + "vlm_logprob": -0.26064865551888944, + "vlm_yes_no_margin": 1.975, + "vlm_token_credit": -0.26064865551888944, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle panoramic shot of a quiet classroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a slender fair wrist with a high ponytail wearing a blue and white school uniform visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of deep black ink drip from a fountain pen and spread across a printed class list?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in to reveal the column title 'Review Class 2' printed in Song typeface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include slanting beams of light from overcast windows with visible dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8289819955825806 + ], + "opposite_logprob_tokens": [ + -0.5789819955825806 + ], + "logprob_reward": -0.8289819955825806, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8289819955825806, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06611805409193039, + 0.05871085822582245, + 0.05495503917336464, + 0.6033394932746887, + 0.03530534729361534, + 0.0740681067109108, + 0.023220913484692574, + 0.01929408311843872, + 0.02855115383863449, + 0.036436960101127625 + ], + "credit_map": [ + 0.007198469713330269, + 0.006895674858242273, + 0.006787050981074572, + 0.011646037921309471, + 0.008338489569723606, + 0.005871491972357035, + 0.010860753245651722, + 0.008520086295902729, + 0.0060782842338085175, + 0.007748637814074755, + 0.0068241311237216, + 0.007925532758235931, + 0.008019046857953072, + 0.010224866680800915, + 0.006345012225210667, + 0.005545342806726694, + 0.005512862466275692, + 0.006417260970920324, + 0.007417941465973854, + 0.007624772842973471, + 0.007543811574578285, + 0.005490735173225403, + 0.00556684285402298, + 0.00938081368803978, + 0.04073221981525421, + 0.24305856227874756, + 0.23024354875087738, + 0.01911265403032303, + 0.0315248966217041, + 0.014300975017249584, + 0.007683455012738705, + 0.016683222725987434, + 0.004668402951210737, + 0.004823376890271902, + 0.006813202518969774, + 0.00559624657034874, + 0.004133423324674368, + 0.004611884243786335, + 0.003010720480233431, + 0.0016480894992128015, + 0.01883240044116974, + 0.015840742737054825, + 0.011334223672747612, + 0.005893654655665159, + 0.006978688761591911, + 0.008318604901432991, + 0.0038559972308576107, + 0.003013794543221593, + 0.0031371689401566982, + 0.0045438664965331554, + 0.004151373170316219, + 0.0030092194210737944, + 0.0019337435951456428, + 0.0024682357907295227, + 0.0020237236749380827, + 0.0019535813480615616, + 0.0016751295188441873, + 0.0028399713337421417, + 0.0024800258688628674, + 0.0029023608658462763, + 0.002995405113324523, + 0.0022522651124745607, + 0.0023198560811579227, + 0.0018290693406015635, + 0.0013967577833682299, + 0.0018166358349844813, + 0.009068436920642853, + 0.009520558640360832, + 0.0010327784111723304, + 0.0014871563762426376, + 0.0016188205918297172, + 0.002610008930787444, + 0.004478156566619873, + 0.003123970702290535, + 0.0059540970250964165, + 0.006066346541047096, + 0.004427656996995211, + 0.00429571932181716, + 0.004452961031347513, + 0.0036380523815751076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.8289819955825806 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle panoramic shot of a quiet classroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a slender fair wrist with a high ponytail wearing a blue and white school uniform visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of deep black ink drip from a fountain pen and spread across a printed class list?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in to reveal the column title 'Review Class 2' printed in Song typeface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include slanting beams of light from overcast windows with visible dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07007783651351929 + ], + "opposite_logprob_tokens": [ + -2.695077896118164 + ], + "logprob_reward": -0.07007783651351929, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07007783651351929, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06991210579872131, + 0.0267343707382679, + 0.2680374085903168, + 0.405304878950119, + 0.06552892923355103, + 0.10506531596183777, + 0.022815505042672157, + 0.016418172046542168, + 0.009851409122347832, + 0.010331886820495129 + ], + "credit_map": [ + 0.007033476140350103, + 0.00715614203363657, + 0.007560221012681723, + 0.010215501300990582, + 0.007446697447448969, + 0.008201063610613346, + 0.01099352166056633, + 0.011305483989417553, + 0.002458650851622224, + 0.002944722305983305, + 0.0030728986021131277, + 0.0033858702518045902, + 0.0027653563302010298, + 0.004236133303493261, + 0.0048047141171991825, + 0.0030660247430205345, + 0.027537044137716293, + 0.046182047575712204, + 0.025767002254724503, + 0.022264987230300903, + 0.0432451069355011, + 0.04887204244732857, + 0.026835013180971146, + 0.02733416110277176, + 0.026743877679109573, + 0.1608094424009323, + 0.15487751364707947, + 0.01322686392813921, + 0.021259941160678864, + 0.01161766517907381, + 0.005099547561258078, + 0.011670061387121677, + 0.009133419021964073, + 0.010322199203073978, + 0.011899866163730621, + 0.009551608934998512, + 0.006712085101753473, + 0.008687236346304417, + 0.005938155576586723, + 0.0032843537628650665, + 0.02425326220691204, + 0.03251473978161812, + 0.007099711801856756, + 0.008561811409890652, + 0.011111299507319927, + 0.010801194235682487, + 0.00632560346275568, + 0.004397701937705278, + 0.0023302314803004265, + 0.0032521954271942377, + 0.004342833999544382, + 0.003079304238781333, + 0.0023557087406516075, + 0.0027540598530322313, + 0.00230162194930017, + 0.0023995491210371256, + 0.0013294123345986009, + 0.002107031876221299, + 0.0021647431422024965, + 0.00226911180652678, + 0.0023387542460113764, + 0.002297469414770603, + 0.0022249300964176655, + 0.001686719013378024, + 0.0008491419139318168, + 0.001310209627263248, + 0.001957691740244627, + 0.0017018335638567805, + 0.0007876551826484501, + 0.0008212567190639675, + 0.0009393367799930274, + 0.0014842834789305925, + 0.0016741919098421931, + 0.001250992063432932, + 0.0017171658109873533, + 0.001727026654407382, + 0.0009867699118331075, + 0.0009642418590374291, + 0.0008102193824015558, + 0.0012012793449684978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 5 + ], + "selected_reward": -0.07007783651351929 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle panoramic shot of a quiet classroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a slender fair wrist with a high ponytail wearing a blue and white school uniform visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of deep black ink drip from a fountain pen and spread across a printed class list?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in to reveal the column title 'Review Class 2' printed in Song typeface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include slanting beams of light from overcast windows with visible dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020596904680132866 + ], + "opposite_logprob_tokens": [ + -3.895596981048584 + ], + "logprob_reward": -0.020596904680132866, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020596904680132866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02854638174176216, + 0.019720664247870445, + 0.09604054689407349, + 0.5108771324157715, + 0.04615965858101845, + 0.18190953135490417, + 0.04623768478631973, + 0.03915470838546753, + 0.013406053185462952, + 0.017947616055607796 + ], + "credit_map": [ + 0.003673653118312359, + 0.0028880247846245766, + 0.003737070132046938, + 0.003903353586792946, + 0.004035890568047762, + 0.0028144430834800005, + 0.0044165970757603645, + 0.0030773505568504333, + 0.0015662952791899443, + 0.0016572745516896248, + 0.002191552659496665, + 0.0028215658385306597, + 0.002026763977482915, + 0.0022302200086414814, + 0.004773296881467104, + 0.0024536976125091314, + 0.009670683182775974, + 0.015507527627050877, + 0.009129884652793407, + 0.009032022207975388, + 0.012669011950492859, + 0.01609731651842594, + 0.010554848238825798, + 0.013379259034991264, + 0.03441581502556801, + 0.205112025141716, + 0.19576337933540344, + 0.01613459549844265, + 0.02663213200867176, + 0.011891727335751057, + 0.00687337014824152, + 0.01405411958694458, + 0.006042164750397205, + 0.006342561915516853, + 0.008906329981982708, + 0.006882120855152607, + 0.004244626499712467, + 0.00757848983630538, + 0.003785734763368964, + 0.002377629978582263, + 0.03428572788834572, + 0.02380700595676899, + 0.04797257483005524, + 0.01732706092298031, + 0.019977226853370667, + 0.02364220656454563, + 0.00875707808881998, + 0.00614064559340477, + 0.004440529737621546, + 0.00404341472312808, + 0.007773198187351227, + 0.005366689059883356, + 0.007518730591982603, + 0.006246501114219427, + 0.0059595960192382336, + 0.004889029543846846, + 0.002848832169547677, + 0.004038145300000906, + 0.007369034457951784, + 0.0074242716655135155, + 0.00480281189084053, + 0.004436529241502285, + 0.00488991430029273, + 0.0033451730851083994, + 0.0016743717715144157, + 0.001737716724164784, + 0.0027591222897171974, + 0.002165704034268856, + 0.0011931132758036256, + 0.0013184644049033523, + 0.0012988398084416986, + 0.0012587225064635277, + 0.0017104805447161198, + 0.0017571687931194901, + 0.0026546430308371782, + 0.0036892606876790524, + 0.00197238614782691, + 0.001409595599398017, + 0.0024464272428303957, + 0.002307655056938529 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.020596904680132866 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle panoramic shot of a quiet classroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a slender fair wrist with a high ponytail wearing a blue and white school uniform visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of deep black ink drip from a fountain pen and spread across a printed class list?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in to reveal the column title 'Review Class 2' printed in Song typeface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include slanting beams of light from overcast windows with visible dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31356027722358704 + ], + "opposite_logprob_tokens": [ + -1.3135602474212646 + ], + "logprob_reward": -0.31356027722358704, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31356027722358704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.041793931275606155, + 0.03086342103779316, + 0.016132093966007233, + 0.7270243763923645, + 0.015913309529423714, + 0.08598450571298599, + 0.016482725739479065, + 0.02489187940955162, + 0.011250934563577175, + 0.029662800952792168 + ], + "credit_map": [ + 0.004435285925865173, + 0.004232096951454878, + 0.004450818058103323, + 0.005793634802103043, + 0.004466597456485033, + 0.00367751345038414, + 0.008349898271262646, + 0.0063880858942866325, + 0.002114345319569111, + 0.0030364806298166513, + 0.0036492880899459124, + 0.004143578931689262, + 0.003020673058927059, + 0.004195714369416237, + 0.006343852262943983, + 0.004359489772468805, + 0.001512310584075749, + 0.002360022161155939, + 0.0018169202376157045, + 0.0028385366313159466, + 0.002154664369300008, + 0.00215331488288939, + 0.0012701955856755376, + 0.002026129513978958, + 0.04832354187965393, + 0.2903851568698883, + 0.2780824899673462, + 0.024805689230561256, + 0.03770573064684868, + 0.016536755487322807, + 0.010809944942593575, + 0.020375091582536697, + 0.00311192381195724, + 0.0030624333303421736, + 0.0026827033143490553, + 0.0017869746079668403, + 0.0014816594775766134, + 0.002081942046061158, + 0.0009060991578735411, + 0.0007995740161277354, + 0.022144099697470665, + 0.017888514325022697, + 0.014576493762433529, + 0.0067461468279361725, + 0.007977647706866264, + 0.009004535153508186, + 0.0042368569411337376, + 0.003410210832953453, + 0.0018966152565553784, + 0.0020992467179894447, + 0.002246356103569269, + 0.00289694732055068, + 0.002399577759206295, + 0.0015877510886639357, + 0.0018411693163216114, + 0.001515061710961163, + 0.0020908736623823643, + 0.0037033744156360626, + 0.004395739175379276, + 0.0056000701151788235, + 0.0015285876579582691, + 0.0015216640895232558, + 0.002546750009059906, + 0.003504821565002203, + 0.00105296541005373, + 0.001375532359816134, + 0.002411120105534792, + 0.0018437468679621816, + 0.0011013249168172479, + 0.0013099790085107088, + 0.0012119682505726814, + 0.000944297993555665, + 0.00213017244823277, + 0.002566782757639885, + 0.004706790670752525, + 0.00828622654080391, + 0.0019914053846150637, + 0.0019098109332844615, + 0.002910349518060684, + 0.0051612635143101215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.31356027722358704 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle panoramic shot of a quiet classroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a slender fair wrist with a high ponytail wearing a blue and white school uniform visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of deep black ink drip from a fountain pen and spread across a printed class list?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in to reveal the column title 'Review Class 2' printed in Song typeface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include slanting beams of light from overcast windows with visible dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002626359462738 + ], + "opposite_logprob_tokens": [ + -2.695026159286499 + ], + "logprob_reward": -0.07002626359462738, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002626359462738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13915622234344482, + 0.08795078843832016, + 0.06754623353481293, + 0.3983193039894104, + 0.03514949977397919, + 0.18087761104106903, + 0.020048001781105995, + 0.016132382676005363, + 0.04055596888065338, + 0.01426400151103735 + ], + "credit_map": [ + 0.016697904095053673, + 0.01660165749490261, + 0.018086662515997887, + 0.019739802926778793, + 0.02568046562373638, + 0.012811300344765186, + 0.016272548586130142, + 0.013265875168144703, + 0.008642367087304592, + 0.009334617294371128, + 0.013969257473945618, + 0.014609023928642273, + 0.010065743699669838, + 0.014260043390095234, + 0.008049403317272663, + 0.009020327590405941, + 0.006918034050613642, + 0.008829171769320965, + 0.007038748357445002, + 0.005440849345177412, + 0.010793191380798817, + 0.012370730750262737, + 0.007621570024639368, + 0.00853393878787756, + 0.026686234399676323, + 0.15974979102611542, + 0.15297377109527588, + 0.012465904466807842, + 0.020746201276779175, + 0.009547226130962372, + 0.005013033747673035, + 0.011137105524539948, + 0.004788043908774853, + 0.0053307549096643925, + 0.006732547655701637, + 0.004789462313055992, + 0.0033325415570288897, + 0.005290500819683075, + 0.003051741048693657, + 0.0018339066300541162, + 0.05550268292427063, + 0.04753165692090988, + 0.029997628182172775, + 0.013660503551363945, + 0.00991438515484333, + 0.01132624875754118, + 0.006781732197850943, + 0.006162777077406645, + 0.002049217466264963, + 0.0023744958452880383, + 0.003865771694108844, + 0.0028460898902267218, + 0.002711620880290866, + 0.002615142147988081, + 0.0018794244388118386, + 0.0017062394181266427, + 0.001573490910232067, + 0.0024667722173035145, + 0.0022034121211618185, + 0.0028613272588700056, + 0.00216917647048831, + 0.0016286509344354272, + 0.0019860442262142897, + 0.0012435086537152529, + 0.002015857957303524, + 0.0026553012430667877, + 0.014565733261406422, + 0.014982125721871853, + 0.0011396557092666626, + 0.0015994207933545113, + 0.0018392200581729412, + 0.0017586573958396912, + 0.0012998469173908234, + 0.0012960272142663598, + 0.0021901270374655724, + 0.002727378625422716, + 0.0016761142760515213, + 0.0015493647661060095, + 0.0016558516072109342, + 0.0018692909507080913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.07002626359462738 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_1/index_tensor([1100, 1100])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_1/index_tensor([1100, 1100])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..87c862a24e0c9fccae50e81a619c4b21d7203faf --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_1/index_tensor([1100, 1100])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "cinematic macro shot captures the intimate details of a lone candle flame, the camera moving in slow, deliberate orbits around its flickering wick. The wick, a delicate, jagged thread, its tip emitting a deep, fiery red glow, gives rise to a dancing flame that sways and dances in the profound darkness. This teardrop of liquid light displays a mesmerizing gradient, from an ethereal blue at its base to a vibrant, pulsating golden and fiery orange body, culminating in a licking, smoky ashtray. The flame rises from a shimmering, translucent pool of melted wax, its surface trembling, perfectly reflecting its hypnotic movement. As the camera rotates smoothly, it reveals the flame's ever-changing shape, stretching, splitting, and condensing in a silent, graceful rhythm. The velvety black abyss of the background creates a stark, high-contrast composition, imbuing the scene with an atmosphere of tranquility, contemplation, and deep intimacy, one completely focused on a unique, enduring source of warmth and light.", + "seed": 6279, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_1/index_1100_seed_6279_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.12110044821165503, + "vlm_logprob": -0.12110044821165503, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.12110044821165503, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flame exhibit a gradient from ethereal blue at the base to golden and fiery orange at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame rising from a shimmering, translucent pool of melted wax that reflects the flame's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in slow, deliberate orbits around the flickering wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a velvety black abyss creating a high-contrast composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008456612005829811 + ], + "opposite_logprob_tokens": [ + -5.258456707000732 + ], + "logprob_reward": -0.008456612005829811, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.008456612005829811, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1962003856897354, + 0.11222098767757416, + 0.07603707164525986, + 0.18970713019371033, + 0.10814075917005539, + 0.05564836785197258, + 0.06518477946519852, + 0.05034845322370529, + 0.04316871985793114, + 0.10334337502717972 + ], + "credit_map": [ + 0.03817914053797722, + 0.01096375286579132, + 0.028995240107178688, + 0.019107002764940262, + 0.01551091019064188, + 0.03210880234837532, + 0.018401673063635826, + 0.03293386474251747, + 0.016072509810328484, + 0.009441145695745945, + 0.01632303185760975, + 0.016291553154587746, + 0.011415665037930012, + 0.017757555469870567, + 0.012333374470472336, + 0.012586147524416447, + 0.01154006365686655, + 0.007314837537705898, + 0.010770627297461033, + 0.01038745604455471, + 0.008321736007928848, + 0.01114604715257883, + 0.007868614979088306, + 0.008687678724527359, + 0.03450803458690643, + 0.01161473523825407, + 0.02658258192241192, + 0.01973087713122368, + 0.01585525833070278, + 0.028737414628267288, + 0.027624044567346573, + 0.025054175406694412, + 0.01655275747179985, + 0.0060313185676932335, + 0.019682317972183228, + 0.01789063960313797, + 0.007747195661067963, + 0.014601616188883781, + 0.012800944969058037, + 0.012833970598876476, + 0.00855224672704935, + 0.0040856036357581615, + 0.008738130331039429, + 0.009249460883438587, + 0.003580517368391156, + 0.00757727213203907, + 0.008187329396605492, + 0.005677810404449701, + 0.008923822082579136, + 0.004747231490910053, + 0.009333751164376736, + 0.01017829030752182, + 0.0047359708696603775, + 0.009890180081129074, + 0.00786846224218607, + 0.009507064707577229, + 0.0077376640401780605, + 0.0037005150225013494, + 0.009093593806028366, + 0.00657449709251523, + 0.003358107293024659, + 0.006568797864019871, + 0.005153050180524588, + 0.008162226527929306, + 0.006664745509624481, + 0.004354940261691809, + 0.005352274980396032, + 0.0059767113998532295, + 0.003986330237239599, + 0.005879753269255161, + 0.006518394220620394, + 0.004435571376234293, + 0.01675545610487461, + 0.00634640222415328, + 0.015366245061159134, + 0.013926577754318714, + 0.006136239040642977, + 0.015112003311514854, + 0.013857797719538212, + 0.015842648223042488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.008456612005829811 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flame exhibit a gradient from ethereal blue at the base to golden and fiery orange at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame rising from a shimmering, translucent pool of melted wax that reflects the flame's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in slow, deliberate orbits around the flickering wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a velvety black abyss creating a high-contrast composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036270327400416136 + ], + "opposite_logprob_tokens": [ + -5.628626823425293 + ], + "logprob_reward": -0.0036270327400416136, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036270327400416136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1627878099679947, + 0.16205084323883057, + 0.12319577485322952, + 0.06662461161613464, + 0.20087210834026337, + 0.035364780575037, + 0.039933763444423676, + 0.0448765829205513, + 0.06538628786802292, + 0.09890739619731903 + ], + "credit_map": [ + 0.021279212087392807, + 0.011597990989685059, + 0.028167366981506348, + 0.01969410665333271, + 0.014761660248041153, + 0.02494051121175289, + 0.017767557874321938, + 0.02457941323518753, + 0.021414313465356827, + 0.01042922306805849, + 0.030080029740929604, + 0.019076628610491753, + 0.014249669387936592, + 0.02237033098936081, + 0.02053780108690262, + 0.023892845958471298, + 0.014675673097372055, + 0.008619175292551517, + 0.026554860174655914, + 0.013979095034301281, + 0.013879119418561459, + 0.017108561471104622, + 0.01412856113165617, + 0.014250734820961952, + 0.011457529850304127, + 0.0044797309674322605, + 0.009814422577619553, + 0.006755220238119364, + 0.005931716412305832, + 0.009592494927346706, + 0.01002595853060484, + 0.008567539975047112, + 0.030107703059911728, + 0.011178132146596909, + 0.03965403512120247, + 0.032093171030282974, + 0.014623786322772503, + 0.02806144952774048, + 0.020093966275453568, + 0.025059856474399567, + 0.004932753276079893, + 0.004531197715550661, + 0.003232222516089678, + 0.004326697438955307, + 0.004752417095005512, + 0.00469574611634016, + 0.0035571653861552477, + 0.0053365821950137615, + 0.0046490454114973545, + 0.0030990936793386936, + 0.004815910942852497, + 0.005822086241096258, + 0.004867002833634615, + 0.005336417350918055, + 0.007307304535061121, + 0.004036902915686369, + 0.0069871945306658745, + 0.003098329296335578, + 0.007246555760502815, + 0.006455201655626297, + 0.0036581060849130154, + 0.006314611993730068, + 0.005102925002574921, + 0.006013659294694662, + 0.009279731661081314, + 0.0038621029816567898, + 0.00953962653875351, + 0.009441239759325981, + 0.0044967178255319595, + 0.009996810927987099, + 0.008700850419700146, + 0.010069203563034534, + 0.017287172377109528, + 0.004198289010673761, + 0.018689479678869247, + 0.012381670996546745, + 0.004282211884856224, + 0.014609240926802158, + 0.009708290919661522, + 0.017751039937138557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.0036270327400416136 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flame exhibit a gradient from ethereal blue at the base to golden and fiery orange at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame rising from a shimmering, translucent pool of melted wax that reflects the flame's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in slow, deliberate orbits around the flickering wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a velvety black abyss creating a high-contrast composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543711811304092 + ], + "opposite_logprob_tokens": [ + -4.387543678283691 + ], + "logprob_reward": -0.012543711811304092, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543711811304092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3748258352279663, + 0.113460972905159, + 0.0763438418507576, + 0.04449811950325966, + 0.053061287850141525, + 0.048308975994586945, + 0.0392538420855999, + 0.06745944917201996, + 0.09612014144659042, + 0.08666752278804779 + ], + "credit_map": [ + 0.05615992099046707, + 0.02094533108174801, + 0.07720969617366791, + 0.04579531028866768, + 0.024687623605132103, + 0.057545505464076996, + 0.03860192745923996, + 0.05388054624199867, + 0.014339758083224297, + 0.008102666586637497, + 0.02000301517546177, + 0.017321458086371422, + 0.009911750443279743, + 0.01626245118677616, + 0.014214648865163326, + 0.013305223546922207, + 0.009808520786464214, + 0.007285812869668007, + 0.014238476753234863, + 0.012792542576789856, + 0.005241028498858213, + 0.011512965895235538, + 0.007668270729482174, + 0.007796221878379583, + 0.006954864598810673, + 0.0025561086367815733, + 0.006321633234620094, + 0.0054607088677585125, + 0.003330037696287036, + 0.006774425506591797, + 0.007555242162197828, + 0.005545096471905708, + 0.00851225946098566, + 0.003755807876586914, + 0.007289722096174955, + 0.004615047946572304, + 0.006667099893093109, + 0.00786016695201397, + 0.007090077269822359, + 0.007271109614521265, + 0.008127385750412941, + 0.002875660778954625, + 0.0058353436179459095, + 0.006960704457014799, + 0.0031349400524049997, + 0.006416934542357922, + 0.007527654059231281, + 0.007430353667587042, + 0.005411911755800247, + 0.0036892162170261145, + 0.006110399030148983, + 0.005757804494351149, + 0.004233937244862318, + 0.005148818250745535, + 0.004382457118481398, + 0.004519294481724501, + 0.012936500832438469, + 0.00457370188087225, + 0.009690617211163044, + 0.00913767609745264, + 0.003991969861090183, + 0.010943107306957245, + 0.00805134978145361, + 0.008134525269269943, + 0.014746802859008312, + 0.005591439548879862, + 0.012315466068685055, + 0.017293225973844528, + 0.0073948753997683525, + 0.013255571015179157, + 0.014077640138566494, + 0.011445119045674801, + 0.014758476987481117, + 0.0040632630698382854, + 0.01324291992932558, + 0.01157293003052473, + 0.005036576651036739, + 0.01280245278030634, + 0.010661685839295387, + 0.014529213309288025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.012543711811304092 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flame exhibit a gradient from ethereal blue at the base to golden and fiery orange at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame rising from a shimmering, translucent pool of melted wax that reflects the flame's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in slow, deliberate orbits around the flickering wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a velvety black abyss creating a high-contrast composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5762386322021484 + ], + "opposite_logprob_tokens": [ + -0.8262386322021484 + ], + "logprob_reward": -0.5762386322021484, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5762386322021484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2087932676076889, + 0.11155533045530319, + 0.04841127246618271, + 0.06263381242752075, + 0.04468676447868347, + 0.14240999519824982, + 0.05007234215736389, + 0.1354614645242691, + 0.134007528424263, + 0.06196817010641098 + ], + "credit_map": [ + 0.03303896263241768, + 0.015386910177767277, + 0.030468042939901352, + 0.020828835666179657, + 0.017028778791427612, + 0.03638660907745361, + 0.021025091409683228, + 0.034630049020051956, + 0.023080890998244286, + 0.006899578962475061, + 0.015485445968806744, + 0.015341974794864655, + 0.008849876932799816, + 0.01864706538617611, + 0.010972700081765652, + 0.012277806177735329, + 0.0062694912776350975, + 0.00417122570797801, + 0.00725953234359622, + 0.007196324411779642, + 0.006407520733773708, + 0.006652953568845987, + 0.006092997267842293, + 0.004361229483038187, + 0.010231559164822102, + 0.0059354412369430065, + 0.007630267180502415, + 0.006579518783837557, + 0.005815256852656603, + 0.009466161020100117, + 0.007476018276065588, + 0.009499596431851387, + 0.0069840396754443645, + 0.004331371746957302, + 0.004783503711223602, + 0.004355916287750006, + 0.005326925311237574, + 0.006441993173211813, + 0.006018899381160736, + 0.006444116123020649, + 0.01960373856127262, + 0.01970762014389038, + 0.013466578908264637, + 0.020693328231573105, + 0.01809176243841648, + 0.018765754997730255, + 0.014044350013136864, + 0.018036868423223495, + 0.006198663264513016, + 0.003735876642167568, + 0.00637875497341156, + 0.007997114211320877, + 0.004459002520889044, + 0.006891115568578243, + 0.009385813027620316, + 0.0050260028801858425, + 0.01950676180422306, + 0.010144773870706558, + 0.02300555445253849, + 0.01715739071369171, + 0.011024023406207561, + 0.019298946484923363, + 0.013618705794215202, + 0.021705327555537224, + 0.018171003088355064, + 0.006565145216882229, + 0.02020876668393612, + 0.020045923069119453, + 0.01044503878802061, + 0.019373545423150063, + 0.020519226789474487, + 0.018678894266486168, + 0.010021395981311798, + 0.00363189447671175, + 0.009241733700037003, + 0.008272962644696236, + 0.0045439135283231735, + 0.008780786767601967, + 0.007983255200088024, + 0.009492232464253902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.5762386322021484 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flame exhibit a gradient from ethereal blue at the base to golden and fiery orange at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame rising from a shimmering, translucent pool of melted wax that reflects the flame's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in slow, deliberate orbits around the flickering wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a velvety black abyss creating a high-contrast composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004636252298951149 + ], + "opposite_logprob_tokens": [ + -5.379636287689209 + ], + "logprob_reward": -0.004636252298951149, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004636252298951149, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13681660592556, + 0.0901113897562027, + 0.1493453085422516, + 0.21675890684127808, + 0.07698439061641693, + 0.08328647166490555, + 0.0387914665043354, + 0.041508749127388, + 0.04153665527701378, + 0.12486005574464798 + ], + "credit_map": [ + 0.021262696012854576, + 0.009419321082532406, + 0.022012535482645035, + 0.017486320808529854, + 0.01129189319908619, + 0.020805776119232178, + 0.012687597423791885, + 0.021850474178791046, + 0.013406841084361076, + 0.006999843288213015, + 0.013834373094141483, + 0.011175928637385368, + 0.008020185865461826, + 0.014430209994316101, + 0.012204968370497227, + 0.010039038956165314, + 0.0239022858440876, + 0.008594238199293613, + 0.024183450266718864, + 0.019516797736287117, + 0.010912997648119926, + 0.024661343544721603, + 0.01600544899702072, + 0.02156873047351837, + 0.043624769896268845, + 0.012954127974808216, + 0.03052782081067562, + 0.02236337400972843, + 0.01625167764723301, + 0.034893058240413666, + 0.024443989619612694, + 0.03170008212327957, + 0.016729535534977913, + 0.0049474057741463184, + 0.009246925823390484, + 0.0073701986111700535, + 0.006948899477720261, + 0.012668692506849766, + 0.007747815456241369, + 0.011324910447001457, + 0.014972773380577564, + 0.007131948135793209, + 0.009656141512095928, + 0.011011584661900997, + 0.004587305709719658, + 0.011740220710635185, + 0.008416054770350456, + 0.0157704446464777, + 0.005044200923293829, + 0.0033488492481410503, + 0.004183553624898195, + 0.006990816444158554, + 0.0034437759313732386, + 0.005546733271330595, + 0.006221420597285032, + 0.004012116231024265, + 0.005247914232313633, + 0.0030327937565743923, + 0.007007780484855175, + 0.006546507589519024, + 0.003590004751458764, + 0.0050713368691504, + 0.00486413761973381, + 0.006148276384919882, + 0.006309273187071085, + 0.0026253596879541874, + 0.005619809031486511, + 0.005738947540521622, + 0.0032706409692764282, + 0.00592917762696743, + 0.005921271629631519, + 0.006122176069766283, + 0.02123960107564926, + 0.005646600853651762, + 0.024000180885195732, + 0.014883174560964108, + 0.006294650491327047, + 0.01759522408246994, + 0.013036726973950863, + 0.022163905203342438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.004636252298951149 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic macro shot captures the intimate details of a lone candle flame, the camera moving in slow, deliberate orbits around its flickering wick. The wick, a delicate, jagged thread, its tip emitting a deep, fiery red glow, gives rise to a dancing flame that sways and dances in the profound darkness. This teardrop of liquid light displays a mesmerizing gradient, from an ethereal blue at its base to a vibrant, pulsating golden and fiery orange body, culminating in a licking, smoky ashtray. The flame rises from a shimmering, translucent pool of melted wax, its surface trembling, perfectly reflecting its hypnotic movement. As the camera rotates smoothly, it reveals the flame's ever-changing shape, stretching, splitting, and condensing in a silent, graceful rhythm. The velvety black abyss of the background creates a stark, high-contrast composition, imbuing the scene with an atmosphere of tranquility, contemplation, and deep intimacy, one completely focused on a unique, enduring source of warmth and light.", + "seed": 6280, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_1/index_1100_seed_6280_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.164414831250906, + "vlm_logprob": -0.164414831250906, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.164414831250906, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flame exhibit a gradient from ethereal blue at the base to golden and fiery orange at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame rising from a shimmering, translucent pool of melted wax that reflects the flame's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in slow, deliberate orbits around the flickering wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a velvety black abyss creating a high-contrast composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009357757866382599 + ], + "opposite_logprob_tokens": [ + -5.259357929229736 + ], + "logprob_reward": -0.009357757866382599, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.009357757866382599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20468279719352722, + 0.1506793051958084, + 0.1536846160888672, + 0.06653565168380737, + 0.09359906613826752, + 0.12179109454154968, + 0.05589696764945984, + 0.08035989850759506, + 0.04787956178188324, + 0.02489101141691208 + ], + "credit_map": [ + 0.03221726417541504, + 0.015500832349061966, + 0.03319135680794716, + 0.018232814967632294, + 0.023668313398957253, + 0.03903454169631004, + 0.016660969704389572, + 0.026176689192652702, + 0.01763046719133854, + 0.015208043158054352, + 0.019866053014993668, + 0.012867432087659836, + 0.022602533921599388, + 0.020799564197659492, + 0.026903975754976273, + 0.014801232144236565, + 0.02257562056183815, + 0.010995808057487011, + 0.015661904588341713, + 0.014647973701357841, + 0.011110128834843636, + 0.01841670460999012, + 0.03615381196141243, + 0.02412266656756401, + 0.010167130269110203, + 0.008264041505753994, + 0.011010332964360714, + 0.006934151519089937, + 0.005627237260341644, + 0.006071900017559528, + 0.011117221787571907, + 0.007343638688325882, + 0.010541914962232113, + 0.014607018791139126, + 0.013888617046177387, + 0.014974168501794338, + 0.005388673860579729, + 0.007706325966864824, + 0.013894434086978436, + 0.012597909197211266, + 0.014487101696431637, + 0.022262368351221085, + 0.014157394878566265, + 0.016240892931818962, + 0.006259174086153507, + 0.011451410129666328, + 0.01815761625766754, + 0.018775129690766335, + 0.006746787112206221, + 0.008472509682178497, + 0.007331730332225561, + 0.007764998357743025, + 0.0027976217679679394, + 0.006422958802431822, + 0.008190004155039787, + 0.008170357905328274, + 0.012472466565668583, + 0.011815398000180721, + 0.010724938474595547, + 0.008759009651839733, + 0.0048256549052894115, + 0.009330594912171364, + 0.007886860519647598, + 0.014544974081218243, + 0.0045222328044474125, + 0.00819430686533451, + 0.005334001965820789, + 0.005345458164811134, + 0.003233131021261215, + 0.007270071655511856, + 0.0072007738053798676, + 0.0067795878276228905, + 0.002276203827932477, + 0.0031999205239117146, + 0.002324283355847001, + 0.003907316364347935, + 0.0015584996435791254, + 0.0038607248570770025, + 0.003541340818628669, + 0.004222721792757511 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.009357757866382599 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flame exhibit a gradient from ethereal blue at the base to golden and fiery orange at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame rising from a shimmering, translucent pool of melted wax that reflects the flame's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in slow, deliberate orbits around the flickering wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a velvety black abyss creating a high-contrast composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016082093119621277 + ], + "opposite_logprob_tokens": [ + -4.141082286834717 + ], + "logprob_reward": -0.016082093119621277, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016082093119621277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11356513947248459, + 0.08625864237546921, + 0.08827570825815201, + 0.10416042804718018, + 0.15480776131153107, + 0.08679374307394028, + 0.05780382081866264, + 0.23001910746097565, + 0.03746591880917549, + 0.04084978997707367 + ], + "credit_map": [ + 0.010270900093019009, + 0.008246204815804958, + 0.01973794400691986, + 0.009287661872804165, + 0.011201191693544388, + 0.016301359981298447, + 0.01158435270190239, + 0.026935521513223648, + 0.00999048538506031, + 0.007224213797599077, + 0.014618964865803719, + 0.010634034872055054, + 0.008839195594191551, + 0.012123089283704758, + 0.01343765389174223, + 0.009391007013618946, + 0.014198966324329376, + 0.0048380158841609955, + 0.00790748093277216, + 0.008519001305103302, + 0.004619456361979246, + 0.01682942919433117, + 0.012647071853280067, + 0.01871628686785698, + 0.017291409894824028, + 0.00982520543038845, + 0.011018217541277409, + 0.010652709752321243, + 0.008800414390861988, + 0.013412166386842728, + 0.022443797439336777, + 0.010716511867940426, + 0.01741853356361389, + 0.02566644363105297, + 0.016419529914855957, + 0.017163263633847237, + 0.011383689008653164, + 0.015966130420565605, + 0.0263315849006176, + 0.02445857599377632, + 0.009513038210570812, + 0.014502284117043018, + 0.009857747703790665, + 0.014619613066315651, + 0.006567318923771381, + 0.007533669471740723, + 0.012727531604468822, + 0.011472536250948906, + 0.008261323906481266, + 0.009295161813497543, + 0.007917859591543674, + 0.006122722756117582, + 0.004032776691019535, + 0.009615628980100155, + 0.006864883471280336, + 0.005693465005606413, + 0.02963593602180481, + 0.028244365006685257, + 0.03100457414984703, + 0.02015349455177784, + 0.012622685171663761, + 0.022804193198680878, + 0.029921414330601692, + 0.055632442235946655, + 0.004235847387462854, + 0.006042745895683765, + 0.006264624651521444, + 0.0051042912527918816, + 0.001621981500647962, + 0.005647832993417978, + 0.0036788340657949448, + 0.004869760479778051, + 0.003138320054858923, + 0.0071584852412343025, + 0.004098530393093824, + 0.005364687647670507, + 0.0015851818025112152, + 0.00762298796325922, + 0.004396394360810518, + 0.007485200650990009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.016082093119621277 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flame exhibit a gradient from ethereal blue at the base to golden and fiery orange at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame rising from a shimmering, translucent pool of melted wax that reflects the flame's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in slow, deliberate orbits around the flickering wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a velvety black abyss creating a high-contrast composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018199946731328964 + ], + "opposite_logprob_tokens": [ + -4.018199920654297 + ], + "logprob_reward": -0.018199946731328964, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018199946731328964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22011999785900116, + 0.05199972912669182, + 0.08584171533584595, + 0.32063689827919006, + 0.08154742419719696, + 0.10267326235771179, + 0.02064679190516472, + 0.05113017186522484, + 0.04999866709113121, + 0.015405399724841118 + ], + "credit_map": [ + 0.026497380807995796, + 0.020166870206594467, + 0.04045526310801506, + 0.021307630464434624, + 0.02552485093474388, + 0.03203893452882767, + 0.021157121285796165, + 0.03297192230820656, + 0.00614070612937212, + 0.003868731437250972, + 0.006962783634662628, + 0.005796702578663826, + 0.005897461902350187, + 0.00687278900295496, + 0.0048492103815078735, + 0.011611337773501873, + 0.012461183592677116, + 0.0065001556649804115, + 0.00817197561264038, + 0.009989848360419273, + 0.005660461261868477, + 0.011901604011654854, + 0.01872941106557846, + 0.012427066452801228, + 0.04432796314358711, + 0.03443349152803421, + 0.06900911033153534, + 0.03296252340078354, + 0.011580233462154865, + 0.03925827890634537, + 0.0340336412191391, + 0.05503162741661072, + 0.009184354916214943, + 0.013010566122829914, + 0.01158172357827425, + 0.010892985388636589, + 0.00508958799764514, + 0.008836120367050171, + 0.010911422781646252, + 0.01204064954072237, + 0.012387030757963657, + 0.019930770620703697, + 0.010622306726872921, + 0.010064376518130302, + 0.006721924524754286, + 0.010359802283346653, + 0.016019323840737343, + 0.016567712649703026, + 0.0024980721063911915, + 0.0029208166524767876, + 0.003310342552140355, + 0.0029129001777619123, + 0.001551992492750287, + 0.0024002790451049805, + 0.0020863492973148823, + 0.002966037020087242, + 0.007429744582623243, + 0.007050426676869392, + 0.008373573422431946, + 0.005901613272726536, + 0.0020933228079229593, + 0.006652985233813524, + 0.004789161961525679, + 0.008839335292577744, + 0.005293768830597401, + 0.0069828396663069725, + 0.00857808068394661, + 0.007981308735907078, + 0.0018616404850035906, + 0.00706560118123889, + 0.005026885773986578, + 0.007208534516394138, + 0.001315703266300261, + 0.0019746434409171343, + 0.0018845812883228064, + 0.0019124047830700874, + 0.0007843641797080636, + 0.0024085345212370157, + 0.0026252742391079664, + 0.002499892609193921 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.018199946731328964 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flame exhibit a gradient from ethereal blue at the base to golden and fiery orange at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame rising from a shimmering, translucent pool of melted wax that reflects the flame's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in slow, deliberate orbits around the flickering wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a velvety black abyss creating a high-contrast composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578536868095398 + ], + "opposite_logprob_tokens": [ + -0.6328536868095398 + ], + "logprob_reward": -0.7578536868095398, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578536868095398, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2475128173828125, + 0.09561057388782501, + 0.15501971542835236, + 0.05916016176342964, + 0.08449375629425049, + 0.12699413299560547, + 0.03353951871395111, + 0.10299258679151535, + 0.04341917484998703, + 0.051257580518722534 + ], + "credit_map": [ + 0.02638828568160534, + 0.01631322130560875, + 0.05247051641345024, + 0.01890784129500389, + 0.02444940060377121, + 0.0351826548576355, + 0.022765377536416054, + 0.051035527139902115, + 0.01234874315559864, + 0.009879798628389835, + 0.012116292491555214, + 0.011135043576359749, + 0.00977669470012188, + 0.012072798795998096, + 0.017702994868159294, + 0.010578202083706856, + 0.024638313800096512, + 0.010310834273695946, + 0.016145342960953712, + 0.01585131697356701, + 0.013543820939958096, + 0.015430871397256851, + 0.04801199212670326, + 0.011087221093475819, + 0.00919210072606802, + 0.0052638063207268715, + 0.009962273761630058, + 0.006106573157012463, + 0.0031498633325099945, + 0.007123372983187437, + 0.008363321423530579, + 0.009998848661780357, + 0.014497068710625172, + 0.017891474068164825, + 0.010279422625899315, + 0.008177630603313446, + 0.004946019500494003, + 0.00844891183078289, + 0.01105986163020134, + 0.009193367324769497, + 0.014924339018762112, + 0.02283148095011711, + 0.013073109090328217, + 0.014718006365001202, + 0.007604215759783983, + 0.011142289265990257, + 0.0165118221193552, + 0.026188869029283524, + 0.0038312694523483515, + 0.005300682038068771, + 0.004212297033518553, + 0.0050521185621619225, + 0.002308337017893791, + 0.003616199130192399, + 0.003920666873455048, + 0.0052979476749897, + 0.014014320448040962, + 0.013446114026010036, + 0.01647847518324852, + 0.010763903148472309, + 0.005285289604216814, + 0.009681107476353645, + 0.014249264262616634, + 0.019074106588959694, + 0.003274125512689352, + 0.006389445159584284, + 0.003571279812604189, + 0.006941622588783503, + 0.003095097839832306, + 0.006718359887599945, + 0.006296576000750065, + 0.0071326675824820995, + 0.004205859731882811, + 0.008205324411392212, + 0.005177757702767849, + 0.005214224569499493, + 0.00254809926263988, + 0.009289353154599667, + 0.008003541268408298, + 0.008613417856395245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.7578536868095398 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flame exhibit a gradient from ethereal blue at the base to golden and fiery orange at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame rising from a shimmering, translucent pool of melted wax that reflects the flame's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in slow, deliberate orbits around the flickering wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a velvety black abyss creating a high-contrast composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020580671727657318 + ], + "opposite_logprob_tokens": [ + -3.895580768585205 + ], + "logprob_reward": -0.020580671727657318, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020580671727657318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.273502379655838, + 0.05999234691262245, + 0.03743291646242142, + 0.07173539698123932, + 0.10901158303022385, + 0.1199803575873375, + 0.0485004223883152, + 0.23514620959758759, + 0.01420391071587801, + 0.030494483187794685 + ], + "credit_map": [ + 0.04612582549452782, + 0.01924811489880085, + 0.05241524428129196, + 0.0243794284760952, + 0.02211938612163067, + 0.04796890541911125, + 0.022804642096161842, + 0.03844086080789566, + 0.008749346248805523, + 0.0034018857404589653, + 0.010848484933376312, + 0.006529834121465683, + 0.0027221490163356066, + 0.010543671436607838, + 0.004076342098414898, + 0.013120630756020546, + 0.005686681717634201, + 0.0023818151094019413, + 0.007953858934342861, + 0.00509325135499239, + 0.002230508718639612, + 0.005005900282412767, + 0.004744311328977346, + 0.004336588084697723, + 0.014636545442044735, + 0.006026549264788628, + 0.009136694483458996, + 0.006609135307371616, + 0.004303319845348597, + 0.010681294836103916, + 0.011675494723021984, + 0.008666369132697582, + 0.0125611936673522, + 0.01966276206076145, + 0.010243367403745651, + 0.013411789201200008, + 0.007637610659003258, + 0.010111409239470959, + 0.01963663101196289, + 0.015746816992759705, + 0.015645962208509445, + 0.023016395047307014, + 0.014425266534090042, + 0.014855888672173023, + 0.007444222457706928, + 0.010717983357608318, + 0.01532730646431446, + 0.01854732818901539, + 0.0057160332798957825, + 0.006850546691566706, + 0.007405842188745737, + 0.005601930432021618, + 0.0032812475692480803, + 0.006958469748497009, + 0.006952503230422735, + 0.005733852740377188, + 0.03775147348642349, + 0.031455159187316895, + 0.03863895684480667, + 0.021808931604027748, + 0.011746698059141636, + 0.023610273376107216, + 0.02746642753481865, + 0.04266829788684845, + 0.0012922750320285559, + 0.002234774874523282, + 0.0012236321344971657, + 0.001545808743685484, + 0.0012728739529848099, + 0.002222792012616992, + 0.001724072266370058, + 0.0026876823976635933, + 0.0027373170014470816, + 0.003381762420758605, + 0.003250209614634514, + 0.004204757045954466, + 0.00222837645560503, + 0.004563018213957548, + 0.005081302486360073, + 0.0050477394834160805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.020580671727657318 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "cinematic macro shot captures the intimate details of a lone candle flame, the camera moving in slow, deliberate orbits around its flickering wick. The wick, a delicate, jagged thread, its tip emitting a deep, fiery red glow, gives rise to a dancing flame that sways and dances in the profound darkness. This teardrop of liquid light displays a mesmerizing gradient, from an ethereal blue at its base to a vibrant, pulsating golden and fiery orange body, culminating in a licking, smoky ashtray. The flame rises from a shimmering, translucent pool of melted wax, its surface trembling, perfectly reflecting its hypnotic movement. As the camera rotates smoothly, it reveals the flame's ever-changing shape, stretching, splitting, and condensing in a silent, graceful rhythm. The velvety black abyss of the background creates a stark, high-contrast composition, imbuing the scene with an atmosphere of tranquility, contemplation, and deep intimacy, one completely focused on a unique, enduring source of warmth and light.", + "seed": 6281, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_1/index_1100_seed_6281_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.11437117932364345, + "vlm_logprob": -0.11437117932364345, + "vlm_yes_no_margin": 3.849999952316284, + "vlm_token_credit": -0.11437117932364345, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flame exhibit a gradient from ethereal blue at the base to golden and fiery orange at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame rising from a shimmering, translucent pool of melted wax that reflects the flame's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in slow, deliberate orbits around the flickering wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a velvety black abyss creating a high-contrast composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015193445608019829 + ], + "opposite_logprob_tokens": [ + -4.515193462371826 + ], + "logprob_reward": -0.015193445608019829, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.015193445608019829, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10814429819583893, + 0.05817896127700806, + 0.06909395754337311, + 0.05958231911063194, + 0.06408753991127014, + 0.13507723808288574, + 0.2930618226528168, + 0.0369226336479187, + 0.10987444967031479, + 0.06597677618265152 + ], + "credit_map": [ + 0.011387872509658337, + 0.005879093427211046, + 0.01757262460887432, + 0.0129353953525424, + 0.00887220073491335, + 0.01570964604616165, + 0.014510651119053364, + 0.02127681113779545, + 0.005110105965286493, + 0.0030125954654067755, + 0.008400851860642433, + 0.006931181065738201, + 0.0057532647624611855, + 0.00840832944959402, + 0.008818088099360466, + 0.011744542978703976, + 0.009174411185085773, + 0.004501793067902327, + 0.011952805332839489, + 0.006582299247384071, + 0.003932610619813204, + 0.010475210845470428, + 0.0067960643209517, + 0.015678757801651955, + 0.007101994939148426, + 0.00377831282094121, + 0.013532028533518314, + 0.004760166630148888, + 0.0034822081215679646, + 0.009817143902182579, + 0.005195954814553261, + 0.011914504691958427, + 0.008570938371121883, + 0.004958049394190311, + 0.009313508868217468, + 0.0072495657950639725, + 0.0066746314987540245, + 0.009995423257350922, + 0.007655618246644735, + 0.009669800288975239, + 0.015244762413203716, + 0.010212407447397709, + 0.027442511171102524, + 0.011494486592710018, + 0.01193744782358408, + 0.018946340307593346, + 0.015236244536936283, + 0.02456304058432579, + 0.039421048015356064, + 0.020215453580021858, + 0.054730918258428574, + 0.03212811425328255, + 0.02000528946518898, + 0.042531728744506836, + 0.027105843648314476, + 0.056923408061265945, + 0.004475530236959457, + 0.0020277162548154593, + 0.0072233895771205425, + 0.0031944075599312782, + 0.003398652421310544, + 0.005856630392372608, + 0.004625932313501835, + 0.0061203776858747005, + 0.012795778922736645, + 0.008685086853802204, + 0.01883269101381302, + 0.00789783988147974, + 0.012022167444229126, + 0.016836438328027725, + 0.010906726121902466, + 0.02189771831035614, + 0.00763075752183795, + 0.00639670854434371, + 0.007500034291297197, + 0.004841374699026346, + 0.007622852455824614, + 0.009327028878033161, + 0.008538340218365192, + 0.014119680970907211 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -0.015193445608019829 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flame exhibit a gradient from ethereal blue at the base to golden and fiery orange at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame rising from a shimmering, translucent pool of melted wax that reflects the flame's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in slow, deliberate orbits around the flickering wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a velvety black abyss creating a high-contrast composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002510493155568838 + ], + "opposite_logprob_tokens": [ + -6.0025105476379395 + ], + "logprob_reward": -0.002510493155568838, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002510493155568838, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.109465591609478, + 0.0564056821167469, + 0.11058749258518219, + 0.04895050451159477, + 0.06149620562791824, + 0.058684591203927994, + 0.13308899104595184, + 0.23506160080432892, + 0.02116117626428604, + 0.165098175406456 + ], + "credit_map": [ + 0.013937538489699364, + 0.006122196558862925, + 0.01847035065293312, + 0.012686336413025856, + 0.007305956911295652, + 0.016776707023382187, + 0.013912132941186428, + 0.020254366099834442, + 0.005540554411709309, + 0.002903193002566695, + 0.00732986256480217, + 0.007016891147941351, + 0.006075652781873941, + 0.008031059987843037, + 0.0074036312289536, + 0.01210483256727457, + 0.014888914301991463, + 0.0049606128595769405, + 0.019337965175509453, + 0.012066958472132683, + 0.006356589496135712, + 0.01586093008518219, + 0.015048454515635967, + 0.022067060694098473, + 0.005875331349670887, + 0.002767113968729973, + 0.012155363336205482, + 0.004289213567972183, + 0.00407757144421339, + 0.007908124476671219, + 0.004369502421468496, + 0.007508280221372843, + 0.007663914933800697, + 0.004291363526135683, + 0.010699858888983727, + 0.006088709458708763, + 0.004334720782935619, + 0.009636850096285343, + 0.0066811637952923775, + 0.012099618092179298, + 0.007321600802242756, + 0.004865424241870642, + 0.009837938472628593, + 0.005998795852065086, + 0.004736519418656826, + 0.008790113963186741, + 0.005855638068169355, + 0.011278553865849972, + 0.017677336931228638, + 0.008265963755548, + 0.026129741221666336, + 0.014068897813558578, + 0.00781629141420126, + 0.01895018108189106, + 0.013051537796854973, + 0.0271290373057127, + 0.025282492861151695, + 0.008923161774873734, + 0.04690951108932495, + 0.021999161690473557, + 0.012737191282212734, + 0.03600599989295006, + 0.018984179943799973, + 0.0642198845744133, + 0.002732387511059642, + 0.0018946570344269276, + 0.0029070614837110043, + 0.0016885838704183698, + 0.0032124361023306847, + 0.0031508062966167927, + 0.0026921723037958145, + 0.0028830671217292547, + 0.02042102999985218, + 0.01004416961222887, + 0.03635045513510704, + 0.013198494911193848, + 0.012179807759821415, + 0.024368222802877426, + 0.02118735760450363, + 0.027348635718226433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.002510493155568838 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flame exhibit a gradient from ethereal blue at the base to golden and fiery orange at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame rising from a shimmering, translucent pool of melted wax that reflects the flame's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in slow, deliberate orbits around the flickering wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a velvety black abyss creating a high-contrast composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819339208304882 + ], + "opposite_logprob_tokens": [ + -4.018193244934082 + ], + "logprob_reward": -0.01819339208304882, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01819339208304882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06309778988361359, + 0.09119681268930435, + 0.06472360342741013, + 0.11870510131120682, + 0.11199473589658737, + 0.07188818603754044, + 0.16361384093761444, + 0.18711136281490326, + 0.0353524424135685, + 0.0923161581158638 + ], + "credit_map": [ + 0.007671345956623554, + 0.002699566073715687, + 0.010097591206431389, + 0.007185086607933044, + 0.004852557089179754, + 0.00951119139790535, + 0.009011328220367432, + 0.012069121934473515, + 0.008681816048920155, + 0.00463368883356452, + 0.014155942015349865, + 0.010524844750761986, + 0.005460737273097038, + 0.014147121459245682, + 0.010403635911643505, + 0.023189019411802292, + 0.008797027170658112, + 0.0037914675194770098, + 0.01040138490498066, + 0.006310727447271347, + 0.0051267920061945915, + 0.009538217447698116, + 0.00774530041962862, + 0.013012687675654888, + 0.014295929111540318, + 0.007356611546128988, + 0.023510918021202087, + 0.011271247640252113, + 0.008130989968776703, + 0.016912592574954033, + 0.015232520177960396, + 0.02199428156018257, + 0.009791657328605652, + 0.007876556366682053, + 0.024187324568629265, + 0.01026374101638794, + 0.005979057867079973, + 0.01749613881111145, + 0.014122729189693928, + 0.02227751910686493, + 0.008269253186881542, + 0.004567570518702269, + 0.015022139996290207, + 0.007199754007160664, + 0.00502744410187006, + 0.009957135654985905, + 0.008517099544405937, + 0.013327782042324543, + 0.02150578424334526, + 0.011771819554269314, + 0.030751165002584457, + 0.017514171078801155, + 0.011278737336397171, + 0.02327716536819935, + 0.01408474799245596, + 0.033430229872465134, + 0.019651316106319427, + 0.008531033992767334, + 0.03854653239250183, + 0.018268238753080368, + 0.0129061508923769, + 0.028163878247141838, + 0.016188759356737137, + 0.04485544189810753, + 0.004176259506493807, + 0.002968838205561042, + 0.004490162245929241, + 0.0020370781421661377, + 0.004139965865761042, + 0.005427240859717131, + 0.003146502887830138, + 0.00896639283746481, + 0.010453388094902039, + 0.010268059559166431, + 0.017299512401223183, + 0.009023409336805344, + 0.01012310478836298, + 0.012696479447185993, + 0.009944157674908638, + 0.012508045881986618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 3 + ], + "selected_reward": -0.01819339208304882 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flame exhibit a gradient from ethereal blue at the base to golden and fiery orange at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame rising from a shimmering, translucent pool of melted wax that reflects the flame's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in slow, deliberate orbits around the flickering wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a velvety black abyss creating a high-contrast composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5234199166297913 + ], + "opposite_logprob_tokens": [ + -0.8984199166297913 + ], + "logprob_reward": -0.5234199166297913, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5234199166297913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16287179291248322, + 0.08818651735782623, + 0.0799683928489685, + 0.029345910996198654, + 0.03670933470129967, + 0.07325484603643417, + 0.15192285180091858, + 0.10411064326763153, + 0.06823571026325226, + 0.2053939700126648 + ], + "credit_map": [ + 0.017759382724761963, + 0.010595515370368958, + 0.022600099444389343, + 0.02503626048564911, + 0.01206236146390438, + 0.023154815658926964, + 0.027368463575839996, + 0.024294888600707054, + 0.008096552453935146, + 0.005191666539758444, + 0.011213023215532303, + 0.012641211040318012, + 0.007803102023899555, + 0.012772774323821068, + 0.0143135916441679, + 0.01615459844470024, + 0.010621136985719204, + 0.005019283387809992, + 0.01362275704741478, + 0.0068237981759011745, + 0.006305044516921043, + 0.012292513623833656, + 0.0077920109033584595, + 0.017491847276687622, + 0.003104365896433592, + 0.0018791980110108852, + 0.005621153395622969, + 0.002996219089254737, + 0.002457082038745284, + 0.004538962617516518, + 0.0030952608212828636, + 0.0056536695919930935, + 0.004006253555417061, + 0.0028298012912273407, + 0.006517934612929821, + 0.0029981706757098436, + 0.0032755767460912466, + 0.005281470250338316, + 0.0043155658058822155, + 0.007484559901058674, + 0.0087823998183012, + 0.005441117566078901, + 0.016263877972960472, + 0.005384388379752636, + 0.006077029276639223, + 0.010798993520438671, + 0.007460720371454954, + 0.013046324253082275, + 0.019979778677225113, + 0.010678381659090519, + 0.028723539784550667, + 0.01632971689105034, + 0.010840753093361855, + 0.024075336754322052, + 0.013548031449317932, + 0.027747320011258125, + 0.012729795649647713, + 0.007455050945281982, + 0.022711167111992836, + 0.011879345402121544, + 0.01211287546902895, + 0.014939732849597931, + 0.011442597955465317, + 0.010840077884495258, + 0.008372747339308262, + 0.003959513269364834, + 0.009082954376935959, + 0.004016280639916658, + 0.005942181684076786, + 0.010562794283032417, + 0.007979749701917171, + 0.01831948757171631, + 0.023358842357993126, + 0.020082680508494377, + 0.03678769990801811, + 0.020544912666082382, + 0.021538907662034035, + 0.028365151956677437, + 0.020083213225007057, + 0.03463255986571312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.5234199166297913 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flame exhibit a gradient from ethereal blue at the base to golden and fiery orange at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame rising from a shimmering, translucent pool of melted wax that reflects the flame's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in slow, deliberate orbits around the flickering wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a velvety black abyss creating a high-contrast composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012538649141788483 + ], + "opposite_logprob_tokens": [ + -4.387538433074951 + ], + "logprob_reward": -0.012538649141788483, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012538649141788483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20159104466438293, + 0.06357552111148834, + 0.24032430350780487, + 0.10453250259160995, + 0.09728098660707474, + 0.03398634493350983, + 0.033586446195840836, + 0.04468246176838875, + 0.11019686609506607, + 0.07024353742599487 + ], + "credit_map": [ + 0.021351251751184464, + 0.01206169929355383, + 0.03548993915319443, + 0.025754457339644432, + 0.014824838377535343, + 0.028416765853762627, + 0.024292951449751854, + 0.03939913213253021, + 0.007046593818813562, + 0.004077473189681768, + 0.010099027305841446, + 0.005196744576096535, + 0.0064252945594489574, + 0.009640967473387718, + 0.004887181334197521, + 0.016202231869101524, + 0.03416739031672478, + 0.012488729320466518, + 0.04507729783654213, + 0.023496821522712708, + 0.01722487434744835, + 0.03268994763493538, + 0.024916987866163254, + 0.05026225373148918, + 0.01194041222333908, + 0.006626445334404707, + 0.023226642981171608, + 0.010134363546967506, + 0.006997699849307537, + 0.014625346288084984, + 0.010887451469898224, + 0.020094141364097595, + 0.010507941246032715, + 0.006419595796614885, + 0.0177058856934309, + 0.007827946916222572, + 0.005788227543234825, + 0.0152285136282444, + 0.012526116333901882, + 0.02127675525844097, + 0.0038955695927143097, + 0.0030191324185580015, + 0.005904909688979387, + 0.0034696960356086493, + 0.0032108863815665245, + 0.004781434778124094, + 0.0037485267966985703, + 0.005956188775599003, + 0.004644820932298899, + 0.002360385609790683, + 0.005415372550487518, + 0.004081315826624632, + 0.0022882905323058367, + 0.004986268933862448, + 0.003963904455304146, + 0.005846088752150536, + 0.005139258224517107, + 0.0031178577337414026, + 0.009182658046483994, + 0.0041283415630459785, + 0.005450363736599684, + 0.006350554525852203, + 0.0058307708241045475, + 0.0054826559498906136, + 0.011873345822095871, + 0.008312762714922428, + 0.022005733102560043, + 0.00792278815060854, + 0.011721701361238956, + 0.017583223059773445, + 0.013960733078420162, + 0.016816578805446625, + 0.008168471045792103, + 0.007951340638101101, + 0.010474990122020245, + 0.008640501648187637, + 0.006830520462244749, + 0.009566009044647217, + 0.00870030652731657, + 0.009911401197314262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.012538649141788483 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic macro shot captures the intimate details of a lone candle flame, the camera moving in slow, deliberate orbits around its flickering wick. The wick, a delicate, jagged thread, its tip emitting a deep, fiery red glow, gives rise to a dancing flame that sways and dances in the profound darkness. This teardrop of liquid light displays a mesmerizing gradient, from an ethereal blue at its base to a vibrant, pulsating golden and fiery orange body, culminating in a licking, smoky ashtray. The flame rises from a shimmering, translucent pool of melted wax, its surface trembling, perfectly reflecting its hypnotic movement. As the camera rotates smoothly, it reveals the flame's ever-changing shape, stretching, splitting, and condensing in a silent, graceful rhythm. The velvety black abyss of the background creates a stark, high-contrast composition, imbuing the scene with an atmosphere of tranquility, contemplation, and deep intimacy, one completely focused on a unique, enduring source of warmth and light.", + "seed": 6282, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_1/index_1100_seed_6282_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1132155055180192, + "vlm_logprob": -0.1132155055180192, + "vlm_yes_no_margin": 3.5499999940395357, + "vlm_token_credit": -0.1132155055180192, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flame exhibit a gradient from ethereal blue at the base to golden and fiery orange at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame rising from a shimmering, translucent pool of melted wax that reflects the flame's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in slow, deliberate orbits around the flickering wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a velvety black abyss creating a high-contrast composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008259662427008152 + ], + "opposite_logprob_tokens": [ + -5.3832597732543945 + ], + "logprob_reward": -0.008259662427008152, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.008259662427008152, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18449167907238007, + 0.0680084303021431, + 0.1597604602575302, + 0.06385619193315506, + 0.056242652237415314, + 0.1309415102005005, + 0.05362803861498833, + 0.13302621245384216, + 0.10281974077224731, + 0.04722508415579796 + ], + "credit_map": [ + 0.028640102595090866, + 0.023771557956933975, + 0.031464457511901855, + 0.020284146070480347, + 0.020526763051748276, + 0.019665414467453957, + 0.017864448949694633, + 0.02227477729320526, + 0.008863790892064571, + 0.008426160551607609, + 0.009250626899302006, + 0.005945052485913038, + 0.009647374972701073, + 0.013013308867812157, + 0.006388920824974775, + 0.0064731924794614315, + 0.014956006780266762, + 0.02013557031750679, + 0.040828265249729156, + 0.016321854665875435, + 0.012129622511565685, + 0.02965715527534485, + 0.010935644619166851, + 0.014796325005590916, + 0.0061044516041874886, + 0.005834903102368116, + 0.011132640764117241, + 0.006957092322409153, + 0.007300872355699539, + 0.01028384454548359, + 0.0077701169066131115, + 0.008472277782857418, + 0.006901843473315239, + 0.007612834218889475, + 0.007270348723977804, + 0.005400487221777439, + 0.005528572015464306, + 0.008898494765162468, + 0.005387511104345322, + 0.009242555126547813, + 0.015685418620705605, + 0.009797216393053532, + 0.020399929955601692, + 0.01731332205235958, + 0.012361008673906326, + 0.018333658576011658, + 0.01289204228669405, + 0.024158909916877747, + 0.0062452941201627254, + 0.0032914995681494474, + 0.013819530606269836, + 0.0052269697189331055, + 0.0033434045035392046, + 0.009330031462013721, + 0.004037098027765751, + 0.008334209211170673, + 0.018338004127144814, + 0.013079922646284103, + 0.01656540483236313, + 0.011863158084452152, + 0.0297018364071846, + 0.022837849333882332, + 0.008878122083842754, + 0.011761912144720554, + 0.015533880330622196, + 0.010518365539610386, + 0.02279076725244522, + 0.00937518198043108, + 0.010578143410384655, + 0.017487864941358566, + 0.008092994801700115, + 0.008442550897598267, + 0.004728217609226704, + 0.0036266634706407785, + 0.013099335134029388, + 0.005239872261881828, + 0.0035229255445301533, + 0.007286819629371166, + 0.005191934760659933, + 0.0045293159782886505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.008259662427008152 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flame exhibit a gradient from ethereal blue at the base to golden and fiery orange at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame rising from a shimmering, translucent pool of melted wax that reflects the flame's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in slow, deliberate orbits around the flickering wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a velvety black abyss creating a high-contrast composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005266008898615837 + ], + "opposite_logprob_tokens": [ + -5.255266189575195 + ], + "logprob_reward": -0.005266008898615837, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005266008898615837, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14769817888736725, + 0.20821210741996765, + 0.13802409172058105, + 0.06417089700698853, + 0.05714420974254608, + 0.051668327301740646, + 0.052553050220012665, + 0.18637661635875702, + 0.06084983050823212, + 0.03330268710851669 + ], + "credit_map": [ + 0.01818600296974182, + 0.013572881929576397, + 0.025511667132377625, + 0.020501991733908653, + 0.016212308779358864, + 0.013092050328850746, + 0.019083164632320404, + 0.021538108587265015, + 0.022052589803934097, + 0.025669818744063377, + 0.03433433920145035, + 0.016814840957522392, + 0.03374975547194481, + 0.03664856031537056, + 0.01829145848751068, + 0.02065075747668743, + 0.012251895852386951, + 0.016079433262348175, + 0.03907185047864914, + 0.01342813577502966, + 0.009120160713791847, + 0.02554827369749546, + 0.010124423541128635, + 0.012399917468428612, + 0.008075732737779617, + 0.006872509606182575, + 0.00932071078568697, + 0.007458314765244722, + 0.007064022123813629, + 0.010702848434448242, + 0.007444439455866814, + 0.00723232002928853, + 0.006918089929968119, + 0.00518525205552578, + 0.009592195972800255, + 0.004352340009063482, + 0.0052105216309428215, + 0.008945576846599579, + 0.005443270783871412, + 0.011496963910758495, + 0.008564832620322704, + 0.00439570564776659, + 0.009785477072000504, + 0.005976163782179356, + 0.004412038717418909, + 0.007496628910303116, + 0.004979867022484541, + 0.006057610735297203, + 0.007629880681633949, + 0.002440313808619976, + 0.01251978985965252, + 0.004111441783607006, + 0.005454277154058218, + 0.007956884801387787, + 0.004356034100055695, + 0.008084431290626526, + 0.026882220059633255, + 0.018612336367368698, + 0.0195604357868433, + 0.01617049053311348, + 0.04351701959967613, + 0.03215973824262619, + 0.010840464383363724, + 0.018633903935551643, + 0.008893275633454323, + 0.00608180183917284, + 0.01280475314706564, + 0.005294906906783581, + 0.006569800898432732, + 0.010115506127476692, + 0.005579365883022547, + 0.005510424263775349, + 0.004482963588088751, + 0.0015120976604521275, + 0.007906782440841198, + 0.004315950442105532, + 0.0024408039171248674, + 0.004979592747986317, + 0.003563044825568795, + 0.0041014510206878185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.005266008898615837 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flame exhibit a gradient from ethereal blue at the base to golden and fiery orange at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame rising from a shimmering, translucent pool of melted wax that reflects the flame's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in slow, deliberate orbits around the flickering wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a velvety black abyss creating a high-contrast composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023292383179068565 + ], + "opposite_logprob_tokens": [ + -3.773292303085327 + ], + "logprob_reward": -0.023292383179068565, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023292383179068565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1044357642531395, + 0.0496554896235466, + 0.1021476462483406, + 0.08984041959047318, + 0.07731363922357559, + 0.16335490345954895, + 0.13987313210964203, + 0.15739935636520386, + 0.05218624323606491, + 0.06379339098930359 + ], + "credit_map": [ + 0.014343516901135445, + 0.011250173673033714, + 0.01616191305220127, + 0.012845620512962341, + 0.01064113900065422, + 0.012046006508171558, + 0.014493881724774837, + 0.012653503566980362, + 0.005138295702636242, + 0.006723932456225157, + 0.008834648877382278, + 0.003942650277167559, + 0.007475142367184162, + 0.007697718683630228, + 0.004768901504576206, + 0.005074202548712492, + 0.01098499447107315, + 0.013598316349089146, + 0.02247765101492405, + 0.009496190585196018, + 0.009200544096529484, + 0.01667163334786892, + 0.009885244071483612, + 0.009833070449531078, + 0.009730970486998558, + 0.007611815817654133, + 0.014965752139687538, + 0.010599653236567974, + 0.011170784942805767, + 0.014879470691084862, + 0.011652067303657532, + 0.009229914285242558, + 0.008053488098084927, + 0.006115620024502277, + 0.011129013262689114, + 0.006365601439028978, + 0.0059737577103078365, + 0.011548899114131927, + 0.007836948148906231, + 0.020290309563279152, + 0.02309585176408291, + 0.00806883443146944, + 0.0342855229973793, + 0.014033396728336811, + 0.010368948802351952, + 0.022556249052286148, + 0.01527079101651907, + 0.035675298422575, + 0.015669748187065125, + 0.0075454795733094215, + 0.03464848920702934, + 0.015002652071416378, + 0.009513364173471928, + 0.023786617442965508, + 0.011560051701962948, + 0.022146735340356827, + 0.02022530697286129, + 0.01331876777112484, + 0.02318539097905159, + 0.013628862798213959, + 0.03273855894804001, + 0.024590732529759407, + 0.013924610801041126, + 0.01578713022172451, + 0.007412099279463291, + 0.004883288871496916, + 0.012301874347031116, + 0.0042142439633607864, + 0.004912532400339842, + 0.008469748310744762, + 0.004482746589928865, + 0.0055097066797316074, + 0.006986135616898537, + 0.005173517856746912, + 0.01719062402844429, + 0.006425514817237854, + 0.004294166341423988, + 0.009475287050008774, + 0.006903594359755516, + 0.007344556041061878 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.023292383179068565 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flame exhibit a gradient from ethereal blue at the base to golden and fiery orange at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame rising from a shimmering, translucent pool of melted wax that reflects the flame's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in slow, deliberate orbits around the flickering wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a velvety black abyss creating a high-contrast composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47431352734565735 + ], + "opposite_logprob_tokens": [ + -0.974313497543335 + ], + "logprob_reward": -0.47431352734565735, + "yes_no_margin_reward": 0.4999999701976776, + "token_credit_reward": -0.47431352734565735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2955452799797058, + 0.07152098417282104, + 0.3023624122142792, + 0.05178289860486984, + 0.038804084062576294, + 0.03206972777843475, + 0.03930862247943878, + 0.13427184522151947, + 0.013208919204771519, + 0.02112521044909954 + ], + "credit_map": [ + 0.04180745407938957, + 0.02477048709988594, + 0.052484672516584396, + 0.03532271087169647, + 0.024386167526245117, + 0.04061988368630409, + 0.04038159176707268, + 0.03577231988310814, + 0.007527428213506937, + 0.009737899526953697, + 0.010454270988702774, + 0.0094668073579669, + 0.007773305289447308, + 0.013000303879380226, + 0.005924748722463846, + 0.007636226713657379, + 0.0236098300665617, + 0.05731813982129097, + 0.06081356853246689, + 0.02761153131723404, + 0.028771715238690376, + 0.0546719953417778, + 0.027385486289858818, + 0.02218017913401127, + 0.005876970011740923, + 0.004824873525649309, + 0.00872684083878994, + 0.004122195765376091, + 0.008935470134019852, + 0.008205162361264229, + 0.006093977019190788, + 0.00499741593375802, + 0.004885267931967974, + 0.003212044481188059, + 0.006451004650443792, + 0.002751352032646537, + 0.004736635833978653, + 0.0058384668081998825, + 0.00374782457947731, + 0.007181490771472454, + 0.0043408190831542015, + 0.0023872104939073324, + 0.006680369842797518, + 0.003542601363733411, + 0.0031003402546048164, + 0.004466885235160589, + 0.0024416595697402954, + 0.005109842866659164, + 0.005648311693221331, + 0.0024600615724921227, + 0.0072954935021698475, + 0.005141111556440592, + 0.004855872131884098, + 0.006088954396545887, + 0.002317957114428282, + 0.005500865168869495, + 0.017230208963155746, + 0.011041365563869476, + 0.02102634124457836, + 0.012123830616474152, + 0.023825595155358315, + 0.020373018458485603, + 0.013203532434999943, + 0.015447963960468769, + 0.0019032444106414914, + 0.0013444687938317657, + 0.0020200731232762337, + 0.0011350181885063648, + 0.0018804350402206182, + 0.0020979391410946846, + 0.001169976661913097, + 0.0016577652422711253, + 0.0023754166904836893, + 0.0010236992966383696, + 0.006660733837634325, + 0.00219068699516356, + 0.0013232139172032475, + 0.0034593890886753798, + 0.0017511979676783085, + 0.0023408751003444195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.47431352734565735 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the flame exhibit a gradient from ethereal blue at the base to golden and fiery orange at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame rising from a shimmering, translucent pool of melted wax that reflects the flame's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in slow, deliberate orbits around the flickering wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a velvety black abyss creating a high-contrast composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054945945739746094 + ], + "opposite_logprob_tokens": [ + -2.929945945739746 + ], + "logprob_reward": -0.054945945739746094, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054945945739746094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19913113117218018, + 0.18746234476566315, + 0.31403884291648865, + 0.03019949421286583, + 0.034264616668224335, + 0.0668700709939003, + 0.024630075320601463, + 0.07944557070732117, + 0.04376697912812233, + 0.020190907642245293 + ], + "credit_map": [ + 0.023808026686310768, + 0.019166558980941772, + 0.03125153109431267, + 0.02699267491698265, + 0.020840706303715706, + 0.01942223682999611, + 0.02972363494336605, + 0.027925748378038406, + 0.02128182165324688, + 0.023452192544937134, + 0.029258595779538155, + 0.015550440177321434, + 0.025545217096805573, + 0.03858448565006256, + 0.016742179170250893, + 0.017047401517629623, + 0.028204625472426414, + 0.03894210606813431, + 0.08284121751785278, + 0.030491577461361885, + 0.023418612778186798, + 0.05650237202644348, + 0.021831346675753593, + 0.03180698677897453, + 0.0033175963908433914, + 0.0038043437525629997, + 0.005489472299814224, + 0.002527757780626416, + 0.00357008446007967, + 0.004905211739242077, + 0.003281343262642622, + 0.0033036815002560616, + 0.004294633865356445, + 0.0030509917996823788, + 0.005483682733029127, + 0.002334924880415201, + 0.004579572472721338, + 0.005046670325100422, + 0.0038217881228774786, + 0.005652347579598427, + 0.008348137140274048, + 0.004709158558398485, + 0.010490001179277897, + 0.007940837182104588, + 0.006507128942757845, + 0.009435555897653103, + 0.006160288117825985, + 0.013278958387672901, + 0.003664368065074086, + 0.001202923245728016, + 0.005188147537410259, + 0.0017419857904314995, + 0.0023604112211614847, + 0.0035962064284831285, + 0.0020977244712412357, + 0.004778306465595961, + 0.010046142153441906, + 0.007524012587964535, + 0.013540683314204216, + 0.006749906577169895, + 0.014557436108589172, + 0.012067090719938278, + 0.007028413005173206, + 0.007931879721581936, + 0.006200925912708044, + 0.004046183079481125, + 0.010720654390752316, + 0.003963241819292307, + 0.0038964084815233946, + 0.0071775163523852825, + 0.0036203612107783556, + 0.004141686949878931, + 0.002154149580746889, + 0.0009169774129986763, + 0.006508337799459696, + 0.0020531327463686466, + 0.0011959627736359835, + 0.0032748167868703604, + 0.0017925945576280355, + 0.0022949338890612125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.054945945739746094 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_10/index_tensor([6756, 6756])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_10/index_tensor([6756, 6756])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..eae438eef5558dbbb22495444bcbf7a9cf245439 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_10/index_tensor([6756, 6756])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "long shot, filmed from a slightly upward angle, slowly pans right across the deep, murky underwater world, revealing an awe-inspiring spectacle: a massive, bioluminescent jellyfish, its translucent parasols threaded with intricate biological veins pulsating with a soft, blue-green glow. Countless slender, luminous tentacles drift gracefully behind it, like silk. With a majestic and deliberate rhythm, the jellyfish, through the rhythmic contraction and expansion of its vast parasols, slowly circles the wreck of an ancient wooden sailing ship, resting quietly on the seabed. The hull of the wreck is thickly coated with algae and barnacles, its broken mast jutting feebly into the darkness, and tiny particles suspended in the water are clearly visible in the light. The jellyfish's own pulsating bioluminescence, the only source of light in the scene, penetrates the nearly inky water around it, casting a shifting, soft, and ethereal light and shadow on the wreck's mottled surface. The color saturation of the entire picture is low, showing a slight film grain and vignette effect, perfectly replicating the unique faded texture of old adventure movies, creating an atmosphere that is both mysterious and magnificent, full of loneliness and a sense of time precipitation.", + "seed": 27370, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_10/index_6756_seed_27370_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03401096221059561, + "vlm_logprob": -0.03401096221059561, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.03401096221059561, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive bioluminescent jellyfish as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish glowing with a soft blue-green light and featuring translucent parasols with visible veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jellyfish slowly circling the wreck of an ancient wooden sailing ship resting on the seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship wreck thickly coated with algae and barnacles, with a broken mast jutting into the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a long shot filmed from a slightly upward angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003122456604614854 + ], + "opposite_logprob_tokens": [ + -7.003122329711914 + ], + "logprob_reward": -0.003122456604614854, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.003122456604614854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28026387095451355, + 0.1087171733379364, + 0.06252790987491608, + 0.06530476361513138, + 0.0708819106221199, + 0.11169562488794327, + 0.07259324938058853, + 0.0660068541765213, + 0.0655529722571373, + 0.09645567089319229 + ], + "credit_map": [ + 0.05522922798991203, + 0.03663434460759163, + 0.037671104073524475, + 0.02826133742928505, + 0.03702647238969803, + 0.027457283809781075, + 0.027507901191711426, + 0.03047621063888073, + 0.016762956976890564, + 0.01090626697987318, + 0.01148368138819933, + 0.018566055223345757, + 0.011974668130278587, + 0.010503433644771576, + 0.010484769940376282, + 0.01803535409271717, + 0.009164532646536827, + 0.006893227808177471, + 0.007983285933732986, + 0.009071020409464836, + 0.007585339248180389, + 0.004808935336768627, + 0.00842836033552885, + 0.008593212813138962, + 0.00927947647869587, + 0.005486397538334131, + 0.010227350518107414, + 0.010232110507786274, + 0.008495314978063107, + 0.006366359535604715, + 0.006816423963755369, + 0.008401332423090935, + 0.010667501017451286, + 0.007173642516136169, + 0.009420300833880901, + 0.014226910658180714, + 0.006649785675108433, + 0.005483635701239109, + 0.008073768578469753, + 0.009186368435621262, + 0.023794785141944885, + 0.009446876123547554, + 0.020530041307210922, + 0.010308383964002132, + 0.01370969321578741, + 0.009076949208974838, + 0.009529517963528633, + 0.015299384482204914, + 0.011405718512833118, + 0.00860882643610239, + 0.011729788966476917, + 0.008134228177368641, + 0.006114221643656492, + 0.006602832116186619, + 0.0068496656604111195, + 0.013147979974746704, + 0.009228769689798355, + 0.007142253220081329, + 0.010136621072888374, + 0.0066672624088823795, + 0.006595879793167114, + 0.007078911177814007, + 0.007639823015779257, + 0.011517343111336231, + 0.0053606475703418255, + 0.005710322875529528, + 0.010441698133945465, + 0.005508648231625557, + 0.008712111972272396, + 0.017843367531895638, + 0.005933232139796019, + 0.006042951717972755, + 0.012341761961579323, + 0.012635164894163609, + 0.012944195419549942, + 0.01326705515384674, + 0.011819520965218544, + 0.009304176084697247, + 0.012072170153260231, + 0.012071636505424976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.003122456604614854 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive bioluminescent jellyfish as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish glowing with a soft blue-green light and featuring translucent parasols with visible veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jellyfish slowly circling the wreck of an ancient wooden sailing ship resting on the seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship wreck thickly coated with algae and barnacles, with a broken mast jutting into the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a long shot filmed from a slightly upward angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002510968828573823 + ], + "opposite_logprob_tokens": [ + -6.002511024475098 + ], + "logprob_reward": -0.002510968828573823, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002510968828573823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.174773171544075, + 0.22539742290973663, + 0.07068043202161789, + 0.07107117772102356, + 0.05301238223910332, + 0.11650785058736801, + 0.091585673391819, + 0.0362425334751606, + 0.034685708582401276, + 0.12604369223117828 + ], + "credit_map": [ + 0.02572721242904663, + 0.025723902508616447, + 0.024979576468467712, + 0.019835514947772026, + 0.02507263980805874, + 0.015506112948060036, + 0.014461195096373558, + 0.023467009887099266, + 0.023246487602591515, + 0.017357267439365387, + 0.030009068548679352, + 0.026735372841358185, + 0.030300982296466827, + 0.020703744143247604, + 0.034964632242918015, + 0.0420798659324646, + 0.007446568924933672, + 0.006587734911590815, + 0.009761047549545765, + 0.011551006697118282, + 0.006482002791017294, + 0.006851506419479847, + 0.00952827651053667, + 0.012472288683056831, + 0.008379107341170311, + 0.004244089126586914, + 0.011059576645493507, + 0.0071620577946305275, + 0.009747615084052086, + 0.007887136191129684, + 0.00815383717417717, + 0.01443775650113821, + 0.005849233828485012, + 0.005418449640274048, + 0.005712798330932856, + 0.010743868537247181, + 0.00780444173142314, + 0.003919415175914764, + 0.007447770796716213, + 0.006116404663771391, + 0.01923060230910778, + 0.00853017158806324, + 0.02716943621635437, + 0.010985924862325191, + 0.007631711661815643, + 0.009464503265917301, + 0.008512292988598347, + 0.024983208626508713, + 0.014299726113677025, + 0.010229126550257206, + 0.014828948304057121, + 0.012619178742170334, + 0.008845418691635132, + 0.006600955966860056, + 0.008151370100677013, + 0.01601095125079155, + 0.00566888740286231, + 0.004841944668442011, + 0.004412193316966295, + 0.005376506131142378, + 0.0036516801919788122, + 0.003192983567714691, + 0.003926957957446575, + 0.00517137860879302, + 0.004355012904852629, + 0.003037905553355813, + 0.00500967912375927, + 0.003194785676896572, + 0.002943493193015456, + 0.008671371266245842, + 0.003097475040704012, + 0.004375985357910395, + 0.01452587265521288, + 0.014703243039548397, + 0.014968526549637318, + 0.01671592891216278, + 0.014293678104877472, + 0.017479874193668365, + 0.01801079884171486, + 0.015345771797001362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.002510968828573823 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive bioluminescent jellyfish as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish glowing with a soft blue-green light and featuring translucent parasols with visible veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jellyfish slowly circling the wreck of an ancient wooden sailing ship resting on the seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship wreck thickly coated with algae and barnacles, with a broken mast jutting into the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a long shot filmed from a slightly upward angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978767640888691 + ], + "opposite_logprob_tokens": [ + -3.52978777885437 + ], + "logprob_reward": -0.02978767640888691, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978767640888691, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.267335444688797, + 0.1651124209165573, + 0.06186792254447937, + 0.05128692463040352, + 0.12534920871257782, + 0.11282939463853836, + 0.043045055121183395, + 0.020837558433413506, + 0.08142007142305374, + 0.07091590762138367 + ], + "credit_map": [ + 0.044415172189474106, + 0.014715742319822311, + 0.05504848062992096, + 0.013280175626277924, + 0.04217342659831047, + 0.020734300836920738, + 0.025034315884113312, + 0.05193384736776352, + 0.01686485856771469, + 0.010976530611515045, + 0.02344902977347374, + 0.013066274113953114, + 0.023762430995702744, + 0.01786881312727928, + 0.028229018673300743, + 0.030895454809069633, + 0.008907975628972054, + 0.005817068740725517, + 0.007763494737446308, + 0.008134927600622177, + 0.009546165354549885, + 0.005264075007289648, + 0.006444461643695831, + 0.009989758022129536, + 0.005363093689084053, + 0.007133232895284891, + 0.006660430692136288, + 0.009679418057203293, + 0.007407268043607473, + 0.004464146215468645, + 0.004811408929526806, + 0.005767927970737219, + 0.013153789564967155, + 0.012889581732451916, + 0.014771533198654652, + 0.022797923535108566, + 0.015475843101739883, + 0.01394622027873993, + 0.018212730064988136, + 0.014101591892540455, + 0.012480334378778934, + 0.005355019588023424, + 0.03221243992447853, + 0.007046168204396963, + 0.004351194016635418, + 0.012415998615324497, + 0.006071831565350294, + 0.03289640694856644, + 0.006313920486718416, + 0.0036671124398708344, + 0.006848777644336224, + 0.004158026538789272, + 0.003474435769021511, + 0.005038563162088394, + 0.0051698884926736355, + 0.008374327793717384, + 0.003235731739550829, + 0.0020419415086507797, + 0.0029901161324232817, + 0.0026976289227604866, + 0.0018232627771794796, + 0.002541303401812911, + 0.002910049632191658, + 0.0025975254829972982, + 0.007566655520349741, + 0.007862131111323833, + 0.012580806389451027, + 0.007737551815807819, + 0.014968862757086754, + 0.016844727098941803, + 0.0071641551330685616, + 0.006695187650620937, + 0.007358918897807598, + 0.010020401328802109, + 0.01071125827729702, + 0.008349565789103508, + 0.00856056623160839, + 0.009968919679522514, + 0.008227246813476086, + 0.00771903432905674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.02978767640888691 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive bioluminescent jellyfish as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish glowing with a soft blue-green light and featuring translucent parasols with visible veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jellyfish slowly circling the wreck of an ancient wooden sailing ship resting on the seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship wreck thickly coated with algae and barnacles, with a broken mast jutting into the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a long shot filmed from a slightly upward angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007626701612025499 + ], + "opposite_logprob_tokens": [ + -4.882626533508301 + ], + "logprob_reward": -0.007626701612025499, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007626701612025499, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17517682909965515, + 0.2033550888299942, + 0.06437767297029495, + 0.12076599150896072, + 0.06062815338373184, + 0.11993606388568878, + 0.04189961403608322, + 0.08218115568161011, + 0.03643573075532913, + 0.09524371474981308 + ], + "credit_map": [ + 0.045622844249010086, + 0.016479257494211197, + 0.045185692608356476, + 0.009898004122078419, + 0.012452146969735622, + 0.011107807978987694, + 0.010570685379207134, + 0.023860381916165352, + 0.020331166684627533, + 0.01477842777967453, + 0.02669370546936989, + 0.018787695094943047, + 0.02615048922598362, + 0.02479424700140953, + 0.033804696053266525, + 0.038014646619558334, + 0.007656084839254618, + 0.006097311619669199, + 0.007665298879146576, + 0.010008004494011402, + 0.009227999486029148, + 0.005195387173444033, + 0.007373989559710026, + 0.011153589934110641, + 0.012879767455160618, + 0.006973723415285349, + 0.023595275357365608, + 0.008971351198852062, + 0.017516624182462692, + 0.011797599494457245, + 0.011695042252540588, + 0.027336591854691505, + 0.007170927245169878, + 0.005882450379431248, + 0.007964998483657837, + 0.00895191915333271, + 0.0071344394236803055, + 0.00748825166374445, + 0.009735776111483574, + 0.006299382075667381, + 0.016059333458542824, + 0.006785016506910324, + 0.03175854682922363, + 0.007343245670199394, + 0.005965749267488718, + 0.012356577441096306, + 0.008764315396547318, + 0.030903268605470657, + 0.005639391951262951, + 0.0036309121642261744, + 0.006097554229199886, + 0.0035867609549313784, + 0.0035193823277950287, + 0.004837046843022108, + 0.006490230560302734, + 0.008098330348730087, + 0.013059305027127266, + 0.006063304841518402, + 0.011675393208861351, + 0.008715114556252956, + 0.004940320271998644, + 0.013350781984627247, + 0.009677942842245102, + 0.014698981307446957, + 0.005472294054925442, + 0.003630576655268669, + 0.00440942170098424, + 0.003293589223176241, + 0.004516158252954483, + 0.005174289457499981, + 0.005440158303827047, + 0.004499239381402731, + 0.012928646057844162, + 0.010630950331687927, + 0.012270808219909668, + 0.011346695944666862, + 0.008658748120069504, + 0.011886809021234512, + 0.01300084125250578, + 0.014520205557346344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.007626701612025499 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive bioluminescent jellyfish as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish glowing with a soft blue-green light and featuring translucent parasols with visible veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jellyfish slowly circling the wreck of an ancient wooden sailing ship resting on the seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship wreck thickly coated with algae and barnacles, with a broken mast jutting into the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a long shot filmed from a slightly upward angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12700700759887695 + ], + "opposite_logprob_tokens": [ + -2.127007007598877 + ], + "logprob_reward": -0.12700700759887695, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12700700759887695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3540796637535095, + 0.07948972284793854, + 0.08701872825622559, + 0.08676104247570038, + 0.10018797963857651, + 0.06294865906238556, + 0.05912415683269501, + 0.045040782541036606, + 0.05848843604326248, + 0.06686082482337952 + ], + "credit_map": [ + 0.060191571712493896, + 0.021181950345635414, + 0.06346256285905838, + 0.027818499132990837, + 0.05817769840359688, + 0.03138915076851845, + 0.032530102878808975, + 0.0593281053006649, + 0.01009155623614788, + 0.004645621404051781, + 0.012724567204713821, + 0.007565307896584272, + 0.009586291387677193, + 0.009442605078220367, + 0.00945673231035471, + 0.0159770417958498, + 0.009081232361495495, + 0.007298097014427185, + 0.01266467198729515, + 0.011224348098039627, + 0.011499440297484398, + 0.010510530322790146, + 0.010138596408069134, + 0.014601817354559898, + 0.008195791393518448, + 0.006294622551649809, + 0.020042235031723976, + 0.010424310341477394, + 0.009240071289241314, + 0.006955395918339491, + 0.00683135911822319, + 0.018777260556817055, + 0.01240956224501133, + 0.010908190160989761, + 0.012678049504756927, + 0.020184388384222984, + 0.009843040257692337, + 0.0072869728319346905, + 0.011840607970952988, + 0.015037174336612225, + 0.006808771751821041, + 0.003960784059017897, + 0.015234023332595825, + 0.005411052610725164, + 0.004681839607656002, + 0.00777750788256526, + 0.004761393181979656, + 0.014313282445073128, + 0.008205686695873737, + 0.0035998227540403605, + 0.011435505002737045, + 0.004229764454066753, + 0.0045685903169214725, + 0.007268465589731932, + 0.006398286670446396, + 0.013418037444353104, + 0.006595632527023554, + 0.004954130854457617, + 0.00641876133158803, + 0.005448987707495689, + 0.003430997021496296, + 0.0048138233833014965, + 0.004404535051435232, + 0.008973914198577404, + 0.005672535393387079, + 0.006457861512899399, + 0.008527335710823536, + 0.004704594612121582, + 0.007084423676133156, + 0.01537489052861929, + 0.004741020500659943, + 0.005925775971263647, + 0.009079251438379288, + 0.006580539513379335, + 0.011795621365308762, + 0.006099623162299395, + 0.00840223953127861, + 0.009251374751329422, + 0.0077287983149290085, + 0.007923377677798271 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.12700700759887695 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, filmed from a slightly upward angle, slowly pans right across the deep, murky underwater world, revealing an awe-inspiring spectacle: a massive, bioluminescent jellyfish, its translucent parasols threaded with intricate biological veins pulsating with a soft, blue-green glow. Countless slender, luminous tentacles drift gracefully behind it, like silk. With a majestic and deliberate rhythm, the jellyfish, through the rhythmic contraction and expansion of its vast parasols, slowly circles the wreck of an ancient wooden sailing ship, resting quietly on the seabed. The hull of the wreck is thickly coated with algae and barnacles, its broken mast jutting feebly into the darkness, and tiny particles suspended in the water are clearly visible in the light. The jellyfish's own pulsating bioluminescence, the only source of light in the scene, penetrates the nearly inky water around it, casting a shifting, soft, and ethereal light and shadow on the wreck's mottled surface. The color saturation of the entire picture is low, showing a slight film grain and vignette effect, perfectly replicating the unique faded texture of old adventure movies, creating an atmosphere that is both mysterious and magnificent, full of loneliness and a sense of time precipitation.", + "seed": 27371, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_10/index_6756_seed_27371_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05490347412414849, + "vlm_logprob": -0.05490347412414849, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.05490347412414849, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive bioluminescent jellyfish as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish glowing with a soft blue-green light and featuring translucent parasols with visible veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jellyfish slowly circling the wreck of an ancient wooden sailing ship resting on the seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship wreck thickly coated with algae and barnacles, with a broken mast jutting into the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a long shot filmed from a slightly upward angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0035415091551840305 + ], + "opposite_logprob_tokens": [ + -7.253541469573975 + ], + "logprob_reward": -0.0035415091551840305, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0035415091551840305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14073194563388824, + 0.1826106458902359, + 0.0636032298207283, + 0.08783119171857834, + 0.07704167068004608, + 0.08176064491271973, + 0.11152008175849915, + 0.08333446830511093, + 0.11665081977844238, + 0.054915327578783035 + ], + "credit_map": [ + 0.012859122827649117, + 0.02692328952252865, + 0.017201649025082588, + 0.015954433009028435, + 0.010368283838033676, + 0.019081782549619675, + 0.020015403628349304, + 0.018327990546822548, + 0.010171648114919662, + 0.012421570718288422, + 0.015190373174846172, + 0.01457107812166214, + 0.009956488385796547, + 0.04483329504728317, + 0.022674111649394035, + 0.05279208719730377, + 0.005297825671732426, + 0.007550528272986412, + 0.007292489055544138, + 0.00916193425655365, + 0.005656425841152668, + 0.007739902939647436, + 0.011754936538636684, + 0.009149184450507164, + 0.009562540799379349, + 0.01542249321937561, + 0.008465304039418697, + 0.016796039417386055, + 0.00673288106918335, + 0.010762033984065056, + 0.009192696772515774, + 0.010897200554609299, + 0.013739490881562233, + 0.011640917509794235, + 0.005777344107627869, + 0.00825302954763174, + 0.013007036410272121, + 0.008562703616917133, + 0.008047838695347309, + 0.008013314567506313, + 0.007763720117509365, + 0.012326700612902641, + 0.009667828679084778, + 0.01161395013332367, + 0.010230532847344875, + 0.009334786795079708, + 0.011161316186189651, + 0.009661815129220486, + 0.009933333843946457, + 0.014193631708621979, + 0.01807424984872341, + 0.01125829666852951, + 0.015040354803204536, + 0.012662373483181, + 0.012978415936231613, + 0.01737942546606064, + 0.006276097148656845, + 0.008013798855245113, + 0.014962435699999332, + 0.008990669623017311, + 0.007497541140764952, + 0.015473661944270134, + 0.008209102787077427, + 0.013911156915128231, + 0.013261562213301659, + 0.022693511098623276, + 0.01596737466752529, + 0.013802208006381989, + 0.012579872272908688, + 0.016154756769537926, + 0.010804574936628342, + 0.011386945843696594, + 0.00423981249332428, + 0.0043076262809336185, + 0.008171219378709793, + 0.006300942972302437, + 0.004078063648194075, + 0.006880545988678932, + 0.008370522409677505, + 0.01256659533828497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.0035415091551840305 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive bioluminescent jellyfish as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish glowing with a soft blue-green light and featuring translucent parasols with visible veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jellyfish slowly circling the wreck of an ancient wooden sailing ship resting on the seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship wreck thickly coated with algae and barnacles, with a broken mast jutting into the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a long shot filmed from a slightly upward angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017308506648987532 + ], + "opposite_logprob_tokens": [ + -6.376730918884277 + ], + "logprob_reward": -0.0017308506648987532, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017308506648987532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1474851816892624, + 0.08794473856687546, + 0.1373097002506256, + 0.14422579109668732, + 0.07721427828073502, + 0.09020961821079254, + 0.1383368968963623, + 0.051861654967069626, + 0.08079138398170471, + 0.04462084546685219 + ], + "credit_map": [ + 0.011706996709108353, + 0.023296082392334938, + 0.02056308463215828, + 0.018533887341618538, + 0.01068247202783823, + 0.02489115670323372, + 0.021696751937270164, + 0.01611473597586155, + 0.006504311691969633, + 0.008633121848106384, + 0.012628959491848946, + 0.010781757533550262, + 0.007636046502739191, + 0.015941224992275238, + 0.013259989209473133, + 0.012559314258396626, + 0.008291915990412235, + 0.011695604771375656, + 0.009716370142996311, + 0.012089337222278118, + 0.00875016488134861, + 0.02445324882864952, + 0.02778911218047142, + 0.034523941576480865, + 0.0191669724881649, + 0.0180066991597414, + 0.012735853902995586, + 0.01762780360877514, + 0.011611849069595337, + 0.01880745030939579, + 0.017725249752402306, + 0.02854388952255249, + 0.008859878405928612, + 0.010376883670687675, + 0.007082991302013397, + 0.009762519039213657, + 0.007925859652459621, + 0.009579544886946678, + 0.010156749747693539, + 0.013469848781824112, + 0.006990657653659582, + 0.010697543621063232, + 0.008411773480474949, + 0.008255534805357456, + 0.01274013053625822, + 0.010014094412326813, + 0.016674483194947243, + 0.016425389796495438, + 0.01198426354676485, + 0.016243478283286095, + 0.022380664944648743, + 0.014160669408738613, + 0.01937776431441307, + 0.017813678830862045, + 0.015391379594802856, + 0.02098497375845909, + 0.004686602856963873, + 0.007699455600231886, + 0.007265463471412659, + 0.0065113212913274765, + 0.006030681077390909, + 0.009297232143580914, + 0.0037875815760344267, + 0.006583311595022678, + 0.008300128392875195, + 0.011766989715397358, + 0.010983016341924667, + 0.011031542904675007, + 0.009655413217842579, + 0.010000589303672314, + 0.009012451395392418, + 0.010041238740086555, + 0.004665601998567581, + 0.0069281491450965405, + 0.006182675715535879, + 0.005525837652385235, + 0.004010915290564299, + 0.006017783656716347, + 0.00555653078481555, + 0.005733349826186895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.0017308506648987532 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive bioluminescent jellyfish as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish glowing with a soft blue-green light and featuring translucent parasols with visible veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jellyfish slowly circling the wreck of an ancient wooden sailing ship resting on the seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship wreck thickly coated with algae and barnacles, with a broken mast jutting into the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a long shot filmed from a slightly upward angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10029231011867523 + ], + "opposite_logprob_tokens": [ + -2.350292205810547 + ], + "logprob_reward": -0.10029231011867523, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10029231011867523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14025536179542542, + 0.04674110189080238, + 0.28247469663619995, + 0.1104486957192421, + 0.06813398003578186, + 0.0644584447145462, + 0.10035683214664459, + 0.04164181277155876, + 0.07489235699176788, + 0.07059673219919205 + ], + "credit_map": [ + 0.00957967434078455, + 0.01658989116549492, + 0.01941887103021145, + 0.015014584176242352, + 0.010700942017138004, + 0.0318773128092289, + 0.019716335460543633, + 0.017357749864459038, + 0.004923960193991661, + 0.005008635111153126, + 0.005999952554702759, + 0.006986204534769058, + 0.0027025670278817415, + 0.006808572448790073, + 0.007909423671662807, + 0.006401783786714077, + 0.01692037656903267, + 0.026079954579472542, + 0.021010151132941246, + 0.030457178130745888, + 0.025265173986554146, + 0.04932127147912979, + 0.05164305120706558, + 0.06177748739719391, + 0.011334266513586044, + 0.017311619594693184, + 0.008673871867358685, + 0.01832902990281582, + 0.009394489228725433, + 0.014422829262912273, + 0.016220925375819206, + 0.014761663973331451, + 0.005172900855541229, + 0.005008700303733349, + 0.007800505496561527, + 0.009489472024142742, + 0.005203064996749163, + 0.009420319460332394, + 0.010540947318077087, + 0.015498076565563679, + 0.005876355338841677, + 0.01007421500980854, + 0.007848662324249744, + 0.005996380001306534, + 0.009289680048823357, + 0.0068114097230136395, + 0.010776772163808346, + 0.0077849687077105045, + 0.009452064521610737, + 0.011409904807806015, + 0.016416039317846298, + 0.009253713302314281, + 0.013829119503498077, + 0.010359276086091995, + 0.012368873693048954, + 0.01726783812046051, + 0.002944475971162319, + 0.005639205686748028, + 0.004996262025088072, + 0.00480693019926548, + 0.003387108212336898, + 0.00911975372582674, + 0.0039221676997840405, + 0.006825908552855253, + 0.007094618398696184, + 0.0115518718957901, + 0.010871272534132004, + 0.008646042086184025, + 0.008753900416195393, + 0.010722285136580467, + 0.007954884320497513, + 0.009297482669353485, + 0.004822653718292713, + 0.005010210908949375, + 0.009891952387988567, + 0.00827031023800373, + 0.005068275146186352, + 0.01022836472839117, + 0.011635851114988327, + 0.01566910929977894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.10029231011867523 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive bioluminescent jellyfish as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish glowing with a soft blue-green light and featuring translucent parasols with visible veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jellyfish slowly circling the wreck of an ancient wooden sailing ship resting on the seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship wreck thickly coated with algae and barnacles, with a broken mast jutting into the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a long shot filmed from a slightly upward angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638513274490833 + ], + "opposite_logprob_tokens": [ + -4.758638381958008 + ], + "logprob_reward": -0.008638513274490833, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638513274490833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15692315995693207, + 0.0826440081000328, + 0.1728212982416153, + 0.13977250456809998, + 0.07757223397493362, + 0.12648682296276093, + 0.06599916517734528, + 0.028469784185290337, + 0.08152951300144196, + 0.06778155267238617 + ], + "credit_map": [ + 0.010370728559792042, + 0.016172071918845177, + 0.023087890818715096, + 0.015261746011674404, + 0.01273111067712307, + 0.037875574082136154, + 0.02361975610256195, + 0.017804274335503578, + 0.005830811336636543, + 0.007314630784094334, + 0.011003744788467884, + 0.011476445943117142, + 0.004976663272827864, + 0.013200249522924423, + 0.01309959776699543, + 0.015741867944598198, + 0.011427583172917366, + 0.01602993719279766, + 0.01297970861196518, + 0.015736091881990433, + 0.016610536724328995, + 0.029408017173409462, + 0.03252524882555008, + 0.038104165345430374, + 0.019111650064587593, + 0.019379641860723495, + 0.011726387776434422, + 0.02114943228662014, + 0.011403340846300125, + 0.01820802316069603, + 0.01536157913506031, + 0.023432450369000435, + 0.005073023494333029, + 0.005048931576311588, + 0.007007810287177563, + 0.009223926812410355, + 0.0076210168190300465, + 0.011819412000477314, + 0.013495717197656631, + 0.018282391130924225, + 0.009665122255682945, + 0.012120557948946953, + 0.010650165379047394, + 0.010485527105629444, + 0.017290357500314713, + 0.01434038020670414, + 0.023980148136615753, + 0.027954552322626114, + 0.005691828206181526, + 0.007791504729539156, + 0.01061477791517973, + 0.005663699004799128, + 0.009283272549510002, + 0.006739944219589233, + 0.00822598859667778, + 0.011988145299255848, + 0.002645868342369795, + 0.00353721110150218, + 0.004181568510830402, + 0.004307514522224665, + 0.0026904058177024126, + 0.0040117050521075726, + 0.0029229489155113697, + 0.004172563552856445, + 0.008205871097743511, + 0.0119484169408679, + 0.011553491465747356, + 0.009274927899241447, + 0.009038770571351051, + 0.011860044673085213, + 0.009126950055360794, + 0.010521041229367256, + 0.007390949875116348, + 0.00613044248893857, + 0.009496040642261505, + 0.010865524411201477, + 0.005352124571800232, + 0.009778698906302452, + 0.010333296842873096, + 0.008434474468231201 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.008638513274490833 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive bioluminescent jellyfish as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish glowing with a soft blue-green light and featuring translucent parasols with visible veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jellyfish slowly circling the wreck of an ancient wooden sailing ship resting on the seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship wreck thickly coated with algae and barnacles, with a broken mast jutting into the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a long shot filmed from a slightly upward angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603141874074936 + ], + "opposite_logprob_tokens": [ + -1.9103142023086548 + ], + "logprob_reward": -0.1603141874074936, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603141874074936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08698083460330963, + 0.0707758218050003, + 0.06602036952972412, + 0.21734635531902313, + 0.10256868600845337, + 0.12440401315689087, + 0.16364696621894836, + 0.06457249075174332, + 0.05741824209690094, + 0.04626627638936043 + ], + "credit_map": [ + 0.007042654324322939, + 0.015475263819098473, + 0.009229538962244987, + 0.008986649103462696, + 0.00953563954681158, + 0.010602285154163837, + 0.012424590066075325, + 0.013684210367500782, + 0.0035617335233837366, + 0.005052117630839348, + 0.006680063903331757, + 0.006693805567920208, + 0.0035726367495954037, + 0.017955679446458817, + 0.007954924367368221, + 0.019304854795336723, + 0.008960312232375145, + 0.006584768649190664, + 0.010559581220149994, + 0.009580406360328197, + 0.005416597705334425, + 0.00773167097941041, + 0.008455518633127213, + 0.00873150210827589, + 0.02797677181661129, + 0.02802368812263012, + 0.018559154123067856, + 0.027047403156757355, + 0.017169775441288948, + 0.026630261912941933, + 0.029952382668852806, + 0.041986893862485886, + 0.018220823258161545, + 0.015347179025411606, + 0.009427826851606369, + 0.010893909260630608, + 0.009842608124017715, + 0.011747830547392368, + 0.010599923320114613, + 0.01648857817053795, + 0.016199609264731407, + 0.02147218957543373, + 0.015882523730397224, + 0.015993429347872734, + 0.01747809909284115, + 0.012352228164672852, + 0.012580943293869495, + 0.012444974854588509, + 0.013597596436738968, + 0.019768578931689262, + 0.03146040067076683, + 0.016823405399918556, + 0.023019522428512573, + 0.019436758011579514, + 0.01720684953033924, + 0.02233383245766163, + 0.004833702463656664, + 0.007932829670608044, + 0.009910065680742264, + 0.007277029100805521, + 0.006683052517473698, + 0.013029729947447777, + 0.005760079249739647, + 0.009145998395979404, + 0.006492498330771923, + 0.010662207379937172, + 0.008134136907756329, + 0.007326326332986355, + 0.006529159378260374, + 0.006191394757479429, + 0.006062895059585571, + 0.006019616965204477, + 0.004703119397163391, + 0.007171275559812784, + 0.006476493552327156, + 0.005796102806925774, + 0.004430541768670082, + 0.007789854891598225, + 0.005296230781823397, + 0.004602657165378332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 5 + ], + "selected_reward": -0.1603141874074936 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, filmed from a slightly upward angle, slowly pans right across the deep, murky underwater world, revealing an awe-inspiring spectacle: a massive, bioluminescent jellyfish, its translucent parasols threaded with intricate biological veins pulsating with a soft, blue-green glow. Countless slender, luminous tentacles drift gracefully behind it, like silk. With a majestic and deliberate rhythm, the jellyfish, through the rhythmic contraction and expansion of its vast parasols, slowly circles the wreck of an ancient wooden sailing ship, resting quietly on the seabed. The hull of the wreck is thickly coated with algae and barnacles, its broken mast jutting feebly into the darkness, and tiny particles suspended in the water are clearly visible in the light. The jellyfish's own pulsating bioluminescence, the only source of light in the scene, penetrates the nearly inky water around it, casting a shifting, soft, and ethereal light and shadow on the wreck's mottled surface. The color saturation of the entire picture is low, showing a slight film grain and vignette effect, perfectly replicating the unique faded texture of old adventure movies, creating an atmosphere that is both mysterious and magnificent, full of loneliness and a sense of time precipitation.", + "seed": 27372, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_10/index_6756_seed_27372_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.028220678167417645, + "vlm_logprob": -0.028220678167417645, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.028220678167417645, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive bioluminescent jellyfish as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish glowing with a soft blue-green light and featuring translucent parasols with visible veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jellyfish slowly circling the wreck of an ancient wooden sailing ship resting on the seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship wreck thickly coated with algae and barnacles, with a broken mast jutting into the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a long shot filmed from a slightly upward angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003894009394571185 + ], + "opposite_logprob_tokens": [ + -6.253893852233887 + ], + "logprob_reward": -0.003894009394571185, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.003894009394571185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09730742871761322, + 0.02030048705637455, + 0.04189847409725189, + 0.018233364447951317, + 0.018439343199133873, + 0.5080938935279846, + 0.06744351983070374, + 0.05428028851747513, + 0.15003834664821625, + 0.023964855819940567 + ], + "credit_map": [ + 0.021242672577500343, + 0.011951586231589317, + 0.023082809522747993, + 0.00823122076690197, + 0.010260446928441525, + 0.007456585764884949, + 0.007353137247264385, + 0.007728962693363428, + 0.0033509512431919575, + 0.0024463001172989607, + 0.0024583027698099613, + 0.0019985497929155827, + 0.0037694668862968683, + 0.002743807854130864, + 0.0018047806806862354, + 0.0017283287597820163, + 0.007673516403883696, + 0.006344587542116642, + 0.00443823728710413, + 0.003128404263406992, + 0.004513057880103588, + 0.005525123327970505, + 0.0040867943316698074, + 0.006188753992319107, + 0.0025265170261263847, + 0.00326091842725873, + 0.001957520144060254, + 0.002131716115400195, + 0.0013882317580282688, + 0.0033387213479727507, + 0.0014006549026817083, + 0.002229083562269807, + 0.0026777118910104036, + 0.0032389257103204727, + 0.0024781026877462864, + 0.0014072805643081665, + 0.003248190274462104, + 0.0021947952918708324, + 0.001730708871036768, + 0.001463627559132874, + 0.07148030400276184, + 0.12282399088144302, + 0.07510623335838318, + 0.03135884180665016, + 0.07309767603874207, + 0.05277518182992935, + 0.04033442959189415, + 0.04111720994114876, + 0.009830796159803867, + 0.012225535698235035, + 0.009761270135641098, + 0.005783384665846825, + 0.005893619265407324, + 0.008683116175234318, + 0.006934370379894972, + 0.008331422694027424, + 0.008904554881155491, + 0.011411076411604881, + 0.00861187744885683, + 0.004929694812744856, + 0.004412615206092596, + 0.007000971585512161, + 0.004505890421569347, + 0.0045036072842776775, + 0.023908482864499092, + 0.01263588946312666, + 0.014887602999806404, + 0.013619521632790565, + 0.009685689583420753, + 0.017085589468479156, + 0.017458565533161163, + 0.04075700789690018, + 0.004043866414576769, + 0.002210786333307624, + 0.0033205533400177956, + 0.0026603287551552057, + 0.0018696438055485487, + 0.0023545806761831045, + 0.00404568063095212, + 0.0034594156313687563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.003894009394571185 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive bioluminescent jellyfish as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish glowing with a soft blue-green light and featuring translucent parasols with visible veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jellyfish slowly circling the wreck of an ancient wooden sailing ship resting on the seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship wreck thickly coated with algae and barnacles, with a broken mast jutting into the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a long shot filmed from a slightly upward angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004136696457862854 + ], + "opposite_logprob_tokens": [ + -5.504136562347412 + ], + "logprob_reward": -0.004136696457862854, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004136696457862854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2504912316799164, + 0.047336213290691376, + 0.05257318168878555, + 0.024293135851621628, + 0.025820037350058556, + 0.08755511790513992, + 0.12758885324001312, + 0.09514684975147247, + 0.13096553087234497, + 0.15822981297969818 + ], + "credit_map": [ + 0.06144706904888153, + 0.02513435296714306, + 0.07408711314201355, + 0.01797184906899929, + 0.021638287231326103, + 0.016366021707654, + 0.01585477404296398, + 0.01799177937209606, + 0.007183209527283907, + 0.006474110297858715, + 0.005932431202381849, + 0.004199046641588211, + 0.007707106880843639, + 0.005783652886748314, + 0.003958089277148247, + 0.006098571699112654, + 0.006262153387069702, + 0.008802390657365322, + 0.005346194840967655, + 0.003425574628636241, + 0.007274274248629808, + 0.007115185260772705, + 0.00476545887067914, + 0.009581954218447208, + 0.0034376175608485937, + 0.0041123488917946815, + 0.003673016093671322, + 0.0021583151537925005, + 0.0027852230705320835, + 0.0037581089418381453, + 0.0016359023284167051, + 0.002732605440542102, + 0.003908860497176647, + 0.00471677677705884, + 0.0038205261807888746, + 0.002536582760512829, + 0.0031409375369548798, + 0.0029778112657368183, + 0.0022030456457287073, + 0.002515499945729971, + 0.013377529568970203, + 0.01909182220697403, + 0.012128668837249279, + 0.006009411998093128, + 0.013303104788064957, + 0.0074198138900101185, + 0.006873894948512316, + 0.009350883774459362, + 0.01728171668946743, + 0.023489924147725105, + 0.01652519963681698, + 0.011337259784340858, + 0.012846225872635841, + 0.016937218606472015, + 0.01591634564101696, + 0.01325497031211853, + 0.01586129330098629, + 0.02061658911406994, + 0.009962393902242184, + 0.00937474425882101, + 0.008315442129969597, + 0.013575495220720768, + 0.00910663977265358, + 0.008334258571267128, + 0.020621277391910553, + 0.02442193776369095, + 0.015727797523140907, + 0.012457973323762417, + 0.015412330627441406, + 0.018252847716212273, + 0.010156099684536457, + 0.01391527708619833, + 0.024012774229049683, + 0.020741481333971024, + 0.028570711612701416, + 0.017483189702033997, + 0.007680252194404602, + 0.019368750974535942, + 0.018468542024493217, + 0.021904122084379196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.004136696457862854 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive bioluminescent jellyfish as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish glowing with a soft blue-green light and featuring translucent parasols with visible veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jellyfish slowly circling the wreck of an ancient wooden sailing ship resting on the seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship wreck thickly coated with algae and barnacles, with a broken mast jutting into the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a long shot filmed from a slightly upward angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537001632153988 + ], + "opposite_logprob_tokens": [ + -4.387537002563477 + ], + "logprob_reward": -0.012537001632153988, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537001632153988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06863968074321747, + 0.033517006784677505, + 0.03547248989343643, + 0.04099161922931671, + 0.03537575900554657, + 0.5827087759971619, + 0.0513063408434391, + 0.05006325617432594, + 0.071112260222435, + 0.03081277199089527 + ], + "credit_map": [ + 0.008424087427556515, + 0.011386292055249214, + 0.01263171061873436, + 0.006277431268244982, + 0.011009478010237217, + 0.006564166396856308, + 0.004458933137357235, + 0.007887584157288074, + 0.007306292653083801, + 0.004425795283168554, + 0.004317064303904772, + 0.003316905116662383, + 0.004154528025537729, + 0.004326455760747194, + 0.0024256675969809294, + 0.0032442964147776365, + 0.00552216125652194, + 0.0054415869526565075, + 0.0044187624007463455, + 0.0030758471693843603, + 0.004066654480993748, + 0.0045143552124500275, + 0.002965040970593691, + 0.0054680840112268925, + 0.004480871371924877, + 0.005888889543712139, + 0.003866777289658785, + 0.004091581329703331, + 0.009481415152549744, + 0.006180582568049431, + 0.00283691193908453, + 0.004164592362940311, + 0.00520601449534297, + 0.004613017197698355, + 0.00816415622830391, + 0.0045780339278280735, + 0.002454604022204876, + 0.004156635142862797, + 0.0027702548541128635, + 0.00343304337002337, + 0.0834525004029274, + 0.1310443878173828, + 0.09404110908508301, + 0.035886768251657486, + 0.08779828250408173, + 0.05244753509759903, + 0.04178109019994736, + 0.05625713989138603, + 0.00777380121871829, + 0.00760632986202836, + 0.006735456641763449, + 0.005157670006155968, + 0.005602195858955383, + 0.0078004710376262665, + 0.005259817466139793, + 0.005370601080358028, + 0.007362721022218466, + 0.00689306203275919, + 0.008662354201078415, + 0.005787302739918232, + 0.004876826424151659, + 0.005317807197570801, + 0.005986938718706369, + 0.005176249425858259, + 0.010323227383196354, + 0.006394722033292055, + 0.006739581003785133, + 0.006447191815823317, + 0.003432365134358406, + 0.008674721233546734, + 0.008217711932957172, + 0.0208827443420887, + 0.004823984112590551, + 0.0035473578609526157, + 0.004659781698137522, + 0.0029748438391834497, + 0.00311582931317389, + 0.004289702512323856, + 0.003738846629858017, + 0.0036624257918447256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.012537001632153988 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive bioluminescent jellyfish as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish glowing with a soft blue-green light and featuring translucent parasols with visible veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jellyfish slowly circling the wreck of an ancient wooden sailing ship resting on the seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship wreck thickly coated with algae and barnacles, with a broken mast jutting into the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a long shot filmed from a slightly upward angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631551939994097 + ], + "opposite_logprob_tokens": [ + -4.882631778717041 + ], + "logprob_reward": -0.007631551939994097, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631551939994097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1488734781742096, + 0.033557381480932236, + 0.07383207231760025, + 0.033524367958307266, + 0.031076006591320038, + 0.13548493385314941, + 0.2792690694332123, + 0.09020785987377167, + 0.10906023532152176, + 0.06511455774307251 + ], + "credit_map": [ + 0.04081781581044197, + 0.01338426023721695, + 0.038023095577955246, + 0.011132758110761642, + 0.012554511427879333, + 0.01081931684166193, + 0.008184431120753288, + 0.013957304880023003, + 0.005042237229645252, + 0.004091419279575348, + 0.00520943570882082, + 0.003452452365309, + 0.005264109466224909, + 0.0038870838470757008, + 0.0023245352786034346, + 0.0042861090041697025, + 0.007310953456908464, + 0.012263267301023006, + 0.007189181633293629, + 0.005558973643928766, + 0.0051328702829778194, + 0.01174027007073164, + 0.0066560786217451096, + 0.017980482429265976, + 0.0057083661668002605, + 0.005422802176326513, + 0.005034240428358316, + 0.0023877350613474846, + 0.004445428494364023, + 0.004998222924768925, + 0.002152667846530676, + 0.0033749074209481478, + 0.004617155529558659, + 0.0046860650181770325, + 0.004272276069968939, + 0.00325858686119318, + 0.0034624210093170404, + 0.003922310192137957, + 0.0034049467649310827, + 0.0034522481728345156, + 0.01949950121343136, + 0.029936080798506737, + 0.0217331200838089, + 0.00846114382147789, + 0.017712486907839775, + 0.013769344426691532, + 0.01182599551975727, + 0.012547274120151997, + 0.03844563663005829, + 0.050587914884090424, + 0.03709583729505539, + 0.026982780545949936, + 0.025498084723949432, + 0.035148877650499344, + 0.035060323774814606, + 0.030449634417891502, + 0.015575436875224113, + 0.014989222399890423, + 0.009245357476174831, + 0.00957409292459488, + 0.011279766447842121, + 0.013000569306313992, + 0.008295690640807152, + 0.00824772845953703, + 0.015980353578925133, + 0.022158946841955185, + 0.01667921431362629, + 0.01286403276026249, + 0.007807936053723097, + 0.0119261359795928, + 0.009536618366837502, + 0.012107008136808872, + 0.011616886593401432, + 0.005426733288913965, + 0.01123790629208088, + 0.006438747514039278, + 0.004272982478141785, + 0.008526851423084736, + 0.007876141928136349, + 0.00971831101924181 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.007631551939994097 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive bioluminescent jellyfish as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish glowing with a soft blue-green light and featuring translucent parasols with visible veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jellyfish slowly circling the wreck of an ancient wooden sailing ship resting on the seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship wreck thickly coated with algae and barnacles, with a broken mast jutting into the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a long shot filmed from a slightly upward angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129041314125061 + ], + "opposite_logprob_tokens": [ + -2.2379040718078613 + ], + "logprob_reward": -0.1129041314125061, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129041314125061, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040489550679922104, + 0.02151380106806755, + 0.0568079799413681, + 0.030084805563092232, + 0.023813864216208458, + 0.4951651990413666, + 0.095130056142807, + 0.10458838939666748, + 0.09356199949979782, + 0.03884435445070267 + ], + "credit_map": [ + 0.006155712064355612, + 0.0057163480669260025, + 0.007629693020135164, + 0.004435531795024872, + 0.004782830830663443, + 0.003921879455447197, + 0.0032285447232425213, + 0.004619010258466005, + 0.00271764793433249, + 0.0032446477562189102, + 0.002496349159628153, + 0.0017769725527614355, + 0.004536242689937353, + 0.0029922083485871553, + 0.0015522221801802516, + 0.0021975121926516294, + 0.008732065558433533, + 0.008781781420111656, + 0.004971944261342287, + 0.004776873625814915, + 0.004099377430975437, + 0.008145160041749477, + 0.005344626493752003, + 0.011956148780882359, + 0.003962724003940821, + 0.0046031279489398, + 0.005072109401226044, + 0.002556583611294627, + 0.003618062473833561, + 0.004042478743940592, + 0.0024036518298089504, + 0.0038260670844465494, + 0.00313859386369586, + 0.003207292640581727, + 0.003286025021225214, + 0.0020674914121627808, + 0.005822465755045414, + 0.0023951365146785975, + 0.001853722846135497, + 0.002043136628344655, + 0.07164467871189117, + 0.1082601249217987, + 0.08241578191518784, + 0.03167606517672539, + 0.07458174973726273, + 0.04203347489237785, + 0.0347573384642601, + 0.0497959740459919, + 0.016249490901827812, + 0.022050200030207634, + 0.010559746995568275, + 0.007566614076495171, + 0.012243147008121014, + 0.011353556998074055, + 0.0064296419732272625, + 0.00867766235023737, + 0.017556389793753624, + 0.011441628448665142, + 0.021345822140574455, + 0.009315584786236286, + 0.009674012660980225, + 0.01306158397346735, + 0.00998431071639061, + 0.012209049426019192, + 0.014104360714554787, + 0.008288917131721973, + 0.00833655335009098, + 0.007868736051023006, + 0.005207335576415062, + 0.009297178126871586, + 0.011807048693299294, + 0.02865186519920826, + 0.006257269065827131, + 0.005457834340631962, + 0.0038184686563909054, + 0.004962695296853781, + 0.002835809485986829, + 0.005534308962523937, + 0.005232392344623804, + 0.004745577462017536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.1129041314125061 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, filmed from a slightly upward angle, slowly pans right across the deep, murky underwater world, revealing an awe-inspiring spectacle: a massive, bioluminescent jellyfish, its translucent parasols threaded with intricate biological veins pulsating with a soft, blue-green glow. Countless slender, luminous tentacles drift gracefully behind it, like silk. With a majestic and deliberate rhythm, the jellyfish, through the rhythmic contraction and expansion of its vast parasols, slowly circles the wreck of an ancient wooden sailing ship, resting quietly on the seabed. The hull of the wreck is thickly coated with algae and barnacles, its broken mast jutting feebly into the darkness, and tiny particles suspended in the water are clearly visible in the light. The jellyfish's own pulsating bioluminescence, the only source of light in the scene, penetrates the nearly inky water around it, casting a shifting, soft, and ethereal light and shadow on the wreck's mottled surface. The color saturation of the entire picture is low, showing a slight film grain and vignette effect, perfectly replicating the unique faded texture of old adventure movies, creating an atmosphere that is both mysterious and magnificent, full of loneliness and a sense of time precipitation.", + "seed": 27373, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_10/index_6756_seed_27373_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03637804957106709, + "vlm_logprob": -0.03637804957106709, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.03637804957106709, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive bioluminescent jellyfish as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish glowing with a soft blue-green light and featuring translucent parasols with visible veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jellyfish slowly circling the wreck of an ancient wooden sailing ship resting on the seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship wreck thickly coated with algae and barnacles, with a broken mast jutting into the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a long shot filmed from a slightly upward angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0045491536147892475 + ], + "opposite_logprob_tokens": [ + -5.629549026489258 + ], + "logprob_reward": -0.0045491536147892475, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0045491536147892475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14434538781642914, + 0.06896481662988663, + 0.05048052594065666, + 0.033790506422519684, + 0.03031284362077713, + 0.24297520518302917, + 0.04293110594153404, + 0.05117223784327507, + 0.032132621854543686, + 0.3028947114944458 + ], + "credit_map": [ + 0.024199344217777252, + 0.010536410845816135, + 0.028012244030833244, + 0.013900214806199074, + 0.011472188867628574, + 0.025453414767980576, + 0.0191301628947258, + 0.011641411110758781, + 0.01456417702138424, + 0.008242974057793617, + 0.008948439732193947, + 0.005484485067427158, + 0.007871200330555439, + 0.01241002231836319, + 0.006290559656918049, + 0.005152969621121883, + 0.006787288933992386, + 0.006569250486791134, + 0.006781847681850195, + 0.005017491988837719, + 0.0066492329351603985, + 0.0057173133827745914, + 0.0061358059756457806, + 0.006822298746556044, + 0.004651633556932211, + 0.002941255923360586, + 0.006638834252953529, + 0.004783718381077051, + 0.003191109048202634, + 0.004532226361334324, + 0.0033573803957551718, + 0.0036943501327186823, + 0.0048854053020477295, + 0.0035635014064610004, + 0.0037130347918719053, + 0.0036601100582629442, + 0.00331871653907001, + 0.0034942561760544777, + 0.00427878787741065, + 0.0033990328665822744, + 0.026117993518710136, + 0.014402768574655056, + 0.055698737502098083, + 0.021271580830216408, + 0.0149735352024436, + 0.07621447741985321, + 0.013157970272004604, + 0.021138161420822144, + 0.006078541278839111, + 0.004555459599941969, + 0.0053853620775043964, + 0.005157222039997578, + 0.004313932731747627, + 0.005869968794286251, + 0.0043970742262899876, + 0.007173549383878708, + 0.007405256852507591, + 0.006332912482321262, + 0.007108019664883614, + 0.0072168693877756596, + 0.00586857320740819, + 0.005199507810175419, + 0.007015001028776169, + 0.0050261011347174644, + 0.005195024888962507, + 0.004063380882143974, + 0.006491722539067268, + 0.004502894356846809, + 0.0029072731267660856, + 0.0028571353759616613, + 0.003344764932990074, + 0.002770428080111742, + 0.028652839362621307, + 0.10380437970161438, + 0.00907383393496275, + 0.017193453386425972, + 0.019323188811540604, + 0.019837485626339912, + 0.08466075360774994, + 0.020348813384771347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.0045491536147892475 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive bioluminescent jellyfish as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish glowing with a soft blue-green light and featuring translucent parasols with visible veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jellyfish slowly circling the wreck of an ancient wooden sailing ship resting on the seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship wreck thickly coated with algae and barnacles, with a broken mast jutting into the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a long shot filmed from a slightly upward angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036359410732984543 + ], + "opposite_logprob_tokens": [ + -5.628635883331299 + ], + "logprob_reward": -0.0036359410732984543, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036359410732984543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23263691365718842, + 0.06652763485908508, + 0.024072708562016487, + 0.04926639050245285, + 0.04165445640683174, + 0.1314297318458557, + 0.06103861331939697, + 0.0738358348608017, + 0.03269454836845398, + 0.2868432402610779 + ], + "credit_map": [ + 0.03568388149142265, + 0.019500551745295525, + 0.033517468720674515, + 0.02518976666033268, + 0.025098279118537903, + 0.03612326830625534, + 0.0358179472386837, + 0.021705709397792816, + 0.00973790418356657, + 0.006017106585204601, + 0.015318108722567558, + 0.004851421806961298, + 0.004989579785615206, + 0.012396822683513165, + 0.007960700429975986, + 0.00525597995147109, + 0.0037463705521076918, + 0.0026205701287835836, + 0.0029521032702177763, + 0.0025134996976703405, + 0.0033701553475111723, + 0.0031270424369722605, + 0.002633181633427739, + 0.00310978339985013, + 0.005541516933590174, + 0.003994673024863005, + 0.010538064874708652, + 0.004987435881048441, + 0.004745893180370331, + 0.009817569516599178, + 0.006204236298799515, + 0.0034369947388768196, + 0.0065767718479037285, + 0.0055724033154547215, + 0.004125922452658415, + 0.004441569093614817, + 0.004613010212779045, + 0.005310089327394962, + 0.006266500800848007, + 0.00474818330258131, + 0.01319833379238844, + 0.009965040720999241, + 0.02895464189350605, + 0.01236012950539589, + 0.008995420299470425, + 0.0402996763586998, + 0.006555628497153521, + 0.011100847274065018, + 0.00858218502253294, + 0.0053011407144367695, + 0.010616512969136238, + 0.008140264078974724, + 0.005420552100986242, + 0.007289383094757795, + 0.005651139188557863, + 0.010037433356046677, + 0.00844380259513855, + 0.010385689325630665, + 0.007208859547972679, + 0.005542363040149212, + 0.010893684811890125, + 0.013613291084766388, + 0.009998072870075703, + 0.007750059012323618, + 0.0051698870956897736, + 0.0038288363721221685, + 0.007013264577835798, + 0.004465907346457243, + 0.003005279926583171, + 0.003230712143704295, + 0.0032010788563638926, + 0.0027795773930847645, + 0.02675553970038891, + 0.09672550857067108, + 0.008932800963521004, + 0.01578075811266899, + 0.014918604865670204, + 0.01899273321032524, + 0.08491457253694534, + 0.019822699949145317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.0036359410732984543 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive bioluminescent jellyfish as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish glowing with a soft blue-green light and featuring translucent parasols with visible veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jellyfish slowly circling the wreck of an ancient wooden sailing ship resting on the seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship wreck thickly coated with algae and barnacles, with a broken mast jutting into the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a long shot filmed from a slightly upward angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000958919525146 + ], + "opposite_logprob_tokens": [ + -2.695009708404541 + ], + "logprob_reward": -0.07000958919525146, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000958919525146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18225695192813873, + 0.17579485476016998, + 0.04272491857409477, + 0.0730176717042923, + 0.062274910509586334, + 0.19663283228874207, + 0.0926775261759758, + 0.04949309304356575, + 0.02590073086321354, + 0.09922648966312408 + ], + "credit_map": [ + 0.029151614755392075, + 0.015311949886381626, + 0.03293042257428169, + 0.0191898662596941, + 0.014967509545385838, + 0.032761964946985245, + 0.02050650492310524, + 0.017437128350138664, + 0.031103434041142464, + 0.018149515613913536, + 0.019459307193756104, + 0.015721630305051804, + 0.019475877285003662, + 0.0273582860827446, + 0.02534390427172184, + 0.01918289251625538, + 0.006692684255540371, + 0.005902206525206566, + 0.005911533255130053, + 0.0034669823944568634, + 0.003132288111373782, + 0.005361039191484451, + 0.004555847961455584, + 0.007702338043600321, + 0.008676479570567608, + 0.005991803947836161, + 0.010760104283690453, + 0.010340871289372444, + 0.00543940206989646, + 0.013018041849136353, + 0.00732022151350975, + 0.011470746248960495, + 0.015956254675984383, + 0.009489260613918304, + 0.00800691731274128, + 0.005751618184149265, + 0.0040577310137450695, + 0.007327888160943985, + 0.0041590058244764805, + 0.007526229601353407, + 0.02129172533750534, + 0.009850827045738697, + 0.04387960210442543, + 0.016923589631915092, + 0.012505329214036465, + 0.06390511989593506, + 0.010876737534999847, + 0.01739991083741188, + 0.013026985339820385, + 0.006370509043335915, + 0.02408270537853241, + 0.00705565232783556, + 0.0060253138653934, + 0.00504917511716485, + 0.007997294887900352, + 0.02306988835334778, + 0.007465856149792671, + 0.007268584333360195, + 0.008417900651693344, + 0.007531046401709318, + 0.0051963599398732185, + 0.004915530793368816, + 0.004981353413313627, + 0.003716463688760996, + 0.003659468609839678, + 0.0035549714230000973, + 0.004865254741162062, + 0.003237010445445776, + 0.0027729871217161417, + 0.0030999945010989904, + 0.0023034121841192245, + 0.002407631604000926, + 0.01038291770964861, + 0.03058123029768467, + 0.005176952108740807, + 0.005424135364592075, + 0.00741304736584425, + 0.010836517438292503, + 0.019907688722014427, + 0.009503999724984169 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.07000958919525146 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive bioluminescent jellyfish as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish glowing with a soft blue-green light and featuring translucent parasols with visible veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jellyfish slowly circling the wreck of an ancient wooden sailing ship resting on the seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship wreck thickly coated with algae and barnacles, with a broken mast jutting into the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a long shot filmed from a slightly upward angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370098024606705 + ], + "opposite_logprob_tokens": [ + -3.408700942993164 + ], + "logprob_reward": -0.03370098024606705, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370098024606705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19210898876190186, + 0.11998621374368668, + 0.06851574033498764, + 0.0790366604924202, + 0.08035974204540253, + 0.0782175287604332, + 0.13540662825107574, + 0.05488761141896248, + 0.023318210616707802, + 0.16816264390945435 + ], + "credit_map": [ + 0.025830110535025597, + 0.014958850108087063, + 0.03842764347791672, + 0.016572246327996254, + 0.018019504845142365, + 0.032950516790151596, + 0.023912664502859116, + 0.021437471732497215, + 0.019676053896546364, + 0.015020034275949001, + 0.019359804689884186, + 0.010782689787447453, + 0.01243180688470602, + 0.0212051160633564, + 0.01296449638903141, + 0.008546220138669014, + 0.007296269293874502, + 0.008945615962147713, + 0.010943944565951824, + 0.007550233043730259, + 0.008837991394102573, + 0.007645376026630402, + 0.009626738727092743, + 0.007669580634683371, + 0.00956210307776928, + 0.005503567401319742, + 0.016913294792175293, + 0.006348043214529753, + 0.007075156085193157, + 0.01875465176999569, + 0.008447568863630295, + 0.0064322794787585735, + 0.01258867047727108, + 0.011480258777737617, + 0.011584676802158356, + 0.009968729689717293, + 0.007568107917904854, + 0.008814280852675438, + 0.009318082593381405, + 0.009036940522491932, + 0.010792005807161331, + 0.007875317707657814, + 0.007216230966150761, + 0.006709687411785126, + 0.005504768341779709, + 0.023507950827479362, + 0.007418738212436438, + 0.00919283926486969, + 0.02124945819377899, + 0.009042833931744099, + 0.036585111171007156, + 0.015097507275640965, + 0.007442752830684185, + 0.008637036196887493, + 0.008800522424280643, + 0.02855142019689083, + 0.0068130469880998135, + 0.008687131106853485, + 0.007331619970500469, + 0.005777838174253702, + 0.0068203844130039215, + 0.0072363982908427715, + 0.0063180518336594105, + 0.00590314669534564, + 0.0038885469548404217, + 0.0030474646482616663, + 0.0034529876429587603, + 0.0021062365267425776, + 0.002863127738237381, + 0.0038875502068549395, + 0.0017319259932264686, + 0.0023403738159686327, + 0.01720547117292881, + 0.05920782312750816, + 0.007228116970509291, + 0.006235597189515829, + 0.008078986778855324, + 0.008010041899979115, + 0.051889609545469284, + 0.010307006537914276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.03370098024606705 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive bioluminescent jellyfish as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish glowing with a soft blue-green light and featuring translucent parasols with visible veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jellyfish slowly circling the wreck of an ancient wooden sailing ship resting on the seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship wreck thickly coated with algae and barnacles, with a broken mast jutting into the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a long shot filmed from a slightly upward angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999458372592926 + ], + "opposite_logprob_tokens": [ + -2.6949946880340576 + ], + "logprob_reward": -0.06999458372592926, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999458372592926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07795663923025131, + 0.07795333117246628, + 0.040358055382966995, + 0.05554379150271416, + 0.07797404378652573, + 0.20092900097370148, + 0.08485075831413269, + 0.06982248276472092, + 0.02924427203834057, + 0.2853676378726959 + ], + "credit_map": [ + 0.01441800594329834, + 0.006484718527644873, + 0.01064709946513176, + 0.007064030505716801, + 0.005903780926018953, + 0.014424741268157959, + 0.00810646079480648, + 0.010907797142863274, + 0.010371723212301731, + 0.008170678280293941, + 0.01356586255133152, + 0.006765738595277071, + 0.0078037516213953495, + 0.013383558951318264, + 0.008694966323673725, + 0.009197047911584377, + 0.005304229445755482, + 0.0058636777102947235, + 0.004885983187705278, + 0.004016237333416939, + 0.0038075039628893137, + 0.004672129172831774, + 0.004765369929373264, + 0.007042920682579279, + 0.006758255884051323, + 0.0056909238919615746, + 0.012407563626766205, + 0.0050331586971879005, + 0.005383690353482962, + 0.008507603779435158, + 0.006396527402102947, + 0.005366064142435789, + 0.009710912592709064, + 0.009896053932607174, + 0.016462696716189384, + 0.010225704871118069, + 0.007441062014549971, + 0.008439118042588234, + 0.009298163466155529, + 0.0065003205090761185, + 0.021775025874376297, + 0.016878953203558922, + 0.05126431956887245, + 0.020929716527462006, + 0.012688485905528069, + 0.046664509922266006, + 0.016970956698060036, + 0.01375702302902937, + 0.012106886133551598, + 0.007534629199653864, + 0.021322526037693024, + 0.008422956801950932, + 0.007308660540729761, + 0.00750796590000391, + 0.0067337434738874435, + 0.013913385570049286, + 0.010338406078517437, + 0.00971735455095768, + 0.009840727783739567, + 0.007205088622868061, + 0.010371003299951553, + 0.007722211070358753, + 0.006940109189599752, + 0.007687577977776527, + 0.004160931333899498, + 0.003849767381325364, + 0.006160594057291746, + 0.003694579703733325, + 0.0031738963443785906, + 0.002778981113806367, + 0.0028322990983724594, + 0.002593220444396138, + 0.025083458051085472, + 0.09793470799922943, + 0.009852977469563484, + 0.013137511909008026, + 0.01338603813201189, + 0.009748422540724277, + 0.09782333672046661, + 0.018401185050606728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.06999458372592926 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_11/index_tensor([4907, 4907])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_11/index_tensor([4907, 4907])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6005850bcb8254c0e458c6483702ad8a388c25f5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_11/index_tensor([4907, 4907])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic 360-degree mid-shot, shot from a slightly upward angle, captures a heavily armored general. He wears a full-body suit of dark metal plate, scarred and weathered. The shoulder plates are engraved with roaring animal heads, reflecting a cold, hard sheen in the golden morning sunlight that filters through the mist. As the camera steadily and smoothly rotates around him, we see him amidst a thicket of waist-high, lush, emerald-green grass. He sinks his weight, his arms flexing with muscle, and with decisive and powerful force, he wields a massive, long-handled sword, its blade gleaming with a cold gleam. With a deep roar, he slashes forward with unstoppable force. The blade instantly tears through the green barrier before him, and in a slow-motion effect, severed blades of grass, glistening dewdrops, and moist earth explode in all directions, clearing a path for him to advance. The golden Jesus light outlines the contours of his armor and reflects a brief rainbow light on the splashing water droplets. The whole picture is full of epic determination and unstoppable momentum, creating a solemn and tense atmosphere.", + "seed": 5363, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_11/index_4907_seed_5363_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08147567035630346, + "vlm_logprob": -0.08147567035630346, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.08147567035630346, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily armored general wearing a full-body suit of dark metal plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a thicket of waist-high, lush, emerald-green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general wield a massive, long-handled sword and slash forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady, smooth 360-degree rotation around the general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a slow-motion effect showing severed grass, glistening dewdrops, and moist earth exploding in all directions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10345373302698135 + ], + "opposite_logprob_tokens": [ + -2.3534536361694336 + ], + "logprob_reward": -0.10345373302698135, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10345373302698135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15814431011676788, + 0.1093926802277565, + 0.044769976288080215, + 0.05583752691745758, + 0.09500834345817566, + 0.06183534115552902, + 0.06051216647028923, + 0.08611617237329483, + 0.20939704775810242, + 0.11898647993803024 + ], + "credit_map": [ + 0.012639695778489113, + 0.011193604208528996, + 0.04199060797691345, + 0.011081178672611713, + 0.009211123920977116, + 0.036977749317884445, + 0.015917625278234482, + 0.0191327054053545, + 0.012169700115919113, + 0.008318178355693817, + 0.02381867915391922, + 0.01279495656490326, + 0.011871572583913803, + 0.010256316512823105, + 0.01391494832932949, + 0.016248328611254692, + 0.003243545303121209, + 0.00540395500138402, + 0.007109952624887228, + 0.006487801671028137, + 0.0059357997961342335, + 0.004357048310339451, + 0.007252868264913559, + 0.004979003220796585, + 0.008262303657829762, + 0.010095831006765366, + 0.007547538727521896, + 0.00696784071624279, + 0.004189172759652138, + 0.004712319001555443, + 0.005043805576860905, + 0.00901870895177126, + 0.009435019455850124, + 0.00863732025027275, + 0.008150496520102024, + 0.025958774611353874, + 0.010142029263079166, + 0.01011444441974163, + 0.014415045268833637, + 0.008155208081007004, + 0.009840654209256172, + 0.005956928711384535, + 0.008674297481775284, + 0.004981221165508032, + 0.011493140831589699, + 0.010778754018247128, + 0.0066489470191299915, + 0.0034613923635333776, + 0.007123436313122511, + 0.0075185662135481834, + 0.0072222803719341755, + 0.006483855191618204, + 0.014696703292429447, + 0.006631426978856325, + 0.005747625604271889, + 0.005088267847895622, + 0.011113173328340054, + 0.018109450116753578, + 0.007704642601311207, + 0.009571864269673824, + 0.011732831597328186, + 0.011593547649681568, + 0.008935442194342613, + 0.007355216890573502, + 0.01362327579408884, + 0.011837729252874851, + 0.06591649353504181, + 0.04438936710357666, + 0.011277059093117714, + 0.03336859121918678, + 0.015569326467812061, + 0.013415182009339333, + 0.010262267664074898, + 0.01088016014546156, + 0.02192198671400547, + 0.01571749709546566, + 0.011653458699584007, + 0.007426407188177109, + 0.027366921305656433, + 0.013757768087089062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.10345373302698135 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily armored general wearing a full-body suit of dark metal plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a thicket of waist-high, lush, emerald-green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general wield a massive, long-handled sword and slash forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady, smooth 360-degree rotation around the general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a slow-motion effect showing severed grass, glistening dewdrops, and moist earth exploding in all directions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005954858381301165 + ], + "opposite_logprob_tokens": [ + -5.130954742431641 + ], + "logprob_reward": -0.005954858381301165, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005954858381301165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11594486236572266, + 0.03930000960826874, + 0.09070073813199997, + 0.12407013028860092, + 0.064681775867939, + 0.10233712196350098, + 0.030200233682990074, + 0.09345410764217377, + 0.31349554657936096, + 0.02581542730331421 + ], + "credit_map": [ + 0.01051250472664833, + 0.008256139233708382, + 0.028184015303850174, + 0.006128367967903614, + 0.007034512236714363, + 0.02074115164577961, + 0.011256887577474117, + 0.023831289261579514, + 0.003593857167288661, + 0.004389788955450058, + 0.006656981538981199, + 0.004197361879050732, + 0.004608067683875561, + 0.003226000349968672, + 0.005868602078408003, + 0.00675934785977006, + 0.004802417941391468, + 0.006750208791345358, + 0.015319614671170712, + 0.018352219834923744, + 0.008078637532889843, + 0.009158109314739704, + 0.018311278894543648, + 0.009928245097398758, + 0.012656762264668941, + 0.01574993133544922, + 0.01700310967862606, + 0.017394891008734703, + 0.009875771589577198, + 0.013233808800578117, + 0.0103001082316041, + 0.027855750173330307, + 0.007877347990870476, + 0.005796000827103853, + 0.00903493631631136, + 0.010566849261522293, + 0.005372013431042433, + 0.008210808970034122, + 0.006488969549536705, + 0.011334851384162903, + 0.015355519019067287, + 0.010002566501498222, + 0.023116078227758408, + 0.011342564597725868, + 0.014698714017868042, + 0.012250007130205631, + 0.008909539319574833, + 0.00666212709620595, + 0.004117180593311787, + 0.005098565947264433, + 0.003903825767338276, + 0.003552968380972743, + 0.005836911965161562, + 0.0035812368150800467, + 0.002163868397474289, + 0.0019456760492175817, + 0.009559872560203075, + 0.02063295803964138, + 0.008411800488829613, + 0.0089424978941679, + 0.01788271963596344, + 0.009151210077106953, + 0.00942491739988327, + 0.009448136202991009, + 0.02001822739839554, + 0.02073916792869568, + 0.09708797186613083, + 0.06749703735113144, + 0.028018806129693985, + 0.02916068397462368, + 0.028563586995005608, + 0.022410038858652115, + 0.0034968298859894276, + 0.0034499415196478367, + 0.004612176213413477, + 0.0031915612053126097, + 0.0030548502691090107, + 0.0015125783393159509, + 0.0026956808287650347, + 0.003801809623837471 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.005954858381301165 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily armored general wearing a full-body suit of dark metal plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a thicket of waist-high, lush, emerald-green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general wield a massive, long-handled sword and slash forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady, smooth 360-degree rotation around the general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a slow-motion effect showing severed grass, glistening dewdrops, and moist earth exploding in all directions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029804455116391182 + ], + "opposite_logprob_tokens": [ + -3.5298044681549072 + ], + "logprob_reward": -0.029804455116391182, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029804455116391182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11922560632228851, + 0.05862785130739212, + 0.04557649791240692, + 0.04184528812766075, + 0.0710359588265419, + 0.04205233231186867, + 0.06954635679721832, + 0.09549175947904587, + 0.2496025115251541, + 0.20699578523635864 + ], + "credit_map": [ + 0.008877036161720753, + 0.012311327271163464, + 0.025459732860326767, + 0.00882194098085165, + 0.010010965168476105, + 0.025853851810097694, + 0.015347150154411793, + 0.01254361029714346, + 0.005569548346102238, + 0.004976088181138039, + 0.010683002881705761, + 0.007747206836938858, + 0.006857698783278465, + 0.004324766807258129, + 0.009815476834774017, + 0.008654068224132061, + 0.004278163425624371, + 0.004671110305935144, + 0.008249196223914623, + 0.004116871859878302, + 0.0049241939559578896, + 0.005362892057746649, + 0.00817776657640934, + 0.005796304438263178, + 0.004477009642869234, + 0.005709860008209944, + 0.005432469304651022, + 0.006887470372021198, + 0.004951829556375742, + 0.0040738447569310665, + 0.006183111108839512, + 0.004129697568714619, + 0.008254156447947025, + 0.008569547906517982, + 0.006300177425146103, + 0.01824982278048992, + 0.007614273112267256, + 0.006398236379027367, + 0.009875806048512459, + 0.005773947574198246, + 0.005070124287158251, + 0.004742854740470648, + 0.004518514033406973, + 0.004985058680176735, + 0.00979635864496231, + 0.005858235061168671, + 0.003966167103499174, + 0.003115021390840411, + 0.008476978167891502, + 0.006848983932286501, + 0.007539356127381325, + 0.009671416133642197, + 0.014263144694268703, + 0.007152447011321783, + 0.008652809076011181, + 0.006941229570657015, + 0.011333408765494823, + 0.015934329479932785, + 0.010791542939841747, + 0.011517094448208809, + 0.012639844790101051, + 0.013070705346763134, + 0.010453053750097752, + 0.009751781821250916, + 0.014178290963172913, + 0.014569719322025776, + 0.08421861380338669, + 0.05647451430559158, + 0.022170986980199814, + 0.024781078100204468, + 0.01686864346265793, + 0.016340678557753563, + 0.02967946231365204, + 0.04469219222664833, + 0.022956322878599167, + 0.022954514250159264, + 0.032732121646404266, + 0.013593149371445179, + 0.023786650970578194, + 0.016601387411355972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.029804455116391182 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily armored general wearing a full-body suit of dark metal plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a thicket of waist-high, lush, emerald-green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general wield a massive, long-handled sword and slash forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady, smooth 360-degree rotation around the general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a slow-motion effect showing severed grass, glistening dewdrops, and moist earth exploding in all directions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25209248065948486 + ], + "opposite_logprob_tokens": [ + -1.5020924806594849 + ], + "logprob_reward": -0.25209248065948486, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25209248065948486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10416410863399506, + 0.09558006376028061, + 0.038653407245874405, + 0.12159927189350128, + 0.06911734491586685, + 0.10673727840185165, + 0.048089347779750824, + 0.11496112495660782, + 0.23356595635414124, + 0.06753209233283997 + ], + "credit_map": [ + 0.008361965417861938, + 0.009364677593111992, + 0.0256744883954525, + 0.007664097473025322, + 0.010395442135632038, + 0.02338266745209694, + 0.010012098588049412, + 0.00930867251008749, + 0.008296268992125988, + 0.00673700449988246, + 0.01533188484609127, + 0.0078601548448205, + 0.010406709276139736, + 0.01716054417192936, + 0.0129309156909585, + 0.016856584697961807, + 0.0028478994499891996, + 0.004006382543593645, + 0.006268199533224106, + 0.005215322133153677, + 0.005200626328587532, + 0.003288122359663248, + 0.006903430912643671, + 0.004923421889543533, + 0.01683065854012966, + 0.020775379613041878, + 0.017406338825821877, + 0.01900944672524929, + 0.01118911150842905, + 0.011325663886964321, + 0.012518243864178658, + 0.012544428929686546, + 0.008463364094495773, + 0.004697449505329132, + 0.012252035550773144, + 0.009538305923342705, + 0.005723109934478998, + 0.00866528321057558, + 0.0072487397119402885, + 0.01252905372530222, + 0.02110791578888893, + 0.012236271053552628, + 0.02171332947909832, + 0.011789169162511826, + 0.012721341103315353, + 0.011921421624720097, + 0.008347575552761555, + 0.006900258827954531, + 0.0051975068636238575, + 0.008375335484743118, + 0.006180782336741686, + 0.006975471507757902, + 0.007855363190174103, + 0.005307542625814676, + 0.003851603716611862, + 0.00434574531391263, + 0.01147717796266079, + 0.027737831696867943, + 0.009467135183513165, + 0.009990396909415722, + 0.023100150749087334, + 0.011262756772339344, + 0.010651194490492344, + 0.011274493299424648, + 0.01984235644340515, + 0.020781485363841057, + 0.054893914610147476, + 0.04617501050233841, + 0.023875880986452103, + 0.020742597058415413, + 0.02665247581899166, + 0.02060222439467907, + 0.010369237512350082, + 0.009247350506484509, + 0.010335025377571583, + 0.008160681463778019, + 0.009338580071926117, + 0.004526756703853607, + 0.006314112339168787, + 0.009240346029400826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 7 + ], + "selected_reward": -0.25209248065948486 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily armored general wearing a full-body suit of dark metal plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a thicket of waist-high, lush, emerald-green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general wield a massive, long-handled sword and slash forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady, smooth 360-degree rotation around the general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a slow-motion effect showing severed grass, glistening dewdrops, and moist earth exploding in all directions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016072824597358704 + ], + "opposite_logprob_tokens": [ + -4.141072750091553 + ], + "logprob_reward": -0.016072824597358704, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016072824597358704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09487561136484146, + 0.08923246711492538, + 0.11081036925315857, + 0.04800436273217201, + 0.10817108303308487, + 0.0921010971069336, + 0.02710839733481407, + 0.11913984268903732, + 0.2154163271188736, + 0.09514039009809494 + ], + "credit_map": [ + 0.009574518539011478, + 0.01183276902884245, + 0.013571243733167648, + 0.007665430661290884, + 0.007898536510765553, + 0.01177198439836502, + 0.011433831416070461, + 0.02112729847431183, + 0.006938324309885502, + 0.008073828183114529, + 0.014716883189976215, + 0.008046738803386688, + 0.01274814736098051, + 0.011150357313454151, + 0.013063748367130756, + 0.01449443306773901, + 0.008114735595881939, + 0.010236062109470367, + 0.01821422018110752, + 0.020538827404379845, + 0.01272447220981121, + 0.009943212382495403, + 0.0194895938038826, + 0.01154924277216196, + 0.006794189568608999, + 0.009306690655648708, + 0.005592033267021179, + 0.006493753287941217, + 0.004010346718132496, + 0.004316115751862526, + 0.004571837373077869, + 0.006919398438185453, + 0.012340720742940903, + 0.012227033264935017, + 0.013696921057999134, + 0.02273080311715603, + 0.012512242421507835, + 0.010659395717084408, + 0.013029024936258793, + 0.010974946431815624, + 0.014659845270216465, + 0.010431395843625069, + 0.019352523609995842, + 0.009059278294444084, + 0.013149432837963104, + 0.01235225610435009, + 0.00789378397166729, + 0.0052025774493813515, + 0.0032046923879534006, + 0.004997050855308771, + 0.0034165068063884974, + 0.0028043657075613737, + 0.005680822301656008, + 0.0028025025967508554, + 0.0023869387805461884, + 0.00181551743298769, + 0.015289919450879097, + 0.016609033569693565, + 0.016504274681210518, + 0.016566485166549683, + 0.011907922104001045, + 0.0161238182336092, + 0.010910088196396828, + 0.01522829756140709, + 0.012940794229507446, + 0.01327332854270935, + 0.07202276587486267, + 0.04786084592342377, + 0.018269527703523636, + 0.019849393516778946, + 0.014792008325457573, + 0.016407668590545654, + 0.01734703779220581, + 0.013079638592898846, + 0.01090901531279087, + 0.009047139436006546, + 0.010926347225904465, + 0.007092835381627083, + 0.012037906795740128, + 0.014700471423566341 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 2 + ], + "selected_reward": -0.016072824597358704 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic 360-degree mid-shot, shot from a slightly upward angle, captures a heavily armored general. He wears a full-body suit of dark metal plate, scarred and weathered. The shoulder plates are engraved with roaring animal heads, reflecting a cold, hard sheen in the golden morning sunlight that filters through the mist. As the camera steadily and smoothly rotates around him, we see him amidst a thicket of waist-high, lush, emerald-green grass. He sinks his weight, his arms flexing with muscle, and with decisive and powerful force, he wields a massive, long-handled sword, its blade gleaming with a cold gleam. With a deep roar, he slashes forward with unstoppable force. The blade instantly tears through the green barrier before him, and in a slow-motion effect, severed blades of grass, glistening dewdrops, and moist earth explode in all directions, clearing a path for him to advance. The golden Jesus light outlines the contours of his armor and reflects a brief rainbow light on the splashing water droplets. The whole picture is full of epic determination and unstoppable momentum, creating a solemn and tense atmosphere.", + "seed": 5364, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_11/index_4907_seed_5364_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0500339912250638, + "vlm_logprob": -0.0500339912250638, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.0500339912250638, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily armored general wearing a full-body suit of dark metal plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a thicket of waist-high, lush, emerald-green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general wield a massive, long-handled sword and slash forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady, smooth 360-degree rotation around the general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a slow-motion effect showing severed grass, glistening dewdrops, and moist earth exploding in all directions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05905262753367424 + ], + "opposite_logprob_tokens": [ + -2.9340527057647705 + ], + "logprob_reward": -0.05905262753367424, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05905262753367424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06806176900863647, + 0.08141853660345078, + 0.04400121793150902, + 0.14664515852928162, + 0.25173917412757874, + 0.04892837628722191, + 0.07293068617582321, + 0.18363885581493378, + 0.06568849831819534, + 0.03694770485162735 + ], + "credit_map": [ + 0.007755996659398079, + 0.008849747478961945, + 0.007514027878642082, + 0.009487473405897617, + 0.006998430471867323, + 0.008248223923146725, + 0.006688638124614954, + 0.012519233860075474, + 0.018759597092866898, + 0.005028554238379002, + 0.006395560223609209, + 0.008491934277117252, + 0.005682672373950481, + 0.02498345635831356, + 0.006424306891858578, + 0.005652459803968668, + 0.00494034867733717, + 0.004718211945146322, + 0.007089647930115461, + 0.004353207070380449, + 0.005312612745910883, + 0.0069694602862000465, + 0.0063988566398620605, + 0.0042188698425889015, + 0.04008930176496506, + 0.011965381912887096, + 0.0069452510215342045, + 0.016989298164844513, + 0.00919764768332243, + 0.049066148698329926, + 0.0076301065273582935, + 0.004762019496411085, + 0.026980260387063026, + 0.01926400139927864, + 0.0705375075340271, + 0.02272680588066578, + 0.028834303840994835, + 0.020599937066435814, + 0.01341771800071001, + 0.0493786484003067, + 0.007767288014292717, + 0.004532665014266968, + 0.009255694225430489, + 0.005873841233551502, + 0.005578408483415842, + 0.005150873679667711, + 0.0036022188141942024, + 0.007167389150708914, + 0.014112783595919609, + 0.016165584325790405, + 0.007256751414388418, + 0.008067233487963676, + 0.007744633127003908, + 0.006955026648938656, + 0.007669015787541866, + 0.004959654528647661, + 0.01632043905556202, + 0.008973103947937489, + 0.04458886384963989, + 0.012369714677333832, + 0.015409919433295727, + 0.06013522669672966, + 0.013915262185037136, + 0.011926316656172276, + 0.005800141952931881, + 0.009156575426459312, + 0.00476848054677248, + 0.011126161552965641, + 0.008930870331823826, + 0.005713082384318113, + 0.011060898192226887, + 0.009132285602390766, + 0.006143213249742985, + 0.003164176130667329, + 0.0025000760797411203, + 0.0023904319386929274, + 0.004000145476311445, + 0.01039959117770195, + 0.005397718865424395, + 0.0029523528646677732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 3 + ], + "selected_reward": -0.05905262753367424 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily armored general wearing a full-body suit of dark metal plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a thicket of waist-high, lush, emerald-green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general wield a massive, long-handled sword and slash forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady, smooth 360-degree rotation around the general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a slow-motion effect showing severed grass, glistening dewdrops, and moist earth exploding in all directions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005956161767244339 + ], + "opposite_logprob_tokens": [ + -5.130956172943115 + ], + "logprob_reward": -0.005956161767244339, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005956161767244339, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10056991875171661, + 0.08618739992380142, + 0.041731178760528564, + 0.3551441729068756, + 0.038177490234375, + 0.049235858023166656, + 0.09616731107234955, + 0.11380194872617722, + 0.04856017976999283, + 0.07042456418275833 + ], + "credit_map": [ + 0.010960583575069904, + 0.008196400478482246, + 0.007084863260388374, + 0.010585615411400795, + 0.011929761618375778, + 0.015249302610754967, + 0.013114441186189651, + 0.023448949679732323, + 0.021983535960316658, + 0.00530234444886446, + 0.004749256651848555, + 0.005194003228098154, + 0.004443230573087931, + 0.0283636637032032, + 0.006463265977799892, + 0.009688098914921284, + 0.004370200913399458, + 0.004532486200332642, + 0.0045383041724562645, + 0.005452009383589029, + 0.005421960726380348, + 0.006473426707088947, + 0.005909982603043318, + 0.005032810382544994, + 0.1273774951696396, + 0.010615040548145771, + 0.007804639171808958, + 0.02833985723555088, + 0.01629478484392166, + 0.1474171280860901, + 0.008125356398522854, + 0.009169881232082844, + 0.004269683267921209, + 0.004237235989421606, + 0.0073492517694830894, + 0.00497845746576786, + 0.005168666131794453, + 0.0033094591926783323, + 0.004370064474642277, + 0.004494672175496817, + 0.006262511480599642, + 0.0046860575675964355, + 0.007855222560465336, + 0.006130545865744352, + 0.005708359181880951, + 0.007379566784948111, + 0.004527153912931681, + 0.006686438340693712, + 0.017580468207597733, + 0.01564856991171837, + 0.01209243107587099, + 0.01160638127475977, + 0.00925802905112505, + 0.011725586839020252, + 0.011030999943614006, + 0.007224844302982092, + 0.012149626389145851, + 0.00481656938791275, + 0.029960691928863525, + 0.00640344200655818, + 0.008397354744374752, + 0.03489864617586136, + 0.009320573881268501, + 0.007855040952563286, + 0.005113951396197081, + 0.009130770340561867, + 0.003981455694884062, + 0.007818534038960934, + 0.005925041623413563, + 0.005577066447585821, + 0.00634941179305315, + 0.00466395216062665, + 0.01054034661501646, + 0.006688435096293688, + 0.004812753293663263, + 0.005601552315056324, + 0.008169273845851421, + 0.01698160544037819, + 0.010881886817514896, + 0.00674870889633894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.005956161767244339 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily armored general wearing a full-body suit of dark metal plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a thicket of waist-high, lush, emerald-green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general wield a massive, long-handled sword and slash forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady, smooth 360-degree rotation around the general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a slow-motion effect showing severed grass, glistening dewdrops, and moist earth exploding in all directions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029812784865498543 + ], + "opposite_logprob_tokens": [ + -3.529812812805176 + ], + "logprob_reward": -0.029812784865498543, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029812784865498543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028959084302186966, + 0.11288030445575714, + 0.033513884991407394, + 0.38653039932250977, + 0.04283025860786438, + 0.07824933528900146, + 0.19268010556697845, + 0.023688508197665215, + 0.038008205592632294, + 0.0626598447561264 + ], + "credit_map": [ + 0.004272809252142906, + 0.0031592529267072678, + 0.003045350778847933, + 0.0033675318118184805, + 0.00330416695214808, + 0.004119144752621651, + 0.002834135899320245, + 0.004856694024056196, + 0.034230414777994156, + 0.006466600112617016, + 0.0044389646500349045, + 0.003948770463466644, + 0.0038554647471755743, + 0.04405021294951439, + 0.006339428015053272, + 0.009550467133522034, + 0.003260453464463353, + 0.004226091783493757, + 0.0043633002787828445, + 0.004886462353169918, + 0.004673361778259277, + 0.004400196950882673, + 0.004645520821213722, + 0.0030585015192627907, + 0.13969875872135162, + 0.011626733466982841, + 0.00788536574691534, + 0.03127303719520569, + 0.016848839819431305, + 0.15935635566711426, + 0.009806979447603226, + 0.010034344159066677, + 0.005800146609544754, + 0.003531204303726554, + 0.010983454994857311, + 0.004768419079482555, + 0.00577381020411849, + 0.003516479628160596, + 0.0045901197008788586, + 0.00386662851087749, + 0.01577422395348549, + 0.007258597761392593, + 0.017216283828020096, + 0.006979151628911495, + 0.007332080975174904, + 0.006605557631701231, + 0.004732177592813969, + 0.012351269833743572, + 0.03601621091365814, + 0.04335559159517288, + 0.02105339989066124, + 0.023625575006008148, + 0.017294995486736298, + 0.019402094185352325, + 0.020057884976267815, + 0.011874365620315075, + 0.003116145497187972, + 0.0014606041368097067, + 0.0041321320459246635, + 0.001687289564870298, + 0.00259519275277853, + 0.006343574728816748, + 0.00232935743406415, + 0.002024211920797825, + 0.0038023910019546747, + 0.00797068141400814, + 0.002780935959890485, + 0.00584430480375886, + 0.004818931221961975, + 0.0040215011686086655, + 0.0048876260407269, + 0.003881836775690317, + 0.010581077076494694, + 0.0032169814221560955, + 0.004043544642627239, + 0.0029257903806865215, + 0.005234438460320234, + 0.022958489134907722, + 0.009368754923343658, + 0.004330775700509548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 1 + ], + "selected_reward": -0.029812784865498543 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily armored general wearing a full-body suit of dark metal plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a thicket of waist-high, lush, emerald-green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general wield a massive, long-handled sword and slash forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady, smooth 360-degree rotation around the general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a slow-motion effect showing severed grass, glistening dewdrops, and moist earth exploding in all directions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14280490577220917 + ], + "opposite_logprob_tokens": [ + -2.0178048610687256 + ], + "logprob_reward": -0.14280490577220917, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14280490577220917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04079166054725647, + 0.2106497883796692, + 0.03587768226861954, + 0.1818990558385849, + 0.09131746739149094, + 0.1888273060321808, + 0.09264418482780457, + 0.057646188884973526, + 0.06683055311441422, + 0.033516090363264084 + ], + "credit_map": [ + 0.005434573162347078, + 0.0065369559451937675, + 0.003227356355637312, + 0.004910506773740053, + 0.0050977496430277824, + 0.005729583092033863, + 0.004311942495405674, + 0.005542994011193514, + 0.0572529137134552, + 0.012396706268191338, + 0.011307157576084137, + 0.013611392118036747, + 0.01151531282812357, + 0.0753403827548027, + 0.011715298518538475, + 0.017510615289211273, + 0.0032766235526651144, + 0.00527748791500926, + 0.004502161405980587, + 0.0034153463784605265, + 0.004961621481925249, + 0.0050017680041491985, + 0.00663395132869482, + 0.0028087219689041376, + 0.062013667076826096, + 0.007764528505504131, + 0.0056923688389360905, + 0.016704512760043144, + 0.009081532247364521, + 0.07094863057136536, + 0.004933055490255356, + 0.004760763142257929, + 0.02571892738342285, + 0.0053761666640639305, + 0.021939003840088844, + 0.007348211947828531, + 0.00469624437391758, + 0.005424618721008301, + 0.005576979834586382, + 0.01523731742054224, + 0.03627266734838486, + 0.012861823663115501, + 0.033993467688560486, + 0.021046536043286324, + 0.016849955543875694, + 0.02334795892238617, + 0.015164656564593315, + 0.02929024212062359, + 0.01924949698150158, + 0.019965611398220062, + 0.009218640625476837, + 0.010145852342247963, + 0.008664360269904137, + 0.009654470719397068, + 0.009197628125548363, + 0.006548122502863407, + 0.005717406515032053, + 0.0029105334542691708, + 0.014551372267305851, + 0.0032106845173984766, + 0.004597090184688568, + 0.01781599037349224, + 0.004410061053931713, + 0.004433049820363522, + 0.007145406678318977, + 0.011694977059960365, + 0.005286591127514839, + 0.00881018117070198, + 0.008208826184272766, + 0.006859010551124811, + 0.010219449177384377, + 0.008606108836829662, + 0.0054082064889371395, + 0.0022061853669583797, + 0.004790324717760086, + 0.0019928677938878536, + 0.0026849349960684776, + 0.010676942765712738, + 0.0038569283206015825, + 0.0018996944418177009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 3 + ], + "selected_reward": -0.14280490577220917 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily armored general wearing a full-body suit of dark metal plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a thicket of waist-high, lush, emerald-green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general wield a massive, long-handled sword and slash forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady, smooth 360-degree rotation around the general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a slow-motion effect showing severed grass, glistening dewdrops, and moist earth exploding in all directions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543476186692715 + ], + "opposite_logprob_tokens": [ + -4.387543678283691 + ], + "logprob_reward": -0.012543476186692715, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543476186692715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09986601024866104, + 0.08509016036987305, + 0.1462061107158661, + 0.09727136790752411, + 0.07924383133649826, + 0.21642445027828217, + 0.0643945038318634, + 0.061972856521606445, + 0.05531623587012291, + 0.09421440958976746 + ], + "credit_map": [ + 0.014604066498577595, + 0.009792035445570946, + 0.008081072010099888, + 0.010705576278269291, + 0.010149452835321426, + 0.013628587126731873, + 0.00980291049927473, + 0.023102309554815292, + 0.017356576398015022, + 0.009012472815811634, + 0.00757979741320014, + 0.006949649192392826, + 0.007266907021403313, + 0.020900681614875793, + 0.00814012996852398, + 0.007883948273956776, + 0.013426586054265499, + 0.018637796863913536, + 0.018164539709687233, + 0.019123036414384842, + 0.022993141785264015, + 0.017669009044766426, + 0.023376625031232834, + 0.012815396301448345, + 0.01611611805856228, + 0.011448120698332787, + 0.007374719250947237, + 0.017006268724799156, + 0.009866239503026009, + 0.020407745614647865, + 0.010253476910293102, + 0.004798682872205973, + 0.019963650032877922, + 0.0036293205339461565, + 0.02135126106441021, + 0.007847955450415611, + 0.005221065133810043, + 0.006023828871548176, + 0.005002407822757959, + 0.01020433846861124, + 0.03548036888241768, + 0.01883849687874317, + 0.03961015120148659, + 0.024165062233805656, + 0.023205768316984177, + 0.027203800156712532, + 0.019328603520989418, + 0.02859221212565899, + 0.01024426706135273, + 0.009865174070000648, + 0.006888906937092543, + 0.007576841861009598, + 0.007061512675136328, + 0.008787605911493301, + 0.008239436894655228, + 0.005730758886784315, + 0.007878197357058525, + 0.003478421363979578, + 0.014834580011665821, + 0.0035225381143391132, + 0.0052764397114515305, + 0.01767938770353794, + 0.004401687532663345, + 0.004901604261249304, + 0.0055804066359996796, + 0.00641328189522028, + 0.005726121366024017, + 0.011494367383420467, + 0.006589903961867094, + 0.004516365472227335, + 0.008003005757927895, + 0.006992781534790993, + 0.011852146126329899, + 0.013478168286383152, + 0.008690400049090385, + 0.011276273056864738, + 0.012660793028771877, + 0.010446685366332531, + 0.017214570194482803, + 0.008595375344157219 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.012543476186692715 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic 360-degree mid-shot, shot from a slightly upward angle, captures a heavily armored general. He wears a full-body suit of dark metal plate, scarred and weathered. The shoulder plates are engraved with roaring animal heads, reflecting a cold, hard sheen in the golden morning sunlight that filters through the mist. As the camera steadily and smoothly rotates around him, we see him amidst a thicket of waist-high, lush, emerald-green grass. He sinks his weight, his arms flexing with muscle, and with decisive and powerful force, he wields a massive, long-handled sword, its blade gleaming with a cold gleam. With a deep roar, he slashes forward with unstoppable force. The blade instantly tears through the green barrier before him, and in a slow-motion effect, severed blades of grass, glistening dewdrops, and moist earth explode in all directions, clearing a path for him to advance. The golden Jesus light outlines the contours of his armor and reflects a brief rainbow light on the splashing water droplets. The whole picture is full of epic determination and unstoppable momentum, creating a solemn and tense atmosphere.", + "seed": 5365, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_11/index_4907_seed_5365_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.14533967114984989, + "vlm_logprob": -0.14533967114984989, + "vlm_yes_no_margin": 2.950000011920929, + "vlm_token_credit": -0.14533967114984989, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily armored general wearing a full-body suit of dark metal plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a thicket of waist-high, lush, emerald-green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general wield a massive, long-handled sword and slash forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady, smooth 360-degree rotation around the general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a slow-motion effect showing severed grass, glistening dewdrops, and moist earth exploding in all directions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3902530074119568 + ], + "opposite_logprob_tokens": [ + -1.1402530670166016 + ], + "logprob_reward": -0.3902530074119568, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.3902530074119568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17930875718593597, + 0.0882631167769432, + 0.07554991543292999, + 0.12158402055501938, + 0.07307180017232895, + 0.18476133048534393, + 0.09572911262512207, + 0.02184233069419861, + 0.03718644008040428, + 0.1227031797170639 + ], + "credit_map": [ + 0.02049485594034195, + 0.015820227563381195, + 0.02148239128291607, + 0.01731909066438675, + 0.021613532677292824, + 0.024777457118034363, + 0.024542182683944702, + 0.03325901925563812, + 0.008457253687083721, + 0.007253194693475962, + 0.009037814103066921, + 0.007228789385408163, + 0.009275313466787338, + 0.01818806678056717, + 0.007726472802460194, + 0.021096209064126015, + 0.010137889534235, + 0.008419277146458626, + 0.008480952121317387, + 0.012542614713311195, + 0.009633757174015045, + 0.008556360378861427, + 0.011194952763617039, + 0.006584116257727146, + 0.018392208963632584, + 0.011907038278877735, + 0.02279227413237095, + 0.013751586899161339, + 0.015065609477460384, + 0.011286801658570766, + 0.019444776698946953, + 0.008943718858063221, + 0.009448380209505558, + 0.008161040022969246, + 0.010950836353003979, + 0.007716802414506674, + 0.011480883695185184, + 0.007748173084110022, + 0.008152151480317116, + 0.009413531981408596, + 0.018215367570519447, + 0.027513325214385986, + 0.01926216296851635, + 0.020247064530849457, + 0.023485103622078896, + 0.019954031333327293, + 0.033489663153886795, + 0.022594597190618515, + 0.009164447896182537, + 0.007474821526557207, + 0.011344184167683125, + 0.008675987832248211, + 0.009299942292273045, + 0.018339617177844048, + 0.020811738446354866, + 0.01061837188899517, + 0.0027925579342991114, + 0.002214191248640418, + 0.004199060145765543, + 0.0035047978162765503, + 0.0027067046612501144, + 0.0018869128543883562, + 0.0024003374855965376, + 0.002137767383828759, + 0.0062725539319217205, + 0.004819485358893871, + 0.004891790449619293, + 0.004037941340357065, + 0.004511906765401363, + 0.004357784986495972, + 0.0038852759171277285, + 0.004409703891724348, + 0.033277157694101334, + 0.010857061482965946, + 0.014305598102509975, + 0.009802502579987049, + 0.009850172325968742, + 0.03157675638794899, + 0.006632429081946611, + 0.006401505321264267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.3902530074119568 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily armored general wearing a full-body suit of dark metal plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a thicket of waist-high, lush, emerald-green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general wield a massive, long-handled sword and slash forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady, smooth 360-degree rotation around the general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a slow-motion effect showing severed grass, glistening dewdrops, and moist earth exploding in all directions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005954621359705925 + ], + "opposite_logprob_tokens": [ + -5.130954742431641 + ], + "logprob_reward": -0.005954621359705925, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005954621359705925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15984401106834412, + 0.06782694160938263, + 0.038385357707738876, + 0.1358395218849182, + 0.03469162434339523, + 0.13764779269695282, + 0.2699306309223175, + 0.06031986325979233, + 0.027161424979567528, + 0.0683528259396553 + ], + "credit_map": [ + 0.018738888204097748, + 0.012266268953680992, + 0.04184941574931145, + 0.011623269878327847, + 0.014890038408339024, + 0.01947716437280178, + 0.024899974465370178, + 0.016098981723189354, + 0.010079121217131615, + 0.0074149733409285545, + 0.01015669945627451, + 0.008030883967876434, + 0.008665322326123714, + 0.008397459983825684, + 0.0060756211169064045, + 0.009006861597299576, + 0.00469621317461133, + 0.003591266693547368, + 0.005211851559579372, + 0.00393499992787838, + 0.00552051467821002, + 0.007151428144425154, + 0.003981940448284149, + 0.004297143779695034, + 0.01675826869904995, + 0.013802425004541874, + 0.02983451634645462, + 0.016562143340706825, + 0.015859026461839676, + 0.012998330406844616, + 0.021583780646324158, + 0.008441019803285599, + 0.00444045290350914, + 0.0031858577858656645, + 0.005278988741338253, + 0.0036701876670122147, + 0.003063285956159234, + 0.004493797197937965, + 0.004639260936528444, + 0.005919792223721743, + 0.01941688172519207, + 0.016472116112709045, + 0.010446085594594479, + 0.013557583093643188, + 0.01571742445230484, + 0.014805622398853302, + 0.015991413965821266, + 0.031240670010447502, + 0.02130153775215149, + 0.014838048256933689, + 0.034003790467977524, + 0.017730671912431717, + 0.01849954202771187, + 0.062094032764434814, + 0.07850483804941177, + 0.022958174347877502, + 0.011559933423995972, + 0.012276124209165573, + 0.007929838262498379, + 0.006659151054918766, + 0.005211338866502047, + 0.0053072054870426655, + 0.005363035947084427, + 0.006013237405568361, + 0.00370712298899889, + 0.0023111943155527115, + 0.0031996096950024366, + 0.002904570894315839, + 0.003412682330235839, + 0.0027265686076134443, + 0.00393845047801733, + 0.004961224272847176, + 0.01726154424250126, + 0.006252765655517578, + 0.007614749018102884, + 0.005041587166488171, + 0.005731500219553709, + 0.01756136305630207, + 0.004794493317604065, + 0.004094823729246855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.005954621359705925 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily armored general wearing a full-body suit of dark metal plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a thicket of waist-high, lush, emerald-green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general wield a massive, long-handled sword and slash forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady, smooth 360-degree rotation around the general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a slow-motion effect showing severed grass, glistening dewdrops, and moist earth exploding in all directions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809879347681999 + ], + "opposite_logprob_tokens": [ + -3.2880988121032715 + ], + "logprob_reward": -0.03809879347681999, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809879347681999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1291552633047104, + 0.03651370480656624, + 0.06303362548351288, + 0.09613915532827377, + 0.056495752185583115, + 0.16299736499786377, + 0.33226197957992554, + 0.0230404082685709, + 0.05235718935728073, + 0.04800558090209961 + ], + "credit_map": [ + 0.016384830698370934, + 0.013116533868014812, + 0.0177842378616333, + 0.016773153096437454, + 0.01569581963121891, + 0.015628820285201073, + 0.01954205147922039, + 0.014229810796678066, + 0.003857884556055069, + 0.003935554064810276, + 0.0035461317747831345, + 0.004571470431983471, + 0.0054779257625341415, + 0.0055983178317546844, + 0.002962690545246005, + 0.006563728209584951, + 0.007761252112686634, + 0.006393535528331995, + 0.010572236031293869, + 0.010624067857861519, + 0.00744757940992713, + 0.007945080287754536, + 0.006994444411247969, + 0.005295427516102791, + 0.00948568619787693, + 0.009837419725954533, + 0.018025772646069527, + 0.012461544014513493, + 0.011100211180746555, + 0.01148136891424656, + 0.01679624803364277, + 0.006950907409191132, + 0.005873825866729021, + 0.005384615156799555, + 0.009283758699893951, + 0.005998410750180483, + 0.009288022294640541, + 0.00606822595000267, + 0.006367796566337347, + 0.00823109783232212, + 0.014920532703399658, + 0.025563281029462814, + 0.01756696030497551, + 0.018860843032598495, + 0.022123293951153755, + 0.020606007426977158, + 0.026996204629540443, + 0.016360249370336533, + 0.02228700928390026, + 0.014461175538599491, + 0.04265045002102852, + 0.021686861291527748, + 0.019131451845169067, + 0.0789014995098114, + 0.10489105433225632, + 0.028252476826310158, + 0.002889958443120122, + 0.0035280657466500998, + 0.003611898049712181, + 0.0021824210416525602, + 0.0033541317097842693, + 0.0024488691706210375, + 0.0026965041179209948, + 0.0023285604547709227, + 0.008242819458246231, + 0.006573115475475788, + 0.00752385426312685, + 0.006789118517190218, + 0.005386856384575367, + 0.004392208531498909, + 0.006057451479136944, + 0.007391762919723988, + 0.010919901542365551, + 0.005043093115091324, + 0.006588837597519159, + 0.0051247719675302505, + 0.004225407727062702, + 0.008976665325462818, + 0.0038108574226498604, + 0.0033160457387566566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.03809879347681999 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily armored general wearing a full-body suit of dark metal plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a thicket of waist-high, lush, emerald-green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general wield a massive, long-handled sword and slash forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady, smooth 360-degree rotation around the general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a slow-motion effect showing severed grass, glistening dewdrops, and moist earth exploding in all directions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813161313533783 + ], + "opposite_logprob_tokens": [ + -1.4063161611557007 + ], + "logprob_reward": -0.2813161313533783, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813161313533783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060307879000902176, + 0.0632525235414505, + 0.022227857261896133, + 0.02454911731183529, + 0.14823678135871887, + 0.08098804950714111, + 0.4439188539981842, + 0.06716510653495789, + 0.03234007954597473, + 0.057013723999261856 + ], + "credit_map": [ + 0.005835536401718855, + 0.004959259182214737, + 0.007258661091327667, + 0.006179728079587221, + 0.006297700572758913, + 0.009708395227789879, + 0.008909272029995918, + 0.011159328743815422, + 0.007780167739838362, + 0.006755562499165535, + 0.008196388371288776, + 0.007782681379467249, + 0.00749146519228816, + 0.008752238936722279, + 0.006454868242144585, + 0.010039150714874268, + 0.0029845850076526403, + 0.0022907292004674673, + 0.003351741936057806, + 0.0038971409667283297, + 0.0029624791350215673, + 0.002303370740264654, + 0.00245379819534719, + 0.0019840137101709843, + 0.0026251685339957476, + 0.0030324335675686598, + 0.005359747912734747, + 0.0029546066652983427, + 0.0027772763278335333, + 0.0022611417807638645, + 0.0039331321604549885, + 0.0016056119929999113, + 0.03135406970977783, + 0.008395504206418991, + 0.04224761947989464, + 0.012877634726464748, + 0.00933490227907896, + 0.017709622159600258, + 0.01181098073720932, + 0.014506444334983826, + 0.009037581272423267, + 0.011769551783800125, + 0.00718654552474618, + 0.008611507713794708, + 0.010901402682065964, + 0.01168886385858059, + 0.011272900737822056, + 0.010519701056182384, + 0.03108982741832733, + 0.01979142799973488, + 0.059592410922050476, + 0.02730763517320156, + 0.022658171132206917, + 0.10259106755256653, + 0.1432354897260666, + 0.03765282779932022, + 0.006731824483722448, + 0.009067082777619362, + 0.009880218654870987, + 0.008852518163621426, + 0.007764870300889015, + 0.007829970680177212, + 0.009062509983778, + 0.007976116612553596, + 0.004557755310088396, + 0.0032862857915461063, + 0.004137084819376469, + 0.0038924184627830982, + 0.004949574824422598, + 0.0054108151234686375, + 0.0028127082623541355, + 0.003293439047411084, + 0.015316971577703953, + 0.004514598287642002, + 0.007128377910703421, + 0.004471782129257917, + 0.004725905600935221, + 0.014768741093575954, + 0.0024679263588041067, + 0.003619422437623143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 5 + ], + "selected_reward": -0.2813161313533783 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily armored general wearing a full-body suit of dark metal plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a thicket of waist-high, lush, emerald-green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general wield a massive, long-handled sword and slash forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady, smooth 360-degree rotation around the general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a slow-motion effect showing severed grass, glistening dewdrops, and moist earth exploding in all directions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075802147388458 + ], + "opposite_logprob_tokens": [ + -4.511075973510742 + ], + "logprob_reward": -0.011075802147388458, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075802147388458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19034714996814728, + 0.05968956649303436, + 0.055638134479522705, + 0.05935773625969887, + 0.09767506271600723, + 0.1322653591632843, + 0.19185493886470795, + 0.048069778829813004, + 0.06129133328795433, + 0.10381095856428146 + ], + "credit_map": [ + 0.028389977291226387, + 0.017239943146705627, + 0.03150593116879463, + 0.022275175899267197, + 0.015760084614157677, + 0.025723401457071304, + 0.031703729182481766, + 0.01774890348315239, + 0.00659211166203022, + 0.006930554751306772, + 0.005421637557446957, + 0.008644822984933853, + 0.008033985272049904, + 0.007625225931406021, + 0.007759967353194952, + 0.00868125818669796, + 0.0065903733484447, + 0.006181128788739443, + 0.008516957983374596, + 0.0070084622129797935, + 0.005970273166894913, + 0.009337007068097591, + 0.005707627162337303, + 0.006326304282993078, + 0.011182628571987152, + 0.004516959190368652, + 0.016606297343969345, + 0.00483738724142313, + 0.006330172531306744, + 0.0036127378698438406, + 0.009251137264072895, + 0.003020413452759385, + 0.0150062907487154, + 0.006985754705965519, + 0.017466848716139793, + 0.009289763867855072, + 0.011146257631480694, + 0.01463159080594778, + 0.011246874928474426, + 0.011901680380105972, + 0.014265795238316059, + 0.019310180097818375, + 0.008741029538214207, + 0.016551382839679718, + 0.013286486268043518, + 0.014953253790736198, + 0.019644301384687424, + 0.025512924417853355, + 0.017665639519691467, + 0.0234820656478405, + 0.017946947365999222, + 0.024679075926542282, + 0.037132952362298965, + 0.019859222695231438, + 0.02634996548295021, + 0.024739066138863564, + 0.008514211513102055, + 0.007526346016675234, + 0.0066716172732412815, + 0.005049058701843023, + 0.00525634502992034, + 0.004397004377096891, + 0.004741639364510775, + 0.005913558881729841, + 0.009345765225589275, + 0.006948157679289579, + 0.00834765937179327, + 0.00766890961676836, + 0.007466000039130449, + 0.005791541654616594, + 0.007848376408219337, + 0.007874923758208752, + 0.031132912263274193, + 0.008419932797551155, + 0.012947184033691883, + 0.006799532566219568, + 0.007144310045987368, + 0.02362043410539627, + 0.007823317311704159, + 0.005923332646489143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.011075802147388458 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic 360-degree mid-shot, shot from a slightly upward angle, captures a heavily armored general. He wears a full-body suit of dark metal plate, scarred and weathered. The shoulder plates are engraved with roaring animal heads, reflecting a cold, hard sheen in the golden morning sunlight that filters through the mist. As the camera steadily and smoothly rotates around him, we see him amidst a thicket of waist-high, lush, emerald-green grass. He sinks his weight, his arms flexing with muscle, and with decisive and powerful force, he wields a massive, long-handled sword, its blade gleaming with a cold gleam. With a deep roar, he slashes forward with unstoppable force. The blade instantly tears through the green barrier before him, and in a slow-motion effect, severed blades of grass, glistening dewdrops, and moist earth explode in all directions, clearing a path for him to advance. The golden Jesus light outlines the contours of his armor and reflects a brief rainbow light on the splashing water droplets. The whole picture is full of epic determination and unstoppable momentum, creating a solemn and tense atmosphere.", + "seed": 5366, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_11/index_4907_seed_5366_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.34173467867076396, + "vlm_logprob": -0.34173467867076396, + "vlm_yes_no_margin": 2.3, + "vlm_token_credit": -0.34173467867076396, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily armored general wearing a full-body suit of dark metal plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a thicket of waist-high, lush, emerald-green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general wield a massive, long-handled sword and slash forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady, smooth 360-degree rotation around the general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a slow-motion effect showing severed grass, glistening dewdrops, and moist earth exploding in all directions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5776833891868591 + ], + "opposite_logprob_tokens": [ + -0.8276833891868591 + ], + "logprob_reward": -0.5776833891868591, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5776833891868591, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09728991985321045, + 0.026614760980010033, + 0.08378588408231735, + 0.09209317713975906, + 0.07213860750198364, + 0.238857239484787, + 0.0283158328384161, + 0.2493951916694641, + 0.0697832703590393, + 0.041726093739271164 + ], + "credit_map": [ + 0.010272647254168987, + 0.011768345721065998, + 0.011322482489049435, + 0.012165170162916183, + 0.011594448238611221, + 0.015015405602753162, + 0.014836120419204235, + 0.01031530275940895, + 0.0018409566255286336, + 0.0021954013500362635, + 0.0040674591436982155, + 0.003691357094794512, + 0.0034771920181810856, + 0.00313539057970047, + 0.0034360873978585005, + 0.004770915489643812, + 0.007415286265313625, + 0.00846697948873043, + 0.017801258713006973, + 0.01013468112796545, + 0.01199419517070055, + 0.009346835315227509, + 0.009997307322919369, + 0.008629345335066319, + 0.009878071025013924, + 0.012170434929430485, + 0.010475565679371357, + 0.010688530281186104, + 0.01621178537607193, + 0.01565094292163849, + 0.008336014114320278, + 0.008681834675371647, + 0.009151769801974297, + 0.01270016748458147, + 0.006078828126192093, + 0.009354496374726295, + 0.012431940995156765, + 0.007720659486949444, + 0.008788703009486198, + 0.005912042688578367, + 0.04155314713716507, + 0.02767062373459339, + 0.014721497893333435, + 0.08178167790174484, + 0.016717087477445602, + 0.013069423846900463, + 0.020985081791877747, + 0.02235870249569416, + 0.0027009123004972935, + 0.0023960727266967297, + 0.0027982713654637337, + 0.003454596037045121, + 0.0016733721131458879, + 0.0020533190108835697, + 0.0031958194449543953, + 0.010043471120297909, + 0.016976462677121162, + 0.034879766404628754, + 0.033646758645772934, + 0.03766736015677452, + 0.03427810221910477, + 0.018780915066599846, + 0.03211487829685211, + 0.041050948202610016, + 0.003803981700912118, + 0.009531104937195778, + 0.00619347020983696, + 0.006834783125668764, + 0.014072783291339874, + 0.01629781164228916, + 0.006765073165297508, + 0.006284263450652361, + 0.002442780649289489, + 0.0037172234151512384, + 0.004328165668994188, + 0.0038390816189348698, + 0.010581564158201218, + 0.0100094648078084, + 0.0037094762083142996, + 0.0030983337201178074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.5776833891868591 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily armored general wearing a full-body suit of dark metal plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a thicket of waist-high, lush, emerald-green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general wield a massive, long-handled sword and slash forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady, smooth 360-degree rotation around the general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a slow-motion effect showing severed grass, glistening dewdrops, and moist earth exploding in all directions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011083230376243591 + ], + "opposite_logprob_tokens": [ + -4.511083126068115 + ], + "logprob_reward": -0.011083230376243591, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011083230376243591, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07554910331964493, + 0.04373272880911827, + 0.04728920757770538, + 0.11541162431240082, + 0.035413917154073715, + 0.11667129397392273, + 0.06177864968776703, + 0.3693673312664032, + 0.07750716060400009, + 0.05727897211909294 + ], + "credit_map": [ + 0.005546023137867451, + 0.008616926148533821, + 0.008792613632977009, + 0.009063051082193851, + 0.007715585641562939, + 0.019828686490654945, + 0.0076825981959700584, + 0.008303619921207428, + 0.004792315419763327, + 0.004356191959232092, + 0.006429553031921387, + 0.004168170038610697, + 0.005042480770498514, + 0.00746840750798583, + 0.006222727242857218, + 0.005252885166555643, + 0.005033545196056366, + 0.0046235001645982265, + 0.005885239690542221, + 0.008582613430917263, + 0.005207086447626352, + 0.006090838462114334, + 0.005194326862692833, + 0.006672054994851351, + 0.012368996627628803, + 0.01348493155092001, + 0.016834136098623276, + 0.011243577115237713, + 0.017897043377161026, + 0.01526515744626522, + 0.00977306254208088, + 0.018544716760516167, + 0.004239479545503855, + 0.006266732234507799, + 0.003118020249530673, + 0.003811368951573968, + 0.005460169166326523, + 0.004035476595163345, + 0.0032258036080747843, + 0.00525686563923955, + 0.019699659198522568, + 0.01283683255314827, + 0.008460018783807755, + 0.04108990728855133, + 0.006880651693791151, + 0.005610825959593058, + 0.010777488350868225, + 0.0113159054890275, + 0.004666258580982685, + 0.00920308567583561, + 0.010697913356125355, + 0.007245182059705257, + 0.006842558737844229, + 0.005925677251070738, + 0.007789839990437031, + 0.009408135898411274, + 0.030351275578141212, + 0.039950478821992874, + 0.06740842759609222, + 0.05354674533009529, + 0.036782365292310715, + 0.029046429321169853, + 0.05626974627375603, + 0.056011877954006195, + 0.0055610439740121365, + 0.013059772551059723, + 0.009629563428461552, + 0.007631730753928423, + 0.014299239963293076, + 0.00821076799184084, + 0.007775307632982731, + 0.011339733377099037, + 0.005905588157474995, + 0.003913095686584711, + 0.010374329052865505, + 0.0042047505266964436, + 0.0115066422149539, + 0.00954731460660696, + 0.00488333310931921, + 0.006943915504962206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 3 + ], + "selected_reward": -0.011083230376243591 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily armored general wearing a full-body suit of dark metal plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a thicket of waist-high, lush, emerald-green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general wield a massive, long-handled sword and slash forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady, smooth 360-degree rotation around the general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a slow-motion effect showing severed grass, glistening dewdrops, and moist earth exploding in all directions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495079979300499 + ], + "opposite_logprob_tokens": [ + -2.929950714111328 + ], + "logprob_reward": -0.05495079979300499, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495079979300499, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044428154826164246, + 0.04787378013134003, + 0.03813903406262398, + 0.07049285620450974, + 0.027175050228834152, + 0.06681975722312927, + 0.02548205852508545, + 0.6147249341011047, + 0.03366849198937416, + 0.031195882707834244 + ], + "credit_map": [ + 0.004710069391876459, + 0.0039512827061116695, + 0.007032415829598904, + 0.005973300896584988, + 0.004094004165381193, + 0.0077906763181090355, + 0.005780888255685568, + 0.005095516797155142, + 0.004197028931230307, + 0.0037614053580909967, + 0.010462209582328796, + 0.0047993166372179985, + 0.006218739785254002, + 0.00635478226467967, + 0.005042499862611294, + 0.007037798408418894, + 0.0032701485324651003, + 0.005491180811077356, + 0.005257760174572468, + 0.005864924285560846, + 0.004206147510558367, + 0.003328436752781272, + 0.005726090632379055, + 0.004994346294552088, + 0.007038847543299198, + 0.007969009689986706, + 0.010037481784820557, + 0.008038358762860298, + 0.012668767012655735, + 0.012036152184009552, + 0.0059858933091163635, + 0.00671834172680974, + 0.0036829651799052954, + 0.006808449514210224, + 0.0020953796338289976, + 0.003026906866580248, + 0.004055042751133442, + 0.002220457885414362, + 0.0031296785455197096, + 0.002156172413378954, + 0.011266621761023998, + 0.00807881448417902, + 0.004300268366932869, + 0.02237713523209095, + 0.004642088431864977, + 0.0038832298014312983, + 0.005822374951094389, + 0.0064492239616811275, + 0.002023284323513508, + 0.0033680361229926348, + 0.0027120516169816256, + 0.0029871833976358175, + 0.0028636674396693707, + 0.001900990610010922, + 0.002535558771342039, + 0.00709128612652421, + 0.041143715381622314, + 0.08415398001670837, + 0.0834750235080719, + 0.09104075282812119, + 0.08380945026874542, + 0.04741111397743225, + 0.07878808677196503, + 0.10490275174379349, + 0.0026811915449798107, + 0.004875354468822479, + 0.003798000980168581, + 0.005298348143696785, + 0.004796359222382307, + 0.004289630800485611, + 0.0034477761946618557, + 0.004481831565499306, + 0.0026631171349436045, + 0.0030035199597477913, + 0.004417593125253916, + 0.002689194865524769, + 0.005288104061037302, + 0.00640901317819953, + 0.003326503559947014, + 0.0033988391514867544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 5 + ], + "selected_reward": -0.05495079979300499 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily armored general wearing a full-body suit of dark metal plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a thicket of waist-high, lush, emerald-green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general wield a massive, long-handled sword and slash forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady, smooth 360-degree rotation around the general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a slow-motion effect showing severed grass, glistening dewdrops, and moist earth exploding in all directions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.053876280784607 + ], + "opposite_logprob_tokens": [ + -0.42887628078460693 + ], + "logprob_reward": -1.053876280784607, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.053876280784607, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10913227498531342, + 0.06282693892717361, + 0.048287343233823776, + 0.04617993161082268, + 0.030746152624487877, + 0.05305382236838341, + 0.05802153795957565, + 0.2744618356227875, + 0.18804723024368286, + 0.1292429268360138 + ], + "credit_map": [ + 0.008946740999817848, + 0.01001301221549511, + 0.017156848683953285, + 0.011736828833818436, + 0.011430162936449051, + 0.020540090277791023, + 0.01790362037718296, + 0.011404969729483128, + 0.006164906546473503, + 0.005402468144893646, + 0.01150087732821703, + 0.007423670031130314, + 0.006154952105134726, + 0.007928432896733284, + 0.009349209256470203, + 0.008902423083782196, + 0.0039128209464251995, + 0.006253700237721205, + 0.0076547968201339245, + 0.007640885189175606, + 0.006277403328567743, + 0.0038818232715129852, + 0.006366300862282515, + 0.006299611181020737, + 0.005926121957600117, + 0.007284830324351788, + 0.0066827028058469296, + 0.00438650231808424, + 0.00770981702953577, + 0.004773563239723444, + 0.003938850946724415, + 0.005477542988955975, + 0.0036477979738265276, + 0.006773944944143295, + 0.002873458433896303, + 0.003583410754799843, + 0.004377330187708139, + 0.003353954292833805, + 0.0030628915410488844, + 0.0030733649618923664, + 0.00634993240237236, + 0.010130530223250389, + 0.005235996562987566, + 0.011907566338777542, + 0.004746412392705679, + 0.0044318074360489845, + 0.005018521565943956, + 0.005233057774603367, + 0.00458004092797637, + 0.007965446449816227, + 0.006605170201510191, + 0.009098691865801811, + 0.005245573818683624, + 0.003945902921259403, + 0.0065018897876143456, + 0.014078824780881405, + 0.01800614222884178, + 0.04151342436671257, + 0.03611085191369057, + 0.04233761131763458, + 0.03666539117693901, + 0.020575376227498055, + 0.03396613523364067, + 0.04528690502047539, + 0.02192903310060501, + 0.025672541931271553, + 0.017916634678840637, + 0.022079817950725555, + 0.028066236525774002, + 0.035675834864377975, + 0.01429747324436903, + 0.022409655153751373, + 0.0076094274409115314, + 0.010531054809689522, + 0.008804211392998695, + 0.010961680673062801, + 0.03623418137431145, + 0.036508169025182724, + 0.008156872354447842, + 0.01043734047561884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -1.053876280784607 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily armored general wearing a full-body suit of dark metal plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a thicket of waist-high, lush, emerald-green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general wield a massive, long-handled sword and slash forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady, smooth 360-degree rotation around the general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a slow-motion effect showing severed grass, glistening dewdrops, and moist earth exploding in all directions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011079693213105202 + ], + "opposite_logprob_tokens": [ + -4.511079788208008 + ], + "logprob_reward": -0.011079693213105202, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011079693213105202, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08079212158918381, + 0.037928953766822815, + 0.04025694727897644, + 0.06715378910303116, + 0.05024472624063492, + 0.18775545060634613, + 0.026501726359128952, + 0.0995091125369072, + 0.08343250304460526, + 0.3264246881008148 + ], + "credit_map": [ + 0.009674353525042534, + 0.007531931158155203, + 0.010793167166411877, + 0.008284367620944977, + 0.007927458733320236, + 0.01932103931903839, + 0.008607827126979828, + 0.008651974610984325, + 0.0034756003879010677, + 0.0026234243996441364, + 0.0057484423741698265, + 0.003694333368912339, + 0.0035571164917200804, + 0.004818746820092201, + 0.0058356658555567265, + 0.008175629191100597, + 0.0032517691142857075, + 0.0034805384930223227, + 0.008619763888418674, + 0.005476443096995354, + 0.006372096017003059, + 0.003497759345918894, + 0.0054788836278021336, + 0.004079695325344801, + 0.006722877733409405, + 0.006399950478225946, + 0.012013355270028114, + 0.005466741975396872, + 0.009310431778430939, + 0.007594826631247997, + 0.006740389857441187, + 0.012905215844511986, + 0.006655353587120771, + 0.01035674475133419, + 0.003948192577809095, + 0.007088307291269302, + 0.007769844960421324, + 0.004737887065857649, + 0.005938123911619186, + 0.0037502769846469164, + 0.03055545501410961, + 0.020848561078310013, + 0.013618703931570053, + 0.061813727021217346, + 0.013201644644141197, + 0.010534089058637619, + 0.01856873370707035, + 0.018614549189805984, + 0.0027810768224298954, + 0.0025211297906935215, + 0.002470481675118208, + 0.002532919170334935, + 0.0030321350786834955, + 0.0020713929552584887, + 0.0024521539453417063, + 0.008640439249575138, + 0.006309471093118191, + 0.01390571054071188, + 0.015065999701619148, + 0.015110572800040245, + 0.011677736416459084, + 0.008145277388393879, + 0.012836059555411339, + 0.016458289697766304, + 0.007495524361729622, + 0.012946907430887222, + 0.006572382058948278, + 0.010095389559864998, + 0.013988218270242214, + 0.009501581080257893, + 0.009933269582688808, + 0.012899238616228104, + 0.019632212817668915, + 0.03132065013051033, + 0.03541388735175133, + 0.029989022761583328, + 0.08816632628440857, + 0.0750863179564476, + 0.023129912093281746, + 0.023686353117227554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 7 + ], + "selected_reward": -0.011079693213105202 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_12/index_tensor([4498, 4498])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_12/index_tensor([4498, 4498])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4f9c135bd93b2f15529de14d3ca56de2c8873427 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_12/index_tensor([4498, 4498])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "series of rapid, dynamic cuts captures the intense collaborative energy of a small, diverse team of engineers. The scene begins with a low-angle, medium shot of the group huddled in a dimly lit, high-tech laboratory with a polished concrete floor. Their silhouettes are silhouetted against a vast, floor-to-ceiling glass wall, where complex mathematical formulas and intricate geometric shapes glow with a bright, electric blue light. The camera then switches to a tight close-up of the face of a young East Asian woman, her dark hair pulled back in a bun, thin-rimmed glasses perched on her nose; as she frowns, the glowing blue equations are reflected in her lens. She is completely absorbed. A subsequent cut reveals a detailed shot of a man's hand, a middle-aged Black engineer in a gray cashmere sweater, decisively pointing his open palm at a specific glowing entity on the glass, his finger tracing its curve in the air. This is followed by another shot, a shoulder-length shot of a young White man with shaggy blond hair, who leans over and quickly uses a stylus to add a new string of glowing white variables to the wall. The figure comes alive upon contact. The entire scene is bathed in cool, high-contrast light from the glass, casting sharp shadows and creating an atmosphere of tension, intelligence, and impending breakthrough.", + "seed": 11023, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_12/index_4498_seed_11023_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05590492659248412, + "vlm_logprob": -0.05590492659248412, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.05590492659248412, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, high-tech laboratory with a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a floor-to-ceiling glass wall displaying glowing electric blue mathematical formulas and geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video include a tight close-up of a young East Asian woman with dark hair in a bun and thin-rimmed glasses reflecting the blue equations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a middle-aged Black man in a gray cashmere sweater pointing his open palm at a glowing entity on the glass wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include a shot of a young White man with shaggy blond hair using a stylus to add glowing white variables to the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10773129016160965 + ], + "opposite_logprob_tokens": [ + -2.357731342315674 + ], + "logprob_reward": -0.10773129016160965, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10773129016160965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16327005624771118, + 0.06202530860900879, + 0.05900045111775398, + 0.024318980053067207, + 0.08070902526378632, + 0.11673198640346527, + 0.09519736468791962, + 0.06955704838037491, + 0.2599859833717346, + 0.06920376420021057 + ], + "credit_map": [ + 0.015365484170615673, + 0.026054415851831436, + 0.018184464424848557, + 0.02024705894291401, + 0.021668773144483566, + 0.022536324337124825, + 0.01926070638000965, + 0.019952846691012383, + 0.006907883565872908, + 0.006369812414050102, + 0.004619193729013205, + 0.007360711228102446, + 0.01098538562655449, + 0.01257515698671341, + 0.004465659148991108, + 0.008741511031985283, + 0.005308082792907953, + 0.008374274708330631, + 0.01150442473590374, + 0.007723264861851931, + 0.00925492774695158, + 0.006472676061093807, + 0.0053246296010911465, + 0.005038175731897354, + 0.0026216628029942513, + 0.0030614801216870546, + 0.0037765558809041977, + 0.004385125357657671, + 0.0030974913388490677, + 0.0019461187766864896, + 0.0034694757778197527, + 0.0019610717426985502, + 0.006730997934937477, + 0.007417476270347834, + 0.021298490464687347, + 0.024963414296507835, + 0.00469420338049531, + 0.0035106197465211153, + 0.006055724341422319, + 0.006038103252649307, + 0.019727220758795738, + 0.010390745475888252, + 0.01723046787083149, + 0.016520695760846138, + 0.012468057684600353, + 0.010668333619832993, + 0.011569323018193245, + 0.018157145008444786, + 0.005876983981579542, + 0.0098769161850214, + 0.006040402688086033, + 0.008368117734789848, + 0.007204744033515453, + 0.005596991162747145, + 0.026074299588799477, + 0.02615891396999359, + 0.005401750560849905, + 0.009428546763956547, + 0.008999590761959553, + 0.006057091522961855, + 0.010005212388932705, + 0.007119597867131233, + 0.008396193385124207, + 0.01414907444268465, + 0.0250453632324934, + 0.030542178079485893, + 0.06451021134853363, + 0.02290264517068863, + 0.028664927929639816, + 0.018837960436940193, + 0.030654584988951683, + 0.03882814943790436, + 0.004101867787539959, + 0.0034584796521812677, + 0.010117997415363789, + 0.003324649529531598, + 0.012673252262175083, + 0.007089056074619293, + 0.0038880605716258287, + 0.024550408124923706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.10773129016160965 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, high-tech laboratory with a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a floor-to-ceiling glass wall displaying glowing electric blue mathematical formulas and geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video include a tight close-up of a young East Asian woman with dark hair in a bun and thin-rimmed glasses reflecting the blue equations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a middle-aged Black man in a gray cashmere sweater pointing his open palm at a glowing entity on the glass wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include a shot of a young White man with shaggy blond hair using a stylus to add glowing white variables to the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002524643437936902 + ], + "opposite_logprob_tokens": [ + -6.0025248527526855 + ], + "logprob_reward": -0.002524643437936902, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002524643437936902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08709798753261566, + 0.03796184062957764, + 0.04944046959280968, + 0.02664029784500599, + 0.10933525115251541, + 0.06687676161527634, + 0.264965295791626, + 0.06010598689317703, + 0.13818785548210144, + 0.15938825905323029 + ], + "credit_map": [ + 0.007970278151333332, + 0.009435707703232765, + 0.008988319896161556, + 0.010691971518099308, + 0.018657494336366653, + 0.010128583759069443, + 0.010867070406675339, + 0.010358559899032116, + 0.005382783245295286, + 0.003910874016582966, + 0.005198344122618437, + 0.004365592263638973, + 0.0038034343160688877, + 0.005768982693552971, + 0.004347673617303371, + 0.005184154491871595, + 0.004561923909932375, + 0.007288267370313406, + 0.00863248948007822, + 0.006834589876234531, + 0.007410827558487654, + 0.005798176862299442, + 0.004591495264321566, + 0.00432269973680377, + 0.0032624336890876293, + 0.003002607962116599, + 0.004859136417508125, + 0.004359430633485317, + 0.0024774300400167704, + 0.0021497393026947975, + 0.004054036922752857, + 0.0024754845071583986, + 0.011271427385509014, + 0.01733110286295414, + 0.026673734188079834, + 0.023995928466320038, + 0.005650518927723169, + 0.007391393184661865, + 0.009252440184354782, + 0.007768712006509304, + 0.006781954318284988, + 0.005797428544610739, + 0.0126642482355237, + 0.010923258028924465, + 0.006942752283066511, + 0.00413509551435709, + 0.00928888376802206, + 0.010343138128519058, + 0.009904601611196995, + 0.013303580693900585, + 0.013065268285572529, + 0.014530595391988754, + 0.011680250987410545, + 0.010313062928617, + 0.09793782234191895, + 0.09423012286424637, + 0.004322110675275326, + 0.009393048472702503, + 0.005891593173146248, + 0.0039632744155824184, + 0.005545238498598337, + 0.005454603582620621, + 0.007937442511320114, + 0.017598673701286316, + 0.011968203820288181, + 0.013608681969344616, + 0.029771657660603523, + 0.01451368723064661, + 0.01390111818909645, + 0.012247628532350063, + 0.017525874078273773, + 0.02465100586414337, + 0.006669738795608282, + 0.0073775905184447765, + 0.018665116280317307, + 0.008673179894685745, + 0.03131066635251045, + 0.018640846014022827, + 0.00793552678078413, + 0.06011558696627617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 8 + ], + "selected_reward": -0.002524643437936902 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, high-tech laboratory with a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a floor-to-ceiling glass wall displaying glowing electric blue mathematical formulas and geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video include a tight close-up of a young East Asian woman with dark hair in a bun and thin-rimmed glasses reflecting the blue equations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a middle-aged Black man in a gray cashmere sweater pointing his open palm at a glowing entity on the glass wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include a shot of a young White man with shaggy blond hair using a stylus to add glowing white variables to the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018198424950242043 + ], + "opposite_logprob_tokens": [ + -4.018198490142822 + ], + "logprob_reward": -0.018198424950242043, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018198424950242043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049109093844890594, + 0.042352043092250824, + 0.07751086354255676, + 0.025924382731318474, + 0.03636697679758072, + 0.027295580133795738, + 0.1675812155008316, + 0.1173572912812233, + 0.1781408190727234, + 0.2783617079257965 + ], + "credit_map": [ + 0.0038440993521362543, + 0.00665900157764554, + 0.004045738372951746, + 0.0049028848297894, + 0.005398561246693134, + 0.006384743377566338, + 0.007070145104080439, + 0.010803919285535812, + 0.0035044627729803324, + 0.004962229635566473, + 0.004052492789924145, + 0.005398330744355917, + 0.0046882713213562965, + 0.008061742410063744, + 0.004483561497181654, + 0.007200952619314194, + 0.005977891385555267, + 0.006645640823990107, + 0.009819375351071358, + 0.009557044133543968, + 0.021692387759685516, + 0.012678866274654865, + 0.006720476783812046, + 0.004419188015162945, + 0.0028645223937928677, + 0.0026787372771650553, + 0.005247199442237616, + 0.0045013525523245335, + 0.0030850388575345278, + 0.0018865496385842562, + 0.0037874430418014526, + 0.0018735367339104414, + 0.0026639041025191545, + 0.0040864236652851105, + 0.008814969100058079, + 0.010525311343371868, + 0.0026380051858723164, + 0.0021949175279587507, + 0.0027729172725230455, + 0.0026705297641456127, + 0.0031615965999662876, + 0.0032468638382852077, + 0.004774964414536953, + 0.005035687703639269, + 0.0032357790041714907, + 0.0015990526881068945, + 0.004513273015618324, + 0.0017283610068261623, + 0.00440828176215291, + 0.006759303621947765, + 0.005361694376915693, + 0.008517030626535416, + 0.006991573143750429, + 0.005310917738825083, + 0.06632427871227264, + 0.06390812993049622, + 0.007140201050788164, + 0.01620090752840042, + 0.010318432934582233, + 0.00684604374691844, + 0.015505789779126644, + 0.0161577258259058, + 0.015068545006215572, + 0.0301196426153183, + 0.017227953299880028, + 0.020271481946110725, + 0.04027803987264633, + 0.01820826157927513, + 0.02121402882039547, + 0.01884370855987072, + 0.0181281715631485, + 0.02396915853023529, + 0.011965191923081875, + 0.009348515421152115, + 0.030782129615545273, + 0.007982542738318443, + 0.06413044035434723, + 0.029901016503572464, + 0.006898803170770407, + 0.11735308170318604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.018198424950242043 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, high-tech laboratory with a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a floor-to-ceiling glass wall displaying glowing electric blue mathematical formulas and geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video include a tight close-up of a young East Asian woman with dark hair in a bun and thin-rimmed glasses reflecting the blue equations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a middle-aged Black man in a gray cashmere sweater pointing his open palm at a glowing entity on the glass wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include a shot of a young White man with shaggy blond hair using a stylus to add glowing white variables to the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902692049741745 + ], + "opposite_logprob_tokens": [ + -2.464026927947998 + ], + "logprob_reward": -0.08902692049741745, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902692049741745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05000048130750656, + 0.049927107989788055, + 0.047823794186115265, + 0.03000466525554657, + 0.052699975669384, + 0.07839009910821915, + 0.34263044595718384, + 0.05293383076786995, + 0.13193140923976898, + 0.16365815699100494 + ], + "credit_map": [ + 0.0040646689012646675, + 0.008968772366642952, + 0.004536492750048637, + 0.004721657373011112, + 0.00885831844061613, + 0.005565362516790628, + 0.009209799580276012, + 0.004075407516211271, + 0.0052355192601680756, + 0.005302041303366423, + 0.003611732739955187, + 0.007210366427898407, + 0.0079306336119771, + 0.010133704170584679, + 0.0039554787799716, + 0.006547632627189159, + 0.003953107167035341, + 0.005670774262398481, + 0.006183507386595011, + 0.005935006774961948, + 0.011481677182018757, + 0.006654823664575815, + 0.004495525266975164, + 0.003449370851740241, + 0.003315137466415763, + 0.0030610638204962015, + 0.005569732282310724, + 0.0046397848054766655, + 0.004129588603973389, + 0.00251378514803946, + 0.004312786273658276, + 0.0024627845268696547, + 0.00585482781752944, + 0.007915856316685677, + 0.011014939285814762, + 0.012606370262801647, + 0.003238614182919264, + 0.002527498872950673, + 0.004647956229746342, + 0.004893912002444267, + 0.00784067902714014, + 0.009220315143465996, + 0.01477007381618023, + 0.013942227698862553, + 0.006450145039707422, + 0.005387794226408005, + 0.012075708247721195, + 0.008703155443072319, + 0.009946847334504128, + 0.015307274647057056, + 0.011985489167273045, + 0.018137529492378235, + 0.01900665834546089, + 0.014736912213265896, + 0.1286100596189499, + 0.12489967793226242, + 0.0035553413908928633, + 0.008696351200342178, + 0.004440814722329378, + 0.002720799995586276, + 0.006368902046233416, + 0.005024522077292204, + 0.007693901192396879, + 0.014433198608458042, + 0.011903371661901474, + 0.014849172905087471, + 0.029080165550112724, + 0.010735530406236649, + 0.01730395294725895, + 0.010578900575637817, + 0.016943203285336494, + 0.020537111908197403, + 0.007216608617454767, + 0.0062808748334646225, + 0.018730975687503815, + 0.005098080728203058, + 0.03619495779275894, + 0.01687893643975258, + 0.005472667049616575, + 0.06778506189584732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 8 + ], + "selected_reward": -0.08902692049741745 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, high-tech laboratory with a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a floor-to-ceiling glass wall displaying glowing electric blue mathematical formulas and geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video include a tight close-up of a young East Asian woman with dark hair in a bun and thin-rimmed glasses reflecting the blue equations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a middle-aged Black man in a gray cashmere sweater pointing his open palm at a glowing entity on the glass wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include a shot of a young White man with shaggy blond hair using a stylus to add glowing white variables to the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204335391521454 + ], + "opposite_logprob_tokens": [ + -2.8120434284210205 + ], + "logprob_reward": -0.06204335391521454, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204335391521454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04845187067985535, + 0.028729140758514404, + 0.03669340908527374, + 0.014367322437465191, + 0.023756295442581177, + 0.03188163414597511, + 0.07605015486478806, + 0.021098358556628227, + 0.04895918443799019, + 0.6700127124786377 + ], + "credit_map": [ + 0.0033456289675086737, + 0.010109934024512768, + 0.004165222868323326, + 0.005162328481674194, + 0.005252761300653219, + 0.005716538056731224, + 0.006159049458801746, + 0.00854040402919054, + 0.002745090052485466, + 0.004074356518685818, + 0.0028488885145634413, + 0.002977908356115222, + 0.0038773384876549244, + 0.0054119909182190895, + 0.0023888670839369297, + 0.00440470315515995, + 0.0031174379400908947, + 0.0029574143700301647, + 0.004958068951964378, + 0.0044968025758862495, + 0.009018444456160069, + 0.005979715846478939, + 0.0032509088050574064, + 0.0029146182350814342, + 0.001585925929248333, + 0.0016693520592525601, + 0.0021554555278271437, + 0.002295017708092928, + 0.0016603732947260141, + 0.0013184938579797745, + 0.001997702056542039, + 0.0016850014217197895, + 0.0024206044618040323, + 0.002131024142727256, + 0.005485985428094864, + 0.006285765208303928, + 0.0017926719738170505, + 0.001346940640360117, + 0.002303952118381858, + 0.0019893499556928873, + 0.003776619443669915, + 0.0030284763779491186, + 0.0051225898787379265, + 0.004877426661550999, + 0.003262309590354562, + 0.0021000653505325317, + 0.0046427748166024685, + 0.0050713722594082355, + 0.004224494099617004, + 0.006554559804499149, + 0.005264641717076302, + 0.007651899475604296, + 0.007496424484997988, + 0.005775465164333582, + 0.01962125673890114, + 0.019461413845419884, + 0.0023199652787297964, + 0.0022500280756503344, + 0.0026204599998891354, + 0.0024841479025781155, + 0.0028969149570912123, + 0.0031408099457621574, + 0.0024836775846779346, + 0.002902354346588254, + 0.004506409168243408, + 0.005735458806157112, + 0.01180192455649376, + 0.003679240122437477, + 0.005870741792023182, + 0.003776764962822199, + 0.00622485252097249, + 0.007363791577517986, + 0.028787560760974884, + 0.022831210866570473, + 0.0748656839132309, + 0.019239412620663643, + 0.15430670976638794, + 0.07179895788431168, + 0.01621151901781559, + 0.28197163343429565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.06204335391521454 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "series of rapid, dynamic cuts captures the intense collaborative energy of a small, diverse team of engineers. The scene begins with a low-angle, medium shot of the group huddled in a dimly lit, high-tech laboratory with a polished concrete floor. Their silhouettes are silhouetted against a vast, floor-to-ceiling glass wall, where complex mathematical formulas and intricate geometric shapes glow with a bright, electric blue light. The camera then switches to a tight close-up of the face of a young East Asian woman, her dark hair pulled back in a bun, thin-rimmed glasses perched on her nose; as she frowns, the glowing blue equations are reflected in her lens. She is completely absorbed. A subsequent cut reveals a detailed shot of a man's hand, a middle-aged Black engineer in a gray cashmere sweater, decisively pointing his open palm at a specific glowing entity on the glass, his finger tracing its curve in the air. This is followed by another shot, a shoulder-length shot of a young White man with shaggy blond hair, who leans over and quickly uses a stylus to add a new string of glowing white variables to the wall. The figure comes alive upon contact. The entire scene is bathed in cool, high-contrast light from the glass, casting sharp shadows and creating an atmosphere of tension, intelligence, and impending breakthrough.", + "seed": 11024, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_12/index_4498_seed_11024_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02373302896739915, + "vlm_logprob": -0.02373302896739915, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.02373302896739915, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, high-tech laboratory with a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a floor-to-ceiling glass wall displaying glowing electric blue mathematical formulas and geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video include a tight close-up of a young East Asian woman with dark hair in a bun and thin-rimmed glasses reflecting the blue equations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a middle-aged Black man in a gray cashmere sweater pointing his open palm at a glowing entity on the glass wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include a shot of a young White man with shaggy blond hair using a stylus to add glowing white variables to the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03497965261340141 + ], + "opposite_logprob_tokens": [ + -3.6599795818328857 + ], + "logprob_reward": -0.03497965261340141, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03497965261340141, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38885897397994995, + 0.0514986626803875, + 0.11980900913476944, + 0.04633847996592522, + 0.03943758085370064, + 0.10905500501394272, + 0.054019175469875336, + 0.050583112984895706, + 0.0826767235994339, + 0.0577232800424099 + ], + "credit_map": [ + 0.06858018785715103, + 0.05818434804677963, + 0.05511484667658806, + 0.05075662210583687, + 0.041740112006664276, + 0.04270168021321297, + 0.01972607709467411, + 0.05205509439110756, + 0.013685878366231918, + 0.007856994867324829, + 0.009017673321068287, + 0.004497359041124582, + 0.0038729379884898663, + 0.004888108000159264, + 0.0035717131104320288, + 0.004107997287064791, + 0.03659733384847641, + 0.04664226993918419, + 0.004896058700978756, + 0.005428093485534191, + 0.009134309366345406, + 0.005584745667874813, + 0.0069375536404550076, + 0.004588654264807701, + 0.008952115662395954, + 0.010768650099635124, + 0.004770570434629917, + 0.003763435408473015, + 0.0041016326285898685, + 0.004964738618582487, + 0.005005889106541872, + 0.004011447541415691, + 0.0034032263793051243, + 0.003638138761743903, + 0.0039043754804879427, + 0.0025029554963111877, + 0.011140367016196251, + 0.005478966981172562, + 0.004157567396759987, + 0.005211981479078531, + 0.02423126809298992, + 0.017111362889409065, + 0.01474200189113617, + 0.012759335339069366, + 0.012274466454982758, + 0.00967533327639103, + 0.00791998952627182, + 0.010341250337660313, + 0.0050215027295053005, + 0.008723931387066841, + 0.004830185789614916, + 0.006597229745239019, + 0.004996354226022959, + 0.010349680669605732, + 0.005142134614288807, + 0.00835815817117691, + 0.0069096083752810955, + 0.005540092941373587, + 0.005619249306619167, + 0.0067083039321005344, + 0.006718502379953861, + 0.007272905670106411, + 0.005842491053044796, + 0.00597195653244853, + 0.00981964636594057, + 0.01209985464811325, + 0.01606575772166252, + 0.0074991886503994465, + 0.01054391823709011, + 0.011656678281724453, + 0.008406398817896843, + 0.006585284136235714, + 0.0062211439944803715, + 0.005435607396066189, + 0.005913992412388325, + 0.008377640508115292, + 0.009079600684344769, + 0.009115717373788357, + 0.0062735299579799175, + 0.007306046783924103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.03497965261340141 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, high-tech laboratory with a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a floor-to-ceiling glass wall displaying glowing electric blue mathematical formulas and geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video include a tight close-up of a young East Asian woman with dark hair in a bun and thin-rimmed glasses reflecting the blue equations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a middle-aged Black man in a gray cashmere sweater pointing his open palm at a glowing entity on the glass wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include a shot of a young White man with shaggy blond hair using a stylus to add glowing white variables to the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007522851810790598 + ], + "opposite_logprob_tokens": [ + -7.2507524490356445 + ], + "logprob_reward": -0.0007522851810790598, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007522851810790598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2882575988769531, + 0.04500814154744148, + 0.30663514137268066, + 0.07372663170099258, + 0.08290892094373703, + 0.07095181196928024, + 0.031865768134593964, + 0.03428417444229126, + 0.03623603656888008, + 0.030125755816698074 + ], + "credit_map": [ + 0.05912918224930763, + 0.05936581268906593, + 0.0326758436858654, + 0.033521294593811035, + 0.04165232554078102, + 0.024957479909062386, + 0.012643191032111645, + 0.02431248500943184, + 0.007873953320086002, + 0.004658966325223446, + 0.009107036516070366, + 0.004295565653592348, + 0.004204700235277414, + 0.005286235362291336, + 0.004748482257127762, + 0.004833204206079245, + 0.10688420385122299, + 0.12822219729423523, + 0.013526701368391514, + 0.011349615640938282, + 0.016993612051010132, + 0.010755796916782856, + 0.010506777092814445, + 0.00839625671505928, + 0.012038429267704487, + 0.01795351691544056, + 0.006644329521805048, + 0.006698506884276867, + 0.006868574768304825, + 0.010349715128540993, + 0.007547715678811073, + 0.005625838413834572, + 0.009153323248028755, + 0.0073778750374913216, + 0.013242403976619244, + 0.006694114301353693, + 0.010341344401240349, + 0.012506981380283833, + 0.009436783380806446, + 0.014156087301671505, + 0.014937080442905426, + 0.007332946639508009, + 0.009634596295654774, + 0.008980737999081612, + 0.005693226121366024, + 0.007602750323712826, + 0.007710305508226156, + 0.009060172364115715, + 0.0042976997792720795, + 0.003202152205631137, + 0.003303198143839836, + 0.004509883001446724, + 0.00330750341527164, + 0.00539973471313715, + 0.0027734411414712667, + 0.005072156898677349, + 0.0029060738161206245, + 0.0033875461667776108, + 0.005568739026784897, + 0.008254523389041424, + 0.00328361545689404, + 0.0037408226635307074, + 0.0036838953383266926, + 0.0034589588176459074, + 0.00437453668564558, + 0.0056203799322247505, + 0.004134892951697111, + 0.004208450671285391, + 0.0046654995530843735, + 0.005623199045658112, + 0.003744193585589528, + 0.0038648813497275114, + 0.0027930454816669226, + 0.003348209196701646, + 0.002717230934649706, + 0.00407768739387393, + 0.005348040722310543, + 0.005066166166216135, + 0.0030856360681355, + 0.0036897400859743357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.0007522851810790598 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, high-tech laboratory with a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a floor-to-ceiling glass wall displaying glowing electric blue mathematical formulas and geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video include a tight close-up of a young East Asian woman with dark hair in a bun and thin-rimmed glasses reflecting the blue equations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a middle-aged Black man in a gray cashmere sweater pointing his open palm at a glowing entity on the glass wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include a shot of a young White man with shaggy blond hair using a stylus to add glowing white variables to the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009793448261916637 + ], + "opposite_logprob_tokens": [ + -4.634793281555176 + ], + "logprob_reward": -0.009793448261916637, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009793448261916637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21960626542568207, + 0.045624326914548874, + 0.27736762166023254, + 0.06392228603363037, + 0.1373000144958496, + 0.058412518352270126, + 0.05897704139351845, + 0.04004098102450371, + 0.07364726811647415, + 0.02510165609419346 + ], + "credit_map": [ + 0.01486958097666502, + 0.02789243869483471, + 0.015033242292702198, + 0.02201128751039505, + 0.05385162681341171, + 0.016219347715377808, + 0.016369296237826347, + 0.05335945263504982, + 0.006257873959839344, + 0.006384760607033968, + 0.00620027631521225, + 0.00581548735499382, + 0.0040000686421990395, + 0.006333129946142435, + 0.006252580787986517, + 0.004380154889076948, + 0.08111932128667831, + 0.09136784076690674, + 0.014179842546582222, + 0.020545953884720802, + 0.027148788794875145, + 0.012664002366364002, + 0.01909397356212139, + 0.011247916147112846, + 0.005773640237748623, + 0.006510686129331589, + 0.008697732351720333, + 0.008334677666425705, + 0.008231112733483315, + 0.007715253159403801, + 0.01350163109600544, + 0.005157556384801865, + 0.01210736483335495, + 0.011000074446201324, + 0.015923641622066498, + 0.00891098566353321, + 0.03429090231657028, + 0.01881137117743492, + 0.013337383978068829, + 0.022918304428458214, + 0.008284440264105797, + 0.011189139448106289, + 0.007418879307806492, + 0.0076003894209861755, + 0.007213291712105274, + 0.006196858827024698, + 0.004215215798467398, + 0.006294306833297014, + 0.004487542435526848, + 0.005905675236135721, + 0.0051799193024635315, + 0.005509458482265472, + 0.005127554293721914, + 0.013686216436326504, + 0.0038045404944568872, + 0.015276139602065086, + 0.003897314192727208, + 0.004396509844809771, + 0.005679575726389885, + 0.007720665540546179, + 0.004329864867031574, + 0.005572631023824215, + 0.004414744209498167, + 0.004029677715152502, + 0.008598936721682549, + 0.009251199662685394, + 0.013161668553948402, + 0.007010700646787882, + 0.010327942669391632, + 0.011043976992368698, + 0.007336176000535488, + 0.006916671060025692, + 0.0029006190598011017, + 0.0021741213276982307, + 0.0023879033979028463, + 0.0030077011324465275, + 0.004367273300886154, + 0.003912366461008787, + 0.003025633981451392, + 0.003326041391119361 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.009793448261916637 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, high-tech laboratory with a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a floor-to-ceiling glass wall displaying glowing electric blue mathematical formulas and geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video include a tight close-up of a young East Asian woman with dark hair in a bun and thin-rimmed glasses reflecting the blue equations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a middle-aged Black man in a gray cashmere sweater pointing his open palm at a glowing entity on the glass wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include a shot of a young White man with shaggy blond hair using a stylus to add glowing white variables to the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018183207139372826 + ], + "opposite_logprob_tokens": [ + -4.01818323135376 + ], + "logprob_reward": -0.018183207139372826, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018183207139372826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10467042028903961, + 0.0509457103908062, + 0.3054323196411133, + 0.08679288625717163, + 0.11964937299489975, + 0.08118133991956711, + 0.07779126614332199, + 0.06029509752988815, + 0.06838346272706985, + 0.044858090579509735 + ], + "credit_map": [ + 0.01023322343826294, + 0.012822717428207397, + 0.007216133177280426, + 0.011350341141223907, + 0.0235510915517807, + 0.007997320033609867, + 0.005010462366044521, + 0.0264891404658556, + 0.006881036330014467, + 0.005773683078587055, + 0.007761574350297451, + 0.0054310280829668045, + 0.004264286253601313, + 0.005825053434818983, + 0.008922871202230453, + 0.006086185574531555, + 0.09723744541406631, + 0.11504463106393814, + 0.010029717348515987, + 0.013588834553956985, + 0.027709495276212692, + 0.01327030174434185, + 0.017817912623286247, + 0.010734026320278645, + 0.006568535231053829, + 0.006559590343385935, + 0.006296803243458271, + 0.008802495896816254, + 0.016975276172161102, + 0.011957413516938686, + 0.02254357375204563, + 0.007089205551892519, + 0.0111228683963418, + 0.009594324976205826, + 0.016456302255392075, + 0.011134929955005646, + 0.02720395103096962, + 0.017050839960575104, + 0.01272148359566927, + 0.014364682137966156, + 0.010612216778099537, + 0.010518187656998634, + 0.010928157716989517, + 0.010915747843682766, + 0.007915236055850983, + 0.009654560126364231, + 0.009714826010167599, + 0.010922415181994438, + 0.00560752023011446, + 0.007019609212875366, + 0.0059662810526788235, + 0.00571030518040061, + 0.005017673596739769, + 0.020280025899410248, + 0.0058830115012824535, + 0.02230685204267502, + 0.005597969517111778, + 0.007117562461644411, + 0.008096032775938511, + 0.01133265532553196, + 0.005331301596015692, + 0.008057409897446632, + 0.008576113730669022, + 0.006186051294207573, + 0.008529377169907093, + 0.009039920754730701, + 0.013098717667162418, + 0.005782937165349722, + 0.008522700518369675, + 0.010488893836736679, + 0.006442722864449024, + 0.006478204391896725, + 0.003190701361745596, + 0.004350938368588686, + 0.004477493464946747, + 0.005569332744926214, + 0.007397234905511141, + 0.009194842539727688, + 0.004547347314655781, + 0.006130202207714319 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.018183207139372826 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, high-tech laboratory with a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a floor-to-ceiling glass wall displaying glowing electric blue mathematical formulas and geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video include a tight close-up of a young East Asian woman with dark hair in a bun and thin-rimmed glasses reflecting the blue equations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a middle-aged Black man in a gray cashmere sweater pointing his open palm at a glowing entity on the glass wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include a shot of a young White man with shaggy blond hair using a stylus to add glowing white variables to the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054956551641225815 + ], + "opposite_logprob_tokens": [ + -2.9299564361572266 + ], + "logprob_reward": -0.054956551641225815, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054956551641225815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22130855917930603, + 0.04186640679836273, + 0.15408837795257568, + 0.08848178386688232, + 0.0820997953414917, + 0.07372907549142838, + 0.07031286507844925, + 0.07978516817092896, + 0.09162107855081558, + 0.09670687466859818 + ], + "credit_map": [ + 0.025632275268435478, + 0.031671732664108276, + 0.011369645595550537, + 0.02077854424715042, + 0.046713586896657944, + 0.017723696306347847, + 0.009919548407196999, + 0.05749952420592308, + 0.005864891689270735, + 0.004794614389538765, + 0.007059016730636358, + 0.0045544458553195, + 0.00413333298638463, + 0.00576763367280364, + 0.0055380770936608315, + 0.004154398571699858, + 0.04229326546192169, + 0.05674953758716583, + 0.006155873648822308, + 0.008767252787947655, + 0.016366785392165184, + 0.007629308849573135, + 0.00833545159548521, + 0.0077909124083817005, + 0.0100549953058362, + 0.01016001682728529, + 0.015257302671670914, + 0.010550763458013535, + 0.00949957873672247, + 0.011459121480584145, + 0.013439662754535675, + 0.008060344494879246, + 0.008192566223442554, + 0.007093150168657303, + 0.012758261524140835, + 0.005448771640658379, + 0.01799158565700054, + 0.01093342900276184, + 0.007927898317575455, + 0.011754144914448261, + 0.008983361534774303, + 0.009752892889082432, + 0.012919468805193901, + 0.008614965714514256, + 0.007350151892751455, + 0.00871667917817831, + 0.006586737465113401, + 0.010804819874465466, + 0.006773096509277821, + 0.005832849070429802, + 0.00595903629437089, + 0.006681093014776707, + 0.00669301301240921, + 0.01649119146168232, + 0.005057168193161488, + 0.016825426369905472, + 0.006440858356654644, + 0.009420040994882584, + 0.009026125073432922, + 0.013009096495807171, + 0.007323167286813259, + 0.014194761402904987, + 0.011219769716262817, + 0.009151353500783443, + 0.008888929150998592, + 0.011102606542408466, + 0.012402317486703396, + 0.009952346794307232, + 0.017288221046328545, + 0.010083350352942944, + 0.010576440952718258, + 0.01132687646895647, + 0.008006942458450794, + 0.007415296509861946, + 0.00832468830049038, + 0.012259231880307198, + 0.02029004879295826, + 0.020053032785654068, + 0.009087760001420975, + 0.011269878596067429 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.054956551641225815 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "series of rapid, dynamic cuts captures the intense collaborative energy of a small, diverse team of engineers. The scene begins with a low-angle, medium shot of the group huddled in a dimly lit, high-tech laboratory with a polished concrete floor. Their silhouettes are silhouetted against a vast, floor-to-ceiling glass wall, where complex mathematical formulas and intricate geometric shapes glow with a bright, electric blue light. The camera then switches to a tight close-up of the face of a young East Asian woman, her dark hair pulled back in a bun, thin-rimmed glasses perched on her nose; as she frowns, the glowing blue equations are reflected in her lens. She is completely absorbed. A subsequent cut reveals a detailed shot of a man's hand, a middle-aged Black engineer in a gray cashmere sweater, decisively pointing his open palm at a specific glowing entity on the glass, his finger tracing its curve in the air. This is followed by another shot, a shoulder-length shot of a young White man with shaggy blond hair, who leans over and quickly uses a stylus to add a new string of glowing white variables to the wall. The figure comes alive upon contact. The entire scene is bathed in cool, high-contrast light from the glass, casting sharp shadows and creating an atmosphere of tension, intelligence, and impending breakthrough.", + "seed": 11025, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_12/index_4498_seed_11025_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03670331793837249, + "vlm_logprob": -0.03670331793837249, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.03670331793837249, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, high-tech laboratory with a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a floor-to-ceiling glass wall displaying glowing electric blue mathematical formulas and geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video include a tight close-up of a young East Asian woman with dark hair in a bun and thin-rimmed glasses reflecting the blue equations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a middle-aged Black man in a gray cashmere sweater pointing his open palm at a glowing entity on the glass wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include a shot of a young White man with shaggy blond hair using a stylus to add glowing white variables to the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019296541810035706 + ], + "opposite_logprob_tokens": [ + -4.394296646118164 + ], + "logprob_reward": -0.019296541810035706, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.019296541810035706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1704588681459427, + 0.0706208124756813, + 0.0372343435883522, + 0.06689997017383575, + 0.03096020221710205, + 0.11257854104042053, + 0.0524708516895771, + 0.2720780670642853, + 0.09205428510904312, + 0.09464410692453384 + ], + "credit_map": [ + 0.01277980487793684, + 0.019572487100958824, + 0.019360685721039772, + 0.018620099872350693, + 0.025968141853809357, + 0.020751232281327248, + 0.024398326873779297, + 0.029008084908127785, + 0.0068656085059046745, + 0.012524113059043884, + 0.008407614193856716, + 0.010158155113458633, + 0.00802832841873169, + 0.007004512473940849, + 0.00781284086406231, + 0.009819638915359974, + 0.0040756347589194775, + 0.003322392702102661, + 0.003992806188762188, + 0.002966189756989479, + 0.004471320193260908, + 0.0060511743649840355, + 0.0049756597727537155, + 0.007379165384918451, + 0.008552183397114277, + 0.004161657299846411, + 0.0026555464137345552, + 0.0047730570659041405, + 0.0049578482285141945, + 0.006414116360247135, + 0.018323350697755814, + 0.01706221140921116, + 0.005019193049520254, + 0.004282600712031126, + 0.002047754591330886, + 0.0025133120361715555, + 0.005172559525817633, + 0.0037505198270082474, + 0.0033924193121492863, + 0.004781841300427914, + 0.011025184765458107, + 0.007468175143003464, + 0.005748205818235874, + 0.004011092241853476, + 0.007166177500039339, + 0.031480029225349426, + 0.034047648310661316, + 0.01163202803581953, + 0.006105914246290922, + 0.006116355303674936, + 0.005950039252638817, + 0.006553571671247482, + 0.00670628622174263, + 0.010718167759478092, + 0.005046224221587181, + 0.005274296272546053, + 0.023047659546136856, + 0.02945227175951004, + 0.05494629219174385, + 0.02606520615518093, + 0.034386228770017624, + 0.02037949115037918, + 0.0108792744576931, + 0.07292164862155914, + 0.009899579919874668, + 0.009719126857817173, + 0.011735048145055771, + 0.01339669432491064, + 0.01376284472644329, + 0.014227457344532013, + 0.008493599481880665, + 0.010819925926625729, + 0.009065147489309311, + 0.011435352265834808, + 0.014036525972187519, + 0.007264537271112204, + 0.015269194729626179, + 0.01485882792621851, + 0.009915297850966454, + 0.012799225747585297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.019296541810035706 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, high-tech laboratory with a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a floor-to-ceiling glass wall displaying glowing electric blue mathematical formulas and geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video include a tight close-up of a young East Asian woman with dark hair in a bun and thin-rimmed glasses reflecting the blue equations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a middle-aged Black man in a gray cashmere sweater pointing his open palm at a glowing entity on the glass wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include a shot of a young White man with shaggy blond hair using a stylus to add glowing white variables to the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032263153698295355 + ], + "opposite_logprob_tokens": [ + -5.753226280212402 + ], + "logprob_reward": -0.0032263153698295355, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032263153698295355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09043989330530167, + 0.10001788288354874, + 0.05412236228585243, + 0.04460906609892845, + 0.026472700759768486, + 0.06444620341062546, + 0.2196034938097, + 0.19963307678699493, + 0.05576491728425026, + 0.14489038288593292 + ], + "credit_map": [ + 0.007666498422622681, + 0.009553361684083939, + 0.011163339018821716, + 0.009932659566402435, + 0.011930208653211594, + 0.009670513682067394, + 0.013551314361393452, + 0.016972005367279053, + 0.007044305093586445, + 0.014965597540140152, + 0.010983770713210106, + 0.01105575356632471, + 0.009689196944236755, + 0.00755327707156539, + 0.01903500221669674, + 0.0196909848600626, + 0.003058183006942272, + 0.003856258001178503, + 0.0038265578914433718, + 0.0031205592676997185, + 0.007734301500022411, + 0.01018417440354824, + 0.009599877521395683, + 0.01274245884269476, + 0.006403099279850721, + 0.0058181267231702805, + 0.0030063858721405268, + 0.003929485101252794, + 0.003284165635704994, + 0.003598484443500638, + 0.009766243398189545, + 0.008803081698715687, + 0.004039211198687553, + 0.0036110186483711004, + 0.0024249807465821505, + 0.0022785207256674767, + 0.00394583772867918, + 0.003284275531768799, + 0.0032348560635000467, + 0.0036540019791573286, + 0.006556347478181124, + 0.008205492049455643, + 0.0055788615718483925, + 0.004634445533156395, + 0.013262799941003323, + 0.00900091603398323, + 0.008920568972826004, + 0.008286776952445507, + 0.024304261431097984, + 0.02596280723810196, + 0.018956881016492844, + 0.024623190984129906, + 0.028303692117333412, + 0.04384222999215126, + 0.023168988525867462, + 0.03044145554304123, + 0.018237436190247536, + 0.013812297023832798, + 0.03354260325431824, + 0.03037344291806221, + 0.025648923590779305, + 0.017360592260956764, + 0.008916519582271576, + 0.05174126848578453, + 0.007849456742405891, + 0.008207050152122974, + 0.007803126238286495, + 0.009325004182755947, + 0.006186765152961016, + 0.0054166256450116634, + 0.004815936088562012, + 0.0061609563417732716, + 0.015521938912570477, + 0.01712051033973694, + 0.02059818059206009, + 0.018105821684002876, + 0.018335022032260895, + 0.020804861560463905, + 0.01765589229762554, + 0.016748176887631416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.0032263153698295355 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, high-tech laboratory with a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a floor-to-ceiling glass wall displaying glowing electric blue mathematical formulas and geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video include a tight close-up of a young East Asian woman with dark hair in a bun and thin-rimmed glasses reflecting the blue equations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a middle-aged Black man in a gray cashmere sweater pointing his open palm at a glowing entity on the glass wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include a shot of a young White man with shaggy blond hair using a stylus to add glowing white variables to the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023299604654312134 + ], + "opposite_logprob_tokens": [ + -3.7732996940612793 + ], + "logprob_reward": -0.023299604654312134, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023299604654312134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06858886778354645, + 0.056087758392095566, + 0.05536287650465965, + 0.06236421316862106, + 0.038069579750299454, + 0.13233321905136108, + 0.1370762437582016, + 0.20472201704978943, + 0.055934276431798935, + 0.18946094810962677 + ], + "credit_map": [ + 0.008019545115530491, + 0.007675472646951675, + 0.008278226479887962, + 0.006732405629009008, + 0.01142592541873455, + 0.008913128636777401, + 0.010554361157119274, + 0.006989805959165096, + 0.0056150248274207115, + 0.004337008111178875, + 0.004759364295750856, + 0.0042767878621816635, + 0.005317583214491606, + 0.010383118875324726, + 0.013416487723588943, + 0.007982384413480759, + 0.004501087125390768, + 0.0040369522757828236, + 0.007420276291668415, + 0.0038006301037967205, + 0.009459295310080051, + 0.012405665591359138, + 0.008312507532536983, + 0.005426463671028614, + 0.00662741856649518, + 0.005128816701471806, + 0.0028587293345481157, + 0.004941182676702738, + 0.0038344128988683224, + 0.005545306019484997, + 0.01647224836051464, + 0.016956103965640068, + 0.004845163784921169, + 0.005802328232675791, + 0.004593563266098499, + 0.0036820850800722837, + 0.0033537985291332006, + 0.00466527184471488, + 0.0049406979233026505, + 0.006186671555042267, + 0.012086587958037853, + 0.005585049279034138, + 0.007576884236186743, + 0.004619856830686331, + 0.010395145043730736, + 0.03904300555586815, + 0.042190514504909515, + 0.010836179368197918, + 0.01783801056444645, + 0.015297896228730679, + 0.01252071838825941, + 0.01577530801296234, + 0.014912184327840805, + 0.03037823922932148, + 0.012040236964821815, + 0.018313657492399216, + 0.016455279663205147, + 0.016131144016981125, + 0.036569178104400635, + 0.023956457152962685, + 0.031168822199106216, + 0.019213195890188217, + 0.008264610543847084, + 0.052963342517614365, + 0.008775442838668823, + 0.007654872722923756, + 0.007430918049067259, + 0.011110837571322918, + 0.005332067608833313, + 0.0052244774997234344, + 0.005119988229125738, + 0.005285678897053003, + 0.02755926549434662, + 0.039757028222084045, + 0.026624534279108047, + 0.020005032420158386, + 0.018987923860549927, + 0.01990499347448349, + 0.01967249996960163, + 0.016949687153100967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.023299604654312134 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, high-tech laboratory with a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a floor-to-ceiling glass wall displaying glowing electric blue mathematical formulas and geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video include a tight close-up of a young East Asian woman with dark hair in a bun and thin-rimmed glasses reflecting the blue equations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a middle-aged Black man in a gray cashmere sweater pointing his open palm at a glowing entity on the glass wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include a shot of a young White man with shaggy blond hair using a stylus to add glowing white variables to the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048649102449417114 + ], + "opposite_logprob_tokens": [ + -3.0486490726470947 + ], + "logprob_reward": -0.048649102449417114, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048649102449417114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04167713597416878, + 0.05057368054986, + 0.05007152631878853, + 0.03782344609498978, + 0.014300270937383175, + 0.08029333502054214, + 0.2551058232784271, + 0.28576308488845825, + 0.07834849506616592, + 0.10604315251111984 + ], + "credit_map": [ + 0.005098639987409115, + 0.004852574784308672, + 0.004434247966855764, + 0.005109048914164305, + 0.0051153479143977165, + 0.0052872407250106335, + 0.005511846859008074, + 0.006268186494708061, + 0.003100036410614848, + 0.0037738471291959286, + 0.00455987174063921, + 0.005428244825452566, + 0.005845645442605019, + 0.007435355801135302, + 0.010441163554787636, + 0.009989517740905285, + 0.0030535489786416292, + 0.003022456308826804, + 0.006301867309957743, + 0.004619217943400145, + 0.007531226146966219, + 0.013781559653580189, + 0.006997624412178993, + 0.004764028824865818, + 0.005105885677039623, + 0.0042007192969322205, + 0.0017588100163266063, + 0.002741314936429262, + 0.001865198602899909, + 0.0030593418050557375, + 0.01001721154898405, + 0.009074965491890907, + 0.001590473810210824, + 0.0017732675187289715, + 0.0018901006551459432, + 0.0016597858630120754, + 0.0015686064725741744, + 0.0016051920829340816, + 0.0025307026226073503, + 0.0016821427270770073, + 0.008725672028958797, + 0.005554167553782463, + 0.005865108221769333, + 0.0035256356932222843, + 0.014092366211116314, + 0.019719069823622704, + 0.015852229669690132, + 0.0069590844213962555, + 0.029123369604349136, + 0.0288892462849617, + 0.02403784543275833, + 0.024307018145918846, + 0.027840906754136086, + 0.05874631926417351, + 0.030839605256915092, + 0.031321533024311066, + 0.026482945308089256, + 0.018459226936101913, + 0.052248336374759674, + 0.03390210121870041, + 0.03446844592690468, + 0.02562168426811695, + 0.011204865761101246, + 0.083375483751297, + 0.011394013650715351, + 0.010484847240149975, + 0.010041885077953339, + 0.012446320615708828, + 0.008663399145007133, + 0.009686581790447235, + 0.007026392500847578, + 0.0086050508543849, + 0.01781383715569973, + 0.0168366190046072, + 0.015723979100584984, + 0.009789086878299713, + 0.011573810130357742, + 0.012854747474193573, + 0.012394453398883343, + 0.009056623093783855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.048649102449417114 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, high-tech laboratory with a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a floor-to-ceiling glass wall displaying glowing electric blue mathematical formulas and geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video include a tight close-up of a young East Asian woman with dark hair in a bun and thin-rimmed glasses reflecting the blue equations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a middle-aged Black man in a gray cashmere sweater pointing his open palm at a glowing entity on the glass wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include a shot of a young White man with shaggy blond hair using a stylus to add glowing white variables to the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904502540826797 + ], + "opposite_logprob_tokens": [ + -2.4640450477600098 + ], + "logprob_reward": -0.08904502540826797, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904502540826797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02621370740234852, + 0.03633036091923714, + 0.05833149701356888, + 0.03350256383419037, + 0.037089888006448746, + 0.05627768486738205, + 0.23353976011276245, + 0.4335959255695343, + 0.03385104611515999, + 0.05126754567027092 + ], + "credit_map": [ + 0.002709000138565898, + 0.0027400224935263395, + 0.003863281337544322, + 0.002820356981828809, + 0.004372438415884972, + 0.003140972461551428, + 0.0036676181480288506, + 0.0029000157956033945, + 0.0034217825159430504, + 0.0032161741983145475, + 0.003200344741344452, + 0.003222859464585781, + 0.003018760122358799, + 0.00534937996417284, + 0.0087859146296978, + 0.006115146446973085, + 0.0036217800807207823, + 0.004172945395112038, + 0.0071593113243579865, + 0.004574573133140802, + 0.00904273334890604, + 0.015088635496795177, + 0.007954282686114311, + 0.006717232055962086, + 0.004261992871761322, + 0.004030593205243349, + 0.0020236403215676546, + 0.0033689348492771387, + 0.004305476788431406, + 0.004038104321807623, + 0.005606875754892826, + 0.005866941995918751, + 0.005932594649493694, + 0.005395194515585899, + 0.0037897592410445213, + 0.0034304694272577763, + 0.005437563639134169, + 0.004810038488358259, + 0.003958526998758316, + 0.004335741046816111, + 0.005886020604521036, + 0.0028931819833815098, + 0.002882755361497402, + 0.0014115599915385246, + 0.0039770593866705894, + 0.01748294197022915, + 0.01734795607626438, + 0.004396202974021435, + 0.020616963505744934, + 0.023395713418722153, + 0.038213517516851425, + 0.03168491646647453, + 0.026005947962403297, + 0.02957708016037941, + 0.0289117731153965, + 0.03513385355472565, + 0.03758087009191513, + 0.024342020973563194, + 0.08526109904050827, + 0.04805059731006622, + 0.058902230113744736, + 0.03631645068526268, + 0.017028965055942535, + 0.12611371278762817, + 0.005605619866400957, + 0.005737444385886192, + 0.004837296437472105, + 0.006380055565387011, + 0.0028343733865767717, + 0.0025777772534638643, + 0.002804004354402423, + 0.0030744767282158136, + 0.0059575289487838745, + 0.008531409315764904, + 0.007925047539174557, + 0.007132755126804113, + 0.004867264535278082, + 0.004458265379071236, + 0.00734754279255867, + 0.005047732498496771 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 2 + ], + "selected_reward": -0.08904502540826797 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "series of rapid, dynamic cuts captures the intense collaborative energy of a small, diverse team of engineers. The scene begins with a low-angle, medium shot of the group huddled in a dimly lit, high-tech laboratory with a polished concrete floor. Their silhouettes are silhouetted against a vast, floor-to-ceiling glass wall, where complex mathematical formulas and intricate geometric shapes glow with a bright, electric blue light. The camera then switches to a tight close-up of the face of a young East Asian woman, her dark hair pulled back in a bun, thin-rimmed glasses perched on her nose; as she frowns, the glowing blue equations are reflected in her lens. She is completely absorbed. A subsequent cut reveals a detailed shot of a man's hand, a middle-aged Black engineer in a gray cashmere sweater, decisively pointing his open palm at a specific glowing entity on the glass, his finger tracing its curve in the air. This is followed by another shot, a shoulder-length shot of a young White man with shaggy blond hair, who leans over and quickly uses a stylus to add a new string of glowing white variables to the wall. The figure comes alive upon contact. The entire scene is bathed in cool, high-contrast light from the glass, casting sharp shadows and creating an atmosphere of tension, intelligence, and impending breakthrough.", + "seed": 11026, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_12/index_4498_seed_11026_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02817827956750989, + "vlm_logprob": -0.02817827956750989, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.02817827956750989, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, high-tech laboratory with a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a floor-to-ceiling glass wall displaying glowing electric blue mathematical formulas and geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video include a tight close-up of a young East Asian woman with dark hair in a bun and thin-rimmed glasses reflecting the blue equations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a middle-aged Black man in a gray cashmere sweater pointing his open palm at a glowing entity on the glass wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include a shot of a young White man with shaggy blond hair using a stylus to add glowing white variables to the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01083337888121605 + ], + "opposite_logprob_tokens": [ + -5.760833263397217 + ], + "logprob_reward": -0.01083337888121605, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.01083337888121605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21488693356513977, + 0.11593293398618698, + 0.0809948593378067, + 0.07704439759254456, + 0.2026372253894806, + 0.054626140743494034, + 0.06590846180915833, + 0.037304218858480453, + 0.07177519053220749, + 0.07888966053724289 + ], + "credit_map": [ + 0.027432311326265335, + 0.030669156461954117, + 0.024152854457497597, + 0.028158608824014664, + 0.03127099946141243, + 0.020976701751351357, + 0.015972672030329704, + 0.03625362738966942, + 0.01264334749430418, + 0.012547871097922325, + 0.02552594430744648, + 0.01675311289727688, + 0.006670094560831785, + 0.0072225360199809074, + 0.0246175155043602, + 0.00995250791311264, + 0.011741257272660732, + 0.009101365692913532, + 0.01167588122189045, + 0.009453863836824894, + 0.011481846682727337, + 0.007902126759290695, + 0.0100628063082695, + 0.00957570131868124, + 0.01338707935065031, + 0.009209821932017803, + 0.008781646378338337, + 0.010357670485973358, + 0.011852264404296875, + 0.00552353635430336, + 0.01078202947974205, + 0.007150341756641865, + 0.011315218172967434, + 0.007124085444957018, + 0.006941467523574829, + 0.013258468359708786, + 0.02136744000017643, + 0.06603074073791504, + 0.010308573953807354, + 0.06629122793674469, + 0.006277818698436022, + 0.006994639523327351, + 0.004353711381554604, + 0.009936004877090454, + 0.00859067402780056, + 0.006236901506781578, + 0.006967318244278431, + 0.0052690664306283, + 0.008498193696141243, + 0.008368490263819695, + 0.011580092832446098, + 0.008645365945994854, + 0.006505957338958979, + 0.008865839801728725, + 0.007555315271019936, + 0.005889197811484337, + 0.004146941006183624, + 0.005941694136708975, + 0.005167018156498671, + 0.007532981224358082, + 0.003925530705600977, + 0.0029222832527011633, + 0.003945421427488327, + 0.00372234801761806, + 0.007932214997708797, + 0.008646691218018532, + 0.00996748823672533, + 0.007853418588638306, + 0.012227300554513931, + 0.006456204690039158, + 0.014296690002083778, + 0.004395178519189358, + 0.007279854733496904, + 0.011192306876182556, + 0.011011087335646152, + 0.014165789820253849, + 0.010151088237762451, + 0.009038863703608513, + 0.009828447364270687, + 0.00622220803052187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.01083337888121605 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, high-tech laboratory with a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a floor-to-ceiling glass wall displaying glowing electric blue mathematical formulas and geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video include a tight close-up of a young East Asian woman with dark hair in a bun and thin-rimmed glasses reflecting the blue equations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a middle-aged Black man in a gray cashmere sweater pointing his open palm at a glowing entity on the glass wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include a shot of a young White man with shaggy blond hair using a stylus to add glowing white variables to the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022273273207247257 + ], + "opposite_logprob_tokens": [ + -6.127227306365967 + ], + "logprob_reward": -0.0022273273207247257, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022273273207247257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09159956872463226, + 0.11001860350370407, + 0.14217860996723175, + 0.060922350734472275, + 0.11367350816726685, + 0.03386133909225464, + 0.16000734269618988, + 0.03363398462533951, + 0.12799423933029175, + 0.12611046433448792 + ], + "credit_map": [ + 0.013891139067709446, + 0.012817876413464546, + 0.013436083681881428, + 0.013391807675361633, + 0.007925757206976414, + 0.00741058774292469, + 0.006312257144600153, + 0.016414059326052666, + 0.013777862302958965, + 0.018100036308169365, + 0.01831326261162758, + 0.022677883505821228, + 0.006820331793278456, + 0.00721427146345377, + 0.012939241714775562, + 0.010175716131925583, + 0.017702652141451836, + 0.01675448752939701, + 0.017527028918266296, + 0.02041778527200222, + 0.015547864139080048, + 0.013419273309409618, + 0.02289525792002678, + 0.01791425794363022, + 0.007580410689115524, + 0.006601108703762293, + 0.006275431253015995, + 0.006385688204318285, + 0.011696175672113895, + 0.004984819330275059, + 0.011031094007194042, + 0.006367625202983618, + 0.006073701195418835, + 0.004230487626045942, + 0.004668698646128178, + 0.005549106281250715, + 0.011292258277535439, + 0.037732575088739395, + 0.005217592231929302, + 0.03890908509492874, + 0.004960719030350447, + 0.004218765534460545, + 0.002698383294045925, + 0.004925754852592945, + 0.004694988019764423, + 0.0042946175672113895, + 0.004654739983379841, + 0.003413371741771698, + 0.0226668119430542, + 0.024030687287449837, + 0.01904759742319584, + 0.022931639105081558, + 0.016266653314232826, + 0.022385872900485992, + 0.018671320751309395, + 0.014006761834025383, + 0.002898742910474539, + 0.005814072210341692, + 0.004852603189647198, + 0.005481068976223469, + 0.004524271935224533, + 0.0030415731016546488, + 0.004579708445817232, + 0.002441942458972335, + 0.015254249796271324, + 0.01755593903362751, + 0.018384994938969612, + 0.017837123945355415, + 0.02120743878185749, + 0.009119909256696701, + 0.02091173827648163, + 0.007722842041403055, + 0.010781564749777317, + 0.017087239772081375, + 0.019371017813682556, + 0.020635850727558136, + 0.018115468323230743, + 0.015579565428197384, + 0.012696168385446072, + 0.011843581683933735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 8 + ], + "selected_reward": -0.0022273273207247257 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, high-tech laboratory with a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a floor-to-ceiling glass wall displaying glowing electric blue mathematical formulas and geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video include a tight close-up of a young East Asian woman with dark hair in a bun and thin-rimmed glasses reflecting the blue equations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a middle-aged Black man in a gray cashmere sweater pointing his open palm at a glowing entity on the glass wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include a shot of a young White man with shaggy blond hair using a stylus to add glowing white variables to the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981775999069214 + ], + "opposite_logprob_tokens": [ + -3.529817819595337 + ], + "logprob_reward": -0.02981775999069214, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981775999069214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11054926365613937, + 0.053696319460868835, + 0.37267911434173584, + 0.056819092482328415, + 0.1684769093990326, + 0.02113388478755951, + 0.03974619880318642, + 0.018792079761624336, + 0.06913945823907852, + 0.0889676958322525 + ], + "credit_map": [ + 0.015463333576917648, + 0.017847426235675812, + 0.010436732321977615, + 0.016869159415364265, + 0.01064821146428585, + 0.014385095797479153, + 0.008265813812613487, + 0.01663348264992237, + 0.006418780889362097, + 0.006965273059904575, + 0.011256386525928974, + 0.006571852136403322, + 0.003114476799964905, + 0.004105948843061924, + 0.010466697625815868, + 0.004796908237040043, + 0.05868123844265938, + 0.04055299237370491, + 0.05559983849525452, + 0.04466880112886429, + 0.0396374836564064, + 0.031337372958660126, + 0.059911735355854034, + 0.04228965565562248, + 0.00836352352052927, + 0.006063300184905529, + 0.007446678355336189, + 0.0063085067085921764, + 0.012048795819282532, + 0.004758189432322979, + 0.0071269129402935505, + 0.0047031887806952, + 0.007779449224472046, + 0.006722557824105024, + 0.0064818668179214, + 0.008336884900927544, + 0.01611778326332569, + 0.05556073412299156, + 0.007537692319601774, + 0.05993994325399399, + 0.003143860260024667, + 0.0032466871198266745, + 0.00168360467068851, + 0.002396129537373781, + 0.002317980630323291, + 0.0024581411853432655, + 0.002689102664589882, + 0.0031983796507120132, + 0.005801829509437084, + 0.005681509152054787, + 0.006603929214179516, + 0.00535882031545043, + 0.004271422512829304, + 0.004808470141142607, + 0.0037816453259438276, + 0.0034385723993182182, + 0.0015478786081075668, + 0.0022643087431788445, + 0.0025840105954557657, + 0.0036850834731012583, + 0.0034487643279135227, + 0.0018376735970377922, + 0.0016791698290035129, + 0.0017451904714107513, + 0.007642847485840321, + 0.010011895559728146, + 0.010391009971499443, + 0.008379739709198475, + 0.008410844020545483, + 0.0069496710784733295, + 0.012379501946270466, + 0.004973953124135733, + 0.011083871126174927, + 0.015264112502336502, + 0.011516683734953403, + 0.014396646991372108, + 0.01252517569810152, + 0.011152169667184353, + 0.006240347865968943, + 0.006788694765418768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.02981775999069214 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, high-tech laboratory with a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a floor-to-ceiling glass wall displaying glowing electric blue mathematical formulas and geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video include a tight close-up of a young East Asian woman with dark hair in a bun and thin-rimmed glasses reflecting the blue equations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a middle-aged Black man in a gray cashmere sweater pointing his open palm at a glowing entity on the glass wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include a shot of a young White man with shaggy blond hair using a stylus to add glowing white variables to the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043051525950431824 + ], + "opposite_logprob_tokens": [ + -3.1680514812469482 + ], + "logprob_reward": -0.043051525950431824, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043051525950431824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1002844050526619, + 0.08821674436330795, + 0.1445249319076538, + 0.0546005554497242, + 0.10086201876401901, + 0.03092554397881031, + 0.09856878221035004, + 0.043675731867551804, + 0.12933552265167236, + 0.20900574326515198 + ], + "credit_map": [ + 0.012711169198155403, + 0.016048206016421318, + 0.012462185695767403, + 0.01679850183427334, + 0.008664087392389774, + 0.006853395141661167, + 0.007408080156892538, + 0.019338790327310562, + 0.00901309959590435, + 0.008301256224513054, + 0.016857989132404327, + 0.009787370450794697, + 0.007880052551627159, + 0.006533718667924404, + 0.020172348245978355, + 0.00967091228812933, + 0.018297143280506134, + 0.014366031624376774, + 0.01738913170993328, + 0.01948430761694908, + 0.023578204214572906, + 0.01215672679245472, + 0.022738421335816383, + 0.016514977440238, + 0.005682608112692833, + 0.0061228615231812, + 0.007699658162891865, + 0.009677194058895111, + 0.00932356994599104, + 0.00424926308915019, + 0.006878793239593506, + 0.004966613836586475, + 0.007105983328074217, + 0.004583798814564943, + 0.007940578274428844, + 0.009079447947442532, + 0.01058268640190363, + 0.026349252089858055, + 0.006087906192988157, + 0.02913237176835537, + 0.005987162701785564, + 0.004635719116777182, + 0.002402411075308919, + 0.003725350834429264, + 0.00288244616240263, + 0.0035476970952004194, + 0.0036948788911104202, + 0.004049880430102348, + 0.011919473297894001, + 0.014744026586413383, + 0.017499327659606934, + 0.015486185438930988, + 0.01069403812289238, + 0.01038991566747427, + 0.008513134904205799, + 0.00932268425822258, + 0.003978554159402847, + 0.007526760455220938, + 0.005152702331542969, + 0.004925338085740805, + 0.006714958231896162, + 0.005367805249989033, + 0.0055557359009981155, + 0.004453879315406084, + 0.011567327193915844, + 0.016840778291225433, + 0.01992896944284439, + 0.021394293755292892, + 0.014114147052168846, + 0.016480455175042152, + 0.014875547960400581, + 0.01413401123136282, + 0.023901240900158882, + 0.03333568572998047, + 0.032872527837753296, + 0.03313254937529564, + 0.019557291641831398, + 0.03056832030415535, + 0.016270799562335014, + 0.019367340952157974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 8 + ], + "selected_reward": -0.043051525950431824 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, high-tech laboratory with a polished concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a floor-to-ceiling glass wall displaying glowing electric blue mathematical formulas and geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video include a tight close-up of a young East Asian woman with dark hair in a bun and thin-rimmed glasses reflecting the blue equations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a middle-aged Black man in a gray cashmere sweater pointing his open palm at a glowing entity on the glass wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include a shot of a young White man with shaggy blond hair using a stylus to add glowing white variables to the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496140569448471 + ], + "opposite_logprob_tokens": [ + -2.9299614429473877 + ], + "logprob_reward": -0.05496140569448471, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496140569448471, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10677773505449295, + 0.04132485389709473, + 0.1601448357105255, + 0.09507092833518982, + 0.12438303232192993, + 0.026031453162431717, + 0.06412416696548462, + 0.03319048509001732, + 0.16494804620742798, + 0.18400448560714722 + ], + "credit_map": [ + 0.015239492990076542, + 0.016761139035224915, + 0.013659561052918434, + 0.01343432255089283, + 0.009694173000752926, + 0.011507801711559296, + 0.0066740126349031925, + 0.019807221367955208, + 0.005275040399283171, + 0.0072455802001059055, + 0.005385054741054773, + 0.005218853242695332, + 0.002918611979112029, + 0.004877554252743721, + 0.005636244080960751, + 0.0047679077833890915, + 0.023092614486813545, + 0.014153026975691319, + 0.02065010368824005, + 0.017566455528140068, + 0.025085527449846268, + 0.014714189805090427, + 0.028101138770580292, + 0.0167817585170269, + 0.011797602288424969, + 0.011700741946697235, + 0.009465665556490421, + 0.010517426766455173, + 0.02087556943297386, + 0.007709936238825321, + 0.013836942613124847, + 0.009167039766907692, + 0.0065299528650939465, + 0.0050275856629014015, + 0.0066983518190681934, + 0.00871439278125763, + 0.012439078651368618, + 0.03796295076608658, + 0.006443637888878584, + 0.0405670702457428, + 0.0036432649940252304, + 0.00420029042288661, + 0.002083021914586425, + 0.0032236604019999504, + 0.002937037730589509, + 0.003169841831550002, + 0.002880384447053075, + 0.003893951652571559, + 0.008519981056451797, + 0.007499165367335081, + 0.012271687388420105, + 0.00944293662905693, + 0.006811443716287613, + 0.006504763849079609, + 0.008074630051851273, + 0.004999556113034487, + 0.002967020496726036, + 0.005503621883690357, + 0.004537551663815975, + 0.004207977093756199, + 0.0068982671946287155, + 0.00188440119381994, + 0.00458298996090889, + 0.002608655020594597, + 0.019505009055137634, + 0.020223243162035942, + 0.022368408739566803, + 0.024175595492124557, + 0.025004025548696518, + 0.015330234542489052, + 0.026565492153167725, + 0.011776014231145382, + 0.01861598715186119, + 0.02524442784488201, + 0.021282214671373367, + 0.024298446252942085, + 0.03036537393927574, + 0.029182184487581253, + 0.01833217404782772, + 0.016683652997016907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.05496140569448471 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_13/index_tensor([8876, 8876])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_13/index_tensor([8876, 8876])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b57ac3c3389deab72d79301f134b3889ca69bba1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_13/index_tensor([8876, 8876])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, medium shot, filmed in cinematic slow motion, the camera steadily following the subject, captures a vigilant red fox, its fur a flaming orange-red, nestled amidst a vast snowy plain bathed in the first golden rays of morning sunlight. The low, slanting sunlight casts long, blue shadows across the pristine white of the snow, while the fox's body, taut and graceful, arches in a powerful arc. Its pointed muzzle nearly touches the snow, its dark eyes fixed on the unseen movements beneath, its muscles poised for action. Suddenly, with a sudden thrust of its hind legs, its entire body leaps vertically into the air, its front paws clasped together and pointed downward like an arrow. With unparalleled swiftness, it plunges headfirst into the deep snow. As it plunges into the drift, a shower of shimmering, sun-lit snow powder explodes in all directions, clearly visible in slow motion. Only the tip of its furry tail remains briefly above the surface. The whole picture, with its sharp contrast of warm golden light and cold blue shadows, presents a primitive, focused and extremely beautiful life tension, full of strong cinematic texture.", + "seed": 11907, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_13/index_8876_seed_11907_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.011539457179605962, + "vlm_logprob": -0.011539457179605962, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.011539457179605962, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast snowy plain bathed in morning sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fox's fur a flaming orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fox leap vertically into the air and then plunge headfirst into the deep snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a shower of shimmering, sun-lit snow powder visible exploding in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video filmed in cinematic slow motion with a low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010591013357043266 + ], + "opposite_logprob_tokens": [ + -5.13559103012085 + ], + "logprob_reward": -0.010591013357043266, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.010591013357043266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08223487436771393, + 0.03281574696302414, + 0.22421714663505554, + 0.06332175433635712, + 0.058498866856098175, + 0.03385593742132187, + 0.12539727985858917, + 0.1900283694267273, + 0.06540621072053909, + 0.12422384321689606 + ], + "credit_map": [ + 0.010478203184902668, + 0.010162083432078362, + 0.008023015223443508, + 0.011311296373605728, + 0.01115056499838829, + 0.01075053308159113, + 0.008602187968790531, + 0.011756985448300838, + 0.0029923676047474146, + 0.0040322039276361465, + 0.005326683633029461, + 0.003972903825342655, + 0.004154040943831205, + 0.0033330742735415697, + 0.005319914780557156, + 0.0036845554132014513, + 0.00929795391857624, + 0.007471263408660889, + 0.008232304826378822, + 0.012819299474358559, + 0.012372026219964027, + 0.07393815368413925, + 0.015390949323773384, + 0.08469520509243011, + 0.008572043851017952, + 0.007376485038548708, + 0.007098815403878689, + 0.00870549213141203, + 0.007587105967104435, + 0.009793691337108612, + 0.003973832819610834, + 0.010214285925030708, + 0.006493471562862396, + 0.004637107718735933, + 0.013140900991857052, + 0.0157085619866848, + 0.002853796584531665, + 0.0040255445055663586, + 0.004561055451631546, + 0.007078428287059069, + 0.003132504178211093, + 0.0031791774090379477, + 0.003780578961595893, + 0.002619297942146659, + 0.002923369174823165, + 0.0053025828674435616, + 0.005589467007666826, + 0.007328961510211229, + 0.01072970125824213, + 0.008061648346483707, + 0.03832606226205826, + 0.009622800163924694, + 0.007649362087249756, + 0.03495223447680473, + 0.008057354018092155, + 0.00799812562763691, + 0.011468987911939621, + 0.005414592567831278, + 0.07292873412370682, + 0.007606619503349066, + 0.004505119286477566, + 0.019987022504210472, + 0.0058126128278672695, + 0.06230466812849045, + 0.007318312302231789, + 0.00950010959059, + 0.008499163202941418, + 0.00882115587592125, + 0.005421220324933529, + 0.006175247952342033, + 0.00622385973110795, + 0.013447137549519539, + 0.012027768418192863, + 0.009709019213914871, + 0.010519454255700111, + 0.03058977983891964, + 0.01196640357375145, + 0.010061266832053661, + 0.016756931319832802, + 0.022593218833208084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 6 + ], + "selected_reward": -0.010591013357043266 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast snowy plain bathed in morning sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fox's fur a flaming orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fox leap vertically into the air and then plunge headfirst into the deep snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a shower of shimmering, sun-lit snow powder visible exploding in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video filmed in cinematic slow motion with a low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014212576672434807 + ], + "opposite_logprob_tokens": [ + -4.264212608337402 + ], + "logprob_reward": -0.014212576672434807, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014212576672434807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1165051981806755, + 0.0816592276096344, + 0.1485641896724701, + 0.102805495262146, + 0.08975733071565628, + 0.11211330443620682, + 0.04884013533592224, + 0.21110038459300995, + 0.04179339483380318, + 0.04686138778924942 + ], + "credit_map": [ + 0.018827397376298904, + 0.019119365140795708, + 0.011302926577627659, + 0.013450555503368378, + 0.011843557469546795, + 0.013977518305182457, + 0.012130122631788254, + 0.015853742137551308, + 0.008966046385467052, + 0.01359170489013195, + 0.009447570890188217, + 0.009627282619476318, + 0.008851887658238411, + 0.00884962361305952, + 0.012032234109938145, + 0.01029286440461874, + 0.013618819415569305, + 0.012944474816322327, + 0.013717048801481724, + 0.013722794130444527, + 0.02018500305712223, + 0.029226556420326233, + 0.01053527183830738, + 0.0346142053604126, + 0.013066677376627922, + 0.009973330423235893, + 0.010235593654215336, + 0.01347841415554285, + 0.012414781376719475, + 0.023425418883562088, + 0.009377432987093925, + 0.010833842679858208, + 0.00788096897304058, + 0.007226136047393084, + 0.02638823352754116, + 0.02150135114789009, + 0.007392426021397114, + 0.008105257526040077, + 0.005435578990727663, + 0.005827369634062052, + 0.005286794155836105, + 0.007328739855438471, + 0.014944089576601982, + 0.008943332359194756, + 0.009195931255817413, + 0.009565597400069237, + 0.018373625352978706, + 0.03847517818212509, + 0.0051704091019928455, + 0.009805689565837383, + 0.006648808717727661, + 0.006055621895939112, + 0.0063066305592656136, + 0.0041392226703464985, + 0.005955791566520929, + 0.004757958929985762, + 0.013146204873919487, + 0.0069141932763159275, + 0.08304383605718613, + 0.007695676293224096, + 0.005669314414262772, + 0.028721990063786507, + 0.005402387585490942, + 0.06050676107406616, + 0.004643679596483707, + 0.005408439785242081, + 0.004460984840989113, + 0.007266293279826641, + 0.0030254286248236895, + 0.003444519592449069, + 0.006233698688447475, + 0.0073103467002511024, + 0.008373108692467213, + 0.0037960316985845566, + 0.004650583956390619, + 0.006222246214747429, + 0.00455878209322691, + 0.004557587672024965, + 0.006953289266675711, + 0.0077497526071965694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.014212576672434807 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast snowy plain bathed in morning sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fox's fur a flaming orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fox leap vertically into the air and then plunge headfirst into the deep snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a shower of shimmering, sun-lit snow powder visible exploding in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video filmed in cinematic slow motion with a low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011074623093008995 + ], + "opposite_logprob_tokens": [ + -4.511074542999268 + ], + "logprob_reward": -0.011074623093008995, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011074623093008995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05504839867353439, + 0.04551474750041962, + 0.03442160412669182, + 0.09204109758138657, + 0.0490180142223835, + 0.06666439026594162, + 0.1818126142024994, + 0.2938491404056549, + 0.09760185331106186, + 0.08402809500694275 + ], + "credit_map": [ + 0.005698475055396557, + 0.003666651900857687, + 0.009697668254375458, + 0.008468015119433403, + 0.0055524311028420925, + 0.007439498323947191, + 0.004761763382703066, + 0.009763899259269238, + 0.005441105458885431, + 0.0052732606418430805, + 0.005469591822475195, + 0.005372142884880304, + 0.007264654152095318, + 0.004853221122175455, + 0.006010282784700394, + 0.005830487236380577, + 0.002371259033679962, + 0.0023549527395516634, + 0.003570150351151824, + 0.003093058941885829, + 0.002569192321971059, + 0.008633315563201904, + 0.0029110321775078773, + 0.00891864113509655, + 0.011883876286447048, + 0.013842412270605564, + 0.01033417321741581, + 0.014447353780269623, + 0.013342193327844143, + 0.01120985671877861, + 0.009301485493779182, + 0.007679745554924011, + 0.004237790126353502, + 0.005202796310186386, + 0.010386763140559196, + 0.010798034258186817, + 0.0038737994618713856, + 0.005678270943462849, + 0.0034440758172422647, + 0.005396487191319466, + 0.007612745743244886, + 0.004661391489207745, + 0.00741552421823144, + 0.007569291163235903, + 0.00521804066374898, + 0.006166168954223394, + 0.013651387766003609, + 0.014369843527674675, + 0.012852303683757782, + 0.011684982106089592, + 0.04367516562342644, + 0.027475960552692413, + 0.011006400920450687, + 0.03857927769422531, + 0.012081139720976353, + 0.02445737086236477, + 0.017908558249473572, + 0.011004194617271423, + 0.11509013921022415, + 0.01026438269764185, + 0.009310814552009106, + 0.04000565782189369, + 0.010347330011427402, + 0.07991807907819748, + 0.009326500818133354, + 0.01648179069161415, + 0.011345017701387405, + 0.011440996080636978, + 0.007474181242287159, + 0.007946964353322983, + 0.009607166051864624, + 0.02397923730313778, + 0.022165924310684204, + 0.007798811420798302, + 0.009085467085242271, + 0.011708186939358711, + 0.0053521194495260715, + 0.006324614863842726, + 0.010482982732355595, + 0.011109990999102592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.011074623093008995 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast snowy plain bathed in morning sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fox's fur a flaming orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fox leap vertically into the air and then plunge headfirst into the deep snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a shower of shimmering, sun-lit snow powder visible exploding in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video filmed in cinematic slow motion with a low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628121413290501 + ], + "opposite_logprob_tokens": [ + -4.882627964019775 + ], + "logprob_reward": -0.007628121413290501, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628121413290501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0369499996304512, + 0.02606029249727726, + 0.037878841161727905, + 0.05710712447762489, + 0.02794247679412365, + 0.28107303380966187, + 0.09959570318460464, + 0.3030886948108673, + 0.059573691338300705, + 0.07073009014129639 + ], + "credit_map": [ + 0.0042061652056872845, + 0.00513882003724575, + 0.0032642227597534657, + 0.005855241324752569, + 0.003143219044432044, + 0.004084540531039238, + 0.0043118842877447605, + 0.006945905741304159, + 0.003533080918714404, + 0.0038269979413598776, + 0.003125829389318824, + 0.003180429805070162, + 0.004555698484182358, + 0.0028995138127356768, + 0.002096776617690921, + 0.002841965528205037, + 0.0045460667461156845, + 0.0036533356178551912, + 0.007070856634527445, + 0.004533481318503618, + 0.004400391131639481, + 0.004608277231454849, + 0.003388581331819296, + 0.005677849054336548, + 0.010217384435236454, + 0.007202254608273506, + 0.004986759275197983, + 0.008409030735492706, + 0.005943167023360729, + 0.005382869392633438, + 0.0036382966209203005, + 0.011327364481985569, + 0.003250685753300786, + 0.0037617331836372614, + 0.005586185958236456, + 0.005374982953071594, + 0.0024495935067534447, + 0.002288838615640998, + 0.0026445293333381414, + 0.002585925394669175, + 0.013441738672554493, + 0.011795169673860073, + 0.03706536814570427, + 0.021033961325883865, + 0.015160994604229927, + 0.02311071753501892, + 0.06127661466598511, + 0.09818848967552185, + 0.009787819348275661, + 0.008917086757719517, + 0.025370387360453606, + 0.008020300418138504, + 0.006439636927098036, + 0.02434375509619713, + 0.009462060406804085, + 0.007254661060869694, + 0.018631869927048683, + 0.01098915096372366, + 0.1193634495139122, + 0.01090113166719675, + 0.007696099579334259, + 0.03761252760887146, + 0.0085277259349823, + 0.08936674147844315, + 0.006043255794793367, + 0.00948481634259224, + 0.0052650729194283485, + 0.016099097207188606, + 0.0047684782184660435, + 0.00467295665293932, + 0.007470426615327597, + 0.005769588053226471, + 0.015698490664362907, + 0.005455020349472761, + 0.007605253718793392, + 0.012039698660373688, + 0.004949397407472134, + 0.004740082658827305, + 0.008306730538606644, + 0.011935418471693993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.007628121413290501 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast snowy plain bathed in morning sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fox's fur a flaming orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fox leap vertically into the air and then plunge headfirst into the deep snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a shower of shimmering, sun-lit snow powder visible exploding in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video filmed in cinematic slow motion with a low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014190951362252235 + ], + "opposite_logprob_tokens": [ + -4.264191150665283 + ], + "logprob_reward": -0.014190951362252235, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014190951362252235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054816726595163345, + 0.02740110643208027, + 0.15842294692993164, + 0.08965937048196793, + 0.07018602639436722, + 0.17017412185668945, + 0.12724699079990387, + 0.2173408567905426, + 0.033196158707141876, + 0.051555708050727844 + ], + "credit_map": [ + 0.005659224931150675, + 0.005067616701126099, + 0.008412966504693031, + 0.009778010658919811, + 0.006908081471920013, + 0.006170519161969423, + 0.0056911190040409565, + 0.007129189558327198, + 0.004132384434342384, + 0.003390163416042924, + 0.0031783413141965866, + 0.0026119397953152657, + 0.003743858076632023, + 0.0032099420204758644, + 0.0030692138243466616, + 0.004065264016389847, + 0.006610805634409189, + 0.006489355582743883, + 0.0076444330625236034, + 0.011783348396420479, + 0.007973003201186657, + 0.04851815477013588, + 0.010478565469384193, + 0.05892527475953102, + 0.017326567322015762, + 0.009479166008532047, + 0.008906039409339428, + 0.012759830802679062, + 0.008128561079502106, + 0.010397403500974178, + 0.006056715734302998, + 0.016605084761977196, + 0.008585117757320404, + 0.007990947924554348, + 0.011170601472258568, + 0.012844611890614033, + 0.008260383270680904, + 0.007296421565115452, + 0.006225061137229204, + 0.007812876254320145, + 0.015492373146116734, + 0.015200641006231308, + 0.01925184391438961, + 0.018238630145788193, + 0.010623752139508724, + 0.012985842302441597, + 0.03224124014377594, + 0.046139802783727646, + 0.012655023485422134, + 0.009646324440836906, + 0.01633717678487301, + 0.02922520972788334, + 0.008441654965281487, + 0.012948418036103249, + 0.010953852906823158, + 0.027039337903261185, + 0.013382204808294773, + 0.010469957254827023, + 0.07058942317962646, + 0.010024870745837688, + 0.006851698271930218, + 0.016235042363405228, + 0.008362838998436928, + 0.08142480999231339, + 0.0031778542324900627, + 0.004015116021037102, + 0.003330862382426858, + 0.007197272032499313, + 0.003370695048943162, + 0.002781268674880266, + 0.003743316512554884, + 0.005579770542681217, + 0.008128492161631584, + 0.005250516347587109, + 0.005448597017675638, + 0.008814942091703415, + 0.004538648761808872, + 0.0036570653319358826, + 0.0067454129457473755, + 0.008972038514912128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 2 + ], + "selected_reward": -0.014190951362252235 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium shot, filmed in cinematic slow motion, the camera steadily following the subject, captures a vigilant red fox, its fur a flaming orange-red, nestled amidst a vast snowy plain bathed in the first golden rays of morning sunlight. The low, slanting sunlight casts long, blue shadows across the pristine white of the snow, while the fox's body, taut and graceful, arches in a powerful arc. Its pointed muzzle nearly touches the snow, its dark eyes fixed on the unseen movements beneath, its muscles poised for action. Suddenly, with a sudden thrust of its hind legs, its entire body leaps vertically into the air, its front paws clasped together and pointed downward like an arrow. With unparalleled swiftness, it plunges headfirst into the deep snow. As it plunges into the drift, a shower of shimmering, sun-lit snow powder explodes in all directions, clearly visible in slow motion. Only the tip of its furry tail remains briefly above the surface. The whole picture, with its sharp contrast of warm golden light and cold blue shadows, presents a primitive, focused and extremely beautiful life tension, full of strong cinematic texture.", + "seed": 11908, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_13/index_8876_seed_11908_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.013233546493574977, + "vlm_logprob": -0.013233546493574977, + "vlm_yes_no_margin": 4.649999952316284, + "vlm_token_credit": -0.013233546493574977, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast snowy plain bathed in morning sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fox's fur a flaming orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fox leap vertically into the air and then plunge headfirst into the deep snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a shower of shimmering, sun-lit snow powder visible exploding in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video filmed in cinematic slow motion with a low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02395787462592125 + ], + "opposite_logprob_tokens": [ + -4.2739577293396 + ], + "logprob_reward": -0.02395787462592125, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.02395787462592125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05080699175596237, + 0.013127093203365803, + 0.009182059206068516, + 0.008200755342841148, + 0.020819325000047684, + 0.019273750483989716, + 0.015241226181387901, + 0.8155792355537415, + 0.029086856171488762, + 0.0186826940625906 + ], + "credit_map": [ + 0.008049334399402142, + 0.006903826259076595, + 0.006393115501850843, + 0.005545728839933872, + 0.005884978920221329, + 0.00633489154279232, + 0.005639735143631697, + 0.006055379752069712, + 0.002412190428003669, + 0.002042502397671342, + 0.0015801340341567993, + 0.0009283012477681041, + 0.0016967101255431771, + 0.0019516603788360953, + 0.0008625349728390574, + 0.0016530592693015933, + 0.0009838675614446402, + 0.0009531358373351395, + 0.0012419109698385, + 0.0015404635341838002, + 0.0012250426225364208, + 0.0011874145129695535, + 0.0008138530538417399, + 0.001236371579580009, + 0.0010699753183871508, + 0.0016996676567941904, + 0.0008572678198106587, + 0.0005559396813623607, + 0.0008839085348881781, + 0.001356682856567204, + 0.0007296045077964664, + 0.0010477094911038876, + 0.0029240071307867765, + 0.0018292319728061557, + 0.002881876192986965, + 0.0022081511560827494, + 0.0017308368114754558, + 0.0023108364548534155, + 0.0019331688527017832, + 0.005001216195523739, + 0.002036679768934846, + 0.0031513168942183256, + 0.0033285552635788918, + 0.002285130089148879, + 0.0017775875749066472, + 0.0016242626588791609, + 0.0023604172747582197, + 0.0027098001446574926, + 0.001814267598092556, + 0.0019034279976040125, + 0.001390293356962502, + 0.0022011231631040573, + 0.0014617235865443945, + 0.0024520577862858772, + 0.001509932684712112, + 0.0025084002409130335, + 0.037605199962854385, + 0.022242702543735504, + 0.030378984287381172, + 0.021078694611787796, + 0.02824682928621769, + 0.025678375735878944, + 0.32502853870391846, + 0.32531994581222534, + 0.004587855190038681, + 0.003399874782189727, + 0.004177242983132601, + 0.0034553962759673595, + 0.003408324671909213, + 0.0028435303829610348, + 0.004281279630959034, + 0.002933352719992399, + 0.0018232589354738593, + 0.002199346898123622, + 0.001623363234102726, + 0.002038446255028248, + 0.002555443672463298, + 0.002738421782851219, + 0.0032768140081316233, + 0.002427600556984544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.02395787462592125 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast snowy plain bathed in morning sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fox's fur a flaming orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fox leap vertically into the air and then plunge headfirst into the deep snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a shower of shimmering, sun-lit snow powder visible exploding in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video filmed in cinematic slow motion with a low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018211066722869873 + ], + "opposite_logprob_tokens": [ + -4.0182108879089355 + ], + "logprob_reward": -0.018211066722869873, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018211066722869873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07150646299123764, + 0.025669192895293236, + 0.01635657623410225, + 0.008956813253462315, + 0.056188248097896576, + 0.028675012290477753, + 0.025962473824620247, + 0.7259994149208069, + 0.025938035920262337, + 0.014747711829841137 + ], + "credit_map": [ + 0.008229332976043224, + 0.014346614480018616, + 0.010635924525558949, + 0.008149789646267891, + 0.01018071360886097, + 0.007193786557763815, + 0.0069299424067139626, + 0.005840359255671501, + 0.003308121347799897, + 0.003996408544480801, + 0.003657749155536294, + 0.0027536433190107346, + 0.0041348617523908615, + 0.003482848173007369, + 0.002725777914747596, + 0.0016097831539809704, + 0.00189785601105541, + 0.0016905356897041202, + 0.0018767376895993948, + 0.0018570143729448318, + 0.002255417639389634, + 0.002547839656472206, + 0.0015493594110012054, + 0.002681816229596734, + 0.001338756293989718, + 0.0012304696720093489, + 0.0009163099457509816, + 0.0008774297311902046, + 0.0011405154364183545, + 0.0013167334254831076, + 0.0007414321298711002, + 0.0013951660366728902, + 0.007771421689540148, + 0.0057825446128845215, + 0.007504538632929325, + 0.005835199728608131, + 0.0046363649889826775, + 0.006342776119709015, + 0.005984609480947256, + 0.012330794706940651, + 0.0026026912964880466, + 0.0027312503661960363, + 0.005720154847949743, + 0.006258710753172636, + 0.0021182673517614603, + 0.002215995453298092, + 0.002628835616633296, + 0.004399106837809086, + 0.004829671233892441, + 0.0033274858724325895, + 0.0022446115035563707, + 0.003090408630669117, + 0.002822056645527482, + 0.003929648548364639, + 0.002128040185198188, + 0.0035905500408262014, + 0.033142898231744766, + 0.019783565774559975, + 0.02789713628590107, + 0.019704975187778473, + 0.02506636269390583, + 0.02281423658132553, + 0.28865495324134827, + 0.2889353036880493, + 0.003553401678800583, + 0.0033041739370673895, + 0.0032569686882197857, + 0.0027800563257187605, + 0.0027405417058616877, + 0.003590010805055499, + 0.004202471114695072, + 0.0025104095693677664, + 0.0017112810164690018, + 0.0021134226117283106, + 0.0012881128350272775, + 0.00146689813118428, + 0.002060794737190008, + 0.0026861857622861862, + 0.001939638750627637, + 0.001481377985328436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.018211066722869873 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast snowy plain bathed in morning sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fox's fur a flaming orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fox leap vertically into the air and then plunge headfirst into the deep snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a shower of shimmering, sun-lit snow powder visible exploding in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video filmed in cinematic slow motion with a low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012539356015622616 + ], + "opposite_logprob_tokens": [ + -4.387539386749268 + ], + "logprob_reward": -0.012539356015622616, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539356015622616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01630312390625477, + 0.01023267675191164, + 0.03335699811577797, + 0.0061654625460505486, + 0.013137184083461761, + 0.009205745533108711, + 0.02052673138678074, + 0.8019928932189941, + 0.06733867526054382, + 0.02174052968621254 + ], + "credit_map": [ + 0.002307275542989373, + 0.0020687421783804893, + 0.0021180722396820784, + 0.0014513592468574643, + 0.002064680214971304, + 0.0021799555979669094, + 0.001940144575200975, + 0.002172894077375531, + 0.001315372297540307, + 0.0014557184185832739, + 0.001211161375977099, + 0.000930230540689081, + 0.0013763942988589406, + 0.0012292966712266207, + 0.0010790291707962751, + 0.0016354736872017384, + 0.0037762450519949198, + 0.0035906117409467697, + 0.0049794395454227924, + 0.006942876148968935, + 0.003869404084980488, + 0.004258790984749794, + 0.0031443368643522263, + 0.0027952929958701134, + 0.0007839397876523435, + 0.0013616925571113825, + 0.0006093308911658823, + 0.0008115808595903218, + 0.0008279986213892698, + 0.0007046317914500833, + 0.000479776383144781, + 0.0005865119164809585, + 0.0012613178696483374, + 0.0013707333710044622, + 0.0014609586214646697, + 0.0014199704164639115, + 0.0018763714469969273, + 0.0013808466028422117, + 0.0011689849197864532, + 0.003198000369593501, + 0.0008693999261595309, + 0.0014033124316483736, + 0.0016184544656425714, + 0.0013499011984094977, + 0.0008390758885070682, + 0.0009195605525746942, + 0.000986858969554305, + 0.0012191819259896874, + 0.0043294052593410015, + 0.0017163287848234177, + 0.0021208375692367554, + 0.0025247063022106886, + 0.0015430654166266322, + 0.0034668066073209047, + 0.001250553410500288, + 0.003575026523321867, + 0.03702935576438904, + 0.021978428587317467, + 0.030544064939022064, + 0.019404996186494827, + 0.028093619272112846, + 0.025282083079218864, + 0.3193494379520416, + 0.32031092047691345, + 0.010488377884030342, + 0.00756532559171319, + 0.00888775009661913, + 0.007285771891474724, + 0.008621550165116787, + 0.008358900435268879, + 0.007932152599096298, + 0.008198847994208336, + 0.003607783932238817, + 0.0030878612305969, + 0.002244209870696068, + 0.002695918083190918, + 0.0025138112250715494, + 0.002936973236501217, + 0.002290718723088503, + 0.00236325291916728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 2 + ], + "selected_reward": -0.012539356015622616 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast snowy plain bathed in morning sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fox's fur a flaming orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fox leap vertically into the air and then plunge headfirst into the deep snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a shower of shimmering, sun-lit snow powder visible exploding in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video filmed in cinematic slow motion with a low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028212766628712416 + ], + "opposite_logprob_tokens": [ + -5.877821445465088 + ], + "logprob_reward": -0.0028212766628712416, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028212766628712416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01464482769370079, + 0.005467719864100218, + 0.014978942461311817, + 0.0016927599208429456, + 0.003337511094287038, + 0.0063036647625267506, + 0.004463612101972103, + 0.9167129993438721, + 0.02471989020705223, + 0.007678051013499498 + ], + "credit_map": [ + 0.0018948112847283483, + 0.0020800933707505465, + 0.001709572272375226, + 0.0017471745377406478, + 0.002197983441874385, + 0.0016312338411808014, + 0.0015757032670080662, + 0.0018082555616274476, + 0.0007421507034450769, + 0.0006577803869731724, + 0.0005598831921815872, + 0.000436952366726473, + 0.0006553141283802688, + 0.0007990160374902189, + 0.0005277670570649207, + 0.0010888563701882958, + 0.0015141242183744907, + 0.0014524245634675026, + 0.00194775708951056, + 0.0021816049702465534, + 0.002183233853429556, + 0.0016134264878928661, + 0.002193893538787961, + 0.0018924787873402238, + 0.0002352688170503825, + 0.00031616369960829616, + 0.00019245872681494802, + 0.0001717746927170083, + 0.00019185447308700532, + 0.000236242325627245, + 0.00016377991414628923, + 0.00018521721358411014, + 0.0005500456900335848, + 0.0003782652784138918, + 0.0004301840963307768, + 0.000558869622182101, + 0.0002938869583886117, + 0.0004050923162139952, + 0.00024444182054139674, + 0.0004767253703903407, + 0.0005486306035891175, + 0.0012037886772304773, + 0.000923854298889637, + 0.0007446615491062403, + 0.0006650873692706227, + 0.0005341641954146326, + 0.0008151591173373163, + 0.0008683187770657241, + 0.0007568587898276746, + 0.0003586279635783285, + 0.0004225410521030426, + 0.0007118824287317693, + 0.0003913000109605491, + 0.0007560431258752942, + 0.00030207171221263707, + 0.0007642870768904686, + 0.04168235883116722, + 0.02490336075425148, + 0.034373342990875244, + 0.023060159757733345, + 0.03238620609045029, + 0.029057413339614868, + 0.3659183084964752, + 0.36533185839653015, + 0.0036475793458521366, + 0.002604811917990446, + 0.003291084198281169, + 0.0026013217866420746, + 0.0024755685590207577, + 0.003283984027802944, + 0.0034076122101396322, + 0.0034079269971698523, + 0.0010116795310750604, + 0.0009231818839907646, + 0.0007216440862976015, + 0.0009131656843237579, + 0.0008357536280527711, + 0.0009989985264837742, + 0.0014313549036160111, + 0.0008422730024904013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 2 + ], + "selected_reward": -0.0028212766628712416 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast snowy plain bathed in morning sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fox's fur a flaming orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fox leap vertically into the air and then plunge headfirst into the deep snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a shower of shimmering, sun-lit snow powder visible exploding in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video filmed in cinematic slow motion with a low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638158440589905 + ], + "opposite_logprob_tokens": [ + -4.758638381958008 + ], + "logprob_reward": -0.008638158440589905, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638158440589905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05593924969434738, + 0.009928199462592602, + 0.030812852084636688, + 0.013589538633823395, + 0.014934340491890907, + 0.04360150545835495, + 0.025180751457810402, + 0.7303897142410278, + 0.04917291924357414, + 0.02645091898739338 + ], + "credit_map": [ + 0.009425238706171513, + 0.007359136827290058, + 0.006701061502099037, + 0.005184265784919262, + 0.005920760333538055, + 0.007099139038473368, + 0.005679585970938206, + 0.008570061065256596, + 0.0015364319551736116, + 0.0012715101474896073, + 0.0012362913694232702, + 0.0008863262482918799, + 0.0014025731943547726, + 0.001343913609161973, + 0.0010117648635059595, + 0.0012393882498145103, + 0.0032638595439493656, + 0.002726148348301649, + 0.003553625661879778, + 0.004458312410861254, + 0.004147734493017197, + 0.0036378009244799614, + 0.0034656175412237644, + 0.0055597517639398575, + 0.001587111852131784, + 0.0023037188220769167, + 0.0015687289414927363, + 0.001983746187761426, + 0.0015217039035633206, + 0.0013363936450332403, + 0.0014071882469579577, + 0.0018809465691447258, + 0.0016972889425233006, + 0.0016267083119601011, + 0.0016143982065841556, + 0.0019649399910122156, + 0.002326953923329711, + 0.0013316392432898283, + 0.0012029183562844992, + 0.0031694932840764523, + 0.0026044396217912436, + 0.002216116990894079, + 0.012878847308456898, + 0.014317498542368412, + 0.0019622426480054855, + 0.002021292457357049, + 0.003021780401468277, + 0.004579287022352219, + 0.00288572465069592, + 0.001814885064959526, + 0.002346492139622569, + 0.004188976716250181, + 0.00243663159199059, + 0.004082628525793552, + 0.002152042230591178, + 0.005273368209600449, + 0.03391105681657791, + 0.019777406007051468, + 0.026297319680452347, + 0.01789517141878605, + 0.02595217153429985, + 0.02299731783568859, + 0.29137811064720154, + 0.2921811640262604, + 0.006644886918365955, + 0.005160012748092413, + 0.008242675103247166, + 0.008101997897028923, + 0.0056216963566839695, + 0.004590905737131834, + 0.00651001650840044, + 0.004300720989704132, + 0.0029813898727297783, + 0.003993846941739321, + 0.0024384104181081057, + 0.0025452440604567528, + 0.003778600599616766, + 0.00335884396918118, + 0.003734610741958022, + 0.00361996958963573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.008638158440589905 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium shot, filmed in cinematic slow motion, the camera steadily following the subject, captures a vigilant red fox, its fur a flaming orange-red, nestled amidst a vast snowy plain bathed in the first golden rays of morning sunlight. The low, slanting sunlight casts long, blue shadows across the pristine white of the snow, while the fox's body, taut and graceful, arches in a powerful arc. Its pointed muzzle nearly touches the snow, its dark eyes fixed on the unseen movements beneath, its muscles poised for action. Suddenly, with a sudden thrust of its hind legs, its entire body leaps vertically into the air, its front paws clasped together and pointed downward like an arrow. With unparalleled swiftness, it plunges headfirst into the deep snow. As it plunges into the drift, a shower of shimmering, sun-lit snow powder explodes in all directions, clearly visible in slow motion. Only the tip of its furry tail remains briefly above the surface. The whole picture, with its sharp contrast of warm golden light and cold blue shadows, presents a primitive, focused and extremely beautiful life tension, full of strong cinematic texture.", + "seed": 11909, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_13/index_8876_seed_11909_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01459219241514802, + "vlm_logprob": -0.01459219241514802, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.01459219241514802, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast snowy plain bathed in morning sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fox's fur a flaming orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fox leap vertically into the air and then plunge headfirst into the deep snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a shower of shimmering, sun-lit snow powder visible exploding in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video filmed in cinematic slow motion with a low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023979637771844864 + ], + "opposite_logprob_tokens": [ + -4.273979663848877 + ], + "logprob_reward": -0.023979637771844864, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.023979637771844864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13529904186725616, + 0.06581611186265945, + 0.04197059944272041, + 0.054410871118307114, + 0.06259181350469589, + 0.07102639973163605, + 0.07423549145460129, + 0.20096205174922943, + 0.17925181984901428, + 0.11443576216697693 + ], + "credit_map": [ + 0.021363122388720512, + 0.016592752188444138, + 0.015308979898691177, + 0.012702996842563152, + 0.010277199558913708, + 0.019583748653531075, + 0.013875490054488182, + 0.02559475600719452, + 0.010934755206108093, + 0.00722061051055789, + 0.005184616893529892, + 0.006183866411447525, + 0.008097253739833832, + 0.009857348166406155, + 0.006220967508852482, + 0.012116698548197746, + 0.007443015463650227, + 0.0038582230918109417, + 0.0037556630559265614, + 0.0035248752683401108, + 0.0034296081867069006, + 0.007512263022363186, + 0.003479493549093604, + 0.008967459201812744, + 0.010314739309251308, + 0.005518103484064341, + 0.0051260157488286495, + 0.0045819818042218685, + 0.006149782333523035, + 0.00913422740995884, + 0.007260005455464125, + 0.006326016969978809, + 0.009169219993054867, + 0.007212724070996046, + 0.006743430159986019, + 0.006542699411511421, + 0.005978464148938656, + 0.008860996924340725, + 0.006307309027761221, + 0.011776967905461788, + 0.011424404568970203, + 0.009658822789788246, + 0.005875043570995331, + 0.005714192520827055, + 0.013741075992584229, + 0.00968242809176445, + 0.006266659591346979, + 0.008663774468004704, + 0.008586933836340904, + 0.01092527061700821, + 0.0071936403401196, + 0.007798794656991959, + 0.008226204663515091, + 0.01103729847818613, + 0.00941561907529831, + 0.011051736772060394, + 0.01583339273929596, + 0.019776077941060066, + 0.015513923950493336, + 0.020466426387429237, + 0.015344580635428429, + 0.03376184031367302, + 0.05229680985212326, + 0.027968978509306908, + 0.01263006404042244, + 0.017571473494172096, + 0.039392389357089996, + 0.03683418035507202, + 0.012157151475548744, + 0.020532390102744102, + 0.017168549820780754, + 0.022965624928474426, + 0.017925284802913666, + 0.011830548755824566, + 0.024049248546361923, + 0.013858080841600895, + 0.009191688150167465, + 0.013200837187469006, + 0.013553920201957226, + 0.01082614716142416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.023979637771844864 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast snowy plain bathed in morning sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fox's fur a flaming orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fox leap vertically into the air and then plunge headfirst into the deep snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a shower of shimmering, sun-lit snow powder visible exploding in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video filmed in cinematic slow motion with a low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00866664107888937 + ], + "opposite_logprob_tokens": [ + -4.758666515350342 + ], + "logprob_reward": -0.00866664107888937, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00866664107888937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09203231334686279, + 0.11010231077671051, + 0.13179351389408112, + 0.036796752363443375, + 0.033780694007873535, + 0.039461683481931686, + 0.3109271228313446, + 0.14280301332473755, + 0.08537004888057709, + 0.01693250983953476 + ], + "credit_map": [ + 0.012163025327026844, + 0.01765877939760685, + 0.010338354855775833, + 0.008631803095340729, + 0.01346959825605154, + 0.012230689637362957, + 0.008882664144039154, + 0.008657405152916908, + 0.01378865446895361, + 0.01902584545314312, + 0.010039941407740116, + 0.017137372866272926, + 0.011305452324450016, + 0.014766496606171131, + 0.011438603512942791, + 0.012599951587617397, + 0.018398741260170937, + 0.017966415733098984, + 0.014761706814169884, + 0.014650248922407627, + 0.015218089334666729, + 0.022570893168449402, + 0.012542076408863068, + 0.01568535529077053, + 0.004641151055693626, + 0.006281754467636347, + 0.005718447733670473, + 0.003525037318468094, + 0.00433563394472003, + 0.004896205849945545, + 0.0037092575803399086, + 0.003689267672598362, + 0.004104175139218569, + 0.0038109298329800367, + 0.004743664059787989, + 0.002906165784224868, + 0.004652814008295536, + 0.004903379362076521, + 0.0036564033944159746, + 0.00500316359102726, + 0.0045915828086435795, + 0.005995844025164843, + 0.0036768282297998667, + 0.00380715006031096, + 0.009710014797747135, + 0.005489143542945385, + 0.0036423976998776197, + 0.0025487274397164583, + 0.036203570663928986, + 0.05694561079144478, + 0.060529351234436035, + 0.04106413945555687, + 0.023986607789993286, + 0.04771003499627113, + 0.02366650477051735, + 0.020821314305067062, + 0.011689883656799793, + 0.016223251819610596, + 0.01279927883297205, + 0.0183212012052536, + 0.011019617319107056, + 0.020275186747312546, + 0.031019818037748337, + 0.021454770117998123, + 0.006573375780135393, + 0.006100380327552557, + 0.011619091965258121, + 0.015218442305922508, + 0.005906454287469387, + 0.009563914500176907, + 0.008668002672493458, + 0.021720387041568756, + 0.0027994599658995867, + 0.002582944231107831, + 0.0026646219193935394, + 0.0018345838179811835, + 0.0014969633193686604, + 0.002177509944885969, + 0.002094884403049946, + 0.001281543867662549 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 2 + ], + "selected_reward": -0.00866664107888937 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast snowy plain bathed in morning sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fox's fur a flaming orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fox leap vertically into the air and then plunge headfirst into the deep snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a shower of shimmering, sun-lit snow powder visible exploding in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video filmed in cinematic slow motion with a low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023289937525987625 + ], + "opposite_logprob_tokens": [ + -3.773289918899536 + ], + "logprob_reward": -0.023289937525987625, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023289937525987625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11847615242004395, + 0.049146488308906555, + 0.025360852479934692, + 0.04975985735654831, + 0.04103652015328407, + 0.08713039755821228, + 0.11714616417884827, + 0.20023031532764435, + 0.1872004270553589, + 0.12451282143592834 + ], + "credit_map": [ + 0.01744070090353489, + 0.022391727194190025, + 0.0112818144261837, + 0.014850460924208164, + 0.01597895845770836, + 0.01519290916621685, + 0.013155690394341946, + 0.008183898404240608, + 0.007714864797890186, + 0.00609353743493557, + 0.005249083507806063, + 0.00568935414776206, + 0.007200835272669792, + 0.007394908927381039, + 0.0048266504891216755, + 0.004977253265678883, + 0.003139589447528124, + 0.003763343906030059, + 0.0026490497402846813, + 0.00340683339163661, + 0.002855726284906268, + 0.0033953271340578794, + 0.0033008172176778316, + 0.002850165357813239, + 0.008962173946201801, + 0.004484947770833969, + 0.004668472800403833, + 0.004893073346465826, + 0.004059350583702326, + 0.008185946382582188, + 0.005546439439058304, + 0.008959454484283924, + 0.0061011239886283875, + 0.004972645081579685, + 0.003643294330686331, + 0.00478474423289299, + 0.004605734720826149, + 0.005984097719192505, + 0.0036849738098680973, + 0.0072599067352712154, + 0.011860431171953678, + 0.009958544746041298, + 0.008138519711792469, + 0.008603540249168873, + 0.020243830978870392, + 0.011910339817404747, + 0.008933983743190765, + 0.007481209933757782, + 0.011584200896322727, + 0.015030805952847004, + 0.017997613176703453, + 0.01172693818807602, + 0.01806008815765381, + 0.017706315964460373, + 0.014523549005389214, + 0.01051664911210537, + 0.016785413026809692, + 0.021596260368824005, + 0.014947299845516682, + 0.021130681037902832, + 0.024733368307352066, + 0.03290213644504547, + 0.042513083666563034, + 0.025622067973017693, + 0.014741378836333752, + 0.01646444946527481, + 0.034138329327106476, + 0.033070195466279984, + 0.018512768670916557, + 0.018007265403866768, + 0.01878112368285656, + 0.03348490968346596, + 0.015462418086826801, + 0.02757798135280609, + 0.013332786038517952, + 0.011515429243445396, + 0.01100395992398262, + 0.010853067971765995, + 0.02300809696316719, + 0.011759085580706596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.023289937525987625 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast snowy plain bathed in morning sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fox's fur a flaming orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fox leap vertically into the air and then plunge headfirst into the deep snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a shower of shimmering, sun-lit snow powder visible exploding in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video filmed in cinematic slow motion with a low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950829479843378 + ], + "opposite_logprob_tokens": [ + -5.130950927734375 + ], + "logprob_reward": -0.005950829479843378, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950829479843378, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09965994954109192, + 0.08115366101264954, + 0.047294117510318756, + 0.03332694619894028, + 0.043618425726890564, + 0.0953090712428093, + 0.10756418108940125, + 0.23744073510169983, + 0.190048947930336, + 0.06458398699760437 + ], + "credit_map": [ + 0.014675810001790524, + 0.014876860193908215, + 0.011324906721711159, + 0.008883892558515072, + 0.010180885903537273, + 0.013779821805655956, + 0.011614239774644375, + 0.014323522336781025, + 0.013556614518165588, + 0.01085406169295311, + 0.008385195396840572, + 0.009680335409939289, + 0.009271454997360706, + 0.011335044167935848, + 0.009093808941543102, + 0.00897714588791132, + 0.00596408499404788, + 0.005775719881057739, + 0.006519155576825142, + 0.005618395283818245, + 0.0050430260598659515, + 0.006306570954620838, + 0.00588789489120245, + 0.0061792670749127865, + 0.004767707549035549, + 0.0035412381403148174, + 0.005402268376201391, + 0.0034367286134511232, + 0.002847633557394147, + 0.005252766888588667, + 0.003543649800121784, + 0.004534950479865074, + 0.006194815970957279, + 0.005898843519389629, + 0.005487009882926941, + 0.005202043801546097, + 0.004939877428114414, + 0.006424174644052982, + 0.004449543543159962, + 0.005022112280130386, + 0.013617856428027153, + 0.00928523764014244, + 0.007246844004839659, + 0.009213645942509174, + 0.024585803970694542, + 0.013532035052776337, + 0.00956910103559494, + 0.008258546702563763, + 0.012350003235042095, + 0.01706273853778839, + 0.014755582436919212, + 0.010710734874010086, + 0.012494893744587898, + 0.016272688284516335, + 0.013045810163021088, + 0.010871714912354946, + 0.016275355592370033, + 0.015933403745293617, + 0.01958838477730751, + 0.02526130899786949, + 0.016591649502515793, + 0.04213936999440193, + 0.06722386926412582, + 0.03442738950252533, + 0.024209385737776756, + 0.02265195921063423, + 0.032843105494976044, + 0.024032043293118477, + 0.014615833759307861, + 0.023736413568258286, + 0.015292626805603504, + 0.032667577266693115, + 0.010284310206770897, + 0.009572413749992847, + 0.009377975948154926, + 0.007073692511767149, + 0.0052099935710430145, + 0.008721999824047089, + 0.007750148419290781, + 0.0065934499725699425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.005950829479843378 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast snowy plain bathed in morning sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fox's fur a flaming orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fox leap vertically into the air and then plunge headfirst into the deep snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a shower of shimmering, sun-lit snow powder visible exploding in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video filmed in cinematic slow motion with a low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011073916219174862 + ], + "opposite_logprob_tokens": [ + -4.511074066162109 + ], + "logprob_reward": -0.011073916219174862, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011073916219174862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09362364560365677, + 0.05730476230382919, + 0.04381934925913811, + 0.03830086812376976, + 0.046450577676296234, + 0.11541001498699188, + 0.13404583930969238, + 0.1788974553346634, + 0.17405876517295837, + 0.1180887222290039 + ], + "credit_map": [ + 0.014450343325734138, + 0.014880777336657047, + 0.009423479437828064, + 0.007069480139762163, + 0.009896671399474144, + 0.013946992345154285, + 0.009728782810270786, + 0.0142271239310503, + 0.008623840287327766, + 0.008446445688605309, + 0.007248320616781712, + 0.009111878462135792, + 0.004405689425766468, + 0.007543511688709259, + 0.006531374994665384, + 0.005393702536821365, + 0.005361965857446194, + 0.005924834869801998, + 0.005362492986023426, + 0.004381494130939245, + 0.003413186175748706, + 0.007142008747905493, + 0.006202473770827055, + 0.006030892953276634, + 0.006818372290581465, + 0.004518318455666304, + 0.003156686667352915, + 0.003120489651337266, + 0.0038855927996337414, + 0.0059433020651340485, + 0.003692758036777377, + 0.007165347225964069, + 0.006637017708271742, + 0.005974085535854101, + 0.004030060023069382, + 0.005234047304838896, + 0.006906931288540363, + 0.006921795196831226, + 0.004953252151608467, + 0.005793384741991758, + 0.01797357201576233, + 0.01272597350180149, + 0.01317665446549654, + 0.013354522176086903, + 0.01723961904644966, + 0.016494370996952057, + 0.013555346988141537, + 0.010889969766139984, + 0.01577584072947502, + 0.018112028017640114, + 0.013844017870724201, + 0.011697616428136826, + 0.01978345960378647, + 0.018765846267342567, + 0.016314920037984848, + 0.019752100110054016, + 0.011499568819999695, + 0.010927996598184109, + 0.009808137081563473, + 0.02626064419746399, + 0.01393189374357462, + 0.028266696259379387, + 0.04729130491614342, + 0.030911212787032127, + 0.02524634078145027, + 0.018114756792783737, + 0.02841128222644329, + 0.023430271074175835, + 0.01827566884458065, + 0.02222413197159767, + 0.020199686288833618, + 0.018156623467803, + 0.019076578319072723, + 0.013969250954687595, + 0.021933663636446, + 0.012683337554335594, + 0.010999423451721668, + 0.01126202940940857, + 0.01785336621105671, + 0.010311071760952473 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.011073916219174862 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium shot, filmed in cinematic slow motion, the camera steadily following the subject, captures a vigilant red fox, its fur a flaming orange-red, nestled amidst a vast snowy plain bathed in the first golden rays of morning sunlight. The low, slanting sunlight casts long, blue shadows across the pristine white of the snow, while the fox's body, taut and graceful, arches in a powerful arc. Its pointed muzzle nearly touches the snow, its dark eyes fixed on the unseen movements beneath, its muscles poised for action. Suddenly, with a sudden thrust of its hind legs, its entire body leaps vertically into the air, its front paws clasped together and pointed downward like an arrow. With unparalleled swiftness, it plunges headfirst into the deep snow. As it plunges into the drift, a shower of shimmering, sun-lit snow powder explodes in all directions, clearly visible in slow motion. Only the tip of its furry tail remains briefly above the surface. The whole picture, with its sharp contrast of warm golden light and cold blue shadows, presents a primitive, focused and extremely beautiful life tension, full of strong cinematic texture.", + "seed": 11910, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_13/index_8876_seed_11910_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.007817065389826895, + "vlm_logprob": -0.007817065389826895, + "vlm_yes_no_margin": 5.175, + "vlm_token_credit": -0.007817065389826895, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast snowy plain bathed in morning sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fox's fur a flaming orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fox leap vertically into the air and then plunge headfirst into the deep snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a shower of shimmering, sun-lit snow powder visible exploding in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video filmed in cinematic slow motion with a low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006927401293069124 + ], + "opposite_logprob_tokens": [ + -5.881927490234375 + ], + "logprob_reward": -0.006927401293069124, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.006927401293069124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11913441866636276, + 0.03116687946021557, + 0.026550015434622765, + 0.06799960881471634, + 0.45751717686653137, + 0.031977877020835876, + 0.09030851721763611, + 0.042101338505744934, + 0.0691928118467331, + 0.06405135244131088 + ], + "credit_map": [ + 0.028170570731163025, + 0.013853257521986961, + 0.01967366226017475, + 0.013392039574682713, + 0.010062512010335922, + 0.011732242070138454, + 0.007520147133618593, + 0.01472997386008501, + 0.005071114748716354, + 0.003604116616770625, + 0.004662860184907913, + 0.0058332993648946285, + 0.0027404120191931725, + 0.0038113633636385202, + 0.002440771786496043, + 0.0030029406771063805, + 0.005447986535727978, + 0.0031893861014395952, + 0.003446368733420968, + 0.0022948000114411116, + 0.0023288412485271692, + 0.004234324675053358, + 0.002841966925188899, + 0.002766337711364031, + 0.0038903849199414253, + 0.0034961847122758627, + 0.004847817122936249, + 0.023326300084590912, + 0.0023002512753009796, + 0.0029658109415322542, + 0.023115718737244606, + 0.00405713589861989, + 0.010784700512886047, + 0.017724361270666122, + 0.022267675027251244, + 0.01235829945653677, + 0.02048361487686634, + 0.019177785143256187, + 0.19802458584308624, + 0.1566961258649826, + 0.006133499555289745, + 0.0041236188262701035, + 0.003597748000174761, + 0.0019810080993920565, + 0.0029518073424696922, + 0.004298632964491844, + 0.0037194506730884314, + 0.005172110628336668, + 0.0169538501650095, + 0.007829800248146057, + 0.014967609196901321, + 0.007113081868737936, + 0.008011141791939735, + 0.009442293085157871, + 0.01739080436527729, + 0.008599933236837387, + 0.0066530839540064335, + 0.004482999909669161, + 0.006377259269356728, + 0.0056280591525137424, + 0.004134329967200756, + 0.003529548179358244, + 0.006185013800859451, + 0.005111043341457844, + 0.004399846773594618, + 0.004326948430389166, + 0.008778375573456287, + 0.008832231163978577, + 0.006624427158385515, + 0.00867553986608982, + 0.017926014959812164, + 0.009629422798752785, + 0.005187441129237413, + 0.005091232247650623, + 0.004799233749508858, + 0.011502103880047798, + 0.008032284677028656, + 0.010305243544280529, + 0.009977553971111774, + 0.00915625598281622 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.006927401293069124 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast snowy plain bathed in morning sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fox's fur a flaming orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fox leap vertically into the air and then plunge headfirst into the deep snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a shower of shimmering, sun-lit snow powder visible exploding in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video filmed in cinematic slow motion with a low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979746226221323 + ], + "opposite_logprob_tokens": [ + -4.6347975730896 + ], + "logprob_reward": -0.00979746226221323, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979746226221323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11933998018503189, + 0.1783267855644226, + 0.06730308383703232, + 0.10408026725053787, + 0.2926076650619507, + 0.01632181741297245, + 0.07117949426174164, + 0.038607075810432434, + 0.06253818422555923, + 0.04969557374715805 + ], + "credit_map": [ + 0.014404784888029099, + 0.012733510695397854, + 0.014880706556141376, + 0.019190501421689987, + 0.011444119736552238, + 0.010494361631572247, + 0.020507046952843666, + 0.015684960409998894, + 0.017001628875732422, + 0.02363641932606697, + 0.032923176884651184, + 0.045184578746557236, + 0.014898829162120819, + 0.013158773072063923, + 0.014031360857188702, + 0.0174920242279768, + 0.005968264304101467, + 0.010308586992323399, + 0.008069491013884544, + 0.008270805701613426, + 0.006869518663734198, + 0.006504385732114315, + 0.011511385440826416, + 0.009800651110708714, + 0.0060008191503584385, + 0.003501582657918334, + 0.005508764181286097, + 0.04095529764890671, + 0.003172561526298523, + 0.0026389285922050476, + 0.03807508945465088, + 0.004227226134389639, + 0.013697817921638489, + 0.02298598550260067, + 0.016027754172682762, + 0.014427716843783855, + 0.0229061059653759, + 0.0178842693567276, + 0.10495111346244812, + 0.07972690463066101, + 0.0016630887985229492, + 0.0025261009577661753, + 0.0016317086992785335, + 0.0010919992346316576, + 0.00228431960567832, + 0.002368596848100424, + 0.0022671297192573547, + 0.002488873666152358, + 0.0059114908799529076, + 0.008185028098523617, + 0.01038312166929245, + 0.008946401998400688, + 0.009738070890307426, + 0.009312532842159271, + 0.011584809981286526, + 0.007118046749383211, + 0.0026836327742785215, + 0.002669059904292226, + 0.0077295503579080105, + 0.009547499939799309, + 0.0037556185852736235, + 0.0035727533977478743, + 0.004242038354277611, + 0.0044069234281778336, + 0.003547725733369589, + 0.004067293368279934, + 0.008326413109898567, + 0.009920968674123287, + 0.004762199707329273, + 0.0065795863047242165, + 0.013330320827662945, + 0.012003684416413307, + 0.004203436896204948, + 0.003208596957847476, + 0.002938066842034459, + 0.014143936336040497, + 0.006059613544493914, + 0.006444168742746115, + 0.007697148714214563, + 0.00500060897320509 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.00979746226221323 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast snowy plain bathed in morning sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fox's fur a flaming orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fox leap vertically into the air and then plunge headfirst into the deep snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a shower of shimmering, sun-lit snow powder visible exploding in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video filmed in cinematic slow motion with a low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012536648660898209 + ], + "opposite_logprob_tokens": [ + -4.387536525726318 + ], + "logprob_reward": -0.012536648660898209, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012536648660898209, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06385181844234467, + 0.039308469742536545, + 0.024491816759109497, + 0.20631834864616394, + 0.17619386315345764, + 0.03119799681007862, + 0.03125843405723572, + 0.18876366317272186, + 0.18955481052398682, + 0.04906078800559044 + ], + "credit_map": [ + 0.008865115232765675, + 0.008358195424079895, + 0.009868867695331573, + 0.005448533687740564, + 0.004514421336352825, + 0.007058612070977688, + 0.011577167548239231, + 0.00816090777516365, + 0.0060019781813025475, + 0.005254266317933798, + 0.007421435788273811, + 0.010660378262400627, + 0.0026685206685215235, + 0.0025548781268298626, + 0.002507799072191119, + 0.0022392140235751867, + 0.0038404129445552826, + 0.0043172165751457214, + 0.0027077484410256147, + 0.0026563620194792747, + 0.002285765251144767, + 0.0038441692013293505, + 0.002644661581143737, + 0.0021954793483018875, + 0.01018733624368906, + 0.008859401568770409, + 0.009694448672235012, + 0.07393491268157959, + 0.007085663266479969, + 0.005706889554858208, + 0.0798214003443718, + 0.011028283275663853, + 0.005085364915430546, + 0.006846276577562094, + 0.009079573675990105, + 0.004855843726545572, + 0.008466063067317009, + 0.007640043739229441, + 0.0759071558713913, + 0.058313533663749695, + 0.005150512792170048, + 0.004783982876688242, + 0.0037405823823064566, + 0.002040341030806303, + 0.002456081099808216, + 0.004168581683188677, + 0.003902634372934699, + 0.0049552819691598415, + 0.003597009228542447, + 0.0037279236130416393, + 0.004458763636648655, + 0.0030865143053233624, + 0.004569745156913996, + 0.004344477783888578, + 0.0052177696488797665, + 0.002256233710795641, + 0.010564152151346207, + 0.014033151790499687, + 0.03809617832303047, + 0.044982828199863434, + 0.020907793194055557, + 0.016491230577230453, + 0.026153400540351868, + 0.01753491908311844, + 0.0157513078302145, + 0.014093756675720215, + 0.02441980130970478, + 0.026530206203460693, + 0.013585787266492844, + 0.02237369678914547, + 0.03235935419797897, + 0.04044089838862419, + 0.00405309721827507, + 0.004025472328066826, + 0.00548725388944149, + 0.006269604433327913, + 0.0060414099134504795, + 0.006032790523022413, + 0.006294300314038992, + 0.010856861248612404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 7 + ], + "selected_reward": -0.012536648660898209 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast snowy plain bathed in morning sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fox's fur a flaming orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fox leap vertically into the air and then plunge headfirst into the deep snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a shower of shimmering, sun-lit snow powder visible exploding in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video filmed in cinematic slow motion with a low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021967582870274782 + ], + "opposite_logprob_tokens": [ + -6.127196788787842 + ], + "logprob_reward": -0.0021967582870274782, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021967582870274782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08080689609050751, + 0.053936783224344254, + 0.02020293101668358, + 0.149550661444664, + 0.08785022795200348, + 0.02542940340936184, + 0.03687875717878342, + 0.10688618570566177, + 0.3706788420677185, + 0.06777926534414291 + ], + "credit_map": [ + 0.011006890796124935, + 0.006705936044454575, + 0.010656646452844143, + 0.014159672893583775, + 0.0058241612277925014, + 0.006935700308531523, + 0.01618521846830845, + 0.00933267641812563, + 0.0083780437707901, + 0.006700558587908745, + 0.009463858790695667, + 0.012883402407169342, + 0.0037318244576454163, + 0.0034075267612934113, + 0.004667464643716812, + 0.004704108461737633, + 0.0028085275553166866, + 0.002274557948112488, + 0.0024092216044664383, + 0.0014977985993027687, + 0.0025394742842763662, + 0.005343362223356962, + 0.00177238998003304, + 0.0015576017322018743, + 0.012719705700874329, + 0.007442166563123465, + 0.008644078858196735, + 0.05466292425990105, + 0.0048874677158892155, + 0.006314872298389673, + 0.047365181148052216, + 0.007514274679124355, + 0.006456464994698763, + 0.01199133787304163, + 0.0037850176449865103, + 0.004179372917860746, + 0.007609655149281025, + 0.00447563361376524, + 0.026937438175082207, + 0.022415315732359886, + 0.003798159072175622, + 0.004283735994249582, + 0.003167543327435851, + 0.0016002606134861708, + 0.001997327199205756, + 0.002320494968444109, + 0.004206298850476742, + 0.004055584315210581, + 0.0065803430043160915, + 0.003766848938539624, + 0.005573136266320944, + 0.002528231590986252, + 0.004392316564917564, + 0.004454638808965683, + 0.005970948841422796, + 0.003612296190112829, + 0.012081000953912735, + 0.009863784536719322, + 0.015834283083677292, + 0.014656406827270985, + 0.01772504486143589, + 0.013371073640882969, + 0.011476841755211353, + 0.011877750046551228, + 0.029410162940621376, + 0.027666902169585228, + 0.04824027791619301, + 0.05669242888689041, + 0.021834952756762505, + 0.05377357080578804, + 0.056397195905447006, + 0.07666337490081787, + 0.004144035279750824, + 0.004585825838148594, + 0.0050249965861439705, + 0.013244317844510078, + 0.008942881599068642, + 0.007710651494562626, + 0.01315830834209919, + 0.010968257673084736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 7 + ], + "selected_reward": -0.0021967582870274782 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast snowy plain bathed in morning sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fox's fur a flaming orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fox leap vertically into the air and then plunge headfirst into the deep snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a shower of shimmering, sun-lit snow powder visible exploding in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video filmed in cinematic slow motion with a low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007627056445926428 + ], + "opposite_logprob_tokens": [ + -4.882627010345459 + ], + "logprob_reward": -0.007627056445926428, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007627056445926428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07890161126852036, + 0.028848936781287193, + 0.02132227271795273, + 0.15141132473945618, + 0.5203471183776855, + 0.010783865116536617, + 0.01966998167335987, + 0.044403593987226486, + 0.04672233760356903, + 0.07758902758359909 + ], + "credit_map": [ + 0.012203371152281761, + 0.009481810964643955, + 0.011571698822081089, + 0.011949511244893074, + 0.005159248132258654, + 0.006177268922328949, + 0.013400368392467499, + 0.008958335965871811, + 0.0043240259401500225, + 0.003716767765581608, + 0.005173211917281151, + 0.007465198170393705, + 0.0019570381846278906, + 0.0017022169195115566, + 0.0020457811187952757, + 0.0024646965321153402, + 0.002913827309384942, + 0.002334628254175186, + 0.002104753628373146, + 0.0016014018328860402, + 0.0027274962048977613, + 0.003607940161600709, + 0.0024720558430999517, + 0.003560170531272888, + 0.011607005260884762, + 0.006691599730402231, + 0.0083118611946702, + 0.05788334831595421, + 0.004340334329754114, + 0.00543601019307971, + 0.04944397136569023, + 0.0076971882954239845, + 0.012689540162682533, + 0.02093730866909027, + 0.02605305053293705, + 0.01441227551549673, + 0.023475848138332367, + 0.021188339218497276, + 0.22600892186164856, + 0.17558178305625916, + 0.0015175086446106434, + 0.0018059097928926349, + 0.001180920284241438, + 0.000969614542555064, + 0.0008818296482786536, + 0.0014391274889931083, + 0.0014869635924696922, + 0.0015019907150417566, + 0.002962093334645033, + 0.001960017951205373, + 0.0031437897123396397, + 0.0016957486514002085, + 0.0022763002198189497, + 0.0020031780004501343, + 0.0033700577914714813, + 0.0022587962448596954, + 0.0042092264629900455, + 0.004312485922127962, + 0.00858083926141262, + 0.011232776567339897, + 0.003319868817925453, + 0.0023248083889484406, + 0.005602112505584955, + 0.004821474198251963, + 0.0032564308494329453, + 0.0036397988442331553, + 0.006015109829604626, + 0.006021863780915737, + 0.004630169365555048, + 0.005254008341580629, + 0.011904234066605568, + 0.006000722758471966, + 0.010193235240876675, + 0.007901521399617195, + 0.009059248492121696, + 0.010681310668587685, + 0.006772697437554598, + 0.00983337964862585, + 0.00993497483432293, + 0.013212655670940876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.007627056445926428 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_14/index_tensor([8633, 8633])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_14/index_tensor([8633, 8633])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..333072c78acde0de8a2ea6cc721d156411890448 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_14/index_tensor([8633, 8633])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, static shot from a slightly elevated angle, focuses on a single, fully ripe fig, which has been cracked open to reveal its inner beauty while resting on a dark, rustic wooden surface. The fig's deep purple skin, with its delicate, powdery blossoms, gives way to a vibrant, velvety flesh within, rich ruby-red pulp and countless tiny, gleaming seeds. From the center of this intricate interior, a thick, shimmering drop of natural syrup slowly emerges. It swells into a perfect, heavy bead, catching a single, soft light source from the side, causing it to shimmer like liquid amber. The syrup begins to ooze out at a deliberate, almost imperceptible pace, its viscous form spreading over the delicate veins of the flesh, carving a viscous, shiny path downward. Dramatic black and white lighting highlights every minute detail—the shimmer of the syrup, the texture of the seeds, the gentle curves of the skin—while a rich palette of deep purple, red, and gold against the dark backdrop creates an atmosphere of profound sensuality and contemplative natural decadence.", + "seed": 1345, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_14/index_8633_seed_1345_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.012732432503253222, + "vlm_logprob": -0.012732432503253222, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.012732432503253222, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single fig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fig cracked open to reveal ruby-red pulp and tiny seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick drop of syrup emerging from the fig's interior and spreading downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fig resting on a dark, rustic wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01427133847028017 + ], + "opposite_logprob_tokens": [ + -4.389271259307861 + ], + "logprob_reward": -0.01427133847028017, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01427133847028017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17519566416740417, + 0.053958021104335785, + 0.5977593660354614, + 0.011474937200546265, + 0.05730826035141945, + 0.015753209590911865, + 0.005768241360783577, + 0.04698699712753296, + 0.023320907726883888, + 0.01247434876859188 + ], + "credit_map": [ + 0.012216305360198021, + 0.012678519822657108, + 0.05768212303519249, + 0.01577858068048954, + 0.011434249579906464, + 0.009519755840301514, + 0.00966253224760294, + 0.04622360318899155, + 0.004423780366778374, + 0.0035914608743041754, + 0.008477801457047462, + 0.004337838385254145, + 0.008472717367112637, + 0.004813202656805515, + 0.004257918335497379, + 0.015583299100399017, + 0.03451480343937874, + 0.02594826929271221, + 0.20452485978603363, + 0.04745764285326004, + 0.025232313200831413, + 0.0330178327858448, + 0.03045833110809326, + 0.1966053694486618, + 0.0011057130759581923, + 0.001110881450586021, + 0.0013212658232077956, + 0.0014445330016314983, + 0.0018515443662181497, + 0.0013108359416946769, + 0.0017413187306374311, + 0.001588843995705247, + 0.003644022159278393, + 0.006285754963755608, + 0.004608626943081617, + 0.012816853821277618, + 0.008993224240839481, + 0.004319357685744762, + 0.00929245911538601, + 0.007347960490733385, + 0.0016623623669147491, + 0.0015593183925375342, + 0.002017236314713955, + 0.001893229316920042, + 0.002380894962698221, + 0.001999130006879568, + 0.0023874195758253336, + 0.0018536185380071402, + 0.0005371564766392112, + 0.0007654330693185329, + 0.00048624389455653727, + 0.001001975266262889, + 0.000559682201128453, + 0.0007180574466474354, + 0.0009647472761571407, + 0.0007349452353082597, + 0.004002160858362913, + 0.0042528132908046246, + 0.003929320722818375, + 0.006231922190636396, + 0.009752354584634304, + 0.005771197844296694, + 0.006340260151773691, + 0.006706969812512398, + 0.0029819344636052847, + 0.002401437610387802, + 0.001900313887745142, + 0.0027773170731961727, + 0.004423231817781925, + 0.0028452598489820957, + 0.002941337414085865, + 0.0030500744469463825, + 0.0015594626311212778, + 0.001866292441263795, + 0.0013706174213439226, + 0.0013564834371209145, + 0.0019528805278241634, + 0.001472413307055831, + 0.0014885670971125364, + 0.0014076309744268656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.01427133847028017 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single fig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fig cracked open to reveal ruby-red pulp and tiny seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick drop of syrup emerging from the fig's interior and spreading downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fig resting on a dark, rustic wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019664489664137363 + ], + "opposite_logprob_tokens": [ + -6.25196647644043 + ], + "logprob_reward": -0.0019664489664137363, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019664489664137363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16555385291576385, + 0.17235271632671356, + 0.3314099907875061, + 0.051818978041410446, + 0.180349200963974, + 0.011540476232767105, + 0.014458967372775078, + 0.03173401951789856, + 0.026349207386374474, + 0.014432533644139767 + ], + "credit_map": [ + 0.009721709415316582, + 0.012412600219249725, + 0.04205923154950142, + 0.02125430479645729, + 0.018451692536473274, + 0.01220701728016138, + 0.015219208784401417, + 0.034228112548589706, + 0.012294630520045757, + 0.00818355567753315, + 0.04429244250059128, + 0.013942764140665531, + 0.02021772786974907, + 0.010625617578625679, + 0.008629094809293747, + 0.05416690185666084, + 0.01917262189090252, + 0.015468322671949863, + 0.1108856275677681, + 0.026206905022263527, + 0.015417497605085373, + 0.01901627890765667, + 0.01825110800564289, + 0.10699167847633362, + 0.0036094391252845526, + 0.005422143265604973, + 0.004522772040218115, + 0.005433094687759876, + 0.01266792044043541, + 0.005958832334727049, + 0.007158168125897646, + 0.007046615704894066, + 0.011606178246438503, + 0.01903141662478447, + 0.014123684726655483, + 0.04146592691540718, + 0.026937663555145264, + 0.013331846334040165, + 0.030200332403182983, + 0.023652156814932823, + 0.001209709094837308, + 0.0010938304476439953, + 0.0012924016918987036, + 0.0010659737745299935, + 0.0018302082316949964, + 0.0017892047762870789, + 0.0016348776407539845, + 0.0016242711571976542, + 0.0015422675060108304, + 0.0013816783903166652, + 0.00144912616815418, + 0.0014099564868956804, + 0.0022233084309846163, + 0.0017212813254445791, + 0.0025578360073268414, + 0.0021735152695327997, + 0.0030097845010459423, + 0.0026362882927060127, + 0.002359633333981037, + 0.004895300604403019, + 0.006648961920291185, + 0.004553890321403742, + 0.0044117094948887825, + 0.003218452911823988, + 0.0022174438927322626, + 0.003198499558493495, + 0.003093808889389038, + 0.0034013453405350447, + 0.004290498327463865, + 0.0031314068473875523, + 0.003130973782390356, + 0.00388523587025702, + 0.002090052468702197, + 0.0014058203669264913, + 0.0016464268555864692, + 0.001412940793670714, + 0.0017202101880684495, + 0.0022812867537140846, + 0.0015721590025350451, + 0.0023036389611661434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 1 + ], + "selected_reward": -0.0019664489664137363 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single fig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fig cracked open to reveal ruby-red pulp and tiny seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick drop of syrup emerging from the fig's interior and spreading downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fig resting on a dark, rustic wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020595386624336243 + ], + "opposite_logprob_tokens": [ + -3.8955953121185303 + ], + "logprob_reward": -0.020595386624336243, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020595386624336243, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17237722873687744, + 0.06573875993490219, + 0.42135491967201233, + 0.04711708053946495, + 0.19052062928676605, + 0.02419719099998474, + 0.012262869626283646, + 0.025010619312524796, + 0.0337982252240181, + 0.007622515317052603 + ], + "credit_map": [ + 0.012339761480689049, + 0.013121926225721836, + 0.041414011269807816, + 0.022349504753947258, + 0.020451383665204048, + 0.011553352698683739, + 0.01627722941339016, + 0.03487005829811096, + 0.004886721260845661, + 0.004618003498762846, + 0.013895762152969837, + 0.005151704885065556, + 0.008251847699284554, + 0.0051609366200864315, + 0.00448283925652504, + 0.019290946424007416, + 0.024314099922776222, + 0.018070893362164497, + 0.1442609280347824, + 0.034038711339235306, + 0.016852984204888344, + 0.02333810180425644, + 0.021102968603372574, + 0.13937625288963318, + 0.003862288547679782, + 0.004937761463224888, + 0.004586573224514723, + 0.006205283105373383, + 0.009903325699269772, + 0.004028771072626114, + 0.0061607505194842815, + 0.007432322949171066, + 0.011693133041262627, + 0.01898976042866707, + 0.013117675669491291, + 0.039726968854665756, + 0.025593100115656853, + 0.0128275565803051, + 0.037228818982839584, + 0.03134360536932945, + 0.0024565858766436577, + 0.0019440328469499946, + 0.0034042266197502613, + 0.0022731716744601727, + 0.003979234956204891, + 0.0026866518892347813, + 0.003824590938165784, + 0.0036286963149905205, + 0.0013582402607426047, + 0.0015529525699093938, + 0.0011116497917100787, + 0.0017395040486007929, + 0.002090696943923831, + 0.0012325076386332512, + 0.0018055426189675927, + 0.0013717758702114224, + 0.0017867442220449448, + 0.002164775040000677, + 0.0019153575412929058, + 0.0037187396083027124, + 0.005912221968173981, + 0.0027031938079744577, + 0.0036134645342826843, + 0.003196125151589513, + 0.0034742921125143766, + 0.003755942452698946, + 0.0034639229997992516, + 0.004520033020526171, + 0.005353090353310108, + 0.004106398206204176, + 0.004684161394834518, + 0.004440384916961193, + 0.000823287176899612, + 0.0006732270121574402, + 0.0007410329999402165, + 0.0009572371491231024, + 0.0012107124784961343, + 0.0011615243274718523, + 0.001032961648888886, + 0.0010225323494523764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.020595386624336243 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single fig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fig cracked open to reveal ruby-red pulp and tiny seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick drop of syrup emerging from the fig's interior and spreading downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fig resting on a dark, rustic wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018189411610364914 + ], + "opposite_logprob_tokens": [ + -4.018189430236816 + ], + "logprob_reward": -0.018189411610364914, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018189411610364914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5411422252655029, + 0.08912363648414612, + 0.1647646576166153, + 0.03509895130991936, + 0.05639424920082092, + 0.06804434955120087, + 0.00897040031850338, + 0.012556868605315685, + 0.004231057595461607, + 0.01967364363372326 + ], + "credit_map": [ + 0.03513164073228836, + 0.03500891849398613, + 0.20474176108837128, + 0.03513667359948158, + 0.023035559803247452, + 0.02952483482658863, + 0.022939803078770638, + 0.15562307834625244, + 0.0064978161826729774, + 0.004843324888497591, + 0.021147629246115685, + 0.007171842269599438, + 0.009936876595020294, + 0.005810788366943598, + 0.00624547153711319, + 0.027469882741570473, + 0.009745460934937, + 0.00738492188975215, + 0.05621309205889702, + 0.013301236554980278, + 0.006891621742397547, + 0.009710872545838356, + 0.0083929980173707, + 0.05312444269657135, + 0.0035050406586378813, + 0.002815250074490905, + 0.0038409815169870853, + 0.004609724506735802, + 0.006556084379553795, + 0.003115109633654356, + 0.004798495676368475, + 0.005858263466507196, + 0.004568526986986399, + 0.003379747038707137, + 0.00547433877363801, + 0.005424330476671457, + 0.0064462823793292046, + 0.004755142144858837, + 0.013665933161973953, + 0.01267994660884142, + 0.007489953655749559, + 0.006752918474376202, + 0.007961536757647991, + 0.0070673637092113495, + 0.008889680728316307, + 0.009941281750798225, + 0.010395769961178303, + 0.009545843116939068, + 0.0009922084864228964, + 0.000844519236125052, + 0.0007778066210448742, + 0.001132001867517829, + 0.00139105215203017, + 0.0009647516999393702, + 0.0014478910015895963, + 0.0014201693702489138, + 0.001161516411229968, + 0.0011524667497724295, + 0.001587899518199265, + 0.0015494701219722629, + 0.0014263354241847992, + 0.002349595306441188, + 0.0018221928039565682, + 0.0015073923859745264, + 0.0005147515330463648, + 0.00041749252704903483, + 0.0004489791754167527, + 0.0005756503669545054, + 0.00048004757263697684, + 0.0005581111763603985, + 0.0005942564457654953, + 0.0006417687982320786, + 0.0023036266211420298, + 0.0030247517861425877, + 0.0021349703893065453, + 0.0023591460194438696, + 0.0025039513129740953, + 0.002145956503227353, + 0.0024533395189791918, + 0.002747899852693081 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.018189411610364914 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single fig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fig cracked open to reveal ruby-red pulp and tiny seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick drop of syrup emerging from the fig's interior and spreading downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fig resting on a dark, rustic wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008639576844871044 + ], + "opposite_logprob_tokens": [ + -4.758639812469482 + ], + "logprob_reward": -0.008639576844871044, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008639576844871044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2922435700893402, + 0.13288697600364685, + 0.20279531180858612, + 0.05103180557489395, + 0.15898390114307404, + 0.029900236055254936, + 0.0196999441832304, + 0.033159881830215454, + 0.03534930944442749, + 0.04394904524087906 + ], + "credit_map": [ + 0.0195713359862566, + 0.019777566194534302, + 0.07602426409721375, + 0.03314924240112305, + 0.03492453321814537, + 0.022422198206186295, + 0.02531246468424797, + 0.061061978340148926, + 0.010010641068220139, + 0.006115826312452555, + 0.033131059259176254, + 0.010814458131790161, + 0.01540762186050415, + 0.010388672351837158, + 0.006506694946438074, + 0.04051199555397034, + 0.012003512121737003, + 0.009860910475254059, + 0.06592948734760284, + 0.016439922153949738, + 0.01027748454362154, + 0.011912830173969269, + 0.011730796657502651, + 0.06464038044214249, + 0.004664069507271051, + 0.005156954284757376, + 0.003957357257604599, + 0.005498112179338932, + 0.011511186137795448, + 0.005022638943046331, + 0.008087360300123692, + 0.007134127896279097, + 0.01089087687432766, + 0.01752062886953354, + 0.011757534928619862, + 0.03734343498945236, + 0.023499781265854836, + 0.011258103884756565, + 0.026113323867321014, + 0.02060021460056305, + 0.003990171477198601, + 0.00278518651612103, + 0.0030876807868480682, + 0.0029913941398262978, + 0.003235423471778631, + 0.004187759943306446, + 0.004594050347805023, + 0.005028570536524057, + 0.002917712787166238, + 0.0018380191177129745, + 0.0020482512190937996, + 0.0016897546593099833, + 0.002320859581232071, + 0.0029697890859097242, + 0.0028537327889353037, + 0.0030618254095315933, + 0.002829005941748619, + 0.002760055474936962, + 0.002720685675740242, + 0.005146809853613377, + 0.005611666012555361, + 0.005020411219447851, + 0.005091537721455097, + 0.003979708533734083, + 0.0033019892871379852, + 0.004149625077843666, + 0.0036338905338197947, + 0.0046439263969659805, + 0.006117477081716061, + 0.003941823728382587, + 0.004875945392996073, + 0.004684630781412125, + 0.005613798275589943, + 0.005983220413327217, + 0.005021072458475828, + 0.004914488643407822, + 0.005851044785231352, + 0.004768109880387783, + 0.00530205387622118, + 0.006495257373899221 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.008639576844871044 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot from a slightly elevated angle, focuses on a single, fully ripe fig, which has been cracked open to reveal its inner beauty while resting on a dark, rustic wooden surface. The fig's deep purple skin, with its delicate, powdery blossoms, gives way to a vibrant, velvety flesh within, rich ruby-red pulp and countless tiny, gleaming seeds. From the center of this intricate interior, a thick, shimmering drop of natural syrup slowly emerges. It swells into a perfect, heavy bead, catching a single, soft light source from the side, causing it to shimmer like liquid amber. The syrup begins to ooze out at a deliberate, almost imperceptible pace, its viscous form spreading over the delicate veins of the flesh, carving a viscous, shiny path downward. Dramatic black and white lighting highlights every minute detail—the shimmer of the syrup, the texture of the seeds, the gentle curves of the skin—while a rich palette of deep purple, red, and gold against the dark backdrop creates an atmosphere of profound sensuality and contemplative natural decadence.", + "seed": 1346, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_14/index_8633_seed_1346_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02627015709877014, + "vlm_logprob": -0.02627015709877014, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.02627015709877014, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single fig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fig cracked open to reveal ruby-red pulp and tiny seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick drop of syrup emerging from the fig's interior and spreading downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fig resting on a dark, rustic wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008083722554147243 + ], + "opposite_logprob_tokens": [ + -5.258083820343018 + ], + "logprob_reward": -0.008083722554147243, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.008083722554147243, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11516840010881424, + 0.13842695951461792, + 0.35901084542274475, + 0.12269498407840729, + 0.03340768814086914, + 0.034528929740190506, + 0.06718699634075165, + 0.04414595291018486, + 0.035736456513404846, + 0.049692850559949875 + ], + "credit_map": [ + 0.016067540273070335, + 0.007137005217373371, + 0.019928568974137306, + 0.013988524675369263, + 0.011935161426663399, + 0.015254758298397064, + 0.014185002073645592, + 0.016671840101480484, + 0.02176220342516899, + 0.011982387863099575, + 0.01967751234769821, + 0.017894349992275238, + 0.012899818830192089, + 0.016669439151883125, + 0.022878196090459824, + 0.014663046225905418, + 0.09529980272054672, + 0.023190904408693314, + 0.05360475182533264, + 0.026967525482177734, + 0.025407638400793076, + 0.03386254608631134, + 0.038753442466259, + 0.061924245208501816, + 0.018326276913285255, + 0.0097267497330904, + 0.018537534400820732, + 0.014800646342337132, + 0.009717914275825024, + 0.013689707964658737, + 0.023839270696043968, + 0.01405689399689436, + 0.004952528979629278, + 0.0024143680930137634, + 0.006752518471330404, + 0.0038888053968548775, + 0.0037531612906605005, + 0.0034983146470040083, + 0.00449648592621088, + 0.003651508828625083, + 0.008159486576914787, + 0.002312926109880209, + 0.006915176287293434, + 0.0030915518291294575, + 0.0023484965786337852, + 0.002900475636124611, + 0.00458362465724349, + 0.004217189271003008, + 0.011216973885893822, + 0.0068939500488340855, + 0.011165057308971882, + 0.007725940551608801, + 0.006290329620242119, + 0.007994147017598152, + 0.0074912430718541145, + 0.008409349247813225, + 0.007471396122127771, + 0.0033050982747226954, + 0.006239798851311207, + 0.005290199536830187, + 0.005706339608877897, + 0.0044419425539672375, + 0.006824749056249857, + 0.004866427276283503, + 0.007983099669218063, + 0.0026084806304425, + 0.005769924260675907, + 0.003155818209052086, + 0.003694825107231736, + 0.003263100516051054, + 0.0039343200623989105, + 0.005326888989657164, + 0.00789764802902937, + 0.004673322197049856, + 0.009195910766720772, + 0.004023808985948563, + 0.005582135636359453, + 0.007258519064635038, + 0.004944506101310253, + 0.006117004435509443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.008083722554147243 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single fig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fig cracked open to reveal ruby-red pulp and tiny seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick drop of syrup emerging from the fig's interior and spreading downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fig resting on a dark, rustic wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008678695186972618 + ], + "opposite_logprob_tokens": [ + -4.758678913116455 + ], + "logprob_reward": -0.008678695186972618, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008678695186972618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1386653631925583, + 0.18085382878780365, + 0.16541190445423126, + 0.08667517453432083, + 0.08159707486629486, + 0.09466569870710373, + 0.051348451524972916, + 0.06881434470415115, + 0.09798986464738846, + 0.033978283405303955 + ], + "credit_map": [ + 0.020691141486167908, + 0.010860545560717583, + 0.024549227207899094, + 0.014233051799237728, + 0.015933148562908173, + 0.018675200641155243, + 0.015863269567489624, + 0.017859771847724915, + 0.029790114611387253, + 0.01891951449215412, + 0.023529794067144394, + 0.02174331247806549, + 0.027379173785448074, + 0.017646826803684235, + 0.02659507468342781, + 0.015250015072524548, + 0.03633937984704971, + 0.012494485825300217, + 0.02157578244805336, + 0.017813757061958313, + 0.017089860513806343, + 0.014842324890196323, + 0.017429497092962265, + 0.027826813980937004, + 0.014342156238853931, + 0.0054293107241392136, + 0.01757647469639778, + 0.011355113238096237, + 0.009030304849147797, + 0.009495705366134644, + 0.00995651539415121, + 0.009489600546658039, + 0.01882515475153923, + 0.004751372616738081, + 0.011771829798817635, + 0.006296112667769194, + 0.010813293047249317, + 0.006992178037762642, + 0.010271691717207432, + 0.011875445954501629, + 0.019506216049194336, + 0.00719138840213418, + 0.015070582740008831, + 0.00928083062171936, + 0.009986188262701035, + 0.009355354122817516, + 0.012132667936384678, + 0.012142466381192207, + 0.009472781792283058, + 0.0032177448738366365, + 0.007866598665714264, + 0.006287753116339445, + 0.007256730459630489, + 0.004529799334704876, + 0.0067877876572310925, + 0.005929252598434687, + 0.01603027991950512, + 0.004669587127864361, + 0.00791037268936634, + 0.0069391257129609585, + 0.00951355416327715, + 0.004401577170938253, + 0.010152100585401058, + 0.009197748266160488, + 0.019730182364583015, + 0.007575082127004862, + 0.012718159705400467, + 0.00903085432946682, + 0.012861714698374271, + 0.007829256355762482, + 0.014186996035277843, + 0.01405762042850256, + 0.005331173073500395, + 0.003751482581719756, + 0.0060357628390192986, + 0.00264719408005476, + 0.004582265857607126, + 0.004706135950982571, + 0.003286458784714341, + 0.0036378123331815004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.008678695186972618 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single fig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fig cracked open to reveal ruby-red pulp and tiny seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick drop of syrup emerging from the fig's interior and spreading downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fig resting on a dark, rustic wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497291311621666 + ], + "opposite_logprob_tokens": [ + -2.9299728870391846 + ], + "logprob_reward": -0.05497291311621666, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497291311621666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1300962120294571, + 0.09959600865840912, + 0.08623603731393814, + 0.07099398225545883, + 0.08761967718601227, + 0.11160081624984741, + 0.06669953465461731, + 0.08761805295944214, + 0.18492192029953003, + 0.07461771368980408 + ], + "credit_map": [ + 0.01817982830107212, + 0.015026065520942211, + 0.02386162243783474, + 0.01659327745437622, + 0.013343366794288158, + 0.014984622597694397, + 0.01730750873684883, + 0.010799933224916458, + 0.013398327864706516, + 0.008609174750745296, + 0.015092962421476841, + 0.016184186562895775, + 0.01129159890115261, + 0.008268970064818859, + 0.01772666536271572, + 0.009024122729897499, + 0.012987352907657623, + 0.007299908436834812, + 0.009347131475806236, + 0.01168231200426817, + 0.011634099297225475, + 0.010705011896789074, + 0.011040794663131237, + 0.011539425700902939, + 0.010865516029298306, + 0.005546396132558584, + 0.013764593750238419, + 0.008143286220729351, + 0.006856015883386135, + 0.006099251098930836, + 0.010548156686127186, + 0.009170769713819027, + 0.014765283092856407, + 0.006474926136434078, + 0.014966730959713459, + 0.009454685263335705, + 0.00921121146529913, + 0.006852836348116398, + 0.013554370030760765, + 0.012339633889496326, + 0.020504944026470184, + 0.00923528615385294, + 0.018912388011813164, + 0.012609126977622509, + 0.010252299718558788, + 0.01074155606329441, + 0.01422971859574318, + 0.015115497633814812, + 0.013722356408834457, + 0.005544265266507864, + 0.012210561893880367, + 0.006990824360400438, + 0.007167107425630093, + 0.006706778425723314, + 0.006665979046374559, + 0.007691655773669481, + 0.014787623658776283, + 0.007581690791994333, + 0.010140822269022465, + 0.011609090492129326, + 0.01177120115607977, + 0.008137628436088562, + 0.013019702397286892, + 0.01057029701769352, + 0.033560916781425476, + 0.013562450185418129, + 0.022663749754428864, + 0.02384321577847004, + 0.02406405285000801, + 0.017802990972995758, + 0.028250861912965775, + 0.021173672750592232, + 0.011862357147037983, + 0.005856472998857498, + 0.011799143627285957, + 0.00950353592634201, + 0.011751599609851837, + 0.007280091755092144, + 0.009878539480268955, + 0.006685978267341852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.05497291311621666 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single fig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fig cracked open to reveal ruby-red pulp and tiny seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick drop of syrup emerging from the fig's interior and spreading downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fig resting on a dark, rustic wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497291311621666 + ], + "opposite_logprob_tokens": [ + -2.9299728870391846 + ], + "logprob_reward": -0.05497291311621666, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497291311621666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16080214083194733, + 0.09425853937864304, + 0.1553594470024109, + 0.17371052503585815, + 0.17107710242271423, + 0.07168512791395187, + 0.03828279674053192, + 0.0384114570915699, + 0.037098173052072525, + 0.05931469798088074 + ], + "credit_map": [ + 0.026583701372146606, + 0.013876885175704956, + 0.020358871668577194, + 0.01634596288204193, + 0.02297712117433548, + 0.01653319038450718, + 0.022646816447377205, + 0.02147960662841797, + 0.017943676561117172, + 0.008790950290858746, + 0.012923899106681347, + 0.011016054078936577, + 0.012304411269724369, + 0.007805704604834318, + 0.01395972166210413, + 0.00951412133872509, + 0.03944851830601692, + 0.010578872635960579, + 0.020417990162968636, + 0.013138525187969208, + 0.010397142730653286, + 0.014398667961359024, + 0.019905541092157364, + 0.027074195444583893, + 0.024994540959596634, + 0.011688224039971828, + 0.03242206946015358, + 0.02072259970009327, + 0.012882799841463566, + 0.01795390248298645, + 0.033318761736154556, + 0.019727632403373718, + 0.02995818667113781, + 0.013127092272043228, + 0.03005235455930233, + 0.02153327316045761, + 0.013590660877525806, + 0.017168769612908363, + 0.02531341277062893, + 0.02033334970474243, + 0.013890487141907215, + 0.006022898480296135, + 0.012498335912823677, + 0.007354444824159145, + 0.006235860753804445, + 0.007795494981110096, + 0.00822770781815052, + 0.009659894742071629, + 0.009215847589075565, + 0.00277880672365427, + 0.007485064212232828, + 0.002746363403275609, + 0.004476572386920452, + 0.003869523759931326, + 0.0032653440721333027, + 0.004445274826139212, + 0.007258098106831312, + 0.002961599500849843, + 0.005155546590685844, + 0.004664184525609016, + 0.005109698977321386, + 0.0030591769609600306, + 0.0058042495511472225, + 0.004398899152874947, + 0.005357641726732254, + 0.0034331842325627804, + 0.005263568367809057, + 0.003999347798526287, + 0.004907270893454552, + 0.004466230981051922, + 0.005689515266567469, + 0.003981415648013353, + 0.007840234786272049, + 0.006129895802587271, + 0.008341768756508827, + 0.005606582388281822, + 0.00832845363765955, + 0.008232943713665009, + 0.006897616200149059, + 0.007937195710837841 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.05497291311621666 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single fig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fig cracked open to reveal ruby-red pulp and tiny seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick drop of syrup emerging from the fig's interior and spreading downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fig resting on a dark, rustic wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004642541520297527 + ], + "opposite_logprob_tokens": [ + -5.379642486572266 + ], + "logprob_reward": -0.004642541520297527, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004642541520297527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18433262407779694, + 0.16709180176258087, + 0.09165308624505997, + 0.09665653109550476, + 0.10742271691560745, + 0.09293065220117569, + 0.051155101507902145, + 0.06671725958585739, + 0.09192539006471634, + 0.05011488497257233 + ], + "credit_map": [ + 0.02832731232047081, + 0.018346203491091728, + 0.028800148516893387, + 0.02136102318763733, + 0.02262762561440468, + 0.01900123618543148, + 0.02515202946960926, + 0.02071702852845192, + 0.029860779643058777, + 0.012212791480123997, + 0.024744411930441856, + 0.014910089783370495, + 0.01757357269525528, + 0.014552297070622444, + 0.01998702809214592, + 0.033250801265239716, + 0.020668301731348038, + 0.00574228260666132, + 0.014847726561129093, + 0.007543296553194523, + 0.00744344899430871, + 0.007895493879914284, + 0.00852183997631073, + 0.018990686163306236, + 0.015120403841137886, + 0.006988879293203354, + 0.024495458230376244, + 0.009908764623105526, + 0.008240163326263428, + 0.009358610957860947, + 0.01065932959318161, + 0.011884903535246849, + 0.014718922786414623, + 0.008631879463791847, + 0.021991508081555367, + 0.01278889924287796, + 0.011069138534367085, + 0.009777807630598545, + 0.014446310698986053, + 0.013998240232467651, + 0.013540011830627918, + 0.008187633939087391, + 0.013920878991484642, + 0.013767514377832413, + 0.010315384715795517, + 0.01059612538665533, + 0.012077734805643559, + 0.010525350458920002, + 0.00906558521091938, + 0.0055677625350654125, + 0.006891992408782244, + 0.006084531545639038, + 0.006055126432329416, + 0.0061991577968001366, + 0.004967755638062954, + 0.006323184818029404, + 0.008186754770576954, + 0.004959442187100649, + 0.007470102049410343, + 0.010573659092187881, + 0.009426109492778778, + 0.009264402091503143, + 0.009909235872328281, + 0.006927537731826305, + 0.02170381136238575, + 0.005584393627941608, + 0.014342297799885273, + 0.008569425903260708, + 0.009701201692223549, + 0.007844317704439163, + 0.010693307034671307, + 0.013486625626683235, + 0.006816760636866093, + 0.004673592746257782, + 0.008879287168383598, + 0.004543220158666372, + 0.007686158642172813, + 0.006616909988224506, + 0.006167162675410509, + 0.004731784574687481 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.004642541520297527 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot from a slightly elevated angle, focuses on a single, fully ripe fig, which has been cracked open to reveal its inner beauty while resting on a dark, rustic wooden surface. The fig's deep purple skin, with its delicate, powdery blossoms, gives way to a vibrant, velvety flesh within, rich ruby-red pulp and countless tiny, gleaming seeds. From the center of this intricate interior, a thick, shimmering drop of natural syrup slowly emerges. It swells into a perfect, heavy bead, catching a single, soft light source from the side, causing it to shimmer like liquid amber. The syrup begins to ooze out at a deliberate, almost imperceptible pace, its viscous form spreading over the delicate veins of the flesh, carving a viscous, shiny path downward. Dramatic black and white lighting highlights every minute detail—the shimmer of the syrup, the texture of the seeds, the gentle curves of the skin—while a rich palette of deep purple, red, and gold against the dark backdrop creates an atmosphere of profound sensuality and contemplative natural decadence.", + "seed": 1347, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_14/index_8633_seed_1347_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.007681435323320329, + "vlm_logprob": -0.007681435323320329, + "vlm_yes_no_margin": 5.225, + "vlm_token_credit": -0.007681435323320329, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single fig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fig cracked open to reveal ruby-red pulp and tiny seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick drop of syrup emerging from the fig's interior and spreading downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fig resting on a dark, rustic wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013276035897433758 + ], + "opposite_logprob_tokens": [ + -4.513276100158691 + ], + "logprob_reward": -0.013276035897433758, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.013276035897433758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10750700533390045, + 0.030383409932255745, + 0.01783795654773712, + 0.03008273057639599, + 0.043952662497758865, + 0.02706567570567131, + 0.10810435563325882, + 0.0949510857462883, + 0.43071505427360535, + 0.10940009355545044 + ], + "credit_map": [ + 0.01262715458869934, + 0.009335897862911224, + 0.013565368950366974, + 0.00936111994087696, + 0.014031141065061092, + 0.021325139328837395, + 0.008856460452079773, + 0.018404727801680565, + 0.004213172011077404, + 0.002037622034549713, + 0.005190305877476931, + 0.003085825592279434, + 0.002399540739133954, + 0.005387494806200266, + 0.0029556879308074713, + 0.005113762337714434, + 0.0023252239916473627, + 0.0011167085031047463, + 0.003415284212678671, + 0.0016872541746124625, + 0.0018066454213112593, + 0.0031865397468209267, + 0.0016610384918749332, + 0.0026392615400254726, + 0.0033884288277477026, + 0.0030989195220172405, + 0.002959739649668336, + 0.0053647356107831, + 0.006527411285787821, + 0.00271434779278934, + 0.0033664703369140625, + 0.0026626759208738804, + 0.0062127066776156425, + 0.0027822658885270357, + 0.005953886080533266, + 0.004313373938202858, + 0.005180673208087683, + 0.004417195450514555, + 0.004882471635937691, + 0.010210092179477215, + 0.003753810189664364, + 0.002485614037141204, + 0.003311139764264226, + 0.004030083306133747, + 0.002822344424203038, + 0.004087723325937986, + 0.002607099013403058, + 0.003967862576246262, + 0.027141928672790527, + 0.005445627961307764, + 0.011175896972417831, + 0.006206626538187265, + 0.009409815073013306, + 0.009893900714814663, + 0.0072668748907744884, + 0.03156368434429169, + 0.018253371119499207, + 0.005152225960046053, + 0.006486181169748306, + 0.005120300687849522, + 0.007162149995565414, + 0.01278742216527462, + 0.00651915417984128, + 0.0334702767431736, + 0.1108565405011177, + 0.019667332991957664, + 0.020797720178961754, + 0.013144011609256268, + 0.03596118837594986, + 0.04409527778625488, + 0.02187015488743782, + 0.1643228381872177, + 0.03193215653300285, + 0.005373572465032339, + 0.005953491199761629, + 0.00417683832347393, + 0.006296725943684578, + 0.009150813333690166, + 0.005985098425298929, + 0.040531404316425323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.013276035897433758 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single fig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fig cracked open to reveal ruby-red pulp and tiny seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick drop of syrup emerging from the fig's interior and spreading downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fig resting on a dark, rustic wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013612775364890695 + ], + "opposite_logprob_tokens": [ + -6.62636137008667 + ], + "logprob_reward": -0.0013612775364890695, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013612775364890695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19335205852985382, + 0.026454277336597443, + 0.013149982318282127, + 0.014889407902956009, + 0.016838455572724342, + 0.027713073417544365, + 0.1091548353433609, + 0.15034878253936768, + 0.28541699051856995, + 0.1626821756362915 + ], + "credit_map": [ + 0.022636080160737038, + 0.015534170903265476, + 0.021028168499469757, + 0.015380538068711758, + 0.021119466051459312, + 0.04441435635089874, + 0.015906989574432373, + 0.03733229637145996, + 0.004023069050163031, + 0.0019034186843782663, + 0.004030859563499689, + 0.0025691904593259096, + 0.0030237967148423195, + 0.004340933635830879, + 0.0034438115544617176, + 0.0031191976740956306, + 0.0016828759107738733, + 0.0009132480481639504, + 0.0023845196701586246, + 0.0016828313237056136, + 0.0017252661054953933, + 0.0012768316082656384, + 0.0014971980126574636, + 0.001987212337553501, + 0.002427884843200445, + 0.0015446648467332125, + 0.0013421486364677548, + 0.0015337055083364248, + 0.0024424861185252666, + 0.002311177784577012, + 0.001993548823520541, + 0.0012937909923493862, + 0.002240453613922, + 0.0013308980269357562, + 0.001806748448871076, + 0.0019373224349692464, + 0.0020984183065593243, + 0.00176442030351609, + 0.0018154451390728354, + 0.003844749415293336, + 0.002888748189434409, + 0.0025553188752382994, + 0.0040014805272221565, + 0.0033961895387619734, + 0.0036064060404896736, + 0.0037694599013775587, + 0.003215463599190116, + 0.00428000558167696, + 0.026906287297606468, + 0.005754946731030941, + 0.009442036040127277, + 0.00764232873916626, + 0.011685059405863285, + 0.010519282892346382, + 0.008204624056816101, + 0.02900027483701706, + 0.026398152112960815, + 0.008037028834223747, + 0.009922170080244541, + 0.006783198565244675, + 0.010981523431837559, + 0.024143723770976067, + 0.01035237219184637, + 0.05373060703277588, + 0.0740424245595932, + 0.013199377804994583, + 0.014108344912528992, + 0.008695925585925579, + 0.02362287975847721, + 0.02873695082962513, + 0.014133326709270477, + 0.10887779295444489, + 0.045913416892290115, + 0.008225234225392342, + 0.009133120998740196, + 0.0073502687737345695, + 0.009740818291902542, + 0.014200618490576744, + 0.008560992777347565, + 0.059557706117630005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.0013612775364890695 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single fig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fig cracked open to reveal ruby-red pulp and tiny seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick drop of syrup emerging from the fig's interior and spreading downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fig resting on a dark, rustic wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014204232022166252 + ], + "opposite_logprob_tokens": [ + -4.264204025268555 + ], + "logprob_reward": -0.014204232022166252, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014204232022166252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16762852668762207, + 0.06943508237600327, + 0.020858094096183777, + 0.06522379070520401, + 0.03947016969323158, + 0.09145887196063995, + 0.1375502049922943, + 0.06905920058488846, + 0.24348565936088562, + 0.09583036601543427 + ], + "credit_map": [ + 0.019602185115218163, + 0.015179409645497799, + 0.022130820900201797, + 0.015720216557383537, + 0.02032230794429779, + 0.026285652071237564, + 0.02079802379012108, + 0.027589906007051468, + 0.006791128311306238, + 0.006203461438417435, + 0.008743400685489178, + 0.008827703073620796, + 0.012217670679092407, + 0.006337603088468313, + 0.01247275061905384, + 0.007841370068490505, + 0.003086807206273079, + 0.00226816744543612, + 0.0025545808020979166, + 0.002609892515465617, + 0.0030863157007843256, + 0.002357034245505929, + 0.0023815466556698084, + 0.0025137499906122684, + 0.00879766047000885, + 0.005947480909526348, + 0.0076492358930408955, + 0.011909577995538712, + 0.011863651685416698, + 0.006961633451282978, + 0.0067728846333920956, + 0.005321658216416836, + 0.004209572449326515, + 0.002445126650854945, + 0.003952749539166689, + 0.005203723907470703, + 0.0051712351851165295, + 0.005484982393682003, + 0.004128223285079002, + 0.008874558843672276, + 0.02006356790661812, + 0.006097364239394665, + 0.007566442247480154, + 0.008023077622056007, + 0.009550001472234726, + 0.024994146078824997, + 0.006787583697587252, + 0.00837668962776661, + 0.02550719492137432, + 0.007277249358594418, + 0.00964121799916029, + 0.007416764739900827, + 0.009566279128193855, + 0.01996385119855404, + 0.00930970348417759, + 0.04886794090270996, + 0.009724322706460953, + 0.005695539992302656, + 0.008918619714677334, + 0.0066935340873897076, + 0.007397864013910294, + 0.008285848423838615, + 0.007861531339585781, + 0.014481943100690842, + 0.06190870702266693, + 0.010996164754033089, + 0.01255710143595934, + 0.007243796717375517, + 0.01884678192436695, + 0.02610345557332039, + 0.012011384591460228, + 0.09381826221942902, + 0.02322082407772541, + 0.005391686223447323, + 0.006150457076728344, + 0.004560726694762707, + 0.007092299405485392, + 0.01766279526054859, + 0.005806237459182739, + 0.02594533935189247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.014204232022166252 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single fig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fig cracked open to reveal ruby-red pulp and tiny seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick drop of syrup emerging from the fig's interior and spreading downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fig resting on a dark, rustic wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00361634255386889 + ], + "opposite_logprob_tokens": [ + -5.6286163330078125 + ], + "logprob_reward": -0.00361634255386889, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00361634255386889, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23777379095554352, + 0.03291548416018486, + 0.013917719945311546, + 0.012586317956447601, + 0.19308333098888397, + 0.21726132929325104, + 0.048422474414110184, + 0.07751490920782089, + 0.08700498938560486, + 0.07951962202787399 + ], + "credit_map": [ + 0.028606396168470383, + 0.021490270271897316, + 0.027257056906819344, + 0.021582767367362976, + 0.026919342577457428, + 0.0494939386844635, + 0.02161547914147377, + 0.0408085472881794, + 0.003988916520029306, + 0.002544423332437873, + 0.0073164356872439384, + 0.003960402682423592, + 0.0030088245403021574, + 0.003983610309660435, + 0.0038711605593562126, + 0.004241712391376495, + 0.0026116918306797743, + 0.0009789180476218462, + 0.0027225534431636333, + 0.001583913923241198, + 0.0012757070362567902, + 0.0017328354297205806, + 0.0014895477797836065, + 0.0015225516399368644, + 0.0015571715775877237, + 0.0010743949096649885, + 0.0018249277491122484, + 0.001649027573876083, + 0.002034929348155856, + 0.0012490638764575124, + 0.0013594962656497955, + 0.0018373074708506465, + 0.021462762728333473, + 0.009762964211404324, + 0.03263729065656662, + 0.01894773356616497, + 0.013799219392240047, + 0.03300350159406662, + 0.01708594523370266, + 0.04638391733169556, + 0.03254624828696251, + 0.007947168312966824, + 0.025298401713371277, + 0.011022120714187622, + 0.008767063729465008, + 0.06653542816638947, + 0.011430095881223679, + 0.053714778274297714, + 0.005582923535257578, + 0.0035424144007265568, + 0.012078830040991306, + 0.004765873774886131, + 0.0035825781524181366, + 0.006539414636790752, + 0.0038317833095788956, + 0.008498653769493103, + 0.013012136332690716, + 0.005826141219586134, + 0.007173274643719196, + 0.005614507012069225, + 0.007823387160897255, + 0.008910697884857655, + 0.006841321010142565, + 0.022313447669148445, + 0.022617939859628677, + 0.004276748280972242, + 0.00497164623811841, + 0.003131261095404625, + 0.007141196634620428, + 0.008177967742085457, + 0.0052358293905854225, + 0.03145240247249603, + 0.010662472806870937, + 0.005159483291208744, + 0.0073025901801884174, + 0.011102303862571716, + 0.007894247770309448, + 0.017726343125104904, + 0.008215317502617836, + 0.011456865817308426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.00361634255386889 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single fig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fig cracked open to reveal ruby-red pulp and tiny seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick drop of syrup emerging from the fig's interior and spreading downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fig resting on a dark, rustic wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005949288606643677 + ], + "opposite_logprob_tokens": [ + -5.1309494972229 + ], + "logprob_reward": -0.005949288606643677, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005949288606643677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14909392595291138, + 0.050251733511686325, + 0.030042653903365135, + 0.09251700341701508, + 0.041502658277750015, + 0.16047708690166473, + 0.1836557686328888, + 0.1336366832256317, + 0.08789185434579849, + 0.07093067467212677 + ], + "credit_map": [ + 0.017365992069244385, + 0.014007432386279106, + 0.018416663631796837, + 0.015301760286092758, + 0.01969216950237751, + 0.02320755459368229, + 0.017520910128951073, + 0.023581452667713165, + 0.006084024906158447, + 0.003823992796242237, + 0.011502948589622974, + 0.004650233779102564, + 0.004374952986836433, + 0.006137812975794077, + 0.006250779144465923, + 0.007426985539495945, + 0.004335210658609867, + 0.0016326025361195207, + 0.00555973406881094, + 0.0024552822578698397, + 0.0025376607663929462, + 0.006178774870932102, + 0.0027255737222731113, + 0.004617815837264061, + 0.011149552650749683, + 0.008501292206346989, + 0.010473987087607384, + 0.017218181863427162, + 0.01601347327232361, + 0.009861964732408524, + 0.010439248755574226, + 0.008859301917254925, + 0.006327024195343256, + 0.002283594338223338, + 0.006668251007795334, + 0.004015270620584488, + 0.004635695368051529, + 0.007323092315346003, + 0.0038405407685786486, + 0.006409191060811281, + 0.03131616860628128, + 0.01250121183693409, + 0.012796700932085514, + 0.015320156700909138, + 0.021383635699748993, + 0.030900465324521065, + 0.014004596509039402, + 0.022254152223467827, + 0.01817563734948635, + 0.011639676988124847, + 0.025199193507432938, + 0.02153567411005497, + 0.022222133353352547, + 0.042224425822496414, + 0.015114526264369488, + 0.027544504031538963, + 0.020726516842842102, + 0.008265423588454723, + 0.010456123389303684, + 0.007958647795021534, + 0.011404372751712799, + 0.02071782387793064, + 0.011385109275579453, + 0.0427226684987545, + 0.022838523611426353, + 0.004270260687917471, + 0.005680541042238474, + 0.0032030362635850906, + 0.007487837690860033, + 0.00841726828366518, + 0.004213834181427956, + 0.031780555844306946, + 0.013562213629484177, + 0.003626474179327488, + 0.00490137655287981, + 0.0032765658106654882, + 0.006711841095238924, + 0.011883814819157124, + 0.005326766520738602, + 0.021641617640852928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.005949288606643677 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot from a slightly elevated angle, focuses on a single, fully ripe fig, which has been cracked open to reveal its inner beauty while resting on a dark, rustic wooden surface. The fig's deep purple skin, with its delicate, powdery blossoms, gives way to a vibrant, velvety flesh within, rich ruby-red pulp and countless tiny, gleaming seeds. From the center of this intricate interior, a thick, shimmering drop of natural syrup slowly emerges. It swells into a perfect, heavy bead, catching a single, soft light source from the side, causing it to shimmer like liquid amber. The syrup begins to ooze out at a deliberate, almost imperceptible pace, its viscous form spreading over the delicate veins of the flesh, carving a viscous, shiny path downward. Dramatic black and white lighting highlights every minute detail—the shimmer of the syrup, the texture of the seeds, the gentle curves of the skin—while a rich palette of deep purple, red, and gold against the dark backdrop creates an atmosphere of profound sensuality and contemplative natural decadence.", + "seed": 1348, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_14/index_8633_seed_1348_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01481231008656323, + "vlm_logprob": -0.01481231008656323, + "vlm_yes_no_margin": 5.05, + "vlm_token_credit": -0.01481231008656323, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single fig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fig cracked open to reveal ruby-red pulp and tiny seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick drop of syrup emerging from the fig's interior and spreading downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fig resting on a dark, rustic wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.051468625664711 + ], + "opposite_logprob_tokens": [ + -3.05146861076355 + ], + "logprob_reward": -0.051468625664711, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.051468625664711, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12034060806035995, + 0.026435067877173424, + 0.09495380520820618, + 0.07314374297857285, + 0.1270948201417923, + 0.054760731756687164, + 0.06864906847476959, + 0.06869979947805405, + 0.08064432442188263, + 0.2852780520915985 + ], + "credit_map": [ + 0.012052926234900951, + 0.018168410286307335, + 0.011117828078567982, + 0.021946998313069344, + 0.008600585162639618, + 0.015058783814311028, + 0.012952827848494053, + 0.020442234352231026, + 0.0024119552690535784, + 0.003961651585996151, + 0.002253369428217411, + 0.005335716996341944, + 0.0020902780815958977, + 0.0028180403169244528, + 0.0030853389762341976, + 0.00447871396318078, + 0.010550225153565407, + 0.009113005362451077, + 0.008338896557688713, + 0.014161142520606518, + 0.011963073164224625, + 0.014169159345328808, + 0.01068754494190216, + 0.0159707423299551, + 0.006005056202411652, + 0.006036521401256323, + 0.005563032813370228, + 0.008969181217253208, + 0.007546230684965849, + 0.0106808478012681, + 0.01014211867004633, + 0.018200743943452835, + 0.010903512127697468, + 0.011372389271855354, + 0.010050666518509388, + 0.018305525183677673, + 0.013180073350667953, + 0.016486825421452522, + 0.01692403480410576, + 0.029871776700019836, + 0.004755851812660694, + 0.004900419618934393, + 0.0048459358513355255, + 0.00824644509702921, + 0.004675550851970911, + 0.007513951510190964, + 0.007684813346713781, + 0.012137760408222675, + 0.010339850559830666, + 0.008687715046107769, + 0.007417129818350077, + 0.008203725330531597, + 0.007462269626557827, + 0.007854355499148369, + 0.007865461520850658, + 0.010818557813763618, + 0.007511640898883343, + 0.005643059033900499, + 0.007279696874320507, + 0.008175717666745186, + 0.005292041692882776, + 0.011531205847859383, + 0.006207386497408152, + 0.017059043049812317, + 0.0065168822184205055, + 0.006831906735897064, + 0.0062946355901658535, + 0.009446900337934494, + 0.005459723528474569, + 0.011885851621627808, + 0.01247469987720251, + 0.021733716130256653, + 0.029878942295908928, + 0.03310444578528404, + 0.03055243007838726, + 0.039058271795511246, + 0.027923503890633583, + 0.030505666509270668, + 0.03675869479775429, + 0.057496100664138794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.051468625664711 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single fig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fig cracked open to reveal ruby-red pulp and tiny seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick drop of syrup emerging from the fig's interior and spreading downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fig resting on a dark, rustic wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013635394861921668 + ], + "opposite_logprob_tokens": [ + -6.626363754272461 + ], + "logprob_reward": -0.0013635394861921668, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013635394861921668, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20269489288330078, + 0.1214003935456276, + 0.06580585986375809, + 0.054907508194446564, + 0.07664228230714798, + 0.034891799092292786, + 0.053503599017858505, + 0.05230943113565445, + 0.05727986618876457, + 0.2805643379688263 + ], + "credit_map": [ + 0.019959187135100365, + 0.03452953323721886, + 0.021890347823500633, + 0.0354568250477314, + 0.015767158940434456, + 0.02501586824655533, + 0.01870756410062313, + 0.031368423253297806, + 0.013890719972550869, + 0.012335670180618763, + 0.013819139450788498, + 0.018211347982287407, + 0.010412972420454025, + 0.019010445103049278, + 0.012970436364412308, + 0.020749665796756744, + 0.007238856051117182, + 0.005644811782985926, + 0.005901604425162077, + 0.010434259660542011, + 0.007593843154609203, + 0.009680421091616154, + 0.006959692109376192, + 0.012352372519671917, + 0.004198869690299034, + 0.0046552252024412155, + 0.004173023626208305, + 0.007389106787741184, + 0.005823261570185423, + 0.008486716076731682, + 0.006617092527449131, + 0.01356421411037445, + 0.0066984002478420734, + 0.007891703397035599, + 0.005475596059113741, + 0.012654687277972698, + 0.009158365428447723, + 0.010140063241124153, + 0.007435743696987629, + 0.01718772016465664, + 0.003860512049868703, + 0.0038848251570016146, + 0.0037002263125032187, + 0.0045322212390601635, + 0.0029856013134121895, + 0.004677406512200832, + 0.004322481341660023, + 0.006928526796400547, + 0.00802927277982235, + 0.005746285896748304, + 0.005854361690580845, + 0.004548355471342802, + 0.005160054191946983, + 0.0068936217576265335, + 0.0064327409490942955, + 0.010838904418051243, + 0.006035021506249905, + 0.0050673414953053, + 0.005626718979328871, + 0.006369654554873705, + 0.004345021676272154, + 0.007958839647471905, + 0.005174463614821434, + 0.011732370592653751, + 0.006772398017346859, + 0.005522940773516893, + 0.006731134373694658, + 0.005714189726859331, + 0.005237407051026821, + 0.006768055260181427, + 0.008771208114922047, + 0.011762531474232674, + 0.031138481572270393, + 0.03272290527820587, + 0.03375016525387764, + 0.035316914319992065, + 0.02788449078798294, + 0.030573181807994843, + 0.03633506968617439, + 0.0528431236743927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.0013635394861921668 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single fig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fig cracked open to reveal ruby-red pulp and tiny seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick drop of syrup emerging from the fig's interior and spreading downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fig resting on a dark, rustic wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012549126520752907 + ], + "opposite_logprob_tokens": [ + -4.387548923492432 + ], + "logprob_reward": -0.012549126520752907, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012549126520752907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14055930078029633, + 0.24460920691490173, + 0.050627678632736206, + 0.12943550944328308, + 0.16005615890026093, + 0.05874381586909294, + 0.06541823595762253, + 0.03393985703587532, + 0.029498418793082237, + 0.08711181581020355 + ], + "credit_map": [ + 0.012596440501511097, + 0.01810538023710251, + 0.014403951354324818, + 0.022460896521806717, + 0.00933882687240839, + 0.019351845607161522, + 0.015500308945775032, + 0.028801659122109413, + 0.028135428205132484, + 0.026559410616755486, + 0.02848874405026436, + 0.04384863004088402, + 0.02250933088362217, + 0.03280903026461601, + 0.028129035606980324, + 0.034129589796066284, + 0.004919720347970724, + 0.004496610723435879, + 0.004250641912221909, + 0.008660932071506977, + 0.005478696897625923, + 0.006545572075992823, + 0.006420701276510954, + 0.009854805655777454, + 0.010904376395046711, + 0.014554537832736969, + 0.008455811068415642, + 0.02126823365688324, + 0.016502201557159424, + 0.019696123898029327, + 0.011665649712085724, + 0.026388579979538918, + 0.015600812621414661, + 0.025747042149305344, + 0.009602734819054604, + 0.023627551272511482, + 0.027161676436662674, + 0.01818305440247059, + 0.01625228114426136, + 0.02388100139796734, + 0.005986496340483427, + 0.008792522363364697, + 0.004000722896307707, + 0.007265207823365927, + 0.00874488614499569, + 0.006820117589086294, + 0.008026153780519962, + 0.009107707999646664, + 0.009898070245981216, + 0.006945031229406595, + 0.007102003786712885, + 0.005675884895026684, + 0.005749431438744068, + 0.008338348008692265, + 0.007951865904033184, + 0.013757600449025631, + 0.0027200167533010244, + 0.0035923549439758062, + 0.002995493356138468, + 0.005915215238928795, + 0.0033528325147926807, + 0.004593148827552795, + 0.003605379955843091, + 0.007165414281189442, + 0.0036381171084940434, + 0.00304242642596364, + 0.0035577823873609304, + 0.0032098263036459684, + 0.002197596710175276, + 0.004062948282808065, + 0.003656181739643216, + 0.006133540067821741, + 0.00927070900797844, + 0.012113469652831554, + 0.006648127920925617, + 0.015474114567041397, + 0.008400209248065948, + 0.00895241741091013, + 0.009996749460697174, + 0.016256015747785568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.012549126520752907 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single fig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fig cracked open to reveal ruby-red pulp and tiny seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick drop of syrup emerging from the fig's interior and spreading downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fig resting on a dark, rustic wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019415830029174685 + ], + "opposite_logprob_tokens": [ + -6.251941680908203 + ], + "logprob_reward": -0.0019415830029174685, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019415830029174685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4165157973766327, + 0.07900834828615189, + 0.057391438633203506, + 0.0371011346578598, + 0.09004373103380203, + 0.09307106584310532, + 0.07239840179681778, + 0.0599740706384182, + 0.028349976986646652, + 0.06614603102207184 + ], + "credit_map": [ + 0.04176291823387146, + 0.05112209543585777, + 0.042597174644470215, + 0.05066503956913948, + 0.029891900718212128, + 0.062479548156261444, + 0.05348384752869606, + 0.08451328426599503, + 0.007167130243033171, + 0.007013992872089148, + 0.006236115004867315, + 0.014459703117609024, + 0.006316450424492359, + 0.012157865799963474, + 0.007142143789678812, + 0.01851494424045086, + 0.006813565734773874, + 0.006806997116655111, + 0.005826099775731564, + 0.007018668577075005, + 0.005974604282528162, + 0.008099360391497612, + 0.006321183871477842, + 0.010530960746109486, + 0.0052939425222575665, + 0.0034224428236484528, + 0.003035711357370019, + 0.005401968490332365, + 0.003251661779358983, + 0.004375454504042864, + 0.0038282033056020737, + 0.008491748943924904, + 0.007987191900610924, + 0.007641802076250315, + 0.006742956582456827, + 0.012313943356275558, + 0.008059244602918625, + 0.012493043206632137, + 0.011836319230496883, + 0.02296922728419304, + 0.008841482922434807, + 0.0073626041412353516, + 0.007232582196593285, + 0.007357765920460224, + 0.006651334930211306, + 0.016322389245033264, + 0.009182061068713665, + 0.030120836570858955, + 0.009920123964548111, + 0.008122507482767105, + 0.0073469774797558784, + 0.007309677079319954, + 0.00704104732722044, + 0.009198516607284546, + 0.008536556735634804, + 0.01492299884557724, + 0.00648243585601449, + 0.004415586125105619, + 0.0058929226361215115, + 0.005912442225962877, + 0.004044162575155497, + 0.010798679664731026, + 0.00548554677516222, + 0.01694229058921337, + 0.0034224800765514374, + 0.002729644300416112, + 0.003531035967171192, + 0.003350482787936926, + 0.002511166501790285, + 0.0035689575597643852, + 0.0038651316426694393, + 0.005371079780161381, + 0.007521106395870447, + 0.006799692753702402, + 0.0075008259154856205, + 0.0079141641035676, + 0.006839020177721977, + 0.009237073361873627, + 0.007050730753690004, + 0.01328341756016016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.0019415830029174685 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single fig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fig cracked open to reveal ruby-red pulp and tiny seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick drop of syrup emerging from the fig's interior and spreading downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fig resting on a dark, rustic wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738675758242607 + ], + "opposite_logprob_tokens": [ + -5.006738662719727 + ], + "logprob_reward": -0.006738675758242607, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738675758242607, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12235286086797714, + 0.06251408904790878, + 0.04607051610946655, + 0.10263010114431381, + 0.11108539253473282, + 0.10681845992803574, + 0.2649441957473755, + 0.022809986025094986, + 0.03651728853583336, + 0.12425711750984192 + ], + "credit_map": [ + 0.013585694134235382, + 0.014161907136440277, + 0.01097860187292099, + 0.014506222680211067, + 0.008536035194993019, + 0.017571862787008286, + 0.01744859665632248, + 0.025563936680555344, + 0.007013706490397453, + 0.005247610155493021, + 0.005573576781898737, + 0.008421401493251324, + 0.005458153784275055, + 0.009659240022301674, + 0.005590345710515976, + 0.015550055541098118, + 0.004499366972595453, + 0.005531917791813612, + 0.0036064707674086094, + 0.007862498052418232, + 0.004747247789055109, + 0.00566779775545001, + 0.0052980659529566765, + 0.008857151493430138, + 0.009343979880213737, + 0.013563988730311394, + 0.007132441736757755, + 0.018973803147673607, + 0.012897411361336708, + 0.010258040390908718, + 0.01259336993098259, + 0.017867064103484154, + 0.009173053316771984, + 0.011140260845422745, + 0.008276964537799358, + 0.019307676702737808, + 0.012878822162747383, + 0.01435980573296547, + 0.011756740510463715, + 0.02419206313788891, + 0.009653883054852486, + 0.013500182889401913, + 0.00852904375642538, + 0.01359616219997406, + 0.018302395939826965, + 0.015523625537753105, + 0.008367999456822872, + 0.01934516802430153, + 0.03672861307859421, + 0.03397592902183533, + 0.029082216322422028, + 0.03568168729543686, + 0.03376298397779465, + 0.029018498957157135, + 0.029995623975992203, + 0.036698661744594574, + 0.002121638273820281, + 0.002509674523025751, + 0.0022331567015498877, + 0.0037659169174730778, + 0.0018596918089315295, + 0.002692685229703784, + 0.0029612844809889793, + 0.004665936343371868, + 0.0045709689147770405, + 0.004688852466642857, + 0.004423319362103939, + 0.004356601275503635, + 0.002919531427323818, + 0.0038819657638669014, + 0.005280417390167713, + 0.006395635660737753, + 0.014170574024319649, + 0.013675784692168236, + 0.014754674397408962, + 0.01606070250272751, + 0.01240838598459959, + 0.015208199620246887, + 0.016237569972872734, + 0.02174123004078865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.006738675758242607 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_15/index_tensor([8393, 8393])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_15/index_tensor([8393, 8393])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7a3fafc74fb2c8106fff979743d8a2913f819a2d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_15/index_tensor([8393, 8393])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, shot from an extremely low angle, almost level with the sand, makes the two crabs facing each other appear incredibly large and intimidating. The last rays of sunset bathe the beach in a warm orange-red hue, and the moist sand reflects the vibrant sunset sky, giving it a mirror-like sheen. In the center of the frame, a crab with a deep russet-brown shell, its surface rough and uneven, covered in tiny dents and the scratches of time, raises its disproportionately large right claw high, the serrations on its inner side gleaming coldly in the golden light, striking a menacing pose. A few inches away, another, smaller but equally aggressive crab, its shell a darker gray, lowers its body, opening and closing its two similarly sized claws on its chest. Cautiously moving its feet sideways, stirring up the thin layer of seawater beneath its feet. The camera remains still, steadily capturing this silent duel. The low-angle sunlight outlines their contours with a bright golden rim of light and casts extremely long, dancing shadows behind them. In the background, the distant waves, dyed a mixture of purple and orange, gently and rhythmically roll towards the shore, forming a sharp contrast between movement and stillness, tranquility and tension with the fierce confrontation on the verge of breaking out in the foreground. The entire scene is filled with the dramatic tension of primal vitality.", + "seed": 8604, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_15/index_8393_seed_8604_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.09379077926278115, + "vlm_logprob": -0.09379077926278115, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.09379077926278115, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle, almost level with the sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame facing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the larger crab raising its right claw high while the smaller crab lowers its body and moves its feet sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a beach at sunset with warm orange-red lighting and a mirror-like sheen on the moist sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02119230106472969 + ], + "opposite_logprob_tokens": [ + -4.3961920738220215 + ], + "logprob_reward": -0.02119230106472969, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.02119230106472969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5203449130058289, + 0.13806362450122833, + 0.13622009754180908, + 0.030270613729953766, + 0.026514798402786255, + 0.03658262640237808, + 0.034514691680669785, + 0.037819184362888336, + 0.021443761885166168, + 0.018225759267807007 + ], + "credit_map": [ + 0.1875603199005127, + 0.2159421443939209, + 0.011035715229809284, + 0.009270831942558289, + 0.02072496898472309, + 0.038114018738269806, + 0.010936782695353031, + 0.026760075241327286, + 0.022982630878686905, + 0.017965471372008324, + 0.01189228892326355, + 0.011085334233939648, + 0.017285559326410294, + 0.03250312805175781, + 0.01083793118596077, + 0.013511250726878643, + 0.048858534544706345, + 0.05294709652662277, + 0.005378527566790581, + 0.005333433859050274, + 0.006376711651682854, + 0.00458088144659996, + 0.004703203681856394, + 0.0080416863784194, + 0.004740863107144833, + 0.005817006807774305, + 0.0032083257101476192, + 0.002587343566119671, + 0.004187906160950661, + 0.0032845819368958473, + 0.002923571038991213, + 0.0035210095811635256, + 0.003740676213055849, + 0.004264281131327152, + 0.00351352640427649, + 0.002479265443980694, + 0.00283520738594234, + 0.0029803519137203693, + 0.003350198734551668, + 0.00335128465667367, + 0.010450047440826893, + 0.009285373613238335, + 0.0035715335980057716, + 0.00193397153634578, + 0.003762209555134177, + 0.0032057971693575382, + 0.0019733491353690624, + 0.0024003395810723305, + 0.0067791081964969635, + 0.004545958712697029, + 0.0034372389782220125, + 0.0022792217787355185, + 0.006037398241460323, + 0.004709962289780378, + 0.0037641862872987986, + 0.002961610909551382, + 0.005513476207852364, + 0.005788445472717285, + 0.00547353271394968, + 0.004097359254956245, + 0.004436633549630642, + 0.00405595637857914, + 0.003799902740865946, + 0.004653870593756437, + 0.002864257898181677, + 0.003125151852145791, + 0.002591995522379875, + 0.0017343584913760424, + 0.0030297564808279276, + 0.002624975750222802, + 0.0019490690901875496, + 0.0035241928417235613, + 0.0034036864526569843, + 0.0027914377860724926, + 0.0016811243258416653, + 0.0014336210442706943, + 0.001731115742586553, + 0.003693845821544528, + 0.0014389881398528814, + 0.0020519383251667023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.02119230106472969 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle, almost level with the sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame facing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the larger crab raising its right claw high while the smaller crab lowers its body and moves its feet sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a beach at sunset with warm orange-red lighting and a mirror-like sheen on the moist sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307584464550018 + ], + "opposite_logprob_tokens": [ + -3.1680757999420166 + ], + "logprob_reward": -0.04307584464550018, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307584464550018, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4303220510482788, + 0.13000844419002533, + 0.16890312731266022, + 0.016788188368082047, + 0.018413176760077477, + 0.01665544882416725, + 0.028284767642617226, + 0.12721216678619385, + 0.02811732143163681, + 0.035295307636260986 + ], + "credit_map": [ + 0.14615482091903687, + 0.16755646467208862, + 0.012895960360765457, + 0.013449328020215034, + 0.017524076625704765, + 0.03818810358643532, + 0.010624815709888935, + 0.023928478360176086, + 0.01842300221323967, + 0.01816132850944996, + 0.012516074813902378, + 0.011534926481544971, + 0.016610050573945045, + 0.029042523354291916, + 0.010515835136175156, + 0.013204703107476234, + 0.060930490493774414, + 0.0680520087480545, + 0.00605118740350008, + 0.005424641538411379, + 0.008054184727370739, + 0.005780811421573162, + 0.00556481396779418, + 0.009045003913342953, + 0.002533762948587537, + 0.002457972848787904, + 0.0025764890015125275, + 0.001603789278306067, + 0.001962409121915698, + 0.0015944044571369886, + 0.002026159316301346, + 0.002033201279118657, + 0.002593323588371277, + 0.003991494420915842, + 0.0015283480752259493, + 0.0016014515422284603, + 0.0030035446397960186, + 0.001927515142597258, + 0.001991942524909973, + 0.0017755557782948017, + 0.0034473391715437174, + 0.003635614411905408, + 0.0018804507562890649, + 0.0012946665519848466, + 0.0019317495170980692, + 0.002131149871274829, + 0.0010942956432700157, + 0.0012401826679706573, + 0.005110256839543581, + 0.003584605176001787, + 0.002660599770024419, + 0.0021229649428278208, + 0.0030675570014864206, + 0.004093692172318697, + 0.0025932956486940384, + 0.005051798652857542, + 0.024773376062512398, + 0.020487673580646515, + 0.010007448494434357, + 0.010708917863667011, + 0.012050404213368893, + 0.023511918261647224, + 0.009707081131637096, + 0.01596534252166748, + 0.0038195496890693903, + 0.00508996844291687, + 0.0030775521881878376, + 0.0024058574344962835, + 0.003614584216848016, + 0.003642830066382885, + 0.0027727102860808372, + 0.0036942684091627598, + 0.0050379009917378426, + 0.0038096937350928783, + 0.0030714846216142178, + 0.0031146889086812735, + 0.006737010087817907, + 0.004338684491813183, + 0.004398049786686897, + 0.004787791986018419 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.04307584464550018 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle, almost level with the sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame facing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the larger crab raising its right claw high while the smaller crab lowers its body and moves its feet sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a beach at sunset with warm orange-red lighting and a mirror-like sheen on the moist sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498193949460983 + ], + "opposite_logprob_tokens": [ + -2.9299819469451904 + ], + "logprob_reward": -0.05498193949460983, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498193949460983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5691202282905579, + 0.08002056926488876, + 0.06112798675894737, + 0.06038834899663925, + 0.018132416531443596, + 0.02579827979207039, + 0.04312756285071373, + 0.08198555558919907, + 0.03495907783508301, + 0.025340022519230843 + ], + "credit_map": [ + 0.20549261569976807, + 0.23733903467655182, + 0.011331616900861263, + 0.010483551770448685, + 0.021424012258648872, + 0.04142959043383598, + 0.012350402772426605, + 0.029269378632307053, + 0.014670655131340027, + 0.010355657897889614, + 0.007590924855321646, + 0.0071950932033360004, + 0.008913137018680573, + 0.01741654798388481, + 0.006586629431694746, + 0.007291923277080059, + 0.012987004593014717, + 0.010620053857564926, + 0.0065794154070317745, + 0.006826477125287056, + 0.004905215930193663, + 0.005875743459910154, + 0.0064164334908127785, + 0.0069176433607935905, + 0.008594617247581482, + 0.006965635344386101, + 0.008297569118440151, + 0.008404248394072056, + 0.006533225066959858, + 0.008397560566663742, + 0.005517410580068827, + 0.007678082212805748, + 0.0034376138355582952, + 0.002774850931018591, + 0.0027212342247366905, + 0.0014117504470050335, + 0.00195395783521235, + 0.0021328572183847427, + 0.0018873118096962571, + 0.0018128394149243832, + 0.007506072521209717, + 0.0069745988585054874, + 0.0017125671729445457, + 0.0012435578973963857, + 0.002414689864963293, + 0.0027008778415620327, + 0.0012940631713718176, + 0.0019518533954396844, + 0.009719560854136944, + 0.0048852283507585526, + 0.003575288923457265, + 0.002937424462288618, + 0.004988725762814283, + 0.006163199432194233, + 0.004002450034022331, + 0.006855687126517296, + 0.01488699484616518, + 0.013543529435992241, + 0.00763607956469059, + 0.007349138148128986, + 0.008158604614436626, + 0.014003800228238106, + 0.006343227345496416, + 0.010064180009067059, + 0.0043376474641263485, + 0.004701630678027868, + 0.0038340771570801735, + 0.0030923376325517893, + 0.005792467854917049, + 0.003702213056385517, + 0.004209409933537245, + 0.005289292894303799, + 0.0038556752260774374, + 0.003313934663310647, + 0.003648943267762661, + 0.0017380635254085064, + 0.0034735682420432568, + 0.004307307302951813, + 0.002309873467311263, + 0.00269265822134912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.05498193949460983 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle, almost level with the sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame facing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the larger crab raising its right claw high while the smaller crab lowers its body and moves its feet sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a beach at sunset with warm orange-red lighting and a mirror-like sheen on the moist sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001042775809764862 + ], + "opposite_logprob_tokens": [ + -6.87604284286499 + ], + "logprob_reward": -0.001042775809764862, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.001042775809764862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6537379026412964, + 0.0580294206738472, + 0.06768888235092163, + 0.018022721633315086, + 0.024469995871186256, + 0.03356681019067764, + 0.019759254530072212, + 0.0405903086066246, + 0.036537133157253265, + 0.04759756475687027 + ], + "credit_map": [ + 0.22902444005012512, + 0.26926329731941223, + 0.01567212864756584, + 0.01324274018406868, + 0.025516172870993614, + 0.05049064755439758, + 0.016016410663723946, + 0.03451211377978325, + 0.009736577980220318, + 0.007581455633044243, + 0.005683596711605787, + 0.0045562442392110825, + 0.008330421522259712, + 0.011409047059714794, + 0.0047389366663992405, + 0.005993148311972618, + 0.012944008223712444, + 0.00880693644285202, + 0.007663886994123459, + 0.008052646182477474, + 0.005711808335036039, + 0.007110379636287689, + 0.008468253538012505, + 0.008930967189371586, + 0.0029308898374438286, + 0.0025467900559306145, + 0.002603522501885891, + 0.0019994911272078753, + 0.002241814509034157, + 0.0019177242647856474, + 0.0014303501229733229, + 0.0023521408438682556, + 0.004837206564843655, + 0.006008436903357506, + 0.0022013678681105375, + 0.0018362612463533878, + 0.003478727536275983, + 0.0024262885563075542, + 0.0016516192117705941, + 0.0020300918258726597, + 0.01175216119736433, + 0.010748724453151226, + 0.0010761801386252046, + 0.0010649716714397073, + 0.002704653888940811, + 0.0030818728264421225, + 0.0012677459744736552, + 0.0018705015536397696, + 0.0035093931946903467, + 0.003954590298235416, + 0.0017007881542667747, + 0.0012265684781596065, + 0.00200932240113616, + 0.003149644238874316, + 0.0014873665058985353, + 0.002721583703532815, + 0.007806367240846157, + 0.006336253136396408, + 0.003262533340603113, + 0.0034646065905690193, + 0.00401603477075696, + 0.007212105207145214, + 0.0032082502730190754, + 0.005284161306917667, + 0.005964743439108133, + 0.006710283923894167, + 0.003180783474817872, + 0.0023937064688652754, + 0.006177723407745361, + 0.004893846344202757, + 0.002948411973193288, + 0.004267639014869928, + 0.010823508724570274, + 0.011084409430623055, + 0.0030403495766222477, + 0.002514350228011608, + 0.004716349765658379, + 0.007639599032700062, + 0.0031710623297840357, + 0.004607942420989275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.001042775809764862 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle, almost level with the sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame facing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the larger crab raising its right claw high while the smaller crab lowers its body and moves its feet sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a beach at sunset with warm orange-red lighting and a mirror-like sheen on the moist sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34866103529930115 + ], + "opposite_logprob_tokens": [ + -1.2236610651016235 + ], + "logprob_reward": -0.34866103529930115, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34866103529930115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36928343772888184, + 0.19331522285938263, + 0.18430952727794647, + 0.06311407685279846, + 0.02838830277323723, + 0.0185210183262825, + 0.022733615711331367, + 0.09565865993499756, + 0.01565231941640377, + 0.00902380608022213 + ], + "credit_map": [ + 0.12205285578966141, + 0.14477862417697906, + 0.011020690202713013, + 0.009728998877108097, + 0.016001861542463303, + 0.033749744296073914, + 0.011597733944654465, + 0.02035292610526085, + 0.03516485169529915, + 0.024580394849181175, + 0.016228217631578445, + 0.014684637077152729, + 0.02434113807976246, + 0.04542400315403938, + 0.014423605985939503, + 0.018468361347913742, + 0.07093904912471771, + 0.07621172815561295, + 0.005177047103643417, + 0.0054223486222326756, + 0.00798983033746481, + 0.005609766114503145, + 0.004081331193447113, + 0.008878414519131184, + 0.00964810699224472, + 0.0073838792741298676, + 0.007628521881997585, + 0.008440888486802578, + 0.007826603949069977, + 0.007797397207468748, + 0.006501888390630484, + 0.007886787876486778, + 0.006871575955301523, + 0.006460864562541246, + 0.0029867244884371758, + 0.0021740475203841925, + 0.0033118962310254574, + 0.0021000136621296406, + 0.002214876702055335, + 0.002268304117023945, + 0.0024590715765953064, + 0.0030152613762766123, + 0.0025544429663568735, + 0.0013534757308661938, + 0.002065518870949745, + 0.0036645398940891027, + 0.0014480478130280972, + 0.001960660330951214, + 0.004960392601788044, + 0.0026201263535767794, + 0.002237842883914709, + 0.001961305970326066, + 0.0019146996783092618, + 0.00299434713087976, + 0.00206271861679852, + 0.003982183523476124, + 0.018811440095305443, + 0.015634048730134964, + 0.007573687471449375, + 0.008550447411835194, + 0.00893139373511076, + 0.017017558217048645, + 0.0070373970083892345, + 0.01210269145667553, + 0.0024042646400630474, + 0.0027876091189682484, + 0.001336070359684527, + 0.0013733895029872656, + 0.0020201231818646193, + 0.0024693922605365515, + 0.0015099133597686887, + 0.0017515571089461446, + 0.0013642477570101619, + 0.0011119480477645993, + 0.001159795792773366, + 0.0009320819517597556, + 0.0008942859130911529, + 0.0013775428524240851, + 0.0010438598692417145, + 0.0011400446528568864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.34866103529930115 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from an extremely low angle, almost level with the sand, makes the two crabs facing each other appear incredibly large and intimidating. The last rays of sunset bathe the beach in a warm orange-red hue, and the moist sand reflects the vibrant sunset sky, giving it a mirror-like sheen. In the center of the frame, a crab with a deep russet-brown shell, its surface rough and uneven, covered in tiny dents and the scratches of time, raises its disproportionately large right claw high, the serrations on its inner side gleaming coldly in the golden light, striking a menacing pose. A few inches away, another, smaller but equally aggressive crab, its shell a darker gray, lowers its body, opening and closing its two similarly sized claws on its chest. Cautiously moving its feet sideways, stirring up the thin layer of seawater beneath its feet. The camera remains still, steadily capturing this silent duel. The low-angle sunlight outlines their contours with a bright golden rim of light and casts extremely long, dancing shadows behind them. In the background, the distant waves, dyed a mixture of purple and orange, gently and rhythmically roll towards the shore, forming a sharp contrast between movement and stillness, tranquility and tension with the fierce confrontation on the verge of breaking out in the foreground. The entire scene is filled with the dramatic tension of primal vitality.", + "seed": 8605, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_15/index_8393_seed_8605_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.10152855643536896, + "vlm_logprob": -0.10152855643536896, + "vlm_yes_no_margin": 3.450000011920929, + "vlm_token_credit": -0.10152855643536896, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle, almost level with the sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame facing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the larger crab raising its right claw high while the smaller crab lowers its body and moves its feet sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a beach at sunset with warm orange-red lighting and a mirror-like sheen on the moist sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05549846589565277 + ], + "opposite_logprob_tokens": [ + -3.1804983615875244 + ], + "logprob_reward": -0.05549846589565277, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.05549846589565277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2452610433101654, + 0.12532258033752441, + 0.06889239698648453, + 0.1115778461098671, + 0.08134647458791733, + 0.029992088675498962, + 0.08657503873109818, + 0.1213584914803505, + 0.0829969197511673, + 0.04667714238166809 + ], + "credit_map": [ + 0.035191360861063004, + 0.02751273475587368, + 0.05406368151307106, + 0.03445176035165787, + 0.025132302194833755, + 0.03140675649046898, + 0.021581757813692093, + 0.01592068560421467, + 0.015268946997821331, + 0.010014987550675869, + 0.02186177298426628, + 0.01329901535063982, + 0.01729169487953186, + 0.02067413739860058, + 0.016429275274276733, + 0.010482742451131344, + 0.010142057202756405, + 0.013632971793413162, + 0.006805483717471361, + 0.007063654717057943, + 0.009542289189994335, + 0.00902027077972889, + 0.007608179003000259, + 0.005077489651739597, + 0.035335443913936615, + 0.03634319081902504, + 0.005041056778281927, + 0.0049454765394330025, + 0.009919948875904083, + 0.007883518002927303, + 0.006327132228761911, + 0.005782082676887512, + 0.011465372517704964, + 0.010104117915034294, + 0.006675128825008869, + 0.007182533387094736, + 0.015743480995297432, + 0.01485836785286665, + 0.007953070104122162, + 0.007364403456449509, + 0.004456513561308384, + 0.005073042586445808, + 0.0032890690490603447, + 0.003776273224502802, + 0.004468491300940514, + 0.0025227698497474194, + 0.0042092278599739075, + 0.002196701942011714, + 0.011649071238934994, + 0.01456504501402378, + 0.00471313064917922, + 0.004890277981758118, + 0.017181815579533577, + 0.011175740510225296, + 0.009546041488647461, + 0.012853924185037613, + 0.020132003352046013, + 0.027890874072909355, + 0.011873931623995304, + 0.009251692332327366, + 0.01517089456319809, + 0.01590520702302456, + 0.009785613045096397, + 0.011348272673785686, + 0.010948960669338703, + 0.01448392029851675, + 0.00934649258852005, + 0.009217451326549053, + 0.013197261840105057, + 0.010493017733097076, + 0.006281672511249781, + 0.00902814231812954, + 0.005891718901693821, + 0.004922463558614254, + 0.005492687691003084, + 0.005288651213049889, + 0.006707330234348774, + 0.005333335604518652, + 0.005551191046833992, + 0.0074897650629282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.05549846589565277 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle, almost level with the sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame facing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the larger crab raising its right claw high while the smaller crab lowers its body and moves its feet sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a beach at sunset with warm orange-red lighting and a mirror-like sheen on the moist sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033716194331645966 + ], + "opposite_logprob_tokens": [ + -3.4087162017822266 + ], + "logprob_reward": -0.033716194331645966, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033716194331645966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11721745878458023, + 0.06476016342639923, + 0.04204590246081352, + 0.22212405502796173, + 0.0673043504357338, + 0.10866879671812057, + 0.07753531634807587, + 0.15695738792419434, + 0.11065658181905746, + 0.03272995725274086 + ], + "credit_map": [ + 0.016391199082136154, + 0.015723086893558502, + 0.015451773069798946, + 0.010328240692615509, + 0.011751734651625156, + 0.018608683720231056, + 0.01586412638425827, + 0.013098623603582382, + 0.009750371798872948, + 0.005885052960366011, + 0.006139188073575497, + 0.0049715726636350155, + 0.01216629333794117, + 0.011929947882890701, + 0.008411668241024017, + 0.005506075453013182, + 0.00755554623901844, + 0.003656454384326935, + 0.004863715264946222, + 0.003742250381037593, + 0.006719948723912239, + 0.005936929024755955, + 0.004830934107303619, + 0.004740129224956036, + 0.07224185764789581, + 0.08192966878414154, + 0.016903793439269066, + 0.009285414591431618, + 0.011122167110443115, + 0.012234240770339966, + 0.00890030711889267, + 0.00950661487877369, + 0.011387518607079983, + 0.01331120915710926, + 0.003897935850545764, + 0.004451544489711523, + 0.01245383732020855, + 0.009767984040081501, + 0.006751087959855795, + 0.005283233243972063, + 0.025195419788360596, + 0.030429759994149208, + 0.007414937485009432, + 0.007859175093472004, + 0.012496696785092354, + 0.007641470991075039, + 0.009981170296669006, + 0.007650173734873533, + 0.012720273807644844, + 0.009896059520542622, + 0.009209327399730682, + 0.00723005598410964, + 0.009137636050581932, + 0.008933342061936855, + 0.010183067992329597, + 0.010225556790828705, + 0.02669306844472885, + 0.024859679862856865, + 0.0198995228856802, + 0.013495009392499924, + 0.01857594959437847, + 0.016910837963223457, + 0.016065144911408424, + 0.02045818231999874, + 0.014800109900534153, + 0.015445555560290813, + 0.01743225008249283, + 0.012296873144805431, + 0.01763278804719448, + 0.0120978569611907, + 0.00988460797816515, + 0.011066547594964504, + 0.004575775470584631, + 0.003712669014930725, + 0.004012905061244965, + 0.004901150707155466, + 0.00402031559497118, + 0.003862066427245736, + 0.0032999804243445396, + 0.004345097579061985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.033716194331645966 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle, almost level with the sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame facing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the larger crab raising its right claw high while the smaller crab lowers its body and moves its feet sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a beach at sunset with warm orange-red lighting and a mirror-like sheen on the moist sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981741353869438 + ], + "opposite_logprob_tokens": [ + -3.5298173427581787 + ], + "logprob_reward": -0.02981741353869438, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981741353869438, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23737740516662598, + 0.0867428183555603, + 0.053387124091386795, + 0.05594554543495178, + 0.09513906389474869, + 0.1606297343969345, + 0.08872518688440323, + 0.05851205438375473, + 0.1150725930929184, + 0.0484684593975544 + ], + "credit_map": [ + 0.030504079535603523, + 0.021576683968305588, + 0.041057225316762924, + 0.02615242637693882, + 0.026321250945329666, + 0.026897471398115158, + 0.042829480022192, + 0.0220387764275074, + 0.016276996582746506, + 0.008779045194387436, + 0.010077586397528648, + 0.008320354856550694, + 0.01491989754140377, + 0.008936808444559574, + 0.00940093956887722, + 0.010031189769506454, + 0.008519209921360016, + 0.006288565695285797, + 0.006752362474799156, + 0.008325763046741486, + 0.007207208313047886, + 0.006410121452063322, + 0.005843892227858305, + 0.0040400042198598385, + 0.0058606774546206, + 0.006908718030899763, + 0.006671014707535505, + 0.00704251928254962, + 0.008873536251485348, + 0.009419321082532406, + 0.0050508067943155766, + 0.006118953227996826, + 0.020530259236693382, + 0.02538691833615303, + 0.008093258365988731, + 0.005521282088011503, + 0.011044437065720558, + 0.009745308198034763, + 0.008226364850997925, + 0.006591237615793943, + 0.04010700061917305, + 0.045556217432022095, + 0.01373487152159214, + 0.008892948739230633, + 0.019728416576981544, + 0.011274025775492191, + 0.011699622496962547, + 0.00963662937283516, + 0.014893297106027603, + 0.01667221076786518, + 0.006686960346996784, + 0.007742894813418388, + 0.013941205106675625, + 0.010000701062381268, + 0.010343526490032673, + 0.008444386534392834, + 0.011330909095704556, + 0.012933628633618355, + 0.005039641633629799, + 0.0035601367708295584, + 0.007109480444341898, + 0.008511512540280819, + 0.006258976645767689, + 0.0037677690852433443, + 0.01812429167330265, + 0.011139631271362305, + 0.01931801624596119, + 0.0176134891808033, + 0.014311929233372211, + 0.009112916886806488, + 0.014601144008338451, + 0.010851175524294376, + 0.006999886129051447, + 0.004252393264323473, + 0.006543030496686697, + 0.0038231092039495707, + 0.009232095442712307, + 0.005742380861192942, + 0.004812789149582386, + 0.007062772288918495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.02981741353869438 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle, almost level with the sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame facing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the larger crab raising its right claw high while the smaller crab lowers its body and moves its feet sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a beach at sunset with warm orange-red lighting and a mirror-like sheen on the moist sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015126940561458468 + ], + "opposite_logprob_tokens": [ + -6.50151252746582 + ], + "logprob_reward": -0.0015126940561458468, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015126940561458468, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08139064162969589, + 0.12163323909044266, + 0.09342944622039795, + 0.1099972352385521, + 0.10183132439851761, + 0.07781587541103363, + 0.10925912111997604, + 0.10902294516563416, + 0.04191216826438904, + 0.15370801091194153 + ], + "credit_map": [ + 0.013029607944190502, + 0.013353894464671612, + 0.008372348733246326, + 0.008110184222459793, + 0.009545711800456047, + 0.01024078018963337, + 0.00748074147850275, + 0.011257367208600044, + 0.016079304739832878, + 0.010662277229130268, + 0.019900977611541748, + 0.015623839572072029, + 0.014850173145532608, + 0.022278402000665665, + 0.011893042363226414, + 0.010345228016376495, + 0.012838258408010006, + 0.02144421450793743, + 0.004969964735209942, + 0.00964453537017107, + 0.011362332850694656, + 0.022950943559408188, + 0.004516824148595333, + 0.005702367052435875, + 0.03447256609797478, + 0.042360980063676834, + 0.00603818241506815, + 0.004433118272572756, + 0.0067145018838346004, + 0.007307338062673807, + 0.004166009370237589, + 0.0045045423321425915, + 0.01859557442367077, + 0.010373982600867748, + 0.009147659875452518, + 0.008995277807116508, + 0.016816703602671623, + 0.01713811792433262, + 0.01017943024635315, + 0.010584576986730099, + 0.023169150575995445, + 0.025392362847924232, + 0.004944671411067247, + 0.003095748135820031, + 0.006665666587650776, + 0.005862566642463207, + 0.004883542191237211, + 0.0038021663203835487, + 0.016228029504418373, + 0.026117131114006042, + 0.006545898038893938, + 0.007010206114500761, + 0.023290101438760757, + 0.011393001303076744, + 0.009867208078503609, + 0.008807547390460968, + 0.026899905875325203, + 0.03356990963220596, + 0.0069644395262002945, + 0.0043789539486169815, + 0.015350914560258389, + 0.008164897561073303, + 0.008208776824176311, + 0.005485144909471273, + 0.005380598828196526, + 0.0072836969047784805, + 0.006732277572154999, + 0.003082445589825511, + 0.007631305605173111, + 0.0052710329182446, + 0.0031391214579343796, + 0.003391687525436282, + 0.019876308739185333, + 0.019741227850317955, + 0.008856352418661118, + 0.009360330179333687, + 0.04336392506957054, + 0.02034446783363819, + 0.015655655413866043, + 0.016509734094142914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 3 + ], + "selected_reward": -0.0015126940561458468 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle, almost level with the sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame facing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the larger crab raising its right claw high while the smaller crab lowers its body and moves its feet sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a beach at sunset with warm orange-red lighting and a mirror-like sheen on the moist sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870980143547058 + ], + "opposite_logprob_tokens": [ + -1.1370980739593506 + ], + "logprob_reward": -0.3870980143547058, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.3870980143547058, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21154406666755676, + 0.09725334495306015, + 0.08834800124168396, + 0.16461656987667084, + 0.0632186159491539, + 0.15020732581615448, + 0.08365170657634735, + 0.03528890386223793, + 0.07195264846086502, + 0.033918797969818115 + ], + "credit_map": [ + 0.0333893820643425, + 0.021751465275883675, + 0.042535584419965744, + 0.02298574708402157, + 0.018432985991239548, + 0.022430356591939926, + 0.03328298404812813, + 0.016735555604100227, + 0.01265967171639204, + 0.008820545859634876, + 0.01667257957160473, + 0.009639844298362732, + 0.011987795121967793, + 0.012690804898738861, + 0.015266463160514832, + 0.009515639394521713, + 0.014485709369182587, + 0.010589392855763435, + 0.013027445413172245, + 0.009853401221334934, + 0.013359662145376205, + 0.009444327093660831, + 0.00964475329965353, + 0.007943308912217617, + 0.048755668103694916, + 0.053446732461452484, + 0.009518181905150414, + 0.008570118807256222, + 0.014214511029422283, + 0.014671685174107552, + 0.008059603162109852, + 0.007380076218396425, + 0.009859149344265461, + 0.00984981283545494, + 0.005084151402115822, + 0.003317863680422306, + 0.013160745613276958, + 0.010590366087853909, + 0.005526753608137369, + 0.005829777102917433, + 0.039187293499708176, + 0.045067623257637024, + 0.008884359151124954, + 0.010373027995228767, + 0.013980022631585598, + 0.008742603473365307, + 0.012114587239921093, + 0.011857802048325539, + 0.011944522149860859, + 0.015980690717697144, + 0.005363612901419401, + 0.006167884450405836, + 0.013154881075024605, + 0.010893416590988636, + 0.008676606230437756, + 0.011470089666545391, + 0.0045404876582324505, + 0.006206698250025511, + 0.0036479411646723747, + 0.0031190593726933002, + 0.004238072317093611, + 0.005629631225019693, + 0.003472778480499983, + 0.004434238653630018, + 0.007880284450948238, + 0.011977804824709892, + 0.007679485250264406, + 0.007223089225590229, + 0.011980989947915077, + 0.01032287161797285, + 0.005823033396154642, + 0.009065086953341961, + 0.0036547877825796604, + 0.003798137651756406, + 0.0054963803850114346, + 0.003935416229069233, + 0.00498433830216527, + 0.004123418126255274, + 0.003995880950242281, + 0.003930439241230488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.3870980143547058 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from an extremely low angle, almost level with the sand, makes the two crabs facing each other appear incredibly large and intimidating. The last rays of sunset bathe the beach in a warm orange-red hue, and the moist sand reflects the vibrant sunset sky, giving it a mirror-like sheen. In the center of the frame, a crab with a deep russet-brown shell, its surface rough and uneven, covered in tiny dents and the scratches of time, raises its disproportionately large right claw high, the serrations on its inner side gleaming coldly in the golden light, striking a menacing pose. A few inches away, another, smaller but equally aggressive crab, its shell a darker gray, lowers its body, opening and closing its two similarly sized claws on its chest. Cautiously moving its feet sideways, stirring up the thin layer of seawater beneath its feet. The camera remains still, steadily capturing this silent duel. The low-angle sunlight outlines their contours with a bright golden rim of light and casts extremely long, dancing shadows behind them. In the background, the distant waves, dyed a mixture of purple and orange, gently and rhythmically roll towards the shore, forming a sharp contrast between movement and stillness, tranquility and tension with the fierce confrontation on the verge of breaking out in the foreground. The entire scene is filled with the dramatic tension of primal vitality.", + "seed": 8606, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_15/index_8393_seed_8606_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.15227307810564525, + "vlm_logprob": -0.15227307810564525, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.15227307810564525, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle, almost level with the sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame facing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the larger crab raising its right claw high while the smaller crab lowers its body and moves its feet sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a beach at sunset with warm orange-red lighting and a mirror-like sheen on the moist sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028586920350790024 + ], + "opposite_logprob_tokens": [ + -4.1535868644714355 + ], + "logprob_reward": -0.028586920350790024, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.028586920350790024, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07393720000982285, + 0.05949382856488228, + 0.03979482501745224, + 0.1589096337556839, + 0.17374524474143982, + 0.18298961222171783, + 0.07997118681669235, + 0.05095191299915314, + 0.05118202790617943, + 0.12902456521987915 + ], + "credit_map": [ + 0.010390648618340492, + 0.010669853538274765, + 0.006261167582124472, + 0.008765297941863537, + 0.010748704895377159, + 0.013646788895130157, + 0.005571951624006033, + 0.007882785983383656, + 0.008610231801867485, + 0.00463970797136426, + 0.005154991522431374, + 0.006843930575996637, + 0.009546855464577675, + 0.011086833663284779, + 0.006491332780569792, + 0.007119945716112852, + 0.0047132293693721294, + 0.004068468231707811, + 0.005185316316783428, + 0.005268682260066271, + 0.004781941883265972, + 0.004910753108561039, + 0.005602971184998751, + 0.005263460334390402, + 0.0485764779150486, + 0.007612250279635191, + 0.005517909768968821, + 0.005544770043343306, + 0.014065101742744446, + 0.06324001401662827, + 0.006658228114247322, + 0.007694872561842203, + 0.05166994780302048, + 0.008220411837100983, + 0.008384931832551956, + 0.00825599767267704, + 0.012813221663236618, + 0.06702952086925507, + 0.008212328888475895, + 0.009158887900412083, + 0.022551629692316055, + 0.015656255185604095, + 0.025216903537511826, + 0.018509266898036003, + 0.016062550246715546, + 0.02652795985341072, + 0.026089981198310852, + 0.032375071197748184, + 0.01644473895430565, + 0.008115645498037338, + 0.008094055578112602, + 0.007781029678881168, + 0.009041376411914825, + 0.015647560358047485, + 0.007697119377553463, + 0.007149658631533384, + 0.008240931667387486, + 0.005146816372871399, + 0.004202104639261961, + 0.0047129495069384575, + 0.008523277938365936, + 0.008968052454292774, + 0.00568603677675128, + 0.005471745505928993, + 0.006158705335110426, + 0.0046439520083367825, + 0.004780454095453024, + 0.004491564352065325, + 0.010066768154501915, + 0.00773076293990016, + 0.007271150127053261, + 0.00603867182508111, + 0.010150830261409283, + 0.009722317568957806, + 0.009638235904276371, + 0.009371529333293438, + 0.03669196739792824, + 0.030442051589488983, + 0.011117218993604183, + 0.011890420690178871 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.028586920350790024 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle, almost level with the sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame facing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the larger crab raising its right claw high while the smaller crab lowers its body and moves its feet sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a beach at sunset with warm orange-red lighting and a mirror-like sheen on the moist sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012553129345178604 + ], + "opposite_logprob_tokens": [ + -4.3875532150268555 + ], + "logprob_reward": -0.012553129345178604, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012553129345178604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045331042259931564, + 0.08212926238775253, + 0.031218726187944412, + 0.06202414259314537, + 0.15043815970420837, + 0.293878972530365, + 0.05498076602816582, + 0.15041111409664154, + 0.056776296347379684, + 0.07281152904033661 + ], + "credit_map": [ + 0.005461842752993107, + 0.005861333571374416, + 0.005776500795036554, + 0.005240811500698328, + 0.006745843216776848, + 0.008317501284182072, + 0.004135189112275839, + 0.003792017698287964, + 0.010799654759466648, + 0.007652118802070618, + 0.007878270000219345, + 0.011990810744464397, + 0.01182401180267334, + 0.013436667621135712, + 0.008623672649264336, + 0.009924056939780712, + 0.00400011008605361, + 0.004376774188131094, + 0.003705281298607588, + 0.0029156880918890238, + 0.0031818244606256485, + 0.004644536878913641, + 0.0034466052893549204, + 0.0049479082226753235, + 0.01648530550301075, + 0.004514776170253754, + 0.004039029590785503, + 0.003218377474695444, + 0.005886546801775694, + 0.02092866599559784, + 0.003084569238126278, + 0.003866871353238821, + 0.04106222093105316, + 0.006797934882342815, + 0.008657783269882202, + 0.008837046101689339, + 0.01093411073088646, + 0.05295894294977188, + 0.010884595103561878, + 0.010305533185601234, + 0.047406453639268875, + 0.02020139992237091, + 0.038253482431173325, + 0.02814750373363495, + 0.022898005321621895, + 0.06006397679448128, + 0.03532438725233078, + 0.041583746671676636, + 0.010697519406676292, + 0.004428948275744915, + 0.004127169027924538, + 0.004001851659268141, + 0.005050727631896734, + 0.015612880699336529, + 0.004998545162379742, + 0.0060631223022937775, + 0.016621273010969162, + 0.012949731200933456, + 0.014119701460003853, + 0.012840015813708305, + 0.023350542411208153, + 0.0180527213960886, + 0.025376150384545326, + 0.027100959792733192, + 0.007255536504089832, + 0.0051049599424004555, + 0.010862398892641068, + 0.005195357836782932, + 0.006924792192876339, + 0.007144255563616753, + 0.007483385037630796, + 0.006805608980357647, + 0.007157165557146072, + 0.00474622892215848, + 0.005057995207607746, + 0.005664918106049299, + 0.017209580168128014, + 0.011445675976574421, + 0.014142824336886406, + 0.007387143559753895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 7 + ], + "selected_reward": -0.012553129345178604 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle, almost level with the sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame facing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the larger crab raising its right claw high while the smaller crab lowers its body and moves its feet sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a beach at sunset with warm orange-red lighting and a mirror-like sheen on the moist sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026366515085101128 + ], + "opposite_logprob_tokens": [ + -3.6513664722442627 + ], + "logprob_reward": -0.026366515085101128, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026366515085101128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04458104819059372, + 0.04127625375986099, + 0.05084456503391266, + 0.04114888235926628, + 0.41993966698646545, + 0.13363724946975708, + 0.06823725253343582, + 0.08386608213186264, + 0.055464405566453934, + 0.061004575341939926 + ], + "credit_map": [ + 0.006664346903562546, + 0.005118614062666893, + 0.004756396636366844, + 0.005335170775651932, + 0.007307325955480337, + 0.006525518372654915, + 0.004840944893658161, + 0.004032731521874666, + 0.00505570275709033, + 0.003978192340582609, + 0.004242989234626293, + 0.00505241472274065, + 0.006353728473186493, + 0.006288380827754736, + 0.00524807907640934, + 0.005056765396147966, + 0.006722685880959034, + 0.005814212374389172, + 0.006597207393497229, + 0.005632131360471249, + 0.005154974292963743, + 0.00849002692848444, + 0.006135490257292986, + 0.006297841668128967, + 0.0052161975763738155, + 0.003451649332419038, + 0.004980810452252626, + 0.004253970924764872, + 0.006870484445244074, + 0.006641788873821497, + 0.005274908151477575, + 0.004459073767066002, + 0.1401066780090332, + 0.017234409227967262, + 0.013734030537307262, + 0.013240928761661053, + 0.0266532264649868, + 0.17821626365184784, + 0.013178319670259953, + 0.017575817182660103, + 0.03309974446892738, + 0.006535406690090895, + 0.00920266192406416, + 0.007377877831459045, + 0.013393237255513668, + 0.04325392097234726, + 0.009116437286138535, + 0.011657982133328915, + 0.01002206839621067, + 0.006334410049021244, + 0.006397129036486149, + 0.006961714942008257, + 0.010171676054596901, + 0.013353313319385052, + 0.008206305094063282, + 0.006790641229599714, + 0.012979993596673012, + 0.007934602908790112, + 0.0068991477601230145, + 0.006565704010426998, + 0.012890157289803028, + 0.01546085812151432, + 0.01056838408112526, + 0.01056724600493908, + 0.006906223949044943, + 0.004794721025973558, + 0.005190389696508646, + 0.0039016513619571924, + 0.009992680512368679, + 0.011375341564416885, + 0.0077429567463696, + 0.005560445599257946, + 0.004884849302470684, + 0.00953215453773737, + 0.0041597988456487656, + 0.003592750057578087, + 0.007608904503285885, + 0.006444158963859081, + 0.016435427591204643, + 0.008346536196768284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -0.026366515085101128 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle, almost level with the sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame facing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the larger crab raising its right claw high while the smaller crab lowers its body and moves its feet sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a beach at sunset with warm orange-red lighting and a mirror-like sheen on the moist sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00043764073052443564 + ], + "opposite_logprob_tokens": [ + -7.7504377365112305 + ], + "logprob_reward": -0.00043764073052443564, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.00043764073052443564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.138625368475914, + 0.050027765333652496, + 0.06892012804746628, + 0.041316259652376175, + 0.23241455852985382, + 0.0685780867934227, + 0.06019085645675659, + 0.05006806179881096, + 0.0626913532614708, + 0.2271675318479538 + ], + "credit_map": [ + 0.016519727185368538, + 0.012734833173453808, + 0.02336275391280651, + 0.018047943711280823, + 0.019647710025310516, + 0.016188297420740128, + 0.015521829016506672, + 0.01660226844251156, + 0.006311080884188414, + 0.005162987392395735, + 0.00582821574062109, + 0.009341193363070488, + 0.005685212556272745, + 0.006781612057238817, + 0.006104516796767712, + 0.004812940955162048, + 0.01307467557489872, + 0.00516478018835187, + 0.005245184991508722, + 0.004452865105122328, + 0.009469174779951572, + 0.01575540378689766, + 0.006264047231525183, + 0.0094939935952425, + 0.006570539437234402, + 0.004367354791611433, + 0.004868823569267988, + 0.004820774309337139, + 0.0067214504815638065, + 0.0056830961257219315, + 0.0043326495215296745, + 0.003951569087803364, + 0.07914582639932632, + 0.009331222623586655, + 0.007941198535263538, + 0.006901040207594633, + 0.013946417719125748, + 0.099297434091568, + 0.007219869643449783, + 0.008631540462374687, + 0.019121669232845306, + 0.004997311159968376, + 0.003836775664240122, + 0.0029682053718715906, + 0.00593873206526041, + 0.022681301459670067, + 0.004305160138756037, + 0.004728930536657572, + 0.01316170021891594, + 0.005037580616772175, + 0.005078232381492853, + 0.003245383035391569, + 0.005834669340401888, + 0.015154479071497917, + 0.006259373389184475, + 0.006419442128390074, + 0.007294856011867523, + 0.004725514445453882, + 0.005061148200184107, + 0.003942304756492376, + 0.00760610168799758, + 0.007314856164157391, + 0.005758171901106834, + 0.008365110494196415, + 0.008146791718900204, + 0.005454353988170624, + 0.006239642854779959, + 0.004111352376639843, + 0.016540242359042168, + 0.010157049633562565, + 0.006069777067750692, + 0.005972142796963453, + 0.016915999352931976, + 0.017446408048272133, + 0.015503128990530968, + 0.015367943793535233, + 0.02682151272892952, + 0.0241941437125206, + 0.08095086365938187, + 0.029967525973916054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.00043764073052443564 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle, almost level with the sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame facing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the larger crab raising its right claw high while the smaller crab lowers its body and moves its feet sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a beach at sunset with warm orange-red lighting and a mirror-like sheen on the moist sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6934211850166321 + ], + "opposite_logprob_tokens": [ + -0.6934211850166321 + ], + "logprob_reward": -0.6934211850166321, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6934211850166321, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.0488918274641037, + 0.07253532111644745, + 0.06658177822828293, + 0.05108053237199783, + 0.3648497760295868, + 0.10225959867238998, + 0.08610926568508148, + 0.04153862223029137, + 0.06015915796160698, + 0.10599419474601746 + ], + "credit_map": [ + 0.00589478574693203, + 0.011355983093380928, + 0.005073371343314648, + 0.005334316752851009, + 0.006978458259254694, + 0.005599198862910271, + 0.0043494123965501785, + 0.004306301008909941, + 0.010503730736672878, + 0.009287526831030846, + 0.00843091681599617, + 0.012107068672776222, + 0.008934293873608112, + 0.008565633557736874, + 0.0075392951257526875, + 0.007166852708905935, + 0.008619646541774273, + 0.009472299367189407, + 0.008028540760278702, + 0.0056152730248868465, + 0.007705305702984333, + 0.010476884432137012, + 0.006831964943557978, + 0.009831869043409824, + 0.01116835419088602, + 0.004852966405451298, + 0.0034887867514044046, + 0.004180665593594313, + 0.0054551223292946815, + 0.015030276030302048, + 0.0031164786778390408, + 0.0037878823932260275, + 0.12269680202007294, + 0.014105133712291718, + 0.012150580063462257, + 0.01086435653269291, + 0.025144826620817184, + 0.15447039902210236, + 0.010842445306479931, + 0.014575188979506493, + 0.025857269763946533, + 0.004876706283539534, + 0.008179835043847561, + 0.005488940980285406, + 0.010349137708544731, + 0.03353768214583397, + 0.006469490472227335, + 0.007500538136810064, + 0.02178150974214077, + 0.00671991566196084, + 0.006034746766090393, + 0.005512386094778776, + 0.007620935328304768, + 0.0253111831843853, + 0.00624131178483367, + 0.006887273397296667, + 0.0045609064400196075, + 0.0035107447765767574, + 0.00402188953012228, + 0.004899152088910341, + 0.007389874663203955, + 0.004993530455976725, + 0.007783070672303438, + 0.004379452206194401, + 0.00670234952121973, + 0.004949995316565037, + 0.007143320515751839, + 0.004659310448914766, + 0.010082107968628407, + 0.013625707477331161, + 0.007797647267580032, + 0.0051987250335514545, + 0.00806050281971693, + 0.009646561928093433, + 0.007022798527032137, + 0.006144162267446518, + 0.012963518500328064, + 0.010555467568337917, + 0.0359640046954155, + 0.01563718356192112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 5 + ], + "selected_reward": -0.6934211850166321 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "medium shot, shot from an extremely low angle, almost level with the sand, makes the two crabs facing each other appear incredibly large and intimidating. The last rays of sunset bathe the beach in a warm orange-red hue, and the moist sand reflects the vibrant sunset sky, giving it a mirror-like sheen. In the center of the frame, a crab with a deep russet-brown shell, its surface rough and uneven, covered in tiny dents and the scratches of time, raises its disproportionately large right claw high, the serrations on its inner side gleaming coldly in the golden light, striking a menacing pose. A few inches away, another, smaller but equally aggressive crab, its shell a darker gray, lowers its body, opening and closing its two similarly sized claws on its chest. Cautiously moving its feet sideways, stirring up the thin layer of seawater beneath its feet. The camera remains still, steadily capturing this silent duel. The low-angle sunlight outlines their contours with a bright golden rim of light and casts extremely long, dancing shadows behind them. In the background, the distant waves, dyed a mixture of purple and orange, gently and rhythmically roll towards the shore, forming a sharp contrast between movement and stillness, tranquility and tension with the fierce confrontation on the verge of breaking out in the foreground. The entire scene is filled with the dramatic tension of primal vitality.", + "seed": 8607, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_15/index_8393_seed_8607_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1064688965678215, + "vlm_logprob": -0.1064688965678215, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.1064688965678215, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle, almost level with the sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame facing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the larger crab raising its right claw high while the smaller crab lowers its body and moves its feet sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a beach at sunset with warm orange-red lighting and a mirror-like sheen on the moist sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023588625714182854 + ], + "opposite_logprob_tokens": [ + -4.39858865737915 + ], + "logprob_reward": -0.023588625714182854, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.023588625714182854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18094100058078766, + 0.10554800927639008, + 0.07538487017154694, + 0.04178233817219734, + 0.02345086820423603, + 0.037449050694704056, + 0.03477736562490463, + 0.18326233327388763, + 0.061281464993953705, + 0.2561226785182953 + ], + "credit_map": [ + 0.027761807665228844, + 0.04293673112988472, + 0.015102032572031021, + 0.019977716729044914, + 0.027194347232580185, + 0.017038390040397644, + 0.019023697823286057, + 0.011906270869076252, + 0.018301453441381454, + 0.011721847578883171, + 0.020080607384443283, + 0.009494069963693619, + 0.014924226328730583, + 0.009089810773730278, + 0.01151292398571968, + 0.010423069819808006, + 0.011700766161084175, + 0.011987892910838127, + 0.008426684886217117, + 0.006198262330144644, + 0.01073573436588049, + 0.009448994882404804, + 0.01081954687833786, + 0.006066988222301006, + 0.006385833956301212, + 0.00920050498098135, + 0.003131517441943288, + 0.0035498535726219416, + 0.00350557710044086, + 0.006165879312902689, + 0.0037559561897069216, + 0.006087214220315218, + 0.002975963754579425, + 0.0037131719291210175, + 0.0027833101339638233, + 0.0026120305992662907, + 0.0030786977149546146, + 0.0023834560997784138, + 0.003314901376143098, + 0.0025893361307680607, + 0.005853958427906036, + 0.009213081561028957, + 0.0035147846210747957, + 0.0026675292756408453, + 0.004728575702756643, + 0.005365130957216024, + 0.0031527667306363583, + 0.0029532236512750387, + 0.0052278549410402775, + 0.003619255730882287, + 0.006280260626226664, + 0.003813514020293951, + 0.0037747996393591166, + 0.005087421741336584, + 0.003113725921139121, + 0.0038605351001024246, + 0.02666628174483776, + 0.023546302691102028, + 0.02875753678381443, + 0.023148179054260254, + 0.024562351405620575, + 0.021500596776604652, + 0.015649503096938133, + 0.0194315817207098, + 0.007060731295496225, + 0.007120907306671143, + 0.009765712544322014, + 0.008784706704318523, + 0.007094323635101318, + 0.008586163632571697, + 0.006673546973615885, + 0.006195375230163336, + 0.030790140852332115, + 0.029164552688598633, + 0.04180726408958435, + 0.0398859977722168, + 0.029937908053398132, + 0.027714809402823448, + 0.034465011209249496, + 0.022357014939188957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.023588625714182854 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle, almost level with the sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame facing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the larger crab raising its right claw high while the smaller crab lowers its body and moves its feet sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a beach at sunset with warm orange-red lighting and a mirror-like sheen on the moist sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048659548163414 + ], + "opposite_logprob_tokens": [ + -3.048659563064575 + ], + "logprob_reward": -0.048659548163414, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048659548163414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1570756584405899, + 0.163355752825737, + 0.20404697954654694, + 0.06449798494577408, + 0.08606210350990295, + 0.07906363904476166, + 0.02995939366519451, + 0.08276346325874329, + 0.033837880939245224, + 0.09933716058731079 + ], + "credit_map": [ + 0.02530556544661522, + 0.039057306945323944, + 0.01425257045775652, + 0.01256764866411686, + 0.02045307494699955, + 0.017278995364904404, + 0.013701916672289371, + 0.01445858459919691, + 0.03118963912129402, + 0.022471019998192787, + 0.026344045996665955, + 0.010128362104296684, + 0.025077151134610176, + 0.014479401521384716, + 0.019334938377141953, + 0.014331205748021603, + 0.034419041126966476, + 0.040248483419418335, + 0.01627335324883461, + 0.016202041879296303, + 0.021764449775218964, + 0.030926082283258438, + 0.022612858563661575, + 0.021600663661956787, + 0.009617662988603115, + 0.014771020039916039, + 0.0049705225974321365, + 0.005163237452507019, + 0.005759627558290958, + 0.006531317252665758, + 0.007821338251233101, + 0.009863256476819515, + 0.011742779053747654, + 0.014177145436406136, + 0.013827754184603691, + 0.010433416813611984, + 0.008154813200235367, + 0.008399615064263344, + 0.010721661150455475, + 0.008604924194514751, + 0.009211286902427673, + 0.018907444551587105, + 0.008953050710260868, + 0.005597296170890331, + 0.009057867340743542, + 0.011136231012642384, + 0.008705911226570606, + 0.007494545541703701, + 0.003924132324755192, + 0.0037362081930041313, + 0.006144698243588209, + 0.003663974581286311, + 0.0030227305833250284, + 0.0033833247143775225, + 0.002923776628449559, + 0.0031605474650859833, + 0.012706884182989597, + 0.012247616425156593, + 0.01257304847240448, + 0.008773678913712502, + 0.00999944657087326, + 0.010864938609302044, + 0.007960835471749306, + 0.00763701181858778, + 0.003955384716391563, + 0.0038085102569311857, + 0.005686212796717882, + 0.00433084974065423, + 0.0035953421611338854, + 0.004265332594513893, + 0.0040453108958899975, + 0.0041509391739964485, + 0.01105122547596693, + 0.013785479590296745, + 0.01475555170327425, + 0.015731146559119225, + 0.011542999185621738, + 0.011163841001689434, + 0.012817100621759892, + 0.00848980713635683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.048659548163414 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle, almost level with the sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame facing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the larger crab raising its right claw high while the smaller crab lowers its body and moves its feet sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a beach at sunset with warm orange-red lighting and a mirror-like sheen on the moist sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02982042171061039 + ], + "opposite_logprob_tokens": [ + -3.529820442199707 + ], + "logprob_reward": -0.02982042171061039, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982042171061039, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20025044679641724, + 0.1201372742652893, + 0.10412225127220154, + 0.07483981549739838, + 0.0451943539083004, + 0.2133633941411972, + 0.02635791338980198, + 0.09336371719837189, + 0.04017144814133644, + 0.08219940215349197 + ], + "credit_map": [ + 0.03587184473872185, + 0.0601075179874897, + 0.017907049506902695, + 0.013196544721722603, + 0.024594144895672798, + 0.01986689865589142, + 0.013452564366161823, + 0.015253894031047821, + 0.01843288354575634, + 0.013148530386388302, + 0.020575689151883125, + 0.00948537327349186, + 0.018900379538536072, + 0.010618547908961773, + 0.016331780701875687, + 0.012644077651202679, + 0.01760861463844776, + 0.01762421429157257, + 0.014161369763314724, + 0.011563263833522797, + 0.013210957869887352, + 0.008434364572167397, + 0.013657653704285622, + 0.007861820980906487, + 0.009870483539998531, + 0.013790390454232693, + 0.007859143428504467, + 0.009522656910121441, + 0.007120316848158836, + 0.008342902176082134, + 0.006577341351658106, + 0.011756584979593754, + 0.006862047594040632, + 0.005996566265821457, + 0.005318549927324057, + 0.004508289974182844, + 0.004568259697407484, + 0.0083465576171875, + 0.005301306489855051, + 0.004292777739465237, + 0.025377463549375534, + 0.0420815646648407, + 0.031308822333812714, + 0.0281654205173254, + 0.025615448132157326, + 0.021757038310170174, + 0.02033323235809803, + 0.018724413588643074, + 0.003413650207221508, + 0.003002607962116599, + 0.0035872769076377153, + 0.004783294629305601, + 0.0032494831830263138, + 0.0034714227076619864, + 0.0019503944786265492, + 0.002899783430621028, + 0.013296694494783878, + 0.01274069119244814, + 0.010958646424114704, + 0.00899520143866539, + 0.013850080780684948, + 0.015745477750897408, + 0.007772408425807953, + 0.010004514828324318, + 0.005774754099547863, + 0.006571324076503515, + 0.0058883968740701675, + 0.003737648716196418, + 0.0057914238423109055, + 0.004736434202641249, + 0.003907368518412113, + 0.0037640947848558426, + 0.009327162988483906, + 0.008966421708464622, + 0.017915453761816025, + 0.012382027693092823, + 0.008292890153825283, + 0.010056920349597931, + 0.008725655265152454, + 0.006532872095704079 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.02982042171061039 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle, almost level with the sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame facing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the larger crab raising its right claw high while the smaller crab lowers its body and moves its feet sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a beach at sunset with warm orange-red lighting and a mirror-like sheen on the moist sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013368725776672363 + ], + "opposite_logprob_tokens": [ + -6.626337051391602 + ], + "logprob_reward": -0.0013368725776672363, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013368725776672363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2970232367515564, + 0.17764101922512054, + 0.1467612385749817, + 0.09955957531929016, + 0.03566568344831467, + 0.08343417942523956, + 0.03143390268087387, + 0.039524246007204056, + 0.049546677619218826, + 0.03941021114587784 + ], + "credit_map": [ + 0.06381744891405106, + 0.06414032727479935, + 0.026610543951392174, + 0.02495068497955799, + 0.04277407005429268, + 0.024346379563212395, + 0.027640394866466522, + 0.02274339087307453, + 0.04136824235320091, + 0.023560795933008194, + 0.025741569697856903, + 0.012835860252380371, + 0.023555658757686615, + 0.013020049780607224, + 0.01819884032011032, + 0.019359983503818512, + 0.0282216165214777, + 0.02854757569730282, + 0.013423353433609009, + 0.011315247043967247, + 0.022714965045452118, + 0.016866877675056458, + 0.014659757725894451, + 0.011011842638254166, + 0.01795376092195511, + 0.0301494263112545, + 0.005664045922458172, + 0.007623204495757818, + 0.008340509608387947, + 0.013840405270457268, + 0.00836376566439867, + 0.007624455261975527, + 0.005587434396147728, + 0.004465207923203707, + 0.004992209374904633, + 0.004273676313459873, + 0.0032853083685040474, + 0.005547138396650553, + 0.004398147109895945, + 0.003116563195362687, + 0.014964195899665356, + 0.010565847158432007, + 0.007029442582279444, + 0.008664949797093868, + 0.007874959148466587, + 0.017448483034968376, + 0.010411600582301617, + 0.006474705412983894, + 0.004150796681642532, + 0.003937127534300089, + 0.006140667945146561, + 0.004717273637652397, + 0.0032832552678883076, + 0.0036086938343942165, + 0.002876895247027278, + 0.002719194395467639, + 0.005619748495519161, + 0.007896877825260162, + 0.004967659246176481, + 0.0035149301402270794, + 0.005638405214995146, + 0.0040714433416724205, + 0.003939845599234104, + 0.00387533544562757, + 0.006497913971543312, + 0.004591680597513914, + 0.008438927121460438, + 0.006543638650327921, + 0.005921164061874151, + 0.007747437804937363, + 0.005543511360883713, + 0.004262405447661877, + 0.004816995933651924, + 0.006288263481110334, + 0.004407916683703661, + 0.004191999323666096, + 0.00524525111541152, + 0.005707158241420984, + 0.004978963173925877, + 0.0037736620288342237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0013368725776672363 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle, almost level with the sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame facing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the larger crab raising its right claw high while the smaller crab lowers its body and moves its feet sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a beach at sunset with warm orange-red lighting and a mirror-like sheen on the moist sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42893901467323303 + ], + "opposite_logprob_tokens": [ + -1.0539389848709106 + ], + "logprob_reward": -0.42893901467323303, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42893901467323303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14181266725063324, + 0.06982806324958801, + 0.11089292168617249, + 0.04287885129451752, + 0.054181285202503204, + 0.1313498169183731, + 0.04646890237927437, + 0.12184871733188629, + 0.030138883739709854, + 0.2505998909473419 + ], + "credit_map": [ + 0.023931747302412987, + 0.03336787596344948, + 0.017580507323145866, + 0.012883355841040611, + 0.01648835279047489, + 0.013942754827439785, + 0.012963693588972092, + 0.010654378682374954, + 0.010945671238005161, + 0.008616844192147255, + 0.009871036745607853, + 0.006779486779123545, + 0.00910597387701273, + 0.0071033574640750885, + 0.01047577429562807, + 0.00692992378026247, + 0.021202271804213524, + 0.021463213488459587, + 0.008037248626351357, + 0.00870536733418703, + 0.01867879554629326, + 0.013968827202916145, + 0.011419720016419888, + 0.007417480461299419, + 0.008612295612692833, + 0.008352629840373993, + 0.004139018710702658, + 0.004291253164410591, + 0.003585046622902155, + 0.005145297851413488, + 0.004322292748838663, + 0.004431017674505711, + 0.007811299990862608, + 0.01299517322331667, + 0.005593038164079189, + 0.005947017576545477, + 0.005853509996086359, + 0.006134059280157089, + 0.0055366926826536655, + 0.004310494288802147, + 0.01681278645992279, + 0.026243947446346283, + 0.015967782586812973, + 0.01744619756937027, + 0.016820460557937622, + 0.015162297524511814, + 0.010791090317070484, + 0.01210525818169117, + 0.007316993083804846, + 0.004154990892857313, + 0.005031492095440626, + 0.004000895656645298, + 0.005688095465302467, + 0.009544217959046364, + 0.004572512581944466, + 0.0061597079038619995, + 0.01889428310096264, + 0.014495032839477062, + 0.01757502555847168, + 0.011349642649292946, + 0.011462501250207424, + 0.027350950986146927, + 0.009591639041900635, + 0.011129633523523808, + 0.0037086026277393103, + 0.0035951307509094477, + 0.005056289490312338, + 0.0034861238673329353, + 0.003512779250741005, + 0.003480789717286825, + 0.003660689340904355, + 0.003638478461652994, + 0.03014272451400757, + 0.02601654827594757, + 0.04294257238507271, + 0.039763011038303375, + 0.02959512174129486, + 0.02715991996228695, + 0.03409974277019501, + 0.020880239084362984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.42893901467323303 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_2/index_tensor([8340, 8340])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_2/index_tensor([8340, 8340])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a89e7435fb707f040caa1e1a40b24cec9c6f072f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_2/index_tensor([8340, 8340])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a steady, medium shot, a lush green meadow bathed in the soft light of a summer afternoon stands a snowman mascot, crowned with a small golden crown. Composed of two plump, round snowballs, the mascot possesses two gleaming coal-black eyes, a playful carrot nose, and a broad, joyful smile. One arm rises, gently and rhythmically wielding a peculiar ice cream scepter—a crisp waffle cone handle topped with a perfect scoop of strawberry-pink ice cream. With this joyful waving, its glistening white body begins to shimmer mysteriously. Countless intricate frost crystals, as if animate, rapidly condense, bloom, and spread from the snow's surface. In the warm, golden sunlight, these newly formed crystals refract a dazzling prismatic light. The entire scene is filled with heartwarming atmosphere and is rendered in a classic Ghibli style, with soft colors, hand-painted textures, and wildflowers in the background softened by a halo, creating a sense of innocence and a magical dreamlike feeling.", + "seed": 19065, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_2/index_8340_seed_19065_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.051789929065853356, + "vlm_logprob": -0.051789929065853356, + "vlm_yes_no_margin": 3.4000000238418577, + "vlm_token_credit": -0.051789929065853356, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush green meadow during a summer afternoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a snowman mascot composed of two round snowballs with a golden crown, coal-black eyes, a carrot nose, and a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowman mascot waving an ice cream scepter with a waffle cone handle and a strawberry-pink scoop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate frost crystals rapidly condense, bloom, and spread from the snowman's surface while refracting prismatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic Ghibli style with soft colors, hand-painted textures, and wildflowers in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14549414813518524 + ], + "opposite_logprob_tokens": [ + -2.020494222640991 + ], + "logprob_reward": -0.14549414813518524, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14549414813518524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3394321799278259, + 0.1396709382534027, + 0.05761132016777992, + 0.04800061881542206, + 0.030379995703697205, + 0.017124682664871216, + 0.055409472435712814, + 0.12870340049266815, + 0.1022973358631134, + 0.0813700407743454 + ], + "credit_map": [ + 0.038046617060899734, + 0.034883834421634674, + 0.04059627279639244, + 0.04418468847870827, + 0.049302954226732254, + 0.052878644317388535, + 0.05213043466210365, + 0.027408719062805176, + 0.015900515019893646, + 0.013458372093737125, + 0.014130452647805214, + 0.01875668577849865, + 0.021237008273601532, + 0.019854536280035973, + 0.022990886121988297, + 0.013342494145035744, + 0.005783836357295513, + 0.00639072572812438, + 0.007041091565042734, + 0.00887231808155775, + 0.0071896449662745, + 0.007120832800865173, + 0.00983714871108532, + 0.005375723820179701, + 0.008209552615880966, + 0.005370601080358028, + 0.004550343845039606, + 0.006080402061343193, + 0.006976571399718523, + 0.005823793355375528, + 0.007054717279970646, + 0.003934632521122694, + 0.009375475347042084, + 0.003291767556220293, + 0.003943409770727158, + 0.0032976854126900434, + 0.0014829522697255015, + 0.003988088574260473, + 0.002163204364478588, + 0.002837409731000662, + 0.004525442607700825, + 0.0016506939427927136, + 0.0018523137550801039, + 0.001933667459525168, + 0.0011623441241681576, + 0.002938052173703909, + 0.0013279098784551024, + 0.0017342573264613748, + 0.015015466138720512, + 0.003792228875681758, + 0.005910281091928482, + 0.004643646534532309, + 0.0035023691598325968, + 0.014356669038534164, + 0.0039156535640358925, + 0.004273159429430962, + 0.02197590470314026, + 0.009488099254667759, + 0.028418639674782753, + 0.011782206594944, + 0.006014646030962467, + 0.018566545099020004, + 0.008393612690269947, + 0.02406374365091324, + 0.03675030544400215, + 0.008910316042602062, + 0.005906027276068926, + 0.007157036568969488, + 0.003327737795189023, + 0.03218602016568184, + 0.0028934995643794537, + 0.005166395101696253, + 0.00791654922068119, + 0.009305492974817753, + 0.00855332426726818, + 0.0169405248016119, + 0.012237277813255787, + 0.009983254596590996, + 0.009737729094922543, + 0.006695888936519623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.14549414813518524 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush green meadow during a summer afternoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a snowman mascot composed of two round snowballs with a golden crown, coal-black eyes, a carrot nose, and a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowman mascot waving an ice cream scepter with a waffle cone handle and a strawberry-pink scoop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate frost crystals rapidly condense, bloom, and spread from the snowman's surface while refracting prismatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic Ghibli style with soft colors, hand-painted textures, and wildflowers in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059514218010008335 + ], + "opposite_logprob_tokens": [ + -5.130951404571533 + ], + "logprob_reward": -0.0059514218010008335, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059514218010008335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.036906104534864426, + 0.024073466658592224, + 0.010048069059848785, + 0.007437503896653652, + 0.011819514445960522, + 0.01684069074690342, + 0.13489820063114166, + 0.07528762519359589, + 0.640324592590332, + 0.042364247143268585 + ], + "credit_map": [ + 0.007776661776006222, + 0.0034945071674883366, + 0.00712883286178112, + 0.004101645667105913, + 0.002865091199055314, + 0.004350569099187851, + 0.002892395481467247, + 0.004296401049941778, + 0.004084499087184668, + 0.0023123198188841343, + 0.004087381064891815, + 0.002510923659428954, + 0.0021497756242752075, + 0.003230765461921692, + 0.0031376006081700325, + 0.0025602022651582956, + 0.0013217752566561103, + 0.0009480526787228882, + 0.0011591289658099413, + 0.001236354699358344, + 0.0010696618119254708, + 0.0021206934470683336, + 0.0012952650431543589, + 0.00089713663328439, + 0.0012966112699359655, + 0.0006609702249988914, + 0.0010898185428231955, + 0.0010243819560855627, + 0.0007597330259159207, + 0.0013765712501481175, + 0.0006250694859772921, + 0.0006043484318070114, + 0.0018961880123242736, + 0.0017652470851317048, + 0.0011492876801639795, + 0.0014435600023716688, + 0.0015765107236802578, + 0.0016013565473258495, + 0.0015831589698791504, + 0.0008042046101763844, + 0.0017061232356354594, + 0.0021413008216768503, + 0.0015396182425320148, + 0.003549779299646616, + 0.00211039325222373, + 0.0020265989005565643, + 0.0023427363485097885, + 0.0014241409953683615, + 0.03645367547869682, + 0.008481761440634727, + 0.008186410181224346, + 0.00855266023427248, + 0.009096675552427769, + 0.04471435397863388, + 0.009525151923298836, + 0.009887514635920525, + 0.01348629966378212, + 0.005768380127847195, + 0.016142116859555244, + 0.00719947274774313, + 0.0032688549254089594, + 0.010929093696177006, + 0.004842239897698164, + 0.013651169836521149, + 0.22649449110031128, + 0.055120110511779785, + 0.03345586359500885, + 0.04565136879682541, + 0.03246285393834114, + 0.19200289249420166, + 0.021944133564829826, + 0.03319288790225983, + 0.003839240176603198, + 0.003561910940334201, + 0.004147286992520094, + 0.010188697837293148, + 0.004507208708673716, + 0.010270855389535427, + 0.003064060816541314, + 0.0027849837206304073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.0059514218010008335 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush green meadow during a summer afternoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a snowman mascot composed of two round snowballs with a golden crown, coal-black eyes, a carrot nose, and a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowman mascot waving an ice cream scepter with a waffle cone handle and a strawberry-pink scoop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate frost crystals rapidly condense, bloom, and spread from the snowman's surface while refracting prismatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic Ghibli style with soft colors, hand-painted textures, and wildflowers in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809810429811478 + ], + "opposite_logprob_tokens": [ + -3.288098096847534 + ], + "logprob_reward": -0.03809810429811478, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809810429811478, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05710954964160919, + 0.02755407616496086, + 0.04702317714691162, + 0.010640955530107021, + 0.011133805848658085, + 0.015447783283889294, + 0.10843503475189209, + 0.11511847376823425, + 0.4895031750202179, + 0.11803390085697174 + ], + "credit_map": [ + 0.011288497596979141, + 0.005355055909603834, + 0.010383396409451962, + 0.00664353696629405, + 0.004619660787284374, + 0.006719714030623436, + 0.006158271338790655, + 0.005941418465226889, + 0.007182638626545668, + 0.0026233098469674587, + 0.004732075147330761, + 0.002342140767723322, + 0.0014357523759827018, + 0.004065431654453278, + 0.0017378524644300342, + 0.0034348785411566496, + 0.005300567485392094, + 0.005131714046001434, + 0.005201342981308699, + 0.007298759184777737, + 0.005462683271616697, + 0.006556043401360512, + 0.007812894880771637, + 0.004259178880602121, + 0.0014792871661484241, + 0.0011045662686228752, + 0.0017547018360346556, + 0.0012928181095048785, + 0.0011508505558595061, + 0.002059277379885316, + 0.0008132009534165263, + 0.0009862548904493451, + 0.001984229777008295, + 0.0010171952890232205, + 0.0017333644209429622, + 0.001378958928398788, + 0.0010433036368340254, + 0.0016728442860767245, + 0.0010474685113877058, + 0.0012564422795549035, + 0.0022742953151464462, + 0.0018433955265209079, + 0.0013529445277526975, + 0.0029293715488165617, + 0.0019537280313670635, + 0.0019076722674071789, + 0.001887087826617062, + 0.001299290219321847, + 0.030314862728118896, + 0.006601122207939625, + 0.006951082963496447, + 0.00661671906709671, + 0.006285029463469982, + 0.03647371008992195, + 0.007248509209603071, + 0.007944007404148579, + 0.017639193683862686, + 0.007000407669693232, + 0.027286265045404434, + 0.008755527436733246, + 0.005188988521695137, + 0.01630334183573723, + 0.008885333314538002, + 0.024059433490037918, + 0.1807917356491089, + 0.04214373603463173, + 0.023168344050645828, + 0.03356209024786949, + 0.01985594816505909, + 0.15363849699497223, + 0.013270621187984943, + 0.02307220734655857, + 0.01135010365396738, + 0.011568811722099781, + 0.011812740005552769, + 0.027304327115416527, + 0.013142752461135387, + 0.026900233700871468, + 0.008494634181261063, + 0.00746030081063509 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.03809810429811478 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush green meadow during a summer afternoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a snowman mascot composed of two round snowballs with a golden crown, coal-black eyes, a carrot nose, and a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowman mascot waving an ice cream scepter with a waffle cone handle and a strawberry-pink scoop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate frost crystals rapidly condense, bloom, and spread from the snowman's surface while refracting prismatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic Ghibli style with soft colors, hand-painted textures, and wildflowers in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306488484144211 + ], + "opposite_logprob_tokens": [ + -3.168064832687378 + ], + "logprob_reward": -0.04306488484144211, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306488484144211, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04723808914422989, + 0.028994079679250717, + 0.03487640619277954, + 0.03303582966327667, + 0.013986999168992043, + 0.012940721586346626, + 0.10080352425575256, + 0.16707630455493927, + 0.2985064685344696, + 0.26254159212112427 + ], + "credit_map": [ + 0.00813367497175932, + 0.004495090339332819, + 0.008598290383815765, + 0.004211841616779566, + 0.005752806086093187, + 0.007335355971008539, + 0.004219380673021078, + 0.0044916486367583275, + 0.0033442785497754812, + 0.002978275064378977, + 0.0037392894737422466, + 0.00425701355561614, + 0.003814967814832926, + 0.0033646435476839542, + 0.004714478738605976, + 0.002781132236123085, + 0.003854871029034257, + 0.0036446924787014723, + 0.0039095766842365265, + 0.005481688305735588, + 0.004144185688346624, + 0.0040472629480063915, + 0.006402585189789534, + 0.0033915459644049406, + 0.005126511678099632, + 0.0031681375112384558, + 0.004041265230625868, + 0.0038306349888443947, + 0.003963435534387827, + 0.005229344125837088, + 0.004439885728061199, + 0.003236614866182208, + 0.0029812746215611696, + 0.0012623192742466927, + 0.00233362871222198, + 0.001536560826934874, + 0.0009126963559538126, + 0.0021305668633431196, + 0.0011058738455176353, + 0.0017240791348740458, + 0.0026079723611474037, + 0.001379912020638585, + 0.0015941448509693146, + 0.0013419578317552805, + 0.0010153314797207713, + 0.0018071496160700917, + 0.0012835630914196372, + 0.0019106903346255422, + 0.030281970277428627, + 0.005384888034313917, + 0.006123427767306566, + 0.004945278167724609, + 0.004971971735358238, + 0.035769522190093994, + 0.0059534357860684395, + 0.007373024709522724, + 0.026121625676751137, + 0.012616765685379505, + 0.03587382286787033, + 0.015361595898866653, + 0.007692420855164528, + 0.02515670657157898, + 0.010361017659306526, + 0.033892348408699036, + 0.10966049134731293, + 0.02556345984339714, + 0.019187750294804573, + 0.020999450236558914, + 0.008714704774320126, + 0.09216112643480301, + 0.006769879255443811, + 0.015449588187038898, + 0.027865007519721985, + 0.036035723984241486, + 0.024129953235387802, + 0.04119173064827919, + 0.04431823268532753, + 0.043816737830638885, + 0.028096303343772888, + 0.0170879103243351 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.04306488484144211 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush green meadow during a summer afternoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a snowman mascot composed of two round snowballs with a golden crown, coal-black eyes, a carrot nose, and a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowman mascot waving an ice cream scepter with a waffle cone handle and a strawberry-pink scoop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate frost crystals rapidly condense, bloom, and spread from the snowman's surface while refracting prismatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic Ghibli style with soft colors, hand-painted textures, and wildflowers in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026341086253523827 + ], + "opposite_logprob_tokens": [ + -3.651341199874878 + ], + "logprob_reward": -0.026341086253523827, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026341086253523827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15098688006401062, + 0.08188040554523468, + 0.01670386828482151, + 0.032680828124284744, + 0.03406130149960518, + 0.02932107448577881, + 0.09321329742670059, + 0.09144952893257141, + 0.38949528336524963, + 0.08020756393671036 + ], + "credit_map": [ + 0.023861754685640335, + 0.016988271847367287, + 0.021560195833444595, + 0.02139158360660076, + 0.015474592335522175, + 0.022931912913918495, + 0.016604911535978317, + 0.012173658236861229, + 0.021771006286144257, + 0.006543200928717852, + 0.01511574536561966, + 0.006919492967426777, + 0.004781509283930063, + 0.01388944499194622, + 0.003836482996121049, + 0.009023522958159447, + 0.0032666337210685015, + 0.0014270704705268145, + 0.002848747419193387, + 0.0020637677516788244, + 0.0015963726909831166, + 0.0020192856900393963, + 0.0020353004802018404, + 0.0014466893626376987, + 0.004643616266548634, + 0.0033031380735337734, + 0.003015700727701187, + 0.004567546304315329, + 0.0038448660634458065, + 0.006301256362348795, + 0.004604104906320572, + 0.0024005991872400045, + 0.00780891440808773, + 0.0036654260475188494, + 0.004747642669826746, + 0.003952316008508205, + 0.0029637692496180534, + 0.004726553335785866, + 0.003101564245298505, + 0.0030951155349612236, + 0.005129454657435417, + 0.003124968847259879, + 0.0029623706359416246, + 0.0044886162504553795, + 0.003239353885874152, + 0.003749560797587037, + 0.004043720196932554, + 0.0025830278173089027, + 0.02352842129766941, + 0.006225688848644495, + 0.00913639273494482, + 0.008632203564047813, + 0.005338296294212341, + 0.026408040896058083, + 0.006024677772074938, + 0.007919574156403542, + 0.019420236349105835, + 0.009154046885669231, + 0.014475062489509583, + 0.00940098986029625, + 0.00614927290007472, + 0.012485293671488762, + 0.007816617377102375, + 0.012548013590276241, + 0.1437753438949585, + 0.03434385359287262, + 0.01705237105488777, + 0.026951637119054794, + 0.016777250915765762, + 0.12325763702392578, + 0.010302596725523472, + 0.01703459955751896, + 0.009420422837138176, + 0.007867041043937206, + 0.00959522183984518, + 0.012955596670508385, + 0.00958027970045805, + 0.01340558473020792, + 0.010494091548025608, + 0.006889324635267258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.026341086253523827 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a steady, medium shot, a lush green meadow bathed in the soft light of a summer afternoon stands a snowman mascot, crowned with a small golden crown. Composed of two plump, round snowballs, the mascot possesses two gleaming coal-black eyes, a playful carrot nose, and a broad, joyful smile. One arm rises, gently and rhythmically wielding a peculiar ice cream scepter—a crisp waffle cone handle topped with a perfect scoop of strawberry-pink ice cream. With this joyful waving, its glistening white body begins to shimmer mysteriously. Countless intricate frost crystals, as if animate, rapidly condense, bloom, and spread from the snow's surface. In the warm, golden sunlight, these newly formed crystals refract a dazzling prismatic light. The entire scene is filled with heartwarming atmosphere and is rendered in a classic Ghibli style, with soft colors, hand-painted textures, and wildflowers in the background softened by a halo, creating a sense of innocence and a magical dreamlike feeling.", + "seed": 19066, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_2/index_8340_seed_19066_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.023414438869804145, + "vlm_logprob": -0.023414438869804145, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.023414438869804145, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush green meadow during a summer afternoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a snowman mascot composed of two round snowballs with a golden crown, coal-black eyes, a carrot nose, and a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowman mascot waving an ice cream scepter with a waffle cone handle and a strawberry-pink scoop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate frost crystals rapidly condense, bloom, and spread from the snowman's surface while refracting prismatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic Ghibli style with soft colors, hand-painted textures, and wildflowers in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01889714226126671 + ], + "opposite_logprob_tokens": [ + -4.14389705657959 + ], + "logprob_reward": -0.01889714226126671, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01889714226126671, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04524511098861694, + 0.03783689811825752, + 0.007863846607506275, + 0.025829147547483444, + 0.03731540963053703, + 0.08589651435613632, + 0.5408398509025574, + 0.021728776395320892, + 0.004291087854653597, + 0.19315332174301147 + ], + "credit_map": [ + 0.0037882288452237844, + 0.0061333319172263145, + 0.006117701530456543, + 0.005485382862389088, + 0.005403824150562286, + 0.00556220905855298, + 0.007292287424206734, + 0.005462143570184708, + 0.0030849874019622803, + 0.005630925763398409, + 0.0056138476356863976, + 0.004523162730038166, + 0.004210706800222397, + 0.006527631543576717, + 0.0035936953499913216, + 0.004651936702430248, + 0.0007739788852632046, + 0.0009757184307090938, + 0.0016431616386398673, + 0.0008384126122109592, + 0.0007787985960021615, + 0.0009189313277602196, + 0.0008082883432507515, + 0.001126556657254696, + 0.0029326423536986113, + 0.00313686509616673, + 0.004214673303067684, + 0.0023707456421107054, + 0.0025849866215139627, + 0.005988352932035923, + 0.0018318138318136334, + 0.002769067184999585, + 0.0043592206202447414, + 0.0023232605308294296, + 0.0061004445888102055, + 0.0032613822259008884, + 0.004253687337040901, + 0.007725811563432217, + 0.004447430372238159, + 0.004844171926379204, + 0.00985841266810894, + 0.007274613715708256, + 0.011278180405497551, + 0.00618763966485858, + 0.0097817899659276, + 0.022789407521486282, + 0.008701722137629986, + 0.01002474594861269, + 0.06622373312711716, + 0.03477897495031357, + 0.053854599595069885, + 0.03433695062994957, + 0.0799490362405777, + 0.10590091347694397, + 0.10604345053434372, + 0.0597522146999836, + 0.0024403496645390987, + 0.002540142275393009, + 0.0029620034620165825, + 0.0016271290369331837, + 0.005041084252297878, + 0.0028676630463451147, + 0.0020563527941703796, + 0.0021940527949482203, + 0.0004970897571183741, + 0.0004384760686662048, + 0.0006127023953013122, + 0.00047017139149829745, + 0.0006923711043782532, + 0.0004929418792016804, + 0.0005869906162843108, + 0.0005003444384783506, + 0.008427503518760204, + 0.008353590965270996, + 0.019033368676900864, + 0.05406363308429718, + 0.013567119836807251, + 0.005832999013364315, + 0.07050275057554245, + 0.013372342102229595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 5 + ], + "selected_reward": -0.01889714226126671 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush green meadow during a summer afternoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a snowman mascot composed of two round snowballs with a golden crown, coal-black eyes, a carrot nose, and a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowman mascot waving an ice cream scepter with a waffle cone handle and a strawberry-pink scoop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate frost crystals rapidly condense, bloom, and spread from the snowman's surface while refracting prismatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic Ghibli style with soft colors, hand-painted textures, and wildflowers in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028259125538170338 + ], + "opposite_logprob_tokens": [ + -5.877825736999512 + ], + "logprob_reward": -0.0028259125538170338, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028259125538170338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04932907596230507, + 0.05426143482327461, + 0.024281350895762444, + 0.021582825109362602, + 0.029699906706809998, + 0.04447346553206444, + 0.2547507882118225, + 0.12156806141138077, + 0.04217790812253952, + 0.35787516832351685 + ], + "credit_map": [ + 0.0036439101677387953, + 0.004235338885337114, + 0.007851810194551945, + 0.006372677627950907, + 0.004545948468148708, + 0.006660103797912598, + 0.007666205987334251, + 0.008353080600500107, + 0.004822141956537962, + 0.008609402924776077, + 0.007545142900198698, + 0.006759645417332649, + 0.005930607207119465, + 0.008995873853564262, + 0.004678700584918261, + 0.006919922772794962, + 0.0019496469758450985, + 0.0031551653519272804, + 0.0037988389376550913, + 0.003278872696682811, + 0.0029944260604679585, + 0.0029794308356940746, + 0.002992828842252493, + 0.0031321425922214985, + 0.0022279280237853527, + 0.003061699215322733, + 0.0025626271963119507, + 0.0025080139748752117, + 0.0021267395932227373, + 0.00352564943023026, + 0.003107555443421006, + 0.0024626129306852818, + 0.0036927934270352125, + 0.0020499182865023613, + 0.0036030313931405544, + 0.003099987516179681, + 0.003879737574607134, + 0.0052010053768754005, + 0.003709117416292429, + 0.004464318975806236, + 0.005865268409252167, + 0.003725275630131364, + 0.005693730898201466, + 0.003484521061182022, + 0.004826220218092203, + 0.011014126241207123, + 0.005042573437094688, + 0.004821745678782463, + 0.034410495311021805, + 0.016169335693120956, + 0.027192234992980957, + 0.015361261554062366, + 0.034068167209625244, + 0.053702861070632935, + 0.04624953493475914, + 0.027596911415457726, + 0.009817545302212238, + 0.014508021995425224, + 0.011158744804561138, + 0.007865123450756073, + 0.037709493190050125, + 0.016179591417312622, + 0.011668837629258633, + 0.012660689651966095, + 0.005671733990311623, + 0.0038473971653729677, + 0.006470378953963518, + 0.00458753015846014, + 0.006464048754423857, + 0.004249020479619503, + 0.0055430373176932335, + 0.005344759672880173, + 0.01709958352148533, + 0.014898998662829399, + 0.04084857553243637, + 0.09252605587244034, + 0.025209510698914528, + 0.011848782189190388, + 0.1303817629814148, + 0.02506190910935402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.0028259125538170338 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush green meadow during a summer afternoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a snowman mascot composed of two round snowballs with a golden crown, coal-black eyes, a carrot nose, and a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowman mascot waving an ice cream scepter with a waffle cone handle and a strawberry-pink scoop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate frost crystals rapidly condense, bloom, and spread from the snowman's surface while refracting prismatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic Ghibli style with soft colors, hand-painted textures, and wildflowers in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195299707353115 + ], + "opposite_logprob_tokens": [ + -4.264195442199707 + ], + "logprob_reward": -0.014195299707353115, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195299707353115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03844855725765228, + 0.2545214891433716, + 0.008549361489713192, + 0.01678105816245079, + 0.06578138470649719, + 0.03440580517053604, + 0.36934778094291687, + 0.0322185680270195, + 0.01403027679771185, + 0.16591577231884003 + ], + "credit_map": [ + 0.0036424086429178715, + 0.005538657773286104, + 0.005472628865391016, + 0.004148620646446943, + 0.004179258830845356, + 0.006057074759155512, + 0.004313863813877106, + 0.005096040200442076, + 0.02282084710896015, + 0.04747866466641426, + 0.03811211884021759, + 0.026649611070752144, + 0.02650630660355091, + 0.04073786363005638, + 0.020092640072107315, + 0.032123398035764694, + 0.0007477826438844204, + 0.0014213836984708905, + 0.0012339722597971559, + 0.0010169311426579952, + 0.0013108128914609551, + 0.00086635397747159, + 0.0009894672548398376, + 0.0009626563987694681, + 0.0014468516455963254, + 0.0015836243983358145, + 0.0025702177081257105, + 0.0017560705309733748, + 0.002052302472293377, + 0.0032179360277950764, + 0.0018872568616643548, + 0.002266795840114355, + 0.0074778334237635136, + 0.004190709907561541, + 0.011289577931165695, + 0.005626581609249115, + 0.006873957347124815, + 0.014664295129477978, + 0.0066505069844424725, + 0.009007911197841167, + 0.0031402623280882835, + 0.0024023104924708605, + 0.0036232718266546726, + 0.0036814112681895494, + 0.004395964089781046, + 0.007200127467513084, + 0.0059790355153381824, + 0.003983417525887489, + 0.046693120151758194, + 0.024283654987812042, + 0.03892872855067253, + 0.023112043738365173, + 0.05179079249501228, + 0.07581201195716858, + 0.06834965944290161, + 0.040377743542194366, + 0.0027664541266858578, + 0.003996560350060463, + 0.0031530021224170923, + 0.002290058648213744, + 0.009759968146681786, + 0.004058874677866697, + 0.0028660588432103395, + 0.0033275866881012917, + 0.0018014218658208847, + 0.0014534792862832546, + 0.0017090565524995327, + 0.0014524009311571717, + 0.0016141175292432308, + 0.001917867106385529, + 0.0023116155061870813, + 0.0017703166231513023, + 0.007481574546545744, + 0.006848122924566269, + 0.022627174854278564, + 0.040398940443992615, + 0.011536980047821999, + 0.005684747826308012, + 0.05939169600605965, + 0.01194651611149311 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 9 + ], + "selected_reward": -0.014195299707353115 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush green meadow during a summer afternoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a snowman mascot composed of two round snowballs with a golden crown, coal-black eyes, a carrot nose, and a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowman mascot waving an ice cream scepter with a waffle cone handle and a strawberry-pink scoop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate frost crystals rapidly condense, bloom, and spread from the snowman's surface while refracting prismatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic Ghibli style with soft colors, hand-painted textures, and wildflowers in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305906221270561 + ], + "opposite_logprob_tokens": [ + -3.1680591106414795 + ], + "logprob_reward": -0.04305906221270561, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305906221270561, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03600262850522995, + 0.0369696281850338, + 0.02992977574467659, + 0.03520835191011429, + 0.010867349803447723, + 0.10079744458198547, + 0.11663994193077087, + 0.05128305032849312, + 0.051353685557842255, + 0.5309481620788574 + ], + "credit_map": [ + 0.003547844709828496, + 0.004901835694909096, + 0.005300846416503191, + 0.004811435006558895, + 0.0037747847381979227, + 0.005031530279666185, + 0.0050405943766236305, + 0.0035937547218054533, + 0.00343373348005116, + 0.00704501336440444, + 0.005488207098096609, + 0.0035933733452111483, + 0.0040800427086651325, + 0.0061738863587379456, + 0.00257995561696589, + 0.004575415980070829, + 0.002466957550495863, + 0.00394994392991066, + 0.0040910388343036175, + 0.0037871357053518295, + 0.003762626089155674, + 0.0037937774322927, + 0.0043935771100223064, + 0.0036847195588052273, + 0.004578146152198315, + 0.0050633735954761505, + 0.0055182757787406445, + 0.0032880331855267286, + 0.003678269684314728, + 0.006772039458155632, + 0.002524462528526783, + 0.003785752458497882, + 0.001470773946493864, + 0.0007516288314945996, + 0.0012930934317409992, + 0.0008786725229583681, + 0.0016917636385187507, + 0.0019105483079329133, + 0.001526786363683641, + 0.0013440829934552312, + 0.011561250314116478, + 0.0088527612388134, + 0.01388680562376976, + 0.007397019304335117, + 0.011504009366035461, + 0.026325087994337082, + 0.009323809295892715, + 0.01194669771939516, + 0.015338224358856678, + 0.007428932934999466, + 0.011704514734447002, + 0.007047816179692745, + 0.015793291851878166, + 0.02406793087720871, + 0.022067144513130188, + 0.013192087411880493, + 0.0035004466772079468, + 0.00631631538271904, + 0.00445476733148098, + 0.0033112415112555027, + 0.016603577882051468, + 0.006921400781720877, + 0.004661614540964365, + 0.005513687618076801, + 0.007764698471873999, + 0.004489298444241285, + 0.005108715035021305, + 0.006007461808621883, + 0.008708924055099487, + 0.005438452586531639, + 0.007872500456869602, + 0.005963635165244341, + 0.024840807542204857, + 0.02385665662586689, + 0.04636240378022194, + 0.14706487953662872, + 0.038723453879356384, + 0.017683815211057663, + 0.19443874061107635, + 0.03797738999128342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 5 + ], + "selected_reward": -0.04305906221270561 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush green meadow during a summer afternoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a snowman mascot composed of two round snowballs with a golden crown, coal-black eyes, a carrot nose, and a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowman mascot waving an ice cream scepter with a waffle cone handle and a strawberry-pink scoop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate frost crystals rapidly condense, bloom, and spread from the snowman's surface while refracting prismatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic Ghibli style with soft colors, hand-painted textures, and wildflowers in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809477761387825 + ], + "opposite_logprob_tokens": [ + -3.2880947589874268 + ], + "logprob_reward": -0.03809477761387825, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809477761387825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09538491070270538, + 0.023965056985616684, + 0.02947312965989113, + 0.11711988598108292, + 0.011838615871965885, + 0.11512648314237595, + 0.24455706775188446, + 0.029744120314717293, + 0.027400262653827667, + 0.30539050698280334 + ], + "credit_map": [ + 0.007104875519871712, + 0.007330254651606083, + 0.013760964386165142, + 0.011275110766291618, + 0.008210241794586182, + 0.01842041127383709, + 0.009111992083489895, + 0.020171063020825386, + 0.0024333999026566744, + 0.0039782230742275715, + 0.003617488080635667, + 0.0022088561672717333, + 0.0025521577335894108, + 0.004185470286756754, + 0.0022855366114526987, + 0.0027039244305342436, + 0.0026415144093334675, + 0.003695292863994837, + 0.006237373687326908, + 0.003318297676742077, + 0.0030636547598987818, + 0.0038430376444011927, + 0.002528527984395623, + 0.004145430866628885, + 0.01348906010389328, + 0.013390162028372288, + 0.02057086117565632, + 0.010525603778660297, + 0.011660323478281498, + 0.025159629061818123, + 0.009024704806506634, + 0.013299543410539627, + 0.0013820806052535772, + 0.0010527120903134346, + 0.0015972851542755961, + 0.001601752475835383, + 0.0013786935014650226, + 0.001611984334886074, + 0.0015278473729267716, + 0.0016862597549334168, + 0.011734416708350182, + 0.009666281752288342, + 0.015281585976481438, + 0.008620982989668846, + 0.013480125926434994, + 0.03043414279818535, + 0.012279667891561985, + 0.013629280030727386, + 0.0296395942568779, + 0.016124585643410683, + 0.025096731260418892, + 0.015400477685034275, + 0.03480611741542816, + 0.04956983029842377, + 0.04650954157114029, + 0.027410168200731277, + 0.002732793567702174, + 0.003379750531166792, + 0.0036899761762470007, + 0.002060622675344348, + 0.006055512465536594, + 0.00527852913364768, + 0.0032299160957336426, + 0.00331702153198421, + 0.005176300648599863, + 0.0024964704643934965, + 0.003668907331302762, + 0.0028522845823317766, + 0.0038669423665851355, + 0.0025057203602045774, + 0.0038252712693065405, + 0.0030083665624260902, + 0.013061633333563805, + 0.014731190167367458, + 0.013290246948599815, + 0.10169508308172226, + 0.021419094875454903, + 0.008888555690646172, + 0.11163879185914993, + 0.020665915682911873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 3 + ], + "selected_reward": -0.03809477761387825 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a steady, medium shot, a lush green meadow bathed in the soft light of a summer afternoon stands a snowman mascot, crowned with a small golden crown. Composed of two plump, round snowballs, the mascot possesses two gleaming coal-black eyes, a playful carrot nose, and a broad, joyful smile. One arm rises, gently and rhythmically wielding a peculiar ice cream scepter—a crisp waffle cone handle topped with a perfect scoop of strawberry-pink ice cream. With this joyful waving, its glistening white body begins to shimmer mysteriously. Countless intricate frost crystals, as if animate, rapidly condense, bloom, and spread from the snow's surface. In the warm, golden sunlight, these newly formed crystals refract a dazzling prismatic light. The entire scene is filled with heartwarming atmosphere and is rendered in a classic Ghibli style, with soft colors, hand-painted textures, and wildflowers in the background softened by a halo, creating a sense of innocence and a magical dreamlike feeling.", + "seed": 19067, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_2/index_8340_seed_19067_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05630369270220399, + "vlm_logprob": -0.05630369270220399, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.05630369270220399, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush green meadow during a summer afternoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a snowman mascot composed of two round snowballs with a golden crown, coal-black eyes, a carrot nose, and a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowman mascot waving an ice cream scepter with a waffle cone handle and a strawberry-pink scoop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate frost crystals rapidly condense, bloom, and spread from the snowman's surface while refracting prismatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic Ghibli style with soft colors, hand-painted textures, and wildflowers in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10272455960512161 + ], + "opposite_logprob_tokens": [ + -2.352724552154541 + ], + "logprob_reward": -0.10272455960512161, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10272455960512161, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2734241783618927, + 0.15512152016162872, + 0.019963175058364868, + 0.03036794438958168, + 0.10841052979230881, + 0.04460298642516136, + 0.22643564641475677, + 0.02229052223265171, + 0.05804116278886795, + 0.06134233623743057 + ], + "credit_map": [ + 0.03757219389081001, + 0.047545865178108215, + 0.044658113270998, + 0.019683577120304108, + 0.02924618311226368, + 0.03855939954519272, + 0.03723735362291336, + 0.01892147772014141, + 0.020387006923556328, + 0.026364615187048912, + 0.021903233602643013, + 0.017350589856505394, + 0.017418470233678818, + 0.01900647208094597, + 0.020437879487872124, + 0.012253252789378166, + 0.0024163122288882732, + 0.004067280795425177, + 0.003965806681662798, + 0.0018449213821440935, + 0.001844216720201075, + 0.002725869184359908, + 0.0011443749535828829, + 0.0019543911330401897, + 0.0030695623718202114, + 0.007160466630011797, + 0.0056287492625415325, + 0.002125951461493969, + 0.0023907816503196955, + 0.004304579459130764, + 0.002678915159776807, + 0.003008936531841755, + 0.017862888053059578, + 0.013328691013157368, + 0.023032134398818016, + 0.007572503294795752, + 0.007413467392325401, + 0.01709606498479843, + 0.009466651827096939, + 0.012638118118047714, + 0.005494847893714905, + 0.010269233956933022, + 0.005654595326632261, + 0.004204338416457176, + 0.004063304513692856, + 0.005897349677979946, + 0.003607451915740967, + 0.005411861464381218, + 0.021302055567502975, + 0.019397150725126266, + 0.009447849355638027, + 0.01915626972913742, + 0.04569700360298157, + 0.03192102164030075, + 0.06357719004154205, + 0.015937093645334244, + 0.004053873009979725, + 0.004160224460065365, + 0.002612003358080983, + 0.0014847933780401945, + 0.0031480013858526945, + 0.002392069436609745, + 0.0027987284120172262, + 0.0016408271621912718, + 0.005360408686101437, + 0.005423315800726414, + 0.009009124711155891, + 0.004516211338341236, + 0.00844055786728859, + 0.006904774345457554, + 0.008332733996212482, + 0.010054031386971474, + 0.008207968436181545, + 0.007301525212824345, + 0.004547572694718838, + 0.0036113739479333162, + 0.017100293189287186, + 0.006300649605691433, + 0.00907660648226738, + 0.005196345038712025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.10272455960512161 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush green meadow during a summer afternoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a snowman mascot composed of two round snowballs with a golden crown, coal-black eyes, a carrot nose, and a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowman mascot waving an ice cream scepter with a waffle cone handle and a strawberry-pink scoop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate frost crystals rapidly condense, bloom, and spread from the snowman's surface while refracting prismatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic Ghibli style with soft colors, hand-painted textures, and wildflowers in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006743293721228838 + ], + "opposite_logprob_tokens": [ + -5.006743431091309 + ], + "logprob_reward": -0.006743293721228838, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006743293721228838, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07215484976768494, + 0.0353316068649292, + 0.02831299602985382, + 0.0358375683426857, + 0.0432383194565773, + 0.01905444823205471, + 0.5120128393173218, + 0.04419947788119316, + 0.12753570079803467, + 0.08232218772172928 + ], + "credit_map": [ + 0.00861503928899765, + 0.013563845306634903, + 0.014243504032492638, + 0.0065305293537676334, + 0.005822821520268917, + 0.01111863274127245, + 0.00495441909879446, + 0.007306058891117573, + 0.0061624846421182156, + 0.009058872237801552, + 0.003979392349720001, + 0.0027010939083993435, + 0.0026317357551306486, + 0.004675083328038454, + 0.0032236624974757433, + 0.002899284241721034, + 0.0034443058539181948, + 0.007741734851151705, + 0.004249434918165207, + 0.0028344066813588142, + 0.0020634070970118046, + 0.0036813640035688877, + 0.001655673491768539, + 0.0026426699478179216, + 0.004182096105068922, + 0.011705854907631874, + 0.0044633205980062485, + 0.003048095852136612, + 0.0034964734222739935, + 0.004299757070839405, + 0.0025316854007542133, + 0.0021102898754179478, + 0.008051246404647827, + 0.007737224455922842, + 0.007707929238677025, + 0.004357447847723961, + 0.0025884860660880804, + 0.006207199301570654, + 0.002942435909062624, + 0.003646349534392357, + 0.0027298692148178816, + 0.004159476608037949, + 0.0015769855817779899, + 0.0012373316567391157, + 0.002864580601453781, + 0.0026244036853313446, + 0.002440578071400523, + 0.0014212228124961257, + 0.041277311742305756, + 0.037083081901073456, + 0.019065752625465393, + 0.04221557080745697, + 0.11610204726457596, + 0.06191782280802727, + 0.1633511334657669, + 0.03100014291703701, + 0.0049403514713048935, + 0.008192195557057858, + 0.005475990939885378, + 0.0034832474775612354, + 0.007066862657666206, + 0.004788591992110014, + 0.006746720522642136, + 0.003505515865981579, + 0.013627276755869389, + 0.013623202219605446, + 0.010133878327906132, + 0.012029492296278477, + 0.028920458629727364, + 0.013381000608205795, + 0.024087805300951004, + 0.011732582934200764, + 0.007684428244829178, + 0.00960136391222477, + 0.003975618630647659, + 0.004274397157132626, + 0.03261140361428261, + 0.007630151696503162, + 0.008687126450240612, + 0.007857690565288067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 9 + ], + "selected_reward": -0.006743293721228838 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush green meadow during a summer afternoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a snowman mascot composed of two round snowballs with a golden crown, coal-black eyes, a carrot nose, and a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowman mascot waving an ice cream scepter with a waffle cone handle and a strawberry-pink scoop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate frost crystals rapidly condense, bloom, and spread from the snowman's surface while refracting prismatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic Ghibli style with soft colors, hand-painted textures, and wildflowers in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043061915785074234 + ], + "opposite_logprob_tokens": [ + -3.1680619716644287 + ], + "logprob_reward": -0.043061915785074234, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043061915785074234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12879689037799835, + 0.054108910262584686, + 0.05753835663199425, + 0.058764707297086716, + 0.18922780454158783, + 0.031115001067519188, + 0.26337772607803345, + 0.046393200755119324, + 0.08081526309251785, + 0.08986212313175201 + ], + "credit_map": [ + 0.017666302621364594, + 0.03247102349996567, + 0.020722828805446625, + 0.009353447705507278, + 0.013615807518362999, + 0.016534699127078056, + 0.007883790880441666, + 0.010548993945121765, + 0.006300424691289663, + 0.008523338474333286, + 0.007203645072877407, + 0.006088513880968094, + 0.006487970240414143, + 0.0072631738148629665, + 0.00768470112234354, + 0.0045571476221084595, + 0.010021325200796127, + 0.019449668005108833, + 0.005955181550234556, + 0.004236084409058094, + 0.004148120526224375, + 0.007224731612950563, + 0.002996023977175355, + 0.00350722623988986, + 0.007792733144015074, + 0.022206729277968407, + 0.005461249966174364, + 0.003656300250440836, + 0.005489378236234188, + 0.006461483892053366, + 0.003775041550397873, + 0.003921785391867161, + 0.040572475641965866, + 0.03867834433913231, + 0.03731710836291313, + 0.013083355501294136, + 0.007285319268703461, + 0.030543042346835136, + 0.006935644429177046, + 0.014812524430453777, + 0.004698856268078089, + 0.005149812903255224, + 0.0024129541125148535, + 0.0018707789713516831, + 0.0035714206751435995, + 0.007631640415638685, + 0.0032402160577476025, + 0.0025393213145434856, + 0.020951081067323685, + 0.019521962851285934, + 0.011565114371478558, + 0.02110963501036167, + 0.06046183407306671, + 0.030325476080179214, + 0.08325802534818649, + 0.01618463359773159, + 0.005369893740862608, + 0.008202598430216312, + 0.004839023575186729, + 0.004431828390806913, + 0.0059833526611328125, + 0.004901680164039135, + 0.009053128771483898, + 0.003611693624407053, + 0.009155001491308212, + 0.011165878735482693, + 0.006582289468497038, + 0.007144349627196789, + 0.020436586812138557, + 0.008033407852053642, + 0.011661091819405556, + 0.0066366554237902164, + 0.008721048943698406, + 0.011525093577802181, + 0.004349211230874062, + 0.005329044070094824, + 0.033486947417259216, + 0.008230039849877357, + 0.01046203263103962, + 0.007758714258670807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.043061915785074234 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush green meadow during a summer afternoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a snowman mascot composed of two round snowballs with a golden crown, coal-black eyes, a carrot nose, and a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowman mascot waving an ice cream scepter with a waffle cone handle and a strawberry-pink scoop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate frost crystals rapidly condense, bloom, and spread from the snowman's surface while refracting prismatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic Ghibli style with soft colors, hand-painted textures, and wildflowers in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292478442192078 + ], + "opposite_logprob_tokens": [ + -2.237924814224243 + ], + "logprob_reward": -0.11292478442192078, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292478442192078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11606524884700775, + 0.10186333954334259, + 0.057764098048210144, + 0.04600638151168823, + 0.07091598212718964, + 0.040587592869997025, + 0.27643999457359314, + 0.08949147164821625, + 0.07731165736913681, + 0.12355431169271469 + ], + "credit_map": [ + 0.02197040244936943, + 0.02311602607369423, + 0.016188358888030052, + 0.00856884103268385, + 0.010068638250231743, + 0.01691196672618389, + 0.009778344072401524, + 0.009462674148380756, + 0.017296435311436653, + 0.028095286339521408, + 0.01658005453646183, + 0.007541019469499588, + 0.005812531802803278, + 0.012768629007041454, + 0.005413798615336418, + 0.008355587720870972, + 0.006614177953451872, + 0.009981636889278889, + 0.008681184612214565, + 0.007445902097970247, + 0.005804806016385555, + 0.007334956433624029, + 0.00721078971400857, + 0.0046906424686312675, + 0.004312889650464058, + 0.008184511214494705, + 0.009283916093409061, + 0.004702161531895399, + 0.00453657703474164, + 0.0067645409144461155, + 0.0031131973955780268, + 0.00510858790948987, + 0.011837072670459747, + 0.012488367967307568, + 0.010225320234894753, + 0.007247096858918667, + 0.004655783995985985, + 0.009998688474297523, + 0.006595027167350054, + 0.007868622429668903, + 0.00750461732968688, + 0.007888870313763618, + 0.0074119484052062035, + 0.002891580807045102, + 0.0028432023245841265, + 0.005689548794180155, + 0.002556164748966694, + 0.0038016599137336016, + 0.02372620441019535, + 0.030807405710220337, + 0.011890515685081482, + 0.021605009213089943, + 0.060197263956069946, + 0.028679389506578445, + 0.08395101130008698, + 0.015583189204335213, + 0.014717844314873219, + 0.01814381219446659, + 0.006841033231467009, + 0.007132026832550764, + 0.016288306564092636, + 0.009201841428875923, + 0.011874588206410408, + 0.005292017478495836, + 0.009960895404219627, + 0.009799998253583908, + 0.01010194979608059, + 0.006483996752649546, + 0.015659887343645096, + 0.00818112026900053, + 0.010700472630560398, + 0.006423340644687414, + 0.012296615168452263, + 0.012276528403162956, + 0.007929263636469841, + 0.007361480966210365, + 0.04506729543209076, + 0.011827091686427593, + 0.017183825373649597, + 0.009612211026251316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.11292478442192078 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush green meadow during a summer afternoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a snowman mascot composed of two round snowballs with a golden crown, coal-black eyes, a carrot nose, and a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowman mascot waving an ice cream scepter with a waffle cone handle and a strawberry-pink scoop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate frost crystals rapidly condense, bloom, and spread from the snowman's surface while refracting prismatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic Ghibli style with soft colors, hand-painted textures, and wildflowers in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016063909977674484 + ], + "opposite_logprob_tokens": [ + -4.141063690185547 + ], + "logprob_reward": -0.016063909977674484, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016063909977674484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16181910037994385, + 0.08808340132236481, + 0.0526232048869133, + 0.06138851121068001, + 0.07661262154579163, + 0.03231192007660866, + 0.3790384829044342, + 0.032048411667346954, + 0.058573007583618164, + 0.05750136822462082 + ], + "credit_map": [ + 0.031040456146001816, + 0.025543635711073875, + 0.02674630843102932, + 0.014295827597379684, + 0.012942833825945854, + 0.025794388726353645, + 0.012882021255791187, + 0.012573632411658764, + 0.014008981175720692, + 0.01732497662305832, + 0.021282754838466644, + 0.004499356262385845, + 0.0038299921434372663, + 0.014079118147492409, + 0.003869222244247794, + 0.009188994765281677, + 0.00572019349783659, + 0.008417739532887936, + 0.012639332562685013, + 0.004795528016984463, + 0.003637362504377961, + 0.00833757221698761, + 0.0028433999978005886, + 0.006232071202248335, + 0.0069269961677491665, + 0.017454765737056732, + 0.010683884844183922, + 0.005310608074069023, + 0.0048026531003415585, + 0.008524216711521149, + 0.00303967809304595, + 0.004645709414035082, + 0.010748039931058884, + 0.011537338607013226, + 0.011294009163975716, + 0.007226168178021908, + 0.006439104210585356, + 0.010486739687621593, + 0.009475884959101677, + 0.009405335411429405, + 0.0043371254578232765, + 0.007777807302772999, + 0.003978980705142021, + 0.002941167214885354, + 0.0038981111720204353, + 0.0041609713807702065, + 0.0024308282881975174, + 0.002786925993859768, + 0.033271368592977524, + 0.030116518959403038, + 0.01660558022558689, + 0.03239486366510391, + 0.08106636255979538, + 0.048011813312768936, + 0.11334481835365295, + 0.02422715723514557, + 0.005938020069152117, + 0.004878128878772259, + 0.0039660194888710976, + 0.002970442408695817, + 0.004343640524893999, + 0.004337648395448923, + 0.003344927681609988, + 0.0022695856168866158, + 0.006280494853854179, + 0.007662395481020212, + 0.004499136935919523, + 0.007592393551021814, + 0.010101016610860825, + 0.005667152348905802, + 0.01258358545601368, + 0.0041868300177156925, + 0.006395718548446894, + 0.006782368756830692, + 0.005321423523128033, + 0.004080421756953001, + 0.014386629685759544, + 0.005742172710597515, + 0.009480546228587627, + 0.005312084220349789 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.016063909977674484 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a steady, medium shot, a lush green meadow bathed in the soft light of a summer afternoon stands a snowman mascot, crowned with a small golden crown. Composed of two plump, round snowballs, the mascot possesses two gleaming coal-black eyes, a playful carrot nose, and a broad, joyful smile. One arm rises, gently and rhythmically wielding a peculiar ice cream scepter—a crisp waffle cone handle topped with a perfect scoop of strawberry-pink ice cream. With this joyful waving, its glistening white body begins to shimmer mysteriously. Countless intricate frost crystals, as if animate, rapidly condense, bloom, and spread from the snow's surface. In the warm, golden sunlight, these newly formed crystals refract a dazzling prismatic light. The entire scene is filled with heartwarming atmosphere and is rendered in a classic Ghibli style, with soft colors, hand-painted textures, and wildflowers in the background softened by a halo, creating a sense of innocence and a magical dreamlike feeling.", + "seed": 19068, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_2/index_8340_seed_19068_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05083874072879553, + "vlm_logprob": -0.05083874072879553, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.05083874072879553, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush green meadow during a summer afternoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a snowman mascot composed of two round snowballs with a golden crown, coal-black eyes, a carrot nose, and a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowman mascot waving an ice cream scepter with a waffle cone handle and a strawberry-pink scoop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate frost crystals rapidly condense, bloom, and spread from the snowman's surface while refracting prismatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic Ghibli style with soft colors, hand-painted textures, and wildflowers in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04442150518298149 + ], + "opposite_logprob_tokens": [ + -3.169421434402466 + ], + "logprob_reward": -0.04442150518298149, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04442150518298149, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3693539798259735, + 0.3698350489139557, + 0.10775915533304214, + 0.00765052018687129, + 0.0067768157459795475, + 0.04712265357375145, + 0.02768722176551819, + 0.02166573703289032, + 0.013625952415168285, + 0.028522945940494537 + ], + "credit_map": [ + 0.058116424828767776, + 0.0329890176653862, + 0.04246307164430618, + 0.03411920741200447, + 0.055791232734918594, + 0.03893672674894333, + 0.05115874856710434, + 0.05577950179576874, + 0.14514517784118652, + 0.01608903519809246, + 0.021660540252923965, + 0.009593892842531204, + 0.01524457149207592, + 0.025161083787679672, + 0.012802798300981522, + 0.12413791567087173, + 0.02626688964664936, + 0.009431558661162853, + 0.010702679865062237, + 0.006797364912927151, + 0.007936865091323853, + 0.010132507421076298, + 0.009312734939157963, + 0.02717854455113411, + 0.0009713716572150588, + 0.0013433974236249924, + 0.0008616215200163424, + 0.0006554717547260225, + 0.0010121912928298116, + 0.0008262564660981297, + 0.0010565767297521234, + 0.0009236331097781658, + 0.001097633969038725, + 0.0004559322551358491, + 0.0011539506958797574, + 0.0007433599093928933, + 0.0008888608426786959, + 0.0007215400692075491, + 0.0006836365791969001, + 0.0010319009888917208, + 0.016786647960543633, + 0.0026619804557412863, + 0.003201041603460908, + 0.0021042826119810343, + 0.0026379369664937258, + 0.003585915081202984, + 0.002790180267766118, + 0.013354665599763393, + 0.003384958254173398, + 0.003578290343284607, + 0.005345179233700037, + 0.0025983264204114676, + 0.0029314388521015644, + 0.0037363036535680294, + 0.0026545801665633917, + 0.003458143677562475, + 0.00311865983530879, + 0.0022053553257137537, + 0.0037466867361217737, + 0.00241289590485394, + 0.0025031836703419685, + 0.002525137737393379, + 0.00222653872333467, + 0.002927277237176895, + 0.0022793570533394814, + 0.0012276415945962071, + 0.0018179884646087885, + 0.0010856345761567354, + 0.0017945178551599383, + 0.0013759838184341788, + 0.002350666094571352, + 0.001694162143394351, + 0.0077018653973937035, + 0.0018556921277195215, + 0.0024419797118753195, + 0.008810488507151604, + 0.0020909367594867945, + 0.0015609989641234279, + 0.002542272675782442, + 0.0015187105163931847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.04442150518298149 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush green meadow during a summer afternoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a snowman mascot composed of two round snowballs with a golden crown, coal-black eyes, a carrot nose, and a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowman mascot waving an ice cream scepter with a waffle cone handle and a strawberry-pink scoop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate frost crystals rapidly condense, bloom, and spread from the snowman's surface while refracting prismatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic Ghibli style with soft colors, hand-painted textures, and wildflowers in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005955095402896404 + ], + "opposite_logprob_tokens": [ + -5.130955219268799 + ], + "logprob_reward": -0.005955095402896404, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005955095402896404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28770697116851807, + 0.20640049874782562, + 0.08505851775407791, + 0.06826569885015488, + 0.06154196336865425, + 0.06234820559620857, + 0.03437052667140961, + 0.0911535695195198, + 0.05202546715736389, + 0.0511285774409771 + ], + "credit_map": [ + 0.045041922479867935, + 0.02932821400463581, + 0.05186207965016365, + 0.027736812829971313, + 0.03021020069718361, + 0.03746432065963745, + 0.021878520026803017, + 0.044184908270835876, + 0.07016295939683914, + 0.01057436317205429, + 0.013012039475142956, + 0.00800321064889431, + 0.015894368290901184, + 0.01618865318596363, + 0.012315304949879646, + 0.06024960055947304, + 0.013905167579650879, + 0.010820447467267513, + 0.008262837305665016, + 0.008587987162172794, + 0.009960154071450233, + 0.008626637980341911, + 0.01450322289019823, + 0.010392061434686184, + 0.010145955719053745, + 0.009466040879487991, + 0.014010226354002953, + 0.008316770195960999, + 0.007714967709034681, + 0.004662038292735815, + 0.006964087951928377, + 0.0069856117479503155, + 0.007043847814202309, + 0.012974971905350685, + 0.005797806195914745, + 0.005782075691968203, + 0.008459476754069328, + 0.007415431085973978, + 0.006629125215113163, + 0.007439229171723127, + 0.014629023149609566, + 0.005639673676341772, + 0.005759070627391338, + 0.005051471758633852, + 0.007047771941870451, + 0.005544865969568491, + 0.0068298159167170525, + 0.011846515350043774, + 0.0040739779360592365, + 0.003275069408118725, + 0.005528359208256006, + 0.002919150749221444, + 0.004665839485824108, + 0.004938208963721991, + 0.004977422766387463, + 0.003992498852312565, + 0.012719565071165562, + 0.01389987114816904, + 0.008077815175056458, + 0.013947137631475925, + 0.013409663923084736, + 0.007525427266955376, + 0.011989399790763855, + 0.009584682062268257, + 0.008236977271735668, + 0.006851255893707275, + 0.005433095619082451, + 0.004722849000245333, + 0.008656969293951988, + 0.004861848894506693, + 0.007665032055228949, + 0.0055974372662603855, + 0.007632777094841003, + 0.006763617042452097, + 0.005879427772015333, + 0.008748364634811878, + 0.005831563379615545, + 0.005310436710715294, + 0.0062870425172150135, + 0.004675347823649645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.005955095402896404 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush green meadow during a summer afternoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a snowman mascot composed of two round snowballs with a golden crown, coal-black eyes, a carrot nose, and a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowman mascot waving an ice cream scepter with a waffle cone handle and a strawberry-pink scoop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate frost crystals rapidly condense, bloom, and spread from the snowman's surface while refracting prismatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic Ghibli style with soft colors, hand-painted textures, and wildflowers in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033701904118061066 + ], + "opposite_logprob_tokens": [ + -3.4087018966674805 + ], + "logprob_reward": -0.033701904118061066, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033701904118061066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15925370156764984, + 0.2592332661151886, + 0.07700622826814651, + 0.08104869723320007, + 0.10900083184242249, + 0.10776013880968094, + 0.02794933132827282, + 0.10437991470098495, + 0.0349515937268734, + 0.039416294544935226 + ], + "credit_map": [ + 0.022751308977603912, + 0.016115745529532433, + 0.03206091746687889, + 0.016016574576497078, + 0.019296804443001747, + 0.020636050030589104, + 0.012209172360599041, + 0.020167116075754166, + 0.10068869590759277, + 0.009871378540992737, + 0.014370442368090153, + 0.006435496266931295, + 0.012856320478022099, + 0.019030656665563583, + 0.009513180702924728, + 0.08646707981824875, + 0.021469082683324814, + 0.005153162870556116, + 0.007213244680315256, + 0.0035092446487396955, + 0.004584027454257011, + 0.010301204398274422, + 0.004566037096083164, + 0.020210225135087967, + 0.01993674784898758, + 0.007128829602152109, + 0.010748296044766903, + 0.003922881558537483, + 0.004597312770783901, + 0.009745880961418152, + 0.00443294970318675, + 0.020535798743367195, + 0.013899828307330608, + 0.02952302061021328, + 0.010886027477681637, + 0.009052515961229801, + 0.012707637622952461, + 0.010651469230651855, + 0.009718009270727634, + 0.012562328018248081, + 0.04085268825292587, + 0.006687829736620188, + 0.007416035048663616, + 0.0035907221026718616, + 0.0036489360500127077, + 0.008422098122537136, + 0.004650009796023369, + 0.032491814345121384, + 0.003297619754448533, + 0.002669251523911953, + 0.006016253959387541, + 0.00215095654129982, + 0.00200884067453444, + 0.0026387041434645653, + 0.00318967760540545, + 0.005978026892989874, + 0.013594201765954494, + 0.014760185964405537, + 0.009003179147839546, + 0.015575631521642208, + 0.016382388770580292, + 0.009088359773159027, + 0.01326440367847681, + 0.01271156594157219, + 0.005877859890460968, + 0.0041914451867341995, + 0.005286829546093941, + 0.0035022266674786806, + 0.003745707683265209, + 0.0037698654923588037, + 0.004473438486456871, + 0.004104220774024725, + 0.005986104719340801, + 0.005117550492286682, + 0.004098576493561268, + 0.007315955124795437, + 0.004712147172540426, + 0.0034674941562116146, + 0.005348851904273033, + 0.003369614714756608 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.033701904118061066 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush green meadow during a summer afternoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a snowman mascot composed of two round snowballs with a golden crown, coal-black eyes, a carrot nose, and a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowman mascot waving an ice cream scepter with a waffle cone handle and a strawberry-pink scoop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate frost crystals rapidly condense, bloom, and spread from the snowman's surface while refracting prismatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic Ghibli style with soft colors, hand-painted textures, and wildflowers in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12705613672733307 + ], + "opposite_logprob_tokens": [ + -2.127056121826172 + ], + "logprob_reward": -0.12705613672733307, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12705613672733307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1806568205356598, + 0.168559268116951, + 0.08358422666788101, + 0.031000742688775063, + 0.04240741208195686, + 0.08361247181892395, + 0.05402670055627823, + 0.11072743684053421, + 0.15323506295681, + 0.09218984097242355 + ], + "credit_map": [ + 0.02538350783288479, + 0.016228655353188515, + 0.04304191842675209, + 0.016456415876746178, + 0.0151608195155859, + 0.027736570686101913, + 0.012320113368332386, + 0.024328820407390594, + 0.051134828478097916, + 0.010529099963605404, + 0.013486510142683983, + 0.010185235179960728, + 0.013908285647630692, + 0.012522758916020393, + 0.010552720166742802, + 0.046239834278821945, + 0.024241089820861816, + 0.006217496935278177, + 0.007651739288121462, + 0.004351749550551176, + 0.00442476524040103, + 0.009551494382321835, + 0.004814537707716227, + 0.02233136259019375, + 0.004945655819028616, + 0.003417344531044364, + 0.006882995367050171, + 0.0028775546234101057, + 0.002421896206215024, + 0.0030287334229797125, + 0.003069150261580944, + 0.0043574124574661255, + 0.00692601827904582, + 0.0036167551297694445, + 0.009718158282339573, + 0.0023864354006946087, + 0.0030096557457000017, + 0.00677439384162426, + 0.0036356099881231785, + 0.006340385414659977, + 0.02358683943748474, + 0.006182454526424408, + 0.005504889879375696, + 0.005234901327639818, + 0.006570917088538408, + 0.00900417659431696, + 0.007166893221437931, + 0.02036140486598015, + 0.008297167718410492, + 0.0056929499842226505, + 0.00964974146336317, + 0.005834667943418026, + 0.0048686605878174305, + 0.007847810164093971, + 0.005485511850565672, + 0.0063501885160803795, + 0.014842611737549305, + 0.015692032873630524, + 0.011077890172600746, + 0.017836883664131165, + 0.015643419697880745, + 0.00904055591672659, + 0.01386233326047659, + 0.012731705792248249, + 0.025070760399103165, + 0.018067646771669388, + 0.018731048330664635, + 0.013621502555906773, + 0.020937662571668625, + 0.01932406611740589, + 0.021082065999507904, + 0.016400307416915894, + 0.013115670531988144, + 0.011892721988260746, + 0.008622551336884499, + 0.01796581596136093, + 0.011479394510388374, + 0.008208525367081165, + 0.013379845768213272, + 0.007525315973907709 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.12705613672733307 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush green meadow during a summer afternoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a snowman mascot composed of two round snowballs with a golden crown, coal-black eyes, a carrot nose, and a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowman mascot waving an ice cream scepter with a waffle cone handle and a strawberry-pink scoop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate frost crystals rapidly condense, bloom, and spread from the snowman's surface while refracting prismatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic Ghibli style with soft colors, hand-painted textures, and wildflowers in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305906221270561 + ], + "opposite_logprob_tokens": [ + -3.1680591106414795 + ], + "logprob_reward": -0.04305906221270561, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305906221270561, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13033130764961243, + 0.3211050033569336, + 0.026097316294908524, + 0.058046676218509674, + 0.06360423564910889, + 0.2642505168914795, + 0.05124188959598541, + 0.016448456794023514, + 0.010748128406703472, + 0.05812650918960571 + ], + "credit_map": [ + 0.024035291746258736, + 0.01129941362887621, + 0.019543878734111786, + 0.00996383186429739, + 0.009438952431082726, + 0.019755670800805092, + 0.008805783465504646, + 0.02748849242925644, + 0.12919911742210388, + 0.011387870647013187, + 0.018231293186545372, + 0.006133724935352802, + 0.012492896988987923, + 0.02268935926258564, + 0.008394081145524979, + 0.11257664859294891, + 0.004958075005561113, + 0.003091279650107026, + 0.0025636281352490187, + 0.002165180863812566, + 0.002916729310527444, + 0.002253446029499173, + 0.003484111512079835, + 0.004664866253733635, + 0.017110157757997513, + 0.0034868279471993446, + 0.007970601320266724, + 0.003271647961810231, + 0.002514360472559929, + 0.004902878310531378, + 0.0029509281739592552, + 0.01583927869796753, + 0.00972255039960146, + 0.007642502896487713, + 0.009842777624726295, + 0.005534414667636156, + 0.008486524224281311, + 0.007738618645817041, + 0.004652233794331551, + 0.009984617121517658, + 0.09737634658813477, + 0.013536020182073116, + 0.016744399443268776, + 0.01016833633184433, + 0.01408572681248188, + 0.02000410109758377, + 0.014559146948158741, + 0.0777764618396759, + 0.005505749024450779, + 0.005442055873572826, + 0.008125120773911476, + 0.004336882848292589, + 0.0045088063925504684, + 0.008869215846061707, + 0.004477467387914658, + 0.009976595640182495, + 0.0025457299780100584, + 0.0017221231246367097, + 0.002232917584478855, + 0.0016077904729172587, + 0.0022675313521176577, + 0.0019033628050237894, + 0.0020351761486381292, + 0.00213382625952363, + 0.0015650114510208368, + 0.0008845663978718221, + 0.0011974076041951776, + 0.0009263388928957283, + 0.0010368952061980963, + 0.00194908631965518, + 0.001658272580243647, + 0.0015305504202842712, + 0.016147391870617867, + 0.004033938981592655, + 0.004557128995656967, + 0.017921457067131996, + 0.004375156480818987, + 0.002749713836237788, + 0.005278016906231642, + 0.0030637120362371206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.04305906221270561 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_3/index_tensor([7883, 7883])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_3/index_tensor([7883, 7883])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..446bcfd93cf9abb36182dff0b7e7589d53261226 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_3/index_tensor([7883, 7883])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up, static shot depicts a young woman, her head tilted noticeably to the left, her features intensely focused as she applies makeup. Her dark brown hair is pulled back from her face, allowing a few fine strands to fall at her temples. The scene is bathed in the bright, even light of a vanity mirror, which casts a soft glow on her clear skin, emphasizing her delicate features. Her eyes are wide open, looking down and out of the frame, her right hand steadily grasping a slender, black-handled makeup brush. The fine bristles at the tip are coated with dark eyeliner, and she carefully draws a precise line along her lower eyelid. As she executes this delicate action with unwavering focus, her mouth unconsciously opens and closes in subtle, rhythmic movements, a physical expression of her deep concentration. The background, a soft, blurred, neutral-toned bathroom, creates a shallow depth of field, isolating her movements and evoking an intimate, tranquil atmosphere centered on the quiet ritual of self-adornment.", + "seed": 29077, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_3/index_7883_seed_29077_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.023828327283263208, + "vlm_logprob": -0.023828327283263208, + "vlm_yes_no_margin": 3.8749999523162844, + "vlm_token_credit": -0.023828327283263208, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a young woman with her head tilted noticeably to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's dark brown hair pulled back with fine strands falling at her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a slender, black-handled makeup brush with dark eyeliner on the bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman drawing a precise line of eyeliner along her lower eyelid while her mouth opens and closes rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, blurred, neutral-toned bathroom creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020405951887369156 + ], + "opposite_logprob_tokens": [ + -4.0204057693481445 + ], + "logprob_reward": -0.020405951887369156, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.020405951887369156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05436350405216217, + 0.06748984754085541, + 0.6088681817054749, + 0.1176389679312706, + 0.022797930985689163, + 0.01791285164654255, + 0.007854935713112354, + 0.038524702191352844, + 0.039941441267728806, + 0.024607708677649498 + ], + "credit_map": [ + 0.0078010051511228085, + 0.006892369594424963, + 0.008233826607465744, + 0.01045936904847622, + 0.006887670606374741, + 0.0038132283370941877, + 0.006312247831374407, + 0.003963785711675882, + 0.016266969963908195, + 0.008686995133757591, + 0.016908012330532074, + 0.006449599750339985, + 0.004699618089944124, + 0.0034862186294049025, + 0.0075853983871638775, + 0.003407033858820796, + 0.24828210473060608, + 0.03745843097567558, + 0.25107982754707336, + 0.027546832337975502, + 0.008265156298875809, + 0.008804040029644966, + 0.015498696826398373, + 0.011933092027902603, + 0.042838551104068756, + 0.008608436211943626, + 0.04223649203777313, + 0.008105983957648277, + 0.0067084720358252525, + 0.002080019796267152, + 0.0041310545057058334, + 0.002929964568465948, + 0.003049275139346719, + 0.0029162417631596327, + 0.003607865422964096, + 0.005511703435331583, + 0.0027056594844907522, + 0.0009498752770014107, + 0.002593369921669364, + 0.0014639409491792321, + 0.005545682739466429, + 0.0012735523050650954, + 0.006060036364942789, + 0.0016580566298216581, + 0.0009997690794989467, + 0.0005470699397847056, + 0.0009543171036057174, + 0.0008743682410567999, + 0.0018411007476970553, + 0.0007918783230707049, + 0.0018764081178233027, + 0.0007642804994247854, + 0.0007703989394940436, + 0.0003866709885187447, + 0.0007112683961167932, + 0.0007129302830435336, + 0.012881862930953503, + 0.002734887646511197, + 0.013564517721533775, + 0.002193114021793008, + 0.0016558816423639655, + 0.0017021553358063102, + 0.0018068754579871893, + 0.001985405571758747, + 0.013160668313503265, + 0.0027545534539967775, + 0.014144081622362137, + 0.0028484193608164787, + 0.0018953023245558143, + 0.0014277981827035546, + 0.001745052868500352, + 0.0019655651412904263, + 0.004345028195530176, + 0.002927379449829459, + 0.004489337559789419, + 0.0019172878237441182, + 0.004302142653614283, + 0.0017744844080880284, + 0.0027937490958720446, + 0.002058299956843257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.020405951887369156 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a young woman with her head tilted noticeably to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's dark brown hair pulled back with fine strands falling at her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a slender, black-handled makeup brush with dark eyeliner on the bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman drawing a precise line of eyeliner along her lower eyelid while her mouth opens and closes rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, blurred, neutral-toned bathroom creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012554306536912918 + ], + "opposite_logprob_tokens": [ + -4.387554168701172 + ], + "logprob_reward": -0.012554306536912918, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012554306536912918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06826309114694595, + 0.04323304817080498, + 0.431096613407135, + 0.03358502686023712, + 0.0715571716427803, + 0.03069227561354637, + 0.044929925352334976, + 0.05912889540195465, + 0.03103763423860073, + 0.18647636473178864 + ], + "credit_map": [ + 0.019434120506048203, + 0.007059010677039623, + 0.018854811787605286, + 0.005934976972639561, + 0.0046209609135985374, + 0.003310865955427289, + 0.004948067013174295, + 0.004100271966308355, + 0.013534987345337868, + 0.002650266280397773, + 0.013913401402533054, + 0.0024401408154517412, + 0.002592271426692605, + 0.0016303665470331907, + 0.004138647578656673, + 0.0023329625837504864, + 0.1740526556968689, + 0.02681444026529789, + 0.17614072561264038, + 0.018809547647833824, + 0.00753795588389039, + 0.006761687342077494, + 0.0123929837718606, + 0.008586635813117027, + 0.010107796639204025, + 0.0031684350688010454, + 0.010148354806005955, + 0.0025088193360716105, + 0.002950165420770645, + 0.0009450101060792804, + 0.002255974570289254, + 0.0015004657907411456, + 0.024337416514754295, + 0.004524363670498133, + 0.025205116719007492, + 0.007510071620345116, + 0.0032599689438939095, + 0.0014832766028121114, + 0.0031804002355784178, + 0.0020565467420965433, + 0.008182809688150883, + 0.0019914552103728056, + 0.009076761081814766, + 0.003476513084024191, + 0.0022605007980018854, + 0.001234732335433364, + 0.002512620761990547, + 0.0019568796269595623, + 0.009099923074245453, + 0.0041089351288974285, + 0.01137508638203144, + 0.003191129770129919, + 0.007628419436514378, + 0.0024279633071273565, + 0.0034190742298960686, + 0.0036793916951864958, + 0.017566179856657982, + 0.004850359633564949, + 0.018579868599772453, + 0.00550658255815506, + 0.0030780311208218336, + 0.002932637231424451, + 0.0027801997493952513, + 0.0038350331597030163, + 0.009097459726035595, + 0.0025261163245886564, + 0.00975703913718462, + 0.0021273258607834578, + 0.001937319990247488, + 0.0019122922094538808, + 0.001599856186658144, + 0.0020802209619432688, + 0.0707741528749466, + 0.013443012721836567, + 0.07399027794599533, + 0.010740101337432861, + 0.0042497022077441216, + 0.0030313427560031414, + 0.006490080151706934, + 0.0037576700560748577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 4 + ], + "selected_reward": -0.012554306536912918 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a young woman with her head tilted noticeably to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's dark brown hair pulled back with fine strands falling at her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a slender, black-handled makeup brush with dark eyeliner on the bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman drawing a precise line of eyeliner along her lower eyelid while her mouth opens and closes rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, blurred, neutral-toned bathroom creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023308105766773224 + ], + "opposite_logprob_tokens": [ + -3.773308038711548 + ], + "logprob_reward": -0.023308105766773224, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023308105766773224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04372510313987732, + 0.012800063006579876, + 0.7142518162727356, + 0.024695497006177902, + 0.06494365632534027, + 0.015272854827344418, + 0.04052140936255455, + 0.015148807317018509, + 0.0035996970254927874, + 0.06504108011722565 + ], + "credit_map": [ + 0.014988572336733341, + 0.0039590755477547646, + 0.014875277876853943, + 0.002837517997249961, + 0.0021787623409181833, + 0.0012693721801042557, + 0.0024173990823328495, + 0.0011991302017122507, + 0.003922358620911837, + 0.0009588193497620523, + 0.0039081755094230175, + 0.0008851214661262929, + 0.0008706531370989978, + 0.0005039824172854424, + 0.0010982692474499345, + 0.0006526862271130085, + 0.29150742292404175, + 0.043967921286821365, + 0.2942293882369995, + 0.030985144898295403, + 0.01025295164436102, + 0.010556253604590893, + 0.019120406359434128, + 0.013632419519126415, + 0.007554837502539158, + 0.0023001155350357294, + 0.007596944458782673, + 0.0023165575694292784, + 0.0020616475958377123, + 0.0006841430440545082, + 0.0012701962841674685, + 0.0009110586252063513, + 0.019983893260359764, + 0.004713782574981451, + 0.020535694435238838, + 0.007490483112633228, + 0.004351378884166479, + 0.0016370425000786781, + 0.003973300568759441, + 0.002258093561977148, + 0.002423941856250167, + 0.0019184132106602192, + 0.002122545149177313, + 0.0019068507244810462, + 0.0022190166637301445, + 0.0009545036009512842, + 0.002253368031233549, + 0.0014742190251126885, + 0.007994960062205791, + 0.0037992866709828377, + 0.010041079483926296, + 0.0029985536821186543, + 0.006628726609051228, + 0.0022268956527113914, + 0.0033128689974546432, + 0.003519040998071432, + 0.003987884148955345, + 0.0011805803515017033, + 0.004230117425322533, + 0.0011328064138069749, + 0.0010690937051549554, + 0.0013171447208151221, + 0.0007834310526959598, + 0.0014477517688646913, + 0.0007117537898011506, + 0.0003458926221355796, + 0.0007604114362038672, + 0.0003600136551540345, + 0.0003546759835444391, + 0.0003864494210574776, + 0.0002345169777981937, + 0.0004459837218746543, + 0.024568844586610794, + 0.004603578709065914, + 0.02559763379395008, + 0.003939990419894457, + 0.001570913358591497, + 0.001084755058400333, + 0.002343745669350028, + 0.001331626670435071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 4 + ], + "selected_reward": -0.023308105766773224 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a young woman with her head tilted noticeably to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's dark brown hair pulled back with fine strands falling at her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a slender, black-handled makeup brush with dark eyeliner on the bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman drawing a precise line of eyeliner along her lower eyelid while her mouth opens and closes rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, blurred, neutral-toned bathroom creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014202939346432686 + ], + "opposite_logprob_tokens": [ + -4.264203071594238 + ], + "logprob_reward": -0.014202939346432686, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202939346432686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16293303668498993, + 0.024096861481666565, + 0.3560052216053009, + 0.06827162206172943, + 0.11050645262002945, + 0.07126573473215103, + 0.04735046997666359, + 0.05396512150764465, + 0.039498470723629, + 0.06610699743032455 + ], + "credit_map": [ + 0.02604566514492035, + 0.02431262470781803, + 0.019633648917078972, + 0.014266462996602058, + 0.03656700998544693, + 0.010971907526254654, + 0.018744131550192833, + 0.012391593307256699, + 0.0031358811538666487, + 0.005026001483201981, + 0.0028898955788463354, + 0.003046605736017227, + 0.002976131159812212, + 0.0017335352022200823, + 0.0031209085136651993, + 0.0021679026540368795, + 0.1444077342748642, + 0.022327130660414696, + 0.1458873301744461, + 0.015717413276433945, + 0.005441506393253803, + 0.005546329077333212, + 0.009604010730981827, + 0.007073764689266682, + 0.02586551010608673, + 0.004412982612848282, + 0.02507133223116398, + 0.004236873704940081, + 0.003211325267329812, + 0.0012208611005917192, + 0.002440375043079257, + 0.0018123664194718003, + 0.028434820473194122, + 0.00974187906831503, + 0.028574194759130478, + 0.01821768842637539, + 0.0087577560916543, + 0.0036801553796976805, + 0.008410310372710228, + 0.004689646884799004, + 0.01452814880758524, + 0.007584914565086365, + 0.015737688168883324, + 0.01609082892537117, + 0.005459133069962263, + 0.002444469602778554, + 0.005716510582715273, + 0.00370404333807528, + 0.010499333962798119, + 0.0042680539190769196, + 0.013020538724958897, + 0.003429286414757371, + 0.006377771962434053, + 0.002493172185495496, + 0.003643130185082555, + 0.0036191835533827543, + 0.015134548768401146, + 0.0038125524297356606, + 0.017603328451514244, + 0.004458224400877953, + 0.0034851098898798227, + 0.0025405767373740673, + 0.002949431538581848, + 0.003981348592787981, + 0.012441292405128479, + 0.003165088128298521, + 0.013348221778869629, + 0.0029525081627070904, + 0.0024891216307878494, + 0.001087771262973547, + 0.0023976273369044065, + 0.0016168418806046247, + 0.01094901654869318, + 0.0097779780626297, + 0.012046386487782001, + 0.011818202212452888, + 0.006985062267631292, + 0.003952796570956707, + 0.0062484159134328365, + 0.004329141695052385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.014202939346432686 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a young woman with her head tilted noticeably to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's dark brown hair pulled back with fine strands falling at her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a slender, black-handled makeup brush with dark eyeliner on the bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman drawing a precise line of eyeliner along her lower eyelid while her mouth opens and closes rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, blurred, neutral-toned bathroom creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867033287882805 + ], + "opposite_logprob_tokens": [ + -3.0486702919006348 + ], + "logprob_reward": -0.04867033287882805, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867033287882805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13701124489307404, + 0.07254951447248459, + 0.1381610780954361, + 0.09824887663125992, + 0.055449459701776505, + 0.04412875697016716, + 0.010444408282637596, + 0.020627031102776527, + 0.012344634160399437, + 0.4110349714756012 + ], + "credit_map": [ + 0.044669825583696365, + 0.01156085915863514, + 0.046457722783088684, + 0.008296238258481026, + 0.009136797860264778, + 0.004266871139407158, + 0.007760701701045036, + 0.004862216766923666, + 0.02081129141151905, + 0.007596931420266628, + 0.0215885229408741, + 0.006075854413211346, + 0.0058272164314985275, + 0.002365772146731615, + 0.004615744575858116, + 0.0036681839264929295, + 0.054601989686489105, + 0.008728796616196632, + 0.05452985689043999, + 0.008000679314136505, + 0.002883097855374217, + 0.002280136104673147, + 0.004038696642965078, + 0.0030978128779679537, + 0.03582826629281044, + 0.007094324566423893, + 0.035749804228544235, + 0.007536215707659721, + 0.0043801055289804935, + 0.0017064472194761038, + 0.003531024092808366, + 0.0024226787500083447, + 0.013789815828204155, + 0.004956530872732401, + 0.013965238817036152, + 0.009625515900552273, + 0.00403505889698863, + 0.002011550823226571, + 0.00446150079369545, + 0.0026042438112199306, + 0.012133372947573662, + 0.003098145592957735, + 0.01376610342413187, + 0.006530153565108776, + 0.0024081645533442497, + 0.0016465107910335064, + 0.0025265247095376253, + 0.002019780222326517, + 0.0014229113003239036, + 0.001564209465868771, + 0.0012531377142295241, + 0.0018795664655044675, + 0.0014992300420999527, + 0.0006047303904779255, + 0.0012345949653536081, + 0.0009860279969871044, + 0.005928123835474253, + 0.0015876626130193472, + 0.0065424018539488316, + 0.00206251023337245, + 0.001180446706712246, + 0.0006972446572035551, + 0.0011711114784702659, + 0.0014575289096683264, + 0.0018595483852550387, + 0.0015291665913537145, + 0.0020947318989783525, + 0.0014484747080132365, + 0.0017473826883360744, + 0.00101634138263762, + 0.0014409245923161507, + 0.0012080642627552152, + 0.15655547380447388, + 0.02928234450519085, + 0.16243714094161987, + 0.024300310760736465, + 0.009477611631155014, + 0.00612094858661294, + 0.014685152098536491, + 0.008175999857485294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.04867033287882805 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot depicts a young woman, her head tilted noticeably to the left, her features intensely focused as she applies makeup. Her dark brown hair is pulled back from her face, allowing a few fine strands to fall at her temples. The scene is bathed in the bright, even light of a vanity mirror, which casts a soft glow on her clear skin, emphasizing her delicate features. Her eyes are wide open, looking down and out of the frame, her right hand steadily grasping a slender, black-handled makeup brush. The fine bristles at the tip are coated with dark eyeliner, and she carefully draws a precise line along her lower eyelid. As she executes this delicate action with unwavering focus, her mouth unconsciously opens and closes in subtle, rhythmic movements, a physical expression of her deep concentration. The background, a soft, blurred, neutral-toned bathroom, creates a shallow depth of field, isolating her movements and evoking an intimate, tranquil atmosphere centered on the quiet ritual of self-adornment.", + "seed": 29078, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_3/index_7883_seed_29078_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.020548623986542225, + "vlm_logprob": -0.020548623986542225, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.020548623986542225, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a young woman with her head tilted noticeably to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's dark brown hair pulled back with fine strands falling at her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a slender, black-handled makeup brush with dark eyeliner on the bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman drawing a precise line of eyeliner along her lower eyelid while her mouth opens and closes rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, blurred, neutral-toned bathroom creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009171971119940281 + ], + "opposite_logprob_tokens": [ + -5.134171962738037 + ], + "logprob_reward": -0.009171971119940281, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.009171971119940281, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20803195238113403, + 0.07011719793081284, + 0.05316454917192459, + 0.09818719327449799, + 0.03238537535071373, + 0.18320950865745544, + 0.10360223799943924, + 0.10267718881368637, + 0.08373183757066727, + 0.0648929551243782 + ], + "credit_map": [ + 0.02293255925178528, + 0.0255989171564579, + 0.031090978533029556, + 0.03820823132991791, + 0.016228437423706055, + 0.018071522936224937, + 0.03808685392141342, + 0.017814451828598976, + 0.008223967626690865, + 0.005606635473668575, + 0.010762723162770271, + 0.009611149318516254, + 0.005845625419169664, + 0.013346527703106403, + 0.00982336513698101, + 0.006897208746522665, + 0.006093350239098072, + 0.006538706365972757, + 0.008929918520152569, + 0.009594112634658813, + 0.004032617900520563, + 0.008123726584017277, + 0.0058668251149356365, + 0.003985293675214052, + 0.010814274661242962, + 0.011854317039251328, + 0.013400853611528873, + 0.01626553013920784, + 0.006945367902517319, + 0.009735471569001675, + 0.022840294986963272, + 0.006331090349704027, + 0.0038654878735542297, + 0.003629745217040181, + 0.004565283190459013, + 0.006503819953650236, + 0.0022688463795930147, + 0.0031222673133015633, + 0.005838251207023859, + 0.002591673983260989, + 0.017848040908575058, + 0.021838456392288208, + 0.033577337861061096, + 0.029390467330813408, + 0.014839095994830132, + 0.03808934986591339, + 0.014180228114128113, + 0.013446533121168613, + 0.010455570183694363, + 0.018063681200146675, + 0.012533881701529026, + 0.019743546843528748, + 0.009523660875856876, + 0.007481500972062349, + 0.016732819378376007, + 0.00906757079064846, + 0.013842858374118805, + 0.011321335099637508, + 0.015866540372371674, + 0.014142335392534733, + 0.0076728747226297855, + 0.008806225843727589, + 0.02192104049026966, + 0.009103972464799881, + 0.012334481813013554, + 0.010000072419643402, + 0.010484742000699043, + 0.012295104563236237, + 0.007608113810420036, + 0.0061750453896820545, + 0.016777368262410164, + 0.008056910708546638, + 0.006563428323715925, + 0.011320069432258606, + 0.009024742059409618, + 0.012048845179378986, + 0.006778558250516653, + 0.004550405777990818, + 0.009867665357887745, + 0.004739245865494013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.009171971119940281 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a young woman with her head tilted noticeably to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's dark brown hair pulled back with fine strands falling at her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a slender, black-handled makeup brush with dark eyeliner on the bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman drawing a precise line of eyeliner along her lower eyelid while her mouth opens and closes rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, blurred, neutral-toned bathroom creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011093014851212502 + ], + "opposite_logprob_tokens": [ + -4.5110931396484375 + ], + "logprob_reward": -0.011093014851212502, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011093014851212502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20393846929073334, + 0.15045534074306488, + 0.11907881498336792, + 0.10239816457033157, + 0.04446123540401459, + 0.033040937036275864, + 0.06981444358825684, + 0.1484638750553131, + 0.03709453344345093, + 0.09125426411628723 + ], + "credit_map": [ + 0.02136237546801567, + 0.03252165764570236, + 0.02391716279089451, + 0.03785378858447075, + 0.016833683475852013, + 0.025015562772750854, + 0.021888740360736847, + 0.024545492604374886, + 0.01940816268324852, + 0.017040861770510674, + 0.02443191036581993, + 0.028334490954875946, + 0.008950297720730305, + 0.01567952334880829, + 0.024643730372190475, + 0.011966364458203316, + 0.017717190086841583, + 0.009473849087953568, + 0.030747363343834877, + 0.01279209740459919, + 0.007110449951142073, + 0.014728773385286331, + 0.014807379804551601, + 0.011701718904078007, + 0.009956641122698784, + 0.010378818027675152, + 0.013622443191707134, + 0.011639918200671673, + 0.009533734992146492, + 0.018488071858882904, + 0.022438520565629005, + 0.006340018939226866, + 0.006298723164945841, + 0.0052354056388139725, + 0.007082202937453985, + 0.007068686652928591, + 0.0032374209258705378, + 0.004366675391793251, + 0.007473176810890436, + 0.003698941320180893, + 0.0036769709549844265, + 0.002804507501423359, + 0.004154672846198082, + 0.004217514768242836, + 0.003499978221952915, + 0.007738857995718718, + 0.0035908338613808155, + 0.0033576011192053556, + 0.00810004211962223, + 0.006748960353434086, + 0.009534562937915325, + 0.00756342988461256, + 0.005270924884825945, + 0.0076400283724069595, + 0.01471666805446148, + 0.010239833034574986, + 0.018422691151499748, + 0.020411530509591103, + 0.018717078492045403, + 0.024592814967036247, + 0.015269937925040722, + 0.015189647674560547, + 0.02101323939859867, + 0.01484693307429552, + 0.004514225758612156, + 0.005421304143965244, + 0.004632849711924791, + 0.007281960919499397, + 0.004221330396831036, + 0.002829352393746376, + 0.005532058421522379, + 0.0026614500675350428, + 0.012813166715204716, + 0.011553525924682617, + 0.013550308533012867, + 0.01082668174058199, + 0.009061320684850216, + 0.007727313321083784, + 0.01109932642430067, + 0.01462261751294136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.011093014851212502 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a young woman with her head tilted noticeably to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's dark brown hair pulled back with fine strands falling at her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a slender, black-handled makeup brush with dark eyeliner on the bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman drawing a precise line of eyeliner along her lower eyelid while her mouth opens and closes rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, blurred, neutral-toned bathroom creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016085024923086166 + ], + "opposite_logprob_tokens": [ + -4.141085147857666 + ], + "logprob_reward": -0.016085024923086166, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016085024923086166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12558682262897491, + 0.08848696947097778, + 0.07249866425991058, + 0.16245819628238678, + 0.09060660004615784, + 0.059698138386011124, + 0.16496683657169342, + 0.13939468562602997, + 0.03712615370750427, + 0.05917702242732048 + ], + "credit_map": [ + 0.013285920023918152, + 0.019815880805253983, + 0.01606481708586216, + 0.025738276541233063, + 0.011860798113048077, + 0.01124994084239006, + 0.0161777101457119, + 0.011393473483622074, + 0.009351789951324463, + 0.012060591951012611, + 0.012876576744019985, + 0.017482858151197433, + 0.006904608570039272, + 0.013253125362098217, + 0.011308105662465096, + 0.005249311216175556, + 0.00949943158775568, + 0.00639825314283371, + 0.01153063215315342, + 0.008627654053270817, + 0.0053849839605391026, + 0.009744290262460709, + 0.013303499668836594, + 0.008009918965399265, + 0.015517937950789928, + 0.01867140829563141, + 0.019167665392160416, + 0.01945454627275467, + 0.01306332740932703, + 0.016145754605531693, + 0.045803047716617584, + 0.014634496532380581, + 0.007945965975522995, + 0.010010569356381893, + 0.016087893396615982, + 0.01066373661160469, + 0.005149237811565399, + 0.013809777796268463, + 0.02048312872648239, + 0.006456281989812851, + 0.006248772609978914, + 0.004370917100459337, + 0.008425900712609291, + 0.008151705376803875, + 0.004522965755313635, + 0.009946433827280998, + 0.010694674216210842, + 0.007336764130741358, + 0.01944015733897686, + 0.011001473292708397, + 0.02748359739780426, + 0.016222259029746056, + 0.010812700726091862, + 0.015073392540216446, + 0.04258386790752411, + 0.022349392995238304, + 0.016972042620182037, + 0.011514747515320778, + 0.019982855767011642, + 0.016898859292268753, + 0.009394537657499313, + 0.012437841854989529, + 0.040414731949567795, + 0.011779061518609524, + 0.0041315373964607716, + 0.004193867556750774, + 0.004783684387803078, + 0.006449922453612089, + 0.004090953152626753, + 0.002371774287894368, + 0.008328843861818314, + 0.0027755694463849068, + 0.005522408057004213, + 0.006369763053953648, + 0.008420664817094803, + 0.009631932713091373, + 0.005989639554172754, + 0.0038717787247151136, + 0.013320430181920528, + 0.00605040043592453 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 7 + ], + "selected_reward": -0.016085024923086166 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a young woman with her head tilted noticeably to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's dark brown hair pulled back with fine strands falling at her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a slender, black-handled makeup brush with dark eyeliner on the bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman drawing a precise line of eyeliner along her lower eyelid while her mouth opens and closes rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, blurred, neutral-toned bathroom creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330356277525425 + ], + "opposite_logprob_tokens": [ + -3.773303508758545 + ], + "logprob_reward": -0.02330356277525425, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330356277525425, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22492225468158722, + 0.0726814791560173, + 0.06453310698270798, + 0.197017103433609, + 0.048881370574235916, + 0.04287771135568619, + 0.06006884574890137, + 0.1577865332365036, + 0.05511796101927757, + 0.07611366361379623 + ], + "credit_map": [ + 0.035752587020397186, + 0.027060000225901604, + 0.04300471022725105, + 0.026645870879292488, + 0.013557204976677895, + 0.015651656314730644, + 0.048969101160764694, + 0.0142811369150877, + 0.00706188939511776, + 0.007224877364933491, + 0.012235965579748154, + 0.00978853553533554, + 0.005045588593930006, + 0.009466759860515594, + 0.013189798220992088, + 0.008668062277138233, + 0.009626111947000027, + 0.005215553101152182, + 0.013735901564359665, + 0.006998908240348101, + 0.004004724323749542, + 0.006127018015831709, + 0.011875052005052567, + 0.006949834059923887, + 0.017453797161579132, + 0.013661947101354599, + 0.02996731735765934, + 0.01934163272380829, + 0.01673918031156063, + 0.03993283212184906, + 0.04197632148861885, + 0.01794406957924366, + 0.0051641883328557014, + 0.0044184112921357155, + 0.00952200498431921, + 0.005269583314657211, + 0.0034836202394217253, + 0.009505411610007286, + 0.007021557539701462, + 0.00449659489095211, + 0.0039868904277682304, + 0.0034795051906257868, + 0.007220495492219925, + 0.0038649733178317547, + 0.004399883560836315, + 0.011401587165892124, + 0.003992100711911917, + 0.0045322757214307785, + 0.007226561661809683, + 0.008108017034828663, + 0.009981982409954071, + 0.01067067589610815, + 0.004484290722757578, + 0.0055985585786402225, + 0.009265781380236149, + 0.004732979927212, + 0.01804601401090622, + 0.016377823427319527, + 0.021334411576390266, + 0.018783194944262505, + 0.013217554427683353, + 0.017515989020466805, + 0.03838993236422539, + 0.014121614396572113, + 0.006653887219727039, + 0.007799260318279266, + 0.0072165969759225845, + 0.010350333526730537, + 0.006113376468420029, + 0.004422548692673445, + 0.008392727933824062, + 0.004169232677668333, + 0.0097105847671628, + 0.009746299125254154, + 0.011004094034433365, + 0.013553368858993053, + 0.005430948454886675, + 0.004856408573687077, + 0.012002509087324142, + 0.00980945024639368 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.02330356277525425 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a young woman with her head tilted noticeably to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's dark brown hair pulled back with fine strands falling at her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a slender, black-handled makeup brush with dark eyeliner on the bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman drawing a precise line of eyeliner along her lower eyelid while her mouth opens and closes rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, blurred, neutral-toned bathroom creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043089546263217926 + ], + "opposite_logprob_tokens": [ + -3.1680896282196045 + ], + "logprob_reward": -0.043089546263217926, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043089546263217926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2037259191274643, + 0.06728997826576233, + 0.06700451672077179, + 0.10262207686901093, + 0.048204146325588226, + 0.0813068374991417, + 0.09319434314966202, + 0.18816129863262177, + 0.06305783987045288, + 0.08543308079242706 + ], + "credit_map": [ + 0.02776985801756382, + 0.01960056647658348, + 0.04370958358049393, + 0.03481845557689667, + 0.012618438340723515, + 0.022441288456320763, + 0.023421330377459526, + 0.01934639737010002, + 0.0081876115873456, + 0.007397804409265518, + 0.008297226391732693, + 0.013279219157993793, + 0.005391768645495176, + 0.007133465260267258, + 0.01021728664636612, + 0.007385593373328447, + 0.009736848063766956, + 0.006734022870659828, + 0.011945532634854317, + 0.009677708148956299, + 0.0039036732632666826, + 0.006748754531145096, + 0.012340675108134747, + 0.005917305592447519, + 0.009466961957514286, + 0.01226532831788063, + 0.014552995562553406, + 0.013294035568833351, + 0.00802561268210411, + 0.017308272421360016, + 0.020723899826407433, + 0.006984968204051256, + 0.006208933889865875, + 0.006131657864898443, + 0.005872105248272419, + 0.005081475712358952, + 0.004703073762357235, + 0.007801935076713562, + 0.006987014319747686, + 0.005417948123067617, + 0.009224221110343933, + 0.008201166056096554, + 0.011939791962504387, + 0.009038010612130165, + 0.006331810727715492, + 0.012955495156347752, + 0.014865024946630001, + 0.008751318790018559, + 0.012870829552412033, + 0.011548045091331005, + 0.020439356565475464, + 0.01570119336247444, + 0.006856725551187992, + 0.007996661588549614, + 0.011521806940436363, + 0.006259722635149956, + 0.0272661279886961, + 0.01783462055027485, + 0.027020927518606186, + 0.01815655082464218, + 0.014764509163796902, + 0.027783449739217758, + 0.02423706091940403, + 0.03109804354608059, + 0.009857266210019588, + 0.008652142249047756, + 0.009045246988534927, + 0.009929200634360313, + 0.00577813433483243, + 0.004673238843679428, + 0.009028811939060688, + 0.006093798670917749, + 0.012974388897418976, + 0.008922908455133438, + 0.014282864518463612, + 0.010961640626192093, + 0.00767818046733737, + 0.007194165140390396, + 0.009235725738108158, + 0.014183208346366882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.043089546263217926 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot depicts a young woman, her head tilted noticeably to the left, her features intensely focused as she applies makeup. Her dark brown hair is pulled back from her face, allowing a few fine strands to fall at her temples. The scene is bathed in the bright, even light of a vanity mirror, which casts a soft glow on her clear skin, emphasizing her delicate features. Her eyes are wide open, looking down and out of the frame, her right hand steadily grasping a slender, black-handled makeup brush. The fine bristles at the tip are coated with dark eyeliner, and she carefully draws a precise line along her lower eyelid. As she executes this delicate action with unwavering focus, her mouth unconsciously opens and closes in subtle, rhythmic movements, a physical expression of her deep concentration. The background, a soft, blurred, neutral-toned bathroom, creates a shallow depth of field, isolating her movements and evoking an intimate, tranquil atmosphere centered on the quiet ritual of self-adornment.", + "seed": 29079, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_3/index_7883_seed_29079_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02255568392574787, + "vlm_logprob": -0.02255568392574787, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.02255568392574787, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a young woman with her head tilted noticeably to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's dark brown hair pulled back with fine strands falling at her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a slender, black-handled makeup brush with dark eyeliner on the bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman drawing a precise line of eyeliner along her lower eyelid while her mouth opens and closes rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, blurred, neutral-toned bathroom creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007939565926790237 + ], + "opposite_logprob_tokens": [ + -5.007939338684082 + ], + "logprob_reward": -0.007939565926790237, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.007939565926790237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17488670349121094, + 0.045612744987010956, + 0.044802773743867874, + 0.029776692390441895, + 0.02996007539331913, + 0.4772917628288269, + 0.027406198903918266, + 0.034247446805238724, + 0.028486648574471474, + 0.10752894729375839 + ], + "credit_map": [ + 0.032292019575834274, + 0.021702008321881294, + 0.033090755343437195, + 0.01905512996017933, + 0.019923219457268715, + 0.011590314097702503, + 0.022455506026744843, + 0.014777746982872486, + 0.005580885335803032, + 0.0056439535692334175, + 0.004979066085070372, + 0.007068526931107044, + 0.006942907348275185, + 0.003161998465657234, + 0.0070674000307917595, + 0.005168003961443901, + 0.0052988710813224316, + 0.008327844552695751, + 0.004589242860674858, + 0.008199319243431091, + 0.004691523965448141, + 0.0034858013968914747, + 0.005886827129870653, + 0.004323345609009266, + 0.003934722859412432, + 0.005221352446824312, + 0.003566912841051817, + 0.005825101397931576, + 0.002781582297757268, + 0.0031137375626713037, + 0.003012279747053981, + 0.002320999978110194, + 0.005614378023892641, + 0.0025573407765477896, + 0.003078363137319684, + 0.003248230554163456, + 0.002693768823519349, + 0.002209076890721917, + 0.004483675118535757, + 0.006075238808989525, + 0.17460231482982635, + 0.01675080694258213, + 0.20231536030769348, + 0.01524812076240778, + 0.010332954116165638, + 0.022605719044804573, + 0.023212295025587082, + 0.012224185280501842, + 0.006060378160327673, + 0.0033182452898472548, + 0.0041527827270329, + 0.0028329796623438597, + 0.003926900681108236, + 0.0019846633076667786, + 0.002735534682869911, + 0.002394712995737791, + 0.003916589077562094, + 0.0059051234275102615, + 0.003786288434639573, + 0.0040793451480567455, + 0.006544659845530987, + 0.003008507424965501, + 0.004208805505186319, + 0.002798124449327588, + 0.004137496463954449, + 0.0037925709038972855, + 0.004331029951572418, + 0.004527350887656212, + 0.00418660556897521, + 0.0022508325055241585, + 0.0035811890847980976, + 0.001679572043940425, + 0.02788248285651207, + 0.008588094264268875, + 0.00995252188295126, + 0.006481412798166275, + 0.01017785631120205, + 0.0038545788265764713, + 0.031506843864917755, + 0.009085152298212051 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.007939565926790237 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a young woman with her head tilted noticeably to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's dark brown hair pulled back with fine strands falling at her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a slender, black-handled makeup brush with dark eyeliner on the bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman drawing a precise line of eyeliner along her lower eyelid while her mouth opens and closes rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, blurred, neutral-toned bathroom creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011089595966041088 + ], + "opposite_logprob_tokens": [ + -4.51108980178833 + ], + "logprob_reward": -0.011089595966041088, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011089595966041088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17229628562927246, + 0.05697794631123543, + 0.036330342292785645, + 0.030936799943447113, + 0.04277859628200531, + 0.48203662037849426, + 0.03683101758360863, + 0.03619588911533356, + 0.02837476320564747, + 0.07724173367023468 + ], + "credit_map": [ + 0.01661013439297676, + 0.01623660698533058, + 0.03978881239891052, + 0.0146145885810256, + 0.031803082674741745, + 0.007501297630369663, + 0.03623571619391441, + 0.009506037458777428, + 0.00728551484644413, + 0.00822876114398241, + 0.007038780488073826, + 0.009138096123933792, + 0.0065548596903681755, + 0.005303591955453157, + 0.008491908200085163, + 0.004936431534588337, + 0.00445175264030695, + 0.006246109493076801, + 0.004183926619589329, + 0.005561911500990391, + 0.004309697076678276, + 0.0038857818581163883, + 0.004332209005951881, + 0.0033589517697691917, + 0.003595268353819847, + 0.005991415120661259, + 0.0032100582029670477, + 0.005970757920295, + 0.004526100121438503, + 0.002341483486816287, + 0.0029885927215218544, + 0.002313125878572464, + 0.0054758815094828606, + 0.005632075946778059, + 0.00471135601401329, + 0.004336666315793991, + 0.010051135905086994, + 0.002682376652956009, + 0.006059070583432913, + 0.003830029396340251, + 0.17757228016853333, + 0.016606684774160385, + 0.2055838704109192, + 0.014884302392601967, + 0.010394909419119358, + 0.023096522316336632, + 0.022469131276011467, + 0.011428892612457275, + 0.004754617810249329, + 0.0062322248704731464, + 0.003938362933695316, + 0.003330803941935301, + 0.009061973541975021, + 0.002712666057050228, + 0.00456413347274065, + 0.0022362382151186466, + 0.003865713020786643, + 0.004701504483819008, + 0.0039335452020168304, + 0.003079470479860902, + 0.009608778171241283, + 0.0023023111280053854, + 0.005375867709517479, + 0.0033286954276263714, + 0.0032980539835989475, + 0.004013121593743563, + 0.0033708158880472183, + 0.002176025416702032, + 0.006649934221059084, + 0.0019221135880798101, + 0.00441801268607378, + 0.0025266860611736774, + 0.01699456200003624, + 0.0066429502330720425, + 0.005570634268224239, + 0.00444911140948534, + 0.006722721271216869, + 0.0033064179588109255, + 0.02589201182126999, + 0.007663325406610966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.011089595966041088 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a young woman with her head tilted noticeably to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's dark brown hair pulled back with fine strands falling at her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a slender, black-handled makeup brush with dark eyeliner on the bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman drawing a precise line of eyeliner along her lower eyelid while her mouth opens and closes rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, blurred, neutral-toned bathroom creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01109513733536005 + ], + "opposite_logprob_tokens": [ + -4.51109504699707 + ], + "logprob_reward": -0.01109513733536005, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01109513733536005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07651089876890182, + 0.1590030938386917, + 0.09634892642498016, + 0.09326706826686859, + 0.019185272976756096, + 0.34843629598617554, + 0.05980951711535454, + 0.046529218554496765, + 0.027047671377658844, + 0.0738620012998581 + ], + "credit_map": [ + 0.01184824574738741, + 0.007189061492681503, + 0.01035026554018259, + 0.008476931601762772, + 0.01617048680782318, + 0.005016551818698645, + 0.01079785730689764, + 0.006661507301032543, + 0.01948131062090397, + 0.02421960048377514, + 0.016248250380158424, + 0.0168499406427145, + 0.026346437633037567, + 0.014995070174336433, + 0.025456475093960762, + 0.015406026504933834, + 0.011033694259822369, + 0.014886794611811638, + 0.010556105524301529, + 0.01429952122271061, + 0.01801471784710884, + 0.006249927915632725, + 0.011200957000255585, + 0.010107218287885189, + 0.010989761911332607, + 0.013539998792111874, + 0.01371595449745655, + 0.010792342945933342, + 0.018712118268013, + 0.006336819846183062, + 0.011871952563524246, + 0.007308128755539656, + 0.0024740993976593018, + 0.0019476263551041484, + 0.0016808001091703773, + 0.0021964889019727707, + 0.0031279344111680984, + 0.0017785365926101804, + 0.0033344433177262545, + 0.0026453444734215736, + 0.125003382563591, + 0.014085556380450726, + 0.14101366698741913, + 0.011797663755714893, + 0.013409829698503017, + 0.016825800761580467, + 0.01661079376935959, + 0.009689632803201675, + 0.010226452723145485, + 0.007080595474690199, + 0.0072031570598483086, + 0.0067366790026426315, + 0.007661108858883381, + 0.003732111072167754, + 0.006940836552530527, + 0.010228578932583332, + 0.008496384136378765, + 0.00575958052650094, + 0.00396893173456192, + 0.004069814458489418, + 0.006503031589090824, + 0.003091588616371155, + 0.00697668083012104, + 0.0076632085256278515, + 0.004205975215882063, + 0.0037953294813632965, + 0.004204367287456989, + 0.0032481830567121506, + 0.0042633116245269775, + 0.0017819813219830394, + 0.0034314983058720827, + 0.0021170282270759344, + 0.014070011675357819, + 0.011465768329799175, + 0.00870484858751297, + 0.008295400068163872, + 0.009843018837273121, + 0.005093904212117195, + 0.01039868127554655, + 0.0059903692454099655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 2 + ], + "selected_reward": -0.01109513733536005 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a young woman with her head tilted noticeably to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's dark brown hair pulled back with fine strands falling at her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a slender, black-handled makeup brush with dark eyeliner on the bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman drawing a precise line of eyeliner along her lower eyelid while her mouth opens and closes rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, blurred, neutral-toned bathroom creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020587796345353127 + ], + "opposite_logprob_tokens": [ + -3.895587682723999 + ], + "logprob_reward": -0.020587796345353127, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020587796345353127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.176026850938797, + 0.09889800101518631, + 0.05007830262184143, + 0.06155006214976311, + 0.05255510285496712, + 0.14507156610488892, + 0.030461370944976807, + 0.1045948714017868, + 0.11135146021842957, + 0.16941238939762115 + ], + "credit_map": [ + 0.01794002577662468, + 0.03383295238018036, + 0.018774032592773438, + 0.02721424028277397, + 0.022236524149775505, + 0.009848824702203274, + 0.028576210141181946, + 0.017604054883122444, + 0.017465418204665184, + 0.016155444085597992, + 0.013638037256896496, + 0.014704128727316856, + 0.010885489173233509, + 0.007705246098339558, + 0.011961517855525017, + 0.006382728926837444, + 0.007733258884400129, + 0.006956308614462614, + 0.0082173440605402, + 0.006624598056077957, + 0.005655136425048113, + 0.0042471024207770824, + 0.005497607868164778, + 0.005146949086338282, + 0.008688950911164284, + 0.011898109689354897, + 0.006363368593156338, + 0.011657655239105225, + 0.006882212590426207, + 0.004553249105811119, + 0.006632617674767971, + 0.00487390486523509, + 0.00680408114567399, + 0.00807913113385439, + 0.0060310810804367065, + 0.00785024743527174, + 0.007213581819087267, + 0.003607693361118436, + 0.00778112793341279, + 0.005188160575926304, + 0.044163044542074203, + 0.008879357948899269, + 0.054852019995450974, + 0.007189854513853788, + 0.008274833671748638, + 0.007761732675135136, + 0.008382621221244335, + 0.0055681029334664345, + 0.004914270713925362, + 0.003702364396303892, + 0.00464634969830513, + 0.003834681585431099, + 0.003874779213219881, + 0.0024715710896998644, + 0.0030646692030131817, + 0.003952688071876764, + 0.013364928774535656, + 0.017399316653609276, + 0.010596306063234806, + 0.011862777173519135, + 0.019781796261668205, + 0.00779181020334363, + 0.013196120038628578, + 0.010601818561553955, + 0.012960504740476608, + 0.014475628733634949, + 0.011956254951655865, + 0.010673017241060734, + 0.02668067254126072, + 0.007326717488467693, + 0.017538651823997498, + 0.0097400126978755, + 0.048062048852443695, + 0.011768316850066185, + 0.012253182008862495, + 0.009263786487281322, + 0.011107855476439, + 0.005334116984158754, + 0.056169744580984116, + 0.015453346073627472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.020587796345353127 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a young woman with her head tilted noticeably to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's dark brown hair pulled back with fine strands falling at her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a slender, black-handled makeup brush with dark eyeliner on the bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman drawing a precise line of eyeliner along her lower eyelid while her mouth opens and closes rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, blurred, neutral-toned bathroom creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062066324055194855 + ], + "opposite_logprob_tokens": [ + -2.8120663166046143 + ], + "logprob_reward": -0.062066324055194855, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062066324055194855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19803990423679352, + 0.045052845031023026, + 0.03057996556162834, + 0.022014647722244263, + 0.039644911885261536, + 0.5212565064430237, + 0.05376376956701279, + 0.024431627243757248, + 0.0384649783372879, + 0.026750823482871056 + ], + "credit_map": [ + 0.03678378835320473, + 0.01666182279586792, + 0.050808001309633255, + 0.01473933830857277, + 0.032706327736377716, + 0.011175925843417645, + 0.02268272638320923, + 0.012481987476348877, + 0.006773270200937986, + 0.0077903601340949535, + 0.005451983772218227, + 0.007112667430192232, + 0.00501592829823494, + 0.004605043679475784, + 0.005441455170512199, + 0.0028621312230825424, + 0.004567560274153948, + 0.0038353484123945236, + 0.005489785689860582, + 0.0037262800615280867, + 0.003170870477333665, + 0.0032622921280562878, + 0.0033522306475788355, + 0.0031755974050611258, + 0.0024076264817267656, + 0.003931224811822176, + 0.002541334368288517, + 0.004011576529592276, + 0.0031757515389472246, + 0.001661120681092143, + 0.0023287804797291756, + 0.0019572321325540543, + 0.005273480899631977, + 0.005004127975553274, + 0.0044639576226472855, + 0.003405694616958499, + 0.008753915317356586, + 0.0028601144440472126, + 0.0059754312969744205, + 0.003908189013600349, + 0.19133418798446655, + 0.01836477778851986, + 0.22270315885543823, + 0.015982355922460556, + 0.01049052830785513, + 0.024614917114377022, + 0.02469288930296898, + 0.013073722831904888, + 0.010714711621403694, + 0.005902471952140331, + 0.008112680166959763, + 0.005871735513210297, + 0.009588716551661491, + 0.0034325034357607365, + 0.005088405683636665, + 0.005052546039223671, + 0.0034320265986025333, + 0.0038145999424159527, + 0.0035846917890012264, + 0.0024901069700717926, + 0.004767975304275751, + 0.0019845503848046064, + 0.0028620439115911722, + 0.0014956352533772588, + 0.004244890995323658, + 0.00513502536341548, + 0.00432712072506547, + 0.002766073914244771, + 0.009711920283734798, + 0.002791636623442173, + 0.005881714168936014, + 0.003606600221246481, + 0.0031022520270198584, + 0.0029347441159188747, + 0.0030786285642534494, + 0.002134566428139806, + 0.0036045664455741644, + 0.0012467848137021065, + 0.007879615761339664, + 0.0027696669567376375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.062066324055194855 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot depicts a young woman, her head tilted noticeably to the left, her features intensely focused as she applies makeup. Her dark brown hair is pulled back from her face, allowing a few fine strands to fall at her temples. The scene is bathed in the bright, even light of a vanity mirror, which casts a soft glow on her clear skin, emphasizing her delicate features. Her eyes are wide open, looking down and out of the frame, her right hand steadily grasping a slender, black-handled makeup brush. The fine bristles at the tip are coated with dark eyeliner, and she carefully draws a precise line along her lower eyelid. As she executes this delicate action with unwavering focus, her mouth unconsciously opens and closes in subtle, rhythmic movements, a physical expression of her deep concentration. The background, a soft, blurred, neutral-toned bathroom, creates a shallow depth of field, isolating her movements and evoking an intimate, tranquil atmosphere centered on the quiet ritual of self-adornment.", + "seed": 29080, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_3/index_7883_seed_29080_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.023114492930471898, + "vlm_logprob": -0.023114492930471898, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.023114492930471898, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a young woman with her head tilted noticeably to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's dark brown hair pulled back with fine strands falling at her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a slender, black-handled makeup brush with dark eyeliner on the bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman drawing a precise line of eyeliner along her lower eyelid while her mouth opens and closes rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, blurred, neutral-toned bathroom creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01937675289809704 + ], + "opposite_logprob_tokens": [ + -4.019376754760742 + ], + "logprob_reward": -0.01937675289809704, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01937675289809704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5583036541938782, + 0.09885593503713608, + 0.04604201391339302, + 0.014276720583438873, + 0.1708507388830185, + 0.0254825409501791, + 0.018708672374486923, + 0.027085434645414352, + 0.02376890368759632, + 0.01662539504468441 + ], + "credit_map": [ + 0.04499885439872742, + 0.05134255811572075, + 0.04556400328874588, + 0.049549251794815063, + 0.11328944563865662, + 0.10750449448823929, + 0.047025345265865326, + 0.09902971982955933, + 0.007032008841633797, + 0.0164068341255188, + 0.009114167653024197, + 0.012927153147757053, + 0.009850529953837395, + 0.01355588249862194, + 0.014069261029362679, + 0.01590009778738022, + 0.0036181097384542227, + 0.006887488998472691, + 0.0035580734256654978, + 0.005240048747509718, + 0.005076051224023104, + 0.006956235971301794, + 0.0065628294833004475, + 0.008143176324665546, + 0.0009287133580073714, + 0.0014890821184962988, + 0.00141623814124614, + 0.001707303337752819, + 0.0014649417717009783, + 0.0019949888810515404, + 0.002532065147534013, + 0.0027433866634964943, + 0.014797376468777657, + 0.015210576355457306, + 0.018968576565384865, + 0.01712830178439617, + 0.018794987350702286, + 0.033474672585725784, + 0.01799461804330349, + 0.034481633454561234, + 0.0015485272742807865, + 0.0028075401205569506, + 0.0025762810837477446, + 0.0024557234719395638, + 0.0029523137491196394, + 0.0038221306167542934, + 0.004684680141508579, + 0.0046353451907634735, + 0.0014687629882246256, + 0.001776959514245391, + 0.003036934183910489, + 0.0024884426966309547, + 0.002016722923144698, + 0.0025239570531994104, + 0.0027470565401017666, + 0.0026498360093683004, + 0.002072045346722007, + 0.0024915377143770456, + 0.0030138243455439806, + 0.0030341921374201775, + 0.002836087020114064, + 0.0037403276655822992, + 0.0035960813984274864, + 0.006301338784396648, + 0.0015018527628853917, + 0.0023939809761941433, + 0.0031008850783109665, + 0.0026207687333226204, + 0.0028913188725709915, + 0.0036489602643996477, + 0.003467148868367076, + 0.004143986850976944, + 0.0008686173241585493, + 0.001450932933948934, + 0.0009654240566305816, + 0.001288781058974564, + 0.0018686994444578886, + 0.0034248630981892347, + 0.002568186493590474, + 0.004189889412373304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.01937675289809704 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a young woman with her head tilted noticeably to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's dark brown hair pulled back with fine strands falling at her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a slender, black-handled makeup brush with dark eyeliner on the bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman drawing a precise line of eyeliner along her lower eyelid while her mouth opens and closes rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, blurred, neutral-toned bathroom creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330007031559944 + ], + "opposite_logprob_tokens": [ + -3.7733001708984375 + ], + "logprob_reward": -0.02330007031559944, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330007031559944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5834977626800537, + 0.11052828282117844, + 0.09817445278167725, + 0.031127477064728737, + 0.04178733378648758, + 0.0459301620721817, + 0.019822794944047928, + 0.02587401494383812, + 0.023538818582892418, + 0.019718872383236885 + ], + "credit_map": [ + 0.0461202971637249, + 0.050551604479551315, + 0.049602922052145004, + 0.05349819362163544, + 0.10999274253845215, + 0.11327333003282547, + 0.05127768963575363, + 0.10918094217777252, + 0.00966773834079504, + 0.010809303261339664, + 0.01167454943060875, + 0.014033902436494827, + 0.015739835798740387, + 0.01585680991411209, + 0.01675080880522728, + 0.015995332971215248, + 0.007950601167976856, + 0.01297232136130333, + 0.008683163672685623, + 0.010300328955054283, + 0.01557763759046793, + 0.017146585509181023, + 0.009754804894328117, + 0.015789007768034935, + 0.0021615324076265097, + 0.002971031004562974, + 0.0036123371683061123, + 0.00476275198161602, + 0.0031106644310057163, + 0.004881104454398155, + 0.003837223630398512, + 0.00579083152115345, + 0.003582602832466364, + 0.005023573990911245, + 0.004402066580951214, + 0.00427966145798564, + 0.005378151312470436, + 0.007528475485742092, + 0.004249286837875843, + 0.007343513425439596, + 0.003566108178347349, + 0.004506543278694153, + 0.0048200166784226894, + 0.005084663163870573, + 0.006241693161427975, + 0.007647756487131119, + 0.004673606716096401, + 0.009389778599143028, + 0.00180214480496943, + 0.002658476121723652, + 0.002184981247410178, + 0.002690528053790331, + 0.0024859365075826645, + 0.0026324805803596973, + 0.002183005679398775, + 0.0031852403189986944, + 0.0023242065217345953, + 0.002714848378673196, + 0.0025298476684838533, + 0.0030286996625363827, + 0.0033287552651017904, + 0.00460779806599021, + 0.0020125899463891983, + 0.005327269434928894, + 0.0016571171581745148, + 0.0022653962951153517, + 0.003041239222511649, + 0.0026870081201195717, + 0.0034729880280792713, + 0.0030283513478934765, + 0.0034885669592767954, + 0.0038981509860605, + 0.0015804577851668, + 0.0023162467405200005, + 0.0015787336742505431, + 0.0024815984070301056, + 0.0030885664746165276, + 0.0028113829903304577, + 0.0023993917275220156, + 0.003462495282292366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.02330007031559944 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a young woman with her head tilted noticeably to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's dark brown hair pulled back with fine strands falling at her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a slender, black-handled makeup brush with dark eyeliner on the bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman drawing a precise line of eyeliner along her lower eyelid while her mouth opens and closes rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, blurred, neutral-toned bathroom creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811256214976311 + ], + "opposite_logprob_tokens": [ + -3.2881126403808594 + ], + "logprob_reward": -0.03811256214976311, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811256214976311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41022801399230957, + 0.12410715967416763, + 0.09090183675289154, + 0.025051191449165344, + 0.17942753434181213, + 0.045552387833595276, + 0.0399387888610363, + 0.03549937158823013, + 0.033203836530447006, + 0.016089892014861107 + ], + "credit_map": [ + 0.030935918912291527, + 0.03576960787177086, + 0.03636767342686653, + 0.03673961013555527, + 0.07842154800891876, + 0.08094107359647751, + 0.03420804813504219, + 0.07684455811977386, + 0.007862688042223454, + 0.021567732095718384, + 0.01061068195849657, + 0.015537945553660393, + 0.01438457053154707, + 0.018120160326361656, + 0.015090957283973694, + 0.020932428538799286, + 0.007368805818259716, + 0.011799922212958336, + 0.009046493098139763, + 0.010136703960597515, + 0.012338051572442055, + 0.014216057024896145, + 0.010258786380290985, + 0.015737013891339302, + 0.001633156556636095, + 0.0030272542499005795, + 0.0028929293621331453, + 0.0035798551980406046, + 0.0021765432320535183, + 0.0040021552704274654, + 0.0028796326369047165, + 0.00485966308042407, + 0.01755361072719097, + 0.016955789178609848, + 0.020710691809654236, + 0.018369993194937706, + 0.022438544780015945, + 0.03451674431562424, + 0.016874782741069794, + 0.03200738504528999, + 0.003143653040751815, + 0.006254303734749556, + 0.004984891042113304, + 0.006551525555551052, + 0.004709102213382721, + 0.00576502550393343, + 0.007209994364529848, + 0.006933892611414194, + 0.0028791038785129786, + 0.004055475350469351, + 0.0044199516996741295, + 0.005721597000956535, + 0.004935929086059332, + 0.006714525166898966, + 0.004291728138923645, + 0.0069204759784042835, + 0.0023923474363982677, + 0.0032102777622640133, + 0.003241361817345023, + 0.004410795401781797, + 0.0037892574910074472, + 0.005861432757228613, + 0.003851512214168906, + 0.008742386475205421, + 0.00216967286542058, + 0.003585382364690304, + 0.004386385902762413, + 0.004157153889536858, + 0.004874405451118946, + 0.004156939219683409, + 0.0048471130430698395, + 0.005026784725487232, + 0.0010388194350525737, + 0.0019285681191831827, + 0.001403676811605692, + 0.0019784357864409685, + 0.0018974329577758908, + 0.0022330600768327713, + 0.002507250290364027, + 0.0031026494689285755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.03811256214976311 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a young woman with her head tilted noticeably to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's dark brown hair pulled back with fine strands falling at her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a slender, black-handled makeup brush with dark eyeliner on the bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman drawing a precise line of eyeliner along her lower eyelid while her mouth opens and closes rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, blurred, neutral-toned bathroom creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020590247586369514 + ], + "opposite_logprob_tokens": [ + -3.895590305328369 + ], + "logprob_reward": -0.020590247586369514, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020590247586369514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19358928501605988, + 0.119041807949543, + 0.06466644257307053, + 0.06778889149427414, + 0.30872029066085815, + 0.042037464678287506, + 0.0786866620182991, + 0.07755067944526672, + 0.027705978602170944, + 0.02021254040300846 + ], + "credit_map": [ + 0.012843010947108269, + 0.021953292191028595, + 0.01725013740360737, + 0.019264711067080498, + 0.02856343239545822, + 0.033430300652980804, + 0.02096657268702984, + 0.039317820221185684, + 0.008796426467597485, + 0.01651078648865223, + 0.011576949618756771, + 0.01905072294175625, + 0.016525493934750557, + 0.016625037416815758, + 0.012865634635090828, + 0.017090758308768272, + 0.004760034382343292, + 0.007023943122476339, + 0.006124706938862801, + 0.007155954372137785, + 0.008229860104620457, + 0.011051536537706852, + 0.008284859359264374, + 0.01203555054962635, + 0.003182247979566455, + 0.005372507963329554, + 0.005437379237264395, + 0.005393135827034712, + 0.005871730390936136, + 0.012085407972335815, + 0.016490640118718147, + 0.013955840840935707, + 0.03048519790172577, + 0.030270058661699295, + 0.037890736013650894, + 0.032289206981658936, + 0.03788544610142708, + 0.05684114620089531, + 0.032318636775016785, + 0.0507398396730423, + 0.0025019394233822823, + 0.00275759887881577, + 0.0034692941699177027, + 0.0034319059923291206, + 0.003302041906863451, + 0.008194292895495892, + 0.007650337647646666, + 0.01073005236685276, + 0.005580871365964413, + 0.008962110616266727, + 0.008819447830319405, + 0.013371018692851067, + 0.009322942234575748, + 0.01188849750906229, + 0.00901318620890379, + 0.011728587560355663, + 0.006288423668593168, + 0.009155036881566048, + 0.008935072459280491, + 0.011748535558581352, + 0.010744081810116768, + 0.010210927575826645, + 0.009341876022517681, + 0.011126724071800709, + 0.0014045442221686244, + 0.002412010682746768, + 0.004005922470241785, + 0.0036906669847667217, + 0.0026040617376565933, + 0.0046294573694467545, + 0.004092796705663204, + 0.004866518545895815, + 0.0018498827703297138, + 0.0030441388953477144, + 0.002459465991705656, + 0.0021644230000674725, + 0.002413175767287612, + 0.0026660740841180086, + 0.002832080703228712, + 0.002783297561109066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.020590247586369514 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a young woman with her head tilted noticeably to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's dark brown hair pulled back with fine strands falling at her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a slender, black-handled makeup brush with dark eyeliner on the bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman drawing a precise line of eyeliner along her lower eyelid while her mouth opens and closes rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, blurred, neutral-toned bathroom creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192831702530384 + ], + "opposite_logprob_tokens": [ + -4.264193058013916 + ], + "logprob_reward": -0.014192831702530384, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192831702530384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2986208200454712, + 0.11871697008609772, + 0.15079516172409058, + 0.08495954424142838, + 0.12422370910644531, + 0.033845532685518265, + 0.0794249027967453, + 0.05185527354478836, + 0.027120329439640045, + 0.030437754467129707 + ], + "credit_map": [ + 0.016036050394177437, + 0.021085500717163086, + 0.019920749589800835, + 0.025701548904180527, + 0.022458691149950027, + 0.057404063642024994, + 0.049865301698446274, + 0.0861489325761795, + 0.00853628572076559, + 0.009391658939421177, + 0.01037536095827818, + 0.013865143992006779, + 0.013140604831278324, + 0.019308898597955704, + 0.015370729379355907, + 0.028728285804390907, + 0.006865031085908413, + 0.00856979563832283, + 0.009346376173198223, + 0.009287442080676556, + 0.011679865419864655, + 0.03211965039372444, + 0.026324426755309105, + 0.04660256579518318, + 0.003318948205560446, + 0.0036085559986531734, + 0.005327175837010145, + 0.004958410747349262, + 0.005854777526110411, + 0.0188254714012146, + 0.013485927134752274, + 0.02958027645945549, + 0.013095976784825325, + 0.011005007661879063, + 0.013625552877783775, + 0.012442647479474545, + 0.014411946758627892, + 0.024011880159378052, + 0.011301432736217976, + 0.024329276755452156, + 0.0013763721799477935, + 0.0022024260833859444, + 0.002234148560091853, + 0.0018373382044956088, + 0.003876814153045416, + 0.007081728428602219, + 0.005768527742475271, + 0.00946817547082901, + 0.008227267302572727, + 0.009798029437661171, + 0.010158671997487545, + 0.012326899915933609, + 0.009777458384633064, + 0.009947560727596283, + 0.009323528036475182, + 0.009865486063063145, + 0.004569696728140116, + 0.00416077533736825, + 0.005265552084892988, + 0.0050634415820240974, + 0.0064874752424657345, + 0.007690757513046265, + 0.006745998281985521, + 0.011871580965816975, + 0.002055601216852665, + 0.0030394932255148888, + 0.0037887871731072664, + 0.003001800272613764, + 0.002709765452891588, + 0.003687990829348564, + 0.003983009606599808, + 0.0048538814298808575, + 0.0014211663510650396, + 0.0018950736848637462, + 0.0015974990092217922, + 0.0028244268614798784, + 0.0027660909108817577, + 0.005673341453075409, + 0.005586686078459024, + 0.008673472329974174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.014192831702530384 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_4/index_tensor([8596, 8596])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_4/index_tensor([8596, 8596])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..eddeaf463c08402cc8853707dc2058e0057bbc18 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_4/index_tensor([8596, 8596])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, sweeping wide shot, rendered in Van Gogh's signature Post-Impressionist style, captures a lone male figure standing atop a precipice amidst a vibrant, swirling wheatfield beneath a turbulent sky. The entire scene is constructed with thick, visible, unsentimental brushstrokes, imbuing both the landscape and the figure with a tactile, three-dimensional reality. The field is an explosive sea of gold and ochre, with the wheat stalks bending and twisting in a chaotic, windswept dance that reflects his inner state. Above, heavy, swirling clouds tumble across a deep blue and cyan sky, the harsh afternoon light casting dramatic, elongated shadows. The man, thin and dressed in worn, dark, rustic clothing, stands frozen. His chest rises and falls with his rapid, labored breaths, a subtle yet powerful micro-movement that suggests he has just stopped fleeing an unseen terror. His eyes, wide open and filled with frantic, manic energy, gaze upon the turbulent landscape as if it were a direct, painterly expression of his own morbid fear and psychological anguish.", + "seed": 20396, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_4/index_8596_seed_20396_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.31064244657754897, + "vlm_logprob": -0.31064244657754897, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.31064244657754897, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Van Gogh Post-Impressionist style with thick, visible brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone male figure standing atop a precipice in a vibrant wheatfield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as turbulent with heavy, swirling clouds in deep blue and cyan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man dressed in worn, dark, rustic clothing and standing frozen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's chest visibly rising and falling with rapid, labored breaths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.035253118723630905 + ], + "opposite_logprob_tokens": [ + -3.4102530479431152 + ], + "logprob_reward": -0.035253118723630905, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.035253118723630905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19121699035167694, + 0.06136420741677284, + 0.01825897768139839, + 0.03754700720310211, + 0.032249558717012405, + 0.2756374478340149, + 0.2602332830429077, + 0.038789406418800354, + 0.022383863106369972, + 0.06231923773884773 + ], + "credit_map": [ + 0.02967599220573902, + 0.02475835755467415, + 0.022640833631157875, + 0.017666233703494072, + 0.018389727920293808, + 0.04271022975444794, + 0.013895280659198761, + 0.021480318158864975, + 0.0070426128804683685, + 0.007112256251275539, + 0.007253294810652733, + 0.005623293574899435, + 0.005603680852800608, + 0.014666642993688583, + 0.004039036575704813, + 0.010023389011621475, + 0.002371452050283551, + 0.002317503560334444, + 0.0026008710265159607, + 0.001924080541357398, + 0.0018686556722968817, + 0.003388083539903164, + 0.001690463162958622, + 0.002097868360579014, + 0.005467702634632587, + 0.006958026438951492, + 0.00276849209330976, + 0.0029192722868174314, + 0.0035889141727238894, + 0.009472143836319447, + 0.002800952410325408, + 0.0035715019330382347, + 0.004259921610355377, + 0.004008246585726738, + 0.0033589035738259554, + 0.0024598187301307917, + 0.003532974747940898, + 0.009280338883399963, + 0.002242111833766103, + 0.003107242751866579, + 0.05322062596678734, + 0.03169795870780945, + 0.02269396372139454, + 0.03303477168083191, + 0.01561151072382927, + 0.050391074270009995, + 0.015546275302767754, + 0.05344124510884285, + 0.06957418471574783, + 0.10121773183345795, + 0.012774861417710781, + 0.011791994795203209, + 0.011522335931658745, + 0.035998497158288956, + 0.008701114915311337, + 0.00865257065743208, + 0.002720216056331992, + 0.005332550499588251, + 0.005814485717564821, + 0.006129919085651636, + 0.0032600986305624247, + 0.004819410387426615, + 0.0043014464899897575, + 0.006411280017346144, + 0.002045109635218978, + 0.0034658960066735744, + 0.0022045806981623173, + 0.0020661079324781895, + 0.0038395021110773087, + 0.0037855824921280146, + 0.002557712374255061, + 0.0024193732533603907, + 0.016881361603736877, + 0.014260047115385532, + 0.0043204305693507195, + 0.0037901790346950293, + 0.0035674539394676685, + 0.013497966341674328, + 0.0022015722934156656, + 0.003800227539613843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.035253118723630905 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Van Gogh Post-Impressionist style with thick, visible brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone male figure standing atop a precipice in a vibrant wheatfield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as turbulent with heavy, swirling clouds in deep blue and cyan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man dressed in worn, dark, rustic clothing and standing frozen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's chest visibly rising and falling with rapid, labored breaths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001526977401226759 + ], + "opposite_logprob_tokens": [ + -6.501526832580566 + ], + "logprob_reward": -0.001526977401226759, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001526977401226759, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0428917370736599, + 0.017105119302868843, + 0.008595609106123447, + 0.011756468564271927, + 0.015735795721411705, + 0.2664077579975128, + 0.41399386525154114, + 0.0421539768576622, + 0.017703408375382423, + 0.16365626454353333 + ], + "credit_map": [ + 0.005164824891835451, + 0.005254375748336315, + 0.005002335645258427, + 0.0055724820122122765, + 0.005718517582863569, + 0.007039066404104233, + 0.004001999273896217, + 0.005138132721185684, + 0.0022027422674000263, + 0.002222831128165126, + 0.0014601987786591053, + 0.0019698580726981163, + 0.0020629223436117172, + 0.002691789297387004, + 0.0015035771066322923, + 0.0029912004247307777, + 0.000765646283980459, + 0.0009339166572317481, + 0.0009407572215422988, + 0.0009247540147043765, + 0.0014278279850259423, + 0.001237334217876196, + 0.001052359351888299, + 0.0013130140723660588, + 0.0012081856839358807, + 0.0018243890954181552, + 0.0011591824004426599, + 0.0011978588299825788, + 0.0009980308823287487, + 0.0027002128772437572, + 0.000988807063549757, + 0.0016798017313703895, + 0.002563521498814225, + 0.0021599021274596453, + 0.0014084192225709558, + 0.001366156036965549, + 0.001426178845576942, + 0.0033442024141550064, + 0.0014449004083871841, + 0.002022514818236232, + 0.04949931800365448, + 0.03088279440999031, + 0.02202678844332695, + 0.034779611974954605, + 0.015132047235965729, + 0.04100067541003227, + 0.01640862412750721, + 0.05667789652943611, + 0.11897395551204681, + 0.11785353720188141, + 0.021322445943951607, + 0.023185206577181816, + 0.022778665646910667, + 0.06677453964948654, + 0.017858007922768593, + 0.025247514247894287, + 0.002514110179618001, + 0.005089374724775553, + 0.0056050559505820274, + 0.006464587524533272, + 0.00350444414652884, + 0.0065128011628985405, + 0.0045629399828612804, + 0.007900663651525974, + 0.0015214185696095228, + 0.0026226502377539873, + 0.002561514964327216, + 0.002814649837091565, + 0.0014809333952143788, + 0.002679029945284128, + 0.0018672550795599818, + 0.002155957743525505, + 0.01444320473819971, + 0.039922114461660385, + 0.01190929114818573, + 0.0059333909302949905, + 0.009709300473332405, + 0.06645040959119797, + 0.004537484608590603, + 0.010751072317361832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 9 + ], + "selected_reward": -0.001526977401226759 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Van Gogh Post-Impressionist style with thick, visible brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone male figure standing atop a precipice in a vibrant wheatfield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as turbulent with heavy, swirling clouds in deep blue and cyan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man dressed in worn, dark, rustic clothing and standing frozen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's chest visibly rising and falling with rapid, labored breaths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003198272082954645 + ], + "opposite_logprob_tokens": [ + -5.753198146820068 + ], + "logprob_reward": -0.003198272082954645, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003198272082954645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09939267486333847, + 0.03921704366803169, + 0.05302444472908974, + 0.08237998932600021, + 0.10919765383005142, + 0.08942244201898575, + 0.2024819403886795, + 0.01806154102087021, + 0.01572178304195404, + 0.29110047221183777 + ], + "credit_map": [ + 0.016807222738862038, + 0.013847636990249157, + 0.010859942063689232, + 0.008890032768249512, + 0.014369617216289043, + 0.015822550281882286, + 0.00895107351243496, + 0.009844613261520863, + 0.0047426666133105755, + 0.004637646023184061, + 0.002895575249567628, + 0.003572750138118863, + 0.0038400732446461916, + 0.010495171882212162, + 0.0034633406903594732, + 0.005569823086261749, + 0.010265271179378033, + 0.005346870515495539, + 0.0053935907781124115, + 0.005989463534206152, + 0.003933871164917946, + 0.009903938509523869, + 0.004085128661245108, + 0.008106310851871967, + 0.015159731730818748, + 0.015764914453029633, + 0.0060039605014026165, + 0.006118384189903736, + 0.005856001749634743, + 0.020852388814091682, + 0.004280651453882456, + 0.008343960158526897, + 0.015277442522346973, + 0.019109176471829414, + 0.009929855354130268, + 0.008680509403347969, + 0.009331120178103447, + 0.03211718797683716, + 0.006189811043441296, + 0.008562548086047173, + 0.01336706429719925, + 0.013968049548566341, + 0.008601111359894276, + 0.0078074331395328045, + 0.0060042766854166985, + 0.023131340742111206, + 0.0047645289450883865, + 0.01177864708006382, + 0.017448946833610535, + 0.05836866423487663, + 0.012430701404809952, + 0.007174067199230194, + 0.009163966402411461, + 0.07760608941316605, + 0.005679349415004253, + 0.014610177837312222, + 0.001548385014757514, + 0.0024095778353512287, + 0.0024908126797527075, + 0.001721863285638392, + 0.0024176789447665215, + 0.0016828984953463078, + 0.0029050270095467567, + 0.0028853006660938263, + 0.0018725186819210649, + 0.001802224898710847, + 0.0018224982777610421, + 0.0018009226769208908, + 0.0017198821296915412, + 0.003487978596240282, + 0.001687286188825965, + 0.0015284707769751549, + 0.024368340149521828, + 0.06449870765209198, + 0.022505493834614754, + 0.010374845005571842, + 0.01725674606859684, + 0.12368041276931763, + 0.008509964682161808, + 0.019905978813767433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 4 + ], + "selected_reward": -0.003198272082954645 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Van Gogh Post-Impressionist style with thick, visible brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone male figure standing atop a precipice in a vibrant wheatfield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as turbulent with heavy, swirling clouds in deep blue and cyan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man dressed in worn, dark, rustic clothing and standing frozen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's chest visibly rising and falling with rapid, labored breaths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011084998026490211 + ], + "opposite_logprob_tokens": [ + -4.511085033416748 + ], + "logprob_reward": -0.011084998026490211, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011084998026490211, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1341186910867691, + 0.045602649450302124, + 0.02216201461851597, + 0.044940169900655746, + 0.02189658209681511, + 0.17650847136974335, + 0.1575750708580017, + 0.0429062694311142, + 0.04173653945326805, + 0.3125535547733307 + ], + "credit_map": [ + 0.019247084856033325, + 0.014263801276683807, + 0.024059463292360306, + 0.012873711995780468, + 0.015380610711872578, + 0.017090197652578354, + 0.015206700190901756, + 0.01599711738526821, + 0.0058411904610693455, + 0.005062180571258068, + 0.006662917789071798, + 0.0061278510838747025, + 0.0035142451524734497, + 0.006602620705962181, + 0.004526518750935793, + 0.0072651212103664875, + 0.0023894088808447123, + 0.002463939366862178, + 0.0024136200081557035, + 0.002133353613317013, + 0.002754863118752837, + 0.004106682259589434, + 0.0026131398044526577, + 0.003287004539743066, + 0.004954248201102018, + 0.006697998382151127, + 0.00472303805872798, + 0.004878378938883543, + 0.005464954301714897, + 0.008377932012081146, + 0.004625072702765465, + 0.0052185445092618465, + 0.001874170615337789, + 0.002599514089524746, + 0.003270550165325403, + 0.002420730423182249, + 0.0028561323415488005, + 0.0028815472032874823, + 0.002496498404070735, + 0.003497437806800008, + 0.015171675011515617, + 0.02612248994410038, + 0.012825089506804943, + 0.018561113625764847, + 0.019037911668419838, + 0.036100395023822784, + 0.022201836109161377, + 0.02648795023560524, + 0.022411705926060677, + 0.05056304112076759, + 0.007928464561700821, + 0.006487713661044836, + 0.015112516470253468, + 0.041456934064626694, + 0.00776706775650382, + 0.005847627762705088, + 0.0033420738764107227, + 0.00454462505877018, + 0.0037035204004496336, + 0.0038213671650737524, + 0.007246994413435459, + 0.006636280100792646, + 0.006710829213261604, + 0.0069005717523396015, + 0.004222167190164328, + 0.005279012490063906, + 0.0037346945609897375, + 0.004432980436831713, + 0.004994939547032118, + 0.007999686524271965, + 0.005291856359690428, + 0.005781196057796478, + 0.032436590641736984, + 0.08296128362417221, + 0.02149796672165394, + 0.012680770829319954, + 0.019318701699376106, + 0.11346227675676346, + 0.010034317150712013, + 0.02016163431107998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.011084998026490211 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Van Gogh Post-Impressionist style with thick, visible brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone male figure standing atop a precipice in a vibrant wheatfield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as turbulent with heavy, swirling clouds in deep blue and cyan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man dressed in worn, dark, rustic clothing and standing frozen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's chest visibly rising and falling with rapid, labored breaths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.5021488666534424 + ], + "opposite_logprob_tokens": [ + -0.25214889645576477 + ], + "logprob_reward": -1.5021488666534424, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.5021488666534424, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.0915398895740509, + 0.03397464379668236, + 0.01892659068107605, + 0.026945117861032486, + 0.04292302951216698, + 0.21687796711921692, + 0.11645988374948502, + 0.06554343551397324, + 0.0327744297683239, + 0.35403501987457275 + ], + "credit_map": [ + 0.015958817675709724, + 0.010857957415282726, + 0.013584078289568424, + 0.009485364891588688, + 0.012611095793545246, + 0.010112389922142029, + 0.010252182371914387, + 0.00867799948900938, + 0.004568883217871189, + 0.0054136645048856735, + 0.00367029570043087, + 0.003007640363648534, + 0.004226312972605228, + 0.0036382044199854136, + 0.0029171323403716087, + 0.006532510742545128, + 0.0019129945430904627, + 0.0025062854401767254, + 0.0017691190587356687, + 0.0016534263268113136, + 0.0033929836936295033, + 0.0025991254951804876, + 0.002101615536957979, + 0.0029910397715866566, + 0.004612556658685207, + 0.0032815448939800262, + 0.003052392043173313, + 0.0021123087499290705, + 0.003245381638407707, + 0.005759596824645996, + 0.0020757336169481277, + 0.0028056057635694742, + 0.004830850753933191, + 0.005350471008569002, + 0.0039275651797652245, + 0.004185115918517113, + 0.004826991353183985, + 0.011040828190743923, + 0.003342514159157872, + 0.005418696906417608, + 0.03259929269552231, + 0.025985008105635643, + 0.016698239371180534, + 0.027589330449700356, + 0.013735238462686539, + 0.03939066454768181, + 0.01788269355893135, + 0.042997490614652634, + 0.028259098529815674, + 0.04139670357108116, + 0.006122540682554245, + 0.004688808228820562, + 0.0064401812851428986, + 0.019155675545334816, + 0.004738722462207079, + 0.005658154375851154, + 0.004063788801431656, + 0.00699040899053216, + 0.006548789795488119, + 0.008686846122145653, + 0.010920562781393528, + 0.01250971294939518, + 0.00816221535205841, + 0.007661107927560806, + 0.003150426782667637, + 0.0033581897150725126, + 0.003302187891677022, + 0.0035576890222728252, + 0.004558578133583069, + 0.0032476794440299273, + 0.004941685125231743, + 0.006657995283603668, + 0.032341208308935165, + 0.08632078021764755, + 0.02599739469587803, + 0.013091427274048328, + 0.02064834162592888, + 0.14213676750659943, + 0.010054662823677063, + 0.023444464430212975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -1.5021488666534424 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "static, sweeping wide shot, rendered in Van Gogh's signature Post-Impressionist style, captures a lone male figure standing atop a precipice amidst a vibrant, swirling wheatfield beneath a turbulent sky. The entire scene is constructed with thick, visible, unsentimental brushstrokes, imbuing both the landscape and the figure with a tactile, three-dimensional reality. The field is an explosive sea of gold and ochre, with the wheat stalks bending and twisting in a chaotic, windswept dance that reflects his inner state. Above, heavy, swirling clouds tumble across a deep blue and cyan sky, the harsh afternoon light casting dramatic, elongated shadows. The man, thin and dressed in worn, dark, rustic clothing, stands frozen. His chest rises and falls with his rapid, labored breaths, a subtle yet powerful micro-movement that suggests he has just stopped fleeing an unseen terror. His eyes, wide open and filled with frantic, manic energy, gaze upon the turbulent landscape as if it were a direct, painterly expression of his own morbid fear and psychological anguish.", + "seed": 20397, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_4/index_8596_seed_20397_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.14121557930484413, + "vlm_logprob": -0.14121557930484413, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.14121557930484413, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Van Gogh Post-Impressionist style with thick, visible brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone male figure standing atop a precipice in a vibrant wheatfield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as turbulent with heavy, swirling clouds in deep blue and cyan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man dressed in worn, dark, rustic clothing and standing frozen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's chest visibly rising and falling with rapid, labored breaths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03204033896327019 + ], + "opposite_logprob_tokens": [ + -3.5320403575897217 + ], + "logprob_reward": -0.03204033896327019, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03204033896327019, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2415090650320053, + 0.11104588210582733, + 0.061448514461517334, + 0.04354530945420265, + 0.050180498510599136, + 0.190582737326622, + 0.10403868556022644, + 0.05004768818616867, + 0.0430111400783062, + 0.10459042340517044 + ], + "credit_map": [ + 0.027462705969810486, + 0.03422416374087334, + 0.026549682021141052, + 0.026742219924926758, + 0.03324195742607117, + 0.029592277482151985, + 0.03427395597100258, + 0.02942209504544735, + 0.010075433179736137, + 0.012683465145528316, + 0.011473986320197582, + 0.023675711825489998, + 0.012222799472510815, + 0.007100842427462339, + 0.022609233856201172, + 0.011204421520233154, + 0.006631689146161079, + 0.008098307065665722, + 0.006119506433606148, + 0.009066862054169178, + 0.008694412186741829, + 0.00736211845651269, + 0.009447638876736164, + 0.006027980707585812, + 0.0032371182460337877, + 0.002792562125250697, + 0.008291376754641533, + 0.007099668495357037, + 0.002244558185338974, + 0.002418036572635174, + 0.006879422347992659, + 0.01058256532996893, + 0.004182102158665657, + 0.0035842228680849075, + 0.009009450674057007, + 0.0126161128282547, + 0.00361617305316031, + 0.004097094293683767, + 0.007080341223627329, + 0.0059950025752186775, + 0.012792794965207577, + 0.011375779286026955, + 0.0535067617893219, + 0.053258638828992844, + 0.006556331180036068, + 0.006363310851156712, + 0.021151408553123474, + 0.025577714666724205, + 0.01305560115724802, + 0.022761400789022446, + 0.010702230967581272, + 0.015018085949122906, + 0.012075098231434822, + 0.007010784465819597, + 0.012686247937381268, + 0.010729237459599972, + 0.007316491100937128, + 0.010418578051030636, + 0.004733689595013857, + 0.005363226868212223, + 0.00824393518269062, + 0.0038615919183939695, + 0.005072419065982103, + 0.005037757568061352, + 0.005027978215366602, + 0.005810176022350788, + 0.00589028000831604, + 0.005866985768079758, + 0.006657445337623358, + 0.0030204798094928265, + 0.004762168973684311, + 0.005975627340376377, + 0.007879607379436493, + 0.005583272781223059, + 0.033030521124601364, + 0.03208792582154274, + 0.0037378710694611073, + 0.00357245234772563, + 0.008816549554467201, + 0.009882231242954731 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.03204033896327019 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Van Gogh Post-Impressionist style with thick, visible brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone male figure standing atop a precipice in a vibrant wheatfield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as turbulent with heavy, swirling clouds in deep blue and cyan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man dressed in worn, dark, rustic clothing and standing frozen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's chest visibly rising and falling with rapid, labored breaths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012559603899717331 + ], + "opposite_logprob_tokens": [ + -4.387559413909912 + ], + "logprob_reward": -0.012559603899717331, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012559603899717331, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15358686447143555, + 0.07075744867324829, + 0.04582327604293823, + 0.04856162518262863, + 0.14882521331310272, + 0.15791118144989014, + 0.03754919022321701, + 0.048025403171777725, + 0.08766704052686691, + 0.20129281282424927 + ], + "credit_map": [ + 0.021031390875577927, + 0.02023640275001526, + 0.013787285424768925, + 0.02257145382463932, + 0.020696517080068588, + 0.014994954690337181, + 0.025744497776031494, + 0.01452435739338398, + 0.008219701237976551, + 0.009181288070976734, + 0.010379836894571781, + 0.007532056421041489, + 0.012084211222827435, + 0.00642409548163414, + 0.007853209041059017, + 0.00908304937183857, + 0.004861265886574984, + 0.005821312312036753, + 0.005060816649347544, + 0.007213311269879341, + 0.005138314329087734, + 0.005477520637214184, + 0.006049578543752432, + 0.00620115315541625, + 0.004133244976401329, + 0.004645123612135649, + 0.006498770788311958, + 0.007842229679226875, + 0.006017882376909256, + 0.003295188071206212, + 0.008444827049970627, + 0.007684354204684496, + 0.0104360431432724, + 0.010290388949215412, + 0.02281242422759533, + 0.047187548130750656, + 0.00553811714053154, + 0.006621690467000008, + 0.02393975295126438, + 0.021999238058924675, + 0.009983156807720661, + 0.010706198401749134, + 0.03868110477924347, + 0.04362792521715164, + 0.010438356548547745, + 0.005087127909064293, + 0.018951117992401123, + 0.02043617144227028, + 0.0034303111024200916, + 0.0038465275429189205, + 0.005612777080386877, + 0.006867740768939257, + 0.004315742291510105, + 0.003458851482719183, + 0.005441484972834587, + 0.004575755447149277, + 0.005401281639933586, + 0.008724777959287167, + 0.0056901047937572, + 0.0069210343062877655, + 0.0061702486127614975, + 0.003932401072233915, + 0.0064043463207781315, + 0.004781208001077175, + 0.008365725167095661, + 0.013772430829703808, + 0.011458058841526508, + 0.011864627711474895, + 0.012361557222902775, + 0.009449596516788006, + 0.011048045940697193, + 0.00934699084609747, + 0.01438379567116499, + 0.013125822879374027, + 0.0563054196536541, + 0.044713713228702545, + 0.01728413626551628, + 0.009176323190331459, + 0.01785256154835224, + 0.028451044112443924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.012559603899717331 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Van Gogh Post-Impressionist style with thick, visible brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone male figure standing atop a precipice in a vibrant wheatfield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as turbulent with heavy, swirling clouds in deep blue and cyan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man dressed in worn, dark, rustic clothing and standing frozen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's chest visibly rising and falling with rapid, labored breaths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005258894059807062 + ], + "opposite_logprob_tokens": [ + -5.255259037017822 + ], + "logprob_reward": -0.005258894059807062, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005258894059807062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4031674563884735, + 0.13252082467079163, + 0.12635838985443115, + 0.04659553989768028, + 0.13142259418964386, + 0.029535865411162376, + 0.01849297806620598, + 0.019486883655190468, + 0.0177102480083704, + 0.07470922917127609 + ], + "credit_map": [ + 0.05140727013349533, + 0.04382308945059776, + 0.047698330134153366, + 0.05595247820019722, + 0.042274657636880875, + 0.05387048423290253, + 0.052622489631175995, + 0.05551868677139282, + 0.013745241798460484, + 0.015848178416490555, + 0.014491104520857334, + 0.03145432099699974, + 0.014553108252584934, + 0.008374055847525597, + 0.021793488413095474, + 0.01226132083684206, + 0.01442150678485632, + 0.014332028105854988, + 0.018404411152005196, + 0.016945764422416687, + 0.014767429791390896, + 0.013243432156741619, + 0.016790924593806267, + 0.01745288446545601, + 0.003119207452982664, + 0.0030916461255401373, + 0.009171344339847565, + 0.014261643402278423, + 0.0024552776012569666, + 0.0017372386064380407, + 0.0075195422396063805, + 0.005239637568593025, + 0.00926021859049797, + 0.008814935572445393, + 0.02458367310464382, + 0.043907858431339264, + 0.004482561722397804, + 0.004370568320155144, + 0.02291715331375599, + 0.013085626997053623, + 0.0018784640124067664, + 0.0021562334150075912, + 0.006453075911849737, + 0.00814908742904663, + 0.0024669405538588762, + 0.0013289825292304158, + 0.0034688804298639297, + 0.0036342008970677853, + 0.0019580780062824488, + 0.0031653305049985647, + 0.0019338402198627591, + 0.002807505428791046, + 0.00220918795093894, + 0.0014624166069552302, + 0.002749709179624915, + 0.002206911100074649, + 0.002388942986726761, + 0.0037021697498857975, + 0.0028401047457009554, + 0.0025059471372514963, + 0.002862226217985153, + 0.0013156934874132276, + 0.002065933309495449, + 0.0018058652058243752, + 0.0013942646328359842, + 0.0014852783642709255, + 0.004483653232455254, + 0.0044245729222893715, + 0.0011948593892157078, + 0.0009901972953230143, + 0.0017909907037392259, + 0.001946431351825595, + 0.005257111042737961, + 0.004057490732520819, + 0.024070214480161667, + 0.02267533726990223, + 0.0026123267598450184, + 0.0020296978764235973, + 0.006232682149857283, + 0.007774378173053265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.005258894059807062 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Van Gogh Post-Impressionist style with thick, visible brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone male figure standing atop a precipice in a vibrant wheatfield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as turbulent with heavy, swirling clouds in deep blue and cyan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man dressed in worn, dark, rustic clothing and standing frozen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's chest visibly rising and falling with rapid, labored breaths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023295992985367775 + ], + "opposite_logprob_tokens": [ + -3.7732958793640137 + ], + "logprob_reward": -0.023295992985367775, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023295992985367775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16631898283958435, + 0.15204904973506927, + 0.03404073417186737, + 0.06783153861761093, + 0.05156799405813217, + 0.18907903134822845, + 0.04760301858186722, + 0.03459370136260986, + 0.1237940564751625, + 0.13312190771102905 + ], + "credit_map": [ + 0.017661120742559433, + 0.01965126395225525, + 0.01942630298435688, + 0.02322976477444172, + 0.01453470066189766, + 0.024232951924204826, + 0.022974830120801926, + 0.024608053267002106, + 0.01121294591575861, + 0.013440401293337345, + 0.017926327884197235, + 0.03776765242218971, + 0.010576442815363407, + 0.008448765613138676, + 0.03553296625614166, + 0.017143556848168373, + 0.0034552812576293945, + 0.0035916788037866354, + 0.00452868128195405, + 0.0047583854757249355, + 0.0034410443622618914, + 0.004021668806672096, + 0.005645368713885546, + 0.00459862919524312, + 0.004146611783653498, + 0.004632624331861734, + 0.009717365726828575, + 0.021721601486206055, + 0.004605886992067099, + 0.003938351757824421, + 0.009530606679618359, + 0.009538497775793076, + 0.0039942278526723385, + 0.0026413029991090298, + 0.01243062037974596, + 0.005354767665266991, + 0.0038220516871660948, + 0.0029809484258294106, + 0.007273311261087656, + 0.013070764020085335, + 0.012102712877094746, + 0.012199604883790016, + 0.04836414009332657, + 0.052432045340538025, + 0.010643203742802143, + 0.00802970863878727, + 0.020466187968850136, + 0.024841446429491043, + 0.005939680151641369, + 0.010355580598115921, + 0.0043723429553210735, + 0.006816933862864971, + 0.006310340017080307, + 0.0033633646089583635, + 0.0059228879399597645, + 0.004521891940385103, + 0.004062585532665253, + 0.006430760957300663, + 0.004325182642787695, + 0.004488136153668165, + 0.004690293222665787, + 0.0035210151690989733, + 0.0030050831846892834, + 0.00407064612954855, + 0.011126527562737465, + 0.012069573625922203, + 0.027775680646300316, + 0.03262561932206154, + 0.0077928281389176846, + 0.008574272505939007, + 0.011375674977898598, + 0.012453884817659855, + 0.00944541022181511, + 0.00761818140745163, + 0.04036653786897659, + 0.042544081807136536, + 0.005959104280918837, + 0.004365337081253529, + 0.011219603009521961, + 0.011603659950196743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.023295992985367775 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Van Gogh Post-Impressionist style with thick, visible brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone male figure standing atop a precipice in a vibrant wheatfield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as turbulent with heavy, swirling clouds in deep blue and cyan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man dressed in worn, dark, rustic clothing and standing frozen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's chest visibly rising and falling with rapid, labored breaths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6329230666160583 + ], + "opposite_logprob_tokens": [ + -0.7579230666160583 + ], + "logprob_reward": -0.6329230666160583, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6329230666160583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.242347851395607, + 0.08018700033426285, + 0.06357842683792114, + 0.1383255124092102, + 0.10679222643375397, + 0.07993613183498383, + 0.04288250952959061, + 0.10288908332586288, + 0.08177980780601501, + 0.061281442642211914 + ], + "credit_map": [ + 0.030415479093790054, + 0.031852640211582184, + 0.030999522656202316, + 0.028974631801247597, + 0.04213256388902664, + 0.02000468038022518, + 0.03152119740843773, + 0.026447154581546783, + 0.007647649385035038, + 0.007112864870578051, + 0.011477390304207802, + 0.012179465964436531, + 0.008460217155516148, + 0.006233384367078543, + 0.01442774198949337, + 0.012648292817175388, + 0.006748440209776163, + 0.006985743995755911, + 0.007024219259619713, + 0.009815883822739124, + 0.006894426420331001, + 0.005678064655512571, + 0.010907703079283237, + 0.00952394399791956, + 0.008761751465499401, + 0.008220246993005276, + 0.02146047353744507, + 0.04279223456978798, + 0.006279599387198687, + 0.007039844989776611, + 0.016957279294729233, + 0.026814110577106476, + 0.007224230095744133, + 0.007178113795816898, + 0.016788747161626816, + 0.03343602642416954, + 0.006189407780766487, + 0.005483993794769049, + 0.01708092913031578, + 0.013410784304141998, + 0.005706343334168196, + 0.006584364455193281, + 0.015624962747097015, + 0.01060002762824297, + 0.008957026526331902, + 0.0064500183798372746, + 0.01096082292497158, + 0.015052573755383492, + 0.0041591813787817955, + 0.006545132491737604, + 0.005267730914056301, + 0.0068281288258731365, + 0.00656278757378459, + 0.0043801674619317055, + 0.005231690593063831, + 0.003907692618668079, + 0.01143477950245142, + 0.019507911056280136, + 0.014280177652835846, + 0.014608890749514103, + 0.012564637698233128, + 0.0055734156630933285, + 0.013279112987220287, + 0.011640168726444244, + 0.009773910976946354, + 0.01028626412153244, + 0.010771455243229866, + 0.01189221628010273, + 0.012421241030097008, + 0.0064522274769842625, + 0.010886689648032188, + 0.009295815601944923, + 0.004873097874224186, + 0.004277375061064959, + 0.014359177090227604, + 0.008730507455766201, + 0.005620401352643967, + 0.004890052601695061, + 0.006479183677583933, + 0.012051653116941452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.6329230666160583 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, sweeping wide shot, rendered in Van Gogh's signature Post-Impressionist style, captures a lone male figure standing atop a precipice amidst a vibrant, swirling wheatfield beneath a turbulent sky. The entire scene is constructed with thick, visible, unsentimental brushstrokes, imbuing both the landscape and the figure with a tactile, three-dimensional reality. The field is an explosive sea of gold and ochre, with the wheat stalks bending and twisting in a chaotic, windswept dance that reflects his inner state. Above, heavy, swirling clouds tumble across a deep blue and cyan sky, the harsh afternoon light casting dramatic, elongated shadows. The man, thin and dressed in worn, dark, rustic clothing, stands frozen. His chest rises and falls with his rapid, labored breaths, a subtle yet powerful micro-movement that suggests he has just stopped fleeing an unseen terror. His eyes, wide open and filled with frantic, manic energy, gaze upon the turbulent landscape as if it were a direct, painterly expression of his own morbid fear and psychological anguish.", + "seed": 20398, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_4/index_8596_seed_20398_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1595720130484551, + "vlm_logprob": -0.1595720130484551, + "vlm_yes_no_margin": 3.924999952316284, + "vlm_token_credit": -0.1595720130484551, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Van Gogh Post-Impressionist style with thick, visible brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone male figure standing atop a precipice in a vibrant wheatfield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as turbulent with heavy, swirling clouds in deep blue and cyan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man dressed in worn, dark, rustic clothing and standing frozen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's chest visibly rising and falling with rapid, labored breaths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014280269853770733 + ], + "opposite_logprob_tokens": [ + -4.389280319213867 + ], + "logprob_reward": -0.014280269853770733, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.014280269853770733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16397733986377716, + 0.04157773405313492, + 0.03500143811106682, + 0.01908743381500244, + 0.24828854203224182, + 0.12287349253892899, + 0.2540578246116638, + 0.03251836448907852, + 0.014309107325971127, + 0.06830879300832748 + ], + "credit_map": [ + 0.020596224814653397, + 0.016767477616667747, + 0.02659342624247074, + 0.026955295354127884, + 0.012159629724919796, + 0.013415919616818428, + 0.02331240475177765, + 0.024176955223083496, + 0.00280999974347651, + 0.0034390126820653677, + 0.009177510626614094, + 0.008182991296052933, + 0.003171508200466633, + 0.003627960104495287, + 0.004972896538674831, + 0.006195856723934412, + 0.0025430540554225445, + 0.002538820030167699, + 0.005208546761423349, + 0.004881351254880428, + 0.007018283009529114, + 0.0026778499595820904, + 0.004893200471997261, + 0.005240333266556263, + 0.001346191274933517, + 0.0016015584114938974, + 0.0044808099046349525, + 0.00251522334292531, + 0.0032904401887208223, + 0.001543753081932664, + 0.0025019401218742132, + 0.0018075170228257775, + 0.007291796151548624, + 0.010748395696282387, + 0.06845834106206894, + 0.10083189606666565, + 0.006761164870113134, + 0.018564127385616302, + 0.01367817260324955, + 0.02195463702082634, + 0.00490311486646533, + 0.012254787608981133, + 0.025843435898423195, + 0.036618947982788086, + 0.005431233439594507, + 0.0232466459274292, + 0.004242884460836649, + 0.010332435369491577, + 0.012328274548053741, + 0.012694905512034893, + 0.09728359431028366, + 0.040018677711486816, + 0.007924759760499, + 0.03945085033774376, + 0.010628137737512589, + 0.033728595823049545, + 0.0029967152513563633, + 0.0026866064872592688, + 0.007372192107141018, + 0.0038836083840578794, + 0.004445218481123447, + 0.002488883910700679, + 0.004240067210048437, + 0.004405070096254349, + 0.001298015471547842, + 0.0014697766164317727, + 0.002568370196968317, + 0.0023314589634537697, + 0.0013548306887969375, + 0.0013175287749618292, + 0.001960174646228552, + 0.002008951734751463, + 0.006722782272845507, + 0.006975478492677212, + 0.010041989386081696, + 0.0071839275769889355, + 0.015092257410287857, + 0.007605648133903742, + 0.008713917806744576, + 0.0059727877378463745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.014280269853770733 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Van Gogh Post-Impressionist style with thick, visible brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone male figure standing atop a precipice in a vibrant wheatfield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as turbulent with heavy, swirling clouds in deep blue and cyan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man dressed in worn, dark, rustic clothing and standing frozen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's chest visibly rising and falling with rapid, labored breaths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022067499812692404 + ], + "opposite_logprob_tokens": [ + -6.127206802368164 + ], + "logprob_reward": -0.0022067499812692404, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022067499812692404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06976161897182465, + 0.03679116442799568, + 0.029189525172114372, + 0.02103876695036888, + 0.4135260581970215, + 0.1177610531449318, + 0.18981656432151794, + 0.01857985369861126, + 0.04702766239643097, + 0.056507762521505356 + ], + "credit_map": [ + 0.004666878376156092, + 0.004708402790129185, + 0.012103213928639889, + 0.013421973213553429, + 0.010628199204802513, + 0.004074565134942532, + 0.010826368816196918, + 0.009332019835710526, + 0.0036444799043238163, + 0.0028360786382108927, + 0.0061264424584805965, + 0.004921521060168743, + 0.005963142029941082, + 0.003234738716855645, + 0.004677994642406702, + 0.005386766046285629, + 0.0020569744519889355, + 0.002375934971496463, + 0.003647006116807461, + 0.003516498953104019, + 0.0055598109029233456, + 0.0026480441447347403, + 0.004850071854889393, + 0.004535184241831303, + 0.0018699794309213758, + 0.001329176826402545, + 0.003998041618615389, + 0.0029501982498914003, + 0.0031031460966914892, + 0.001174470060504973, + 0.003940411377698183, + 0.002673342591151595, + 0.010979914106428623, + 0.014846144244074821, + 0.1205989420413971, + 0.17509515583515167, + 0.009027489461004734, + 0.030444074422121048, + 0.022314339876174927, + 0.03022002801299095, + 0.007149296812713146, + 0.015478946268558502, + 0.011575108394026756, + 0.022169025614857674, + 0.008017380721867085, + 0.03740973398089409, + 0.006247008219361305, + 0.009714544750750065, + 0.011646304279565811, + 0.008900749497115612, + 0.0690709576010704, + 0.022264979779720306, + 0.008932020515203476, + 0.025674154981970787, + 0.010574481450021267, + 0.032752908766269684, + 0.0012312845792621374, + 0.0017863159300759435, + 0.0041639311239123344, + 0.0028911340050399303, + 0.002499423222616315, + 0.0013506150571629405, + 0.002634153002873063, + 0.0020229960791766644, + 0.0034179333597421646, + 0.006230797152966261, + 0.007079664617776871, + 0.008150934241712093, + 0.005906370468437672, + 0.006065359339118004, + 0.004765699151903391, + 0.005410903599113226, + 0.004704753402620554, + 0.008868744596838951, + 0.008257562294602394, + 0.010942312888801098, + 0.00448918342590332, + 0.007564404979348183, + 0.005950066726654768, + 0.0057307337410748005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 5 + ], + "selected_reward": -0.0022067499812692404 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Van Gogh Post-Impressionist style with thick, visible brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone male figure standing atop a precipice in a vibrant wheatfield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as turbulent with heavy, swirling clouds in deep blue and cyan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man dressed in worn, dark, rustic clothing and standing frozen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's chest visibly rising and falling with rapid, labored breaths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005259842611849308 + ], + "opposite_logprob_tokens": [ + -5.255259990692139 + ], + "logprob_reward": -0.005259842611849308, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005259842611849308, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08171536028385162, + 0.03618216887116432, + 0.02801455743610859, + 0.019014613702893257, + 0.3494701385498047, + 0.06595660746097565, + 0.16758722066879272, + 0.014472298324108124, + 0.013116220943629742, + 0.22447079420089722 + ], + "credit_map": [ + 0.007733045145869255, + 0.006808731704950333, + 0.010437563993036747, + 0.013212214224040508, + 0.013607166707515717, + 0.005352035164833069, + 0.012617163360118866, + 0.011947434395551682, + 0.004320149775594473, + 0.0025422885082662106, + 0.005316749215126038, + 0.005614818539470434, + 0.004944631364196539, + 0.0034911721013486385, + 0.004827470052987337, + 0.005124883260577917, + 0.0039566075429320335, + 0.0018837612587958574, + 0.0037780876737087965, + 0.00403627147898078, + 0.002461802214384079, + 0.003494673641398549, + 0.0029762755148112774, + 0.005427074618637562, + 0.002076878445222974, + 0.0015217525651678443, + 0.003019237658008933, + 0.002387874759733677, + 0.0033905187156051397, + 0.0015173160936683416, + 0.002081584418192506, + 0.0030194486025720835, + 0.00966778676956892, + 0.012204804457724094, + 0.09547517448663712, + 0.14977684617042542, + 0.00792776606976986, + 0.024327477440238, + 0.020125988870859146, + 0.029964260756969452, + 0.004017534200102091, + 0.012582916766405106, + 0.005360741168260574, + 0.0044519929215312, + 0.005210569594055414, + 0.025268396362662315, + 0.0034766884054988623, + 0.005587767343968153, + 0.007534791715443134, + 0.01273019053041935, + 0.057794272899627686, + 0.025540389120578766, + 0.006208889186382294, + 0.03252967819571495, + 0.006619580555707216, + 0.01862940564751625, + 0.0010051450226455927, + 0.001383833005093038, + 0.002679675118997693, + 0.0021372565533965826, + 0.002234966494143009, + 0.0011191779049113393, + 0.001985084731131792, + 0.001927157398313284, + 0.0011364540550857782, + 0.001200360246002674, + 0.0018137021688744426, + 0.0016096773324534297, + 0.002873137826099992, + 0.0011846917914226651, + 0.001970857847481966, + 0.0013273381628096104, + 0.014039980247616768, + 0.02899204194545746, + 0.03383687511086464, + 0.04039328917860985, + 0.026112107560038567, + 0.027611982077360153, + 0.027626140043139458, + 0.025858353823423386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 6 + ], + "selected_reward": -0.005259842611849308 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Van Gogh Post-Impressionist style with thick, visible brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone male figure standing atop a precipice in a vibrant wheatfield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as turbulent with heavy, swirling clouds in deep blue and cyan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man dressed in worn, dark, rustic clothing and standing frozen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's chest visibly rising and falling with rapid, labored breaths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018198659643530846 + ], + "opposite_logprob_tokens": [ + -4.018198490142822 + ], + "logprob_reward": -0.018198659643530846, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018198659643530846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06750527769327164, + 0.028796618804335594, + 0.018622133880853653, + 0.014649255201220512, + 0.4236408472061157, + 0.291937917470932, + 0.07559269666671753, + 0.012804285623133183, + 0.03750419244170189, + 0.02894682064652443 + ], + "credit_map": [ + 0.005411733407527208, + 0.0051429723389446735, + 0.008344373665750027, + 0.010632028803229332, + 0.011108111590147018, + 0.0066427988931536674, + 0.012332294136285782, + 0.007890958338975906, + 0.0036361028905957937, + 0.002393500879406929, + 0.005838882178068161, + 0.004832107573747635, + 0.0036087350454181433, + 0.0021908774506300688, + 0.003125567454844713, + 0.003170842071995139, + 0.0013227390591055155, + 0.0010979529470205307, + 0.003078912850469351, + 0.0035275949630886316, + 0.002981470664963126, + 0.0013395411660894752, + 0.0022463304921984673, + 0.0030275906901806593, + 0.001587901497259736, + 0.001282734447158873, + 0.002620777115225792, + 0.0014384485548362136, + 0.0025338437408208847, + 0.0014511637855321169, + 0.0018461209256201982, + 0.0018882645526900887, + 0.011849891394376755, + 0.015057897195219994, + 0.12078708410263062, + 0.1793491393327713, + 0.011241136118769646, + 0.029976170510053635, + 0.0236479751765728, + 0.0317315049469471, + 0.01846877858042717, + 0.030074525624513626, + 0.03444439917802811, + 0.061143890023231506, + 0.01776961423456669, + 0.0822710171341896, + 0.01615706831216812, + 0.03160860016942024, + 0.004710743669420481, + 0.007665139157325029, + 0.017191609367728233, + 0.013513774611055851, + 0.010508122853934765, + 0.006997804623097181, + 0.00862040463835001, + 0.006385087966918945, + 0.0010399287566542625, + 0.0012692312011495233, + 0.001857321010902524, + 0.0020728032104671, + 0.002145499223843217, + 0.001558830146677792, + 0.0014876421773806214, + 0.0013730295468121767, + 0.0024598927702754736, + 0.004698923788964748, + 0.0053841425105929375, + 0.006672620307654142, + 0.004591979086399078, + 0.0045584095641970634, + 0.00420067785307765, + 0.004937548190355301, + 0.002610850380733609, + 0.004028293304145336, + 0.004503826145082712, + 0.0052941469475626945, + 0.0035570517648011446, + 0.003257114440202713, + 0.0029284991323947906, + 0.0027670366689562798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.018198659643530846 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Van Gogh Post-Impressionist style with thick, visible brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone male figure standing atop a precipice in a vibrant wheatfield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as turbulent with heavy, swirling clouds in deep blue and cyan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man dressed in worn, dark, rustic clothing and standing frozen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's chest visibly rising and falling with rapid, labored breaths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7579145431518555 + ], + "opposite_logprob_tokens": [ + -0.6329145431518555 + ], + "logprob_reward": -0.7579145431518555, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7579145431518555, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.04734412953257561, + 0.015219754539430141, + 0.011411020532250404, + 0.00988987646996975, + 0.5416532754898071, + 0.18235906958580017, + 0.11967021971940994, + 0.008479656651616096, + 0.04792001098394394, + 0.016052929684519768 + ], + "credit_map": [ + 0.006455000024288893, + 0.0038800553884357214, + 0.0069205304607748985, + 0.007489666808396578, + 0.005771878641098738, + 0.004005869850516319, + 0.005758509039878845, + 0.007062622811645269, + 0.002434060676023364, + 0.0011740088229998946, + 0.0024881556164473295, + 0.0017391863511875272, + 0.0019285293528810143, + 0.0016179012600332499, + 0.0019622237887233496, + 0.0018756884383037686, + 0.0009870728245005012, + 0.0009291169117204845, + 0.0016622331459075212, + 0.0013534564059227705, + 0.0025110074784606695, + 0.0011305527295917273, + 0.0016902693314477801, + 0.0011473129270598292, + 0.0007232766947709024, + 0.0006034920806996524, + 0.0016805148916319013, + 0.0013106812257319689, + 0.0019958247430622578, + 0.000717603019438684, + 0.0017461315728724003, + 0.0011123539879918098, + 0.014527928084135056, + 0.017609909176826477, + 0.15618006885051727, + 0.23317940533161163, + 0.01314642932265997, + 0.03277628496289253, + 0.030102739110589027, + 0.044130533933639526, + 0.011175180785357952, + 0.01834341324865818, + 0.020721081644296646, + 0.041116632521152496, + 0.010903042741119862, + 0.05055404081940651, + 0.010121861472725868, + 0.019423849880695343, + 0.005753194913268089, + 0.007540302351117134, + 0.043510183691978455, + 0.01750020682811737, + 0.003762805135920644, + 0.020735779777169228, + 0.00466673169285059, + 0.016201026737689972, + 0.0008705658256076276, + 0.0010106024565175176, + 0.0018669990822672844, + 0.0009163892245851457, + 0.0011355768656358123, + 0.0005765273235738277, + 0.0008757032337598503, + 0.0012272928142920136, + 0.003415808081626892, + 0.006387599743902683, + 0.006447379942983389, + 0.008394538424909115, + 0.005813094321638346, + 0.006340451538562775, + 0.005575748160481453, + 0.005545397754758596, + 0.001636343658901751, + 0.0016076999017968774, + 0.0028127548284828663, + 0.0021931063383817673, + 0.0029609715566039085, + 0.0014504718128591776, + 0.0018938189605250955, + 0.0014977639075368643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.7579145431518555 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "static, sweeping wide shot, rendered in Van Gogh's signature Post-Impressionist style, captures a lone male figure standing atop a precipice amidst a vibrant, swirling wheatfield beneath a turbulent sky. The entire scene is constructed with thick, visible, unsentimental brushstrokes, imbuing both the landscape and the figure with a tactile, three-dimensional reality. The field is an explosive sea of gold and ochre, with the wheat stalks bending and twisting in a chaotic, windswept dance that reflects his inner state. Above, heavy, swirling clouds tumble across a deep blue and cyan sky, the harsh afternoon light casting dramatic, elongated shadows. The man, thin and dressed in worn, dark, rustic clothing, stands frozen. His chest rises and falls with his rapid, labored breaths, a subtle yet powerful micro-movement that suggests he has just stopped fleeing an unseen terror. His eyes, wide open and filled with frantic, manic energy, gaze upon the turbulent landscape as if it were a direct, painterly expression of his own morbid fear and psychological anguish.", + "seed": 20399, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_4/index_8596_seed_20399_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2765143098309636, + "vlm_logprob": -0.2765143098309636, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.2765143098309636, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Van Gogh Post-Impressionist style with thick, visible brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone male figure standing atop a precipice in a vibrant wheatfield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as turbulent with heavy, swirling clouds in deep blue and cyan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man dressed in worn, dark, rustic clothing and standing frozen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's chest visibly rising and falling with rapid, labored breaths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01618027500808239 + ], + "opposite_logprob_tokens": [ + -4.266180038452148 + ], + "logprob_reward": -0.01618027500808239, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01618027500808239, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5033165812492371, + 0.07550562918186188, + 0.039515551179647446, + 0.06617449223995209, + 0.06485000252723694, + 0.06472426652908325, + 0.03349060192704201, + 0.02728981338441372, + 0.02243882417678833, + 0.10269427299499512 + ], + "credit_map": [ + 0.04903789982199669, + 0.03946471959352493, + 0.1181488186120987, + 0.023409320041537285, + 0.04685039818286896, + 0.07581236213445663, + 0.05913882330060005, + 0.09145420044660568, + 0.00780415628105402, + 0.0080737778916955, + 0.019976099953055382, + 0.004836483392864466, + 0.006415842100977898, + 0.011045747436583042, + 0.005872984882444143, + 0.011480530723929405, + 0.003914817236363888, + 0.002843653317540884, + 0.011497319675981998, + 0.0028271093033254147, + 0.0027419058606028557, + 0.0038366999942809343, + 0.0035987391602247953, + 0.008255301974713802, + 0.006729469634592533, + 0.0056229932233691216, + 0.014435159973800182, + 0.00495934160426259, + 0.006540308706462383, + 0.009083127602934837, + 0.004733514506369829, + 0.014070572331547737, + 0.006704244296997786, + 0.006200467702001333, + 0.010205541737377644, + 0.0044808764941990376, + 0.008119490928947926, + 0.013372237794101238, + 0.005423324182629585, + 0.010343815200030804, + 0.007454011123627424, + 0.010218522511422634, + 0.01139570027589798, + 0.0038193496875464916, + 0.007367268204689026, + 0.007481355685740709, + 0.004809224512428045, + 0.01217882800847292, + 0.0034884572960436344, + 0.0034336838871240616, + 0.007631311193108559, + 0.0021924173925071955, + 0.0029720175080001354, + 0.004253477789461613, + 0.0030565494671463966, + 0.006462686229497194, + 0.002499652560800314, + 0.0028861104510724545, + 0.004834100138396025, + 0.0018332561012357473, + 0.0026309287641197443, + 0.004073940683156252, + 0.003432301338762045, + 0.0050995200872421265, + 0.0017898693913593888, + 0.0030028813052922487, + 0.004834556952118874, + 0.0020522510167211294, + 0.0020599954295903444, + 0.00160486891400069, + 0.0027652219869196415, + 0.004329179413616657, + 0.009748168289661407, + 0.008311621844768524, + 0.0251933503895998, + 0.0060681612230837345, + 0.011234775185585022, + 0.01126655749976635, + 0.008803904987871647, + 0.022067733108997345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.01618027500808239 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Van Gogh Post-Impressionist style with thick, visible brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone male figure standing atop a precipice in a vibrant wheatfield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as turbulent with heavy, swirling clouds in deep blue and cyan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man dressed in worn, dark, rustic clothing and standing frozen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's chest visibly rising and falling with rapid, labored breaths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036328528076410294 + ], + "opposite_logprob_tokens": [ + -5.62863302230835 + ], + "logprob_reward": -0.0036328528076410294, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036328528076410294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3798404932022095, + 0.10645151138305664, + 0.06863169372081757, + 0.037544846534729004, + 0.07468920201063156, + 0.06786224246025085, + 0.0373590923845768, + 0.04221954941749573, + 0.08547462522983551, + 0.09992673993110657 + ], + "credit_map": [ + 0.03378545492887497, + 0.04908427223563194, + 0.108219675719738, + 0.023276951164007187, + 0.027611741796135902, + 0.029419198632240295, + 0.03468303009867668, + 0.07376016676425934, + 0.011969998478889465, + 0.025159567594528198, + 0.015032467432320118, + 0.010204042308032513, + 0.010811615735292435, + 0.009337243624031544, + 0.01130891777575016, + 0.012627660296857357, + 0.006762912031263113, + 0.007610737346112728, + 0.01666606031358242, + 0.004799787420779467, + 0.005711392033845186, + 0.007949299179017544, + 0.0061891269870102406, + 0.012942379340529442, + 0.003504115156829357, + 0.004922946449369192, + 0.008458787575364113, + 0.003750912845134735, + 0.0025495095178484917, + 0.0033358638174831867, + 0.003516118973493576, + 0.007506588939577341, + 0.006939344108104706, + 0.008326745592057705, + 0.01547249685972929, + 0.00464703468605876, + 0.008007895201444626, + 0.008603078313171864, + 0.007151895668357611, + 0.015540710650384426, + 0.006623922381550074, + 0.004104031249880791, + 0.018603887408971786, + 0.005829863715916872, + 0.003593503963202238, + 0.004960756748914719, + 0.007040603086352348, + 0.017105668783187866, + 0.003644551383331418, + 0.007186501286923885, + 0.005902557168155909, + 0.00445503368973732, + 0.0026903548277914524, + 0.003287946805357933, + 0.003454806050285697, + 0.006737339310348034, + 0.004695521201938391, + 0.0038816030137240887, + 0.00992915965616703, + 0.002995437942445278, + 0.005389055702835321, + 0.003717064391821623, + 0.004307209979742765, + 0.007304498925805092, + 0.009152551181614399, + 0.00896022655069828, + 0.02439860999584198, + 0.00514037674292922, + 0.004527004901319742, + 0.005526668392121792, + 0.008748098276555538, + 0.019021090120077133, + 0.01059116329997778, + 0.007299467921257019, + 0.029698124155402184, + 0.0069686234928667545, + 0.008552069775760174, + 0.007318912539631128, + 0.009830273687839508, + 0.019668105989694595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0036328528076410294 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Van Gogh Post-Impressionist style with thick, visible brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone male figure standing atop a precipice in a vibrant wheatfield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as turbulent with heavy, swirling clouds in deep blue and cyan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man dressed in worn, dark, rustic clothing and standing frozen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's chest visibly rising and falling with rapid, labored breaths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812633082270622 + ], + "opposite_logprob_tokens": [ + -3.288126230239868 + ], + "logprob_reward": -0.03812633082270622, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812633082270622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29100358486175537, + 0.1303035020828247, + 0.08334393054246902, + 0.053603291511535645, + 0.07955127209424973, + 0.03494363650679588, + 0.02570337988436222, + 0.026375416666269302, + 0.08064407855272293, + 0.19452796876430511 + ], + "credit_map": [ + 0.030021172016859055, + 0.030848732218146324, + 0.07031777501106262, + 0.014731225557625294, + 0.02485688403248787, + 0.03227553516626358, + 0.03436732292175293, + 0.05358489975333214, + 0.011761008761823177, + 0.008669080212712288, + 0.03544346243143082, + 0.0073885368183255196, + 0.010117905214428902, + 0.011739896610379219, + 0.01188019197434187, + 0.03330341354012489, + 0.008187972009181976, + 0.006064866203814745, + 0.022009996697306633, + 0.006610000506043434, + 0.006192933302372694, + 0.00866173580288887, + 0.007038583047688007, + 0.018577834591269493, + 0.004591061733663082, + 0.004489146173000336, + 0.015568218193948269, + 0.0028803939931094646, + 0.0034939751494675875, + 0.004089608788490295, + 0.004124773200601339, + 0.014366110786795616, + 0.007829603739082813, + 0.005976221524178982, + 0.01575133390724659, + 0.004567962605506182, + 0.008245746605098248, + 0.012347475625574589, + 0.007288528606295586, + 0.017544390633702278, + 0.003616554895415902, + 0.004097261466085911, + 0.008185691200196743, + 0.002611317206174135, + 0.002342667430639267, + 0.0032915968913584948, + 0.0031015987042337656, + 0.007696943357586861, + 0.0023200272116810083, + 0.0023505219724029303, + 0.006619465071707964, + 0.0024444942828267813, + 0.001385910320095718, + 0.0019873313140124083, + 0.0021631820127367973, + 0.006432446651160717, + 0.002692884299904108, + 0.0021654947195202112, + 0.007228993810713291, + 0.0018311308231204748, + 0.0020000378135591745, + 0.0027222305070608854, + 0.002345783170312643, + 0.00538885872811079, + 0.00858567375689745, + 0.00477557722479105, + 0.023904617875814438, + 0.00439744908362627, + 0.004837110638618469, + 0.006448282394558191, + 0.006619276013225317, + 0.021076079457998276, + 0.01767486147582531, + 0.012782554142177105, + 0.05686338618397713, + 0.011389859952032566, + 0.018748508766293526, + 0.01839720830321312, + 0.01912170648574829, + 0.039549849927425385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.03812633082270622 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Van Gogh Post-Impressionist style with thick, visible brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone male figure standing atop a precipice in a vibrant wheatfield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as turbulent with heavy, swirling clouds in deep blue and cyan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man dressed in worn, dark, rustic clothing and standing frozen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's chest visibly rising and falling with rapid, labored breaths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011088770814239979 + ], + "opposite_logprob_tokens": [ + -4.511088848114014 + ], + "logprob_reward": -0.011088770814239979, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011088770814239979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3629322052001953, + 0.10158345848321915, + 0.06108982115983963, + 0.03874053433537483, + 0.1032939925789833, + 0.03814354166388512, + 0.030346186831593513, + 0.06800983101129532, + 0.07709463685750961, + 0.11876582354307175 + ], + "credit_map": [ + 0.0345158651471138, + 0.061738356947898865, + 0.06292878091335297, + 0.05224348232150078, + 0.026691237464547157, + 0.02503611147403717, + 0.04958348348736763, + 0.050194889307022095, + 0.008936767466366291, + 0.016739292070269585, + 0.017174160107970238, + 0.01068887859582901, + 0.008278374560177326, + 0.008262705989181995, + 0.012035660445690155, + 0.019467607140541077, + 0.004958294797688723, + 0.008183097466826439, + 0.0123710623010993, + 0.007652288302779198, + 0.003992641344666481, + 0.004264400340616703, + 0.007869496010243893, + 0.011798538267612457, + 0.0033966635819524527, + 0.0047632846981287, + 0.0071959178894758224, + 0.006168781314045191, + 0.0018772657494992018, + 0.0025822727475315332, + 0.004185563884675503, + 0.008570784702897072, + 0.010136658325791359, + 0.023210112005472183, + 0.013544834218919277, + 0.016198834404349327, + 0.006353091914206743, + 0.008338065817952156, + 0.012773817405104637, + 0.012738567776978016, + 0.004533950239419937, + 0.006296123843640089, + 0.005414262413978577, + 0.005950829479843378, + 0.003410014323890209, + 0.0027847110759466887, + 0.004165357444435358, + 0.0055882916785776615, + 0.002847788855433464, + 0.004026768263429403, + 0.006806378252804279, + 0.0033602535258978605, + 0.002708199666813016, + 0.002667477121576667, + 0.003174907062202692, + 0.004754411522299051, + 0.006247154902666807, + 0.011721054092049599, + 0.008258931338787079, + 0.011521990410983562, + 0.005093283485621214, + 0.00609447481110692, + 0.008743803948163986, + 0.010329139418900013, + 0.010053368285298347, + 0.005645197816193104, + 0.021775785833597183, + 0.004284876864403486, + 0.004971651826053858, + 0.005986390635371208, + 0.006534087471663952, + 0.017843274399638176, + 0.010942935943603516, + 0.015707528218626976, + 0.02958589233458042, + 0.012482077814638615, + 0.008073313161730766, + 0.007966599427163601, + 0.014808262698352337, + 0.019199201837182045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.011088770814239979 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Van Gogh Post-Impressionist style with thick, visible brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone male figure standing atop a precipice in a vibrant wheatfield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as turbulent with heavy, swirling clouds in deep blue and cyan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man dressed in worn, dark, rustic clothing and standing frozen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's chest visibly rising and falling with rapid, labored breaths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.3135433197021484 + ], + "opposite_logprob_tokens": [ + -0.31354328989982605 + ], + "logprob_reward": -1.3135433197021484, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.3135433197021484, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.4119330942630768, + 0.10706514865159988, + 0.06158054247498512, + 0.0467873252928257, + 0.06344550102949142, + 0.06025230884552002, + 0.034459665417671204, + 0.032992903143167496, + 0.08088372647762299, + 0.10059982538223267 + ], + "credit_map": [ + 0.0433233268558979, + 0.05830458179116249, + 0.08570882678031921, + 0.03787277266383171, + 0.04068068414926529, + 0.049574244767427444, + 0.043761659413576126, + 0.0527070090174675, + 0.01331092044711113, + 0.027712132781744003, + 0.01833214983344078, + 0.0102461539208889, + 0.009358963929116726, + 0.005927570164203644, + 0.008217458613216877, + 0.013959798030555248, + 0.007138053886592388, + 0.013675118796527386, + 0.010949044488370419, + 0.008118598721921444, + 0.005551149602979422, + 0.003280751407146454, + 0.005078153219074011, + 0.007789671886712313, + 0.004900783766061068, + 0.007511802017688751, + 0.008074763230979443, + 0.005584120284765959, + 0.003920052200555801, + 0.0048900083638727665, + 0.003995606210082769, + 0.007910188287496567, + 0.006227624136954546, + 0.0061987629160285, + 0.011689001694321632, + 0.007515501696616411, + 0.006771173793822527, + 0.008177492767572403, + 0.006528932601213455, + 0.010337013751268387, + 0.006180348806083202, + 0.013273094780743122, + 0.007981694303452969, + 0.010168038308620453, + 0.004468008875846863, + 0.00410050293430686, + 0.00794204417616129, + 0.0061385752633214, + 0.003590224776417017, + 0.005344193894416094, + 0.007583942264318466, + 0.004000683315098286, + 0.0021925061009824276, + 0.0026774967554956675, + 0.003924214746803045, + 0.005146403796970844, + 0.0034583245869725943, + 0.005444983020424843, + 0.0055110570974648, + 0.004156799521297216, + 0.0026599306147545576, + 0.002403029939159751, + 0.003778657875955105, + 0.005580121651291847, + 0.007239916827529669, + 0.008253989741206169, + 0.022052055224776268, + 0.005267719738185406, + 0.0049012755043804646, + 0.006297449581325054, + 0.006415875628590584, + 0.02045544981956482, + 0.009395993314683437, + 0.006435994524508715, + 0.02586979791522026, + 0.00838012807071209, + 0.009711726568639278, + 0.009249391034245491, + 0.01067867036908865, + 0.020878124982118607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -1.3135433197021484 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_5/index_tensor([1182, 1182])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_5/index_tensor([1182, 1182])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..66c7b53ee4de3bf9a3c6aada94e3e6e02fa0f80b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_5/index_tensor([1182, 1182])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, upward-tilting long shot pierces the pitch-black, nearly lightless depths of the ocean, following the graceful ascent of a school of luminous jellyfish. This silent procession, comprised of dozens of jellyfish of varying shapes, emits a soft, rhythmic bioluminescence from within their translucent, gelatinous umbrellas, shifting between ethereal cyan, dreamy pink-purple, and soft emerald green. As they propel upward through graceful contractions and expansions, countless long, slender, silken ribbon-like tentacles drift lazily behind them, carving trails of shimmering light through the darkness. Their own light, the sole source of illumination in the inky black waters, faintly illuminates the tiny particles of marine snow suspended around them, creating an effect reminiscent of stardust. The entire scene is shrouded in a profound sense of tranquility and mystery. The jellyfish's effortless ascent contrasts sharply with the endless darkness of the deep, creating a visual spectacle both otherworldly and vibrant.", + "seed": 5530, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_5/index_1182_seed_5530_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.13807003661058842, + "vlm_logprob": -0.13807003661058842, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.13807003661058842, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment pitch-black with no external light sources?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dozens of jellyfish with varying shapes visible ascending in a school?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jellyfish emit bioluminescence in cyan, pink-purple, and emerald green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jellyfish possess long, slender, silken ribbon-like tentacles drifting behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42137470841407776 + ], + "opposite_logprob_tokens": [ + -1.1713746786117554 + ], + "logprob_reward": -0.42137470841407776, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.42137470841407776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14828695356845856, + 0.07448448240756989, + 0.15183129906654358, + 0.061198119074106216, + 0.07592066377401352, + 0.05716188997030258, + 0.07148940861225128, + 0.11385967582464218, + 0.063267283141613, + 0.18250027298927307 + ], + "credit_map": [ + 0.03055853210389614, + 0.024726426228880882, + 0.02511805295944214, + 0.020519793033599854, + 0.012101570144295692, + 0.010540093295276165, + 0.015156111679971218, + 0.009566374123096466, + 0.010974649339914322, + 0.0066206082701683044, + 0.014352208934724331, + 0.012407874688506126, + 0.006398130673915148, + 0.007218524347990751, + 0.006721318233758211, + 0.009791163727641106, + 0.025064289569854736, + 0.041097208857536316, + 0.021380988880991936, + 0.018565574660897255, + 0.012778921984136105, + 0.011548126116394997, + 0.011058269999921322, + 0.010337897576391697, + 0.006892816629260778, + 0.013911159709095955, + 0.010129902511835098, + 0.01076507568359375, + 0.0054588583298027515, + 0.00295761926099658, + 0.005303279496729374, + 0.005779405124485493, + 0.009506015107035637, + 0.013115041889250278, + 0.013381228782236576, + 0.01660345122218132, + 0.005054693203419447, + 0.006301199551671743, + 0.006616777274757624, + 0.0053422474302351475, + 0.007031456101685762, + 0.006133398972451687, + 0.008904870599508286, + 0.0145479217171669, + 0.00426436634734273, + 0.004082715604454279, + 0.006952603347599506, + 0.005244553089141846, + 0.009896529838442802, + 0.009069584310054779, + 0.01043655164539814, + 0.01120683178305626, + 0.012570459395647049, + 0.005929412320256233, + 0.006821661256253719, + 0.005558369681239128, + 0.018996933475136757, + 0.017348842695355415, + 0.020530832931399345, + 0.013107082806527615, + 0.010628576390445232, + 0.009884153492748737, + 0.011088474653661251, + 0.012274762615561485, + 0.010087946429848671, + 0.004982237238436937, + 0.017564866691827774, + 0.009562231600284576, + 0.006492112297564745, + 0.0031702162232249975, + 0.00398299191147089, + 0.0074246712028980255, + 0.017246538773179054, + 0.015341242775321007, + 0.03303994610905647, + 0.028917130082845688, + 0.015705617144703865, + 0.01884704828262329, + 0.01947651617228985, + 0.033926211297512054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.42137470841407776 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment pitch-black with no external light sources?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dozens of jellyfish with varying shapes visible ascending in a school?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jellyfish emit bioluminescence in cyan, pink-purple, and emerald green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jellyfish possess long, slender, silken ribbon-like tentacles drifting behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22561147809028625 + ], + "opposite_logprob_tokens": [ + -1.6006114482879639 + ], + "logprob_reward": -0.22561147809028625, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22561147809028625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10177405923604965, + 0.02690792642533779, + 0.1093263030052185, + 0.07218457758426666, + 0.04408521205186844, + 0.23134787380695343, + 0.22386550903320312, + 0.12584556639194489, + 0.024519184604287148, + 0.04014376923441887 + ], + "credit_map": [ + 0.015154179185628891, + 0.016069291159510612, + 0.009864217601716518, + 0.008494479581713676, + 0.03456902131438255, + 0.005729853641241789, + 0.006390515714883804, + 0.005502498708665371, + 0.0040343995206058025, + 0.0027389286551624537, + 0.00392657658085227, + 0.003997769206762314, + 0.0031269884202629328, + 0.0025017093867063522, + 0.0034523289650678635, + 0.0031292270869016647, + 0.01883910782635212, + 0.029057037085294724, + 0.014110477641224861, + 0.016389960423111916, + 0.008261926472187042, + 0.006222526077181101, + 0.008676033467054367, + 0.007769227959215641, + 0.012493591755628586, + 0.010905568487942219, + 0.011403257958590984, + 0.009031625464558601, + 0.007744347210973501, + 0.005591329652816057, + 0.007445794530212879, + 0.0075690564699471, + 0.005513540934771299, + 0.006940828170627356, + 0.007451819721609354, + 0.008307790383696556, + 0.0031528661493211985, + 0.003730745054781437, + 0.004527108743786812, + 0.0044605121947824955, + 0.03003636747598648, + 0.023631462827324867, + 0.03912742808461189, + 0.05794684588909149, + 0.017134785652160645, + 0.013828875496983528, + 0.026257501915097237, + 0.02338460646569729, + 0.05419403687119484, + 0.033410705626010895, + 0.04248164966702461, + 0.03668787330389023, + 0.0164361372590065, + 0.011250411160290241, + 0.013427204452455044, + 0.015977485105395317, + 0.021891212090849876, + 0.019165659323334694, + 0.028976766392588615, + 0.009716706350445747, + 0.017732126638293266, + 0.00666123628616333, + 0.010540841147303581, + 0.01116101536899805, + 0.0032846666872501373, + 0.0022170175798237324, + 0.005270709749311209, + 0.0031222484540194273, + 0.002633688971400261, + 0.0015353818889707327, + 0.0025606846902519464, + 0.0038947872817516327, + 0.0031882792245596647, + 0.005111717619001865, + 0.011715605854988098, + 0.0070551251992583275, + 0.0033496085088700056, + 0.0029022928792983294, + 0.0027471277862787247, + 0.004074011463671923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.22561147809028625 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment pitch-black with no external light sources?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dozens of jellyfish with varying shapes visible ascending in a school?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jellyfish emit bioluminescence in cyan, pink-purple, and emerald green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jellyfish possess long, slender, silken ribbon-like tentacles drifting behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026353858411312103 + ], + "opposite_logprob_tokens": [ + -3.6513538360595703 + ], + "logprob_reward": -0.026353858411312103, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026353858411312103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1814551204442978, + 0.05698392540216446, + 0.0642990693449974, + 0.08500899374485016, + 0.05297299847006798, + 0.07115937024354935, + 0.1423056423664093, + 0.10313298553228378, + 0.12737733125686646, + 0.1153046041727066 + ], + "credit_map": [ + 0.02509636990725994, + 0.03694716840982437, + 0.0178126972168684, + 0.013304117135703564, + 0.056128453463315964, + 0.009623944759368896, + 0.013932004570960999, + 0.008610357530415058, + 0.0064394730143249035, + 0.0073920320719480515, + 0.009204492904245853, + 0.006818020716309547, + 0.007348951883614063, + 0.005060121882706881, + 0.0058977422304451466, + 0.008823086507618427, + 0.010231146588921547, + 0.013905934989452362, + 0.008938945829868317, + 0.007895583286881447, + 0.005041476339101791, + 0.005673547741025686, + 0.006338859908282757, + 0.006273570936173201, + 0.011423974297940731, + 0.014325480908155441, + 0.013747542165219784, + 0.010821778327226639, + 0.007748987525701523, + 0.0058971275575459, + 0.011384799145162106, + 0.009659306146204472, + 0.00881276000291109, + 0.009234366938471794, + 0.010266747325658798, + 0.0076036290265619755, + 0.0029768324457108974, + 0.005803598556667566, + 0.0040181041695177555, + 0.0042569562792778015, + 0.009753315709531307, + 0.00810668058693409, + 0.012201880104839802, + 0.015066657215356827, + 0.005303400568664074, + 0.005032092332839966, + 0.006740872748196125, + 0.008954470045864582, + 0.0264599546790123, + 0.023482339456677437, + 0.023800255730748177, + 0.022558704018592834, + 0.016191279515624046, + 0.009847160428762436, + 0.00827997736632824, + 0.011685975827276707, + 0.01674181967973709, + 0.016074182465672493, + 0.020162364467978477, + 0.00989600084722042, + 0.009995799511671066, + 0.008346442133188248, + 0.008509283885359764, + 0.013407088816165924, + 0.015810098499059677, + 0.01361165102571249, + 0.019764242693781853, + 0.01558870542794466, + 0.01420027855783701, + 0.010599237866699696, + 0.013646474108099937, + 0.02415664866566658, + 0.011864644475281239, + 0.013623179867863655, + 0.017065707594156265, + 0.018138226121664047, + 0.00826551578938961, + 0.012822238728404045, + 0.012554837390780449, + 0.020970256999135017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.026353858411312103 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment pitch-black with no external light sources?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dozens of jellyfish with varying shapes visible ascending in a school?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jellyfish emit bioluminescence in cyan, pink-purple, and emerald green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jellyfish possess long, slender, silken ribbon-like tentacles drifting behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002823535120114684 + ], + "opposite_logprob_tokens": [ + -5.877823352813721 + ], + "logprob_reward": -0.002823535120114684, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002823535120114684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20720522105693817, + 0.03221858665347099, + 0.2542670667171478, + 0.06459742784500122, + 0.24183537065982819, + 0.03782360255718231, + 0.03613506630063057, + 0.045708026736974716, + 0.013954751193523407, + 0.0662548765540123 + ], + "credit_map": [ + 0.030476996675133705, + 0.029543409124016762, + 0.02566090226173401, + 0.022662585601210594, + 0.02760792337357998, + 0.026185564696788788, + 0.022869404405355453, + 0.022198451682925224, + 0.004562124609947205, + 0.002888327930122614, + 0.005683228839188814, + 0.00492089195176959, + 0.0029759726021438837, + 0.0037070843391120434, + 0.003544021863490343, + 0.003936932422220707, + 0.0414094403386116, + 0.0683414489030838, + 0.032155148684978485, + 0.038606613874435425, + 0.019003411754965782, + 0.015117011964321136, + 0.018417613580822945, + 0.02121635712683201, + 0.007934224791824818, + 0.012407420203089714, + 0.010019224137067795, + 0.007514589466154575, + 0.007540979888290167, + 0.004129149951040745, + 0.007603901904076338, + 0.0074479361064732075, + 0.03701848164200783, + 0.04105856269598007, + 0.04610668122768402, + 0.05466800928115845, + 0.010279889218509197, + 0.013158204033970833, + 0.02444419451057911, + 0.01510135643184185, + 0.005472553428262472, + 0.003810775000602007, + 0.006272505037486553, + 0.006788392551243305, + 0.0032900171354413033, + 0.002812961582094431, + 0.004459398798644543, + 0.004916996229439974, + 0.005659724585711956, + 0.004896142054349184, + 0.004464353900402784, + 0.0034024950582534075, + 0.0039976611733436584, + 0.005187148228287697, + 0.003651995211839676, + 0.004875545855611563, + 0.008757655508816242, + 0.0059078228659927845, + 0.008486149832606316, + 0.005522858817130327, + 0.004202206153422594, + 0.00475348113104701, + 0.004120478872209787, + 0.003957376349717379, + 0.0013172124745324254, + 0.0013852565316483378, + 0.0018846327438950539, + 0.0027594908606261015, + 0.0015048759523779154, + 0.0014294376596808434, + 0.0015332034090533853, + 0.002140641678124666, + 0.005832456052303314, + 0.0042953467927873135, + 0.020868197083473206, + 0.012117169797420502, + 0.006377511657774448, + 0.003712097182869911, + 0.007098904810845852, + 0.005953185725957155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.002823535120114684 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment pitch-black with no external light sources?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dozens of jellyfish with varying shapes visible ascending in a school?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jellyfish emit bioluminescence in cyan, pink-purple, and emerald green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jellyfish possess long, slender, silken ribbon-like tentacles drifting behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014186603017151356 + ], + "opposite_logprob_tokens": [ + -4.264186382293701 + ], + "logprob_reward": -0.014186603017151356, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014186603017151356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26721319556236267, + 0.06620854884386063, + 0.09895665943622589, + 0.043560076504945755, + 0.10435044765472412, + 0.026137391105294228, + 0.04195134714245796, + 0.06500973552465439, + 0.03495370224118233, + 0.2516588568687439 + ], + "credit_map": [ + 0.05384048819541931, + 0.025203879922628403, + 0.032131727784872055, + 0.025939567014575005, + 0.06880342960357666, + 0.01953211985528469, + 0.026344366371631622, + 0.015417632646858692, + 0.009182016365230083, + 0.006911640986800194, + 0.011489967815577984, + 0.008626601658761501, + 0.004841635935008526, + 0.010489086620509624, + 0.007279522251337767, + 0.007388085126876831, + 0.015166722238063812, + 0.021605107933282852, + 0.011867821216583252, + 0.01634194701910019, + 0.006624497007578611, + 0.006430433131754398, + 0.0074668279848992825, + 0.013453314080834389, + 0.004354844801127911, + 0.009123321622610092, + 0.006237772759050131, + 0.0066052828915417194, + 0.004734037909656763, + 0.0029090612661093473, + 0.005090383812785149, + 0.004505376797169447, + 0.012635069899260998, + 0.015558479353785515, + 0.016905594617128372, + 0.01713571511209011, + 0.006624350789934397, + 0.01496919710189104, + 0.011952237226068974, + 0.0085698077455163, + 0.0036493013612926006, + 0.0030470264609903097, + 0.005060269962996244, + 0.0024349025916308165, + 0.0028428579680621624, + 0.0025642032269388437, + 0.002308351220563054, + 0.00423048110678792, + 0.004749218001961708, + 0.005902160424739122, + 0.0059791449457407, + 0.0043285200372338295, + 0.005390392150729895, + 0.004967111628502607, + 0.003661821596324444, + 0.006972985807806253, + 0.00975767895579338, + 0.009975939989089966, + 0.009825579822063446, + 0.006815552711486816, + 0.008741609752178192, + 0.007231874857097864, + 0.004863470792770386, + 0.00779804028570652, + 0.00329079432412982, + 0.0044215950183570385, + 0.006028472911566496, + 0.005661080591380596, + 0.004902187269181013, + 0.0032546021975576878, + 0.0028463557828217745, + 0.004548618104308844, + 0.01977935992181301, + 0.02659767121076584, + 0.0670916810631752, + 0.04179420694708824, + 0.022367458790540695, + 0.01755155436694622, + 0.02775583043694496, + 0.028721120208501816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.014186603017151356 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, upward-tilting long shot pierces the pitch-black, nearly lightless depths of the ocean, following the graceful ascent of a school of luminous jellyfish. This silent procession, comprised of dozens of jellyfish of varying shapes, emits a soft, rhythmic bioluminescence from within their translucent, gelatinous umbrellas, shifting between ethereal cyan, dreamy pink-purple, and soft emerald green. As they propel upward through graceful contractions and expansions, countless long, slender, silken ribbon-like tentacles drift lazily behind them, carving trails of shimmering light through the darkness. Their own light, the sole source of illumination in the inky black waters, faintly illuminates the tiny particles of marine snow suspended around them, creating an effect reminiscent of stardust. The entire scene is shrouded in a profound sense of tranquility and mystery. The jellyfish's effortless ascent contrasts sharply with the endless darkness of the deep, creating a visual spectacle both otherworldly and vibrant.", + "seed": 5531, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_5/index_1182_seed_5531_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.13705831868574023, + "vlm_logprob": -0.13705831868574023, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.13705831868574023, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment pitch-black with no external light sources?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dozens of jellyfish with varying shapes visible ascending in a school?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jellyfish emit bioluminescence in cyan, pink-purple, and emerald green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jellyfish possess long, slender, silken ribbon-like tentacles drifting behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5527670383453369 + ], + "opposite_logprob_tokens": [ + -0.9277670383453369 + ], + "logprob_reward": -0.5527670383453369, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5527670383453369, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1717628538608551, + 0.1311352699995041, + 0.07866349071264267, + 0.12346699088811874, + 0.06507483124732971, + 0.1387491226196289, + 0.04645434021949768, + 0.08297843486070633, + 0.048772308975458145, + 0.11294227838516235 + ], + "credit_map": [ + 0.01731397956609726, + 0.020069314166903496, + 0.015475251711905003, + 0.03044297732412815, + 0.029603006318211555, + 0.019014058634638786, + 0.01565207540988922, + 0.024192210286855698, + 0.011539263650774956, + 0.013624642044305801, + 0.013711565174162388, + 0.014327764511108398, + 0.02052135020494461, + 0.02148282341659069, + 0.015318331308662891, + 0.0206095390021801, + 0.008592338301241398, + 0.008778147399425507, + 0.011028585955500603, + 0.015727009624242783, + 0.010916154831647873, + 0.007940073497593403, + 0.005563448183238506, + 0.010117736645042896, + 0.020938815549016, + 0.009893584996461868, + 0.006950760260224342, + 0.007611388340592384, + 0.04029032215476036, + 0.008860473521053791, + 0.010244620963931084, + 0.018677029758691788, + 0.009332816116511822, + 0.016902878880500793, + 0.007784388959407806, + 0.007196678314357996, + 0.005418294109404087, + 0.006683570332825184, + 0.006129929795861244, + 0.005626279395073652, + 0.030055303126573563, + 0.015518751926720142, + 0.011824830435216427, + 0.009919059462845325, + 0.013572507537901402, + 0.016394857317209244, + 0.033719681203365326, + 0.007744140923023224, + 0.007087637670338154, + 0.01017388328909874, + 0.004408066626638174, + 0.004979779943823814, + 0.00219691707752645, + 0.004267398733645678, + 0.0074943858198821545, + 0.005846276879310608, + 0.010603380389511585, + 0.010236122645437717, + 0.00995215866714716, + 0.008250555954873562, + 0.01491359993815422, + 0.007956579327583313, + 0.008722606115043163, + 0.012343432754278183, + 0.008045325987040997, + 0.008023543283343315, + 0.007925701327621937, + 0.004868706688284874, + 0.003215803997591138, + 0.0035654506646096706, + 0.007914222776889801, + 0.005213555879890919, + 0.0168523658066988, + 0.026271510869264603, + 0.016647782176733017, + 0.010295071639120579, + 0.008548559620976448, + 0.005888419225811958, + 0.01649239845573902, + 0.011946185491979122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.5527670383453369 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment pitch-black with no external light sources?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dozens of jellyfish with varying shapes visible ascending in a school?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jellyfish emit bioluminescence in cyan, pink-purple, and emerald green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jellyfish possess long, slender, silken ribbon-like tentacles drifting behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0790298655629158 + ], + "opposite_logprob_tokens": [ + -2.5790297985076904 + ], + "logprob_reward": -0.0790298655629158, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0790298655629158, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08842021971940994, + 0.24961349368095398, + 0.19278912246227264, + 0.08476360142230988, + 0.06769156455993652, + 0.06258124858140945, + 0.051913075149059296, + 0.08428362756967545, + 0.09202469885349274, + 0.02591932751238346 + ], + "credit_map": [ + 0.012266491539776325, + 0.007026459090411663, + 0.009747957810759544, + 0.015377134084701538, + 0.011205272749066353, + 0.00982026755809784, + 0.009696285240352154, + 0.013280353508889675, + 0.029632722958922386, + 0.02107927016913891, + 0.021634627133607864, + 0.02797064371407032, + 0.04238089546561241, + 0.04786444082856178, + 0.020881757140159607, + 0.038169123232364655, + 0.014290961436927319, + 0.010441647842526436, + 0.014495947398245335, + 0.019046766683459282, + 0.0637713149189949, + 0.014564485289156437, + 0.011962536722421646, + 0.04421545937657356, + 0.020696282386779785, + 0.004279032815247774, + 0.004263923969119787, + 0.0036721457727253437, + 0.022451117634773254, + 0.0066898902878165245, + 0.006430215202271938, + 0.016280993819236755, + 0.010330941528081894, + 0.010860017500817776, + 0.010642523877322674, + 0.010527999140322208, + 0.004051190335303545, + 0.007352366577833891, + 0.007807442452758551, + 0.006119078025221825, + 0.016057213768363, + 0.0043392665684223175, + 0.005690423771739006, + 0.005105379968881607, + 0.003694601822644472, + 0.006067200098186731, + 0.01744563691318035, + 0.004181526601314545, + 0.012419736944139004, + 0.006002261769026518, + 0.0034068592358380556, + 0.005093962419778109, + 0.0028905454091727734, + 0.003203056752681732, + 0.012433391064405441, + 0.0064632585272192955, + 0.010111318901181221, + 0.013558760285377502, + 0.012329407036304474, + 0.009056237526237965, + 0.006791521795094013, + 0.009854977950453758, + 0.011293435469269753, + 0.011287963949143887, + 0.015622933395206928, + 0.012327615171670914, + 0.02040361426770687, + 0.011289593763649464, + 0.007617658469825983, + 0.008395595476031303, + 0.009301163256168365, + 0.007066520396620035, + 0.005239417776465416, + 0.0037198453210294247, + 0.005109292920678854, + 0.002451528562232852, + 0.0019641092512756586, + 0.0010626293951645494, + 0.0034533750731498003, + 0.002919131191447377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 8 + ], + "selected_reward": -0.0790298655629158 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment pitch-black with no external light sources?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dozens of jellyfish with varying shapes visible ascending in a school?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jellyfish emit bioluminescence in cyan, pink-purple, and emerald green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jellyfish possess long, slender, silken ribbon-like tentacles drifting behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007634864654392004 + ], + "opposite_logprob_tokens": [ + -4.88263463973999 + ], + "logprob_reward": -0.007634864654392004, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007634864654392004, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11541559547185898, + 0.06993298977613449, + 0.12892411649227142, + 0.05163484811782837, + 0.11257565021514893, + 0.25007379055023193, + 0.08049710839986801, + 0.02951602078974247, + 0.12902846932411194, + 0.03240141272544861 + ], + "credit_map": [ + 0.009029240347445011, + 0.014028769917786121, + 0.013681511394679546, + 0.01620045118033886, + 0.014891832135617733, + 0.015479161404073238, + 0.012911238707602024, + 0.019193388521671295, + 0.007128545083105564, + 0.0073140645399689674, + 0.007583757862448692, + 0.009902452118694782, + 0.008723908104002476, + 0.012381304986774921, + 0.007473591715097427, + 0.009425370022654533, + 0.009375001303851604, + 0.009780213236808777, + 0.015665311366319656, + 0.03204318508505821, + 0.03311438858509064, + 0.009545881301164627, + 0.008392530493438244, + 0.011007615365087986, + 0.01016821339726448, + 0.0034020787570625544, + 0.004347665701061487, + 0.005223128478974104, + 0.011546771042048931, + 0.003580355318263173, + 0.007483231369405985, + 0.0058834063820540905, + 0.01676071435213089, + 0.024898458272218704, + 0.015893815085291862, + 0.010392382740974426, + 0.00820449274033308, + 0.01315520890057087, + 0.011443120427429676, + 0.011827461421489716, + 0.0733199343085289, + 0.015512158162891865, + 0.01629798486828804, + 0.013562158681452274, + 0.01357676088809967, + 0.022017287090420723, + 0.08427312970161438, + 0.011514363810420036, + 0.012623448856174946, + 0.018465740606188774, + 0.006695928517729044, + 0.01242478284984827, + 0.007424482610076666, + 0.004455117974430323, + 0.01192550640553236, + 0.006482097785919905, + 0.0037378345150500536, + 0.0035660648718476295, + 0.004349654074758291, + 0.003778581041842699, + 0.002371008275076747, + 0.0022155342157930136, + 0.003433645237237215, + 0.00606369785964489, + 0.01985478214919567, + 0.01700497232377529, + 0.02589125744998455, + 0.01631910167634487, + 0.012612135149538517, + 0.010051416233181953, + 0.01745614968240261, + 0.009838663041591644, + 0.005067845340818167, + 0.005056401714682579, + 0.004527093376964331, + 0.003179459599778056, + 0.0025084218941628933, + 0.002123773330822587, + 0.004490930587053299, + 0.0054474868811666965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 2 + ], + "selected_reward": -0.007634864654392004 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment pitch-black with no external light sources?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dozens of jellyfish with varying shapes visible ascending in a school?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jellyfish emit bioluminescence in cyan, pink-purple, and emerald green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jellyfish possess long, slender, silken ribbon-like tentacles drifting behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607411541044712 + ], + "opposite_logprob_tokens": [ + -4.141074180603027 + ], + "logprob_reward": -0.01607411541044712, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607411541044712, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10669949650764465, + 0.0406174510717392, + 0.25562262535095215, + 0.08337841182947159, + 0.08281362801790237, + 0.15095584094524384, + 0.12182801961898804, + 0.022382183000445366, + 0.08853393793106079, + 0.047168415039777756 + ], + "credit_map": [ + 0.020886750891804695, + 0.00819534994661808, + 0.015011526644229889, + 0.013901333324611187, + 0.012413023971021175, + 0.011286268010735512, + 0.008563084527850151, + 0.016442159190773964, + 0.005077867768704891, + 0.004380465019494295, + 0.003850489854812622, + 0.0052112857811152935, + 0.0066913762129843235, + 0.005460215266793966, + 0.004178021568804979, + 0.0057677277363836765, + 0.018451416864991188, + 0.015523086301982403, + 0.015147245489060879, + 0.02220151573419571, + 0.08974895626306534, + 0.016778690740466118, + 0.012675929814577103, + 0.06509578227996826, + 0.012006345205008984, + 0.004969226196408272, + 0.0033845698926597834, + 0.004060877952724695, + 0.024918029084801674, + 0.006861990783363581, + 0.006065386813133955, + 0.0211119893938303, + 0.013798091560602188, + 0.025911839678883553, + 0.006194073241204023, + 0.006328904069960117, + 0.007354406639933586, + 0.00981945265084505, + 0.006383254658430815, + 0.007023608777672052, + 0.0475524440407753, + 0.007529554422944784, + 0.007042421028017998, + 0.006689241621643305, + 0.005757317412644625, + 0.01271603349596262, + 0.056051142513751984, + 0.0076176817528903484, + 0.015157890506088734, + 0.027543047443032265, + 0.01224613469094038, + 0.025230474770069122, + 0.007942236959934235, + 0.006498713046312332, + 0.01583278365433216, + 0.011376737616956234, + 0.0027312245219945908, + 0.003361037001013756, + 0.0030418673995882273, + 0.00203007273375988, + 0.0033553631510585546, + 0.0019791012164205313, + 0.0031572759617120028, + 0.0027262421790510416, + 0.013120102696120739, + 0.011273817159235477, + 0.018131570890545845, + 0.00871234480291605, + 0.008595925755798817, + 0.005837529432028532, + 0.01660185493528843, + 0.006260786205530167, + 0.006207604892551899, + 0.0060176909901201725, + 0.008600578643381596, + 0.004649740178138018, + 0.003090376267209649, + 0.0030782045796513557, + 0.00816489290446043, + 0.007359327282756567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 6 + ], + "selected_reward": -0.01607411541044712 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment pitch-black with no external light sources?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dozens of jellyfish with varying shapes visible ascending in a school?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jellyfish emit bioluminescence in cyan, pink-purple, and emerald green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jellyfish possess long, slender, silken ribbon-like tentacles drifting behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978570945560932 + ], + "opposite_logprob_tokens": [ + -3.529785633087158 + ], + "logprob_reward": -0.02978570945560932, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978570945560932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12145765125751495, + 0.08342040330171585, + 0.07818169891834259, + 0.14152903854846954, + 0.025380879640579224, + 0.16144248843193054, + 0.14414635300636292, + 0.11852464079856873, + 0.09305573254823685, + 0.032861120998859406 + ], + "credit_map": [ + 0.0166628360748291, + 0.0106486976146698, + 0.015811361372470856, + 0.01704154536128044, + 0.014666196890175343, + 0.01652669906616211, + 0.013368808664381504, + 0.0167315024882555, + 0.006773416418582201, + 0.008107761852443218, + 0.007167682517319918, + 0.010377618484199047, + 0.015399578958749771, + 0.01535410899668932, + 0.006864086724817753, + 0.013376145623624325, + 0.006094736512750387, + 0.004987477324903011, + 0.007788588758558035, + 0.0072972094640135765, + 0.0224024448543787, + 0.006437340285629034, + 0.004627423360943794, + 0.018546467646956444, + 0.021480178460478783, + 0.007553953677415848, + 0.004847903270274401, + 0.005926697514951229, + 0.04433765262365341, + 0.011080448515713215, + 0.010525332763791084, + 0.03577686846256256, + 0.0037592933513224125, + 0.003080046735703945, + 0.003979379311203957, + 0.0023299690801650286, + 0.0022021622862666845, + 0.0026145305018872023, + 0.0036169441882520914, + 0.0037985527887940407, + 0.047491610050201416, + 0.01031738892197609, + 0.010729881934821606, + 0.007827882654964924, + 0.007049389183521271, + 0.014477359130978584, + 0.05557173490524292, + 0.007977236062288284, + 0.03819815814495087, + 0.012610998935997486, + 0.010590135119855404, + 0.011041170917451382, + 0.0063707055523991585, + 0.009336162358522415, + 0.029843436554074287, + 0.026155581697821617, + 0.01682760939002037, + 0.0142960911616683, + 0.013015038333833218, + 0.011333425529301167, + 0.0185229554772377, + 0.011712322942912579, + 0.016461268067359924, + 0.016355928033590317, + 0.014488167129456997, + 0.014867277815937996, + 0.014318468980491161, + 0.010895443148911, + 0.008770238608121872, + 0.008316904306411743, + 0.012999593280255795, + 0.008399632759392262, + 0.005321003962308168, + 0.006965920329093933, + 0.004045348614454269, + 0.00339933461509645, + 0.0022065280936658382, + 0.0017777800094336271, + 0.004578772466629744, + 0.004566431976854801 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 3 + ], + "selected_reward": -0.02978570945560932 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, upward-tilting long shot pierces the pitch-black, nearly lightless depths of the ocean, following the graceful ascent of a school of luminous jellyfish. This silent procession, comprised of dozens of jellyfish of varying shapes, emits a soft, rhythmic bioluminescence from within their translucent, gelatinous umbrellas, shifting between ethereal cyan, dreamy pink-purple, and soft emerald green. As they propel upward through graceful contractions and expansions, countless long, slender, silken ribbon-like tentacles drift lazily behind them, carving trails of shimmering light through the darkness. Their own light, the sole source of illumination in the inky black waters, faintly illuminates the tiny particles of marine snow suspended around them, creating an effect reminiscent of stardust. The entire scene is shrouded in a profound sense of tranquility and mystery. The jellyfish's effortless ascent contrasts sharply with the endless darkness of the deep, creating a visual spectacle both otherworldly and vibrant.", + "seed": 5532, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_5/index_1182_seed_5532_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1338852945715189, + "vlm_logprob": -0.1338852945715189, + "vlm_yes_no_margin": 3.199999988079071, + "vlm_token_credit": -0.1338852945715189, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment pitch-black with no external light sources?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dozens of jellyfish with varying shapes visible ascending in a school?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jellyfish emit bioluminescence in cyan, pink-purple, and emerald green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jellyfish possess long, slender, silken ribbon-like tentacles drifting behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.43520456552505493 + ], + "opposite_logprob_tokens": [ + -1.1852045059204102 + ], + "logprob_reward": -0.43520456552505493, + "yes_no_margin_reward": 0.7499999403953552, + "token_credit_reward": -0.43520456552505493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11245287954807281, + 0.19201117753982544, + 0.08751145750284195, + 0.09700263291597366, + 0.07606637477874756, + 0.05028822645545006, + 0.07289060205221176, + 0.09327781945466995, + 0.09737080335617065, + 0.12112800776958466 + ], + "credit_map": [ + 0.008321966975927353, + 0.008218297734856606, + 0.024713335558772087, + 0.0166610199958086, + 0.014930288307368755, + 0.012817624025046825, + 0.012578201480209827, + 0.014212135225534439, + 0.015971235930919647, + 0.025056224316358566, + 0.014474897645413876, + 0.01998606137931347, + 0.04086790606379509, + 0.02866608090698719, + 0.024478672072291374, + 0.022510109469294548, + 0.007514395285397768, + 0.011225031688809395, + 0.00765989487990737, + 0.011834094300866127, + 0.017834188416600227, + 0.011969443410634995, + 0.012929676100611687, + 0.006544735748320818, + 0.009150560013949871, + 0.012287609279155731, + 0.009594161994755268, + 0.010582983493804932, + 0.016577547416090965, + 0.014175409451127052, + 0.009742199443280697, + 0.014892159961163998, + 0.007098976988345385, + 0.006825939752161503, + 0.008417196571826935, + 0.013276222161948681, + 0.004982305690646172, + 0.007243210915476084, + 0.015860797837376595, + 0.01236172579228878, + 0.005305133294314146, + 0.007820147089660168, + 0.007388402707874775, + 0.004456712864339352, + 0.007945094257593155, + 0.007244117092341185, + 0.0047198012471199036, + 0.0054088169708848, + 0.01899006776511669, + 0.005844421219080687, + 0.013402199372649193, + 0.004629849456250668, + 0.006005096714943647, + 0.005680713336914778, + 0.007905798964202404, + 0.010432451963424683, + 0.010288766585290432, + 0.0073533253744244576, + 0.012028342112898827, + 0.012047389522194862, + 0.012729206122457981, + 0.013325858861207962, + 0.016552919521927834, + 0.008952022530138493, + 0.009721556678414345, + 0.010873825289309025, + 0.017754901200532913, + 0.0134290661662817, + 0.007913844659924507, + 0.014847604557871819, + 0.010777106508612633, + 0.012052909471094608, + 0.012096593156456947, + 0.008058386854827404, + 0.023648742586374283, + 0.028522314503788948, + 0.009276789613068104, + 0.01651742123067379, + 0.01038382574915886, + 0.012623928487300873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.43520456552505493 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment pitch-black with no external light sources?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dozens of jellyfish with varying shapes visible ascending in a school?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jellyfish emit bioluminescence in cyan, pink-purple, and emerald green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jellyfish possess long, slender, silken ribbon-like tentacles drifting behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2016233503818512 + ], + "opposite_logprob_tokens": [ + -1.7016233205795288 + ], + "logprob_reward": -0.2016233503818512, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2016233503818512, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08806483447551727, + 0.26765012741088867, + 0.12375032901763916, + 0.03696843981742859, + 0.08452748507261276, + 0.03613041713833809, + 0.1326708048582077, + 0.03135303407907486, + 0.10003948211669922, + 0.09884500503540039 + ], + "credit_map": [ + 0.007989773526787758, + 0.009548063389956951, + 0.011953270062804222, + 0.010432670824229717, + 0.012342270463705063, + 0.011970045045018196, + 0.01429717242717743, + 0.009531572461128235, + 0.018531745299696922, + 0.043466318398714066, + 0.01907329633831978, + 0.032587725669145584, + 0.0561063326895237, + 0.03968024626374245, + 0.036137815564870834, + 0.022066647186875343, + 0.00886340718716383, + 0.018135415390133858, + 0.012629442848265171, + 0.017218945547938347, + 0.02055056393146515, + 0.017659345641732216, + 0.01674972102046013, + 0.011943494901061058, + 0.003385805757716298, + 0.0040367599576711655, + 0.004424370359629393, + 0.003218855708837509, + 0.004059663042426109, + 0.007922268472611904, + 0.005174433812499046, + 0.004746283870190382, + 0.007576609496027231, + 0.005572889931499958, + 0.009920603595674038, + 0.015198641456663609, + 0.005543030332773924, + 0.010740477591753006, + 0.01685110107064247, + 0.013124143704771996, + 0.0037358065601438284, + 0.005982242524623871, + 0.0034207685384899378, + 0.0031519278418272734, + 0.0064735556952655315, + 0.005096284206956625, + 0.0035279076546430588, + 0.004741929937154055, + 0.032688453793525696, + 0.008689617738127708, + 0.025363730266690254, + 0.008057949133217335, + 0.011299999430775642, + 0.012328061275184155, + 0.015772994607686996, + 0.01847000978887081, + 0.0028852138202637434, + 0.0019597685895860195, + 0.0034614363685250282, + 0.0038544824346899986, + 0.0029687611386179924, + 0.006696286611258984, + 0.006513217464089394, + 0.0030138669535517693, + 0.00993959978222847, + 0.008552717976272106, + 0.014160881750285625, + 0.017994390800595284, + 0.00881696306169033, + 0.01529751718044281, + 0.013577280566096306, + 0.011700132861733437, + 0.009280177764594555, + 0.007288786116987467, + 0.01800561510026455, + 0.021854320541024208, + 0.005990447010844946, + 0.014125837944447994, + 0.008774411864578724, + 0.013525409623980522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 2 + ], + "selected_reward": -0.2016233503818512 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment pitch-black with no external light sources?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dozens of jellyfish with varying shapes visible ascending in a school?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jellyfish emit bioluminescence in cyan, pink-purple, and emerald green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jellyfish possess long, slender, silken ribbon-like tentacles drifting behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006748148240149021 + ], + "opposite_logprob_tokens": [ + -5.006748199462891 + ], + "logprob_reward": -0.006748148240149021, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006748148240149021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1330084204673767, + 0.17007669806480408, + 0.1594836413860321, + 0.06558223068714142, + 0.1810428351163864, + 0.04673362150788307, + 0.04874847084283829, + 0.06492064893245697, + 0.08951298147439957, + 0.04089047759771347 + ], + "credit_map": [ + 0.011654487811028957, + 0.021329615265130997, + 0.016052335500717163, + 0.015466895885765553, + 0.01594267413020134, + 0.013556917198002338, + 0.018031127750873566, + 0.020974351093173027, + 0.011695764027535915, + 0.02365998737514019, + 0.01122428011149168, + 0.01746230572462082, + 0.036614082753658295, + 0.026004834100604057, + 0.024832965806126595, + 0.01858246512711048, + 0.010466116480529308, + 0.0235750712454319, + 0.01394739467650652, + 0.024014798924326897, + 0.029615281149744987, + 0.021987926214933395, + 0.022828130051493645, + 0.013048914261162281, + 0.007083437871187925, + 0.0072555094957351685, + 0.007404831238090992, + 0.007629518397152424, + 0.008573324419558048, + 0.008815387263894081, + 0.006663378793746233, + 0.012156832963228226, + 0.017315257340669632, + 0.020284779369831085, + 0.018353525549173355, + 0.016132785007357597, + 0.021307216957211494, + 0.03121519647538662, + 0.027999557554721832, + 0.028434498235583305, + 0.004776638466864824, + 0.0043380241841077805, + 0.005134889855980873, + 0.004047174472361803, + 0.008077370002865791, + 0.00754951499402523, + 0.0058912308886647224, + 0.006918775383383036, + 0.009999800473451614, + 0.00446318881586194, + 0.008202832192182541, + 0.004164518788456917, + 0.003953730687499046, + 0.003544582985341549, + 0.008868606761097908, + 0.005551203154027462, + 0.007236603647470474, + 0.005585853476077318, + 0.006578107364475727, + 0.013262836262583733, + 0.006219646893441677, + 0.01180946547538042, + 0.00878905039280653, + 0.0054390826262533665, + 0.009178867563605309, + 0.007549881003797054, + 0.016720205545425415, + 0.010135858319699764, + 0.010655412450432777, + 0.018784113228321075, + 0.00820956565439701, + 0.008279070258140564, + 0.004047454800456762, + 0.005349235143512487, + 0.006515082903206348, + 0.005625464953482151, + 0.005347951781004667, + 0.004411009605973959, + 0.00490476144477725, + 0.004689514171332121 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.006748148240149021 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment pitch-black with no external light sources?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dozens of jellyfish with varying shapes visible ascending in a school?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jellyfish emit bioluminescence in cyan, pink-purple, and emerald green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jellyfish possess long, slender, silken ribbon-like tentacles drifting behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978908035904169 + ], + "opposite_logprob_tokens": [ + -4.634788990020752 + ], + "logprob_reward": -0.00978908035904169, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978908035904169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09901434183120728, + 0.10761323571205139, + 0.4030986428260803, + 0.04729876667261124, + 0.05897374078631401, + 0.04827181622385979, + 0.09748020023107529, + 0.020135710015892982, + 0.06679178029298782, + 0.05132177472114563 + ], + "credit_map": [ + 0.008022012189030647, + 0.01737913116812706, + 0.012087252922356129, + 0.011026634834706783, + 0.010339757427573204, + 0.009067008271813393, + 0.01853647269308567, + 0.012556065805256367, + 0.007683676667511463, + 0.012205797247588634, + 0.011024069972336292, + 0.013250663876533508, + 0.012648090720176697, + 0.015164724551141262, + 0.01750115491449833, + 0.018135059624910355, + 0.02877434715628624, + 0.055441856384277344, + 0.03644854575395584, + 0.054541219025850296, + 0.08033245801925659, + 0.053526561707258224, + 0.06337614357471466, + 0.030657535418868065, + 0.004764859098941088, + 0.0039018213283270597, + 0.006054960656911135, + 0.004528854973614216, + 0.004098628181964159, + 0.008439617231488228, + 0.004730216693133116, + 0.010779807344079018, + 0.00732883857563138, + 0.004750299733132124, + 0.007474773097783327, + 0.009790043346583843, + 0.0049119312316179276, + 0.006466539576649666, + 0.008842641487717628, + 0.009408671408891678, + 0.005217633675783873, + 0.0048791635781526566, + 0.005368324927985668, + 0.004504448268562555, + 0.008450417779386044, + 0.006784504279494286, + 0.006722223944962025, + 0.006345100700855255, + 0.023307377472519875, + 0.004923129919916391, + 0.021171119064092636, + 0.005987863522022963, + 0.0066958884708583355, + 0.008501406759023666, + 0.009096511639654636, + 0.017796901986002922, + 0.002426099730655551, + 0.002051129238680005, + 0.0025539605412632227, + 0.0030325590632855892, + 0.0027856039814651012, + 0.002348765032365918, + 0.002924056723713875, + 0.002013534540310502, + 0.007731742225587368, + 0.007323472294956446, + 0.00847291573882103, + 0.007649833336472511, + 0.010881193913519382, + 0.009262965992093086, + 0.007054726127535105, + 0.008414928801357746, + 0.004418938420712948, + 0.006659995298832655, + 0.008263681083917618, + 0.0069234478287398815, + 0.005813817493617535, + 0.004528204910457134, + 0.0075827487744390965, + 0.0071309409104287624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.00978908035904169 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment pitch-black with no external light sources?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dozens of jellyfish with varying shapes visible ascending in a school?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jellyfish emit bioluminescence in cyan, pink-purple, and emerald green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jellyfish possess long, slender, silken ribbon-like tentacles drifting behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606132835149765 + ], + "opposite_logprob_tokens": [ + -4.141061305999756 + ], + "logprob_reward": -0.01606132835149765, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606132835149765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10556228458881378, + 0.12566153705120087, + 0.09270777553319931, + 0.08748932927846909, + 0.31501540541648865, + 0.02482661046087742, + 0.029701806604862213, + 0.08682381361722946, + 0.09521904587745667, + 0.03699247166514397 + ], + "credit_map": [ + 0.009410370141267776, + 0.012733079493045807, + 0.014148133806884289, + 0.01326065044850111, + 0.014640230685472488, + 0.01052482333034277, + 0.01588444970548153, + 0.01496055256575346, + 0.010052121244370937, + 0.020390894263982773, + 0.009471616707742214, + 0.01724405214190483, + 0.01685250550508499, + 0.017464376986026764, + 0.021656952798366547, + 0.012529025785624981, + 0.009330865927040577, + 0.011624062433838844, + 0.013611877337098122, + 0.015036400407552719, + 0.008344009518623352, + 0.012861991301178932, + 0.011585641652345657, + 0.010312925092875957, + 0.007903550751507282, + 0.007446207571774721, + 0.010482139885425568, + 0.011093458160758018, + 0.00900280475616455, + 0.015340982005000114, + 0.009242990054190159, + 0.01697719655930996, + 0.035475991666316986, + 0.030547868460416794, + 0.041171930730342865, + 0.030290568247437477, + 0.036772750318050385, + 0.053340520709753036, + 0.0391857773065567, + 0.04822998493909836, + 0.002821261528879404, + 0.0026478106155991554, + 0.003646389814093709, + 0.0017930581234395504, + 0.0026987376622855663, + 0.004438256379216909, + 0.0030588386580348015, + 0.0037222583778202534, + 0.006174990441650152, + 0.002537709893658757, + 0.005848240107297897, + 0.0028480899054557085, + 0.002558814361691475, + 0.0026983486022800207, + 0.0028556485194712877, + 0.004179963376373053, + 0.009629936888813972, + 0.006532672792673111, + 0.009852498769760132, + 0.012228776700794697, + 0.009702339768409729, + 0.016755785793066025, + 0.01186578068882227, + 0.010256024077534676, + 0.012955787591636181, + 0.009234429337084293, + 0.010077029466629028, + 0.009985045529901981, + 0.016298554837703705, + 0.01271132379770279, + 0.01102183572947979, + 0.012935040518641472, + 0.0038946380373090506, + 0.0036384949926286936, + 0.006300076376646757, + 0.0043814158998429775, + 0.004710146691650152, + 0.003465544432401657, + 0.005783745087683201, + 0.004818410146981478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.01606132835149765 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, upward-tilting long shot pierces the pitch-black, nearly lightless depths of the ocean, following the graceful ascent of a school of luminous jellyfish. This silent procession, comprised of dozens of jellyfish of varying shapes, emits a soft, rhythmic bioluminescence from within their translucent, gelatinous umbrellas, shifting between ethereal cyan, dreamy pink-purple, and soft emerald green. As they propel upward through graceful contractions and expansions, countless long, slender, silken ribbon-like tentacles drift lazily behind them, carving trails of shimmering light through the darkness. Their own light, the sole source of illumination in the inky black waters, faintly illuminates the tiny particles of marine snow suspended around them, creating an effect reminiscent of stardust. The entire scene is shrouded in a profound sense of tranquility and mystery. The jellyfish's effortless ascent contrasts sharply with the endless darkness of the deep, creating a visual spectacle both otherworldly and vibrant.", + "seed": 5533, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_5/index_1182_seed_5533_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.15396537445485592, + "vlm_logprob": -0.15396537445485592, + "vlm_yes_no_margin": 3.025000011920929, + "vlm_token_credit": -0.15396537445485592, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment pitch-black with no external light sources?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dozens of jellyfish with varying shapes visible ascending in a school?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jellyfish emit bioluminescence in cyan, pink-purple, and emerald green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jellyfish possess long, slender, silken ribbon-like tentacles drifting behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3832782506942749 + ], + "opposite_logprob_tokens": [ + -1.258278250694275 + ], + "logprob_reward": -0.3832782506942749, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3832782506942749, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19868367910385132, + 0.06820843368768692, + 0.06114611402153969, + 0.19302430748939514, + 0.029868245124816895, + 0.2931128442287445, + 0.07452239096164703, + 0.02154732309281826, + 0.03255871310830116, + 0.027327949181199074 + ], + "credit_map": [ + 0.01433632429689169, + 0.019954968243837357, + 0.03171681985259056, + 0.027687426656484604, + 0.024442128837108612, + 0.03537488356232643, + 0.02075604535639286, + 0.024415070191025734, + 0.006175164598971605, + 0.005978610832244158, + 0.006344318389892578, + 0.009619377553462982, + 0.008901450783014297, + 0.011757919564843178, + 0.009518750943243504, + 0.00991283729672432, + 0.009549508802592754, + 0.005025122780352831, + 0.010272582992911339, + 0.00684668542817235, + 0.011213939636945724, + 0.005529890302568674, + 0.005132946185767651, + 0.007575438357889652, + 0.0220057163387537, + 0.008345169946551323, + 0.014818966388702393, + 0.011352969333529472, + 0.04613649845123291, + 0.00800870731472969, + 0.05890066921710968, + 0.023455621674656868, + 0.003347872057929635, + 0.0014493638882413507, + 0.005395581480115652, + 0.0020004985854029655, + 0.0026809254195541143, + 0.0046363105066120625, + 0.004692337475717068, + 0.005665355361998081, + 0.02764151617884636, + 0.037816498428583145, + 0.04811111465096474, + 0.03006008453667164, + 0.02834206447005272, + 0.06379229575395584, + 0.03599380701780319, + 0.02135544829070568, + 0.005346482619643211, + 0.008250780403614044, + 0.005332115106284618, + 0.007251705974340439, + 0.023715626448392868, + 0.0092374412342906, + 0.01034913957118988, + 0.005039098672568798, + 0.0021320770028978586, + 0.002536425832659006, + 0.0034583061933517456, + 0.0018088867655023932, + 0.0022962132934480906, + 0.004474091809242964, + 0.0026978831738233566, + 0.0021434398368000984, + 0.0027908915653824806, + 0.005435301456600428, + 0.002779482863843441, + 0.003709352109581232, + 0.004472763743251562, + 0.004931515082716942, + 0.004286908078938723, + 0.004152501467615366, + 0.0020546589512377977, + 0.002788390265777707, + 0.002751303371042013, + 0.0030835617799311876, + 0.005601570941507816, + 0.004701509606093168, + 0.0038353921845555305, + 0.002511561382561922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.3832782506942749 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment pitch-black with no external light sources?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dozens of jellyfish with varying shapes visible ascending in a school?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jellyfish emit bioluminescence in cyan, pink-purple, and emerald green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jellyfish possess long, slender, silken ribbon-like tentacles drifting behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34867972135543823 + ], + "opposite_logprob_tokens": [ + -1.223679780960083 + ], + "logprob_reward": -0.34867972135543823, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.34867972135543823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3377983272075653, + 0.064668208360672, + 0.02628621645271778, + 0.2349526286125183, + 0.025050636380910873, + 0.06643858551979065, + 0.04864474758505821, + 0.13232232630252838, + 0.03274821117520332, + 0.03109012171626091 + ], + "credit_map": [ + 0.023061752319335938, + 0.019198326393961906, + 0.039216745644807816, + 0.03626057878136635, + 0.046935390681028366, + 0.08670501410961151, + 0.02857075445353985, + 0.057849761098623276, + 0.005867519415915012, + 0.0047499616630375385, + 0.0065462831407785416, + 0.00750533863902092, + 0.007977447472512722, + 0.013827431946992874, + 0.00835524220019579, + 0.009838985279202461, + 0.004036808852106333, + 0.0019718336407095194, + 0.0029192559886723757, + 0.002546275733038783, + 0.005765181966125965, + 0.0027662820648401976, + 0.0027635504957288504, + 0.003517027013003826, + 0.029369017109274864, + 0.008744942955672741, + 0.014133809134364128, + 0.01175261102616787, + 0.03483342379331589, + 0.010830041021108627, + 0.08854276686906815, + 0.03674599528312683, + 0.002563873305916786, + 0.002354529919102788, + 0.00346408411860466, + 0.0020569446496665478, + 0.003206337569281459, + 0.005430193617939949, + 0.0038190586492419243, + 0.0021556138526648283, + 0.0050785536877810955, + 0.004913515876978636, + 0.015219275839626789, + 0.004334626719355583, + 0.005594966467469931, + 0.0168614499270916, + 0.007809311617165804, + 0.006626882590353489, + 0.003156619379296899, + 0.005840691737830639, + 0.0034826728515326977, + 0.005253454204648733, + 0.012287990190088749, + 0.0070749856531620026, + 0.0076905423775315285, + 0.0038577860686928034, + 0.009506979025900364, + 0.013911142013967037, + 0.023398464545607567, + 0.016169752925634384, + 0.013721899129450321, + 0.020677082240581512, + 0.022381577640771866, + 0.012555410154163837, + 0.0021011352073401213, + 0.00587028544396162, + 0.002307634335011244, + 0.0032239279244095087, + 0.004430100321769714, + 0.006265092641115189, + 0.004785975441336632, + 0.0037640570662915707, + 0.002356243785470724, + 0.0051381816156208515, + 0.0031104793306440115, + 0.0044542038813233376, + 0.0033526006154716015, + 0.005194404628127813, + 0.004287505056709051, + 0.0031965014059096575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.34867972135543823 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment pitch-black with no external light sources?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dozens of jellyfish with varying shapes visible ascending in a school?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jellyfish emit bioluminescence in cyan, pink-purple, and emerald green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jellyfish possess long, slender, silken ribbon-like tentacles drifting behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864796806126833 + ], + "opposite_logprob_tokens": [ + -4.758647918701172 + ], + "logprob_reward": -0.00864796806126833, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864796806126833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23051717877388, + 0.14284534752368927, + 0.09064936637878418, + 0.16421489417552948, + 0.04141280800104141, + 0.03154429420828819, + 0.03911932185292244, + 0.04647985100746155, + 0.07603683322668076, + 0.1371801644563675 + ], + "credit_map": [ + 0.015918783843517303, + 0.01838206872344017, + 0.04647105187177658, + 0.02561182901263237, + 0.027118248865008354, + 0.04094062000513077, + 0.02399362623691559, + 0.03208094462752342, + 0.015818748623132706, + 0.010792477056384087, + 0.014017775654792786, + 0.01683039218187332, + 0.01717221550643444, + 0.028527850285172462, + 0.01745360717177391, + 0.02223227173089981, + 0.011389127932488918, + 0.009845918975770473, + 0.014472304843366146, + 0.010629829950630665, + 0.010253233835101128, + 0.008641156367957592, + 0.007912873290479183, + 0.017504923045635223, + 0.018519075587391853, + 0.00925079919397831, + 0.010283507406711578, + 0.012023012153804302, + 0.014615979045629501, + 0.012640154920518398, + 0.04998430237174034, + 0.03689805418252945, + 0.004097968339920044, + 0.004971015267074108, + 0.005724071059376001, + 0.0044165984727442265, + 0.007350434083491564, + 0.006211875472217798, + 0.005274040624499321, + 0.003366803750395775, + 0.0029365825466811657, + 0.004366949666291475, + 0.0063918340019881725, + 0.004338715225458145, + 0.0030768164433538914, + 0.003486479399725795, + 0.0049238260835409164, + 0.0020230922382324934, + 0.0025680926628410816, + 0.0039391289465129375, + 0.005052097607403994, + 0.005381776485592127, + 0.006748352665454149, + 0.005585597828030586, + 0.005557594820857048, + 0.004286684095859528, + 0.0052900477312505245, + 0.00444426666945219, + 0.004447817336767912, + 0.006035114172846079, + 0.0064835199154913425, + 0.006691238842904568, + 0.006801331881433725, + 0.00628651212900877, + 0.004515020176768303, + 0.012991134077310562, + 0.008048691786825657, + 0.009206805378198624, + 0.009836430661380291, + 0.01339783612638712, + 0.010823952034115791, + 0.007216963451355696, + 0.011002990417182446, + 0.023185689002275467, + 0.015772465616464615, + 0.018346691504120827, + 0.019034715369343758, + 0.023058362305164337, + 0.01266503892838955, + 0.014114213176071644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.00864796806126833 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment pitch-black with no external light sources?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dozens of jellyfish with varying shapes visible ascending in a school?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jellyfish emit bioluminescence in cyan, pink-purple, and emerald green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jellyfish possess long, slender, silken ribbon-like tentacles drifting behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864796806126833 + ], + "opposite_logprob_tokens": [ + -4.758647918701172 + ], + "logprob_reward": -0.00864796806126833, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864796806126833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22308747470378876, + 0.23351316154003143, + 0.10192835330963135, + 0.1196543350815773, + 0.04621381685137749, + 0.04342781752347946, + 0.08552064746618271, + 0.027008939534425735, + 0.06289598345756531, + 0.05674941465258598 + ], + "credit_map": [ + 0.019407715648412704, + 0.018931420519948006, + 0.04206246882677078, + 0.03256005048751831, + 0.02536911331117153, + 0.029366929084062576, + 0.02292504720389843, + 0.03246473893523216, + 0.019838890060782433, + 0.037032730877399445, + 0.022257819771766663, + 0.03054264932870865, + 0.048575837165117264, + 0.02589419297873974, + 0.021140985190868378, + 0.02823006920516491, + 0.018064608797430992, + 0.006862196605652571, + 0.014668927527964115, + 0.008057265542447567, + 0.02715662121772766, + 0.005877459887415171, + 0.007615457754582167, + 0.013625824823975563, + 0.02001652866601944, + 0.006765701808035374, + 0.008620952256023884, + 0.008083966560661793, + 0.01681589148938656, + 0.007100176066160202, + 0.0412711463868618, + 0.010979975573718548, + 0.003983668517321348, + 0.003355356166139245, + 0.005416227038949728, + 0.0034645164851099253, + 0.006563483737409115, + 0.01269577257335186, + 0.006687437649816275, + 0.004047357477247715, + 0.0032317808363586664, + 0.0036268960684537888, + 0.008431283757090569, + 0.0031954057049006224, + 0.0035127957817167044, + 0.010561875067651272, + 0.006053171586245298, + 0.004814613610506058, + 0.005155366845428944, + 0.006653555668890476, + 0.005491516087204218, + 0.007843365892767906, + 0.030337341129779816, + 0.010565696284174919, + 0.013607575558125973, + 0.005866239313036203, + 0.002710232976824045, + 0.00331841129809618, + 0.004055925644934177, + 0.0030527289491146803, + 0.003226825036108494, + 0.004024522844702005, + 0.003817787393927574, + 0.0028025072533637285, + 0.003915900364518166, + 0.0073326057754457, + 0.005494843237102032, + 0.008031860925257206, + 0.012785383500158787, + 0.009710507467389107, + 0.009780580177903175, + 0.005844302009791136, + 0.004443768877536058, + 0.0067241862416267395, + 0.006673644762486219, + 0.006342858541756868, + 0.006811489816755056, + 0.00947850476950407, + 0.010994271375238895, + 0.00528069119900465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.00864796806126833 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment pitch-black with no external light sources?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dozens of jellyfish with varying shapes visible ascending in a school?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the jellyfish emit bioluminescence in cyan, pink-purple, and emerald green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the jellyfish possess long, slender, silken ribbon-like tentacles drifting behind them?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0205729641020298 + ], + "opposite_logprob_tokens": [ + -3.8955729007720947 + ], + "logprob_reward": -0.0205729641020298, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.0205729641020298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2608513832092285, + 0.10162241756916046, + 0.11716599017381668, + 0.12155161798000336, + 0.046506594866514206, + 0.06741341203451157, + 0.052176691591739655, + 0.03950973600149155, + 0.04594138637185097, + 0.14726075530052185 + ], + "credit_map": [ + 0.02203601971268654, + 0.0224146731197834, + 0.05246628075838089, + 0.036961622536182404, + 0.031226376071572304, + 0.03670728579163551, + 0.0283596720546484, + 0.03067946806550026, + 0.009740025736391544, + 0.013113587163388729, + 0.010128469206392765, + 0.022186947986483574, + 0.012952694669365883, + 0.011813194490969181, + 0.00897128600627184, + 0.012716214172542095, + 0.0188977662473917, + 0.009177906438708305, + 0.02031378634274006, + 0.010331710800528526, + 0.021504521369934082, + 0.009263523854315281, + 0.008551014587283134, + 0.019125768914818764, + 0.027502788230776787, + 0.00573414983227849, + 0.007525595370680094, + 0.007362530566751957, + 0.016820937395095825, + 0.005368763580918312, + 0.03313726186752319, + 0.018099607899785042, + 0.0038535406347364187, + 0.004986131098121405, + 0.00566842732951045, + 0.006449532229453325, + 0.00623294385150075, + 0.0074324337765574455, + 0.006894615478813648, + 0.004988973028957844, + 0.0060789273120462894, + 0.004511467646807432, + 0.02016971819102764, + 0.004634002689272165, + 0.004714377224445343, + 0.012077448889613152, + 0.007497861050069332, + 0.0077296108938753605, + 0.004658625926822424, + 0.007010121829807758, + 0.005839788354933262, + 0.0060083926655352116, + 0.009871023707091808, + 0.006127407308667898, + 0.006401050370186567, + 0.006260287947952747, + 0.0034860088489949703, + 0.003748381743207574, + 0.006064190529286861, + 0.00472622225061059, + 0.004950443748384714, + 0.006221826188266277, + 0.005669913720339537, + 0.004642750136554241, + 0.0034434369299560785, + 0.006682536564767361, + 0.003429804462939501, + 0.004935426637530327, + 0.006925581023097038, + 0.00850970670580864, + 0.007002600468695164, + 0.0050122980028390884, + 0.011277314275503159, + 0.028224200010299683, + 0.01697595976293087, + 0.018564943224191666, + 0.025779135525226593, + 0.023667341098189354, + 0.014130450785160065, + 0.008641424588859081 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.0205729641020298 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_6/index_tensor([3007, 3007])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_6/index_tensor([3007, 3007])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1d2dc257ce2b488393ae7ddf5ad41950cb15bf65 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_6/index_tensor([3007, 3007])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic medium shot, shot from a slightly upward angle, smoothly pans sideways, closely following a man and woman engaged in a fierce confrontation within a dense, deep, emerald green bamboo forest. The man, with his long, dark hair tied high with a headband, is handsome, his eyes piercing like an eagle. He wears a flowing white silk robe, its wide sleeves fluttering in the air with his swift movements. The woman, also wearing a high ponytail, has a beautiful face and a focused, stern expression. She wears a sharply tailored red suit, creating a stark and striking visual contrast with the man's white robe. The afternoon sun struggles to penetrate the gaps between the layers of bamboo leaves, casting a dappled, flickering light on the ground, shrouded in a light mist. The two men's figures were as fast as lightning, deftly shuttling, moving and spinning between the straight and tall bamboo poles. The fine steel swords in their hands drew deadly and graceful silver arcs in the air. Every precise and fierce collision instantly burst out dazzling golden sparks, briefly illuminating their focused faces and the cold light flowing on the sword blades. The whole fighting process was full of power and like a carefully choreographed dance, swift and graceful, creating a martial arts atmosphere that was both intense and poetic against the backdrop of the quiet bamboo forest.", + "seed": 23681, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_6/index_3007_seed_23681_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.024424251960590483, + "vlm_logprob": -0.024424251960590483, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.024424251960590483, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense emerald green bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures, a man in a white silk robe and a woman in a red suit, engaged in a sword fight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the characters perform swift movements including spinning and dodging between bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sword collisions produce dazzling golden sparks that briefly illuminate the characters' faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and smoothly panning sideways to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02012021653354168 + ], + "opposite_logprob_tokens": [ + -4.270120143890381 + ], + "logprob_reward": -0.02012021653354168, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.02012021653354168, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14086641371250153, + 0.04426158592104912, + 0.1757919192314148, + 0.12137466669082642, + 0.05328122526407242, + 0.035043865442276, + 0.13765768706798553, + 0.14814895391464233, + 0.05262669175863266, + 0.09094694256782532 + ], + "credit_map": [ + 0.02293325774371624, + 0.03051590546965599, + 0.015275194309651852, + 0.009470767341554165, + 0.01084041316062212, + 0.02016099914908409, + 0.012056976556777954, + 0.019612908363342285, + 0.009797769598662853, + 0.0043723308481276035, + 0.00931694358587265, + 0.004286834504455328, + 0.00470832921564579, + 0.004668045789003372, + 0.0030947597697377205, + 0.004016574937850237, + 0.014751633629202843, + 0.007095863576978445, + 0.05259214714169502, + 0.007982490584254265, + 0.02062712237238884, + 0.06040360406041145, + 0.005799870938062668, + 0.006539203692227602, + 0.015470319427549839, + 0.012338614091277122, + 0.008719160221517086, + 0.0065655894577503204, + 0.023751135915517807, + 0.0414545014500618, + 0.0038216104730963707, + 0.009253746829926968, + 0.0064122602343559265, + 0.008646789006888866, + 0.0072298189625144005, + 0.004751433152705431, + 0.007154053542762995, + 0.007523112930357456, + 0.005436618346720934, + 0.006127147004008293, + 0.00516852829605341, + 0.006025247275829315, + 0.0038453852757811546, + 0.004037331789731979, + 0.00402262294664979, + 0.005073170643299818, + 0.0033127525821328163, + 0.003558830823749304, + 0.0268795657902956, + 0.016224563121795654, + 0.02213546261191368, + 0.009292743168771267, + 0.013808305375277996, + 0.028326841071248055, + 0.009340917691588402, + 0.011649299412965775, + 0.028198007494211197, + 0.03521592915058136, + 0.02453538030385971, + 0.01676257513463497, + 0.012364163994789124, + 0.009872633963823318, + 0.010502979159355164, + 0.01069729682058096, + 0.006260599009692669, + 0.008631054311990738, + 0.00847164448350668, + 0.005594511050730944, + 0.006775666493922472, + 0.0076514920219779015, + 0.004987142980098724, + 0.004254583735018969, + 0.009711922146379948, + 0.010398917831480503, + 0.010458833537995815, + 0.005346381571143866, + 0.016064416617155075, + 0.014008465223014355, + 0.014477009885013103, + 0.010481001809239388 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.02012021653354168 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense emerald green bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures, a man in a white silk robe and a woman in a red suit, engaged in a sword fight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the characters perform swift movements including spinning and dodging between bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sword collisions produce dazzling golden sparks that briefly illuminate the characters' faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and smoothly panning sideways to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329878881573677 + ], + "opposite_logprob_tokens": [ + -3.773298740386963 + ], + "logprob_reward": -0.02329878881573677, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329878881573677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1062222570180893, + 0.06679962575435638, + 0.12300044298171997, + 0.1827312409877777, + 0.15872696042060852, + 0.052661437541246414, + 0.0897424966096878, + 0.06336383521556854, + 0.07062719762325287, + 0.08612450212240219 + ], + "credit_map": [ + 0.01727517694234848, + 0.016288651153445244, + 0.016466569155454636, + 0.007306086830794811, + 0.012335233390331268, + 0.009237891994416714, + 0.012477530166506767, + 0.014835115522146225, + 0.010401271283626556, + 0.012846406549215317, + 0.007547103799879551, + 0.0071493457071483135, + 0.010107330046594143, + 0.006428546272218227, + 0.008292843587696552, + 0.004026781301945448, + 0.01259747426956892, + 0.012049497105181217, + 0.021895287558436394, + 0.011640273965895176, + 0.01552154403179884, + 0.035732630640268326, + 0.006746381055563688, + 0.0068173473700881, + 0.02418115735054016, + 0.018975453451275826, + 0.010537045076489449, + 0.010950910858809948, + 0.03611200302839279, + 0.06416189670562744, + 0.004264874383807182, + 0.013547901064157486, + 0.024754652753472328, + 0.01908128894865513, + 0.022587841376662254, + 0.014926853589713573, + 0.01932763122022152, + 0.022243557497859, + 0.015388871543109417, + 0.0204162634909153, + 0.008825826458632946, + 0.009193490259349346, + 0.007333931978791952, + 0.005232295021414757, + 0.0057469867169857025, + 0.006106802262365818, + 0.004351321142166853, + 0.005870787426829338, + 0.01904015801846981, + 0.010382645763456821, + 0.014466186054050922, + 0.007374350912868977, + 0.009366785176098347, + 0.01575174555182457, + 0.0058928304351866245, + 0.007467790972441435, + 0.006992954295128584, + 0.009397224523127079, + 0.008797118440270424, + 0.009107849560678005, + 0.008979853242635727, + 0.00655206385999918, + 0.0068254852667450905, + 0.006711287889629602, + 0.007797290105372667, + 0.008814283646643162, + 0.012448621913790703, + 0.009524082764983177, + 0.010732192546129227, + 0.00846090167760849, + 0.005943359807133675, + 0.00690646143630147, + 0.010210306383669376, + 0.013356692157685757, + 0.010998792946338654, + 0.007414302788674831, + 0.012653465382754803, + 0.012203840538859367, + 0.011728254146873951, + 0.0075588528998196125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 2 + ], + "selected_reward": -0.02329878881573677 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense emerald green bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures, a man in a white silk robe and a woman in a red suit, engaged in a sword fight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the characters perform swift movements including spinning and dodging between bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sword collisions produce dazzling golden sparks that briefly illuminate the characters' faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and smoothly panning sideways to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024878999684005976 + ], + "opposite_logprob_tokens": [ + -6.002488136291504 + ], + "logprob_reward": -0.0024878999684005976, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024878999684005976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1390077769756317, + 0.06833375990390778, + 0.11462762951850891, + 0.1981797218322754, + 0.028287535533308983, + 0.03908460587263107, + 0.13257460296154022, + 0.13649702072143555, + 0.06198975071310997, + 0.08141756802797318 + ], + "credit_map": [ + 0.023921867832541466, + 0.0210476815700531, + 0.011026942171156406, + 0.008319420740008354, + 0.018376780673861504, + 0.023018766194581985, + 0.01620914600789547, + 0.0170871801674366, + 0.01768558658659458, + 0.0047684647142887115, + 0.01709613762795925, + 0.003905914956703782, + 0.008771882392466068, + 0.00722107756882906, + 0.005532817915081978, + 0.0033518855925649405, + 0.015129223465919495, + 0.007591882720589638, + 0.03499029949307442, + 0.006143747828900814, + 0.023804491385817528, + 0.014039563946425915, + 0.007438480854034424, + 0.0054899416863918304, + 0.027357907965779305, + 0.01906287483870983, + 0.012434766627848148, + 0.012771974317729473, + 0.039259955286979675, + 0.066593237221241, + 0.0054890429601073265, + 0.015209966339170933, + 0.0030879832338541746, + 0.005247447174042463, + 0.0039521474391222, + 0.00261096959002316, + 0.003711601486429572, + 0.0035715026315301657, + 0.0029213952366262674, + 0.003184487344697118, + 0.005845380946993828, + 0.006097650155425072, + 0.005082137417048216, + 0.003681396134197712, + 0.00521038519218564, + 0.005629041697829962, + 0.003112289123237133, + 0.004426324740052223, + 0.021072981879115105, + 0.013558922335505486, + 0.023205867037177086, + 0.010376126505434513, + 0.014740278013050556, + 0.027349507436156273, + 0.010168732143938541, + 0.012102187611162663, + 0.015078512020409107, + 0.019950175657868385, + 0.019247358664870262, + 0.02003774791955948, + 0.012387225404381752, + 0.02492079697549343, + 0.010099655017256737, + 0.014775550924241543, + 0.006756606977432966, + 0.00643479498103261, + 0.009650311432778835, + 0.00736596342176199, + 0.009513596072793007, + 0.007966289296746254, + 0.007115130312740803, + 0.007187057752162218, + 0.008736803196370602, + 0.01107820961624384, + 0.00760633172467351, + 0.0052055842243134975, + 0.011162747628986835, + 0.012737897224724293, + 0.015040040016174316, + 0.00984995998442173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.0024878999684005976 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense emerald green bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures, a man in a white silk robe and a woman in a red suit, engaged in a sword fight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the characters perform swift movements including spinning and dodging between bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sword collisions produce dazzling golden sparks that briefly illuminate the characters' faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and smoothly panning sideways to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203080713748932 + ], + "opposite_logprob_tokens": [ + -2.812030792236328 + ], + "logprob_reward": -0.06203080713748932, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203080713748932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0930027961730957, + 0.11809360980987549, + 0.05958978086709976, + 0.4047899842262268, + 0.04073168337345123, + 0.03200095146894455, + 0.05844533443450928, + 0.07370048016309738, + 0.028933092951774597, + 0.0907122939825058 + ], + "credit_map": [ + 0.015714390203356743, + 0.01687896065413952, + 0.0071945819072425365, + 0.005239641759544611, + 0.012515108101069927, + 0.010601854883134365, + 0.009434415958821774, + 0.015423841774463654, + 0.022944122552871704, + 0.018138647079467773, + 0.018933573737740517, + 0.011904653161764145, + 0.01765088364481926, + 0.011942841112613678, + 0.010638396255671978, + 0.005940490867942572, + 0.008058927021920681, + 0.006731098052114248, + 0.00963683146983385, + 0.004197980277240276, + 0.008542916737496853, + 0.013691543601453304, + 0.003809021320194006, + 0.00492146285250783, + 0.045219920575618744, + 0.03799581900238991, + 0.0326177217066288, + 0.026164591312408447, + 0.08163095265626907, + 0.14414985477924347, + 0.010075689293444157, + 0.026935430243611336, + 0.004514352884143591, + 0.005487001035362482, + 0.006141778081655502, + 0.003714528400450945, + 0.0057518370449543, + 0.005218180362135172, + 0.004795235116034746, + 0.005108773708343506, + 0.0055717104114592075, + 0.0038358611054718494, + 0.003010306740179658, + 0.0023214034736156464, + 0.0051343622617423534, + 0.006201417651027441, + 0.002869865857064724, + 0.003056024434044957, + 0.011232137680053711, + 0.0052496339194476604, + 0.009351539425551891, + 0.004888166207820177, + 0.006564431823790073, + 0.010009577497839928, + 0.0063175419345498085, + 0.004832302685827017, + 0.008112451992928982, + 0.011586139909923077, + 0.010204585269093513, + 0.010686827823519707, + 0.009519812650978565, + 0.010380366817116737, + 0.005545260850340128, + 0.007665041368454695, + 0.003153928555548191, + 0.004328445997089148, + 0.004775907378643751, + 0.002721890341490507, + 0.004306929185986519, + 0.0034549152478575706, + 0.002987821353599429, + 0.0032032548915594816, + 0.009328344836831093, + 0.010984533466398716, + 0.011243373155593872, + 0.006226551253348589, + 0.012408203445374966, + 0.013318579643964767, + 0.016126174479722977, + 0.011076540686190128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.06203080713748932 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense emerald green bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures, a man in a white silk robe and a woman in a red suit, engaged in a sword fight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the characters perform swift movements including spinning and dodging between bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sword collisions produce dazzling golden sparks that briefly illuminate the characters' faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and smoothly panning sideways to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014183547347784042 + ], + "opposite_logprob_tokens": [ + -4.264183521270752 + ], + "logprob_reward": -0.014183547347784042, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014183547347784042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15662120282649994, + 0.0913042277097702, + 0.09936602413654327, + 0.04203666001558304, + 0.0947740450501442, + 0.07770812511444092, + 0.1836223602294922, + 0.11245738714933395, + 0.07554011791944504, + 0.06656985729932785 + ], + "credit_map": [ + 0.029739605262875557, + 0.03466524928808212, + 0.015628661960363388, + 0.009476018138229847, + 0.017107341438531876, + 0.020479679107666016, + 0.013397445902228355, + 0.016127195209264755, + 0.01625194400548935, + 0.009814763441681862, + 0.014615275897085667, + 0.010065644979476929, + 0.014436905272305012, + 0.010474834591150284, + 0.009900417178869247, + 0.005744438152760267, + 0.011833622120320797, + 0.016703197732567787, + 0.010020832531154156, + 0.008552612736821175, + 0.014865792356431484, + 0.022257119417190552, + 0.008102009072899818, + 0.007030840497463942, + 0.007805946283042431, + 0.006256737280637026, + 0.006591293960809708, + 0.0032535558566451073, + 0.0046859304420650005, + 0.005487637594342232, + 0.002632486168295145, + 0.005323072895407677, + 0.014005620032548904, + 0.01767696440219879, + 0.009483513422310352, + 0.010631526820361614, + 0.010715585201978683, + 0.00965750589966774, + 0.010143859311938286, + 0.01245946902781725, + 0.01241297833621502, + 0.01327733788639307, + 0.012784171849489212, + 0.010356098413467407, + 0.009533359669148922, + 0.008507722988724709, + 0.006208931095898151, + 0.0046275220811367035, + 0.03390298783779144, + 0.016486424952745438, + 0.03133292868733406, + 0.014278337359428406, + 0.019822800531983376, + 0.0327957384288311, + 0.015107611194252968, + 0.019895533099770546, + 0.012602832168340683, + 0.020108753815293312, + 0.015282615087926388, + 0.016310734674334526, + 0.015432026237249374, + 0.011104651726782322, + 0.010406449437141418, + 0.011209321208298206, + 0.009278645738959312, + 0.015217334032058716, + 0.009507675655186176, + 0.007592575624585152, + 0.013059915974736214, + 0.011439519934356213, + 0.00440726475790143, + 0.005037181079387665, + 0.007310650777071714, + 0.007360118441283703, + 0.012704120948910713, + 0.0051153493113815784, + 0.011440076865255833, + 0.008829222992062569, + 0.00918054860085249, + 0.004629773553460836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.014183547347784042 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic medium shot, shot from a slightly upward angle, smoothly pans sideways, closely following a man and woman engaged in a fierce confrontation within a dense, deep, emerald green bamboo forest. The man, with his long, dark hair tied high with a headband, is handsome, his eyes piercing like an eagle. He wears a flowing white silk robe, its wide sleeves fluttering in the air with his swift movements. The woman, also wearing a high ponytail, has a beautiful face and a focused, stern expression. She wears a sharply tailored red suit, creating a stark and striking visual contrast with the man's white robe. The afternoon sun struggles to penetrate the gaps between the layers of bamboo leaves, casting a dappled, flickering light on the ground, shrouded in a light mist. The two men's figures were as fast as lightning, deftly shuttling, moving and spinning between the straight and tall bamboo poles. The fine steel swords in their hands drew deadly and graceful silver arcs in the air. Every precise and fierce collision instantly burst out dazzling golden sparks, briefly illuminating their focused faces and the cold light flowing on the sword blades. The whole fighting process was full of power and like a carefully choreographed dance, swift and graceful, creating a martial arts atmosphere that was both intense and poetic against the backdrop of the quiet bamboo forest.", + "seed": 23682, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_6/index_3007_seed_23682_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.022834821511059998, + "vlm_logprob": -0.022834821511059998, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.022834821511059998, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense emerald green bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures, a man in a white silk robe and a woman in a red suit, engaged in a sword fight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the characters perform swift movements including spinning and dodging between bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sword collisions produce dazzling golden sparks that briefly illuminate the characters' faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and smoothly panning sideways to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020916594192385674 + ], + "opposite_logprob_tokens": [ + -4.27091646194458 + ], + "logprob_reward": -0.020916594192385674, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.020916594192385674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2609925866127014, + 0.15391555428504944, + 0.04661073908209801, + 0.036799926310777664, + 0.030878938734531403, + 0.06855875998735428, + 0.13127626478672028, + 0.055708277970552444, + 0.07846337556838989, + 0.13679559528827667 + ], + "credit_map": [ + 0.06891351193189621, + 0.02406570501625538, + 0.0351322777569294, + 0.018491230905056, + 0.05196317285299301, + 0.025167346000671387, + 0.018911780789494514, + 0.018347561359405518, + 0.01752731390297413, + 0.022701427340507507, + 0.010570527985692024, + 0.010151800699532032, + 0.02624456398189068, + 0.04454255849123001, + 0.00982630718499422, + 0.012351053766906261, + 0.005339929834008217, + 0.006678040139377117, + 0.005252604838460684, + 0.005535479169338942, + 0.0072586433961987495, + 0.005522372201085091, + 0.006124841049313545, + 0.004898823332041502, + 0.0031144674867391586, + 0.0031012617982923985, + 0.004970306996256113, + 0.003215489210560918, + 0.0083681121468544, + 0.008326889015734196, + 0.0029907922726124525, + 0.0027126113418489695, + 0.004308325704187155, + 0.004343699663877487, + 0.003345950273796916, + 0.004823080264031887, + 0.0037640694063156843, + 0.004193077329546213, + 0.00338494754396379, + 0.0027157901786267757, + 0.014334161765873432, + 0.010153823532164097, + 0.004790017381310463, + 0.0034911981783807278, + 0.01349665317684412, + 0.008678821846842766, + 0.006822746712714434, + 0.006791339255869389, + 0.019578702747821808, + 0.023967266082763672, + 0.011844445019960403, + 0.01623213291168213, + 0.018861955031752586, + 0.019751250743865967, + 0.013543134555220604, + 0.007497368846088648, + 0.0043630655854940414, + 0.011786065064370632, + 0.002995571354404092, + 0.005779031664133072, + 0.009610123001039028, + 0.006693854462355375, + 0.0077270097099244595, + 0.006753560621291399, + 0.004887288436293602, + 0.02229819819331169, + 0.0044867354445159435, + 0.0035854578018188477, + 0.02604985050857067, + 0.0075349360704422, + 0.006656005047261715, + 0.0029649098869413137, + 0.02392280101776123, + 0.019443996250629425, + 0.01387383509427309, + 0.012644409202039242, + 0.026551369577646255, + 0.015091247856616974, + 0.012568177655339241, + 0.01269975770264864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.020916594192385674 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense emerald green bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures, a man in a white silk robe and a woman in a red suit, engaged in a sword fight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the characters perform swift movements including spinning and dodging between bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sword collisions produce dazzling golden sparks that briefly illuminate the characters' faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and smoothly panning sideways to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016086315736174583 + ], + "opposite_logprob_tokens": [ + -4.141086101531982 + ], + "logprob_reward": -0.016086315736174583, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016086315736174583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3760063946247101, + 0.03206348419189453, + 0.14055754244327545, + 0.08209919929504395, + 0.03869512304663658, + 0.04310663044452667, + 0.05165835842490196, + 0.0547245629131794, + 0.11848143488168716, + 0.06260731816291809 + ], + "credit_map": [ + 0.08289210498332977, + 0.030063331127166748, + 0.03669017553329468, + 0.019197212532162666, + 0.10375962406396866, + 0.036220841109752655, + 0.035067882388830185, + 0.03211523965001106, + 0.005880990065634251, + 0.005069445818662643, + 0.0024506240151822567, + 0.0016589033184573054, + 0.004350976087152958, + 0.004954928997904062, + 0.003912295214831829, + 0.0037853228859603405, + 0.011266927234828472, + 0.016109157353639603, + 0.010956725105643272, + 0.017326654866337776, + 0.03202948346734047, + 0.017605403438210487, + 0.01868181861937046, + 0.016581371426582336, + 0.006817482877522707, + 0.008870684541761875, + 0.015803789719939232, + 0.011400403454899788, + 0.01285590324550867, + 0.015760065987706184, + 0.005407455377280712, + 0.005183407571166754, + 0.005678340792655945, + 0.0048313201405107975, + 0.002899812301620841, + 0.004085302818566561, + 0.00504288962110877, + 0.006299642380326986, + 0.005258714314550161, + 0.004599100910127163, + 0.0066763670183718204, + 0.008209428749978542, + 0.004026912618428469, + 0.0037423938047140837, + 0.007140559144318104, + 0.005454007536172867, + 0.004343534354120493, + 0.0035134253557771444, + 0.005573862232267857, + 0.010416689328849316, + 0.0027697139885276556, + 0.005559287033975124, + 0.012334437109529972, + 0.006711174733936787, + 0.004675315693020821, + 0.0036178804002702236, + 0.0065503232181072235, + 0.009028682485222816, + 0.004745475482195616, + 0.006130596622824669, + 0.0074309962801635265, + 0.006206794176250696, + 0.00721355527639389, + 0.007418142631649971, + 0.009280871599912643, + 0.032176725566387177, + 0.02883545309305191, + 0.011283853091299534, + 0.011448814533650875, + 0.013278212398290634, + 0.008092259988188744, + 0.004085243213921785, + 0.009015494957566261, + 0.006079564802348614, + 0.007929680868983269, + 0.0063868616707623005, + 0.009531505405902863, + 0.007445215247571468, + 0.0066318209283053875, + 0.009587175212800503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.016086315736174583 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense emerald green bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures, a man in a white silk robe and a woman in a red suit, engaged in a sword fight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the characters perform swift movements including spinning and dodging between bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sword collisions produce dazzling golden sparks that briefly illuminate the characters' faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and smoothly panning sideways to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251067224889994 + ], + "opposite_logprob_tokens": [ + -5.255250930786133 + ], + "logprob_reward": -0.005251067224889994, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251067224889994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22526559233665466, + 0.057282429188489914, + 0.07816357910633087, + 0.26502177119255066, + 0.042262665927410126, + 0.06583591550588608, + 0.033838000148534775, + 0.043669234961271286, + 0.1447588950395584, + 0.04390191286802292 + ], + "credit_map": [ + 0.04184092581272125, + 0.02152911387383938, + 0.037459537386894226, + 0.020117437466979027, + 0.04247964173555374, + 0.02304082177579403, + 0.017532337456941605, + 0.021265793591737747, + 0.007995743304491043, + 0.0052950442768633366, + 0.005740404594689608, + 0.004582738503813744, + 0.009225642308592796, + 0.012557282112538815, + 0.005095878150314093, + 0.006789700128138065, + 0.009441068395972252, + 0.007485213689506054, + 0.007059017196297646, + 0.008895087987184525, + 0.00942965503782034, + 0.011719855479896069, + 0.012411144562065601, + 0.011722537688910961, + 0.020192505791783333, + 0.028279075399041176, + 0.05302167311310768, + 0.03258845582604408, + 0.02344954013824463, + 0.07149794697761536, + 0.01820559985935688, + 0.01778694987297058, + 0.004190720617771149, + 0.004934426862746477, + 0.0030321672093123198, + 0.005943595431745052, + 0.005804524756968021, + 0.007967920042574406, + 0.005635668523609638, + 0.004753646440804005, + 0.011325346305966377, + 0.011847791261970997, + 0.0046861302107572556, + 0.005032982677221298, + 0.00804950762540102, + 0.009747931733727455, + 0.006676852703094482, + 0.008469371125102043, + 0.0046276855282485485, + 0.005394861567765474, + 0.0029464170802384615, + 0.0037577676121145487, + 0.0061177597381174564, + 0.004414963535964489, + 0.003829674795269966, + 0.0027488681953400373, + 0.005998218897730112, + 0.006540603470057249, + 0.003536337288096547, + 0.004552118480205536, + 0.0052277203649282455, + 0.0056831021793186665, + 0.006544818636029959, + 0.005586313083767891, + 0.0072495718486607075, + 0.05260041356086731, + 0.01915663667023182, + 0.008321440778672695, + 0.03291182965040207, + 0.01307288371026516, + 0.0075046648271381855, + 0.0039414577186107635, + 0.0053646196611225605, + 0.0066888402216136456, + 0.004298501648008823, + 0.004770063795149326, + 0.007588203065097332, + 0.006786476820707321, + 0.004262406378984451, + 0.004142802208662033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.005251067224889994 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense emerald green bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures, a man in a white silk robe and a woman in a red suit, engaged in a sword fight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the characters perform swift movements including spinning and dodging between bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sword collisions produce dazzling golden sparks that briefly illuminate the characters' faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and smoothly panning sideways to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048646148294210434 + ], + "opposite_logprob_tokens": [ + -3.0486462116241455 + ], + "logprob_reward": -0.048646148294210434, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048646148294210434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1331702619791031, + 0.040885504335165024, + 0.09727615863084793, + 0.12256071716547012, + 0.02386315166950226, + 0.05264420434832573, + 0.11700081825256348, + 0.03812307491898537, + 0.3238108158111572, + 0.0506652407348156 + ], + "credit_map": [ + 0.03271747753024101, + 0.010036162100732327, + 0.014958384446799755, + 0.008927812799811363, + 0.02433738484978676, + 0.01588604785501957, + 0.012222089804708958, + 0.01408490538597107, + 0.005933528300374746, + 0.005110230762511492, + 0.005311476532369852, + 0.004126464948058128, + 0.005152678582817316, + 0.006207325030118227, + 0.003604287514463067, + 0.005439512897282839, + 0.009852348826825619, + 0.008534999564290047, + 0.014023363590240479, + 0.010516466572880745, + 0.011758816428482533, + 0.012258973903954029, + 0.016051240265369415, + 0.014279943890869617, + 0.008055498823523521, + 0.02499289996922016, + 0.012149043381214142, + 0.012948587536811829, + 0.022903814911842346, + 0.025691432878375053, + 0.007908505387604237, + 0.007910925894975662, + 0.00361925782635808, + 0.0033280358184129, + 0.0023680522572249174, + 0.003376787295565009, + 0.0030075509566813707, + 0.00310858734883368, + 0.002842811169102788, + 0.002212068298831582, + 0.0064449612982571125, + 0.0068587386049330235, + 0.004287468735128641, + 0.0032954635098576546, + 0.010472719557583332, + 0.00884066428989172, + 0.0066058021038770676, + 0.005838387180119753, + 0.011229705065488815, + 0.03096764348447323, + 0.009343460202217102, + 0.011824904009699821, + 0.017450110986828804, + 0.013813056983053684, + 0.014235241338610649, + 0.008136704564094543, + 0.0034323600120842457, + 0.004566719755530357, + 0.002458291593939066, + 0.004723685327917337, + 0.006739777512848377, + 0.004921740852296352, + 0.0054503134451806545, + 0.005830185487866402, + 0.01249272096902132, + 0.12300076335668564, + 0.014260506257414818, + 0.00890841893851757, + 0.11234167963266373, + 0.03156103938817978, + 0.015783704817295074, + 0.005461996421217918, + 0.006856785621494055, + 0.008582833223044872, + 0.005289675667881966, + 0.005603603087365627, + 0.008274005725979805, + 0.005937862675637007, + 0.004440162796527147, + 0.00568031519651413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.048646148294210434 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense emerald green bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures, a man in a white silk robe and a woman in a red suit, engaged in a sword fight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the characters perform swift movements including spinning and dodging between bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sword collisions produce dazzling golden sparks that briefly illuminate the characters' faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and smoothly panning sideways to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023273982107639313 + ], + "opposite_logprob_tokens": [ + -3.7732739448547363 + ], + "logprob_reward": -0.023273982107639313, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023273982107639313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1864953190088272, + 0.061712272465229034, + 0.08663659542798996, + 0.0681365355849266, + 0.053377002477645874, + 0.06802719831466675, + 0.06673239171504974, + 0.06323874741792679, + 0.3028477132320404, + 0.04279624670743942 + ], + "credit_map": [ + 0.03396093100309372, + 0.026800110936164856, + 0.01694040186703205, + 0.014013230800628662, + 0.025732845067977905, + 0.02764979563653469, + 0.019242756068706512, + 0.022155245766043663, + 0.0091407410800457, + 0.009189498610794544, + 0.004244567360728979, + 0.0030464523006230593, + 0.009402777999639511, + 0.014374412596225739, + 0.005925850011408329, + 0.0063879769295454025, + 0.006676395423710346, + 0.011992543935775757, + 0.009669843129813671, + 0.014629941433668137, + 0.013430288061499596, + 0.010714071802794933, + 0.009201066568493843, + 0.010322454385459423, + 0.007451117504388094, + 0.00960110779851675, + 0.006229302380234003, + 0.007858139462769032, + 0.013699464499950409, + 0.012833322398364544, + 0.004967475309967995, + 0.005496610421687365, + 0.007854919880628586, + 0.0061920443549752235, + 0.004976666998118162, + 0.009112689644098282, + 0.005720983259379864, + 0.0070988330990076065, + 0.007028291933238506, + 0.005392576102167368, + 0.009141306392848492, + 0.009059745818376541, + 0.005308343097567558, + 0.004691002424806356, + 0.015215516090393066, + 0.01034725271165371, + 0.006488594692200422, + 0.007775441277772188, + 0.009098240174353123, + 0.01082173828035593, + 0.004184518475085497, + 0.006831602659076452, + 0.012467877008020878, + 0.010117941536009312, + 0.008853726089000702, + 0.00435674749314785, + 0.00534052774310112, + 0.010443140752613544, + 0.004654940217733383, + 0.005257320590317249, + 0.009129472076892853, + 0.0078382333740592, + 0.011023903265595436, + 0.009551212191581726, + 0.013275064527988434, + 0.10235073417425156, + 0.013439529575407505, + 0.01175587810575962, + 0.10723560303449631, + 0.0318341888487339, + 0.017178768292069435, + 0.005777968093752861, + 0.005737312603741884, + 0.006613212637603283, + 0.0046352664940059185, + 0.005718981847167015, + 0.006843946408480406, + 0.005178655963391066, + 0.003979123663157225, + 0.004089751280844212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.023273982107639313 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic medium shot, shot from a slightly upward angle, smoothly pans sideways, closely following a man and woman engaged in a fierce confrontation within a dense, deep, emerald green bamboo forest. The man, with his long, dark hair tied high with a headband, is handsome, his eyes piercing like an eagle. He wears a flowing white silk robe, its wide sleeves fluttering in the air with his swift movements. The woman, also wearing a high ponytail, has a beautiful face and a focused, stern expression. She wears a sharply tailored red suit, creating a stark and striking visual contrast with the man's white robe. The afternoon sun struggles to penetrate the gaps between the layers of bamboo leaves, casting a dappled, flickering light on the ground, shrouded in a light mist. The two men's figures were as fast as lightning, deftly shuttling, moving and spinning between the straight and tall bamboo poles. The fine steel swords in their hands drew deadly and graceful silver arcs in the air. Every precise and fierce collision instantly burst out dazzling golden sparks, briefly illuminating their focused faces and the cold light flowing on the sword blades. The whole fighting process was full of power and like a carefully choreographed dance, swift and graceful, creating a martial arts atmosphere that was both intense and poetic against the backdrop of the quiet bamboo forest.", + "seed": 23683, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_6/index_3007_seed_23683_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01866451231762767, + "vlm_logprob": -0.01866451231762767, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.01866451231762767, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense emerald green bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures, a man in a white silk robe and a woman in a red suit, engaged in a sword fight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the characters perform swift movements including spinning and dodging between bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sword collisions produce dazzling golden sparks that briefly illuminate the characters' faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and smoothly panning sideways to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012261842377483845 + ], + "opposite_logprob_tokens": [ + -4.887261867523193 + ], + "logprob_reward": -0.012261842377483845, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.012261842377483845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4094081223011017, + 0.04905974864959717, + 0.07807599753141403, + 0.07289887964725494, + 0.04402053728699684, + 0.03956029936671257, + 0.06433403491973877, + 0.09687687456607819, + 0.059308119118213654, + 0.08645740896463394 + ], + "credit_map": [ + 0.040839213877916336, + 0.08581187576055527, + 0.04812600091099739, + 0.1095651239156723, + 0.03192644193768501, + 0.027273833751678467, + 0.02635331265628338, + 0.03951229527592659, + 0.007146832533180714, + 0.00535054923966527, + 0.006887222174555063, + 0.0053069936111569405, + 0.0058130547404289246, + 0.00666114641353488, + 0.005415352061390877, + 0.006478596944361925, + 0.007347295060753822, + 0.010088586248457432, + 0.005529998801648617, + 0.006491734180599451, + 0.009246667847037315, + 0.0163094624876976, + 0.006157152820378542, + 0.01690509542822838, + 0.00705259433016181, + 0.01043227780610323, + 0.007965115830302238, + 0.004928769543766975, + 0.01161721721291542, + 0.014327692799270153, + 0.005572935100644827, + 0.01100227888673544, + 0.00650780601426959, + 0.0046461704187095165, + 0.00784936361014843, + 0.004278369713574648, + 0.005213066469877958, + 0.005214236211031675, + 0.004707796964794397, + 0.005603731144219637, + 0.00586646469309926, + 0.009632987901568413, + 0.002808180171996355, + 0.005987321957945824, + 0.005993612576276064, + 0.004250145982950926, + 0.0027171741239726543, + 0.0023044112604111433, + 0.010788921266794205, + 0.007781909313052893, + 0.009105158969759941, + 0.007788662798702717, + 0.009880203753709793, + 0.005516015458852053, + 0.009092322550714016, + 0.0043808394111692905, + 0.015169182792305946, + 0.011638703756034374, + 0.010340766981244087, + 0.012977266684174538, + 0.017998773604631424, + 0.008383585140109062, + 0.013768687844276428, + 0.006599913816899061, + 0.009189790114760399, + 0.00878897961229086, + 0.009467806667089462, + 0.009336891584098339, + 0.003946831449866295, + 0.006531762424856424, + 0.006009832490235567, + 0.0060362257063388824, + 0.01435467042028904, + 0.026566511020064354, + 0.011055502109229565, + 0.004988677334040403, + 0.0110361622646451, + 0.007404941599816084, + 0.006396748591214418, + 0.004654204472899437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.012261842377483845 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense emerald green bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures, a man in a white silk robe and a woman in a red suit, engaged in a sword fight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the characters perform swift movements including spinning and dodging between bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sword collisions produce dazzling golden sparks that briefly illuminate the characters' faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and smoothly panning sideways to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026348400861024857 + ], + "opposite_logprob_tokens": [ + -3.651348352432251 + ], + "logprob_reward": -0.026348400861024857, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026348400861024857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0804460197687149, + 0.02846420742571354, + 0.19826142489910126, + 0.19738061726093292, + 0.05177997052669525, + 0.030622603371739388, + 0.09411635994911194, + 0.11801701784133911, + 0.1629670411348343, + 0.037944793701171875 + ], + "credit_map": [ + 0.01036849245429039, + 0.010784833692014217, + 0.015004809014499187, + 0.010025075636804104, + 0.010316036641597748, + 0.00813105795532465, + 0.007955124601721764, + 0.007860582321882248, + 0.005139684770256281, + 0.003137940540909767, + 0.003302199300378561, + 0.0030643343925476074, + 0.004415073432028294, + 0.002737592440098524, + 0.0034267576411366463, + 0.00324062118306756, + 0.0195561982691288, + 0.023929640650749207, + 0.009463418275117874, + 0.01374767441302538, + 0.04089973866939545, + 0.031257983297109604, + 0.022176606580615044, + 0.03723013773560524, + 0.02145678922533989, + 0.016695361584424973, + 0.014667438343167305, + 0.019987106323242188, + 0.034135349094867706, + 0.029105808585882187, + 0.03473812714219093, + 0.026594627648591995, + 0.007991086691617966, + 0.012586233206093311, + 0.007108316291123629, + 0.005848176311701536, + 0.005043879151344299, + 0.005167942959815264, + 0.002976885763928294, + 0.005057444330304861, + 0.0038372944109141827, + 0.006530803628265858, + 0.0021116980351507664, + 0.004294377285987139, + 0.004030033014714718, + 0.004781715571880341, + 0.0022690831683576107, + 0.0027675952296704054, + 0.013573747128248215, + 0.011420130729675293, + 0.011246339417994022, + 0.01306673139333725, + 0.01308559812605381, + 0.012642563320696354, + 0.011204365640878677, + 0.007876879535615444, + 0.01116118673235178, + 0.014244511723518372, + 0.01635637693107128, + 0.011325458064675331, + 0.009881305508315563, + 0.017352771013975143, + 0.01829206570982933, + 0.019403325393795967, + 0.029128853231668472, + 0.025797763839364052, + 0.020652180537581444, + 0.024532904848456383, + 0.010615987703204155, + 0.017124783247709274, + 0.015076497569680214, + 0.0200380589812994, + 0.005930683109909296, + 0.005358582362532616, + 0.008845699019730091, + 0.004810144193470478, + 0.004197296220809221, + 0.003114011837169528, + 0.002854615915566683, + 0.0028337561525404453 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 8 + ], + "selected_reward": -0.026348400861024857 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense emerald green bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures, a man in a white silk robe and a woman in a red suit, engaged in a sword fight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the characters perform swift movements including spinning and dodging between bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sword collisions produce dazzling golden sparks that briefly illuminate the characters' faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and smoothly panning sideways to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040937201119959354 + ], + "opposite_logprob_tokens": [ + -5.504093647003174 + ], + "logprob_reward": -0.0040937201119959354, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040937201119959354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2032930999994278, + 0.05605386197566986, + 0.17632721364498138, + 0.14307992160320282, + 0.09578650444746017, + 0.09758006036281586, + 0.050816573202610016, + 0.09560446441173553, + 0.04244169965386391, + 0.039016660302877426 + ], + "credit_map": [ + 0.027504906058311462, + 0.04582396522164345, + 0.022869639098644257, + 0.04362225532531738, + 0.022355305030941963, + 0.017544589936733246, + 0.0119108185172081, + 0.01166162546724081, + 0.010729864239692688, + 0.00404420355334878, + 0.010967186652123928, + 0.005283975508064032, + 0.005104438401758671, + 0.007295021787285805, + 0.006182023324072361, + 0.006447149906307459, + 0.018506966531276703, + 0.020495766773819923, + 0.008607265539467335, + 0.01238085888326168, + 0.03977737948298454, + 0.027420934289693832, + 0.01620345748960972, + 0.032934579998254776, + 0.015587035566568375, + 0.015218222513794899, + 0.012221311219036579, + 0.012148040346801281, + 0.021747063845396042, + 0.022355876863002777, + 0.020356332883238792, + 0.02344602905213833, + 0.01454144436866045, + 0.02347293682396412, + 0.013835071586072445, + 0.012294295243918896, + 0.008349313400685787, + 0.008045053109526634, + 0.006162656005471945, + 0.009085734374821186, + 0.012749570421874523, + 0.013952216133475304, + 0.008182382211089134, + 0.009019155986607075, + 0.01727667637169361, + 0.016611512750387192, + 0.011120371520519257, + 0.008668173104524612, + 0.0073635331355035305, + 0.006479863543063402, + 0.005123828537762165, + 0.005355094093829393, + 0.009026464074850082, + 0.007686090189963579, + 0.004728392697870731, + 0.00505330553278327, + 0.009191008284687996, + 0.011486939154565334, + 0.009047801606357098, + 0.007415585685521364, + 0.011926557868719101, + 0.016259172931313515, + 0.015020197257399559, + 0.01525720115751028, + 0.00523077929392457, + 0.0043237521313130856, + 0.004565294366329908, + 0.005573204252868891, + 0.004093158058822155, + 0.00571654736995697, + 0.009055288508534431, + 0.0038836784660816193, + 0.006341563072055578, + 0.004945102147758007, + 0.007152393460273743, + 0.003959824331104755, + 0.005363355856388807, + 0.0037544274237006903, + 0.003661077469587326, + 0.003838916076347232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.0040937201119959354 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense emerald green bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures, a man in a white silk robe and a woman in a red suit, engaged in a sword fight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the characters perform swift movements including spinning and dodging between bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sword collisions produce dazzling golden sparks that briefly illuminate the characters' faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and smoothly panning sideways to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809007257223129 + ], + "opposite_logprob_tokens": [ + -3.2880899906158447 + ], + "logprob_reward": -0.03809007257223129, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809007257223129, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2081238478422165, + 0.060050711035728455, + 0.0786106139421463, + 0.09329516440629959, + 0.1847548484802246, + 0.05400864779949188, + 0.04873375594615936, + 0.108828105032444, + 0.06211179494857788, + 0.10148250311613083 + ], + "credit_map": [ + 0.04716083034873009, + 0.028982508927583694, + 0.02540457248687744, + 0.023669641464948654, + 0.028070811182260513, + 0.014227124862372875, + 0.018837830051779747, + 0.02177051641047001, + 0.008688744157552719, + 0.005127079784870148, + 0.00856990460306406, + 0.005607527680695057, + 0.007986287586390972, + 0.009888041764497757, + 0.0049791159108281136, + 0.009204007685184479, + 0.010755591094493866, + 0.008668804541230202, + 0.00914493016898632, + 0.005717175547033548, + 0.008778266608715057, + 0.013103960081934929, + 0.0067863184958696365, + 0.01565556414425373, + 0.013946648687124252, + 0.011697801761329174, + 0.0069287363439798355, + 0.006526513025164604, + 0.010526116006076336, + 0.016864554956555367, + 0.00985317025333643, + 0.01695161871612072, + 0.028248554095625877, + 0.04637458920478821, + 0.023418083786964417, + 0.022914357483386993, + 0.016788911074399948, + 0.018540633842349052, + 0.011175090447068214, + 0.017294619232416153, + 0.008638043887913227, + 0.00913381576538086, + 0.005076503846794367, + 0.006219387985765934, + 0.008115310221910477, + 0.008169768378138542, + 0.0048980507999658585, + 0.0037577636539936066, + 0.006073646247386932, + 0.005149464588612318, + 0.00465631578117609, + 0.00545542873442173, + 0.007443086244165897, + 0.007101274095475674, + 0.004119472578167915, + 0.008735067211091518, + 0.015693800523877144, + 0.01133954618126154, + 0.01198501791805029, + 0.014666273258626461, + 0.01762247085571289, + 0.010385067202150822, + 0.018998179584741592, + 0.008137746714055538, + 0.010367128998041153, + 0.011190921999514103, + 0.00904871802777052, + 0.008616382256150246, + 0.0035602448042482138, + 0.00497742323204875, + 0.008688844740390778, + 0.0056621236726641655, + 0.014545919373631477, + 0.021380499005317688, + 0.019255658611655235, + 0.01157446950674057, + 0.010441656224429607, + 0.0080962423235178, + 0.0065536838956177235, + 0.00963437557220459 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.03809007257223129 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense emerald green bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures, a man in a white silk robe and a woman in a red suit, engaged in a sword fight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the characters perform swift movements including spinning and dodging between bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sword collisions produce dazzling golden sparks that briefly illuminate the characters' faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and smoothly panning sideways to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012528525665402412 + ], + "opposite_logprob_tokens": [ + -4.387528419494629 + ], + "logprob_reward": -0.012528525665402412, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012528525665402412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2419736385345459, + 0.04874614626169205, + 0.05712012201547623, + 0.1518842726945877, + 0.07139290869235992, + 0.05994708836078644, + 0.10070062428712845, + 0.07326655089855194, + 0.07398612797260284, + 0.12098252773284912 + ], + "credit_map": [ + 0.03919415548443794, + 0.03910965472459793, + 0.02880259044468403, + 0.05272497981786728, + 0.022457551211118698, + 0.019715191796422005, + 0.0158429853618145, + 0.02412652038037777, + 0.009435485117137432, + 0.004791760817170143, + 0.0061879283748567104, + 0.0043794927187263966, + 0.007356615271419287, + 0.006136408541351557, + 0.005401495844125748, + 0.005056959111243486, + 0.006621960550546646, + 0.005940682254731655, + 0.005533130839467049, + 0.004135871771723032, + 0.00948597677052021, + 0.010764592327177525, + 0.005120438989251852, + 0.009517470374703407, + 0.016629472374916077, + 0.017655784264206886, + 0.012553459033370018, + 0.013691280037164688, + 0.027283888310194016, + 0.01887086220085621, + 0.023222925141453743, + 0.021976586431264877, + 0.011469338089227676, + 0.016330288723111153, + 0.010803098790347576, + 0.007123323157429695, + 0.007121481001377106, + 0.007534220814704895, + 0.004399543162435293, + 0.006611617282032967, + 0.006694324780255556, + 0.01282140240073204, + 0.005856581963598728, + 0.005423890892416239, + 0.01231275126338005, + 0.00733666867017746, + 0.004564751870930195, + 0.004936722107231617, + 0.017602048814296722, + 0.015073028393089771, + 0.013277540914714336, + 0.009509066119790077, + 0.01366045605391264, + 0.01220528781414032, + 0.007869734428822994, + 0.011503460817039013, + 0.011638334952294827, + 0.0076767937280237675, + 0.008478201925754547, + 0.006528518162667751, + 0.013294289819896221, + 0.0071179866790771484, + 0.01178412139415741, + 0.006748308427631855, + 0.011126533150672913, + 0.012800070457160473, + 0.00876575242727995, + 0.009112978354096413, + 0.0053484756499528885, + 0.010816285386681557, + 0.009298907592892647, + 0.006717119365930557, + 0.021144231781363487, + 0.02704729698598385, + 0.024410126730799675, + 0.0110939284786582, + 0.012550914660096169, + 0.008689260110259056, + 0.008617380633950233, + 0.007429380435496569 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.012528525665402412 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic medium shot, shot from a slightly upward angle, smoothly pans sideways, closely following a man and woman engaged in a fierce confrontation within a dense, deep, emerald green bamboo forest. The man, with his long, dark hair tied high with a headband, is handsome, his eyes piercing like an eagle. He wears a flowing white silk robe, its wide sleeves fluttering in the air with his swift movements. The woman, also wearing a high ponytail, has a beautiful face and a focused, stern expression. She wears a sharply tailored red suit, creating a stark and striking visual contrast with the man's white robe. The afternoon sun struggles to penetrate the gaps between the layers of bamboo leaves, casting a dappled, flickering light on the ground, shrouded in a light mist. The two men's figures were as fast as lightning, deftly shuttling, moving and spinning between the straight and tall bamboo poles. The fine steel swords in their hands drew deadly and graceful silver arcs in the air. Every precise and fierce collision instantly burst out dazzling golden sparks, briefly illuminating their focused faces and the cold light flowing on the sword blades. The whole fighting process was full of power and like a carefully choreographed dance, swift and graceful, creating a martial arts atmosphere that was both intense and poetic against the backdrop of the quiet bamboo forest.", + "seed": 23684, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_6/index_3007_seed_23684_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02409580275416374, + "vlm_logprob": -0.02409580275416374, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.02409580275416374, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense emerald green bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures, a man in a white silk robe and a woman in a red suit, engaged in a sword fight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the characters perform swift movements including spinning and dodging between bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sword collisions produce dazzling golden sparks that briefly illuminate the characters' faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and smoothly panning sideways to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01955925114452839 + ], + "opposite_logprob_tokens": [ + -4.644559383392334 + ], + "logprob_reward": -0.01955925114452839, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.01955925114452839, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09691886603832245, + 0.05382800102233887, + 0.0222331490367651, + 0.02544679306447506, + 0.01757017895579338, + 0.34822601079940796, + 0.20142462849617004, + 0.16380707919597626, + 0.024236498400568962, + 0.04630879685282707 + ], + "credit_map": [ + 0.018650110810995102, + 0.015113179571926594, + 0.014898017980158329, + 0.009513272903859615, + 0.010611383244395256, + 0.009978853166103363, + 0.009984876029193401, + 0.00816918071359396, + 0.012774648144841194, + 0.007630026899278164, + 0.003989735618233681, + 0.00529377069324255, + 0.005969327874481678, + 0.0057714250870049, + 0.008227541111409664, + 0.004171527922153473, + 0.003815628122538328, + 0.003332747844979167, + 0.0017325971275568008, + 0.0020006869453936815, + 0.002980560762807727, + 0.0024417478125542402, + 0.0029111874755471945, + 0.003017992712557316, + 0.002564928960055113, + 0.004751448053866625, + 0.0021595428697764874, + 0.00281934579834342, + 0.003005767473950982, + 0.005316259805113077, + 0.00196077860891819, + 0.002868722192943096, + 0.00366700766608119, + 0.0020519932731986046, + 0.003001607721671462, + 0.0016808734508231282, + 0.0018664891831576824, + 0.0018973114201799035, + 0.0022096987813711166, + 0.0011951976921409369, + 0.03893876448273659, + 0.01108053419739008, + 0.07845071703195572, + 0.031126927584409714, + 0.026036223396658897, + 0.014986886642873287, + 0.12094660103321075, + 0.026659376919269562, + 0.021743252873420715, + 0.010847161523997784, + 0.05933203548192978, + 0.01086306944489479, + 0.01575946994125843, + 0.0679875984787941, + 0.005835334304720163, + 0.00905671063810587, + 0.027569685131311417, + 0.008008590899407864, + 0.009556584991514683, + 0.02488846145570278, + 0.012037436477839947, + 0.06160230189561844, + 0.008666369132697582, + 0.011477652005851269, + 0.0039701880887150764, + 0.0030182639602571726, + 0.004163097590208054, + 0.0020739086903631687, + 0.0033824671991169453, + 0.0036293710581958294, + 0.0024244238156825304, + 0.0015747793950140476, + 0.005887479521334171, + 0.007144318427890539, + 0.005393957253545523, + 0.0038524125702679157, + 0.008215884678065777, + 0.00577588239684701, + 0.005840872414410114, + 0.004197991918772459 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.01955925114452839 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense emerald green bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures, a man in a white silk robe and a woman in a red suit, engaged in a sword fight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the characters perform swift movements including spinning and dodging between bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sword collisions produce dazzling golden sparks that briefly illuminate the characters' faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and smoothly panning sideways to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014214457012712955 + ], + "opposite_logprob_tokens": [ + -4.264214515686035 + ], + "logprob_reward": -0.014214457012712955, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014214457012712955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06955669075250626, + 0.2306968867778778, + 0.04685446619987488, + 0.05890266224741936, + 0.04852183908224106, + 0.20645330846309662, + 0.07820548117160797, + 0.1574152559041977, + 0.048105474561452866, + 0.055287931114435196 + ], + "credit_map": [ + 0.014289750717580318, + 0.009273956529796124, + 0.009738209657371044, + 0.005971934180706739, + 0.007726916112005711, + 0.00822862982749939, + 0.008047164417803288, + 0.006280131172388792, + 0.028520962223410606, + 0.017507493495941162, + 0.018269497901201248, + 0.016974182799458504, + 0.05376805365085602, + 0.04109295830130577, + 0.03127923235297203, + 0.02328452281653881, + 0.004806194920092821, + 0.006216695066541433, + 0.003248128341510892, + 0.0032938187941908836, + 0.0064545380882918835, + 0.004602895118296146, + 0.007705465890467167, + 0.010526733472943306, + 0.008105388842523098, + 0.011139651760458946, + 0.00403412664309144, + 0.0042482418939471245, + 0.0060170674696564674, + 0.009549370966851711, + 0.006202299147844315, + 0.009606515988707542, + 0.012821379117667675, + 0.004558466374874115, + 0.007364506833255291, + 0.004997231997549534, + 0.0055246613919734955, + 0.005652414634823799, + 0.004292027559131384, + 0.0033111481461673975, + 0.02268628589808941, + 0.008323297835886478, + 0.04704129323363304, + 0.018119407817721367, + 0.015369538217782974, + 0.00848591048270464, + 0.07033194601535797, + 0.016095632687211037, + 0.007108971010893583, + 0.004024909809231758, + 0.022259963676333427, + 0.003795279422774911, + 0.0067060585133731365, + 0.02699812687933445, + 0.0028856745921075344, + 0.004426499363034964, + 0.02630479820072651, + 0.008861889131367207, + 0.008403785526752472, + 0.020128389820456505, + 0.012793000787496567, + 0.05924644321203232, + 0.009030072018504143, + 0.012646878138184547, + 0.009167206473648548, + 0.007119001820683479, + 0.007331832777708769, + 0.004076935816556215, + 0.006150847300887108, + 0.0047704060561954975, + 0.0059317732229828835, + 0.0035574710927903652, + 0.005824890453368425, + 0.00885953102260828, + 0.004752529785037041, + 0.0060103340074419975, + 0.007408280856907368, + 0.00902037788182497, + 0.0057687475346028805, + 0.007643235381692648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 7 + ], + "selected_reward": -0.014214457012712955 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense emerald green bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures, a man in a white silk robe and a woman in a red suit, engaged in a sword fight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the characters perform swift movements including spinning and dodging between bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sword collisions produce dazzling golden sparks that briefly illuminate the characters' faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and smoothly panning sideways to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040963320061564445 + ], + "opposite_logprob_tokens": [ + -5.504096508026123 + ], + "logprob_reward": -0.0040963320061564445, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040963320061564445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22576697170734406, + 0.09034192562103271, + 0.06517057865858078, + 0.0877753272652626, + 0.06082268804311752, + 0.06539025902748108, + 0.14767540991306305, + 0.15471389889717102, + 0.06344349682331085, + 0.0388994924724102 + ], + "credit_map": [ + 0.028336049988865852, + 0.03783227875828743, + 0.018937768414616585, + 0.02005733922123909, + 0.047224946320056915, + 0.03503049537539482, + 0.024248991161584854, + 0.014099108055233955, + 0.00915639940649271, + 0.007906312122941017, + 0.007717655971646309, + 0.006883381400257349, + 0.023063523694872856, + 0.014238747768104076, + 0.012099445797502995, + 0.00927646178752184, + 0.008266627788543701, + 0.014196109026670456, + 0.0032077522482722998, + 0.004408170469105244, + 0.0099458834156394, + 0.008519234135746956, + 0.007128754630684853, + 0.009498047642409801, + 0.00953592173755169, + 0.016556352376937866, + 0.006642374210059643, + 0.005754262674599886, + 0.009813791140913963, + 0.014107842929661274, + 0.010390754789113998, + 0.01497401762753725, + 0.010244905017316341, + 0.0075055379420518875, + 0.006720163393765688, + 0.008635500445961952, + 0.00949623342603445, + 0.0067381723783910275, + 0.007689208723604679, + 0.0037929655518382788, + 0.008980559185147285, + 0.00585483992472291, + 0.0109682222828269, + 0.006020255386829376, + 0.004870517645031214, + 0.004753998946398497, + 0.015865907073020935, + 0.00807595532387495, + 0.013687945902347565, + 0.006860358640551567, + 0.04515179991722107, + 0.007254024036228657, + 0.01188285369426012, + 0.053019799292087555, + 0.004040809813886881, + 0.005777813494205475, + 0.025137685239315033, + 0.008671925403177738, + 0.010474910028278828, + 0.023482074961066246, + 0.011360681615769863, + 0.05624108761548996, + 0.008147355169057846, + 0.011198188178241253, + 0.009275883436203003, + 0.011052737012505531, + 0.014517896808683872, + 0.005852516740560532, + 0.0075876349583268166, + 0.006253126543015242, + 0.00501966942101717, + 0.0038840286433696747, + 0.0052282800897955894, + 0.006256076041609049, + 0.0038641919381916523, + 0.004885348491370678, + 0.006895812693983316, + 0.004080047365278006, + 0.00451460899785161, + 0.0031751268543303013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.0040963320061564445 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense emerald green bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures, a man in a white silk robe and a woman in a red suit, engaged in a sword fight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the characters perform swift movements including spinning and dodging between bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sword collisions produce dazzling golden sparks that briefly illuminate the characters' faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and smoothly panning sideways to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062037527561187744 + ], + "opposite_logprob_tokens": [ + -2.812037467956543 + ], + "logprob_reward": -0.062037527561187744, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062037527561187744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20092947781085968, + 0.07345468550920486, + 0.08685201406478882, + 0.11949115991592407, + 0.04787944257259369, + 0.20796559751033783, + 0.08680541813373566, + 0.07614514231681824, + 0.04779280349612236, + 0.052684277296066284 + ], + "credit_map": [ + 0.033253058791160583, + 0.03686498478055, + 0.02269710786640644, + 0.01714080199599266, + 0.029182100668549538, + 0.028388725593686104, + 0.016667591407895088, + 0.01673509180545807, + 0.009246746078133583, + 0.006948929745703936, + 0.005503046326339245, + 0.006148763932287693, + 0.015518421307206154, + 0.011778942309319973, + 0.01020137034356594, + 0.008108467794954777, + 0.010382027365267277, + 0.012769326567649841, + 0.007134763989597559, + 0.007633816916495562, + 0.011629940010607243, + 0.009284375235438347, + 0.012664892710745335, + 0.015352873131632805, + 0.012204894796013832, + 0.016754718497395515, + 0.00868806429207325, + 0.008369172923266888, + 0.013986613601446152, + 0.023208435624837875, + 0.014874025247991085, + 0.021405238658189774, + 0.011647866107523441, + 0.004288669675588608, + 0.006321942433714867, + 0.005717426538467407, + 0.00568432966247201, + 0.005364918615669012, + 0.005460871383547783, + 0.0033934174571186304, + 0.023919200524687767, + 0.007551215589046478, + 0.04442279413342476, + 0.01833520270884037, + 0.014709945768117905, + 0.009426782839000225, + 0.06876157969236374, + 0.020838866010308266, + 0.009459019638597965, + 0.005043342709541321, + 0.025502057746052742, + 0.004954206291586161, + 0.006965428590774536, + 0.02767363004386425, + 0.0031724239233881235, + 0.004035312682390213, + 0.011651715263724327, + 0.006675602402538061, + 0.006764802150428295, + 0.012916585430502892, + 0.005911856424063444, + 0.022885529324412346, + 0.004386129789054394, + 0.004952920135110617, + 0.010505583137273788, + 0.008272066712379456, + 0.005378559697419405, + 0.0034465789794921875, + 0.00593955023214221, + 0.008042808622121811, + 0.003562541212886572, + 0.002645116299390793, + 0.006229968275874853, + 0.005745770409703255, + 0.0071192095056176186, + 0.005734461825340986, + 0.005500428378582001, + 0.01058807224035263, + 0.005122891627252102, + 0.006643474102020264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.062037527561187744 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense emerald green bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two figures, a man in a white silk robe and a woman in a red suit, engaged in a sword fight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the characters perform swift movements including spinning and dodging between bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sword collisions produce dazzling golden sparks that briefly illuminate the characters' faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and smoothly panning sideways to follow the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020571446046233177 + ], + "opposite_logprob_tokens": [ + -3.89557147026062 + ], + "logprob_reward": -0.020571446046233177, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020571446046233177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12871153652668, + 0.05726718157529831, + 0.09502355754375458, + 0.10240726172924042, + 0.0329788476228714, + 0.21520379185676575, + 0.08855769038200378, + 0.1775585561990738, + 0.07294601947069168, + 0.029345553368330002 + ], + "credit_map": [ + 0.016926618292927742, + 0.020854713395237923, + 0.015155917964875698, + 0.011097817681729794, + 0.017938029021024704, + 0.018602700904011726, + 0.016800520941615105, + 0.011335229501128197, + 0.008814799599349499, + 0.006480593234300613, + 0.004763180855661631, + 0.0044937580823898315, + 0.010297753848135471, + 0.007780944928526878, + 0.0083705373108387, + 0.006265610456466675, + 0.011101962067186832, + 0.018323006108403206, + 0.005230084992945194, + 0.00818141084164381, + 0.014615513384342194, + 0.014658212661743164, + 0.009831073693931103, + 0.013082293793559074, + 0.014874656684696674, + 0.01931825838983059, + 0.013250740244984627, + 0.008555342443287373, + 0.010888129472732544, + 0.012724780477583408, + 0.009980087168514729, + 0.0128152621909976, + 0.004740801174193621, + 0.004313353449106216, + 0.00342470477335155, + 0.004347510170191526, + 0.004408269189298153, + 0.0037708384916186333, + 0.005414774175733328, + 0.0025585980620235205, + 0.023289918899536133, + 0.009119914844632149, + 0.04806685447692871, + 0.02066320925951004, + 0.015892500057816505, + 0.009334102272987366, + 0.07207641005516052, + 0.016760889440774918, + 0.009291048161685467, + 0.006750369910150766, + 0.024950727820396423, + 0.005355472676455975, + 0.007148128468543291, + 0.02901778556406498, + 0.0027979393489658833, + 0.0032462221570312977, + 0.028402913361787796, + 0.009941965341567993, + 0.01276459265500307, + 0.028011439368128777, + 0.01316071767359972, + 0.06390176713466644, + 0.00956715177744627, + 0.0118080023676157, + 0.0197847168892622, + 0.00736710662022233, + 0.007762521505355835, + 0.005879611242562532, + 0.011929076164960861, + 0.010623270645737648, + 0.005483128130435944, + 0.004116587340831757, + 0.003743907902389765, + 0.005198505241423845, + 0.002364673651754856, + 0.002783041913062334, + 0.004753234330564737, + 0.004976486321538687, + 0.002807877492159605, + 0.0027178265154361725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.020571446046233177 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_7/index_tensor([4896, 4896])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_7/index_tensor([4896, 4896])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3072f106079902e73b8c24f2eeaaef46c10612c5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_7/index_tensor([4896, 4896])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extremely close-up, static shot, positioned at a slightly elevated angle, focuses intensely on a shimmering, ruby-red gelatin dessert centered on a pristine white porcelain plate. Bright, warm sidelighting illuminates the scene, casting brilliant, mirror-like highlights on the dessert's smooth, moist surface and penetrating its translucent body, causing it to glow from within like a gem. The dessert possesses a slow, subtle sway, its entire form trembling with a mesmerizing rhythm. This movement emanates from its center in a continuous cascade of concentric ripples that radiate outward across its tense surface, distorting the reflections of light. A shallow depth of field, completely obscuring the background, ensures that every detail of the gelatin's jostling motion and the hypnotic play of light through its form are captured, creating a captivating and satisfyingly tactile atmosphere.", + "seed": 32999, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_7/index_4896_seed_32999_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.034423560835421085, + "vlm_logprob": -0.034423560835421085, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.034423560835421085, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a pristine white porcelain plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, warm sidelighting that creates mirror-like highlights on the dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dessert exhibit a slow, subtle sway with concentric ripples radiating outward from its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely obscured by a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006162451580166817 + ], + "opposite_logprob_tokens": [ + -6.506162643432617 + ], + "logprob_reward": -0.006162451580166817, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.006162451580166817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1575005054473877, + 0.05124741792678833, + 0.2778305113315582, + 0.04559894651174545, + 0.15147827565670013, + 0.06623919308185577, + 0.10776381939649582, + 0.04175873473286629, + 0.029733791947364807, + 0.07084875553846359 + ], + "credit_map": [ + 0.007130715064704418, + 0.01751561462879181, + 0.006094361189752817, + 0.02722204476594925, + 0.0495377853512764, + 0.007888189516961575, + 0.0287379901856184, + 0.013373811729252338, + 0.0031800265423953533, + 0.0066839526407420635, + 0.002509340178221464, + 0.00784276332706213, + 0.014985446818172932, + 0.003048742888495326, + 0.0060988436453044415, + 0.006898307707160711, + 0.004687041975557804, + 0.010169590823352337, + 0.006537689361721277, + 0.11546886712312698, + 0.012307293713092804, + 0.004992199596017599, + 0.10906799882650375, + 0.014599867165088654, + 0.0019505375530570745, + 0.006042304448783398, + 0.0025404354091733694, + 0.007710978388786316, + 0.014439954422414303, + 0.0017291840631514788, + 0.00689631886780262, + 0.0042892396450042725, + 0.005196894984692335, + 0.020511319860816002, + 0.00542896194383502, + 0.027049023658037186, + 0.05732542648911476, + 0.005782358348369598, + 0.02228962816298008, + 0.007894671522080898, + 0.004499821458011866, + 0.007373754400759935, + 0.0030224968213588, + 0.013695016503334045, + 0.015802709385752678, + 0.0038166504818946123, + 0.010720482096076012, + 0.007308267056941986, + 0.005617781076580286, + 0.012055174447596073, + 0.004282784182578325, + 0.022771336138248444, + 0.03157290071249008, + 0.006619755178689957, + 0.017773877829313278, + 0.00707022100687027, + 0.002971509238705039, + 0.004769043065607548, + 0.0022355427499860525, + 0.009081786498427391, + 0.009281984530389309, + 0.0028095957823097706, + 0.006769122090190649, + 0.0038401540368795395, + 0.0018376922234892845, + 0.0025584353134036064, + 0.0026005927938967943, + 0.00659093726426363, + 0.004962580278515816, + 0.0016746246255934238, + 0.006250802893191576, + 0.0032581319101154804, + 0.004353296011686325, + 0.010056477971374989, + 0.003949545789510012, + 0.011871565133333206, + 0.02043835073709488, + 0.004585131071507931, + 0.009953954257071018, + 0.0056404354982078075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.006162451580166817 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a pristine white porcelain plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, warm sidelighting that creates mirror-like highlights on the dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dessert exhibit a slow, subtle sway with concentric ripples radiating outward from its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely obscured by a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05503564700484276 + ], + "opposite_logprob_tokens": [ + -2.9300355911254883 + ], + "logprob_reward": -0.05503564700484276, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05503564700484276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.026782872155308723, + 0.026758555322885513, + 0.3443605303764343, + 0.019273240119218826, + 0.1006886437535286, + 0.06377853453159332, + 0.27740079164505005, + 0.12003515660762787, + 0.004158273339271545, + 0.016763396561145782 + ], + "credit_map": [ + 0.0013538755010813475, + 0.0031922475900501013, + 0.0008047468727454543, + 0.0037249287124723196, + 0.009091360494494438, + 0.0015405721496790648, + 0.004219521302729845, + 0.002855621511116624, + 0.0017730663530528545, + 0.0035321873147040606, + 0.0014333392027765512, + 0.006647679954767227, + 0.0032780193723738194, + 0.0014842808013781905, + 0.0056303939782083035, + 0.002979589393362403, + 0.003411300713196397, + 0.011969185434281826, + 0.004242544062435627, + 0.1505340039730072, + 0.019199712201952934, + 0.004154878668487072, + 0.14122259616851807, + 0.009626304730772972, + 0.0008449720335192978, + 0.002525908639654517, + 0.0008416963391937315, + 0.004263175185769796, + 0.005256005097180605, + 0.0010680198902264237, + 0.003390504512935877, + 0.0010829579550772905, + 0.004320312291383743, + 0.014132190495729446, + 0.0032184619922190905, + 0.01587645895779133, + 0.036943912506103516, + 0.005006770137697458, + 0.016115805134177208, + 0.005074727348983288, + 0.003823224687948823, + 0.007414909545332193, + 0.0034701216500252485, + 0.012305405922234058, + 0.018900515511631966, + 0.0027153990231454372, + 0.010240764357149601, + 0.00490818964317441, + 0.014583542011678219, + 0.033917076885700226, + 0.013856161385774612, + 0.04708494246006012, + 0.09350806474685669, + 0.015331107191741467, + 0.04075850546360016, + 0.01836141012609005, + 0.01057013962417841, + 0.018560387194156647, + 0.010889211669564247, + 0.018900353461503983, + 0.017744289711117744, + 0.009742273017764091, + 0.01551734283566475, + 0.01811116561293602, + 0.00026282179169356823, + 0.0007126132841221988, + 0.0002100100100506097, + 0.0012035269755870104, + 0.0006041591987013817, + 0.00018045317847281694, + 0.0007741242297925055, + 0.00021056467085145414, + 0.0016526562394574285, + 0.00274150469340384, + 0.0016028754180297256, + 0.002319557825103402, + 0.0030959509313106537, + 0.001347828540019691, + 0.002256547100841999, + 0.0017464755801483989 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 7 + ], + "selected_reward": -0.05503564700484276 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a pristine white porcelain plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, warm sidelighting that creates mirror-like highlights on the dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dessert exhibit a slow, subtle sway with concentric ripples radiating outward from its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely obscured by a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864796806126833 + ], + "opposite_logprob_tokens": [ + -4.758647918701172 + ], + "logprob_reward": -0.00864796806126833, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864796806126833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12568572163581848, + 0.12062407284975052, + 0.3293801248073578, + 0.0734572634100914, + 0.05257508158683777, + 0.06606335937976837, + 0.13231799006462097, + 0.02769525535404682, + 0.02195119857788086, + 0.05024993047118187 + ], + "credit_map": [ + 0.005611656233668327, + 0.009739095345139503, + 0.006201695650815964, + 0.03183836117386818, + 0.022205788642168045, + 0.007502846419811249, + 0.030037080869078636, + 0.012549201026558876, + 0.006721482612192631, + 0.016366424039006233, + 0.005790420807898045, + 0.019089223816990852, + 0.034700773656368256, + 0.008365526795387268, + 0.01943148858845234, + 0.010158732533454895, + 0.0121059725061059, + 0.03248549625277519, + 0.016727017238736153, + 0.08051913976669312, + 0.056618835777044296, + 0.02746061235666275, + 0.08104092627763748, + 0.02242213860154152, + 0.0031402346212416887, + 0.010461029596626759, + 0.004878859501332045, + 0.012919562868773937, + 0.022624507546424866, + 0.003394829574972391, + 0.011354347690939903, + 0.004683890845626593, + 0.0019864486530423164, + 0.007240324281156063, + 0.0014314647996798158, + 0.009501243941485882, + 0.01942957378923893, + 0.002242106245830655, + 0.007826412096619606, + 0.002917506266385317, + 0.0037610467988997698, + 0.009058275260031223, + 0.002887091599404812, + 0.012812401168048382, + 0.019826527684926987, + 0.003893852001056075, + 0.009769521653652191, + 0.004054640419781208, + 0.007488806266337633, + 0.015776190906763077, + 0.007336304057389498, + 0.022232942283153534, + 0.04352826997637749, + 0.007071526721119881, + 0.01906401850283146, + 0.009819942526519299, + 0.003098311135545373, + 0.00406982097774744, + 0.0017304443754255772, + 0.004111920017749071, + 0.006332889664918184, + 0.0022940647322684526, + 0.004051608964800835, + 0.002006197115406394, + 0.0014965523732826114, + 0.002153996378183365, + 0.0018757780781015754, + 0.004908545408397913, + 0.0031451138202100992, + 0.0016988118877634406, + 0.004003843292593956, + 0.002668556058779359, + 0.0027178518939763308, + 0.00591720687225461, + 0.0037200707010924816, + 0.008441767655313015, + 0.016214145347476006, + 0.0018079603323712945, + 0.00787905603647232, + 0.0035518689546734095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.00864796806126833 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a pristine white porcelain plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, warm sidelighting that creates mirror-like highlights on the dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dessert exhibit a slow, subtle sway with concentric ripples radiating outward from its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely obscured by a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897446304559708 + ], + "opposite_logprob_tokens": [ + -2.578974485397339 + ], + "logprob_reward": -0.07897446304559708, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897446304559708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15055696666240692, + 0.041002340614795685, + 0.6038471460342407, + 0.036074697971343994, + 0.011399363167583942, + 0.022212602198123932, + 0.06898624449968338, + 0.029087914153933525, + 0.008201834745705128, + 0.028630897402763367 + ], + "credit_map": [ + 0.006533653941005468, + 0.012274722568690777, + 0.008728956803679466, + 0.03283069655299187, + 0.02805032767355442, + 0.00792904756963253, + 0.03623499721288681, + 0.017974579706788063, + 0.0019443766213953495, + 0.005160114727914333, + 0.001836850424297154, + 0.008129979483783245, + 0.012256177142262459, + 0.0018961039604619145, + 0.007461152505129576, + 0.0023175901733338833, + 0.010006574913859367, + 0.022232845425605774, + 0.012296302244067192, + 0.2515528202056885, + 0.03855498880147934, + 0.009734940715134144, + 0.23710283637046814, + 0.022365882992744446, + 0.002764136530458927, + 0.005162359680980444, + 0.0031496540177613497, + 0.005738891661167145, + 0.009053913876414299, + 0.0022323825396597385, + 0.0053121414966881275, + 0.002661222591996193, + 0.0008216004353016615, + 0.0017537388484925032, + 0.000928074587136507, + 0.0016814590198919177, + 0.002735816640779376, + 0.0006616968894377351, + 0.001554741756990552, + 0.0012622365029528737, + 0.001392195001244545, + 0.0028166037518531084, + 0.0012239894131198525, + 0.004014582838863134, + 0.006872338708490133, + 0.0010383016197010875, + 0.003483510809019208, + 0.001371082034893334, + 0.003617498790845275, + 0.008425145410001278, + 0.003794885240495205, + 0.011470562778413296, + 0.023498931899666786, + 0.0036388824228197336, + 0.009992003440856934, + 0.00454834196716547, + 0.002323080552741885, + 0.004035793244838715, + 0.0028030690737068653, + 0.0045243604108691216, + 0.005354104097932577, + 0.0021691834554076195, + 0.003710262943059206, + 0.004168064333498478, + 0.0005587746272794902, + 0.0012977616861462593, + 0.0005438061198219657, + 0.0015530333621427417, + 0.0018609219696372747, + 0.00038589839823544025, + 0.0013897165190428495, + 0.0006119232275523245, + 0.0014847861602902412, + 0.003541566664353013, + 0.0009747354779392481, + 0.005876400042325258, + 0.008758889511227608, + 0.0013315080432221293, + 0.004531760234385729, + 0.0021312562748789787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.07897446304559708 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a pristine white porcelain plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, warm sidelighting that creates mirror-like highlights on the dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dessert exhibit a slow, subtle sway with concentric ripples radiating outward from its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely obscured by a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023297274485230446 + ], + "opposite_logprob_tokens": [ + -3.7732973098754883 + ], + "logprob_reward": -0.023297274485230446, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023297274485230446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11853840202093124, + 0.059742510318756104, + 0.18013186752796173, + 0.0507335364818573, + 0.1404392123222351, + 0.08593415468931198, + 0.23081187903881073, + 0.08349578082561493, + 0.022850362583994865, + 0.027322271838784218 + ], + "credit_map": [ + 0.004749209154397249, + 0.012519467622041702, + 0.005366112105548382, + 0.02065764181315899, + 0.031313516199588776, + 0.00846408773213625, + 0.02317178249359131, + 0.012296590022742748, + 0.0027267246041446924, + 0.006604581605643034, + 0.003646479919552803, + 0.010613316670060158, + 0.018000805750489235, + 0.0030122878961265087, + 0.00903273094445467, + 0.006105578970164061, + 0.007412908133119345, + 0.01084562111645937, + 0.009745665825903416, + 0.05718199908733368, + 0.01770435832440853, + 0.006177711300551891, + 0.05432447791099548, + 0.016739120706915855, + 0.0025872925762087107, + 0.007556298281997442, + 0.001802112557925284, + 0.010555355809628963, + 0.01134527288377285, + 0.0030268260743469, + 0.008920906111598015, + 0.004939474165439606, + 0.006116940174251795, + 0.019825465977191925, + 0.0041513158939778805, + 0.022269217297434807, + 0.05171601474285126, + 0.0074777305126190186, + 0.021608995273709297, + 0.007273523136973381, + 0.0056862421333789825, + 0.009964442811906338, + 0.006833021994680166, + 0.014608751982450485, + 0.024013707414269447, + 0.004045125097036362, + 0.013040381483733654, + 0.007742485962808132, + 0.01153985969722271, + 0.027893666177988052, + 0.011923621408641338, + 0.03972962498664856, + 0.07743401825428009, + 0.013451995328068733, + 0.034075308591127396, + 0.014763775281608105, + 0.0069059706293046474, + 0.013214921578764915, + 0.006592465098947287, + 0.013154161162674427, + 0.013208670541644096, + 0.00634642643854022, + 0.010476923547685146, + 0.013596242293715477, + 0.0012076572747901082, + 0.003499017795547843, + 0.0011097568785771728, + 0.004831820260733366, + 0.0057100639678537846, + 0.0008622573805041611, + 0.003900073701515794, + 0.0017297135200351477, + 0.0017337710596621037, + 0.003789326874539256, + 0.0011896779760718346, + 0.004716815426945686, + 0.0083188246935606, + 0.0014896404463797808, + 0.003755697514861822, + 0.002328516449779272 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 4 + ], + "selected_reward": -0.023297274485230446 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extremely close-up, static shot, positioned at a slightly elevated angle, focuses intensely on a shimmering, ruby-red gelatin dessert centered on a pristine white porcelain plate. Bright, warm sidelighting illuminates the scene, casting brilliant, mirror-like highlights on the dessert's smooth, moist surface and penetrating its translucent body, causing it to glow from within like a gem. The dessert possesses a slow, subtle sway, its entire form trembling with a mesmerizing rhythm. This movement emanates from its center in a continuous cascade of concentric ripples that radiate outward across its tense surface, distorting the reflections of light. A shallow depth of field, completely obscuring the background, ensures that every detail of the gelatin's jostling motion and the hypnotic play of light through its form are captured, creating a captivating and satisfyingly tactile atmosphere.", + "seed": 33000, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_7/index_4896_seed_33000_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07408962650224567, + "vlm_logprob": -0.07408962650224567, + "vlm_yes_no_margin": 3.824999988079071, + "vlm_token_credit": -0.07408962650224567, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a pristine white porcelain plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, warm sidelighting that creates mirror-like highlights on the dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dessert exhibit a slow, subtle sway with concentric ripples radiating outward from its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely obscured by a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008887732401490211 + ], + "opposite_logprob_tokens": [ + -5.258887767791748 + ], + "logprob_reward": -0.008887732401490211, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.008887732401490211, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23211218416690826, + 0.11769966036081314, + 0.08286302536725998, + 0.06600338220596313, + 0.06770668923854828, + 0.04223133251070976, + 0.03843572735786438, + 0.10592544823884964, + 0.100196972489357, + 0.14682555198669434 + ], + "credit_map": [ + 0.019774867221713066, + 0.03173571825027466, + 0.022381193935871124, + 0.03790813684463501, + 0.01807158626616001, + 0.015848688781261444, + 0.034479521214962006, + 0.05191248282790184, + 0.01123045478016138, + 0.01804323121905327, + 0.013708318583667278, + 0.01703370176255703, + 0.013668968342244625, + 0.010078744031488895, + 0.018041208386421204, + 0.015895046293735504, + 0.008176886476576328, + 0.019524812698364258, + 0.007904418744146824, + 0.014502473175525665, + 0.006581916008144617, + 0.006928741466253996, + 0.010247299447655678, + 0.008996477350592613, + 0.00608542887493968, + 0.010635346174240112, + 0.006422342266887426, + 0.011397644877433777, + 0.005283684004098177, + 0.0061345635913312435, + 0.009449443779885769, + 0.010594931431114674, + 0.006176007445901632, + 0.009369098581373692, + 0.00393290352076292, + 0.014384893700480461, + 0.004923741798847914, + 0.00855911523103714, + 0.012675573118031025, + 0.007685358170419931, + 0.0052339727990329266, + 0.007589960470795631, + 0.0034911634866148233, + 0.007225003559142351, + 0.004427202045917511, + 0.004777455702424049, + 0.0048781200312078, + 0.00460845697671175, + 0.003554240334779024, + 0.005448310635983944, + 0.002016849583014846, + 0.007975459098815918, + 0.004047425929456949, + 0.0035903912503272295, + 0.008498840034008026, + 0.003304216545075178, + 0.010633324272930622, + 0.024291349574923515, + 0.008977852761745453, + 0.016960399225354195, + 0.008892943151295185, + 0.008685328997671604, + 0.01685572788119316, + 0.01062853541225195, + 0.011454830877482891, + 0.019119197502732277, + 0.006773939356207848, + 0.01756800152361393, + 0.008961685933172703, + 0.007289143279194832, + 0.018443716689944267, + 0.010586456395685673, + 0.014835944399237633, + 0.028909742832183838, + 0.012415511533617973, + 0.023770857602357864, + 0.012101085856556892, + 0.0139574334025383, + 0.02125510200858116, + 0.019579868763685226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.008887732401490211 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a pristine white porcelain plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, warm sidelighting that creates mirror-like highlights on the dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dessert exhibit a slow, subtle sway with concentric ripples radiating outward from its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely obscured by a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3141244053840637 + ], + "opposite_logprob_tokens": [ + -1.314124345779419 + ], + "logprob_reward": -0.3141244053840637, + "yes_no_margin_reward": 0.9999999403953552, + "token_credit_reward": -0.3141244053840637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12879128754138947, + 0.19381417334079742, + 0.07069556415081024, + 0.04258473962545395, + 0.08290013670921326, + 0.061772748827934265, + 0.03136992081999779, + 0.0626041442155838, + 0.048491865396499634, + 0.27697548270225525 + ], + "credit_map": [ + 0.009479054249823093, + 0.01848701201379299, + 0.006439140997827053, + 0.02295561507344246, + 0.011229858733713627, + 0.011026867665350437, + 0.034970737993717194, + 0.014202999882400036, + 0.021979719400405884, + 0.026694929227232933, + 0.018695220351219177, + 0.025118403136730194, + 0.021717581897974014, + 0.02748764120042324, + 0.02843846194446087, + 0.023682206869125366, + 0.009106884710490704, + 0.01224706880748272, + 0.006982980761677027, + 0.009458874352276325, + 0.006098353303968906, + 0.008191543631255627, + 0.0075203245505690575, + 0.011089536361396313, + 0.003707833355292678, + 0.006271469406783581, + 0.004138053860515356, + 0.00709932716563344, + 0.0027524561155587435, + 0.0037227252032607794, + 0.006450489163398743, + 0.008442382328212261, + 0.00677567208185792, + 0.016215849667787552, + 0.007394131738692522, + 0.01331536378711462, + 0.0075012813322246075, + 0.009782349690794945, + 0.012390677817165852, + 0.009524810127913952, + 0.006592826917767525, + 0.011370213702321053, + 0.004966313019394875, + 0.010326779447495937, + 0.004943711217492819, + 0.007781969849020243, + 0.007889402098953724, + 0.007901533506810665, + 0.003110151505097747, + 0.003471581730991602, + 0.0029295599088072777, + 0.005868339911103249, + 0.00226411409676075, + 0.0025702379643917084, + 0.005239082500338554, + 0.005916852504014969, + 0.0070725250989198685, + 0.012657898478209972, + 0.005738791543990374, + 0.010082591325044632, + 0.005019993055611849, + 0.0046539087779819965, + 0.009647390805184841, + 0.007731045596301556, + 0.004805123899132013, + 0.007966872304677963, + 0.003726980881765485, + 0.007288850843906403, + 0.002970891073346138, + 0.004289921838790178, + 0.005782078020274639, + 0.011661143973469734, + 0.02876364253461361, + 0.05543559417128563, + 0.024216949939727783, + 0.042286861687898636, + 0.023029044270515442, + 0.025091664865612984, + 0.03737001493573189, + 0.04078169912099838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.3141244053840637 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a pristine white porcelain plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, warm sidelighting that creates mirror-like highlights on the dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dessert exhibit a slow, subtle sway with concentric ripples radiating outward from its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely obscured by a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005955332424491644 + ], + "opposite_logprob_tokens": [ + -5.130955219268799 + ], + "logprob_reward": -0.005955332424491644, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005955332424491644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2726726830005646, + 0.24952512979507446, + 0.04235175997018814, + 0.06644047796726227, + 0.08505263924598694, + 0.05599048361182213, + 0.03596825897693634, + 0.0764898806810379, + 0.041876763105392456, + 0.07363193482160568 + ], + "credit_map": [ + 0.018190454691648483, + 0.027727946639060974, + 0.014770600944757462, + 0.048085615038871765, + 0.0211574025452137, + 0.024295229464769363, + 0.06984276324510574, + 0.04860266298055649, + 0.02047388255596161, + 0.03488238900899887, + 0.02554696425795555, + 0.03524239733815193, + 0.026011856272816658, + 0.03816813975572586, + 0.04103256016969681, + 0.02816692553460598, + 0.005385946948081255, + 0.009068755432963371, + 0.0027344359550625086, + 0.0058922297321259975, + 0.00430627865716815, + 0.003377934917807579, + 0.006897304207086563, + 0.0046888734214007854, + 0.006923621520400047, + 0.009624456986784935, + 0.0070003666914999485, + 0.01015099324285984, + 0.004890800453722477, + 0.004722066223621368, + 0.0105182696133852, + 0.012609913013875484, + 0.006965100299566984, + 0.012689854018390179, + 0.005962883122265339, + 0.019392795860767365, + 0.0062830098904669285, + 0.009584513492882252, + 0.017274657264351845, + 0.006899824365973473, + 0.0055575501173734665, + 0.008994643576443195, + 0.005147974006831646, + 0.007485303096473217, + 0.005755958613008261, + 0.006665787193924189, + 0.007794319186359644, + 0.008588947355747223, + 0.005103487521409988, + 0.005250389687716961, + 0.0042851087637245655, + 0.006125450599938631, + 0.002671437105163932, + 0.00277382112108171, + 0.004342534113675356, + 0.005416030064225197, + 0.00977273378521204, + 0.015191934071481228, + 0.006201321724802256, + 0.01193238329142332, + 0.0064155845902860165, + 0.0068854037672281265, + 0.010696052573621273, + 0.009394469670951366, + 0.003016431350260973, + 0.005390744656324387, + 0.0043935165740549564, + 0.007275047246366739, + 0.002655447693541646, + 0.004875994753092527, + 0.00669864472001791, + 0.007570939138531685, + 0.007192664314061403, + 0.010603264905512333, + 0.005160022061318159, + 0.01124787237495184, + 0.0045484378933906555, + 0.006092654075473547, + 0.00743852136656642, + 0.0213484950363636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.005955332424491644 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a pristine white porcelain plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, warm sidelighting that creates mirror-like highlights on the dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dessert exhibit a slow, subtle sway with concentric ripples radiating outward from its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely obscured by a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023285046219825745 + ], + "opposite_logprob_tokens": [ + -3.773285150527954 + ], + "logprob_reward": -0.023285046219825745, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023285046219825745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32143914699554443, + 0.15199348330497742, + 0.03821491077542305, + 0.06663362681865692, + 0.0736866295337677, + 0.053385332226753235, + 0.038096651434898376, + 0.1878766268491745, + 0.031114699319005013, + 0.0375588983297348 + ], + "credit_map": [ + 0.019260162487626076, + 0.04455110803246498, + 0.014734056778252125, + 0.05631816387176514, + 0.026206478476524353, + 0.028363049030303955, + 0.07840919494628906, + 0.053596921265125275, + 0.013190420344471931, + 0.029614249244332314, + 0.008053278550505638, + 0.022533509880304337, + 0.013234022073447704, + 0.02327873557806015, + 0.02964721992611885, + 0.012442050501704216, + 0.003956574015319347, + 0.0070038847625255585, + 0.0031738735269755125, + 0.0060913111083209515, + 0.0033249238040298223, + 0.0043943943455815315, + 0.00515559921041131, + 0.005114351399242878, + 0.008520584553480148, + 0.00866625364869833, + 0.005582116078585386, + 0.010420683771371841, + 0.004355148877948523, + 0.00793648324906826, + 0.007908032275736332, + 0.013244325295090675, + 0.008493738248944283, + 0.011218402534723282, + 0.006076375488191843, + 0.010400757193565369, + 0.007282407488673925, + 0.011622156947851181, + 0.011085350066423416, + 0.00750744016841054, + 0.007267612032592297, + 0.009185193106532097, + 0.004292001016438007, + 0.008948015049099922, + 0.005509697832167149, + 0.005435295403003693, + 0.006561970803886652, + 0.006185546051710844, + 0.0037109439726918936, + 0.004767588805407286, + 0.0030308805871754885, + 0.006248232442885637, + 0.004502648487687111, + 0.0029326314106583595, + 0.008694392628967762, + 0.004209334962069988, + 0.024806257337331772, + 0.03897683322429657, + 0.01460507232695818, + 0.03198671713471413, + 0.016238855198025703, + 0.014517838135361671, + 0.026332123205065727, + 0.020412933081388474, + 0.0022742885630577803, + 0.00388150243088603, + 0.0028528987895697355, + 0.00543963210657239, + 0.0019100341014564037, + 0.0033107742201536894, + 0.005315727554261684, + 0.006129839923232794, + 0.0038960508536547422, + 0.004746421705931425, + 0.0022727856412529945, + 0.006907584611326456, + 0.0024099755100905895, + 0.0034804102033376694, + 0.005290738772600889, + 0.008554928936064243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.023285046219825745 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a pristine white porcelain plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, warm sidelighting that creates mirror-like highlights on the dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dessert exhibit a slow, subtle sway with concentric ripples radiating outward from its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely obscured by a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018195616081357002 + ], + "opposite_logprob_tokens": [ + -4.018195629119873 + ], + "logprob_reward": -0.018195616081357002, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018195616081357002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11399227380752563, + 0.12883451581001282, + 0.10887277126312256, + 0.14710277318954468, + 0.09945914894342422, + 0.058888014405965805, + 0.06370119005441666, + 0.026879603043198586, + 0.0969538539648056, + 0.1553158462047577 + ], + "credit_map": [ + 0.015586107969284058, + 0.017654141411185265, + 0.008216291666030884, + 0.017845088616013527, + 0.01289218757301569, + 0.007843253202736378, + 0.015889765694737434, + 0.018065422773361206, + 0.014849647879600525, + 0.02117462269961834, + 0.009771923534572124, + 0.018568286672234535, + 0.017912156879901886, + 0.012853395193815231, + 0.01725071854889393, + 0.016453759744763374, + 0.013108531013131142, + 0.022365868091583252, + 0.009925703518092632, + 0.0160438884049654, + 0.007540008053183556, + 0.009313012473285198, + 0.016927756369113922, + 0.013648000545799732, + 0.013696150854229927, + 0.021258004009723663, + 0.013224652037024498, + 0.02600807696580887, + 0.010798037983477116, + 0.012401758693158627, + 0.018804386258125305, + 0.03091171197593212, + 0.010355319827795029, + 0.014002133160829544, + 0.009312712587416172, + 0.01715296506881714, + 0.00662407698109746, + 0.013883188366889954, + 0.01163350511342287, + 0.016495250165462494, + 0.006684981286525726, + 0.012169783934950829, + 0.004849492106586695, + 0.010112481191754341, + 0.004549891222268343, + 0.006751290522515774, + 0.007436972577124834, + 0.0063331229612231255, + 0.006275408901274204, + 0.008599916473031044, + 0.007798437029123306, + 0.009658140130341053, + 0.006414253730326891, + 0.006180714350193739, + 0.009846009314060211, + 0.008928311057388783, + 0.002974108560010791, + 0.0052078659646213055, + 0.002228837227448821, + 0.004709357861429453, + 0.002457010792568326, + 0.0023643807508051395, + 0.004021029453724623, + 0.0029170115012675524, + 0.009028504602611065, + 0.01596958376467228, + 0.011944864876568317, + 0.01620125211775303, + 0.0072180419228971004, + 0.007961221039295197, + 0.01855916529893875, + 0.010071221739053726, + 0.02096458151936531, + 0.02373335137963295, + 0.0132346972823143, + 0.026675764471292496, + 0.009965267032384872, + 0.013034245930612087, + 0.01619364507496357, + 0.031514305621385574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 1 + ], + "selected_reward": -0.018195616081357002 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extremely close-up, static shot, positioned at a slightly elevated angle, focuses intensely on a shimmering, ruby-red gelatin dessert centered on a pristine white porcelain plate. Bright, warm sidelighting illuminates the scene, casting brilliant, mirror-like highlights on the dessert's smooth, moist surface and penetrating its translucent body, causing it to glow from within like a gem. The dessert possesses a slow, subtle sway, its entire form trembling with a mesmerizing rhythm. This movement emanates from its center in a continuous cascade of concentric ripples that radiate outward across its tense surface, distorting the reflections of light. A shallow depth of field, completely obscuring the background, ensures that every detail of the gelatin's jostling motion and the hypnotic play of light through its form are captured, creating a captivating and satisfyingly tactile atmosphere.", + "seed": 33001, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_7/index_4896_seed_33001_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.11565354075282812, + "vlm_logprob": -0.11565354075282812, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.11565354075282812, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a pristine white porcelain plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, warm sidelighting that creates mirror-like highlights on the dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dessert exhibit a slow, subtle sway with concentric ripples radiating outward from its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely obscured by a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009163112379610538 + ], + "opposite_logprob_tokens": [ + -5.134162902832031 + ], + "logprob_reward": -0.009163112379610538, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.009163112379610538, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029259951785206795, + 0.15919584035873413, + 0.02932409755885601, + 0.1451355516910553, + 0.05297527462244034, + 0.02753848396241665, + 0.20762482285499573, + 0.031378019601106644, + 0.29743441939353943, + 0.020133627578616142 + ], + "credit_map": [ + 0.00189651467371732, + 0.0045168763026595116, + 0.004027911461889744, + 0.004016744904220104, + 0.005127193406224251, + 0.0032277805730700493, + 0.003239785321056843, + 0.0032071454916149378, + 0.01634538359940052, + 0.0078501608222723, + 0.0633259043097496, + 0.039467502385377884, + 0.004937053192406893, + 0.007442011032253504, + 0.011482815258204937, + 0.008344997651875019, + 0.004137049429118633, + 0.005967292003333569, + 0.0028733829967677593, + 0.004813113249838352, + 0.0034116802271455526, + 0.0020125953014940023, + 0.003073514671996236, + 0.003035469213500619, + 0.01313456054776907, + 0.008764119818806648, + 0.05563389137387276, + 0.02473047748208046, + 0.005993093363940716, + 0.0067430599592626095, + 0.010108033195137978, + 0.020028311759233475, + 0.010262285359203815, + 0.006408156827092171, + 0.007176547776907682, + 0.011411893181502819, + 0.003950855229049921, + 0.004206402227282524, + 0.00390984583646059, + 0.005649287719279528, + 0.002409266773611307, + 0.0028553036972880363, + 0.003017391311004758, + 0.009444748982787132, + 0.0014318185858428478, + 0.0016930087003856897, + 0.0027092217933386564, + 0.00397772341966629, + 0.01924753002822399, + 0.01814604550600052, + 0.08070120215415955, + 0.024156352505087852, + 0.011207254603505135, + 0.012287731282413006, + 0.016386281698942184, + 0.02549242600798607, + 0.0017206217162311077, + 0.0070739309303462505, + 0.004813540726900101, + 0.007872547022998333, + 0.0034732052590698004, + 0.0021173576824367046, + 0.002351542469114065, + 0.0019552710000425577, + 0.05206351727247238, + 0.019376292824745178, + 0.13746783137321472, + 0.02823842689394951, + 0.011217664927244186, + 0.010107035748660564, + 0.021073605865240097, + 0.017890028655529022, + 0.002252163365483284, + 0.0017079452518373728, + 0.004567018710076809, + 0.0036512259393930435, + 0.0023364706430584192, + 0.0013321266742423177, + 0.0018040882423520088, + 0.0024825897999107838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 1 + ], + "selected_reward": -0.009163112379610538 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a pristine white porcelain plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, warm sidelighting that creates mirror-like highlights on the dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dessert exhibit a slow, subtle sway with concentric ripples radiating outward from its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely obscured by a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47485166788101196 + ], + "opposite_logprob_tokens": [ + -0.974851667881012 + ], + "logprob_reward": -0.47485166788101196, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47485166788101196, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07690523564815521, + 0.05196494981646538, + 0.08717013895511627, + 0.40436986088752747, + 0.072527214884758, + 0.03368929401040077, + 0.06163371726870537, + 0.032604798674583435, + 0.14869804680347443, + 0.03043675422668457 + ], + "credit_map": [ + 0.009813882410526276, + 0.009780782274901867, + 0.018230538815259933, + 0.01084376685321331, + 0.00893032643944025, + 0.007237871177494526, + 0.005681043490767479, + 0.006387020461261272, + 0.004506320226937532, + 0.0071976929903030396, + 0.013154016807675362, + 0.011551843956112862, + 0.005138779059052467, + 0.0032756582368165255, + 0.004374868236482143, + 0.0027657717000693083, + 0.007765050046145916, + 0.017133554443717003, + 0.010636946186423302, + 0.020865999162197113, + 0.008471708744764328, + 0.005881218705326319, + 0.009014515206217766, + 0.007401145529001951, + 0.034652434289455414, + 0.023073267191648483, + 0.15098334848880768, + 0.0968090072274208, + 0.019142506644129753, + 0.017897136509418488, + 0.025683099403977394, + 0.03612903133034706, + 0.006671710405498743, + 0.012744017876684666, + 0.011923382058739662, + 0.015292635187506676, + 0.006624165456742048, + 0.007491330616176128, + 0.005500417202711105, + 0.00627955561503768, + 0.005132309161126614, + 0.005032933317124844, + 0.005723963491618633, + 0.004484998527914286, + 0.002717704512178898, + 0.0037716892547905445, + 0.0029314770363271236, + 0.0038942177779972553, + 0.00838005542755127, + 0.003999409731477499, + 0.027338413521647453, + 0.007455022539943457, + 0.002967239124700427, + 0.0025337834376841784, + 0.004529532510787249, + 0.004430263303220272, + 0.002951470436528325, + 0.006888591684401035, + 0.003942546434700489, + 0.005158074200153351, + 0.005124673247337341, + 0.0025790813378989697, + 0.0032280103769153357, + 0.002732353750616312, + 0.024462752044200897, + 0.009436829015612602, + 0.06889025121927261, + 0.014172553084790707, + 0.00580105185508728, + 0.005667833611369133, + 0.01076835673302412, + 0.009498422034084797, + 0.005373010411858559, + 0.0023640054278075695, + 0.006340535823255777, + 0.006593083497136831, + 0.003094349056482315, + 0.0017236886778846383, + 0.0026190902572125196, + 0.002328991424292326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 2 + ], + "selected_reward": -0.47485166788101196 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a pristine white porcelain plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, warm sidelighting that creates mirror-like highlights on the dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dessert exhibit a slow, subtle sway with concentric ripples radiating outward from its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely obscured by a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029805263504385948 + ], + "opposite_logprob_tokens": [ + -3.5298051834106445 + ], + "logprob_reward": -0.029805263504385948, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029805263504385948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05080501735210419, + 0.03981193155050278, + 0.046303585171699524, + 0.2769218981266022, + 0.04997309669852257, + 0.09437954425811768, + 0.14660388231277466, + 0.039597850292921066, + 0.22881928086280823, + 0.026783909648656845 + ], + "credit_map": [ + 0.0051148380152881145, + 0.009984617121517658, + 0.005714901722967625, + 0.007234317250549793, + 0.007499784231185913, + 0.0039030066691339016, + 0.007431411184370518, + 0.003922143951058388, + 0.0059639764949679375, + 0.00351624540053308, + 0.014826790429651737, + 0.005253650713711977, + 0.002174317603930831, + 0.0018482414307072759, + 0.0036751963198184967, + 0.0025535156019032, + 0.009284669533371925, + 0.006302274763584137, + 0.006398085039108992, + 0.00552560156211257, + 0.004238278605043888, + 0.004972800146788359, + 0.003818385535851121, + 0.0057634939439594746, + 0.02504895068705082, + 0.016252143308520317, + 0.11231371015310287, + 0.050754349678754807, + 0.010243757627904415, + 0.012166127562522888, + 0.01732131466269493, + 0.03282152861356735, + 0.010087392292916775, + 0.005583981983363628, + 0.00838580634444952, + 0.009438437409698963, + 0.0043108523823320866, + 0.004119192250072956, + 0.0032382584176957607, + 0.004809176083654165, + 0.0070974864065647125, + 0.0040623825043439865, + 0.038726139813661575, + 0.028409719467163086, + 0.00320087349973619, + 0.0029553151689469814, + 0.0031683556735515594, + 0.006759266834706068, + 0.020513368770480156, + 0.013342431746423244, + 0.040152471512556076, + 0.016107546165585518, + 0.009643398225307465, + 0.010685927234590054, + 0.0109635004773736, + 0.025195248425006866, + 0.003208656795322895, + 0.005923504941165447, + 0.004271017387509346, + 0.01572880521416664, + 0.003081390867009759, + 0.0018037104746326804, + 0.0034494451247155666, + 0.002131320536136627, + 0.03801947087049484, + 0.015077408403158188, + 0.10503973066806793, + 0.022519215941429138, + 0.009182472713291645, + 0.008110834285616875, + 0.01635410636663437, + 0.014516040682792664, + 0.006093576084822416, + 0.0021448496263474226, + 0.0050172703340649605, + 0.0045106979086995125, + 0.002513761632144451, + 0.001683056354522705, + 0.0022447085939347744, + 0.002575988881289959 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 6 + ], + "selected_reward": -0.029805263504385948 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a pristine white porcelain plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, warm sidelighting that creates mirror-like highlights on the dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dessert exhibit a slow, subtle sway with concentric ripples radiating outward from its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely obscured by a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809879347681999 + ], + "opposite_logprob_tokens": [ + -3.2880988121032715 + ], + "logprob_reward": -0.03809879347681999, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809879347681999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09542371332645416, + 0.0774797722697258, + 0.039980847388505936, + 0.18409378826618195, + 0.08362708985805511, + 0.12364206463098526, + 0.1667877435684204, + 0.19846142828464508, + 0.014944223687052727, + 0.01555933989584446 + ], + "credit_map": [ + 0.011485246941447258, + 0.015450578182935715, + 0.016804514452815056, + 0.014244256541132927, + 0.011962872929871082, + 0.0092886071652174, + 0.00868466217070818, + 0.0075029730796813965, + 0.009416910819709301, + 0.004845205694437027, + 0.03104514069855213, + 0.014934425242245197, + 0.003944076132029295, + 0.003436689032241702, + 0.005896399728953838, + 0.003960923757404089, + 0.003995348233729601, + 0.007206686772406101, + 0.006156567018479109, + 0.00769078591838479, + 0.004152202978730202, + 0.003026930382475257, + 0.0040986319072544575, + 0.003653695108368993, + 0.01576206088066101, + 0.011269624345004559, + 0.07684634625911713, + 0.03756581246852875, + 0.005211376119405031, + 0.0061773075722157955, + 0.011321509256958961, + 0.01993974670767784, + 0.02074136957526207, + 0.00805708859115839, + 0.011621562764048576, + 0.015076114796102047, + 0.0077108885161578655, + 0.00502000143751502, + 0.0044665345922112465, + 0.010933526791632175, + 0.007542724721133709, + 0.011874646879732609, + 0.04248300567269325, + 0.04053470492362976, + 0.005886869505047798, + 0.004937042482197285, + 0.004787422716617584, + 0.005595650058239698, + 0.015693768858909607, + 0.016950761899352074, + 0.06495799124240875, + 0.019516928121447563, + 0.010261202231049538, + 0.00909253116697073, + 0.01274330448359251, + 0.017571263015270233, + 0.01733129471540451, + 0.017016494646668434, + 0.027564281597733498, + 0.08886509388685226, + 0.009854612872004509, + 0.010984697379171848, + 0.015360444784164429, + 0.011484505608677864, + 0.0017022142419591546, + 0.0012790450127795339, + 0.005381158087402582, + 0.0016699766274541616, + 0.00113564298953861, + 0.0009077878203243017, + 0.0014528072206303477, + 0.0014155928511172533, + 0.0014131542993709445, + 0.0013014290016144514, + 0.003842126578092575, + 0.0029816331807523966, + 0.0015041320584714413, + 0.0009536518482491374, + 0.0015752253821119666, + 0.0019879876635968685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.03809879347681999 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a pristine white porcelain plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, warm sidelighting that creates mirror-like highlights on the dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dessert exhibit a slow, subtle sway with concentric ripples radiating outward from its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely obscured by a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026348866522312164 + ], + "opposite_logprob_tokens": [ + -3.651348829269409 + ], + "logprob_reward": -0.026348866522312164, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026348866522312164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06052205339074135, + 0.07101595401763916, + 0.0473058708012104, + 0.19366617500782013, + 0.05808359757065773, + 0.05337578430771828, + 0.15681204199790955, + 0.09706372767686844, + 0.24146580696105957, + 0.020688949152827263 + ], + "credit_map": [ + 0.008369666524231434, + 0.011390119791030884, + 0.009878171607851982, + 0.008164212107658386, + 0.005595922004431486, + 0.006093386095017195, + 0.004863264504820108, + 0.00616731122136116, + 0.007355893962085247, + 0.007140870206058025, + 0.018287548795342445, + 0.015977509319782257, + 0.005668634083122015, + 0.003996773157268763, + 0.0053519378416240215, + 0.007236788515001535, + 0.004686661995947361, + 0.010020511224865913, + 0.0056502241641283035, + 0.008146545849740505, + 0.005876149982213974, + 0.004452143795788288, + 0.004485791549086571, + 0.0039878422394394875, + 0.01676703430712223, + 0.01225032564252615, + 0.07654091715812683, + 0.04067235440015793, + 0.006195855792611837, + 0.007376509252935648, + 0.011574937030673027, + 0.02228824608027935, + 0.007984776049852371, + 0.006926884409040213, + 0.006583640351891518, + 0.014294571243226528, + 0.005166050978004932, + 0.006530956830829382, + 0.004255848471075296, + 0.00634086923673749, + 0.005832790397107601, + 0.0055244555696845055, + 0.00665886327624321, + 0.01439183484762907, + 0.003252078779041767, + 0.0044170101173222065, + 0.00520574813708663, + 0.008092998526990414, + 0.017056258395314217, + 0.01574760675430298, + 0.05127584934234619, + 0.017967481166124344, + 0.009434747509658337, + 0.010712554678320885, + 0.011986802332103252, + 0.022630756720900536, + 0.007604997605085373, + 0.009014983661472797, + 0.01496700756251812, + 0.04131101816892624, + 0.004680495243519545, + 0.006433023139834404, + 0.006899891421198845, + 0.006152311339974403, + 0.04270439222455025, + 0.015518708154559135, + 0.11173827946186066, + 0.023139281198382378, + 0.008725961670279503, + 0.008077511563897133, + 0.017103374004364014, + 0.014458292163908482, + 0.0020027037244290113, + 0.0015860123094171286, + 0.006185495760291815, + 0.004246023017913103, + 0.0016719052800908685, + 0.0011708003003150225, + 0.0015077544376254082, + 0.0023182551376521587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 6 + ], + "selected_reward": -0.026348866522312164 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extremely close-up, static shot, positioned at a slightly elevated angle, focuses intensely on a shimmering, ruby-red gelatin dessert centered on a pristine white porcelain plate. Bright, warm sidelighting illuminates the scene, casting brilliant, mirror-like highlights on the dessert's smooth, moist surface and penetrating its translucent body, causing it to glow from within like a gem. The dessert possesses a slow, subtle sway, its entire form trembling with a mesmerizing rhythm. This movement emanates from its center in a continuous cascade of concentric ripples that radiate outward across its tense surface, distorting the reflections of light. A shallow depth of field, completely obscuring the background, ensures that every detail of the gelatin's jostling motion and the hypnotic play of light through its form are captured, creating a captivating and satisfyingly tactile atmosphere.", + "seed": 33002, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_7/index_4896_seed_33002_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03818198433145881, + "vlm_logprob": -0.03818198433145881, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.03818198433145881, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a pristine white porcelain plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, warm sidelighting that creates mirror-like highlights on the dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dessert exhibit a slow, subtle sway with concentric ripples radiating outward from its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely obscured by a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0069295321591198444 + ], + "opposite_logprob_tokens": [ + -5.881929397583008 + ], + "logprob_reward": -0.0069295321591198444, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0069295321591198444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13623842597007751, + 0.16526660323143005, + 0.11679983884096146, + 0.03235694766044617, + 0.08898569643497467, + 0.0653093233704567, + 0.13508537411689758, + 0.06957948207855225, + 0.13176345825195312, + 0.0586148276925087 + ], + "credit_map": [ + 0.015428978949785233, + 0.02101036161184311, + 0.011086760088801384, + 0.01696089096367359, + 0.019362634047865868, + 0.014853645116090775, + 0.016176382079720497, + 0.021358786150813103, + 0.013549406081438065, + 0.01922224834561348, + 0.011937406845390797, + 0.0302454624325037, + 0.014246626757085323, + 0.02200053259730339, + 0.02550402283668518, + 0.02856089547276497, + 0.008676271885633469, + 0.014020968228578568, + 0.0077375746332108974, + 0.01479294989258051, + 0.019081205129623413, + 0.024021316319704056, + 0.014614499174058437, + 0.013855061493813992, + 0.003274878254160285, + 0.004474195651710033, + 0.0035421077627688646, + 0.0054857246577739716, + 0.003930909559130669, + 0.003803773783147335, + 0.0035730688832700253, + 0.004272289574146271, + 0.0073331850580871105, + 0.009802176617085934, + 0.007025476545095444, + 0.013403357937932014, + 0.009637164883315563, + 0.013358493335545063, + 0.010941149666905403, + 0.017484696581959724, + 0.006382250692695379, + 0.00817715935409069, + 0.006105733569711447, + 0.009497487917542458, + 0.008611578494310379, + 0.007673349231481552, + 0.009552338160574436, + 0.009309425950050354, + 0.010484394617378712, + 0.019807130098342896, + 0.01333617977797985, + 0.02103332057595253, + 0.015170987695455551, + 0.021083340048789978, + 0.01716383546590805, + 0.017006199806928635, + 0.005332745146006346, + 0.011490837670862675, + 0.007257577031850815, + 0.0110845398157835, + 0.006394845899194479, + 0.010054939426481724, + 0.007380824536085129, + 0.010583178140223026, + 0.009774471633136272, + 0.040016815066337585, + 0.005837940610945225, + 0.015062003396451473, + 0.01467911060899496, + 0.014323936775326729, + 0.013734439387917519, + 0.01833474449813366, + 0.005499847233295441, + 0.006516523193567991, + 0.005876773037016392, + 0.010298115201294422, + 0.007454345002770424, + 0.007384844124317169, + 0.007537777069956064, + 0.008046604692935944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.0069295321591198444 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a pristine white porcelain plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, warm sidelighting that creates mirror-like highlights on the dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dessert exhibit a slow, subtle sway with concentric ripples radiating outward from its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely obscured by a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048717908561229706 + ], + "opposite_logprob_tokens": [ + -3.048717975616455 + ], + "logprob_reward": -0.048717908561229706, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048717908561229706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15149688720703125, + 0.056378573179244995, + 0.09573336690664291, + 0.08157011866569519, + 0.1043609157204628, + 0.22268493473529816, + 0.04122891277074814, + 0.044525016099214554, + 0.0377303883433342, + 0.16429083049297333 + ], + "credit_map": [ + 0.015622733160853386, + 0.02579864114522934, + 0.0152095602825284, + 0.022477779537439346, + 0.017577657476067543, + 0.014001875184476376, + 0.018804218620061874, + 0.022004425525665283, + 0.00508361728861928, + 0.008591300807893276, + 0.005381834227591753, + 0.009294433519244194, + 0.0052971141412854195, + 0.005452351178973913, + 0.008677450940012932, + 0.00860047247260809, + 0.0070298234932124615, + 0.012156298384070396, + 0.00701569439843297, + 0.015393971465528011, + 0.018375620245933533, + 0.017955714836716652, + 0.007826347835361958, + 0.00997989997267723, + 0.0059624603018164635, + 0.01047595962882042, + 0.007789501454681158, + 0.013003144413232803, + 0.0076166968792676926, + 0.01029862742871046, + 0.01146459300071001, + 0.014959143474698067, + 0.007497474085539579, + 0.010352491401135921, + 0.006964388769119978, + 0.012923749163746834, + 0.012375506572425365, + 0.01771458610892296, + 0.012470372952520847, + 0.024062346667051315, + 0.028969069942831993, + 0.027964968234300613, + 0.024206239730119705, + 0.029153380542993546, + 0.02497238479554653, + 0.021014221012592316, + 0.02926400676369667, + 0.03714066371321678, + 0.0024690856225788593, + 0.008206699974834919, + 0.003665597876533866, + 0.005610023625195026, + 0.00444681104272604, + 0.006396726705133915, + 0.004765958990901709, + 0.005668009165674448, + 0.003112050238996744, + 0.005897340830415487, + 0.0036265715025365353, + 0.00719417305663228, + 0.006356413941830397, + 0.0067456066608428955, + 0.005502544809132814, + 0.006090318318456411, + 0.0029275515116751194, + 0.011671393178403378, + 0.001771029899828136, + 0.004300507716834545, + 0.004005052614957094, + 0.003967151045799255, + 0.003768846858292818, + 0.005318856798112392, + 0.017665492370724678, + 0.022974904626607895, + 0.019069522619247437, + 0.024881809949874878, + 0.02215423434972763, + 0.017228534445166588, + 0.02075737714767456, + 0.019558951258659363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.048717908561229706 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a pristine white porcelain plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, warm sidelighting that creates mirror-like highlights on the dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dessert exhibit a slow, subtle sway with concentric ripples radiating outward from its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely obscured by a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979321263730526 + ], + "opposite_logprob_tokens": [ + -4.634793281555176 + ], + "logprob_reward": -0.00979321263730526, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979321263730526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07721948623657227, + 0.07784108817577362, + 0.2605287432670593, + 0.14163658022880554, + 0.14544916152954102, + 0.14660266041755676, + 0.037743788212537766, + 0.018706869333982468, + 0.03976377472281456, + 0.05450783669948578 + ], + "credit_map": [ + 0.009137621149420738, + 0.009644799865782261, + 0.006631361320614815, + 0.011090944521129131, + 0.008108257316052914, + 0.007520007435232401, + 0.014628546312451363, + 0.010457946918904781, + 0.007681204006075859, + 0.010103806853294373, + 0.009111794643104076, + 0.012997271493077278, + 0.00888331700116396, + 0.006105725187808275, + 0.008986076340079308, + 0.013971889391541481, + 0.017451154068112373, + 0.031141864135861397, + 0.016936369240283966, + 0.03492221608757973, + 0.040386635810136795, + 0.058504458516836166, + 0.028655922040343285, + 0.03253011032938957, + 0.013437096029520035, + 0.021135158836841583, + 0.0142583716660738, + 0.022308629006147385, + 0.012900886125862598, + 0.01869402453303337, + 0.015903444960713387, + 0.02299896627664566, + 0.01290107611566782, + 0.015203318558633327, + 0.009276122786104679, + 0.019876763224601746, + 0.01247395109385252, + 0.021583067253232002, + 0.01849200949072838, + 0.0356428399682045, + 0.019217316061258316, + 0.01553951296955347, + 0.015894809737801552, + 0.021402472630143166, + 0.015496873296797276, + 0.012512633576989174, + 0.01706068217754364, + 0.02947835810482502, + 0.0026036978233605623, + 0.00790383480489254, + 0.0033831645268946886, + 0.0057947165332734585, + 0.004431435838341713, + 0.00496771140024066, + 0.004448884632438421, + 0.004210338927805424, + 0.0018092691898345947, + 0.0030684364028275013, + 0.0021017969120293856, + 0.0022297713439911604, + 0.002683330327272415, + 0.002387054031714797, + 0.0019255856750532985, + 0.0025016244035214186, + 0.0034598822239786386, + 0.0073065925389528275, + 0.0029140133410692215, + 0.005286009516566992, + 0.0038579879328608513, + 0.0046617318876087666, + 0.004872889723628759, + 0.007404666393995285, + 0.004882663954049349, + 0.007419451139867306, + 0.005235837772488594, + 0.009054789319634438, + 0.008028226904571056, + 0.0064941346645355225, + 0.0055151148699223995, + 0.007877619005739689 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 4 + ], + "selected_reward": -0.00979321263730526 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a pristine white porcelain plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, warm sidelighting that creates mirror-like highlights on the dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dessert exhibit a slow, subtle sway with concentric ripples radiating outward from its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely obscured by a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292532086372375 + ], + "opposite_logprob_tokens": [ + -2.2379252910614014 + ], + "logprob_reward": -0.11292532086372375, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292532086372375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0709051638841629, + 0.1515849083662033, + 0.2542581260204315, + 0.049045104533433914, + 0.09231267869472504, + 0.06325388699769974, + 0.038838017731904984, + 0.028042830526828766, + 0.2208772897720337, + 0.030881965532898903 + ], + "credit_map": [ + 0.00807876605540514, + 0.008166817016899586, + 0.00563075440004468, + 0.009831077419221401, + 0.00826208759099245, + 0.009232668206095695, + 0.008523623459041119, + 0.013179373927414417, + 0.015235734172165394, + 0.03274823725223541, + 0.017038583755493164, + 0.019638346508145332, + 0.014833256602287292, + 0.011291594244539738, + 0.018596315756440163, + 0.022202853113412857, + 0.01880287379026413, + 0.03096529096364975, + 0.01633627898991108, + 0.03376426547765732, + 0.034992631524801254, + 0.05543562397360802, + 0.03160090371966362, + 0.032360274344682693, + 0.006337433122098446, + 0.005749829579144716, + 0.0072589754126966, + 0.006904419511556625, + 0.005430049262940884, + 0.005102501716464758, + 0.007043029647320509, + 0.005218872334808111, + 0.006448307074606419, + 0.009572116658091545, + 0.0063599878922104836, + 0.010784261859953403, + 0.011689926497638226, + 0.016051404178142548, + 0.01064263004809618, + 0.020764052867889404, + 0.007090339437127113, + 0.007865824736654758, + 0.0068341647274792194, + 0.010840412229299545, + 0.005674668122082949, + 0.005462102126330137, + 0.008693593554198742, + 0.010792789980769157, + 0.003206677967682481, + 0.006583111826330423, + 0.004715329967439175, + 0.0055840471759438515, + 0.004681512713432312, + 0.005571234505623579, + 0.0048554642125964165, + 0.0036406435538083315, + 0.003344199853017926, + 0.0028144116513431072, + 0.003521140431985259, + 0.0032559907995164394, + 0.00308573292568326, + 0.003009598469361663, + 0.006051057484000921, + 0.0029607019387185574, + 0.01355056744068861, + 0.07152074575424194, + 0.009782553650438786, + 0.021936364471912384, + 0.030988771468400955, + 0.01768057979643345, + 0.024285253137350082, + 0.031132467091083527, + 0.0022960288915783167, + 0.0037945685908198357, + 0.0028398281428962946, + 0.005527525674551725, + 0.003229983616620302, + 0.0049631018191576, + 0.0031366629991680384, + 0.005094272084534168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 1 + ], + "selected_reward": -0.11292532086372375 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a pristine white porcelain plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, warm sidelighting that creates mirror-like highlights on the dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dessert exhibit a slow, subtle sway with concentric ripples radiating outward from its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely obscured by a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254394743591547 + ], + "opposite_logprob_tokens": [ + -4.38754415512085 + ], + "logprob_reward": -0.01254394743591547, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254394743591547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13613945245742798, + 0.14029648900032043, + 0.05976466089487076, + 0.08100728690624237, + 0.16692602634429932, + 0.2460402399301529, + 0.0623447522521019, + 0.04549260810017586, + 0.03392418101429939, + 0.028064224869012833 + ], + "credit_map": [ + 0.013378164730966091, + 0.014585396274924278, + 0.010838637128472328, + 0.01934240572154522, + 0.0129541652277112, + 0.015881381928920746, + 0.021111849695444107, + 0.028047455474734306, + 0.014866873621940613, + 0.016466910019516945, + 0.013018762692809105, + 0.023925120010972023, + 0.013606902211904526, + 0.016339097172021866, + 0.020265748724341393, + 0.02180708944797516, + 0.0043359482660889626, + 0.008050009608268738, + 0.0039047931786626577, + 0.008251884020864964, + 0.015161948278546333, + 0.0062648276798427105, + 0.006176031660288572, + 0.007619230076670647, + 0.009787642396986485, + 0.011103271506726742, + 0.009353497996926308, + 0.013329684734344482, + 0.008499975316226482, + 0.008723207749426365, + 0.009717442095279694, + 0.010492565110325813, + 0.011751308105885983, + 0.015341119840741158, + 0.010084237903356552, + 0.017944589257240295, + 0.02091057412326336, + 0.028802523389458656, + 0.023918116465210915, + 0.03817357122898102, + 0.02451266534626484, + 0.025660838931798935, + 0.021902767941355705, + 0.046935614198446274, + 0.021730557084083557, + 0.02303233928978443, + 0.03338029608130455, + 0.04888519272208214, + 0.004392407834529877, + 0.011127332225441933, + 0.006152771413326263, + 0.008091249503195286, + 0.00653299642726779, + 0.009730814956128597, + 0.007631734013557434, + 0.00868545938283205, + 0.005081623326987028, + 0.005892358720302582, + 0.0062002926133573055, + 0.00632561556994915, + 0.004028446972370148, + 0.005830894224345684, + 0.0063190460205078125, + 0.005814331583678722, + 0.0028022904880344868, + 0.006436346564441919, + 0.0021246764808893204, + 0.005124611314386129, + 0.003207132685929537, + 0.00301765208132565, + 0.0047124107368290424, + 0.006499063689261675, + 0.0024140120949596167, + 0.003320306306704879, + 0.0027308037970215082, + 0.004169575870037079, + 0.0036181851755827665, + 0.004358846228569746, + 0.0032882827799767256, + 0.00416421378031373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 1 + ], + "selected_reward": -0.01254394743591547 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_8/index_tensor([6449, 6449])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_8/index_tensor([6449, 6449])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1ab6259127c6b9c3a45ed5083d66c9f80dd6f75e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_8/index_tensor([6449, 6449])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, extreme macro shot, taken at a slightly elevated angle, captures the deliberate journey of a common garden snail across a vibrant, emerald green leaf. Soft, diffused morning light filters through the invisible canopy, illuminating the scene with a gentle glow. The snail's shell is a delicate spiral of amber and beige hues, displaying fine concentric growth lines on its polished surface. Its soft, translucent body, a pale, fleshy gray, shimmers as it extends its two upper tentacles, each tipped with a tiny black eye that cautiously senses the path ahead. The creature moves in an almost imperceptible, rhythmic wave, steadily skimming across the leaf's textured surface, where the prominent veins form a delicate network of branches. In its wake, it leaves a thick, sticky trail of silvery mucus that catches the light and refracts it into a stunning rainbow of iridescent colors that shimmers with every subtle shift in the light. The shallow depth of field blurs the background into a soft bokeh of greens and browns, creating an atmosphere of tranquility, contemplation, and wonder, focusing entirely on this miniature, thoughtful journey.", + "seed": 21309, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_8/index_6449_seed_21309_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03701221626251936, + "vlm_logprob": -0.03701221626251936, + "vlm_yes_no_margin": 4.249999976158142, + "vlm_token_credit": -0.03701221626251936, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of a common garden snail on a vibrant emerald green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's shell a delicate spiral of amber and beige hues with fine concentric growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snail move by extending its two upper tentacles tipped with tiny black eyes while leaving a thick, sticky trail of silvery mucus that refracts light into rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft bokeh of greens and browns due to a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009367087855935097 + ], + "opposite_logprob_tokens": [ + -5.259366989135742 + ], + "logprob_reward": -0.009367087855935097, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.009367087855935097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21281056106090546, + 0.16205641627311707, + 0.10442986339330673, + 0.10994640737771988, + 0.06143338233232498, + 0.10869396477937698, + 0.05801355466246605, + 0.05402900278568268, + 0.06406480818986893, + 0.06452209502458572 + ], + "credit_map": [ + 0.013669749721884727, + 0.017683204263448715, + 0.018427997827529907, + 0.022137321531772614, + 0.011943398974835873, + 0.051898546516895294, + 0.045967619866132736, + 0.03108268603682518, + 0.014788826927542686, + 0.012193187139928341, + 0.011267938651144505, + 0.02558663859963417, + 0.009942236356437206, + 0.051210034638643265, + 0.015869857743382454, + 0.021197672933340073, + 0.015060887672007084, + 0.015249067917466164, + 0.010906238108873367, + 0.008753928355872631, + 0.011326620355248451, + 0.016707753762602806, + 0.019028348848223686, + 0.0073970104567706585, + 0.012740498408675194, + 0.01295781321823597, + 0.013346254825592041, + 0.01368612702935934, + 0.011064760386943817, + 0.02224479801952839, + 0.011062978766858578, + 0.012843163684010506, + 0.011042498983442783, + 0.007895260117948055, + 0.008739477954804897, + 0.007708506193011999, + 0.004359889775514603, + 0.007400640286505222, + 0.007533975411206484, + 0.006753126159310341, + 0.010782154276967049, + 0.0123597402125597, + 0.01152682863175869, + 0.009964318946003914, + 0.007632408291101456, + 0.012759838253259659, + 0.023404773324728012, + 0.02026388980448246, + 0.008861586451530457, + 0.008876382373273373, + 0.008884563110768795, + 0.0074846516363322735, + 0.003662245348095894, + 0.00609175581485033, + 0.007807991001754999, + 0.006344372406601906, + 0.00888805091381073, + 0.008292685262858868, + 0.007299702148884535, + 0.008789612911641598, + 0.004998392891138792, + 0.0041052065789699554, + 0.00599702587351203, + 0.005658320616930723, + 0.011737735942006111, + 0.009552400559186935, + 0.007706181146204472, + 0.008304695598781109, + 0.006292644888162613, + 0.005933547392487526, + 0.006704202387481928, + 0.007833389565348625, + 0.007249706890434027, + 0.006393164396286011, + 0.007571202237159014, + 0.009972047992050648, + 0.005174504593014717, + 0.007988751865923405, + 0.012099502608180046, + 0.008073209784924984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.009367087855935097 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of a common garden snail on a vibrant emerald green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's shell a delicate spiral of amber and beige hues with fine concentric growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snail move by extending its two upper tentacles tipped with tiny black eyes while leaving a thick, sticky trail of silvery mucus that refracts light into rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft bokeh of greens and browns due to a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076462216675281525 + ], + "opposite_logprob_tokens": [ + -4.882646083831787 + ], + "logprob_reward": -0.0076462216675281525, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076462216675281525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16564008593559265, + 0.10468845069408417, + 0.12365739047527313, + 0.049402181059122086, + 0.12370696663856506, + 0.13770438730716705, + 0.10767887532711029, + 0.07434111833572388, + 0.058238957077264786, + 0.05494154989719391 + ], + "credit_map": [ + 0.013700716197490692, + 0.011223042383790016, + 0.03153495490550995, + 0.01925607956945896, + 0.011754961684346199, + 0.03058326244354248, + 0.034698665142059326, + 0.01288842037320137, + 0.009604694321751595, + 0.008442037738859653, + 0.009100035764276981, + 0.015567618422210217, + 0.008219355717301369, + 0.028123116120696068, + 0.012825843878090382, + 0.01280574593693018, + 0.017965972423553467, + 0.01761201024055481, + 0.013324436731636524, + 0.012043831869959831, + 0.013766990974545479, + 0.02089003287255764, + 0.018932152539491653, + 0.0091219712048769, + 0.005453521851450205, + 0.004743907134979963, + 0.005657528527081013, + 0.006045354530215263, + 0.004077286459505558, + 0.011776143684983253, + 0.004675094038248062, + 0.006973346695303917, + 0.021294858306646347, + 0.01732768304646015, + 0.013058752752840519, + 0.018328772857785225, + 0.008315978571772575, + 0.01137422677129507, + 0.016773752868175507, + 0.017232945188879967, + 0.019017191603779793, + 0.020714223384857178, + 0.016255347058176994, + 0.0158805213868618, + 0.011528899893164635, + 0.015338950790464878, + 0.021347099915146828, + 0.01762216165661812, + 0.02433129958808422, + 0.020877806469798088, + 0.013424503616988659, + 0.011382135562598705, + 0.007189316675066948, + 0.011483118869364262, + 0.00807513203471899, + 0.010915568098425865, + 0.008839172311127186, + 0.014131452888250351, + 0.010415715165436268, + 0.01489517092704773, + 0.0051284548826515675, + 0.00573834078386426, + 0.007587483152747154, + 0.007605333346873522, + 0.007980458438396454, + 0.0077900877222418785, + 0.005832756869494915, + 0.007300438825041056, + 0.00581041956320405, + 0.00540629867464304, + 0.007695571053773165, + 0.010422932915389538, + 0.006538897752761841, + 0.005334844812750816, + 0.007137476932257414, + 0.00909667182713747, + 0.004979485180228949, + 0.00702669657766819, + 0.006501379422843456, + 0.008326099254190922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.0076462216675281525 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of a common garden snail on a vibrant emerald green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's shell a delicate spiral of amber and beige hues with fine concentric growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snail move by extending its two upper tentacles tipped with tiny black eyes while leaving a thick, sticky trail of silvery mucus that refracts light into rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft bokeh of greens and browns due to a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14282402396202087 + ], + "opposite_logprob_tokens": [ + -2.0178239345550537 + ], + "logprob_reward": -0.14282402396202087, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14282402396202087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16186603903770447, + 0.2545803189277649, + 0.06214828044176102, + 0.05330028012394905, + 0.09001907706260681, + 0.0831260085105896, + 0.07125918567180634, + 0.08361466228961945, + 0.07142682373523712, + 0.06865932047367096 + ], + "credit_map": [ + 0.013608860783278942, + 0.018616782501339912, + 0.0179930217564106, + 0.021374667063355446, + 0.01262360904365778, + 0.03334805741906166, + 0.026887616142630577, + 0.01741342805325985, + 0.011651357635855675, + 0.010562123730778694, + 0.020721444860100746, + 0.01632050797343254, + 0.010180470533668995, + 0.09649652242660522, + 0.025084028020501137, + 0.06356385350227356, + 0.006314482074230909, + 0.005955338478088379, + 0.00884699821472168, + 0.007501959800720215, + 0.005975248757749796, + 0.011767568066716194, + 0.009738045744597912, + 0.006048636510968208, + 0.004506152123212814, + 0.007105207070708275, + 0.007908131927251816, + 0.013422663323581219, + 0.0028416782151907682, + 0.009217556565999985, + 0.004267225041985512, + 0.0040316651575267315, + 0.012622405774891376, + 0.012911878526210785, + 0.010460009798407555, + 0.009645294398069382, + 0.006658134516328573, + 0.010696282610297203, + 0.013703727163374424, + 0.013321353122591972, + 0.00968631636351347, + 0.011073010973632336, + 0.01115969754755497, + 0.010472944937646389, + 0.00718646077439189, + 0.007815951481461525, + 0.015885280445218086, + 0.009846349246799946, + 0.010978282429277897, + 0.01098883617669344, + 0.009940506890416145, + 0.009612422436475754, + 0.006133751478046179, + 0.007156409788876772, + 0.009187628515064716, + 0.007261347025632858, + 0.011369701474905014, + 0.012059750966727734, + 0.01446663960814476, + 0.016742007806897163, + 0.006955444347113371, + 0.006258086301386356, + 0.007683309726417065, + 0.00807972252368927, + 0.013743453659117222, + 0.011667522601783276, + 0.006829400081187487, + 0.007670049089938402, + 0.007284917868673801, + 0.006598383653908968, + 0.00741153908893466, + 0.010221562348306179, + 0.006084647495299578, + 0.00605248985812068, + 0.009604321792721748, + 0.01318326499313116, + 0.004818226210772991, + 0.011406531557440758, + 0.0071189189329743385, + 0.010390917770564556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.14282402396202087 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of a common garden snail on a vibrant emerald green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's shell a delicate spiral of amber and beige hues with fine concentric growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snail move by extending its two upper tentacles tipped with tiny black eyes while leaving a thick, sticky trail of silvery mucus that refracts light into rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft bokeh of greens and browns due to a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00463933777064085 + ], + "opposite_logprob_tokens": [ + -5.379639148712158 + ], + "logprob_reward": -0.00463933777064085, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00463933777064085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0960640087723732, + 0.11512492597103119, + 0.04993240907788277, + 0.06045079603791237, + 0.046148210763931274, + 0.34859442710876465, + 0.05320918932557106, + 0.07964573800563812, + 0.019596436992287636, + 0.1312338262796402 + ], + "credit_map": [ + 0.005524022039026022, + 0.006134295370429754, + 0.016489865258336067, + 0.010551853105425835, + 0.004485155921429396, + 0.023420095443725586, + 0.014711680822074413, + 0.014747034758329391, + 0.004749756772071123, + 0.004013364668935537, + 0.008949090726673603, + 0.008789397776126862, + 0.00451561389490962, + 0.0451558381319046, + 0.011655375361442566, + 0.027296487241983414, + 0.003688918426632881, + 0.002934236777946353, + 0.004886106122285128, + 0.005321530159562826, + 0.004848862532526255, + 0.011305383406579494, + 0.012002373114228249, + 0.004945000633597374, + 0.0039097610861063, + 0.00331727578304708, + 0.00628461129963398, + 0.0118822380900383, + 0.003563488833606243, + 0.0180363729596138, + 0.005902607459574938, + 0.007554438430815935, + 0.006266945973038673, + 0.004499513655900955, + 0.005389280617237091, + 0.0048849014565348625, + 0.004965240601450205, + 0.004768659360706806, + 0.01015183050185442, + 0.005221841391175985, + 0.04330261051654816, + 0.04606739431619644, + 0.03393292427062988, + 0.03153926879167557, + 0.02993582934141159, + 0.039210256189107895, + 0.06632251292467117, + 0.05828361585736275, + 0.010059690102934837, + 0.007467776536941528, + 0.005331404507160187, + 0.004993101116269827, + 0.003659584326669574, + 0.007751592434942722, + 0.006159025710076094, + 0.00778701389208436, + 0.015459870919585228, + 0.015014944598078728, + 0.009991836734116077, + 0.008737456053495407, + 0.008062467910349369, + 0.006060374900698662, + 0.007918138056993484, + 0.008400644175708294, + 0.0024818626698106527, + 0.0027281297370791435, + 0.0025349785573780537, + 0.002481298055499792, + 0.001635734224691987, + 0.0015676391776651144, + 0.0027309090364724398, + 0.00343588599935174, + 0.010673921555280685, + 0.010005030781030655, + 0.017862914130091667, + 0.027729758992791176, + 0.007055114954710007, + 0.016121603548526764, + 0.017728246748447418, + 0.02405724674463272 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 1 + ], + "selected_reward": -0.00463933777064085 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of a common garden snail on a vibrant emerald green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's shell a delicate spiral of amber and beige hues with fine concentric growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snail move by extending its two upper tentacles tipped with tiny black eyes while leaving a thick, sticky trail of silvery mucus that refracts light into rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft bokeh of greens and browns due to a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020584410056471825 + ], + "opposite_logprob_tokens": [ + -3.8955843448638916 + ], + "logprob_reward": -0.020584410056471825, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020584410056471825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2031480073928833, + 0.058947913348674774, + 0.07490139454603195, + 0.024349231272935867, + 0.04368818178772926, + 0.04333114996552467, + 0.06166508048772812, + 0.057747259736061096, + 0.035034894943237305, + 0.3971869647502899 + ], + "credit_map": [ + 0.012029512785375118, + 0.014685004949569702, + 0.03990316018462181, + 0.03588387370109558, + 0.01224020030349493, + 0.024032915011048317, + 0.04571590572595596, + 0.018657434731721878, + 0.004518166184425354, + 0.006162199657410383, + 0.006859573069959879, + 0.005833177361637354, + 0.00429215282201767, + 0.01201117504388094, + 0.007786176633089781, + 0.011485293507575989, + 0.014533436857163906, + 0.011734812520444393, + 0.009111423045396805, + 0.007075523026287556, + 0.007518360391259193, + 0.009302329272031784, + 0.008479471318423748, + 0.007146027870476246, + 0.0020342194475233555, + 0.0019146085251122713, + 0.002967903623357415, + 0.005399119574576616, + 0.0017250359524041414, + 0.006026370916515589, + 0.002298751613125205, + 0.001983222784474492, + 0.006071926094591618, + 0.005790404509752989, + 0.0047672041691839695, + 0.0053125061094760895, + 0.00315736117772758, + 0.004355029668658972, + 0.0068727475591003895, + 0.007361002266407013, + 0.0064482660964131355, + 0.004514424130320549, + 0.003900524228811264, + 0.004371352028101683, + 0.0035566838923841715, + 0.004731064662337303, + 0.005460802465677261, + 0.010348032228648663, + 0.011506138369441032, + 0.009372295811772346, + 0.006038956809788942, + 0.006649355869740248, + 0.004057224839925766, + 0.008757717907428741, + 0.006796198897063732, + 0.008487186394631863, + 0.008593009784817696, + 0.009554107673466206, + 0.00826851837337017, + 0.006227048579603434, + 0.005037063267081976, + 0.0047471169382333755, + 0.007092275656759739, + 0.008228122256696224, + 0.005888023879379034, + 0.003712323261424899, + 0.003975250292569399, + 0.004247691947966814, + 0.0034904400818049908, + 0.0034342606086283922, + 0.004801776260137558, + 0.005485126748681068, + 0.033415786921978, + 0.029525963589549065, + 0.05231262370944023, + 0.08414816111326218, + 0.02300160750746727, + 0.04826585575938225, + 0.05263005942106247, + 0.07388689368963242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.020584410056471825 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme macro shot, taken at a slightly elevated angle, captures the deliberate journey of a common garden snail across a vibrant, emerald green leaf. Soft, diffused morning light filters through the invisible canopy, illuminating the scene with a gentle glow. The snail's shell is a delicate spiral of amber and beige hues, displaying fine concentric growth lines on its polished surface. Its soft, translucent body, a pale, fleshy gray, shimmers as it extends its two upper tentacles, each tipped with a tiny black eye that cautiously senses the path ahead. The creature moves in an almost imperceptible, rhythmic wave, steadily skimming across the leaf's textured surface, where the prominent veins form a delicate network of branches. In its wake, it leaves a thick, sticky trail of silvery mucus that catches the light and refracts it into a stunning rainbow of iridescent colors that shimmers with every subtle shift in the light. The shallow depth of field blurs the background into a soft bokeh of greens and browns, creating an atmosphere of tranquility, contemplation, and wonder, focusing entirely on this miniature, thoughtful journey.", + "seed": 21310, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_8/index_6449_seed_21310_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.048091234266757966, + "vlm_logprob": -0.048091234266757966, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.048091234266757966, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of a common garden snail on a vibrant emerald green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's shell a delicate spiral of amber and beige hues with fine concentric growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snail move by extending its two upper tentacles tipped with tiny black eyes while leaving a thick, sticky trail of silvery mucus that refracts light into rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft bokeh of greens and browns due to a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012296108528971672 + ], + "opposite_logprob_tokens": [ + -4.887296199798584 + ], + "logprob_reward": -0.012296108528971672, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.012296108528971672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17568451166152954, + 0.08825922757387161, + 0.050499267876148224, + 0.23205126821994781, + 0.043653834611177444, + 0.07618534564971924, + 0.04379105940461159, + 0.04943494871258736, + 0.1422010362148285, + 0.09823951125144958 + ], + "credit_map": [ + 0.015737267211079597, + 0.023196879774332047, + 0.02012030780315399, + 0.025317832827568054, + 0.02092796564102173, + 0.020869173109531403, + 0.02807626500725746, + 0.02143881283700466, + 0.008143946528434753, + 0.011521890759468079, + 0.01115774642676115, + 0.010460572317242622, + 0.011672837659716606, + 0.01163126528263092, + 0.012374242767691612, + 0.011296719312667847, + 0.005290528759360313, + 0.006414263043552637, + 0.007389755919575691, + 0.004735737107694149, + 0.006030707154422998, + 0.0047537493519485, + 0.008844006806612015, + 0.007040518801659346, + 0.013563987798988819, + 0.016542576253414154, + 0.011935626156628132, + 0.012513156048953533, + 0.025713348761200905, + 0.0771244466304779, + 0.018238697201013565, + 0.056419409811496735, + 0.004936535842716694, + 0.006034294608980417, + 0.004831543192267418, + 0.0042474837973713875, + 0.005215393379330635, + 0.0070806085132062435, + 0.006720974575728178, + 0.0045869979076087475, + 0.008249777369201183, + 0.009568559005856514, + 0.009759679436683655, + 0.008529688231647015, + 0.011969045735895634, + 0.010876971296966076, + 0.010142910294234753, + 0.007088709156960249, + 0.004087036941200495, + 0.0054157390259206295, + 0.004563953261822462, + 0.005165388807654381, + 0.006992429029196501, + 0.006255268584936857, + 0.005206695757806301, + 0.00610454473644495, + 0.005735799670219421, + 0.005601289216428995, + 0.004555540159344673, + 0.005175446160137653, + 0.00815427303314209, + 0.005928859580308199, + 0.008349712938070297, + 0.005934026557952166, + 0.014482006430625916, + 0.013850351795554161, + 0.02011384814977646, + 0.016433313488960266, + 0.01641162857413292, + 0.018779868260025978, + 0.019367726519703865, + 0.02276228368282318, + 0.0075068725273013115, + 0.010067019611597061, + 0.0075204698368906975, + 0.008119001053273678, + 0.013379340060055256, + 0.019508779048919678, + 0.014173844829201698, + 0.01796417310833931 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.012296108528971672 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of a common garden snail on a vibrant emerald green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's shell a delicate spiral of amber and beige hues with fine concentric growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snail move by extending its two upper tentacles tipped with tiny black eyes while leaving a thick, sticky trail of silvery mucus that refracts light into rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft bokeh of greens and browns due to a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014226209372282028 + ], + "opposite_logprob_tokens": [ + -4.26422643661499 + ], + "logprob_reward": -0.014226209372282028, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014226209372282028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10365031659603119, + 0.05606662854552269, + 0.05191318690776825, + 0.48893600702285767, + 0.036114685237407684, + 0.023288173601031303, + 0.04281197488307953, + 0.05378922447562218, + 0.08863871544599533, + 0.05479109659790993 + ], + "credit_map": [ + 0.011342967860400677, + 0.01180250197649002, + 0.01317704375833273, + 0.0146675705909729, + 0.010938679799437523, + 0.012626883573830128, + 0.015450169332325459, + 0.013644497841596603, + 0.005006357561796904, + 0.008631208911538124, + 0.007627640850841999, + 0.007287648506462574, + 0.007329157553613186, + 0.0072138430550694466, + 0.006961519364267588, + 0.006009252741932869, + 0.004159676376730204, + 0.006519815884530544, + 0.007293663453310728, + 0.005843537859618664, + 0.006523784715682268, + 0.004723773803561926, + 0.011155912652611732, + 0.005693021230399609, + 0.01611456647515297, + 0.01442432589828968, + 0.017152853310108185, + 0.012981128878891468, + 0.049411624670028687, + 0.19942320883274078, + 0.022463174536824226, + 0.15696515142917633, + 0.0030029057525098324, + 0.003542529186233878, + 0.0032586015295237303, + 0.0028567134868353605, + 0.004912372678518295, + 0.00824466347694397, + 0.005460894666612148, + 0.004836004227399826, + 0.0019887269008904696, + 0.0026054198388010263, + 0.0028233258053660393, + 0.0026220313739031553, + 0.0031464642379432917, + 0.003626155201345682, + 0.003547336207702756, + 0.0029287156648933887, + 0.0038353095296770334, + 0.0041525703854858875, + 0.003451199270784855, + 0.0034295651130378246, + 0.009920408949255943, + 0.005632399115711451, + 0.004361375235021114, + 0.008029148913919926, + 0.004252838436514139, + 0.004480683244764805, + 0.005369667429476976, + 0.0057629854418337345, + 0.009516509249806404, + 0.006857470143586397, + 0.009378520771861076, + 0.008170552551746368, + 0.01008432824164629, + 0.010444816201925278, + 0.012567195110023022, + 0.010551752522587776, + 0.009210752323269844, + 0.010930291377007961, + 0.01244593970477581, + 0.012403635308146477, + 0.004870247561484575, + 0.006299635861068964, + 0.004996963776648045, + 0.005588131956756115, + 0.010054762475192547, + 0.005972488317638636, + 0.008464200422167778, + 0.008544662967324257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.014226209372282028 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of a common garden snail on a vibrant emerald green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's shell a delicate spiral of amber and beige hues with fine concentric growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snail move by extending its two upper tentacles tipped with tiny black eyes while leaving a thick, sticky trail of silvery mucus that refracts light into rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft bokeh of greens and browns due to a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1799413412809372 + ], + "opposite_logprob_tokens": [ + -1.8049412965774536 + ], + "logprob_reward": -0.1799413412809372, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1799413412809372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07663784176111221, + 0.034127041697502136, + 0.02021447755396366, + 0.6665283441543579, + 0.02598470263183117, + 0.03119213879108429, + 0.037051837891340256, + 0.050096120685338974, + 0.017562832683324814, + 0.04060464724898338 + ], + "credit_map": [ + 0.007241158280521631, + 0.009096866473555565, + 0.00811129529029131, + 0.01047505158931017, + 0.009248959831893444, + 0.010222859680652618, + 0.014415500685572624, + 0.007826152257621288, + 0.0033547813072800636, + 0.004455591551959515, + 0.004329516552388668, + 0.004933146759867668, + 0.003836617572233081, + 0.003761765081435442, + 0.004836089909076691, + 0.004619534593075514, + 0.002061563776805997, + 0.0019925537053495646, + 0.0022334102541208267, + 0.0025017934385687113, + 0.0022351203951984644, + 0.0021537819411605597, + 0.004464486613869667, + 0.002571767196059227, + 0.021390646696090698, + 0.013609502464532852, + 0.021481873467564583, + 0.013172789476811886, + 0.06686189025640488, + 0.28069570660591125, + 0.027807457372546196, + 0.22150850296020508, + 0.0023874479811638594, + 0.002265318064019084, + 0.0032023603562265635, + 0.002279384294524789, + 0.003118582535535097, + 0.004521287977695465, + 0.004269100725650787, + 0.003941223490983248, + 0.003970603924244642, + 0.00373852695338428, + 0.004249407909810543, + 0.003135517705231905, + 0.004364096559584141, + 0.0038805443327873945, + 0.004552361089736223, + 0.003301081946119666, + 0.002986245322972536, + 0.00305192731320858, + 0.0030496991239488125, + 0.0030643732752650976, + 0.00839298591017723, + 0.004236948676407337, + 0.0039128935895860195, + 0.008356764912605286, + 0.006164314690977335, + 0.006410327740013599, + 0.006580669898539782, + 0.004724784754216671, + 0.006879314314574003, + 0.0060792602598667145, + 0.007257811259478331, + 0.0059996419586241245, + 0.0016485858941450715, + 0.0020038182847201824, + 0.0015250627184286714, + 0.0023389048874378204, + 0.00238101021386683, + 0.002594555029645562, + 0.0022963518276810646, + 0.0027745431289076805, + 0.002593891229480505, + 0.0034687959123402834, + 0.0031279774848371744, + 0.0031326685566455126, + 0.007215901277959347, + 0.007121017202734947, + 0.005215677432715893, + 0.008728714659810066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.1799413412809372 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of a common garden snail on a vibrant emerald green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's shell a delicate spiral of amber and beige hues with fine concentric growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snail move by extending its two upper tentacles tipped with tiny black eyes while leaving a thick, sticky trail of silvery mucus that refracts light into rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft bokeh of greens and browns due to a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076326169073581696 + ], + "opposite_logprob_tokens": [ + -4.882632732391357 + ], + "logprob_reward": -0.0076326169073581696, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076326169073581696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06045402213931084, + 0.03435668349266052, + 0.035320524126291275, + 0.4944450259208679, + 0.06049083545804024, + 0.027778783813118935, + 0.02645370550453663, + 0.06805665045976639, + 0.028943784534931183, + 0.16369998455047607 + ], + "credit_map": [ + 0.004320414736866951, + 0.0059814671985805035, + 0.0075204456225037575, + 0.007005492690950632, + 0.005292950663715601, + 0.007713638711720705, + 0.00836850143969059, + 0.014251106418669224, + 0.0034763796720653772, + 0.0027969637885689735, + 0.0034840665757656097, + 0.0028442544862627983, + 0.003411069745197892, + 0.004908069036900997, + 0.004622946958988905, + 0.00881293136626482, + 0.0028193725738674402, + 0.004120005294680595, + 0.004116129130125046, + 0.003639243310317397, + 0.003784686792641878, + 0.0040231067687273026, + 0.007221178151667118, + 0.005596796050667763, + 0.015984870493412018, + 0.00822083093225956, + 0.015643779188394547, + 0.008198585361242294, + 0.04920461028814316, + 0.2105325162410736, + 0.020121578127145767, + 0.16653822362422943, + 0.005697049666196108, + 0.00674594659358263, + 0.0055848946794867516, + 0.005651806481182575, + 0.009792931377887726, + 0.011581292375922203, + 0.00999594945460558, + 0.0054409620352089405, + 0.002634986536577344, + 0.0024956727866083384, + 0.0035452693700790405, + 0.0027715563774108887, + 0.004887474235147238, + 0.004266162868589163, + 0.0039053529035300016, + 0.0032723061740398407, + 0.002310049021616578, + 0.0021728582214564085, + 0.003049431834369898, + 0.003035549772903323, + 0.0038785948418080807, + 0.003254491603001952, + 0.0049675037153065205, + 0.0037852271925657988, + 0.00796412117779255, + 0.007722438778728247, + 0.007861778140068054, + 0.004855819046497345, + 0.009727731347084045, + 0.008570513688027859, + 0.011099597439169884, + 0.01025464478880167, + 0.0018911226652562618, + 0.0025918330065906048, + 0.0020451515447348356, + 0.0023851655423641205, + 0.003764538327232003, + 0.00531275849789381, + 0.00431086216121912, + 0.006642351858317852, + 0.008863871917128563, + 0.014367234893143177, + 0.011548059061169624, + 0.010506393387913704, + 0.021222935989499092, + 0.038363564759492874, + 0.023687200620770454, + 0.03514070436358452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 7 + ], + "selected_reward": -0.0076326169073581696 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of a common garden snail on a vibrant emerald green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's shell a delicate spiral of amber and beige hues with fine concentric growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snail move by extending its two upper tentacles tipped with tiny black eyes while leaving a thick, sticky trail of silvery mucus that refracts light into rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft bokeh of greens and browns due to a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635989524424076 + ], + "opposite_logprob_tokens": [ + -3.651359796524048 + ], + "logprob_reward": -0.02635989524424076, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635989524424076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04242340847849846, + 0.03492150828242302, + 0.02353588491678238, + 0.5672507286071777, + 0.02436058409512043, + 0.043296851217746735, + 0.1133042722940445, + 0.07161229103803635, + 0.03616669774055481, + 0.04312777519226074 + ], + "credit_map": [ + 0.004123274702578783, + 0.007208571303635836, + 0.004600392654538155, + 0.004322114400565624, + 0.0038334219716489315, + 0.004905087873339653, + 0.004764188546687365, + 0.008666353300213814, + 0.003896601963788271, + 0.004712170921266079, + 0.004633370786905289, + 0.0052189091220498085, + 0.003200879553332925, + 0.0037850632797926664, + 0.005555959418416023, + 0.0039185527712106705, + 0.002710876287892461, + 0.0027049302589148283, + 0.0027806356083601713, + 0.002645186148583889, + 0.0022138177882879972, + 0.002858224557712674, + 0.0035458828788250685, + 0.004076329059898853, + 0.018227223306894302, + 0.010809993371367455, + 0.017979497089982033, + 0.010180352255702019, + 0.0556168258190155, + 0.24029482901096344, + 0.02356664091348648, + 0.19057537615299225, + 0.001717864302918315, + 0.0022483677603304386, + 0.0030126364436000586, + 0.0018610891420394182, + 0.0031257893424481153, + 0.006554202642291784, + 0.001835956354625523, + 0.004004676826298237, + 0.004855007864534855, + 0.005564885679632425, + 0.004789894912391901, + 0.005264453589916229, + 0.0056677404791116714, + 0.00520311389118433, + 0.006768834311515093, + 0.00518292048946023, + 0.007786783389747143, + 0.009686575271189213, + 0.014244900085031986, + 0.01197279617190361, + 0.022487757727503777, + 0.011076949536800385, + 0.017049677670001984, + 0.018998827785253525, + 0.006130258087068796, + 0.005755613092333078, + 0.0064459750428795815, + 0.006182674318552017, + 0.01336204819381237, + 0.008336493745446205, + 0.01188204437494278, + 0.013517185114324093, + 0.0038509853184223175, + 0.0036022344138473272, + 0.0032903377432376146, + 0.003053926397114992, + 0.004955269396305084, + 0.005634914617985487, + 0.00495522515848279, + 0.006823805160820484, + 0.00524742528796196, + 0.006671111099421978, + 0.00785340927541256, + 0.004534075502306223, + 0.004141557961702347, + 0.00464207585901022, + 0.005231037270277739, + 0.004807083867490292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 7 + ], + "selected_reward": -0.02635989524424076 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme macro shot, taken at a slightly elevated angle, captures the deliberate journey of a common garden snail across a vibrant, emerald green leaf. Soft, diffused morning light filters through the invisible canopy, illuminating the scene with a gentle glow. The snail's shell is a delicate spiral of amber and beige hues, displaying fine concentric growth lines on its polished surface. Its soft, translucent body, a pale, fleshy gray, shimmers as it extends its two upper tentacles, each tipped with a tiny black eye that cautiously senses the path ahead. The creature moves in an almost imperceptible, rhythmic wave, steadily skimming across the leaf's textured surface, where the prominent veins form a delicate network of branches. In its wake, it leaves a thick, sticky trail of silvery mucus that catches the light and refracts it into a stunning rainbow of iridescent colors that shimmers with every subtle shift in the light. The shallow depth of field blurs the background into a soft bokeh of greens and browns, creating an atmosphere of tranquility, contemplation, and wonder, focusing entirely on this miniature, thoughtful journey.", + "seed": 21311, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_8/index_6449_seed_21311_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.019692574627697468, + "vlm_logprob": -0.019692574627697468, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.019692574627697468, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of a common garden snail on a vibrant emerald green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's shell a delicate spiral of amber and beige hues with fine concentric growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snail move by extending its two upper tentacles tipped with tiny black eyes while leaving a thick, sticky trail of silvery mucus that refracts light into rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft bokeh of greens and browns due to a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007906096056103706 + ], + "opposite_logprob_tokens": [ + -5.132905960083008 + ], + "logprob_reward": -0.007906096056103706, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.007906096056103706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14073510468006134, + 0.02075260877609253, + 0.01835539937019348, + 0.5954714417457581, + 0.05692671239376068, + 0.028014985844492912, + 0.037023358047008514, + 0.04893995448946953, + 0.016057681292295456, + 0.03772270679473877 + ], + "credit_map": [ + 0.011726828292012215, + 0.010808674618601799, + 0.03267206624150276, + 0.011804173700511456, + 0.010466760024428368, + 0.012328681536018848, + 0.010721403174102306, + 0.04020652920007706, + 0.0021984109189361334, + 0.002015134319663048, + 0.003976133186370134, + 0.0024756393395364285, + 0.002084128325805068, + 0.0019281897693872452, + 0.0018223135266453028, + 0.0042526600882411, + 0.0019797994755208492, + 0.0027190628461539745, + 0.0024948709178715944, + 0.002446959028020501, + 0.0033208841923624277, + 0.0019418394658714533, + 0.0016388267977163196, + 0.001813155715353787, + 0.04525192826986313, + 0.03932894766330719, + 0.16910889744758606, + 0.047859612852334976, + 0.030395012348890305, + 0.044549476355314255, + 0.027087925001978874, + 0.1918896734714508, + 0.00479418458417058, + 0.005165745969861746, + 0.007157491520047188, + 0.009150774218142033, + 0.006928245071321726, + 0.00675021531060338, + 0.006589829921722412, + 0.010390227660536766, + 0.003505517728626728, + 0.0024696795735508204, + 0.0030696443282067776, + 0.005257362499833107, + 0.002843694295734167, + 0.003502946812659502, + 0.0029104985296726227, + 0.004455642309039831, + 0.004077681340277195, + 0.004574617836624384, + 0.00439877063035965, + 0.0045784758403897285, + 0.004885392729192972, + 0.004401675425469875, + 0.0047768051736056805, + 0.00532993720844388, + 0.00452535692602396, + 0.005778228864073753, + 0.005338911898434162, + 0.008703570812940598, + 0.008461701683700085, + 0.005062107462435961, + 0.005498119164258242, + 0.005571953486651182, + 0.0014116134261712432, + 0.0018913880921900272, + 0.0016716552199795842, + 0.002609523478895426, + 0.0032198212575167418, + 0.00228627840988338, + 0.0013289686758071184, + 0.0016384315676987171, + 0.002982502104714513, + 0.003828030079603195, + 0.003916276153177023, + 0.006351770833134651, + 0.00778464088216424, + 0.004499670118093491, + 0.0038000699132680893, + 0.004559745080769062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.007906096056103706 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of a common garden snail on a vibrant emerald green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's shell a delicate spiral of amber and beige hues with fine concentric growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snail move by extending its two upper tentacles tipped with tiny black eyes while leaving a thick, sticky trail of silvery mucus that refracts light into rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft bokeh of greens and browns due to a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006750871893018484 + ], + "opposite_logprob_tokens": [ + -5.00675106048584 + ], + "logprob_reward": -0.006750871893018484, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006750871893018484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10885333269834518, + 0.015679847449064255, + 0.023007266223430634, + 0.5965967178344727, + 0.04151513800024986, + 0.026626691222190857, + 0.037055209279060364, + 0.0663377121090889, + 0.03733285516500473, + 0.04699527844786644 + ], + "credit_map": [ + 0.009024782106280327, + 0.008788557723164558, + 0.02487758733332157, + 0.010640259832143784, + 0.007388008758425713, + 0.009816830977797508, + 0.00893454160541296, + 0.029382755979895592, + 0.001590520259924233, + 0.0022159768268465996, + 0.002756791189312935, + 0.0014546910533681512, + 0.001800435595214367, + 0.0014206235064193606, + 0.0014318082248792052, + 0.003008998231962323, + 0.002258459571748972, + 0.004769607447087765, + 0.0033991497475653887, + 0.002955809934064746, + 0.003836217802017927, + 0.0017771953716874123, + 0.0018391372868791223, + 0.002171686151996255, + 0.04614323377609253, + 0.03966988995671272, + 0.17231467366218567, + 0.05137897655367851, + 0.03076697699725628, + 0.0357087105512619, + 0.027266038581728935, + 0.19334813952445984, + 0.0033225473016500473, + 0.003239997196942568, + 0.004532374907284975, + 0.005397008266299963, + 0.00880042277276516, + 0.003509742906317115, + 0.005817847792059183, + 0.006895194761455059, + 0.002920325845479965, + 0.002556569641456008, + 0.0033577207941561937, + 0.004996891133487225, + 0.004016682505607605, + 0.002842558082193136, + 0.003015547525137663, + 0.0029203917365521193, + 0.0032267163041979074, + 0.0037418592255562544, + 0.006678567733615637, + 0.005607372149825096, + 0.004342207219451666, + 0.004109309054911137, + 0.0037205347325652838, + 0.005628638900816441, + 0.010260325856506824, + 0.008899000473320484, + 0.006773469038307667, + 0.01002427563071251, + 0.00779491662979126, + 0.0064375875517725945, + 0.007102899719029665, + 0.00904522929340601, + 0.0033857703674584627, + 0.0041579375974833965, + 0.005041429772973061, + 0.005251702852547169, + 0.009066798724234104, + 0.00418838020414114, + 0.0028517737518996, + 0.0033890600316226482, + 0.002925679786130786, + 0.00502027478069067, + 0.004984400700777769, + 0.005701103247702122, + 0.011241571977734566, + 0.007228075992316008, + 0.004872572608292103, + 0.005021593067795038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.006750871893018484 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of a common garden snail on a vibrant emerald green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's shell a delicate spiral of amber and beige hues with fine concentric growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snail move by extending its two upper tentacles tipped with tiny black eyes while leaving a thick, sticky trail of silvery mucus that refracts light into rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft bokeh of greens and browns due to a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0549643374979496 + ], + "opposite_logprob_tokens": [ + -2.929964303970337 + ], + "logprob_reward": -0.0549643374979496, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549643374979496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08201108127832413, + 0.017389601096510887, + 0.033647432923316956, + 0.5500841736793518, + 0.05427084490656853, + 0.05860583856701851, + 0.0323282852768898, + 0.08613752573728561, + 0.052982304245233536, + 0.03254282847046852 + ], + "credit_map": [ + 0.0067051067017018795, + 0.007319997530430555, + 0.01762303151190281, + 0.008153198286890984, + 0.004905506037175655, + 0.010639213025569916, + 0.00664822943508625, + 0.020016798749566078, + 0.0016153688775375485, + 0.0021925040055066347, + 0.0032429173588752747, + 0.0018130007665604353, + 0.002358866622671485, + 0.0016499620396643877, + 0.001764492830261588, + 0.002752488013356924, + 0.002820858033373952, + 0.004802053328603506, + 0.007494197227060795, + 0.0033445642329752445, + 0.004511865321546793, + 0.003148651449009776, + 0.0018320585368201137, + 0.005693183746188879, + 0.04279780015349388, + 0.03491619974374771, + 0.16318583488464355, + 0.04727189615368843, + 0.026091979816555977, + 0.030022528022527695, + 0.024453507736325264, + 0.18134447932243347, + 0.005571012385189533, + 0.004011756740510464, + 0.007693462073802948, + 0.008455975912511349, + 0.004413241986185312, + 0.006032040808349848, + 0.007888670079410076, + 0.010204683989286423, + 0.0071264044381678104, + 0.0038837208412587643, + 0.008374089375138283, + 0.010187028907239437, + 0.005220339633524418, + 0.006221643649041653, + 0.00649169972166419, + 0.011100907810032368, + 0.00296515179798007, + 0.003326793434098363, + 0.006325260270386934, + 0.0053056771866977215, + 0.0033574088010936975, + 0.003408031538128853, + 0.0030648610554635525, + 0.004575103521347046, + 0.011779439635574818, + 0.010930580087006092, + 0.009644431993365288, + 0.014637582935392857, + 0.014057828113436699, + 0.00846912432461977, + 0.008155384100973606, + 0.008463151752948761, + 0.004325026646256447, + 0.00665242737159133, + 0.005801872815936804, + 0.009088701568543911, + 0.00993844959884882, + 0.0061066001653671265, + 0.004958329722285271, + 0.006110901013016701, + 0.00255792704410851, + 0.003137858584523201, + 0.003270938294008374, + 0.004111269023269415, + 0.003991767764091492, + 0.00696180434897542, + 0.0034233317710459232, + 0.005087934900075197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.0549643374979496 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of a common garden snail on a vibrant emerald green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's shell a delicate spiral of amber and beige hues with fine concentric growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snail move by extending its two upper tentacles tipped with tiny black eyes while leaving a thick, sticky trail of silvery mucus that refracts light into rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft bokeh of greens and browns due to a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00248968368396163 + ], + "opposite_logprob_tokens": [ + -6.0024895668029785 + ], + "logprob_reward": -0.00248968368396163, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.00248968368396163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03397872671484947, + 0.050498273223638535, + 0.019859064370393753, + 0.7827669382095337, + 0.026446158066391945, + 0.024684229865670204, + 0.014090853743255138, + 0.011085977777838707, + 0.015361124649643898, + 0.021228697150945663 + ], + "credit_map": [ + 0.003429671749472618, + 0.004936561454087496, + 0.005292516201734543, + 0.003449764335528016, + 0.003324780846014619, + 0.002956808777526021, + 0.003695693099871278, + 0.006892932578921318, + 0.007191679440438747, + 0.004574339371174574, + 0.008139120414853096, + 0.007030705455690622, + 0.004912607837468386, + 0.006182420067489147, + 0.004599560517817736, + 0.007867839187383652, + 0.002091925358399749, + 0.0031691603362560272, + 0.0028815013356506824, + 0.0028886988293379545, + 0.00223901798017323, + 0.003146015340462327, + 0.0014127077301964164, + 0.002030037110671401, + 0.06398958712816238, + 0.07050091028213501, + 0.20534130930900574, + 0.0682026743888855, + 0.0504021942615509, + 0.04953332245349884, + 0.04434475302696228, + 0.23045216500759125, + 0.003187311114743352, + 0.0019509042613208294, + 0.0030206942465156317, + 0.003031826578080654, + 0.004808490630239248, + 0.0027352082543075085, + 0.003280441975221038, + 0.004431283567100763, + 0.0032141103874891996, + 0.002035856945440173, + 0.0030258672777563334, + 0.004038725048303604, + 0.002274609636515379, + 0.0035042697563767433, + 0.0022068542893975973, + 0.004383936990052462, + 0.0014193312963470817, + 0.0016922042705118656, + 0.0023251676466315985, + 0.0020599807612597942, + 0.0017737123416736722, + 0.0016130992444232106, + 0.0013954541645944118, + 0.0018119043670594692, + 0.0013287437614053488, + 0.001900790142826736, + 0.0012020310387015343, + 0.001770326984114945, + 0.001289485371671617, + 0.0009261001832783222, + 0.0011271997354924679, + 0.0015413005603477359, + 0.0013420783216133714, + 0.0018930556252598763, + 0.0014681729953736067, + 0.002531805308535695, + 0.0028572201263159513, + 0.0016594880726188421, + 0.001769344205968082, + 0.001839961507357657, + 0.0018229620764032006, + 0.002323987428098917, + 0.0022181563545018435, + 0.002872647251933813, + 0.004592581652104855, + 0.0029277177527546883, + 0.0020222885068506002, + 0.0024483550805598497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.00248968368396163 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of a common garden snail on a vibrant emerald green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's shell a delicate spiral of amber and beige hues with fine concentric growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snail move by extending its two upper tentacles tipped with tiny black eyes while leaving a thick, sticky trail of silvery mucus that refracts light into rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft bokeh of greens and browns due to a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635188400745392 + ], + "opposite_logprob_tokens": [ + -3.6513519287109375 + ], + "logprob_reward": -0.02635188400745392, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635188400745392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11480366438627243, + 0.05094272270798683, + 0.0158904567360878, + 0.5793436169624329, + 0.04057285189628601, + 0.03625250235199928, + 0.0368584543466568, + 0.033577773720026016, + 0.03174488991498947, + 0.060013096779584885 + ], + "credit_map": [ + 0.011319592595100403, + 0.010186592116951942, + 0.0252927765250206, + 0.011500897817313671, + 0.008640987798571587, + 0.011457092128694057, + 0.009451719932258129, + 0.02695399895310402, + 0.00619001779705286, + 0.003809356363490224, + 0.009195973165333271, + 0.007322344928979874, + 0.0038280622102320194, + 0.005496346391737461, + 0.004238014575093985, + 0.01086260937154293, + 0.0016307835467159748, + 0.0022081893403083086, + 0.002480587689206004, + 0.002649450208991766, + 0.0020191080402582884, + 0.00214834394864738, + 0.0012598587200045586, + 0.001494135009124875, + 0.04416602849960327, + 0.04248620197176933, + 0.16141213476657867, + 0.04999074339866638, + 0.03317015618085861, + 0.03727905452251434, + 0.02832983247935772, + 0.18250945210456848, + 0.0047208829782903194, + 0.0038698334246873856, + 0.005525786895304918, + 0.0057473513297736645, + 0.00404332997277379, + 0.005127216689288616, + 0.004598761908710003, + 0.0069396900944411755, + 0.00495842844247818, + 0.0024791753385215998, + 0.005607421509921551, + 0.006710710469633341, + 0.0023539066314697266, + 0.0037049122620373964, + 0.00391344865784049, + 0.006524499971419573, + 0.003931086510419846, + 0.0037923564668744802, + 0.004340405110269785, + 0.006507302634418011, + 0.0035930511076003313, + 0.004285835660994053, + 0.0048667495138943195, + 0.005541669204831123, + 0.003333783010020852, + 0.0031760206911712885, + 0.004365069791674614, + 0.006975261494517326, + 0.004388162866234779, + 0.0031463203486055136, + 0.0034881222527474165, + 0.004705033265054226, + 0.0031345009338110685, + 0.005120156332850456, + 0.0036273719742894173, + 0.005061231553554535, + 0.0054495809599757195, + 0.003139443462714553, + 0.003044290468096733, + 0.003168314229696989, + 0.006883637513965368, + 0.005099664442241192, + 0.006013520993292332, + 0.009650051593780518, + 0.00872040819376707, + 0.010150627233088017, + 0.0065917205065488815, + 0.006903467699885368 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.02635188400745392 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme macro shot, taken at a slightly elevated angle, captures the deliberate journey of a common garden snail across a vibrant, emerald green leaf. Soft, diffused morning light filters through the invisible canopy, illuminating the scene with a gentle glow. The snail's shell is a delicate spiral of amber and beige hues, displaying fine concentric growth lines on its polished surface. Its soft, translucent body, a pale, fleshy gray, shimmers as it extends its two upper tentacles, each tipped with a tiny black eye that cautiously senses the path ahead. The creature moves in an almost imperceptible, rhythmic wave, steadily skimming across the leaf's textured surface, where the prominent veins form a delicate network of branches. In its wake, it leaves a thick, sticky trail of silvery mucus that catches the light and refracts it into a stunning rainbow of iridescent colors that shimmers with every subtle shift in the light. The shallow depth of field blurs the background into a soft bokeh of greens and browns, creating an atmosphere of tranquility, contemplation, and wonder, focusing entirely on this miniature, thoughtful journey.", + "seed": 21312, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_8/index_6449_seed_21312_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10695454641245306, + "vlm_logprob": -0.10695454641245306, + "vlm_yes_no_margin": 3.699999952316284, + "vlm_token_credit": -0.10695454641245306, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of a common garden snail on a vibrant emerald green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's shell a delicate spiral of amber and beige hues with fine concentric growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snail move by extending its two upper tentacles tipped with tiny black eyes while leaving a thick, sticky trail of silvery mucus that refracts light into rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft bokeh of greens and browns due to a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012626232579350471 + ], + "opposite_logprob_tokens": [ + -4.6376261711120605 + ], + "logprob_reward": -0.012626232579350471, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.012626232579350471, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055750127881765366, + 0.025179948657751083, + 0.0165064986795187, + 0.036802466958761215, + 0.01500174030661583, + 0.02261228673160076, + 0.02333747409284115, + 0.021369364112615585, + 0.7240331172943115, + 0.05940694734454155 + ], + "credit_map": [ + 0.004680091515183449, + 0.004601188004016876, + 0.00495966337621212, + 0.005517215002328157, + 0.008926576003432274, + 0.012391205877065659, + 0.00973277073353529, + 0.004941417835652828, + 0.0030131335370242596, + 0.003022813703864813, + 0.0032561307307332754, + 0.002933311276137829, + 0.0026542809791862965, + 0.0027419226244091988, + 0.005573215428739786, + 0.0019851410761475563, + 0.0019712147768586874, + 0.0014934170758351684, + 0.0016617832006886601, + 0.002462799195200205, + 0.0024451755452901125, + 0.002430540742352605, + 0.0016426789807155728, + 0.0023988906759768724, + 0.0030218553729355335, + 0.0035781930200755596, + 0.008490527048707008, + 0.007557120639830828, + 0.0025441977195441723, + 0.003924040123820305, + 0.003948306664824486, + 0.0037382221780717373, + 0.0015121364267542958, + 0.002048784401267767, + 0.0017813834128901362, + 0.0014601954026147723, + 0.0017411591252312064, + 0.0019693863578140736, + 0.0026078398805111647, + 0.0018808571621775627, + 0.0019768725614994764, + 0.0025768782943487167, + 0.0020455815829336643, + 0.0020924117416143417, + 0.0034715384244918823, + 0.003007316030561924, + 0.0037905157078057528, + 0.0036511728540062904, + 0.0027093889657408, + 0.003596134716644883, + 0.0026854793541133404, + 0.0029203633312135935, + 0.0028048334643244743, + 0.002652861177921295, + 0.0032186349853873253, + 0.002749778563156724, + 0.0022101858630776405, + 0.002241050358861685, + 0.0027909164782613516, + 0.0018174522556364536, + 0.003733063582330942, + 0.0024319784715771675, + 0.003296049078926444, + 0.002848670119419694, + 0.025437498465180397, + 0.023608272895216942, + 0.01998753286898136, + 0.03309012949466705, + 0.07079879194498062, + 0.295888215303421, + 0.034682344645261765, + 0.2205403596162796, + 0.008396984077990055, + 0.00995730422437191, + 0.005678711459040642, + 0.00569495651870966, + 0.008289597928524017, + 0.008685104548931122, + 0.0049487571232020855, + 0.007755532395094633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.012626232579350471 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of a common garden snail on a vibrant emerald green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's shell a delicate spiral of amber and beige hues with fine concentric growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snail move by extending its two upper tentacles tipped with tiny black eyes while leaving a thick, sticky trail of silvery mucus that refracts light into rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft bokeh of greens and browns due to a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018219612538814545 + ], + "opposite_logprob_tokens": [ + -4.018219470977783 + ], + "logprob_reward": -0.018219612538814545, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018219612538814545, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061443738639354706, + 0.024583429098129272, + 0.019276311621069908, + 0.016288118436932564, + 0.035444628447294235, + 0.025671210139989853, + 0.03453989699482918, + 0.039887987077236176, + 0.6917247772216797, + 0.05113999545574188 + ], + "credit_map": [ + 0.005553837399929762, + 0.005612208507955074, + 0.007814798504114151, + 0.00696905329823494, + 0.007844565436244011, + 0.011019882746040821, + 0.011530604213476181, + 0.005098781548440456, + 0.002827603369951248, + 0.003196952398866415, + 0.004036750644445419, + 0.0029182208236306906, + 0.0023876952473074198, + 0.00277012144215405, + 0.004427239764481783, + 0.0020188437774777412, + 0.002781872870400548, + 0.002293760422617197, + 0.001983105204999447, + 0.0017837134655565023, + 0.003296790411695838, + 0.002450882690027356, + 0.0026615085080266, + 0.002024676650762558, + 0.0021218338515609503, + 0.0013603850966319442, + 0.001591646927408874, + 0.0017254542326554656, + 0.0017754138680174947, + 0.0019476067973300815, + 0.002768885809928179, + 0.0029968908056616783, + 0.003625722136348486, + 0.005061492323875427, + 0.0035341186448931694, + 0.002925385255366564, + 0.004577958956360817, + 0.006687985733151436, + 0.004880205728113651, + 0.004151754546910524, + 0.0022234751377254725, + 0.002799183828756213, + 0.0023763319477438927, + 0.002720463555306196, + 0.0034547154791653156, + 0.0037025990895926952, + 0.004246287979185581, + 0.004148151725530624, + 0.004128590226173401, + 0.004851071164011955, + 0.004078573081642389, + 0.003143438370898366, + 0.004452102817595005, + 0.004380162339657545, + 0.0042704190127551556, + 0.005235535092651844, + 0.0034094310831278563, + 0.00325274420902133, + 0.004469116684049368, + 0.0037575128953903913, + 0.00662267254665494, + 0.0065507302060723305, + 0.006288581062108278, + 0.005537194665521383, + 0.02456223964691162, + 0.023207692429423332, + 0.019250812008976936, + 0.03134222328662872, + 0.06815297901630402, + 0.28213873505592346, + 0.033188965171575546, + 0.20988105237483978, + 0.007113568019121885, + 0.008425374515354633, + 0.004922565072774887, + 0.004996343515813351, + 0.006930551491677761, + 0.007695484906435013, + 0.0043102530762553215, + 0.006745854392647743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.018219612538814545 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of a common garden snail on a vibrant emerald green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's shell a delicate spiral of amber and beige hues with fine concentric growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snail move by extending its two upper tentacles tipped with tiny black eyes while leaving a thick, sticky trail of silvery mucus that refracts light into rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft bokeh of greens and browns due to a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4743767976760864 + ], + "opposite_logprob_tokens": [ + -0.9743767976760864 + ], + "logprob_reward": -0.4743767976760864, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4743767976760864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06500127166509628, + 0.01853804662823677, + 0.027122793719172478, + 0.02275816537439823, + 0.021921759471297264, + 0.03765050694346428, + 0.028179997578263283, + 0.04184022918343544, + 0.670956015586853, + 0.06603113561868668 + ], + "credit_map": [ + 0.005952082108706236, + 0.00604521669447422, + 0.007462188135832548, + 0.009503640234470367, + 0.00770096480846405, + 0.01129225268959999, + 0.011731875129044056, + 0.005313051864504814, + 0.002386340405791998, + 0.0025636828504502773, + 0.00233767949976027, + 0.0026713437400758266, + 0.0018537689466029406, + 0.0019036956364288926, + 0.003365229582414031, + 0.0014563074801117182, + 0.0029325291980057955, + 0.00248621660284698, + 0.002011964563280344, + 0.003602624638006091, + 0.003813903545960784, + 0.003780432278290391, + 0.004507291130721569, + 0.003987833391875029, + 0.0017001493833959103, + 0.0023779836483299732, + 0.003399529494345188, + 0.0036418228410184383, + 0.0019295350648462772, + 0.0030132599640637636, + 0.0034877415746450424, + 0.0032081448007375, + 0.0022066067904233932, + 0.0022749050986021757, + 0.0023721493780612946, + 0.0023691097740083933, + 0.0034422720782458782, + 0.0027333758771419525, + 0.0033735353499650955, + 0.0031498067546635866, + 0.002653113566339016, + 0.002974337199702859, + 0.0030784436967223883, + 0.004060270264744759, + 0.0041131009347736835, + 0.004993567708879709, + 0.007715310901403427, + 0.008062362670898438, + 0.0036027305759489536, + 0.004373413510620594, + 0.002944663166999817, + 0.0028337689582258463, + 0.0035476151388138533, + 0.0033037445973604918, + 0.003840724239125848, + 0.003733338788151741, + 0.0038018948398530483, + 0.0032488710712641478, + 0.003870193613693118, + 0.0039550140500068665, + 0.006613864563405514, + 0.008679633028805256, + 0.0059760273434221745, + 0.005694734398275614, + 0.02369239367544651, + 0.022697174921631813, + 0.01894913986325264, + 0.03070916049182415, + 0.06529674679040909, + 0.273377925157547, + 0.031958840787410736, + 0.2042747139930725, + 0.010442073456943035, + 0.01129617914557457, + 0.006230877246707678, + 0.006355200428515673, + 0.009349341504275799, + 0.008976789191365242, + 0.005117649678140879, + 0.00826303195208311 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.4743767976760864 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of a common garden snail on a vibrant emerald green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's shell a delicate spiral of amber and beige hues with fine concentric growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snail move by extending its two upper tentacles tipped with tiny black eyes while leaving a thick, sticky trail of silvery mucus that refracts light into rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft bokeh of greens and browns due to a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031953013967722654 + ], + "opposite_logprob_tokens": [ + -5.753195285797119 + ], + "logprob_reward": -0.0031953013967722654, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031953013967722654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07227220386266708, + 0.07692078500986099, + 0.044355958700180054, + 0.032903533428907394, + 0.058560483157634735, + 0.06346667557954788, + 0.045393116772174835, + 0.0817575678229332, + 0.47630107402801514, + 0.04806867614388466 + ], + "credit_map": [ + 0.0065202871337533, + 0.0067953444086015224, + 0.006743256468325853, + 0.005469425115734339, + 0.01372700184583664, + 0.01721842586994171, + 0.008462876081466675, + 0.0073355757631361485, + 0.007034521549940109, + 0.007028648164123297, + 0.007402525283396244, + 0.01498475857079029, + 0.00765519542619586, + 0.007687616627663374, + 0.01821889728307724, + 0.006908618379384279, + 0.004334022756665945, + 0.004641055595129728, + 0.00557434419170022, + 0.005047385115176439, + 0.007034687325358391, + 0.00506992731243372, + 0.005299924407154322, + 0.007354608736932278, + 0.00308491219766438, + 0.003469094866886735, + 0.005376964341849089, + 0.005197674967348576, + 0.0031433934345841408, + 0.0034629651345312595, + 0.004390073474496603, + 0.004778448957949877, + 0.0033873228821903467, + 0.005810811184346676, + 0.0064281197264790535, + 0.008927494287490845, + 0.006397509481757879, + 0.008895660750567913, + 0.007471258752048016, + 0.01124229934066534, + 0.005136527121067047, + 0.00677080312743783, + 0.005799748469144106, + 0.0051266406662762165, + 0.009982241317629814, + 0.01038676779717207, + 0.008299565874040127, + 0.011964372359216213, + 0.0030687672551721334, + 0.004359577316790819, + 0.006404159124940634, + 0.0055678896605968475, + 0.004253513179719448, + 0.006385287269949913, + 0.0051697599701583385, + 0.010184157639741898, + 0.006639281753450632, + 0.006640112493187189, + 0.011916587129235268, + 0.00687866285443306, + 0.0109309833496809, + 0.015399589203298092, + 0.006078595761209726, + 0.017273753881454468, + 0.01702231727540493, + 0.015918945893645287, + 0.013072444126009941, + 0.0216079019010067, + 0.047004472464323044, + 0.1947167068719864, + 0.022552473470568657, + 0.14440575242042542, + 0.006560058332979679, + 0.008177569136023521, + 0.004721645265817642, + 0.0044331117533147335, + 0.006621008273214102, + 0.0068213120102882385, + 0.004555815365165472, + 0.006178149022161961 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 1 + ], + "selected_reward": -0.0031953013967722654 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of a common garden snail on a vibrant emerald green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's shell a delicate spiral of amber and beige hues with fine concentric growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snail move by extending its two upper tentacles tipped with tiny black eyes while leaving a thick, sticky trail of silvery mucus that refracts light into rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft bokeh of greens and browns due to a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635478787124157 + ], + "opposite_logprob_tokens": [ + -3.6513547897338867 + ], + "logprob_reward": -0.02635478787124157, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635478787124157, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0897679328918457, + 0.082905612885952, + 0.0472041480243206, + 0.07961898297071457, + 0.03865751251578331, + 0.0329822301864624, + 0.04167986288666725, + 0.08987967669963837, + 0.47288018465042114, + 0.02442382648587227 + ], + "credit_map": [ + 0.009643440134823322, + 0.00855562649667263, + 0.011453970335423946, + 0.00900621060281992, + 0.010242902673780918, + 0.017801137641072273, + 0.015658188611268997, + 0.007406466640532017, + 0.007822657935321331, + 0.007851812988519669, + 0.009994547814130783, + 0.013397621922194958, + 0.008553948253393173, + 0.009998572990298271, + 0.017513278871774673, + 0.007773173972964287, + 0.00687130494043231, + 0.00610025878995657, + 0.007243142928928137, + 0.0044053103774785995, + 0.005116126034408808, + 0.007145429961383343, + 0.005603552330285311, + 0.0047190249897539616, + 0.006905794143676758, + 0.006255425978451967, + 0.018736176192760468, + 0.01428704522550106, + 0.00449677649885416, + 0.011328313499689102, + 0.01012811902910471, + 0.007481342181563377, + 0.0032442016527056694, + 0.004138837568461895, + 0.00658015999943018, + 0.005929095204919577, + 0.002228035591542721, + 0.005378172732889652, + 0.004548316355794668, + 0.006610691547393799, + 0.003361488226801157, + 0.003482887288555503, + 0.002785953227430582, + 0.0026772445999085903, + 0.005157907493412495, + 0.004468897357583046, + 0.006073338445276022, + 0.004974514711648226, + 0.0040962244383990765, + 0.003933293279260397, + 0.00495221046730876, + 0.004993663635104895, + 0.005965821444988251, + 0.00494302436709404, + 0.006602989975363016, + 0.0061926390044391155, + 0.0063912831246852875, + 0.007686179131269455, + 0.008152334950864315, + 0.008828992024064064, + 0.011866751126945019, + 0.018978280946612358, + 0.015986859798431396, + 0.011989003047347069, + 0.01670331321656704, + 0.016259416937828064, + 0.014926428906619549, + 0.021754426881670952, + 0.04608041048049927, + 0.19121122360229492, + 0.02272188477218151, + 0.1432230919599533, + 0.0035063137765973806, + 0.004177760798484087, + 0.0022327881306409836, + 0.0022319841664284468, + 0.0031241653487086296, + 0.0034606684930622578, + 0.002230128040537238, + 0.0034600181970745325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.02635478787124157 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_9/index_tensor([3232, 3232])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_9/index_tensor([3232, 3232])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b393dddd3b2cd187818b1c22b5bab88ada89da62 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_9/index_tensor([3232, 3232])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, static shot captures the scene from a sideways, eye-level angle: the end of a clear plastic ruler, its transparent surface showing subtle wear and precise black millimeter markings, rests firmly against the rough, dark edge of a wooden workbench. Dramatic, high-contrast sidelighting from the left casts a bright, sharp highlight along the ruler's curved shape, while casting the surrounding area into deep shadow, highlighting the wood's grain. Suddenly, the ruler is released and rebounds with explosive force. Captured in ultra-slow motion, its body contorts into rapid waves of vibration, blurring into oscillating arcs that repeatedly strike the wood surface with sharp impacts. The violent vibrations gradually subside, and the movement slows until the ruler comes to a complete stillness, its clear form once again sharp, still against the wood, creating an atmosphere of tension and intense concentration, a controlled release of physical energy.", + "seed": 28027, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_9/index_3232_seed_28027_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.13670502044260502, + "vlm_logprob": -0.13670502044260502, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.13670502044260502, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a clear plastic ruler with black millimeter markings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting against the rough, dark edge of a wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by dramatic, high-contrast sidelighting from the left creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ruler released and shown rebounding with explosive force in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ruler's body contort into rapid waves of vibration that blur into oscillating arcs before coming to a complete stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.034375593066215515 + ], + "opposite_logprob_tokens": [ + -3.7843756675720215 + ], + "logprob_reward": -0.034375593066215515, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.034375593066215515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25201693177223206, + 0.11859681457281113, + 0.3528760075569153, + 0.09171883016824722, + 0.053135283291339874, + 0.029927199706435204, + 0.023960812017321587, + 0.01693105511367321, + 0.030383046716451645, + 0.030454017221927643 + ], + "credit_map": [ + 0.027593884617090225, + 0.027627484872937202, + 0.022850438952445984, + 0.019874515011906624, + 0.06290625035762787, + 0.04666182026267052, + 0.028178546577692032, + 0.016323991119861603, + 0.008264630101621151, + 0.008589423261582851, + 0.02193329483270645, + 0.008470556698739529, + 0.012529517523944378, + 0.014592678286135197, + 0.031031673774123192, + 0.013185037299990654, + 0.046610768884420395, + 0.07367024570703506, + 0.024727389216423035, + 0.02802105061709881, + 0.07383642345666885, + 0.058820776641368866, + 0.02597597800195217, + 0.02121337689459324, + 0.009897207841277122, + 0.014818904921412468, + 0.008402391336858273, + 0.011041881516575813, + 0.009544855915009975, + 0.020761599764227867, + 0.007758073974400759, + 0.009493912570178509, + 0.002402402926236391, + 0.014741623774170876, + 0.0020940154790878296, + 0.003717551240697503, + 0.01567254029214382, + 0.007258305326104164, + 0.003954513464123011, + 0.0032943293917924166, + 0.001872338354587555, + 0.0036735846661031246, + 0.0028713091742247343, + 0.0019530178979039192, + 0.005853606387972832, + 0.008272327482700348, + 0.003158978186547756, + 0.0022720375563949347, + 0.0011582328006625175, + 0.004582116845995188, + 0.001281160395592451, + 0.0018814801005646586, + 0.006758031435310841, + 0.005146002396941185, + 0.0014005685225129128, + 0.0017532182391732931, + 0.0011956383241340518, + 0.0034684224519878626, + 0.0016644112765789032, + 0.0017788315890356898, + 0.003825927386060357, + 0.0023414900060743093, + 0.0015439152484759688, + 0.0011124185984954238, + 0.0015789712779223919, + 0.0024279870558530092, + 0.002239529276266694, + 0.0019195477943867445, + 0.00887770764529705, + 0.0074561298824846745, + 0.0028966700192540884, + 0.0029865028336644173, + 0.0018145012436434627, + 0.007983879186213017, + 0.0019264566944912076, + 0.0022870847024023533, + 0.005605297163128853, + 0.005173936486244202, + 0.0027017747052013874, + 0.002961087040603161 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.034375593066215515 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a clear plastic ruler with black millimeter markings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting against the rough, dark edge of a wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by dramatic, high-contrast sidelighting from the left creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ruler released and shown rebounding with explosive force in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ruler's body contort into rapid waves of vibration that blur into oscillating arcs before coming to a complete stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026382071897387505 + ], + "opposite_logprob_tokens": [ + -3.6513819694519043 + ], + "logprob_reward": -0.026382071897387505, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026382071897387505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22631949186325073, + 0.12627899646759033, + 0.33119073510169983, + 0.034305356442928314, + 0.07123623788356781, + 0.05396914482116699, + 0.02236979827284813, + 0.05285804346203804, + 0.035411346703767776, + 0.04606091231107712 + ], + "credit_map": [ + 0.022662809118628502, + 0.022196386009454727, + 0.035339098423719406, + 0.026210220530629158, + 0.04437718167901039, + 0.0337555967271328, + 0.023783553391695023, + 0.01799464039504528, + 0.006984045263379812, + 0.006256020627915859, + 0.03782423213124275, + 0.006729018408805132, + 0.00998820923268795, + 0.008130318485200405, + 0.04281134903430939, + 0.00755581259727478, + 0.027702001854777336, + 0.043208394199609756, + 0.02304825372993946, + 0.030273353680968285, + 0.08476579934358597, + 0.07121434807777405, + 0.031050609424710274, + 0.019927948713302612, + 0.003712934209033847, + 0.004961859900504351, + 0.0032136249355971813, + 0.004797789268195629, + 0.0037018938455730677, + 0.005962003488093615, + 0.0034933737479150295, + 0.0044618756510317326, + 0.004146502818912268, + 0.01489465031772852, + 0.0069238608703017235, + 0.006729822605848312, + 0.01826293021440506, + 0.008557329885661602, + 0.006468307692557573, + 0.005252833943814039, + 0.0024872091598808765, + 0.006894025020301342, + 0.0032516196370124817, + 0.0036193544510751963, + 0.017426589503884315, + 0.012272817082703114, + 0.004630319308489561, + 0.0033872080966830254, + 0.0011850947048515081, + 0.006341947708278894, + 0.0016608460573479533, + 0.0011284244246780872, + 0.006765411235392094, + 0.0026775833684951067, + 0.0013181028189137578, + 0.001292389235459268, + 0.0015244993846863508, + 0.016279494389891624, + 0.0013686510501429439, + 0.002904353430494666, + 0.02135092206299305, + 0.005897096823900938, + 0.002213574480265379, + 0.0013194516068324447, + 0.0014752957504242659, + 0.010032502003014088, + 0.001807353226467967, + 0.0018974123522639275, + 0.0032080605160444975, + 0.011977325193583965, + 0.0025956693571060896, + 0.0024177273735404015, + 0.0012935471022501588, + 0.013608378358185291, + 0.0014271378749981523, + 0.001559691154398024, + 0.018701404333114624, + 0.005643223412334919, + 0.002043199259787798, + 0.0017843313980847597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.026382071897387505 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a clear plastic ruler with black millimeter markings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting against the rough, dark edge of a wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by dramatic, high-contrast sidelighting from the left creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ruler released and shown rebounding with explosive force in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ruler's body contort into rapid waves of vibration that blur into oscillating arcs before coming to a complete stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330007031559944 + ], + "opposite_logprob_tokens": [ + -3.7733001708984375 + ], + "logprob_reward": -0.02330007031559944, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330007031559944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1783921867609024, + 0.4190065562725067, + 0.14063498377799988, + 0.04402921721339226, + 0.0822358950972557, + 0.03693639114499092, + 0.015553751960396767, + 0.027170514687895775, + 0.042693402618169785, + 0.013347108848392963 + ], + "credit_map": [ + 0.023462750017642975, + 0.022521547973155975, + 0.01765611581504345, + 0.0269002802670002, + 0.02356259897351265, + 0.031183060258626938, + 0.016469091176986694, + 0.016636744141578674, + 0.022085299715399742, + 0.014867446385324001, + 0.1338820457458496, + 0.01780027151107788, + 0.019718017429113388, + 0.027208968997001648, + 0.1615927666425705, + 0.02185172587633133, + 0.010338954627513885, + 0.014344146475195885, + 0.012210488319396973, + 0.01911691389977932, + 0.02577512338757515, + 0.025271132588386536, + 0.024227486923336983, + 0.00935074221342802, + 0.004168545361608267, + 0.005745260510593653, + 0.005710875149816275, + 0.005871538072824478, + 0.004869939759373665, + 0.006847644690424204, + 0.005270366091281176, + 0.005545048508793116, + 0.0033631278201937675, + 0.027333160862326622, + 0.0031031405087560415, + 0.004118139855563641, + 0.02851056307554245, + 0.007845721207559109, + 0.004149098880589008, + 0.003812940325587988, + 0.0021518634166568518, + 0.008531676605343819, + 0.002278015948832035, + 0.0031350059434771538, + 0.008859754540026188, + 0.005282858852297068, + 0.004167109727859497, + 0.002530104713514447, + 0.0007613486959598958, + 0.0028565165121108294, + 0.0008209857624024153, + 0.0013959347270429134, + 0.0045715924352407455, + 0.003232009010389447, + 0.0009922365425154567, + 0.0009231276926584542, + 0.0007779669831506908, + 0.00983871053904295, + 0.0005101282731629908, + 0.0009459881694056094, + 0.011490645818412304, + 0.002173104789108038, + 0.0006619410705752671, + 0.0007720306748524308, + 0.0011498879175633192, + 0.004188050515949726, + 0.0009351259213872254, + 0.0011021223617717624, + 0.015187324024736881, + 0.018191443756222725, + 0.0009865384781733155, + 0.0009529099334031343, + 0.0008536737295798957, + 0.002350140130147338, + 0.0010546583216637373, + 0.0016237744130194187, + 0.002142935059964657, + 0.0025115334428846836, + 0.0015776078216731548, + 0.001232785522006452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.02330007031559944 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a clear plastic ruler with black millimeter markings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting against the rough, dark edge of a wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by dramatic, high-contrast sidelighting from the left creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ruler released and shown rebounding with explosive force in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ruler's body contort into rapid waves of vibration that blur into oscillating arcs before coming to a complete stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761703252792358 + ], + "opposite_logprob_tokens": [ + -0.8261703252792358 + ], + "logprob_reward": -0.5761703252792358, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761703252792358, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11160608381032944, + 0.09092848002910614, + 0.2284710854291916, + 0.07554832845926285, + 0.3297154903411865, + 0.06790999323129654, + 0.02346438355743885, + 0.01959766075015068, + 0.026631586253643036, + 0.026126902550458908 + ], + "credit_map": [ + 0.0189061276614666, + 0.011143720708787441, + 0.013708208687603474, + 0.013586741872131824, + 0.01950390264391899, + 0.019905149936676025, + 0.00789526104927063, + 0.006956973113119602, + 0.007242310792207718, + 0.007307020947337151, + 0.02082139439880848, + 0.00647364929318428, + 0.011853846721351147, + 0.012309910729527473, + 0.01763213239610195, + 0.007288218010216951, + 0.022862909361720085, + 0.03545895591378212, + 0.017919044941663742, + 0.025564050301909447, + 0.040078964084386826, + 0.040265459567308426, + 0.030915269628167152, + 0.015406442806124687, + 0.0092714112251997, + 0.010130826383829117, + 0.00865935068577528, + 0.01213846541941166, + 0.008589622564613819, + 0.010238101705908775, + 0.009538782760500908, + 0.006981770973652601, + 0.010581940412521362, + 0.12533317506313324, + 0.011619025841355324, + 0.009337414987385273, + 0.13326671719551086, + 0.020382478833198547, + 0.011980816721916199, + 0.007213916163891554, + 0.003108648583292961, + 0.012114710174500942, + 0.003501166822388768, + 0.00326383113861084, + 0.023559996858239174, + 0.014801327139139175, + 0.004085001535713673, + 0.0034753107465803623, + 0.0014501767000183463, + 0.0028433629777282476, + 0.0018553034169599414, + 0.0024890797212719917, + 0.005884888116270304, + 0.004394131246954203, + 0.002291179494932294, + 0.002256261883303523, + 0.0009944725316017866, + 0.005185924936085939, + 0.0015604678774252534, + 0.0018935747211799026, + 0.005736109800636768, + 0.0018199254991486669, + 0.001327850273810327, + 0.0010793340625241399, + 0.00118415302131325, + 0.008009295910596848, + 0.0011830027215182781, + 0.0014762881910428405, + 0.002156219445168972, + 0.00952262245118618, + 0.0016786166233941913, + 0.0014213889371603727, + 0.0012214557500556111, + 0.00817263126373291, + 0.0009989498648792505, + 0.0013739117421209812, + 0.007660543546080589, + 0.0030806937720626593, + 0.002098067896440625, + 0.0015206476673483849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.5761703252792358 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a clear plastic ruler with black millimeter markings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting against the rough, dark edge of a wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by dramatic, high-contrast sidelighting from the left creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ruler released and shown rebounding with explosive force in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ruler's body contort into rapid waves of vibration that blur into oscillating arcs before coming to a complete stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023297041654586792 + ], + "opposite_logprob_tokens": [ + -3.773297071456909 + ], + "logprob_reward": -0.023297041654586792, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023297041654586792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08830412477254868, + 0.2090003788471222, + 0.1496318131685257, + 0.04987279698252678, + 0.2984711527824402, + 0.03804326429963112, + 0.05500493198633194, + 0.06713824719190598, + 0.02041749842464924, + 0.024115828797221184 + ], + "credit_map": [ + 0.01051456667482853, + 0.009835980832576752, + 0.009791717864573002, + 0.01266055554151535, + 0.013752114959061146, + 0.016001764684915543, + 0.007869759574532509, + 0.00787766370922327, + 0.011062774807214737, + 0.005590230226516724, + 0.06931794434785843, + 0.008763371966779232, + 0.010044226422905922, + 0.013535869307816029, + 0.08097849041223526, + 0.009707462973892689, + 0.009657620452344418, + 0.01816006563603878, + 0.008504582569003105, + 0.010346715338528156, + 0.04736771062016487, + 0.037942904978990555, + 0.010772217996418476, + 0.006879990454763174, + 0.006517880596220493, + 0.006931880954653025, + 0.003417891450226307, + 0.006221538409590721, + 0.004904455970972776, + 0.012005755677819252, + 0.004627106711268425, + 0.005246286746114492, + 0.0083755599334836, + 0.11396095156669617, + 0.007661149371415377, + 0.007012289483100176, + 0.12849678099155426, + 0.02007976360619068, + 0.007345978636294603, + 0.005538640543818474, + 0.003222833154723048, + 0.007055058144032955, + 0.004190407693386078, + 0.0031419217120856047, + 0.008235515095293522, + 0.004628932569175959, + 0.0033410245086997747, + 0.004227569326758385, + 0.0023363670334219933, + 0.01580442488193512, + 0.0014820062788203359, + 0.002026176545768976, + 0.020021656528115273, + 0.006862279027700424, + 0.0025770249776542187, + 0.003894991474226117, + 0.0020139175467193127, + 0.02803763747215271, + 0.0015103464247658849, + 0.0016859823372215033, + 0.026413483545184135, + 0.0045412625186145306, + 0.0017063963459804654, + 0.0012292182072997093, + 0.00140046002343297, + 0.0030053425580263138, + 0.001700631924904883, + 0.001369177713058889, + 0.003438987536355853, + 0.005988746881484985, + 0.0020647854544222355, + 0.0014493652852252126, + 0.0015267328126356006, + 0.00456639938056469, + 0.001793334144167602, + 0.0017715098802000284, + 0.006292259786278009, + 0.003454207442700863, + 0.00256171147339046, + 0.002149673644453287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.023297041654586792 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot captures the scene from a sideways, eye-level angle: the end of a clear plastic ruler, its transparent surface showing subtle wear and precise black millimeter markings, rests firmly against the rough, dark edge of a wooden workbench. Dramatic, high-contrast sidelighting from the left casts a bright, sharp highlight along the ruler's curved shape, while casting the surrounding area into deep shadow, highlighting the wood's grain. Suddenly, the ruler is released and rebounds with explosive force. Captured in ultra-slow motion, its body contorts into rapid waves of vibration, blurring into oscillating arcs that repeatedly strike the wood surface with sharp impacts. The violent vibrations gradually subside, and the movement slows until the ruler comes to a complete stillness, its clear form once again sharp, still against the wood, creating an atmosphere of tension and intense concentration, a controlled release of physical energy.", + "seed": 28028, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_9/index_3232_seed_28028_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05474901609122753, + "vlm_logprob": -0.05474901609122753, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.05474901609122753, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a clear plastic ruler with black millimeter markings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting against the rough, dark edge of a wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by dramatic, high-contrast sidelighting from the left creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ruler released and shown rebounding with explosive force in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ruler's body contort into rapid waves of vibration that blur into oscillating arcs before coming to a complete stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024723060429096222 + ], + "opposite_logprob_tokens": [ + -4.149723052978516 + ], + "logprob_reward": -0.024723060429096222, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.024723060429096222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08579513430595398, + 0.2232993245124817, + 0.05348781496286392, + 0.06512454152107239, + 0.0997198298573494, + 0.22017477452754974, + 0.14221559464931488, + 0.030864229425787926, + 0.023628821596503258, + 0.05568988248705864 + ], + "credit_map": [ + 0.019817989319562912, + 0.007804528344422579, + 0.01119261421263218, + 0.006621273700147867, + 0.00456335349008441, + 0.005503454245626926, + 0.01858525723218918, + 0.01170666515827179, + 0.04798699915409088, + 0.015584133565425873, + 0.011899597011506557, + 0.012740980833768845, + 0.0732293650507927, + 0.031631167978048325, + 0.01329146046191454, + 0.01693562977015972, + 0.002347054425626993, + 0.014336283318698406, + 0.003281188430264592, + 0.002395153744146228, + 0.019242240116000175, + 0.006464273668825626, + 0.0023347241804003716, + 0.0030868968460708857, + 0.005108375567942858, + 0.010659515857696533, + 0.006664040498435497, + 0.008586030453443527, + 0.010920903645455837, + 0.012628692202270031, + 0.004804141353815794, + 0.005752843338996172, + 0.0036704055964946747, + 0.03301452472805977, + 0.0036068253684788942, + 0.0025004083290696144, + 0.010240122675895691, + 0.03730370104312897, + 0.004394955467432737, + 0.004988892935216427, + 0.008465268649160862, + 0.08413761109113693, + 0.011422330513596535, + 0.007395540829747915, + 0.013779114000499249, + 0.08098834753036499, + 0.007696610409766436, + 0.006289957091212273, + 0.005498732440173626, + 0.05998509004712105, + 0.0042166439816355705, + 0.0031792274676263332, + 0.007706906646490097, + 0.05612548440694809, + 0.0031201220117509365, + 0.0023833943996578455, + 0.001882144482806325, + 0.009898911230266094, + 0.0021066798362880945, + 0.0015957957366481423, + 0.0020963854622095823, + 0.00907048862427473, + 0.002037467435002327, + 0.0021763532422482967, + 0.0018408827017992735, + 0.005153448320925236, + 0.002274397062137723, + 0.001878425944596529, + 0.0030087484046816826, + 0.0038699267897754908, + 0.0027248007245361805, + 0.002878190716728568, + 0.0031586303375661373, + 0.018810508772730827, + 0.003485711757093668, + 0.0029083951376378536, + 0.004060027189552784, + 0.01491293404251337, + 0.004069159273058176, + 0.004284518770873547 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 6 + ], + "selected_reward": -0.024723060429096222 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a clear plastic ruler with black millimeter markings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting against the rough, dark edge of a wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by dramatic, high-contrast sidelighting from the left creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ruler released and shown rebounding with explosive force in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ruler's body contort into rapid waves of vibration that blur into oscillating arcs before coming to a complete stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018209779635071754 + ], + "opposite_logprob_tokens": [ + -4.018209934234619 + ], + "logprob_reward": -0.018209779635071754, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018209779635071754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.025270311161875725, + 0.674055814743042, + 0.031102700158953667, + 0.0407932847738266, + 0.021028948947787285, + 0.06619901210069656, + 0.08402196317911148, + 0.014007621444761753, + 0.016192376613616943, + 0.02732793055474758 + ], + "credit_map": [ + 0.005885319318622351, + 0.0019402302568778396, + 0.00560756353661418, + 0.0016454204451292753, + 0.0012512124376371503, + 0.0013218744425103068, + 0.00477282889187336, + 0.002845864975824952, + 0.18571755290031433, + 0.046353306621313095, + 0.022268474102020264, + 0.011977148242294788, + 0.2818938195705414, + 0.09737391769886017, + 0.015587840229272842, + 0.012883852235972881, + 0.0027726562693715096, + 0.006224413402378559, + 0.0026252607349306345, + 0.0027329563163220882, + 0.005189645104110241, + 0.006782648619264364, + 0.001962896203622222, + 0.0028122253715991974, + 0.0020883798133581877, + 0.012243400327861309, + 0.0028729101177304983, + 0.0034665712155401707, + 0.006154973525553942, + 0.00953215267509222, + 0.0023926531430333853, + 0.002042247913777828, + 0.0011110063642263412, + 0.0067031048238277435, + 0.001086720614694059, + 0.0009720574016682804, + 0.0031039665918797255, + 0.004778025206178427, + 0.0014897596556693316, + 0.001784312422387302, + 0.00197949749417603, + 0.026579007506370544, + 0.0016581222880631685, + 0.0012675216421484947, + 0.00397764565423131, + 0.028005728498101234, + 0.0014106297167018056, + 0.0013208710588514805, + 0.0031372900120913982, + 0.03462306410074234, + 0.002730051288381219, + 0.0024004466831684113, + 0.004791215993463993, + 0.03329676389694214, + 0.0016386127099394798, + 0.0014045241987332702, + 0.0008902756962925196, + 0.0035681873559951782, + 0.0011404529213905334, + 0.000677294097840786, + 0.0016837210860103369, + 0.0038941313978284597, + 0.0011565210297703743, + 0.0009970392566174269, + 0.0010930707212537527, + 0.003227009205147624, + 0.0016945351380854845, + 0.0009528279188089073, + 0.001664264127612114, + 0.00350641249679029, + 0.0019320027204230428, + 0.002122255740687251, + 0.0016928734257817268, + 0.008034225553274155, + 0.0018674834864214063, + 0.0019700645934790373, + 0.002928537782281637, + 0.006260322872549295, + 0.0022492941934615374, + 0.0023251317907124758 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 5 + ], + "selected_reward": -0.018209779635071754 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a clear plastic ruler with black millimeter markings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting against the rough, dark edge of a wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by dramatic, high-contrast sidelighting from the left creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ruler released and shown rebounding with explosive force in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ruler's body contort into rapid waves of vibration that blur into oscillating arcs before coming to a complete stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008648086339235306 + ], + "opposite_logprob_tokens": [ + -4.758647918701172 + ], + "logprob_reward": -0.008648086339235306, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008648086339235306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.027270980179309845, + 0.04776030778884888, + 0.09844730794429779, + 0.03131113946437836, + 0.5117958784103394, + 0.0628424808382988, + 0.14887464046478271, + 0.03042791225016117, + 0.016866328194737434, + 0.024402974173426628 + ], + "credit_map": [ + 0.00614417577162385, + 0.0022594763431698084, + 0.004014021251350641, + 0.0014952986966818571, + 0.0015548716764897108, + 0.0024617817252874374, + 0.0057554119266569614, + 0.003585943952202797, + 0.008408410474658012, + 0.004267480224370956, + 0.003316739108413458, + 0.0036582211032509804, + 0.012958446517586708, + 0.006915845908224583, + 0.003996464889496565, + 0.004238699562847614, + 0.002426381688565016, + 0.03796003386378288, + 0.0020406979601830244, + 0.002093088813126087, + 0.042963795363903046, + 0.007151752710342407, + 0.00208490714430809, + 0.0017266551731154323, + 0.0025378242135047913, + 0.006824662443250418, + 0.0021664893720299006, + 0.0031227527651935816, + 0.004687808919698, + 0.006913080811500549, + 0.0025549791753292084, + 0.0025035396683961153, + 0.01667872443795204, + 0.23145337402820587, + 0.004940657410770655, + 0.007170829921960831, + 0.024826940149068832, + 0.21831929683685303, + 0.00273088738322258, + 0.005675158929079771, + 0.0023229136131703854, + 0.022105643525719643, + 0.0024098253343254328, + 0.0019109449349343777, + 0.004906018264591694, + 0.024244453758001328, + 0.0025497081223875284, + 0.0023929765447974205, + 0.005562842357903719, + 0.06307229399681091, + 0.004360442515462637, + 0.0029888213612139225, + 0.008412029594182968, + 0.05853305384516716, + 0.002942566527053714, + 0.003002611920237541, + 0.001425622496753931, + 0.011943587101995945, + 0.001237131655216217, + 0.0009870172943919897, + 0.001892049447633326, + 0.010673102922737598, + 0.0012361403787508607, + 0.0010332620004191995, + 0.0011674027191475034, + 0.004821443930268288, + 0.0010914081940427423, + 0.0009578218450769782, + 0.0015870134811848402, + 0.003929280210286379, + 0.0015437626279890537, + 0.00176819390617311, + 0.0018425461603328586, + 0.005597488023340702, + 0.0025791453663259745, + 0.001813940005376935, + 0.002211625687777996, + 0.0063056848011910915, + 0.0023051213938742876, + 0.0017474211053922772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 2 + ], + "selected_reward": -0.008648086339235306 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a clear plastic ruler with black millimeter markings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting against the rough, dark edge of a wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by dramatic, high-contrast sidelighting from the left creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ruler released and shown rebounding with explosive force in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ruler's body contort into rapid waves of vibration that blur into oscillating arcs before coming to a complete stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20157881081104279 + ], + "opposite_logprob_tokens": [ + -1.7015788555145264 + ], + "logprob_reward": -0.20157881081104279, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20157881081104279, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05001522973179817, + 0.19017313420772552, + 0.09848877042531967, + 0.04955856874585152, + 0.4881853759288788, + 0.03495238348841667, + 0.020843135192990303, + 0.03154445067048073, + 0.018046891316771507, + 0.018192023038864136 + ], + "credit_map": [ + 0.012254969216883183, + 0.003669035155326128, + 0.004144277889281511, + 0.0030421321280300617, + 0.0030820374377071857, + 0.0026357658207416534, + 0.012874288484454155, + 0.008312727324664593, + 0.05299484729766846, + 0.013993731699883938, + 0.006890517193824053, + 0.003434971207752824, + 0.07879049330949783, + 0.02591540850698948, + 0.004774302709847689, + 0.003378871828317642, + 0.0023996036034077406, + 0.03864962235093117, + 0.002168088685721159, + 0.001993221929296851, + 0.04259508103132248, + 0.006567005068063736, + 0.0023213340900838375, + 0.001794820767827332, + 0.0032074262853711843, + 0.01476349774748087, + 0.003901941003277898, + 0.004697241820394993, + 0.005791342351585627, + 0.010918989777565002, + 0.0029579431284219027, + 0.003320185700431466, + 0.016848072409629822, + 0.21672384440898895, + 0.0074346186593174934, + 0.008678634651005268, + 0.02670334093272686, + 0.1948191374540329, + 0.007634581532329321, + 0.00934319943189621, + 0.0030837340746074915, + 0.007323268800973892, + 0.004412277601659298, + 0.003049407387152314, + 0.006211374886333942, + 0.00484911072999239, + 0.003289297688752413, + 0.0027339125517755747, + 0.0015418098773807287, + 0.003829628461971879, + 0.00298462831415236, + 0.0028751478530466557, + 0.0020233497489243746, + 0.0036864199209958315, + 0.001830412307754159, + 0.002071741037070751, + 0.0010831301333382726, + 0.012358401902019978, + 0.000816486484836787, + 0.0010134880430996418, + 0.0019352207891643047, + 0.012401938438415527, + 0.000739266281016171, + 0.0011965191224589944, + 0.0008342523360624909, + 0.006319638807326555, + 0.0008392439340241253, + 0.0009241049992851913, + 0.0013222784036770463, + 0.006285964045673609, + 0.0007500271894969046, + 0.0007713812519796193, + 0.0012384520377963781, + 0.0046523865312337875, + 0.00154962670058012, + 0.0010865667136386037, + 0.001798948273062706, + 0.00523723429068923, + 0.0013862550258636475, + 0.0012425545137375593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.20157881081104279 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a clear plastic ruler with black millimeter markings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting against the rough, dark edge of a wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by dramatic, high-contrast sidelighting from the left creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ruler released and shown rebounding with explosive force in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ruler's body contort into rapid waves of vibration that blur into oscillating arcs before coming to a complete stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058534324169159 + ], + "opposite_logprob_tokens": [ + -3.895585298538208 + ], + "logprob_reward": -0.02058534324169159, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058534324169159, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01760263554751873, + 0.11368905007839203, + 0.01990106701850891, + 0.05251424014568329, + 0.6225636601448059, + 0.05124947801232338, + 0.0617637075483799, + 0.019138559699058533, + 0.0112311327829957, + 0.030346404761075974 + ], + "credit_map": [ + 0.0028692372143268585, + 0.0014123539440333843, + 0.00280586164444685, + 0.0015045532491058111, + 0.0011534825898706913, + 0.0017175422981381416, + 0.003016887465491891, + 0.003122717374935746, + 0.03003847226500511, + 0.007951969280838966, + 0.004226159770041704, + 0.0027749575674533844, + 0.04682609438896179, + 0.015958135947585106, + 0.003287714207544923, + 0.0026255485136061907, + 0.0017127387691289186, + 0.005429194774478674, + 0.0012087816139683127, + 0.0012762334663420916, + 0.003178128506988287, + 0.003746513044461608, + 0.001229376532137394, + 0.002120101125910878, + 0.002248253906145692, + 0.02003144659101963, + 0.00241269008256495, + 0.0033071893267333508, + 0.005295539274811745, + 0.014852328225970268, + 0.0024238196201622486, + 0.0019429732346907258, + 0.020442413166165352, + 0.2807391881942749, + 0.007001128047704697, + 0.009098843671381474, + 0.03061509318649769, + 0.26318445801734924, + 0.004385279957205057, + 0.007097269408404827, + 0.0029429139103740454, + 0.015327339060604572, + 0.005162667483091354, + 0.003688520984724164, + 0.00478783855214715, + 0.012515508569777012, + 0.004065418150275946, + 0.002759271766990423, + 0.0021930322982370853, + 0.02480766735970974, + 0.0022973401937633753, + 0.001780397491529584, + 0.0034289031755179167, + 0.023760227486491203, + 0.0014107312308624387, + 0.002085411921143532, + 0.0008938759565353394, + 0.005871225148439407, + 0.0012703185202553868, + 0.0009181557688862085, + 0.0013431943953037262, + 0.005830566864460707, + 0.0013750824145972729, + 0.001636141212657094, + 0.0009488931391388178, + 0.003051779232919216, + 0.0009995766449719667, + 0.0011907275766134262, + 0.0017055287025868893, + 0.0016450188122689724, + 0.0009908354841172695, + 0.0006987715605646372, + 0.0019376061391085386, + 0.007819636724889278, + 0.0026419369969516993, + 0.0025913529098033905, + 0.002601939020678401, + 0.006580053362995386, + 0.0033016828820109367, + 0.0028721969574689865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 6 + ], + "selected_reward": -0.02058534324169159 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot captures the scene from a sideways, eye-level angle: the end of a clear plastic ruler, its transparent surface showing subtle wear and precise black millimeter markings, rests firmly against the rough, dark edge of a wooden workbench. Dramatic, high-contrast sidelighting from the left casts a bright, sharp highlight along the ruler's curved shape, while casting the surrounding area into deep shadow, highlighting the wood's grain. Suddenly, the ruler is released and rebounds with explosive force. Captured in ultra-slow motion, its body contorts into rapid waves of vibration, blurring into oscillating arcs that repeatedly strike the wood surface with sharp impacts. The violent vibrations gradually subside, and the movement slows until the ruler comes to a complete stillness, its clear form once again sharp, still against the wood, creating an atmosphere of tension and intense concentration, a controlled release of physical energy.", + "seed": 28029, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_9/index_3232_seed_28029_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1232971167191863, + "vlm_logprob": -0.1232971167191863, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.1232971167191863, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a clear plastic ruler with black millimeter markings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting against the rough, dark edge of a wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by dramatic, high-contrast sidelighting from the left creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ruler released and shown rebounding with explosive force in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ruler's body contort into rapid waves of vibration that blur into oscillating arcs before coming to a complete stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.032284025102853775 + ], + "opposite_logprob_tokens": [ + -5.157284259796143 + ], + "logprob_reward": -0.032284025102853775, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.032284025102853775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15521195530891418, + 0.03046070784330368, + 0.09722615778446198, + 0.31415629386901855, + 0.048368312418460846, + 0.13308511674404144, + 0.057586669921875, + 0.0474487729370594, + 0.0597209557890892, + 0.05673506483435631 + ], + "credit_map": [ + 0.031086115166544914, + 0.04665464535355568, + 0.00858292356133461, + 0.006380056496709585, + 0.021155906841158867, + 0.0069821919314563274, + 0.010546136647462845, + 0.023823976516723633, + 0.0074438839219510555, + 0.006905704271048307, + 0.003298660274595022, + 0.0014067032607272267, + 0.004287186078727245, + 0.0021841395646333694, + 0.0026021196972578764, + 0.0023323087953031063, + 0.011580861173570156, + 0.004347741603851318, + 0.03243274986743927, + 0.0046673910692334175, + 0.011187269352376461, + 0.007054960820823908, + 0.0038624086882919073, + 0.022092774510383606, + 0.10584649443626404, + 0.027291100472211838, + 0.10676545649766922, + 0.018492477014660835, + 0.008778559975326061, + 0.014008785597980022, + 0.01046362891793251, + 0.022509779781103134, + 0.005663122516125441, + 0.007850845344364643, + 0.005489193368703127, + 0.006510732229799032, + 0.007671143859624863, + 0.0049057272262871265, + 0.005922913551330566, + 0.004354633390903473, + 0.015969419851899147, + 0.017592281103134155, + 0.01943986862897873, + 0.009944391436874866, + 0.017991840839385986, + 0.028370823711156845, + 0.013024945743381977, + 0.010751538909971714, + 0.005716775078326464, + 0.010545937344431877, + 0.004384064581245184, + 0.005186096299439669, + 0.010674960911273956, + 0.006332502234727144, + 0.007190813776105642, + 0.007555521093308926, + 0.00445817643776536, + 0.00600025849416852, + 0.004471166525036097, + 0.005722477566450834, + 0.006325206253677607, + 0.010027137584984303, + 0.005778050981462002, + 0.004666298162192106, + 0.007747952360659838, + 0.012832763604819775, + 0.006587306037545204, + 0.005569609813392162, + 0.010187194682657719, + 0.007896924391388893, + 0.003958918619900942, + 0.0049402848817408085, + 0.006122989114373922, + 0.008903141133487225, + 0.006656965706497431, + 0.005921678151935339, + 0.010117817670106888, + 0.006454452406615019, + 0.006929932162165642, + 0.005628093145787716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.032284025102853775 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a clear plastic ruler with black millimeter markings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting against the rough, dark edge of a wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by dramatic, high-contrast sidelighting from the left creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ruler released and shown rebounding with explosive force in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ruler's body contort into rapid waves of vibration that blur into oscillating arcs before coming to a complete stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014231263659894466 + ], + "opposite_logprob_tokens": [ + -4.264231204986572 + ], + "logprob_reward": -0.014231263659894466, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014231263659894466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23443645238876343, + 0.04158319905400276, + 0.11518484354019165, + 0.09252788126468658, + 0.14844441413879395, + 0.10718392580747604, + 0.061351463198661804, + 0.09739620983600616, + 0.03803958371281624, + 0.06385204195976257 + ], + "credit_map": [ + 0.0660710260272026, + 0.09751911461353302, + 0.011556986719369888, + 0.004877907689660788, + 0.0352821908891201, + 0.0065512750297784805, + 0.004327510483562946, + 0.00825045257806778, + 0.006690262816846371, + 0.00699553731828928, + 0.005807084031403065, + 0.0030980852898210287, + 0.006046546623110771, + 0.0048797582276165485, + 0.0037180257495492697, + 0.004347898997366428, + 0.013415218330919743, + 0.0062471237033605576, + 0.03748954460024834, + 0.005753508768975735, + 0.007050197571516037, + 0.00864951591938734, + 0.004298526793718338, + 0.03228120878338814, + 0.016266977414488792, + 0.017974600195884705, + 0.02059592306613922, + 0.010727874003350735, + 0.006052684038877487, + 0.0056999153457582, + 0.005257011391222477, + 0.009952896274626255, + 0.014043347910046577, + 0.017033951357007027, + 0.03216171637177467, + 0.02155897580087185, + 0.018419427797198296, + 0.019895780831575394, + 0.013189892284572124, + 0.012141325511038303, + 0.012305453419685364, + 0.017829226329922676, + 0.00792263075709343, + 0.009859599173069, + 0.022625120356678963, + 0.012549321167171001, + 0.014013291336596012, + 0.010079273954033852, + 0.006553609389811754, + 0.007703397888690233, + 0.014121330343186855, + 0.007087265606969595, + 0.006756606511771679, + 0.008642425760626793, + 0.005656760185956955, + 0.004830068442970514, + 0.010483356192708015, + 0.01389357727020979, + 0.013832156546413898, + 0.008283577859401703, + 0.019671037793159485, + 0.014891081489622593, + 0.00955091044306755, + 0.006790510378777981, + 0.004645771346986294, + 0.004101516678929329, + 0.0069931885227561, + 0.0038685493636876345, + 0.005156650207936764, + 0.006698122248053551, + 0.0030035434756428003, + 0.0035722460597753525, + 0.006220623850822449, + 0.009526711888611317, + 0.008965948596596718, + 0.006397434510290623, + 0.010022848844528198, + 0.009743975475430489, + 0.006185002159327269, + 0.006789499428123236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.014231263659894466 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a clear plastic ruler with black millimeter markings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting against the rough, dark edge of a wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by dramatic, high-contrast sidelighting from the left creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ruler released and shown rebounding with explosive force in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ruler's body contort into rapid waves of vibration that blur into oscillating arcs before coming to a complete stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023316258564591408 + ], + "opposite_logprob_tokens": [ + -3.7733161449432373 + ], + "logprob_reward": -0.023316258564591408, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023316258564591408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1014159545302391, + 0.040725890547037125, + 0.13173457980155945, + 0.1971193104982376, + 0.10699392855167389, + 0.14641577005386353, + 0.06878525763750076, + 0.08913712948560715, + 0.05345365032553673, + 0.06421847641468048 + ], + "credit_map": [ + 0.01622699201107025, + 0.02373528853058815, + 0.013483012095093727, + 0.0067506227642297745, + 0.016670987010002136, + 0.011240769177675247, + 0.0056621539406478405, + 0.007646133191883564, + 0.006360913626849651, + 0.005402958486229181, + 0.006207753904163837, + 0.0034350031055510044, + 0.005692918784916401, + 0.004132031928747892, + 0.0041411267593503, + 0.005353187210857868, + 0.020366735756397247, + 0.013168214820325375, + 0.01535126008093357, + 0.017959831282496452, + 0.01428268849849701, + 0.022776618599891663, + 0.010316607542335987, + 0.017512639984488487, + 0.039447516202926636, + 0.037515588104724884, + 0.03370094299316406, + 0.02098681777715683, + 0.01489571575075388, + 0.01526032853871584, + 0.01342863216996193, + 0.02188379317522049, + 0.01164541207253933, + 0.023206613957881927, + 0.01792289689183235, + 0.010507572442293167, + 0.012147489003837109, + 0.011331209912896156, + 0.009534423239529133, + 0.010698319412767887, + 0.01481939759105444, + 0.024605853483080864, + 0.01232107076793909, + 0.013110639527440071, + 0.029852889478206635, + 0.017517438158392906, + 0.018944784998893738, + 0.015243705362081528, + 0.007314153015613556, + 0.01222825888544321, + 0.009644370526075363, + 0.005588610656559467, + 0.009698786772787571, + 0.011554025113582611, + 0.006260501686483622, + 0.0064965588971972466, + 0.011376469396054745, + 0.016097277402877808, + 0.007011704612523317, + 0.007686796598136425, + 0.019588559865951538, + 0.010875500738620758, + 0.00930225569754839, + 0.007198571693152189, + 0.005519454833120108, + 0.004918999038636684, + 0.011715099215507507, + 0.005879928823560476, + 0.005881131626665592, + 0.009043121710419655, + 0.005869400687515736, + 0.004626516252756119, + 0.006322288885712624, + 0.009624601341784, + 0.008898326195776463, + 0.006142076104879379, + 0.012030115351080894, + 0.007060613948851824, + 0.007552719675004482, + 0.006587738636881113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.023316258564591408 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a clear plastic ruler with black millimeter markings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting against the rough, dark edge of a wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by dramatic, high-contrast sidelighting from the left creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ruler released and shown rebounding with explosive force in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ruler's body contort into rapid waves of vibration that blur into oscillating arcs before coming to a complete stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233553647994995 + ], + "opposite_logprob_tokens": [ + -0.8983553647994995 + ], + "logprob_reward": -0.5233553647994995, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233553647994995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34459570050239563, + 0.01648828200995922, + 0.09472031146287918, + 0.18398341536521912, + 0.03821367770433426, + 0.06012579798698425, + 0.07034502923488617, + 0.11536555737257004, + 0.052167858928442, + 0.02399427257478237 + ], + "credit_map": [ + 0.1000082716345787, + 0.14816439151763916, + 0.012699028477072716, + 0.006140485405921936, + 0.05244089290499687, + 0.008510603569447994, + 0.005843767896294594, + 0.010788291692733765, + 0.0028434586711227894, + 0.0035357296001166105, + 0.001912405015900731, + 0.0012501634191721678, + 0.0022336840629577637, + 0.0015321800019592047, + 0.0017498264787718654, + 0.0014308361569419503, + 0.016594093292951584, + 0.010317141190171242, + 0.013181212358176708, + 0.012708425521850586, + 0.008401736617088318, + 0.011776572093367577, + 0.007089389953762293, + 0.014651751145720482, + 0.04714338853955269, + 0.032464951276779175, + 0.04089966416358948, + 0.017120666801929474, + 0.01080398727208376, + 0.008936691097915173, + 0.010667812079191208, + 0.015946270897984505, + 0.004842132795602083, + 0.006762648932635784, + 0.003663562936708331, + 0.0036616483703255653, + 0.006395487114787102, + 0.006510793697088957, + 0.0034923271741718054, + 0.0028850806411355734, + 0.007048836909234524, + 0.006931247189640999, + 0.00886493269354105, + 0.005097787827253342, + 0.00849401205778122, + 0.012608706019818783, + 0.00651191733777523, + 0.004568362142890692, + 0.008506289683282375, + 0.01231673639267683, + 0.0061169760301709175, + 0.005685894284397364, + 0.01161750964820385, + 0.010334563441574574, + 0.008322712033987045, + 0.0074443519115448, + 0.012524711899459362, + 0.01889508031308651, + 0.007964830845594406, + 0.01008633803576231, + 0.023497648537158966, + 0.01729043573141098, + 0.014149713329970837, + 0.010956813581287861, + 0.004874985199421644, + 0.008500151336193085, + 0.006825474090874195, + 0.005438563413918018, + 0.007864262908697128, + 0.007270319387316704, + 0.006003446411341429, + 0.005390659440308809, + 0.0022119509521871805, + 0.0038318352308124304, + 0.0020113387145102024, + 0.0022140173241496086, + 0.005032981745898724, + 0.002806569216772914, + 0.0032128659076988697, + 0.0026727153453975916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.5233553647994995 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a clear plastic ruler with black millimeter markings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting against the rough, dark edge of a wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by dramatic, high-contrast sidelighting from the left creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ruler released and shown rebounding with explosive force in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ruler's body contort into rapid waves of vibration that blur into oscillating arcs before coming to a complete stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329867146909237 + ], + "opposite_logprob_tokens": [ + -3.773298740386963 + ], + "logprob_reward": -0.02329867146909237, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329867146909237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25781792402267456, + 0.02576012723147869, + 0.10731387883424759, + 0.07490811496973038, + 0.07910683751106262, + 0.0726858600974083, + 0.14440622925758362, + 0.1306861788034439, + 0.07483337819576263, + 0.03248143196105957 + ], + "credit_map": [ + 0.06873658299446106, + 0.10849660634994507, + 0.009581172838807106, + 0.005091536790132523, + 0.04070020467042923, + 0.007351615000516176, + 0.0062403553165495396, + 0.011619887314736843, + 0.004291963763535023, + 0.005251826252788305, + 0.0031336445827037096, + 0.0015342566184699535, + 0.0034770467318594456, + 0.0026392568834125996, + 0.0030450448393821716, + 0.0023870887234807014, + 0.016046717762947083, + 0.006286736112087965, + 0.028650915250182152, + 0.007013717666268349, + 0.01535766664892435, + 0.010828285478055477, + 0.004999573342502117, + 0.018130283802747726, + 0.018442895263433456, + 0.009477458894252777, + 0.02051413245499134, + 0.007222321815788746, + 0.0036933592054992914, + 0.005203078966587782, + 0.00413799611851573, + 0.0062168799340724945, + 0.00975426658987999, + 0.025152109563350677, + 0.010658368468284607, + 0.004967277869582176, + 0.014710911549627781, + 0.004541149828583002, + 0.005124363582581282, + 0.004198394250124693, + 0.00872946809977293, + 0.011470659635961056, + 0.006553152576088905, + 0.006805716082453728, + 0.012740220874547958, + 0.008970016613602638, + 0.010650914162397385, + 0.006765720434486866, + 0.015019973739981651, + 0.02823401615023613, + 0.019380224868655205, + 0.011715203523635864, + 0.022968526929616928, + 0.02012771926820278, + 0.013015695847570896, + 0.013944889418780804, + 0.015681734308600426, + 0.024682676419615746, + 0.008028394542634487, + 0.011634906753897667, + 0.029765795916318893, + 0.016544964164495468, + 0.014154459349811077, + 0.010193263180553913, + 0.006600256077945232, + 0.006722758524119854, + 0.01634337194263935, + 0.008516354486346245, + 0.010358057916164398, + 0.01086349319666624, + 0.009308977983891964, + 0.0061201113276183605, + 0.004049844574183226, + 0.006068017333745956, + 0.0037291287444531918, + 0.002832351950928569, + 0.005697143264114857, + 0.00447717122733593, + 0.002402359852567315, + 0.0032254192046821117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.02329867146909237 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot captures the scene from a sideways, eye-level angle: the end of a clear plastic ruler, its transparent surface showing subtle wear and precise black millimeter markings, rests firmly against the rough, dark edge of a wooden workbench. Dramatic, high-contrast sidelighting from the left casts a bright, sharp highlight along the ruler's curved shape, while casting the surrounding area into deep shadow, highlighting the wood's grain. Suddenly, the ruler is released and rebounds with explosive force. Captured in ultra-slow motion, its body contorts into rapid waves of vibration, blurring into oscillating arcs that repeatedly strike the wood surface with sharp impacts. The violent vibrations gradually subside, and the movement slows until the ruler comes to a complete stillness, its clear form once again sharp, still against the wood, creating an atmosphere of tension and intense concentration, a controlled release of physical energy.", + "seed": 28030, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_9/index_3232_seed_28030_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1517843198031187, + "vlm_logprob": -0.1517843198031187, + "vlm_yes_no_margin": 2.4, + "vlm_token_credit": -0.1517843198031187, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a clear plastic ruler with black millimeter markings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting against the rough, dark edge of a wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by dramatic, high-contrast sidelighting from the left creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ruler released and shown rebounding with explosive force in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ruler's body contort into rapid waves of vibration that blur into oscillating arcs before coming to a complete stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5292124152183533 + ], + "opposite_logprob_tokens": [ + -0.9042124152183533 + ], + "logprob_reward": -0.5292124152183533, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5292124152183533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15470193326473236, + 0.09788062423467636, + 0.08730249851942062, + 0.08966394513845444, + 0.06368442624807358, + 0.06764905154705048, + 0.2548169493675232, + 0.05401945486664772, + 0.02188807725906372, + 0.10839305818080902 + ], + "credit_map": [ + 0.015414498746395111, + 0.038690801709890366, + 0.007827578112483025, + 0.00816235039383173, + 0.018773803487420082, + 0.0334102138876915, + 0.0177304744720459, + 0.014692212454974651, + 0.010109743103384972, + 0.009509301744401455, + 0.010824819095432758, + 0.00951540470123291, + 0.018540140241384506, + 0.0115805147215724, + 0.01391934510320425, + 0.013881352730095387, + 0.008215299807488918, + 0.017346519976854324, + 0.005878000054508448, + 0.007262370549142361, + 0.014083066023886204, + 0.01056066993623972, + 0.012409919872879982, + 0.011546649038791656, + 0.008203675970435143, + 0.007345519028604031, + 0.007948058657348156, + 0.006090258248150349, + 0.009855958633124828, + 0.015271496959030628, + 0.020116720348596573, + 0.014832250773906708, + 0.006188568659126759, + 0.004924994893372059, + 0.006335839629173279, + 0.006601246539503336, + 0.0111982561647892, + 0.012059811502695084, + 0.009548536501824856, + 0.00682717002928257, + 0.004153790418058634, + 0.004823313094675541, + 0.0025127858389168978, + 0.002675585215911269, + 0.00795317254960537, + 0.024860695004463196, + 0.004030753392726183, + 0.016638953238725662, + 0.008508452214300632, + 0.009115944616496563, + 0.002839518478140235, + 0.0019573974423110485, + 0.027275245636701584, + 0.11649781465530396, + 0.008014887571334839, + 0.08060767501592636, + 0.0019312838558107615, + 0.0020346748642623425, + 0.0011039760429412127, + 0.000831127050332725, + 0.006116911303251982, + 0.023228289559483528, + 0.00231677433475852, + 0.016456415876746178, + 0.0028577055782079697, + 0.003155126003548503, + 0.0016833972185850143, + 0.001647396944463253, + 0.004250812344253063, + 0.0028286681044846773, + 0.002646749373525381, + 0.0028182188980281353, + 0.005423200316727161, + 0.006790978834033012, + 0.004762241151183844, + 0.004053540527820587, + 0.011581405997276306, + 0.039153385907411575, + 0.006368630100041628, + 0.030259663239121437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.5292124152183533 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a clear plastic ruler with black millimeter markings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting against the rough, dark edge of a wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by dramatic, high-contrast sidelighting from the left creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ruler released and shown rebounding with explosive force in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ruler's body contort into rapid waves of vibration that blur into oscillating arcs before coming to a complete stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048741865903139114 + ], + "opposite_logprob_tokens": [ + -3.0487418174743652 + ], + "logprob_reward": -0.048741865903139114, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048741865903139114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1412222981452942, + 0.09144500643014908, + 0.05462869256734848, + 0.029669394716620445, + 0.04778089374303818, + 0.11307022720575333, + 0.2671111226081848, + 0.09439053386449814, + 0.07093500345945358, + 0.08974678814411163 + ], + "credit_map": [ + 0.015946082770824432, + 0.03928240016102791, + 0.006089671049267054, + 0.007370726205408573, + 0.02251248247921467, + 0.02362561598420143, + 0.01490591000765562, + 0.011489417403936386, + 0.011177612468600273, + 0.008940253406763077, + 0.01162587758153677, + 0.011744567193090916, + 0.009724984876811504, + 0.010988344438374043, + 0.013605514541268349, + 0.013637865893542767, + 0.005695664323866367, + 0.009915352798998356, + 0.004886869341135025, + 0.004790667910128832, + 0.008104158565402031, + 0.005975379608571529, + 0.007816669531166553, + 0.0074439323507249355, + 0.003086245618760586, + 0.0038061998784542084, + 0.002278407569974661, + 0.002248707925900817, + 0.0038563006091862917, + 0.0049210176803171635, + 0.0041928221471607685, + 0.005279696546494961, + 0.005539970006793737, + 0.004563090391457081, + 0.005397227592766285, + 0.005493448581546545, + 0.007609001826494932, + 0.006556441076099873, + 0.007069661747664213, + 0.005552058108150959, + 0.005179859232157469, + 0.005787597503513098, + 0.002194550819694996, + 0.0026188078336417675, + 0.011050867848098278, + 0.04744106903672218, + 0.004958384204655886, + 0.03383910283446312, + 0.008895114995539188, + 0.009240129962563515, + 0.003111329162493348, + 0.001493395189754665, + 0.028661837801337242, + 0.12216313183307648, + 0.008659418672323227, + 0.08488677442073822, + 0.004761406220495701, + 0.00527174910530448, + 0.002654685638844967, + 0.002704424085095525, + 0.011182097718119621, + 0.03585555776953697, + 0.005679586436599493, + 0.02628103457391262, + 0.0027469436172395945, + 0.0029618593398481607, + 0.0015549593372270465, + 0.0016040368936955929, + 0.007681389804929495, + 0.02950989454984665, + 0.0034859769511967897, + 0.021389950066804886, + 0.005798692815005779, + 0.006922242231667042, + 0.0035248524509370327, + 0.0035873902961611748, + 0.010836325585842133, + 0.03000727668404579, + 0.0059225899167358875, + 0.023147424682974815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.048741865903139114 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a clear plastic ruler with black millimeter markings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting against the rough, dark edge of a wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by dramatic, high-contrast sidelighting from the left creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ruler released and shown rebounding with explosive force in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ruler's body contort into rapid waves of vibration that blur into oscillating arcs before coming to a complete stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062086936086416245 + ], + "opposite_logprob_tokens": [ + -2.812086820602417 + ], + "logprob_reward": -0.062086936086416245, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062086936086416245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19682128727436066, + 0.05525621771812439, + 0.04253305867314339, + 0.08792375028133392, + 0.16312754154205322, + 0.14800333976745605, + 0.164862260222435, + 0.0323980450630188, + 0.0544305145740509, + 0.05464402586221695 + ], + "credit_map": [ + 0.006694526877254248, + 0.012639997527003288, + 0.01102725975215435, + 0.009197726845741272, + 0.01883266679942608, + 0.07700075209140778, + 0.01288659404963255, + 0.048541758209466934, + 0.005801206920295954, + 0.004946700297296047, + 0.007050906307995319, + 0.005835480988025665, + 0.007022445555776358, + 0.0056006330996751785, + 0.008520659990608692, + 0.010478181764483452, + 0.0030978810973465443, + 0.006654928904026747, + 0.004177895374596119, + 0.0038272642996162176, + 0.006795988883823156, + 0.005282048135995865, + 0.0070098224096000195, + 0.00568723026663065, + 0.006142815109342337, + 0.006520148832350969, + 0.007206737529486418, + 0.005024739075452089, + 0.00914816465228796, + 0.019080182537436485, + 0.01965651474893093, + 0.015144439414143562, + 0.016047973185777664, + 0.015258507803082466, + 0.016865545883774757, + 0.0153671158477664, + 0.03045007586479187, + 0.02603480964899063, + 0.025328656658530235, + 0.017774848267436028, + 0.004364694468677044, + 0.0076302126981318, + 0.002151911146938801, + 0.0025428340304642916, + 0.015007641166448593, + 0.06512568891048431, + 0.005751612596213818, + 0.045428745448589325, + 0.0070080021396279335, + 0.006332704331725836, + 0.003943335264921188, + 0.002565887989476323, + 0.018643205985426903, + 0.06969107687473297, + 0.007819080725312233, + 0.048858966678380966, + 0.00196625548414886, + 0.0016013047425076365, + 0.0015616848831996322, + 0.0011554454686120152, + 0.0035595460794866085, + 0.011669727973639965, + 0.002339462051168084, + 0.0085446173325181, + 0.0037745924200862646, + 0.004475359804928303, + 0.0025013384874910116, + 0.002254160586744547, + 0.007534765172749758, + 0.015845734626054764, + 0.005023491103202105, + 0.013021075166761875, + 0.0031389507930725813, + 0.003030222374945879, + 0.003575310343876481, + 0.0030822649132460356, + 0.005441701039671898, + 0.018219053745269775, + 0.004559084307402372, + 0.013597437180578709 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.062086936086416245 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a clear plastic ruler with black millimeter markings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting against the rough, dark edge of a wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by dramatic, high-contrast sidelighting from the left creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ruler released and shown rebounding with explosive force in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ruler's body contort into rapid waves of vibration that blur into oscillating arcs before coming to a complete stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906007558107376 + ], + "opposite_logprob_tokens": [ + -2.464060068130493 + ], + "logprob_reward": -0.08906007558107376, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906007558107376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19601626694202423, + 0.1348138451576233, + 0.12438136339187622, + 0.020943285897374153, + 0.05152793228626251, + 0.18824034929275513, + 0.09888968616724014, + 0.0870637595653534, + 0.048874225467443466, + 0.04924933984875679 + ], + "credit_map": [ + 0.010288015007972717, + 0.015440590679645538, + 0.010456155985593796, + 0.010181526653468609, + 0.020743083208799362, + 0.07601744681596756, + 0.01641026698052883, + 0.03647919371724129, + 0.012258305214345455, + 0.012859568931162357, + 0.010341443121433258, + 0.010105391032993793, + 0.025749890133738518, + 0.022225577384233475, + 0.019370362162590027, + 0.021903319284319878, + 0.014538773335516453, + 0.026440367102622986, + 0.007195084355771542, + 0.010800834745168686, + 0.01632492057979107, + 0.013581572100520134, + 0.018547117710113525, + 0.01695268787443638, + 0.0027571488171815872, + 0.002207729732617736, + 0.0017033962067216635, + 0.0017700023017823696, + 0.002872266573831439, + 0.0033702384680509567, + 0.0030951034277677536, + 0.003167400835081935, + 0.0046937777660787106, + 0.004099383484572172, + 0.005217785947024822, + 0.004456766415387392, + 0.007016780320554972, + 0.01156616397202015, + 0.005822630599141121, + 0.008654646575450897, + 0.007309213280677795, + 0.008596531115472317, + 0.0034785985480993986, + 0.0037044170312583447, + 0.01829335279762745, + 0.08189843595027924, + 0.007438424043357372, + 0.057521361857652664, + 0.0032300392631441355, + 0.004297728184610605, + 0.0017686164937913418, + 0.0014969812473282218, + 0.010645314119756222, + 0.043148983269929886, + 0.00406484492123127, + 0.030237173661589622, + 0.0034529271069914103, + 0.0039605689235031605, + 0.0018689408898353577, + 0.001822274993173778, + 0.00978242326527834, + 0.03564518317580223, + 0.0043819197453558445, + 0.026149526238441467, + 0.00313216308131814, + 0.0023695409763604403, + 0.0015289461007341743, + 0.0011207138886675239, + 0.004899145103991032, + 0.01901870034635067, + 0.0038205075543373823, + 0.012984505854547024, + 0.002764694392681122, + 0.002702629892155528, + 0.0019607646390795708, + 0.0018500564619898796, + 0.00600453419610858, + 0.017302434891462326, + 0.0032886238768696785, + 0.013375600799918175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.08906007558107376 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a clear plastic ruler with black millimeter markings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting against the rough, dark edge of a wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by dramatic, high-contrast sidelighting from the left creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ruler released and shown rebounding with explosive force in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ruler's body contort into rapid waves of vibration that blur into oscillating arcs before coming to a complete stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029820306226611137 + ], + "opposite_logprob_tokens": [ + -3.529820203781128 + ], + "logprob_reward": -0.029820306226611137, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029820306226611137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08835427463054657, + 0.06227238103747368, + 0.10123367607593536, + 0.05511828884482384, + 0.11732486635446548, + 0.23480281233787537, + 0.17602720856666565, + 0.04649920389056206, + 0.07411408424377441, + 0.04425325617194176 + ], + "credit_map": [ + 0.005764992907643318, + 0.006922460161149502, + 0.007396204397082329, + 0.005606438033282757, + 0.011483711190521717, + 0.02394256927073002, + 0.008413530886173248, + 0.01882435381412506, + 0.00683539267629385, + 0.006486210506409407, + 0.0058020176365971565, + 0.00557423522695899, + 0.010207626968622208, + 0.00860651209950447, + 0.009334814734756947, + 0.00942557118833065, + 0.009015093557536602, + 0.020115099847316742, + 0.007642178796231747, + 0.00912848673760891, + 0.015589691698551178, + 0.012997591868042946, + 0.0138130197301507, + 0.012932498008012772, + 0.004790866747498512, + 0.004103310406208038, + 0.0028453106060624123, + 0.003382124938070774, + 0.007281881291419268, + 0.011942851357161999, + 0.01020857971161604, + 0.010563362389802933, + 0.008850513026118279, + 0.007181266322731972, + 0.009844641201198101, + 0.013094485737383366, + 0.01711275987327099, + 0.027638662606477737, + 0.016069261357188225, + 0.01753327250480652, + 0.009853742085397243, + 0.011696217581629753, + 0.0056637986563146114, + 0.005746558308601379, + 0.022927748039364815, + 0.09836739301681519, + 0.010258514434099197, + 0.07028883695602417, + 0.006315215490758419, + 0.006534446496516466, + 0.0026319283060729504, + 0.0017951359041035175, + 0.018358798697590828, + 0.07974530756473541, + 0.005741644185036421, + 0.05490470677614212, + 0.002219015033915639, + 0.002170265419408679, + 0.0019798551220446825, + 0.0017210915684700012, + 0.0055063385516405106, + 0.017226921394467354, + 0.003579953918233514, + 0.012095760554075241, + 0.00398225337266922, + 0.0031175175681710243, + 0.002147831255570054, + 0.0021386437583714724, + 0.008403092622756958, + 0.029827719554305077, + 0.00384691683575511, + 0.020650101825594902, + 0.0038305034395307302, + 0.003979959990829229, + 0.004169278312474489, + 0.0035583782009780407, + 0.004729624837636948, + 0.01009651180356741, + 0.005063975695520639, + 0.00882501807063818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.029820306226611137 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_0/index_tensor([5288, 5288])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_0/index_tensor([5288, 5288])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..cba643463730136acf01ac65d0b9f5d4e6647ed0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_0/index_tensor([5288, 5288])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, captured at eye level, the camera remains stationary, focusing squarely on the blazing core of a cast-iron furnace. Deep orange flames lick wildly at the glowing charcoal. Suddenly, a fistful of incandescent sparks bursts violently from the flame's center. These tiny, burning particles radiate outward, trailing a brilliant orange trail. Like a swarm of miniature comets, they rush swiftly and chaotically into the deep, nearly lightless darkness surrounding them. They trace brief, graceful arcs across the sky, their light rapidly cooling from incandescent white to red before extinguishing themselves, one by one, within seconds, vanishing into nothingness. The scene is imbued with a sense of primal power, the intense light contrasting with the absolute darkness, creating a fleeting spectacle that is both violent and mesmerizing.", + "seed": 8132, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_0/index_5288_seed_8132_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.2590270911343396, + "vlm_logprob": -0.2590270911343396, + "vlm_yes_no_margin": 2.1499999880790712, + "vlm_token_credit": -0.2590270911343396, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the core of a cast-iron furnace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep orange flames visible licking at glowing charcoal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do incandescent sparks burst violently from the center of the flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sparks cool from incandescent white to red before extinguishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera stationary at eye level throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7062562704086304 + ], + "opposite_logprob_tokens": [ + -0.7062562704086304 + ], + "logprob_reward": -0.7062562704086304, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.7062562704086304, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05503831431269646, + 0.06686811149120331, + 0.04545082151889801, + 0.07310610264539719, + 0.21629779040813446, + 0.1710386723279953, + 0.09508126974105835, + 0.16777342557907104, + 0.08225001394748688, + 0.0270954892039299 + ], + "credit_map": [ + 0.009244880639016628, + 0.012981520034372807, + 0.008663213811814785, + 0.0051389653235673904, + 0.005279745906591415, + 0.004430894739925861, + 0.004470274318009615, + 0.004828817211091518, + 0.011907418258488178, + 0.006541091948747635, + 0.008863837458193302, + 0.008321717381477356, + 0.006237877998501062, + 0.004953894764184952, + 0.014728925190865993, + 0.005313341971486807, + 0.004419307690113783, + 0.008207274600863457, + 0.005535792093724012, + 0.006116863340139389, + 0.005084525793790817, + 0.006490957457572222, + 0.005409232806414366, + 0.00418686680495739, + 0.01593349687755108, + 0.012994862161576748, + 0.018696069717407227, + 0.005757465958595276, + 0.0043783001601696014, + 0.004436134360730648, + 0.005183668807148933, + 0.005726102739572525, + 0.02952224761247635, + 0.02632603421807289, + 0.026335610076785088, + 0.03299838677048683, + 0.021165935322642326, + 0.03482022136449814, + 0.019889555871486664, + 0.02523978240787983, + 0.022874703630805016, + 0.02752157486975193, + 0.014656487852334976, + 0.01753632165491581, + 0.03171175345778465, + 0.01732228882610798, + 0.027888478711247444, + 0.011527053080499172, + 0.009870261885225773, + 0.010964234359562397, + 0.012329109013080597, + 0.011995814740657806, + 0.01595926284790039, + 0.012591434642672539, + 0.014018315821886063, + 0.007352837827056646, + 0.018065446987748146, + 0.01839694380760193, + 0.020557954907417297, + 0.016679730266332626, + 0.016947362571954727, + 0.032594069838523865, + 0.01582362689077854, + 0.02870827168226242, + 0.008558496832847595, + 0.008664663881063461, + 0.008935066871345043, + 0.007860200479626656, + 0.01941845379769802, + 0.010747394524514675, + 0.008584355935454369, + 0.009481383487582207, + 0.0026753200218081474, + 0.0024243637453764677, + 0.004736713599413633, + 0.002151140244677663, + 0.005209899041801691, + 0.004056959878653288, + 0.003120120381936431, + 0.002720970893278718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -0.7062562704086304 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the core of a cast-iron furnace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep orange flames visible licking at glowing charcoal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do incandescent sparks burst violently from the center of the flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sparks cool from incandescent white to red before extinguishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera stationary at eye level throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004668171051889658 + ], + "opposite_logprob_tokens": [ + -5.379668235778809 + ], + "logprob_reward": -0.004668171051889658, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004668171051889658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08443869650363922, + 0.09743195027112961, + 0.05225274711847305, + 0.06445116549730301, + 0.1891484409570694, + 0.1496346890926361, + 0.12052560597658157, + 0.1356986165046692, + 0.06530773639678955, + 0.041110359132289886 + ], + "credit_map": [ + 0.011838998645544052, + 0.019916396588087082, + 0.010809589177370071, + 0.008913185447454453, + 0.00883653573691845, + 0.006235430482774973, + 0.007111905142664909, + 0.010776665061712265, + 0.016989832744002342, + 0.009853284806013107, + 0.013343272730708122, + 0.01323570217937231, + 0.009576236829161644, + 0.006937822327017784, + 0.019474200904369354, + 0.008021606132388115, + 0.005759828258305788, + 0.008378921076655388, + 0.00938313640654087, + 0.004707087762653828, + 0.006345702335238457, + 0.006427232176065445, + 0.004755726549774408, + 0.006495118606835604, + 0.012598087079823017, + 0.00843744445592165, + 0.018665501847863197, + 0.005942267365753651, + 0.004673581104725599, + 0.003917854279279709, + 0.0048065404407680035, + 0.005409893114119768, + 0.025710823014378548, + 0.023033253848552704, + 0.02429189160466194, + 0.02914321795105934, + 0.01884693279862404, + 0.029693949967622757, + 0.017344661056995392, + 0.021083729341626167, + 0.02105403132736683, + 0.018924379721283913, + 0.011656706221401691, + 0.017696795985102654, + 0.03152056038379669, + 0.015788152813911438, + 0.022126687690615654, + 0.010867386125028133, + 0.012496850453317165, + 0.014871719293296337, + 0.014565282501280308, + 0.01434212177991867, + 0.021710971370339394, + 0.01586538553237915, + 0.016803337261080742, + 0.009869948960840702, + 0.014048801735043526, + 0.01451045274734497, + 0.01580493152141571, + 0.013656015507876873, + 0.01437418069690466, + 0.02446988970041275, + 0.013247761875391006, + 0.025586599484086037, + 0.006558532360941172, + 0.007400737144052982, + 0.007745936047285795, + 0.006035265047103167, + 0.015324144624173641, + 0.008990570902824402, + 0.00737789086997509, + 0.0058746663853526115, + 0.0035006960388273, + 0.004023455549031496, + 0.006623302586376667, + 0.004060302395373583, + 0.007658061571419239, + 0.005767031107097864, + 0.005289570894092321, + 0.004187940154224634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -0.004668171051889658 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the core of a cast-iron furnace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep orange flames visible licking at glowing charcoal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do incandescent sparks burst violently from the center of the flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sparks cool from incandescent white to red before extinguishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera stationary at eye level throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270747184753418 + ], + "opposite_logprob_tokens": [ + -2.127074718475342 + ], + "logprob_reward": -0.1270747184753418, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270747184753418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23342394828796387, + 0.0748237818479538, + 0.09294744580984116, + 0.09865093976259232, + 0.09286624938249588, + 0.07654200494289398, + 0.11963066458702087, + 0.10550795495510101, + 0.06061809882521629, + 0.04498893395066261 + ], + "credit_map": [ + 0.033809371292591095, + 0.047002214938402176, + 0.03581336513161659, + 0.02591891586780548, + 0.025791732594370842, + 0.018256014212965965, + 0.024533066898584366, + 0.02229926735162735, + 0.012202310375869274, + 0.006680062506347895, + 0.011138071306049824, + 0.008507715538144112, + 0.008219853974878788, + 0.008827726356685162, + 0.011875604279339314, + 0.007372437976300716, + 0.012437636964023113, + 0.011724274605512619, + 0.01971026510000229, + 0.008335045538842678, + 0.010612391866743565, + 0.011896724812686443, + 0.007381505332887173, + 0.010849597863852978, + 0.01673179864883423, + 0.018067944794893265, + 0.024902984499931335, + 0.00832083635032177, + 0.007096755318343639, + 0.00872792024165392, + 0.007281510159373283, + 0.007521185092628002, + 0.014890504069626331, + 0.01252992358058691, + 0.015093880705535412, + 0.01293523795902729, + 0.007024561520665884, + 0.011057735420763493, + 0.010384614579379559, + 0.008949795737862587, + 0.012482581660151482, + 0.009105595760047436, + 0.008480711840093136, + 0.005958987399935722, + 0.01362200640141964, + 0.0077804639004170895, + 0.010972830466926098, + 0.008138828910887241, + 0.011138444766402245, + 0.016834452748298645, + 0.014797651208937168, + 0.0134731475263834, + 0.019993074238300323, + 0.01505882479250431, + 0.014262310229241848, + 0.014072761870920658, + 0.01126253791153431, + 0.013748018071055412, + 0.011934174224734306, + 0.011806335300207138, + 0.011331642046570778, + 0.017279090359807014, + 0.00940712820738554, + 0.018739033490419388, + 0.005521388724446297, + 0.006526170764118433, + 0.007973534055054188, + 0.005831022281199694, + 0.014153305441141129, + 0.00764385936781764, + 0.006808331236243248, + 0.0061604841612279415, + 0.005306670442223549, + 0.004088930319994688, + 0.007411071099340916, + 0.005495953839272261, + 0.005867542698979378, + 0.006580768618732691, + 0.004747994244098663, + 0.005489998497068882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.1270747184753418 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the core of a cast-iron furnace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep orange flames visible licking at glowing charcoal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do incandescent sparks burst violently from the center of the flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sparks cool from incandescent white to red before extinguishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera stationary at eye level throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07006271928548813 + ], + "opposite_logprob_tokens": [ + -2.6950626373291016 + ], + "logprob_reward": -0.07006271928548813, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07006271928548813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12132783979177475, + 0.20814646780490875, + 0.16448649764060974, + 0.06198722869157791, + 0.07907664030790329, + 0.18625657260417938, + 0.036214735358953476, + 0.06984014064073563, + 0.02754964493215084, + 0.04511421546339989 + ], + "credit_map": [ + 0.020387504249811172, + 0.02782348357141018, + 0.020207567140460014, + 0.010095196776092052, + 0.010658739134669304, + 0.009404623880982399, + 0.009629207663238049, + 0.013121511787176132, + 0.02918078936636448, + 0.032881129533052444, + 0.02599327638745308, + 0.023280862718820572, + 0.02294902130961418, + 0.021128643304109573, + 0.03554890304803848, + 0.017183825373649597, + 0.02493792027235031, + 0.032035067677497864, + 0.022327646613121033, + 0.012115361168980598, + 0.01655213162302971, + 0.023426514118909836, + 0.012036486528813839, + 0.021055348217487335, + 0.009870274923741817, + 0.00757514638826251, + 0.01681646890938282, + 0.0050843581557273865, + 0.006208665203303099, + 0.005708398763090372, + 0.005062333308160305, + 0.00566157978028059, + 0.012532709166407585, + 0.00883388426154852, + 0.0118579575791955, + 0.010282890871167183, + 0.007674012333154678, + 0.009247790090739727, + 0.009950626641511917, + 0.008696762844920158, + 0.02575014904141426, + 0.02145667001605034, + 0.01580473780632019, + 0.021396564319729805, + 0.04261399805545807, + 0.016326285898685455, + 0.028457781299948692, + 0.014450372196733952, + 0.005780682433396578, + 0.003830595873296261, + 0.0047281659208238125, + 0.0030123291071504354, + 0.005144256167113781, + 0.004521728493273258, + 0.004500194918364286, + 0.0046967798843979836, + 0.009058916009962559, + 0.00731523847207427, + 0.009131370112299919, + 0.004685520194470882, + 0.00731568643823266, + 0.012925262562930584, + 0.0045073810033500195, + 0.014900760725140572, + 0.0033059804700315, + 0.0038761773612350225, + 0.003342937445268035, + 0.0025938081089407206, + 0.003899744711816311, + 0.004055988043546677, + 0.0034643448889255524, + 0.003010659711435437, + 0.003717054845765233, + 0.004141274373978376, + 0.0072978283278644085, + 0.003896295325830579, + 0.0077749500051140785, + 0.006497681140899658, + 0.00531840231269598, + 0.006470724008977413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 2 + ], + "selected_reward": -0.07006271928548813 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the core of a cast-iron furnace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep orange flames visible licking at glowing charcoal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do incandescent sparks burst violently from the center of the flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sparks cool from incandescent white to red before extinguishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera stationary at eye level throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870735764503479 + ], + "opposite_logprob_tokens": [ + -1.1370735168457031 + ], + "logprob_reward": -0.3870735764503479, + "yes_no_margin_reward": 0.7499999403953552, + "token_credit_reward": -0.3870735764503479, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14509217441082, + 0.1190093606710434, + 0.07162509113550186, + 0.04199553281068802, + 0.047900717705488205, + 0.2560366690158844, + 0.07265134155750275, + 0.10235442221164703, + 0.1121770441532135, + 0.031157633289694786 + ], + "credit_map": [ + 0.02357599325478077, + 0.04569065943360329, + 0.02070828154683113, + 0.010731815360486507, + 0.013975829817354679, + 0.00981395784765482, + 0.009641703218221664, + 0.010953952558338642, + 0.018940631300210953, + 0.013531263917684555, + 0.016342923045158386, + 0.01754070818424225, + 0.012293213978409767, + 0.007868931628763676, + 0.023339392617344856, + 0.00915231741964817, + 0.008363528177142143, + 0.014289787970483303, + 0.011274761520326138, + 0.006511961575597525, + 0.006531702354550362, + 0.010712957940995693, + 0.006580017507076263, + 0.007360383402556181, + 0.006753236521035433, + 0.006366056390106678, + 0.007215216290205717, + 0.004864940885454416, + 0.004492866341024637, + 0.004525431431829929, + 0.004421283956617117, + 0.003356505651026964, + 0.006510825827717781, + 0.007032996043562889, + 0.005460962187498808, + 0.008506456390023232, + 0.0047188811004161835, + 0.004731368273496628, + 0.004298421088606119, + 0.006640810519456863, + 0.035122014582157135, + 0.04405377432703972, + 0.023825621232390404, + 0.02368791028857231, + 0.04181019216775894, + 0.029762517660856247, + 0.04034516215324402, + 0.017429495230317116, + 0.0073618642054498196, + 0.011043745093047619, + 0.008130363188683987, + 0.0068974425084888935, + 0.013438254594802856, + 0.010100280866026878, + 0.0069520012475550175, + 0.008727406151592731, + 0.014884673990309238, + 0.012171563692390919, + 0.012771686539053917, + 0.006539926864206791, + 0.010409570299088955, + 0.02554708532989025, + 0.009449834004044533, + 0.010580088943243027, + 0.014501863159239292, + 0.010989096015691757, + 0.013021320104598999, + 0.01061259862035513, + 0.025086674839258194, + 0.015974529087543488, + 0.010904466733336449, + 0.011086498387157917, + 0.0033371765166521072, + 0.0030346321873366833, + 0.004487949889153242, + 0.0021607116796076298, + 0.005787823814898729, + 0.004649464972317219, + 0.003398643573746085, + 0.004301234148442745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.3870735764503479 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, captured at eye level, the camera remains stationary, focusing squarely on the blazing core of a cast-iron furnace. Deep orange flames lick wildly at the glowing charcoal. Suddenly, a fistful of incandescent sparks bursts violently from the flame's center. These tiny, burning particles radiate outward, trailing a brilliant orange trail. Like a swarm of miniature comets, they rush swiftly and chaotically into the deep, nearly lightless darkness surrounding them. They trace brief, graceful arcs across the sky, their light rapidly cooling from incandescent white to red before extinguishing themselves, one by one, within seconds, vanishing into nothingness. The scene is imbued with a sense of primal power, the intense light contrasting with the absolute darkness, creating a fleeting spectacle that is both violent and mesmerizing.", + "seed": 8133, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_0/index_5288_seed_8133_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.3679335568100214, + "vlm_logprob": -0.3679335568100214, + "vlm_yes_no_margin": 1.925, + "vlm_token_credit": -0.3679335568100214, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the core of a cast-iron furnace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep orange flames visible licking at glowing charcoal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do incandescent sparks burst violently from the center of the flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sparks cool from incandescent white to red before extinguishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera stationary at eye level throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.3232213258743286 + ], + "opposite_logprob_tokens": [ + -0.3232212960720062 + ], + "logprob_reward": -1.3232213258743286, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.3232213258743286, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06862729042768478, + 0.02074131928384304, + 0.025077374652028084, + 0.024586711078882217, + 0.03691887483000755, + 0.03047812171280384, + 0.049710530787706375, + 0.685097336769104, + 0.03155355527997017, + 0.02720896154642105 + ], + "credit_map": [ + 0.006736986804753542, + 0.007360674906522036, + 0.010543151758611202, + 0.009628839790821075, + 0.009488789364695549, + 0.00826937798410654, + 0.010825080797076225, + 0.005774385761469603, + 0.002916815923526883, + 0.002323207911103964, + 0.0018556511495262384, + 0.001860633841715753, + 0.002276973333209753, + 0.004312477540224791, + 0.00215517939068377, + 0.0030403784476220608, + 0.005934092681854963, + 0.0024378784000873566, + 0.0022685728035867214, + 0.0020698225125670433, + 0.002229737350717187, + 0.004336310550570488, + 0.002582923509180546, + 0.003218037774786353, + 0.0062189302407205105, + 0.0014714276185259223, + 0.0025526275858283043, + 0.0017891584429889917, + 0.0017572690267115831, + 0.004124730825424194, + 0.002721745753660798, + 0.003950821235775948, + 0.004136896692216396, + 0.006473738234490156, + 0.004510866943746805, + 0.003879487980157137, + 0.004607042297720909, + 0.004785844124853611, + 0.004401496611535549, + 0.004123501013964415, + 0.0030004519503563643, + 0.002875217702239752, + 0.004315570928156376, + 0.0026226278860121965, + 0.0031172146555036306, + 0.004425219260156155, + 0.004009656608104706, + 0.006112163420766592, + 0.006277939770370722, + 0.009332988411188126, + 0.0056090643629431725, + 0.0043172952719032764, + 0.007298483978956938, + 0.006529869511723518, + 0.005116574466228485, + 0.005228319671005011, + 0.2598300278186798, + 0.05519936978816986, + 0.2547893524169922, + 0.01889820769429207, + 0.01118153240531683, + 0.04751507192850113, + 0.013175885193049908, + 0.02450784668326378, + 0.0033543785102665424, + 0.0074306330643594265, + 0.0038432746659964323, + 0.003064734162762761, + 0.004308977164328098, + 0.003594861598685384, + 0.003051048144698143, + 0.002905646339058876, + 0.0030747661367058754, + 0.0029575731605291367, + 0.004030365031212568, + 0.0022121856454759836, + 0.0037295955698937178, + 0.0030893890652805567, + 0.004624807741492987, + 0.003490276401862502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -1.3232213258743286 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the core of a cast-iron furnace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep orange flames visible licking at glowing charcoal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do incandescent sparks burst violently from the center of the flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sparks cool from incandescent white to red before extinguishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera stationary at eye level throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008668886497616768 + ], + "opposite_logprob_tokens": [ + -4.758668899536133 + ], + "logprob_reward": -0.008668886497616768, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008668886497616768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06748593598604202, + 0.029622456058859825, + 0.0752093717455864, + 0.02324693650007248, + 0.05841894820332527, + 0.03961224853992462, + 0.044355954974889755, + 0.5819310545921326, + 0.03597412630915642, + 0.04414292424917221 + ], + "credit_map": [ + 0.0076942844316363335, + 0.007862183265388012, + 0.007929692044854164, + 0.009306559339165688, + 0.00996301881968975, + 0.008005095645785332, + 0.010169762186706066, + 0.006555344443768263, + 0.004867371637374163, + 0.003511396935209632, + 0.002889834577217698, + 0.0027338622603565454, + 0.003556725336238742, + 0.004480849951505661, + 0.003037356538698077, + 0.004545056726783514, + 0.010694424621760845, + 0.004739914555102587, + 0.005553515162318945, + 0.004142456687986851, + 0.008842610754072666, + 0.018736423924565315, + 0.01107002142816782, + 0.011429998092353344, + 0.005583913065493107, + 0.0019314208766445518, + 0.0021609135437756777, + 0.0017337019089609385, + 0.00265208980999887, + 0.003882740158587694, + 0.0026366894599050283, + 0.002665468957275152, + 0.006162252277135849, + 0.007891975343227386, + 0.006514008156955242, + 0.0066024926491081715, + 0.00873643159866333, + 0.006745574995875359, + 0.009049597196280956, + 0.006716615986078978, + 0.0036513216327875853, + 0.0037141614593565464, + 0.006118267774581909, + 0.002931371796876192, + 0.0031657989602535963, + 0.006093030795454979, + 0.004608898889273405, + 0.009329396300017834, + 0.006098438519984484, + 0.008898001164197922, + 0.004918959923088551, + 0.0036936078686267138, + 0.0061823404394090176, + 0.005986127536743879, + 0.004354584496468306, + 0.004223894793540239, + 0.22023342549800873, + 0.048823095858097076, + 0.21559451520442963, + 0.01529165729880333, + 0.009398876689374447, + 0.03973914310336113, + 0.011476809158921242, + 0.02137354575097561, + 0.0047043669037520885, + 0.0076305498369038105, + 0.0051602693274617195, + 0.003202043240889907, + 0.0037266917061060667, + 0.00473231915384531, + 0.0032198727130889893, + 0.003598012961447239, + 0.0045864698477089405, + 0.003911992534995079, + 0.006010985001921654, + 0.003605398116633296, + 0.00624441122636199, + 0.004902845248579979, + 0.009887643158435822, + 0.004993177950382233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.008668886497616768 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the core of a cast-iron furnace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep orange flames visible licking at glowing charcoal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do incandescent sparks burst violently from the center of the flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sparks cool from incandescent white to red before extinguishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera stationary at eye level throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004182040691376 + ], + "opposite_logprob_tokens": [ + -2.6950418949127197 + ], + "logprob_reward": -0.07004182040691376, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004182040691376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06117857247591019, + 0.02330034412443638, + 0.040892794728279114, + 0.008699599653482437, + 0.021174121648073196, + 0.01548135094344616, + 0.011268243193626404, + 0.794790506362915, + 0.015367424115538597, + 0.00784712191671133 + ], + "credit_map": [ + 0.008049538359045982, + 0.005114586558192968, + 0.014144680462777615, + 0.0053143673576414585, + 0.006017544772475958, + 0.010615506209433079, + 0.0060121603310108185, + 0.005910185165703297, + 0.004137109033763409, + 0.0029288851656019688, + 0.0027944932226091623, + 0.0018786281580105424, + 0.002246470656245947, + 0.0023911839816719294, + 0.0020356846507638693, + 0.00488788727670908, + 0.008153771050274372, + 0.0048799398355185986, + 0.003929500002413988, + 0.0033392722252756357, + 0.002884101355448365, + 0.005930206272751093, + 0.005979032255709171, + 0.005796968936920166, + 0.0018151425756514072, + 0.0006773509085178375, + 0.0009730864549055696, + 0.0006997314631007612, + 0.0006234810571186244, + 0.001511450158432126, + 0.0008098259568214417, + 0.001589530729688704, + 0.002310421084985137, + 0.0020832428708672523, + 0.00212289672344923, + 0.0021344462875276804, + 0.002282969653606415, + 0.004128327593207359, + 0.0028526922687888145, + 0.0032591228373348713, + 0.0019163819961249828, + 0.00132767076138407, + 0.0016714641824364662, + 0.0011571183567866683, + 0.0012105187634006143, + 0.002932407194748521, + 0.0016079987399280071, + 0.0036577910650521517, + 0.0012535307323560119, + 0.002162506338208914, + 0.0012310502352192998, + 0.0013004651991650462, + 0.0013811163371428847, + 0.0014704337809234858, + 0.0011007837019860744, + 0.0013683567522093654, + 0.2885737121105194, + 0.05187833681702614, + 0.31564295291900635, + 0.018195707350969315, + 0.005798906087875366, + 0.08016694337129593, + 0.010535563342273235, + 0.02399832382798195, + 0.0022523810621351004, + 0.001937920693308115, + 0.0023826323449611664, + 0.0017680216114968061, + 0.0014034765772521496, + 0.002284154761582613, + 0.0014472021721303463, + 0.0018916334956884384, + 0.0010249231709167361, + 0.0007141279056668282, + 0.0009675385663285851, + 0.000564676767680794, + 0.0011069074971601367, + 0.001269664615392685, + 0.0009524395572952926, + 0.00124684302136302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.07004182040691376 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the core of a cast-iron furnace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep orange flames visible licking at glowing charcoal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do incandescent sparks burst violently from the center of the flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sparks cool from incandescent white to red before extinguishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera stationary at eye level throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08908362686634064 + ], + "opposite_logprob_tokens": [ + -2.464083671569824 + ], + "logprob_reward": -0.08908362686634064, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08908362686634064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06745413690805435, + 0.01417471095919609, + 0.03926204890012741, + 0.009569077752530575, + 0.020600324496626854, + 0.01936391182243824, + 0.00845988281071186, + 0.8000385165214539, + 0.015488049946725368, + 0.0055892919190227985 + ], + "credit_map": [ + 0.011295581236481667, + 0.008130581118166447, + 0.010386915877461433, + 0.008892222307622433, + 0.00553708802908659, + 0.007688227109611034, + 0.009625363163650036, + 0.005898158065974712, + 0.0020265160128474236, + 0.0017717756563797593, + 0.0013364804908633232, + 0.0011236763093620539, + 0.0017796617466956377, + 0.001776237040758133, + 0.0015976416179910302, + 0.00276272208429873, + 0.008993339724838734, + 0.0039820922538638115, + 0.004822383169084787, + 0.003286689752712846, + 0.00315930787473917, + 0.0056583997793495655, + 0.0031584075186401606, + 0.0062014264985919, + 0.0014488387387245893, + 0.0009413810912519693, + 0.0009033418609760702, + 0.0008965849410742521, + 0.0010013030841946602, + 0.0013158059446141124, + 0.0013053779257461429, + 0.0017564438749104738, + 0.003006816143169999, + 0.0023987414315342903, + 0.002202509203925729, + 0.0011123190633952618, + 0.0029079080559313297, + 0.003995633218437433, + 0.001826810883358121, + 0.0031495860312134027, + 0.0027467685285955667, + 0.0011112792417407036, + 0.0023561976850032806, + 0.0010307235643267632, + 0.0008879279484972358, + 0.0036942362785339355, + 0.001299212104640901, + 0.006237567402422428, + 0.0009773371275514364, + 0.0010234052315354347, + 0.0009020697325468063, + 0.0009726178832352161, + 0.001486357650719583, + 0.0009764572023414075, + 0.0010798893636092544, + 0.0010417482117190957, + 0.28652089834213257, + 0.05033161863684654, + 0.3187728524208069, + 0.018351981416344643, + 0.006798507180064917, + 0.08460618555545807, + 0.010202353820204735, + 0.024454167112708092, + 0.0017791928257793188, + 0.0018935962580144405, + 0.0027817643713206053, + 0.001946095610037446, + 0.0015870729694142938, + 0.002125452272593975, + 0.001582697732374072, + 0.0017921774415299296, + 0.000867287686560303, + 0.0005831928574480116, + 0.0007102658855728805, + 0.00034845032496377826, + 0.0006700937519781291, + 0.0007576870848424733, + 0.0006259302026592195, + 0.0010263840667903423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.08908362686634064 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the core of a cast-iron furnace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep orange flames visible licking at glowing charcoal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do incandescent sparks burst violently from the center of the flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sparks cool from incandescent white to red before extinguishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera stationary at eye level throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486521244049072 + ], + "opposite_logprob_tokens": [ + -1.2236521244049072 + ], + "logprob_reward": -0.3486521244049072, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486521244049072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037355951964855194, + 0.029965685680508614, + 0.04042989760637283, + 0.023958737030625343, + 0.0235119815915823, + 0.04567290470004082, + 0.028245238587260246, + 0.6938015818595886, + 0.06431221961975098, + 0.012745737098157406 + ], + "credit_map": [ + 0.004092304036021233, + 0.004755027592182159, + 0.007831142283976078, + 0.004364673513919115, + 0.0035934108309447765, + 0.005434968508780003, + 0.003746955655515194, + 0.003537470242008567, + 0.006346024107187986, + 0.00442235404625535, + 0.003918549511581659, + 0.0020156516693532467, + 0.0025101150386035442, + 0.003175122430548072, + 0.0020745291840285063, + 0.005503337364643812, + 0.007152789272367954, + 0.004027757328003645, + 0.004132658243179321, + 0.0033904907759279013, + 0.004388714674860239, + 0.006490090861916542, + 0.0051019322127103806, + 0.0057454644702374935, + 0.005151310004293919, + 0.001996972830966115, + 0.002633209340274334, + 0.002338235266506672, + 0.0025518613401800394, + 0.0037304451689124107, + 0.002282334491610527, + 0.0032743681222200394, + 0.00351975136436522, + 0.0027130336966365576, + 0.002397441305220127, + 0.0020172588992863894, + 0.002236621920019388, + 0.004130108747631311, + 0.0025638039223849773, + 0.003933963365852833, + 0.005815305281430483, + 0.0025912774726748466, + 0.006197113543748856, + 0.0021349461749196053, + 0.0017630856018513441, + 0.006891471799463034, + 0.002413942012935877, + 0.01786576211452484, + 0.003159069688990712, + 0.003570315893739462, + 0.0035156570374965668, + 0.0028709962498396635, + 0.004640921484678984, + 0.0034780865535140038, + 0.0035148346796631813, + 0.0034953567665070295, + 0.23668576776981354, + 0.04288221523165703, + 0.27503183484077454, + 0.015830518677830696, + 0.010001761838793755, + 0.08256470412015915, + 0.009575892239809036, + 0.021228894591331482, + 0.006771681364625692, + 0.010239171795547009, + 0.0112239895388484, + 0.00872107595205307, + 0.006055776029825211, + 0.008420265279710293, + 0.006008070427924395, + 0.006872184574604034, + 0.0015890175709500909, + 0.0012274745386093855, + 0.0020699109882116318, + 0.0011401562951505184, + 0.0019971339497715235, + 0.0017307751113548875, + 0.0015509322984144092, + 0.0014403358800336719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.3486521244049072 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, captured at eye level, the camera remains stationary, focusing squarely on the blazing core of a cast-iron furnace. Deep orange flames lick wildly at the glowing charcoal. Suddenly, a fistful of incandescent sparks bursts violently from the flame's center. These tiny, burning particles radiate outward, trailing a brilliant orange trail. Like a swarm of miniature comets, they rush swiftly and chaotically into the deep, nearly lightless darkness surrounding them. They trace brief, graceful arcs across the sky, their light rapidly cooling from incandescent white to red before extinguishing themselves, one by one, within seconds, vanishing into nothingness. The scene is imbued with a sense of primal power, the intense light contrasting with the absolute darkness, creating a fleeting spectacle that is both violent and mesmerizing.", + "seed": 8134, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_0/index_5288_seed_8134_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.23263153084553778, + "vlm_logprob": -0.23263153084553778, + "vlm_yes_no_margin": 2.6, + "vlm_token_credit": -0.23263153084553778, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the core of a cast-iron furnace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep orange flames visible licking at glowing charcoal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do incandescent sparks burst violently from the center of the flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sparks cool from incandescent white to red before extinguishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera stationary at eye level throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14845328032970428 + ], + "opposite_logprob_tokens": [ + -2.0234532356262207 + ], + "logprob_reward": -0.14845328032970428, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14845328032970428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10009567439556122, + 0.03511613607406616, + 0.07733362168073654, + 0.1206246092915535, + 0.24386171996593475, + 0.15013207495212555, + 0.11584363132715225, + 0.04983876273036003, + 0.06379392743110657, + 0.04335980862379074 + ], + "credit_map": [ + 0.012156562879681587, + 0.011312548071146011, + 0.011206177994608879, + 0.014106889255344868, + 0.012203573249280453, + 0.013643619604408741, + 0.008947051130235195, + 0.016519246622920036, + 0.0031931004486978054, + 0.003301800461485982, + 0.0062330905348062515, + 0.005591298919171095, + 0.0035147322341799736, + 0.004556065425276756, + 0.004603838548064232, + 0.00412220973521471, + 0.018823064863681793, + 0.008316022343933582, + 0.008949114941060543, + 0.005426950287073851, + 0.0073369890451431274, + 0.012119069695472717, + 0.006768188439309597, + 0.009594224393367767, + 0.020379286259412766, + 0.009634329937398434, + 0.017558187246322632, + 0.012080317363142967, + 0.013271093368530273, + 0.015878096222877502, + 0.01270787138491869, + 0.019115427508950233, + 0.02345750294625759, + 0.033795665949583054, + 0.02448907494544983, + 0.020823964849114418, + 0.05849222466349602, + 0.03576520085334778, + 0.019075769931077957, + 0.027962321415543556, + 0.012456999160349369, + 0.018865132704377174, + 0.02214415930211544, + 0.016075417399406433, + 0.018746647983789444, + 0.021960953250527382, + 0.02260040119290352, + 0.017282361164689064, + 0.010504527948796749, + 0.011771874502301216, + 0.011747071519494057, + 0.017397643998265266, + 0.012925740331411362, + 0.016207899898290634, + 0.010637203231453896, + 0.024651670828461647, + 0.0044977422803640366, + 0.006122861057519913, + 0.006209466606378555, + 0.005113991908729076, + 0.009695906192064285, + 0.007085863035172224, + 0.005153937265276909, + 0.005958995316177607, + 0.007742368150502443, + 0.004682203754782677, + 0.008559150621294975, + 0.01127233263105154, + 0.004846304655075073, + 0.008929776027798653, + 0.008933699689805508, + 0.008828090503811836, + 0.00471900962293148, + 0.0036636346485465765, + 0.00646628625690937, + 0.005215939600020647, + 0.004890845622867346, + 0.005463650915771723, + 0.008083619177341461, + 0.004856820683926344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.14845328032970428 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the core of a cast-iron furnace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep orange flames visible licking at glowing charcoal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do incandescent sparks burst violently from the center of the flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sparks cool from incandescent white to red before extinguishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera stationary at eye level throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032137196976691484 + ], + "opposite_logprob_tokens": [ + -5.753213882446289 + ], + "logprob_reward": -0.0032137196976691484, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032137196976691484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16749586164951324, + 0.05182250589132309, + 0.08759474754333496, + 0.09006956964731216, + 0.11157495528459549, + 0.11218627542257309, + 0.07975488156080246, + 0.1085018515586853, + 0.08288515359163284, + 0.10811421275138855 + ], + "credit_map": [ + 0.02138330228626728, + 0.02984192594885826, + 0.014669109135866165, + 0.023223722353577614, + 0.020387478172779083, + 0.02192719280719757, + 0.01710044965147972, + 0.018962673842906952, + 0.00447738217189908, + 0.005188411101698875, + 0.0077790915966033936, + 0.008556964807212353, + 0.005091871600598097, + 0.006176793947815895, + 0.00733846565708518, + 0.00721352593973279, + 0.019405679777264595, + 0.009581419639289379, + 0.010172863490879536, + 0.007947280071675777, + 0.009828228503465652, + 0.013019900768995285, + 0.007529327645897865, + 0.010110046714544296, + 0.0159084964543581, + 0.007485521491616964, + 0.0175904780626297, + 0.006830819882452488, + 0.00962348934262991, + 0.012682400643825531, + 0.007629540748894215, + 0.012318823486566544, + 0.010958824306726456, + 0.013643588870763779, + 0.012440758757293224, + 0.008070483803749084, + 0.028324319049715996, + 0.01653396710753441, + 0.00831395573914051, + 0.013289056718349457, + 0.00852228794246912, + 0.01613629423081875, + 0.012846162542700768, + 0.011133364401757717, + 0.01541530154645443, + 0.01799217239022255, + 0.013894026167690754, + 0.016246674582362175, + 0.008855312131345272, + 0.007143069989979267, + 0.008566356264054775, + 0.0071706101298332214, + 0.010276492685079575, + 0.010355460457503796, + 0.007457802537828684, + 0.019929775968194008, + 0.008630894124507904, + 0.010664599947631359, + 0.011960215866565704, + 0.014809459447860718, + 0.012968714348971844, + 0.014615476131439209, + 0.0193482656031847, + 0.015504218637943268, + 0.009607091546058655, + 0.005680382717400789, + 0.010516423732042313, + 0.009716609492897987, + 0.00973653607070446, + 0.01112786028534174, + 0.01651277020573616, + 0.009987477213144302, + 0.011304267682135105, + 0.010179096832871437, + 0.014298416674137115, + 0.012707707472145557, + 0.011137277819216251, + 0.013425896875560284, + 0.021764690056443214, + 0.013296867720782757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.0032137196976691484 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the core of a cast-iron furnace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep orange flames visible licking at glowing charcoal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do incandescent sparks burst violently from the center of the flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sparks cool from incandescent white to red before extinguishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera stationary at eye level throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004726678133011 + ], + "opposite_logprob_tokens": [ + -2.695047378540039 + ], + "logprob_reward": -0.07004726678133011, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004726678133011, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13861998915672302, + 0.09037846326828003, + 0.045182064175605774, + 0.0881197527050972, + 0.05817136913537979, + 0.20944511890411377, + 0.11987859010696411, + 0.13117773830890656, + 0.055235713720321655, + 0.0637911856174469 + ], + "credit_map": [ + 0.018917400389909744, + 0.022505685687065125, + 0.014729644171893597, + 0.01893097721040249, + 0.0177552979439497, + 0.014961577951908112, + 0.013737578876316547, + 0.01708182319998741, + 0.007868990302085876, + 0.009384612552821636, + 0.01623394899070263, + 0.010931136086583138, + 0.01113270316272974, + 0.010409774258732796, + 0.01330721378326416, + 0.011110082268714905, + 0.010220223106443882, + 0.004867980722337961, + 0.006105650682002306, + 0.003220495767891407, + 0.003345974488183856, + 0.00684689823538065, + 0.003924786113202572, + 0.006650056689977646, + 0.0170944444835186, + 0.006365357432514429, + 0.01689291000366211, + 0.006412241607904434, + 0.0075437347404658794, + 0.012656174600124359, + 0.007456517778337002, + 0.01369837298989296, + 0.008165566250681877, + 0.006263653747737408, + 0.008309858851134777, + 0.006280707661062479, + 0.009948707185685635, + 0.0075814430601894855, + 0.005431183148175478, + 0.00619024783372879, + 0.020492609590291977, + 0.026715850457549095, + 0.029816510155797005, + 0.019687315449118614, + 0.028235526755452156, + 0.031222224235534668, + 0.025317002087831497, + 0.027958082035183907, + 0.014296411536633968, + 0.015349148772656918, + 0.012902254238724709, + 0.01622638665139675, + 0.018386302515864372, + 0.015957428142428398, + 0.01029421016573906, + 0.016466451808810234, + 0.01116300467401743, + 0.012785268016159534, + 0.013756214641034603, + 0.01880810037255287, + 0.017582502216100693, + 0.017182856798171997, + 0.01735270582139492, + 0.022547079250216484, + 0.006197492592036724, + 0.006368734408169985, + 0.006623547524213791, + 0.010225861333310604, + 0.005429811775684357, + 0.006703002844005823, + 0.007955999113619328, + 0.0057312664575874805, + 0.0063411882147192955, + 0.006879178807139397, + 0.009074322879314423, + 0.006782861426472664, + 0.007666286081075668, + 0.008907454088330269, + 0.01257444266229868, + 0.005565452389419079 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.07004726678133011 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the core of a cast-iron furnace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep orange flames visible licking at glowing charcoal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do incandescent sparks burst violently from the center of the flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sparks cool from incandescent white to red before extinguishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera stationary at eye level throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308897629380226 + ], + "opposite_logprob_tokens": [ + -3.168088912963867 + ], + "logprob_reward": -0.04308897629380226, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308897629380226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12116070091724396, + 0.04937220737338066, + 0.04523564130067825, + 0.12241914123296738, + 0.1007312536239624, + 0.1224040687084198, + 0.08366160094738007, + 0.11329244822263718, + 0.12503699958324432, + 0.1166859045624733 + ], + "credit_map": [ + 0.00862449686974287, + 0.015135264955461025, + 0.022420493885874748, + 0.020001612603664398, + 0.011788394302129745, + 0.013275764882564545, + 0.012477092444896698, + 0.017437588423490524, + 0.0045539396815001965, + 0.005530709866434336, + 0.008330648764967918, + 0.008912208490073681, + 0.0056031132116913795, + 0.0060198609717190266, + 0.004829775542020798, + 0.005591952707618475, + 0.006511308718472719, + 0.005305867176502943, + 0.004327552858740091, + 0.006320790853351355, + 0.004368508234620094, + 0.006500989198684692, + 0.0040476140566170216, + 0.007853012531995773, + 0.01391445379704237, + 0.011964407749474049, + 0.02813926339149475, + 0.010254348628222942, + 0.010363055393099785, + 0.01900649070739746, + 0.00958356074988842, + 0.01919356919825077, + 0.01060449704527855, + 0.012931525707244873, + 0.010476944968104362, + 0.010130525566637516, + 0.020785929635167122, + 0.014994392171502113, + 0.00992269441485405, + 0.010884745977818966, + 0.008485501632094383, + 0.012304735369980335, + 0.015406442806124687, + 0.011509539559483528, + 0.021675674244761467, + 0.020639915019273758, + 0.015412607230246067, + 0.01696966215968132, + 0.009116938337683678, + 0.00943452026695013, + 0.009672330692410469, + 0.010793316178023815, + 0.01326759159564972, + 0.013394458219408989, + 0.008598200045526028, + 0.009384242817759514, + 0.015374845825135708, + 0.013096139766275883, + 0.018067246302962303, + 0.010578924790024757, + 0.020861897617578506, + 0.016622394323349, + 0.00803828239440918, + 0.010652724653482437, + 0.012836223468184471, + 0.011904753744602203, + 0.015039852820336819, + 0.01806984469294548, + 0.013866097666323185, + 0.015632277354598045, + 0.022605156525969505, + 0.015082783997058868, + 0.01348282303661108, + 0.013641517609357834, + 0.01494783628731966, + 0.012706618756055832, + 0.012837408110499382, + 0.013602297753095627, + 0.019632425159215927, + 0.015834975987672806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 5 + ], + "selected_reward": -0.04308897629380226 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the core of a cast-iron furnace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep orange flames visible licking at glowing charcoal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do incandescent sparks burst violently from the center of the flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sparks cool from incandescent white to red before extinguishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera stationary at eye level throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983544111251831 + ], + "opposite_logprob_tokens": [ + -0.5233544111251831 + ], + "logprob_reward": -0.8983544111251831, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983544111251831, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07458008825778961, + 0.05052762106060982, + 0.0887506902217865, + 0.06647548079490662, + 0.11552626639604568, + 0.14044220745563507, + 0.22028669714927673, + 0.10606329143047333, + 0.06613753736019135, + 0.07121007144451141 + ], + "credit_map": [ + 0.0066206613555550575, + 0.007210100069642067, + 0.010368434712290764, + 0.01036056037992239, + 0.008753719739615917, + 0.0091481888666749, + 0.011534783989191055, + 0.010583640076220036, + 0.003952683415263891, + 0.004141575191169977, + 0.006657748017460108, + 0.0063507878221571445, + 0.004539225250482559, + 0.0070203207433223724, + 0.007367487531155348, + 0.01049779262393713, + 0.01912328414618969, + 0.011800101958215237, + 0.011300390586256981, + 0.007339800707995892, + 0.010467397049069405, + 0.01394027192145586, + 0.005840781610459089, + 0.008938658982515335, + 0.010493053123354912, + 0.0069618006236851215, + 0.007849277928471565, + 0.006623771041631699, + 0.005238008219748735, + 0.010204358026385307, + 0.00785713829100132, + 0.01124807633459568, + 0.012947759591042995, + 0.009785258211195469, + 0.016804946586489677, + 0.015231188386678696, + 0.0142279714345932, + 0.01614154316484928, + 0.01013800036162138, + 0.02024959772825241, + 0.011241032741963863, + 0.025105250999331474, + 0.02214101329445839, + 0.013879369013011456, + 0.013182509690523148, + 0.020220475271344185, + 0.020283229649066925, + 0.014389334246516228, + 0.021770648658275604, + 0.026274675503373146, + 0.024619264528155327, + 0.03439856693148613, + 0.034253526479005814, + 0.03171291574835777, + 0.01849299669265747, + 0.028764113783836365, + 0.011385154910385609, + 0.011286652646958828, + 0.014095312915742397, + 0.012539767660200596, + 0.017030656337738037, + 0.014094507321715355, + 0.011635242961347103, + 0.013995993882417679, + 0.006662952248007059, + 0.009092152118682861, + 0.009336450137197971, + 0.008882314898073673, + 0.006719156168401241, + 0.010325144045054913, + 0.007120955735445023, + 0.00799841433763504, + 0.007707493845373392, + 0.00701252231374383, + 0.0101692583411932, + 0.007061538752168417, + 0.007563667371869087, + 0.009111061692237854, + 0.011341804638504982, + 0.011242721229791641 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.8983544111251831 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "An extreme close-up, captured at eye level, the camera remains stationary, focusing squarely on the blazing core of a cast-iron furnace. Deep orange flames lick wildly at the glowing charcoal. Suddenly, a fistful of incandescent sparks bursts violently from the flame's center. These tiny, burning particles radiate outward, trailing a brilliant orange trail. Like a swarm of miniature comets, they rush swiftly and chaotically into the deep, nearly lightless darkness surrounding them. They trace brief, graceful arcs across the sky, their light rapidly cooling from incandescent white to red before extinguishing themselves, one by one, within seconds, vanishing into nothingness. The scene is imbued with a sense of primal power, the intense light contrasting with the absolute darkness, creating a fleeting spectacle that is both violent and mesmerizing.", + "seed": 8135, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_0/index_5288_seed_8135_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.27370451148599384, + "vlm_logprob": -0.27370451148599384, + "vlm_yes_no_margin": 2.55, + "vlm_token_credit": -0.27370451148599384, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the core of a cast-iron furnace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep orange flames visible licking at glowing charcoal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do incandescent sparks burst violently from the center of the flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sparks cool from incandescent white to red before extinguishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera stationary at eye level throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.287947416305542 + ], + "opposite_logprob_tokens": [ + -1.412947416305542 + ], + "logprob_reward": -0.287947416305542, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.287947416305542, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19650177657604218, + 0.08733562380075455, + 0.0769650861620903, + 0.08067566901445389, + 0.10148268938064575, + 0.2816782593727112, + 0.04553081840276718, + 0.04222050681710243, + 0.049141161143779755, + 0.03846840560436249 + ], + "credit_map": [ + 0.018212411552667618, + 0.03726349398493767, + 0.023485612124204636, + 0.021686699241399765, + 0.028963273391127586, + 0.021162573248147964, + 0.023384835571050644, + 0.022342903539538383, + 0.013593092560768127, + 0.012301241047680378, + 0.013839112594723701, + 0.007003542967140675, + 0.01827167719602585, + 0.00862450897693634, + 0.006051267962902784, + 0.007651184685528278, + 0.01356788631528616, + 0.008129402995109558, + 0.009607736952602863, + 0.008056262508034706, + 0.007358607370406389, + 0.007870936766266823, + 0.006750430446118116, + 0.015623828396201134, + 0.013845141045749187, + 0.00914722215384245, + 0.013837332837283611, + 0.006900022272020578, + 0.012043917551636696, + 0.01085324864834547, + 0.0076213413849473, + 0.00642744405195117, + 0.017009535804390907, + 0.013908368535339832, + 0.011885646730661392, + 0.007603722158819437, + 0.009593230672180653, + 0.01666978932917118, + 0.010608705691993237, + 0.014203699305653572, + 0.04661029949784279, + 0.041619546711444855, + 0.028758112341165543, + 0.023775221779942513, + 0.06881236284971237, + 0.0231610257178545, + 0.029369963333010674, + 0.01957174576818943, + 0.007302180863916874, + 0.005715587642043829, + 0.006755313370376825, + 0.004365426488220692, + 0.0037273147609084845, + 0.007884310558438301, + 0.005912770517170429, + 0.0038679190911352634, + 0.007152199745178223, + 0.005951269064098597, + 0.007657308131456375, + 0.004401434678584337, + 0.0033376626670360565, + 0.006085574626922607, + 0.0034908882807940245, + 0.004144175909459591, + 0.012361142784357071, + 0.0038669437635689974, + 0.013886128552258015, + 0.0037730205804109573, + 0.0047449045814573765, + 0.004591967444866896, + 0.0031018340960144997, + 0.00281522492878139, + 0.0064991493709385395, + 0.004214370157569647, + 0.006613048259168863, + 0.006615846883505583, + 0.0027642734348773956, + 0.004039764869958162, + 0.00455277506262064, + 0.0031691775657236576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.287947416305542 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the core of a cast-iron furnace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep orange flames visible licking at glowing charcoal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do incandescent sparks burst violently from the center of the flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sparks cool from incandescent white to red before extinguishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera stationary at eye level throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046577295288443565 + ], + "opposite_logprob_tokens": [ + -5.379657745361328 + ], + "logprob_reward": -0.0046577295288443565, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046577295288443565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15371821820735931, + 0.07821452617645264, + 0.0654374212026596, + 0.0581936277449131, + 0.18923048675060272, + 0.2045406699180603, + 0.03669196739792824, + 0.08053549379110336, + 0.10101964324712753, + 0.03241797164082527 + ], + "credit_map": [ + 0.01588473655283451, + 0.03064107708632946, + 0.013583467341959476, + 0.017972202971577644, + 0.026561174541711807, + 0.018075641244649887, + 0.013696808367967606, + 0.0173031035810709, + 0.0121859610080719, + 0.008434562012553215, + 0.01402608584612608, + 0.010881522670388222, + 0.008679925464093685, + 0.008991372771561146, + 0.0054818009957671165, + 0.009533287957310677, + 0.00949705857783556, + 0.006773251108825207, + 0.008287856355309486, + 0.007103359792381525, + 0.008472273126244545, + 0.006543736904859543, + 0.005820567719638348, + 0.01293931994587183, + 0.009871207177639008, + 0.011081629432737827, + 0.0070024700835347176, + 0.004119338002055883, + 0.006246668752282858, + 0.006539341527968645, + 0.007803167682141066, + 0.005529805552214384, + 0.03231978416442871, + 0.034332823008298874, + 0.020113548263907433, + 0.012713993899524212, + 0.02445225976407528, + 0.023615701124072075, + 0.018396740779280663, + 0.023285631090402603, + 0.03069138340651989, + 0.03338085860013962, + 0.019699545577168465, + 0.019272267818450928, + 0.05014174431562424, + 0.01590961590409279, + 0.020528610795736313, + 0.014916638843715191, + 0.005813256371766329, + 0.004152368288487196, + 0.005156298168003559, + 0.003739646403118968, + 0.004185137338936329, + 0.005660110153257847, + 0.004778224974870682, + 0.003206924069672823, + 0.013559783808887005, + 0.015592312440276146, + 0.009286710061132908, + 0.007198031060397625, + 0.007452308665961027, + 0.011764246970415115, + 0.006822994444519281, + 0.008859112858772278, + 0.02959904819726944, + 0.006834343075752258, + 0.0298264492303133, + 0.005852904170751572, + 0.008770073764026165, + 0.008833050727844238, + 0.006498104892671108, + 0.004805667791515589, + 0.005979157984256744, + 0.003223174251616001, + 0.005235443823039532, + 0.005538542754948139, + 0.00240167323499918, + 0.003601381555199623, + 0.0039192200638353825, + 0.002519379137083888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.0046577295288443565 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the core of a cast-iron furnace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep orange flames visible licking at glowing charcoal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do incandescent sparks burst violently from the center of the flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sparks cool from incandescent white to red before extinguishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera stationary at eye level throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545477598905563 + ], + "opposite_logprob_tokens": [ + -4.387545585632324 + ], + "logprob_reward": -0.012545477598905563, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545477598905563, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14612171053886414, + 0.2294009029865265, + 0.07684459537267685, + 0.13344024121761322, + 0.08874697238206863, + 0.16343489289283752, + 0.03491197153925896, + 0.05110250413417816, + 0.03774581849575043, + 0.03825042024254799 + ], + "credit_map": [ + 0.016857894137501717, + 0.03205820918083191, + 0.015947502106428146, + 0.012258977629244328, + 0.01995597407221794, + 0.01679062470793724, + 0.015726424753665924, + 0.016526099294424057, + 0.03579963371157646, + 0.02611803449690342, + 0.03180380165576935, + 0.028847893700003624, + 0.039089545607566833, + 0.024076493456959724, + 0.01797342114150524, + 0.025692077353596687, + 0.011890731751918793, + 0.008689324371516705, + 0.009083340875804424, + 0.006370518356561661, + 0.01499171182513237, + 0.006980755366384983, + 0.005925718229264021, + 0.012912494130432606, + 0.021956516429781914, + 0.02535528503358364, + 0.016270341351628304, + 0.010343651287257671, + 0.017593281343579292, + 0.012321028858423233, + 0.013336711563169956, + 0.01626342348754406, + 0.015556749887764454, + 0.023286893963813782, + 0.010650057345628738, + 0.0056183733977377415, + 0.012389615178108215, + 0.007030822336673737, + 0.00623770896345377, + 0.007976749911904335, + 0.02645423263311386, + 0.029493138194084167, + 0.015385907143354416, + 0.014963516034185886, + 0.039673082530498505, + 0.011663258075714111, + 0.014394878409802914, + 0.011406888253986835, + 0.006197804119437933, + 0.004957782104611397, + 0.005415887571871281, + 0.0031682662665843964, + 0.003925041761249304, + 0.005300904158502817, + 0.0031141838990151882, + 0.0028320983983576298, + 0.009311829693615437, + 0.006022593937814236, + 0.013450597412884235, + 0.006013863254338503, + 0.002727413084357977, + 0.0058268229477107525, + 0.004545631818473339, + 0.0032037487253546715, + 0.010603675618767738, + 0.0024243483785539865, + 0.012103662826120853, + 0.0030796644277870655, + 0.0018900720169767737, + 0.0032873942982405424, + 0.002025074092671275, + 0.002331928815692663, + 0.006666873581707478, + 0.0041855196468532085, + 0.006021928042173386, + 0.0071759335696697235, + 0.002749326406046748, + 0.0036621775943785906, + 0.004549151286482811, + 0.0032395098824054003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.012545477598905563 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the core of a cast-iron furnace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep orange flames visible licking at glowing charcoal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do incandescent sparks burst violently from the center of the flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sparks cool from incandescent white to red before extinguishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera stationary at eye level throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906487375497818 + ], + "opposite_logprob_tokens": [ + -2.464064836502075 + ], + "logprob_reward": -0.08906487375497818, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906487375497818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17231303453445435, + 0.07999353855848312, + 0.11866723001003265, + 0.03566475585103035, + 0.11029462516307831, + 0.1111859455704689, + 0.06092935800552368, + 0.04299992695450783, + 0.21461503207683563, + 0.05333653464913368 + ], + "credit_map": [ + 0.01653645932674408, + 0.03506525233387947, + 0.018544834107160568, + 0.01952448859810829, + 0.025083715096116066, + 0.020208213478326797, + 0.018845604732632637, + 0.018504468724131584, + 0.01118896808475256, + 0.012971226125955582, + 0.009923402220010757, + 0.010228250175714493, + 0.012403804808855057, + 0.007479446940124035, + 0.006116889417171478, + 0.009681549854576588, + 0.01869562827050686, + 0.011736351065337658, + 0.014231747016310692, + 0.009803214110434055, + 0.025997182354331017, + 0.01208110898733139, + 0.007019288372248411, + 0.019102707505226135, + 0.006065836176276207, + 0.008317780680954456, + 0.005781060084700584, + 0.002849396551027894, + 0.0029443465173244476, + 0.002147186081856489, + 0.0036998752038925886, + 0.0038592747878283262, + 0.018471380695700645, + 0.026134662330150604, + 0.015563813969492912, + 0.00566934235394001, + 0.014216248877346516, + 0.011203058995306492, + 0.007709933910518885, + 0.01132618635892868, + 0.016182919964194298, + 0.02334035187959671, + 0.010615146718919277, + 0.011698804795742035, + 0.022909317165613174, + 0.007164319511502981, + 0.00892405491322279, + 0.010351026430726051, + 0.01045244187116623, + 0.007314589340239763, + 0.008170219138264656, + 0.005906553007662296, + 0.005291779525578022, + 0.010877973400056362, + 0.006578223779797554, + 0.006337576545774937, + 0.007695342879742384, + 0.008743724785745144, + 0.003667350858449936, + 0.004546174313873053, + 0.0047438014298677444, + 0.00536795100197196, + 0.003984592854976654, + 0.004250988829880953, + 0.07252881675958633, + 0.011407100595533848, + 0.07160753756761551, + 0.010685778222978115, + 0.009792604483664036, + 0.016856729984283447, + 0.010420872829854488, + 0.01131559070199728, + 0.009386658668518066, + 0.00569720845669508, + 0.008145678788423538, + 0.009506241418421268, + 0.004429245367646217, + 0.005056118592619896, + 0.0066352346912026405, + 0.004480146802961826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.08906487375497818 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the core of a cast-iron furnace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep orange flames visible licking at glowing charcoal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do incandescent sparks burst violently from the center of the flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sparks cool from incandescent white to red before extinguishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera stationary at eye level throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9743070602416992 + ], + "opposite_logprob_tokens": [ + -0.4743070602416992 + ], + "logprob_reward": -0.9743070602416992, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9743070602416992, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.039564937353134155, + 0.062380190938711166, + 0.07793258130550385, + 0.05930466949939728, + 0.08538778871297836, + 0.129134863615036, + 0.12395426630973816, + 0.04273147135972977, + 0.3349629044532776, + 0.04464632272720337 + ], + "credit_map": [ + 0.00507837999612093, + 0.0052573843859136105, + 0.00403415085747838, + 0.003485695691779256, + 0.005232269875705242, + 0.005263989791274071, + 0.0047647240571677685, + 0.0064483387395739555, + 0.010599492117762566, + 0.008452282287180424, + 0.009126126766204834, + 0.00643414631485939, + 0.010446165688335896, + 0.007306511048227549, + 0.004125624895095825, + 0.0058898418210446835, + 0.010387091897428036, + 0.006197554524987936, + 0.010605650953948498, + 0.007383083924651146, + 0.013191502541303635, + 0.011809224262833595, + 0.005107092205435038, + 0.01325137447565794, + 0.01237335056066513, + 0.011308470740914345, + 0.005595067050307989, + 0.004232729319483042, + 0.006300700828433037, + 0.006006230134516954, + 0.006460787262767553, + 0.007027335464954376, + 0.014921838417649269, + 0.013166832737624645, + 0.009705777280032635, + 0.01000694278627634, + 0.011275057680904865, + 0.008029211312532425, + 0.009330647066235542, + 0.00895148515701294, + 0.022071009501814842, + 0.019148793071508408, + 0.020397106185555458, + 0.017645392566919327, + 0.010735339485108852, + 0.01346763875335455, + 0.01047383714467287, + 0.015195753425359726, + 0.02101236581802368, + 0.016721924766898155, + 0.014544320292770863, + 0.012774139642715454, + 0.017902366816997528, + 0.01867455057799816, + 0.009881088510155678, + 0.01244350802153349, + 0.00743330642580986, + 0.007525479421019554, + 0.004950235597789288, + 0.0031849616207182407, + 0.0034552724100649357, + 0.007845250889658928, + 0.002355302916839719, + 0.005981660448014736, + 0.11807612329721451, + 0.016900673508644104, + 0.1156640499830246, + 0.01565798558294773, + 0.011026127263903618, + 0.02560623735189438, + 0.014594326727092266, + 0.017437389120459557, + 0.008416903205215931, + 0.005028389394283295, + 0.007776806131005287, + 0.0077976444736123085, + 0.0029682370368391275, + 0.0036678845062851906, + 0.005398640874773264, + 0.003591816406697035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 6 + ], + "selected_reward": -0.9743070602416992 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_1/index_tensor([4591, 4591])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_1/index_tensor([4591, 4591])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e9dd8bb75dc61a0a5015d50d7cc437a3a5032f89 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_1/index_tensor([4591, 4591])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-angle, handheld shot closely follows a heavily modified, metallic cherry-red race car as it performs an extreme drift on an asphalt track scorched by the afternoon sun. The car, adorned with a widebody kit and a massive carbon fiber rear wing, pitches violently to one side at high speed. Its rear wheels instantly lose grip, rubbing violently against the cracked, rough surface. With a piercing tire screeching, a thick, nearly opaque cloud of white smoke billows out, instantly engulfing the rear of the car. A cloud of yellow dust and white smoke swirls beneath the glaring sunlight, which reflects off the car's smooth, curved surface in flowing patterns. The camera pans with the car's violent slide, capturing the car's graceful yet perilous sideways glide through the curve on the edge of control and loss of control. The entire scene is filled with raw mechanical power and adrenaline-pumping tension.", + "seed": 20844, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_1/index_4591_seed_20844_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.014787007356062531, + "vlm_logprob": -0.014787007356062531, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.014787007356062531, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic cherry-red race car with a widebody kit and a massive carbon fiber rear wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car performing an extreme drift on an asphalt track that appears scorched by the afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pitch violently to one side while its rear wheels lose grip and rub against the cracked, rough surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, nearly opaque cloud of white smoke billowing from the rear wheels, mixed with yellow dust swirling beneath the glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld view that pans with the car's violent slide?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02396869845688343 + ], + "opposite_logprob_tokens": [ + -4.273968696594238 + ], + "logprob_reward": -0.02396869845688343, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.02396869845688343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2953091561794281, + 0.09839516878128052, + 0.05544620007276535, + 0.05299288034439087, + 0.03633768483996391, + 0.058942243456840515, + 0.07573144137859344, + 0.0991816446185112, + 0.13043764233589172, + 0.09722597151994705 + ], + "credit_map": [ + 0.024234503507614136, + 0.022625714540481567, + 0.023992013186216354, + 0.021462973207235336, + 0.018997959792613983, + 0.038785893470048904, + 0.06676038354635239, + 0.07844971120357513, + 0.009730285964906216, + 0.015818104147911072, + 0.010426913388073444, + 0.012913956306874752, + 0.0139676658436656, + 0.01103176735341549, + 0.012415137141942978, + 0.01209133118391037, + 0.006116747856140137, + 0.006168801337480545, + 0.00604692567139864, + 0.006548456382006407, + 0.009918125346302986, + 0.006340248975902796, + 0.0046111964620649815, + 0.009695702232420444, + 0.006429445464164019, + 0.00500094797462225, + 0.0054207369685173035, + 0.00843887124210596, + 0.006943476386368275, + 0.007326759397983551, + 0.005729143042117357, + 0.007703497540205717, + 0.0036575929261744022, + 0.005249171517789364, + 0.005396147258579731, + 0.004896862898021936, + 0.004522669594734907, + 0.0036313075106590986, + 0.004820149391889572, + 0.0041637844406068325, + 0.004766030702739954, + 0.006024807225912809, + 0.0041444310918450356, + 0.0034851552918553352, + 0.009609694592654705, + 0.01625365950167179, + 0.007051592692732811, + 0.007606872823089361, + 0.003160082036629319, + 0.003601657459512353, + 0.003778040176257491, + 0.004513208754360676, + 0.014342249371111393, + 0.0073940041474998, + 0.012287743389606476, + 0.026654452085494995, + 0.005837402772158384, + 0.00836469791829586, + 0.013490616343915462, + 0.01325556356459856, + 0.012170986272394657, + 0.018915215507149696, + 0.014274862594902515, + 0.012872309423983097, + 0.007696515880525112, + 0.00702890008687973, + 0.013749197125434875, + 0.01798640936613083, + 0.017644913867115974, + 0.01715754345059395, + 0.0343906506896019, + 0.014783510938286781, + 0.00775675056502223, + 0.008226203732192516, + 0.022467998787760735, + 0.008239123970270157, + 0.007352183572947979, + 0.009427509270608425, + 0.009806924499571323, + 0.023949269205331802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.02396869845688343 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic cherry-red race car with a widebody kit and a massive carbon fiber rear wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car performing an extreme drift on an asphalt track that appears scorched by the afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pitch violently to one side while its rear wheels lose grip and rub against the cracked, rough surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, nearly opaque cloud of white smoke billowing from the rear wheels, mixed with yellow dust swirling beneath the glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld view that pans with the car's violent slide?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014212341979146004 + ], + "opposite_logprob_tokens": [ + -4.264212131500244 + ], + "logprob_reward": -0.014212341979146004, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014212341979146004, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1671665757894516, + 0.1838955134153366, + 0.13077804446220398, + 0.05216704308986664, + 0.026968875899910927, + 0.1329032927751541, + 0.08848937600851059, + 0.05344239994883537, + 0.11744522303342819, + 0.046743713319301605 + ], + "credit_map": [ + 0.02054835669696331, + 0.0209454707801342, + 0.022831162437796593, + 0.017734907567501068, + 0.015247298404574394, + 0.024145394563674927, + 0.018864305689930916, + 0.026849668473005295, + 0.018180081620812416, + 0.022794440388679504, + 0.014980463311076164, + 0.012344561517238617, + 0.028859857469797134, + 0.03654002770781517, + 0.02117273584008217, + 0.029023345559835434, + 0.019985808059573174, + 0.012569177895784378, + 0.007882428355515003, + 0.011746562086045742, + 0.017187191173434258, + 0.017229709774255753, + 0.013582413084805012, + 0.030594754964113235, + 0.004896795377135277, + 0.0051834168843925, + 0.006389477290213108, + 0.00805729255080223, + 0.0065972157754004, + 0.005140500143170357, + 0.008594527840614319, + 0.007307817228138447, + 0.0033138422295451164, + 0.003477684687823057, + 0.004003183916211128, + 0.003648071549832821, + 0.00341683067381382, + 0.0021720416843891144, + 0.003779205959290266, + 0.003158015664666891, + 0.012632256373763084, + 0.011878696270287037, + 0.009166172705590725, + 0.007309915963560343, + 0.0222335048019886, + 0.037719618529081345, + 0.01567043364048004, + 0.016292694956064224, + 0.0033851424232125282, + 0.003723529865965247, + 0.004001399967819452, + 0.0051003447733819485, + 0.017017174512147903, + 0.00965970940887928, + 0.014045491814613342, + 0.03155657649040222, + 0.0030449077021330595, + 0.003950537648051977, + 0.006410825066268444, + 0.006074731703847647, + 0.00824674591422081, + 0.010336442850530148, + 0.008495278656482697, + 0.0068829297088086605, + 0.006928245536983013, + 0.006869365926831961, + 0.012696725316345692, + 0.019346138462424278, + 0.01401230413466692, + 0.01252732053399086, + 0.03172046318650246, + 0.013344663195312023, + 0.004800775554031134, + 0.003828297136351466, + 0.0112083051353693, + 0.00586493918672204, + 0.004894669633358717, + 0.0034724432043731213, + 0.004926789086312056, + 0.0077474964782595634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.014212341979146004 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic cherry-red race car with a widebody kit and a massive carbon fiber rear wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car performing an extreme drift on an asphalt track that appears scorched by the afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pitch violently to one side while its rear wheels lose grip and rub against the cracked, rough surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, nearly opaque cloud of white smoke billowing from the rear wheels, mixed with yellow dust swirling beneath the glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld view that pans with the car's violent slide?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016078339889645576 + ], + "opposite_logprob_tokens": [ + -4.141078472137451 + ], + "logprob_reward": -0.016078339889645576, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016078339889645576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11580463498830795, + 0.07302616536617279, + 0.07245771586894989, + 0.045657262206077576, + 0.05159584432840347, + 0.0267251655459404, + 0.2198009490966797, + 0.04871263727545738, + 0.14114677906036377, + 0.2050727903842926 + ], + "credit_map": [ + 0.012379176914691925, + 0.022047752514481544, + 0.01353546790778637, + 0.010404692031443119, + 0.01249849982559681, + 0.012863234616816044, + 0.015081368386745453, + 0.016994457691907883, + 0.008486832492053509, + 0.014900082722306252, + 0.005326594226062298, + 0.005607749335467815, + 0.009483729489147663, + 0.014474730007350445, + 0.006778031121939421, + 0.007968426682054996, + 0.011164341121912003, + 0.00494633661583066, + 0.006599860731512308, + 0.005397781264036894, + 0.010013645514845848, + 0.011155487038195133, + 0.006584036163985729, + 0.016596227884292603, + 0.006021010689437389, + 0.0036838098894804716, + 0.0038770565297454596, + 0.005601674783974886, + 0.004498846363276243, + 0.009595129638910294, + 0.005651460494846106, + 0.006728277541697025, + 0.00562738161534071, + 0.004081898834556341, + 0.004947288427501917, + 0.004851994104683399, + 0.0056275189854204655, + 0.01090528815984726, + 0.008463380858302116, + 0.007091097999364138, + 0.002303360728546977, + 0.0025925333611667156, + 0.0030250486452132463, + 0.002687903353944421, + 0.002954599680379033, + 0.004737779963761568, + 0.004684909712523222, + 0.0037390354555100203, + 0.008918586187064648, + 0.007889926433563232, + 0.009677805937826633, + 0.007147185038775206, + 0.045924074947834015, + 0.022171229124069214, + 0.033616382628679276, + 0.08445578813552856, + 0.0023472525645047426, + 0.0032370402477681637, + 0.009214235469698906, + 0.007560295984148979, + 0.005478943232446909, + 0.007160622626543045, + 0.007836845703423023, + 0.00587740633636713, + 0.00684959813952446, + 0.00820155069231987, + 0.01627206616103649, + 0.023179909214377403, + 0.017572807148098946, + 0.014821012504398823, + 0.038246285170316696, + 0.0160035602748394, + 0.015075888484716415, + 0.012654009275138378, + 0.05973378196358681, + 0.014586289413273335, + 0.009513301774859428, + 0.01607699505984783, + 0.01497361809015274, + 0.062458932399749756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 8 + ], + "selected_reward": -0.016078339889645576 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic cherry-red race car with a widebody kit and a massive carbon fiber rear wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car performing an extreme drift on an asphalt track that appears scorched by the afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pitch violently to one side while its rear wheels lose grip and rub against the cracked, rough surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, nearly opaque cloud of white smoke billowing from the rear wheels, mixed with yellow dust swirling beneath the glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld view that pans with the car's violent slide?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036145609337836504 + ], + "opposite_logprob_tokens": [ + -5.62861442565918 + ], + "logprob_reward": -0.0036145609337836504, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036145609337836504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12473020702600479, + 0.194101020693779, + 0.06768719851970673, + 0.05125192552804947, + 0.05237150937318802, + 0.11245352029800415, + 0.17528316378593445, + 0.047814711928367615, + 0.08701260387897491, + 0.0872940719127655 + ], + "credit_map": [ + 0.018450114876031876, + 0.018494056537747383, + 0.016243577003479004, + 0.010290861129760742, + 0.012519176118075848, + 0.010478993877768517, + 0.020139070227742195, + 0.018114354461431503, + 0.018792539834976196, + 0.02679155208170414, + 0.01628325879573822, + 0.019843319430947304, + 0.029466737061738968, + 0.03129124641418457, + 0.026691587641835213, + 0.024940794333815575, + 0.01045125350356102, + 0.005635786801576614, + 0.005044681951403618, + 0.005439154803752899, + 0.008280699141323566, + 0.009977162815630436, + 0.006831022445112467, + 0.016027437523007393, + 0.004103243816643953, + 0.0037018777802586555, + 0.004826556425541639, + 0.005192554090172052, + 0.008350714109838009, + 0.00866756122559309, + 0.008385385386645794, + 0.008024035952985287, + 0.006576219107955694, + 0.005702303256839514, + 0.006343637127429247, + 0.006248691119253635, + 0.008461819030344486, + 0.007387121208012104, + 0.0055170124396681786, + 0.006134707480669022, + 0.008199814707040787, + 0.01222267746925354, + 0.010809852741658688, + 0.007102000527083874, + 0.016602441668510437, + 0.027816249057650566, + 0.014955306425690651, + 0.014745179563760757, + 0.007031551096588373, + 0.0056903851218521595, + 0.00719053391367197, + 0.005682329647243023, + 0.03788452968001366, + 0.014748474583029747, + 0.026764042675495148, + 0.07029131799936295, + 0.003183955792337656, + 0.0036223705392330885, + 0.006381174549460411, + 0.006818914785981178, + 0.008370469324290752, + 0.007892601191997528, + 0.006553059909492731, + 0.004992170259356499, + 0.005174103192985058, + 0.005705375224351883, + 0.010010232217609882, + 0.015412287786602974, + 0.009783754125237465, + 0.009068714454770088, + 0.02213207259774208, + 0.009726057760417461, + 0.006690044421702623, + 0.0074226511642336845, + 0.016256211325526237, + 0.013038000091910362, + 0.0084256362169981, + 0.00659506069496274, + 0.013356776908040047, + 0.015509694814682007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.0036145609337836504 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic cherry-red race car with a widebody kit and a massive carbon fiber rear wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car performing an extreme drift on an asphalt track that appears scorched by the afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pitch violently to one side while its rear wheels lose grip and rub against the cracked, rough surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, nearly opaque cloud of white smoke billowing from the rear wheels, mixed with yellow dust swirling beneath the glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld view that pans with the car's violent slide?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016061095520853996 + ], + "opposite_logprob_tokens": [ + -4.141061305999756 + ], + "logprob_reward": -0.016061095520853996, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016061095520853996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17375829815864563, + 0.09801027923822403, + 0.09208317846059799, + 0.032091282308101654, + 0.037302374839782715, + 0.1294848918914795, + 0.05027610808610916, + 0.04702451452612877, + 0.13079579174518585, + 0.209173321723938 + ], + "credit_map": [ + 0.020312707871198654, + 0.02801373600959778, + 0.02515380084514618, + 0.018063250929117203, + 0.01232602633535862, + 0.016501756384968758, + 0.024229491129517555, + 0.029157519340515137, + 0.00960332527756691, + 0.018791528418660164, + 0.008096223697066307, + 0.007436043582856655, + 0.01419502217322588, + 0.018889185041189194, + 0.011139217764139175, + 0.009859731420874596, + 0.010475926101207733, + 0.00775126414373517, + 0.008666522800922394, + 0.008072691038250923, + 0.009592020884156227, + 0.017823336645960808, + 0.007458101958036423, + 0.022243311628699303, + 0.0027596966829150915, + 0.0025948595721274614, + 0.004535222891718149, + 0.004370472393929958, + 0.004254397936165333, + 0.0039625996723771095, + 0.0042050122283399105, + 0.00540902279317379, + 0.004248859826475382, + 0.0026341602206230164, + 0.00621233182027936, + 0.0060199289582669735, + 0.0035368057433515787, + 0.005837022792547941, + 0.004587646573781967, + 0.0042256200686097145, + 0.011313140392303467, + 0.01068006083369255, + 0.010000820271670818, + 0.007084283512085676, + 0.021972237154841423, + 0.0371386744081974, + 0.015857189893722534, + 0.015438495203852654, + 0.002323821419849992, + 0.0026200164575129747, + 0.003962446935474873, + 0.004237438552081585, + 0.006450558081269264, + 0.006795734632760286, + 0.00965499971061945, + 0.014231093227863312, + 0.003136121667921543, + 0.002608364913612604, + 0.005001783836632967, + 0.004600238054990768, + 0.007235449738800526, + 0.01116457860916853, + 0.006665156222879887, + 0.006612822413444519, + 0.007358986884355545, + 0.007345747202634811, + 0.015771668404340744, + 0.019444873556494713, + 0.016871189698576927, + 0.014530123211443424, + 0.03483687713742256, + 0.014636334963142872, + 0.015650320798158646, + 0.013502046465873718, + 0.062312204390764236, + 0.018198814243078232, + 0.01097668707370758, + 0.015247107483446598, + 0.017003171145915985, + 0.056282974779605865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.016061095520853996 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle, handheld shot closely follows a heavily modified, metallic cherry-red race car as it performs an extreme drift on an asphalt track scorched by the afternoon sun. The car, adorned with a widebody kit and a massive carbon fiber rear wing, pitches violently to one side at high speed. Its rear wheels instantly lose grip, rubbing violently against the cracked, rough surface. With a piercing tire screeching, a thick, nearly opaque cloud of white smoke billows out, instantly engulfing the rear of the car. A cloud of yellow dust and white smoke swirls beneath the glaring sunlight, which reflects off the car's smooth, curved surface in flowing patterns. The camera pans with the car's violent slide, capturing the car's graceful yet perilous sideways glide through the curve on the edge of control and loss of control. The entire scene is filled with raw mechanical power and adrenaline-pumping tension.", + "seed": 20845, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_1/index_4591_seed_20845_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.018134845606982707, + "vlm_logprob": -0.018134845606982707, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.018134845606982707, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic cherry-red race car with a widebody kit and a massive carbon fiber rear wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car performing an extreme drift on an asphalt track that appears scorched by the afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pitch violently to one side while its rear wheels lose grip and rub against the cracked, rough surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, nearly opaque cloud of white smoke billowing from the rear wheels, mixed with yellow dust swirling beneath the glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld view that pans with the car's violent slide?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015255319885909557 + ], + "opposite_logprob_tokens": [ + -4.890255451202393 + ], + "logprob_reward": -0.015255319885909557, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.015255319885909557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21432682871818542, + 0.1468583196401596, + 0.10692649334669113, + 0.08565930277109146, + 0.04279671981930733, + 0.05526932701468468, + 0.13485491275787354, + 0.0947406217455864, + 0.05861472338438034, + 0.059952739626169205 + ], + "credit_map": [ + 0.03334340080618858, + 0.021250298246741295, + 0.035055313259363174, + 0.02439400553703308, + 0.01939241960644722, + 0.04085726663470268, + 0.021227464079856873, + 0.018806662410497665, + 0.008419537916779518, + 0.010465097613632679, + 0.03705315664410591, + 0.03901594132184982, + 0.009293797425925732, + 0.009479091502726078, + 0.021562382578849792, + 0.011569306254386902, + 0.00863331276923418, + 0.0048995912075042725, + 0.019922615960240364, + 0.014896158128976822, + 0.007331978064030409, + 0.010565772652626038, + 0.01977609470486641, + 0.02090097777545452, + 0.0103448536247015, + 0.008532069623470306, + 0.01396200992166996, + 0.010508577339351177, + 0.013393032364547253, + 0.011148602701723576, + 0.008475886657834053, + 0.009294276125729084, + 0.004845977295190096, + 0.006280743982642889, + 0.006394182797521353, + 0.005125364288687706, + 0.004242365248501301, + 0.0048581743612885475, + 0.0049143824726343155, + 0.006135528441518545, + 0.0039611561223864555, + 0.004377744626253843, + 0.005766765214502811, + 0.010013796389102936, + 0.005653854459524155, + 0.006438825279474258, + 0.01157683227211237, + 0.007480355445295572, + 0.00665689492598176, + 0.010716949589550495, + 0.015459910966455936, + 0.041163358837366104, + 0.009797693230211735, + 0.00885955523699522, + 0.012341567315161228, + 0.029858984053134918, + 0.009478486143052578, + 0.007086379453539848, + 0.021579105406999588, + 0.00981080625206232, + 0.009868678636848927, + 0.0119281942024827, + 0.009075610898435116, + 0.01591336913406849, + 0.007153303828090429, + 0.008085436187684536, + 0.005634104833006859, + 0.0073385280556976795, + 0.0071975900791585445, + 0.008619265630841255, + 0.006407603621482849, + 0.008178891614079475, + 0.005861434154212475, + 0.007777060382068157, + 0.005781315267086029, + 0.006232187617570162, + 0.009895067662000656, + 0.009670346044003963, + 0.006107605993747711, + 0.008627720177173615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.015255319885909557 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic cherry-red race car with a widebody kit and a massive carbon fiber rear wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car performing an extreme drift on an asphalt track that appears scorched by the afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pitch violently to one side while its rear wheels lose grip and rub against the cracked, rough surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, nearly opaque cloud of white smoke billowing from the rear wheels, mixed with yellow dust swirling beneath the glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld view that pans with the car's violent slide?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029807347804307938 + ], + "opposite_logprob_tokens": [ + -3.5298073291778564 + ], + "logprob_reward": -0.029807347804307938, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029807347804307938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2660037577152252, + 0.20191465318202972, + 0.0902988463640213, + 0.022778106853365898, + 0.06940699368715286, + 0.05309195816516876, + 0.08312278240919113, + 0.11551094800233841, + 0.056680355221033096, + 0.041191600263118744 + ], + "credit_map": [ + 0.04510059207677841, + 0.026460185647010803, + 0.042067334055900574, + 0.028626373037695885, + 0.02066608890891075, + 0.041557785123586655, + 0.03308665752410889, + 0.02843872271478176, + 0.014187411405146122, + 0.013361625373363495, + 0.05448273569345474, + 0.056124065071344376, + 0.0091087706387043, + 0.017961254343390465, + 0.019082749262452126, + 0.017606021836400032, + 0.009467418305575848, + 0.00478782644495368, + 0.01673458144068718, + 0.011162007227540016, + 0.006422423757612705, + 0.011902387253940105, + 0.01592218317091465, + 0.013900007121264935, + 0.0022898116149008274, + 0.0020394090097397566, + 0.0028078891336917877, + 0.003184214700013399, + 0.0028783271554857492, + 0.0036365902051329613, + 0.0029771055560559034, + 0.002964755054563284, + 0.005533257964998484, + 0.005326785147190094, + 0.01738893613219261, + 0.007436138112097979, + 0.0046335915103554726, + 0.011373482644557953, + 0.005083273630589247, + 0.012631521560251713, + 0.005677428562194109, + 0.004304748959839344, + 0.006154696922749281, + 0.007960870862007141, + 0.0066584632731974125, + 0.006472123321145773, + 0.010015616193413734, + 0.005848006810992956, + 0.005465061403810978, + 0.012182280421257019, + 0.00840025208890438, + 0.013270528987050056, + 0.013112262822687626, + 0.008862082846462727, + 0.010912938043475151, + 0.010917369276285172, + 0.01238225121051073, + 0.012537609785795212, + 0.017075417563319206, + 0.011486990377306938, + 0.018219588324427605, + 0.012071176432073116, + 0.017936522141098976, + 0.013801388442516327, + 0.004803513176739216, + 0.0057191611267626286, + 0.006052128504961729, + 0.0045721507631242275, + 0.009744786657392979, + 0.014152233488857746, + 0.005276494659483433, + 0.006359884515404701, + 0.004701296798884869, + 0.004923530388623476, + 0.003435468999668956, + 0.004575672093778849, + 0.007124882657080889, + 0.005307952407747507, + 0.00366442883387208, + 0.0074583617970347404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.029807347804307938 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic cherry-red race car with a widebody kit and a massive carbon fiber rear wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car performing an extreme drift on an asphalt track that appears scorched by the afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pitch violently to one side while its rear wheels lose grip and rub against the cracked, rough surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, nearly opaque cloud of white smoke billowing from the rear wheels, mixed with yellow dust swirling beneath the glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld view that pans with the car's violent slide?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026348749175667763 + ], + "opposite_logprob_tokens": [ + -3.651348829269409 + ], + "logprob_reward": -0.026348749175667763, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026348749175667763, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10092154890298843, + 0.17386722564697266, + 0.06698504090309143, + 0.024930546060204506, + 0.12121041864156723, + 0.06098998710513115, + 0.1566978543996811, + 0.10882776975631714, + 0.09685350954532623, + 0.08871607482433319 + ], + "credit_map": [ + 0.017017196863889694, + 0.009381461888551712, + 0.014328189194202423, + 0.011695974506437778, + 0.007231688592582941, + 0.018888402730226517, + 0.013136149384081364, + 0.009242481552064419, + 0.012382913380861282, + 0.011839302256703377, + 0.04789173603057861, + 0.05103505030274391, + 0.008487414568662643, + 0.014286522753536701, + 0.014891041442751884, + 0.01305325422435999, + 0.008707018569111824, + 0.004069893155246973, + 0.012369082309305668, + 0.009029442444443703, + 0.0052222926169633865, + 0.007531610783189535, + 0.009807955473661423, + 0.010247751139104366, + 0.0029720175080001354, + 0.0022045981604605913, + 0.002983173355460167, + 0.0025626246351748705, + 0.0033548357896506786, + 0.004771161358803511, + 0.002768878825008869, + 0.0033132554963231087, + 0.011206288821995258, + 0.008123786188662052, + 0.029754487797617912, + 0.011552873998880386, + 0.005704000126570463, + 0.02089637517929077, + 0.010949350893497467, + 0.02302325889468193, + 0.0038746038917452097, + 0.00701920548453927, + 0.0064946990460157394, + 0.011135080829262733, + 0.005367562640458345, + 0.006485256832093, + 0.01357670221477747, + 0.007036874536424875, + 0.009950336068868637, + 0.025963999330997467, + 0.013368953950703144, + 0.03469124436378479, + 0.01491186860948801, + 0.015076647512614727, + 0.01128927432000637, + 0.03144552558660507, + 0.008031012490391731, + 0.01452526357024908, + 0.00997399352490902, + 0.012669633142650127, + 0.023369070142507553, + 0.012703314423561096, + 0.017951715737581253, + 0.009603757411241531, + 0.009750823490321636, + 0.013106796890497208, + 0.011375902220606804, + 0.00884324125945568, + 0.014391744509339333, + 0.019483067095279694, + 0.008934615179896355, + 0.01096732821315527, + 0.006855224724858999, + 0.01152803935110569, + 0.006098256912082434, + 0.008654769510030746, + 0.016642356291413307, + 0.018755173310637474, + 0.008502406999468803, + 0.011679837480187416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 4 + ], + "selected_reward": -0.026348749175667763 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic cherry-red race car with a widebody kit and a massive carbon fiber rear wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car performing an extreme drift on an asphalt track that appears scorched by the afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pitch violently to one side while its rear wheels lose grip and rub against the cracked, rough surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, nearly opaque cloud of white smoke billowing from the rear wheels, mixed with yellow dust swirling beneath the glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld view that pans with the car's violent slide?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006732873618602753 + ], + "opposite_logprob_tokens": [ + -5.006732940673828 + ], + "logprob_reward": -0.006732873618602753, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006732873618602753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15468533337116241, + 0.06083521991968155, + 0.04920433461666107, + 0.30791792273521423, + 0.03058331087231636, + 0.031212378293275833, + 0.07383713871240616, + 0.04562343284487724, + 0.14179472625255585, + 0.10430615395307541 + ], + "credit_map": [ + 0.025478864088654518, + 0.0172699186950922, + 0.018855327740311623, + 0.01957785151898861, + 0.012396165169775486, + 0.024905916303396225, + 0.01657111756503582, + 0.019630184397101402, + 0.0033577082213014364, + 0.004250986501574516, + 0.014348607510328293, + 0.014639025554060936, + 0.004756411537528038, + 0.004027594346553087, + 0.010564799420535564, + 0.00489008566364646, + 0.004834095947444439, + 0.0028672427870333195, + 0.00874340906739235, + 0.005977154243737459, + 0.004097089171409607, + 0.004520552698522806, + 0.008838895708322525, + 0.009325899183750153, + 0.037384770810604095, + 0.02884269505739212, + 0.04291544482111931, + 0.03459667041897774, + 0.04524758458137512, + 0.05039006844162941, + 0.029551060870289803, + 0.03898962587118149, + 0.0025839507579803467, + 0.003174376208335161, + 0.00692337192595005, + 0.0029374328441917896, + 0.0017490487080067396, + 0.005177247803658247, + 0.0027833590283989906, + 0.005254524759948254, + 0.002138240961357951, + 0.0032516641076654196, + 0.003947664983570576, + 0.010266275145113468, + 0.0016993512399494648, + 0.0022023979108780622, + 0.004966408479958773, + 0.002740375464782119, + 0.00538046145811677, + 0.012989137321710587, + 0.008349239826202393, + 0.020391814410686493, + 0.005861770361661911, + 0.005247245542705059, + 0.0035896706394851208, + 0.012027806602418423, + 0.0042335111647844315, + 0.006915456615388393, + 0.00639782240614295, + 0.005735164042562246, + 0.006515796296298504, + 0.00442116055637598, + 0.006468596868216991, + 0.00493592256680131, + 0.015661155804991722, + 0.021469134837388992, + 0.0124479029327631, + 0.014899729751050472, + 0.020107120275497437, + 0.02773544378578663, + 0.01349509134888649, + 0.015979159623384476, + 0.010813926346600056, + 0.013582928106188774, + 0.006213060114532709, + 0.009735377505421638, + 0.01998293027281761, + 0.017288392409682274, + 0.00941629521548748, + 0.01727324165403843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.006732873618602753 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic cherry-red race car with a widebody kit and a massive carbon fiber rear wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car performing an extreme drift on an asphalt track that appears scorched by the afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pitch violently to one side while its rear wheels lose grip and rub against the cracked, rough surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, nearly opaque cloud of white smoke billowing from the rear wheels, mixed with yellow dust swirling beneath the glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld view that pans with the car's violent slide?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01252993755042553 + ], + "opposite_logprob_tokens": [ + -4.3875298500061035 + ], + "logprob_reward": -0.01252993755042553, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01252993755042553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08553388714790344, + 0.19727402925491333, + 0.08392680436372757, + 0.1895349621772766, + 0.019506171345710754, + 0.06153016909956932, + 0.13632483780384064, + 0.05475611984729767, + 0.12232322990894318, + 0.049289751797914505 + ], + "credit_map": [ + 0.011467385105788708, + 0.005306513048708439, + 0.010884934104979038, + 0.0143379932269454, + 0.0068081701174378395, + 0.010240275412797928, + 0.014741270802915096, + 0.011747345328330994, + 0.010621124878525734, + 0.011043035425245762, + 0.06521505862474442, + 0.0645698606967926, + 0.007791475858539343, + 0.009878360666334629, + 0.017688320949673653, + 0.010466796346008778, + 0.011075972579419613, + 0.005048924591392279, + 0.01012262050062418, + 0.009833352640271187, + 0.0071984538808465, + 0.013587010093033314, + 0.011109991930425167, + 0.015950482338666916, + 0.024854354560375214, + 0.01723223552107811, + 0.028828086331486702, + 0.02005413919687271, + 0.028022319078445435, + 0.028408439829945564, + 0.018548430874943733, + 0.02358696609735489, + 0.002278197556734085, + 0.0019073503790423274, + 0.0022443956695497036, + 0.0028875989373773336, + 0.0023494965862482786, + 0.0023115254007279873, + 0.002666361164301634, + 0.002861245535314083, + 0.003725437680259347, + 0.007744085043668747, + 0.007732747122645378, + 0.021994635462760925, + 0.002522416179999709, + 0.0028891959227621555, + 0.009548663161695004, + 0.005372995510697365, + 0.01349262148141861, + 0.028728695586323738, + 0.014273295179009438, + 0.03324643895030022, + 0.010715498588979244, + 0.00866594910621643, + 0.0063193850219249725, + 0.020882969722151756, + 0.004682991188019514, + 0.005407795310020447, + 0.010751787573099136, + 0.0039844634011387825, + 0.007799656596034765, + 0.006532575469464064, + 0.007135454099625349, + 0.008461399003863335, + 0.010177128948271275, + 0.014752875082194805, + 0.012164219282567501, + 0.016944825649261475, + 0.015690883621573448, + 0.024935908615589142, + 0.013672788627445698, + 0.013984602876007557, + 0.00580280227586627, + 0.0074190846644341946, + 0.004618308041244745, + 0.003845604369416833, + 0.009681091643869877, + 0.007878938689827919, + 0.004025114234536886, + 0.006018808577209711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 6 + ], + "selected_reward": -0.01252993755042553 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle, handheld shot closely follows a heavily modified, metallic cherry-red race car as it performs an extreme drift on an asphalt track scorched by the afternoon sun. The car, adorned with a widebody kit and a massive carbon fiber rear wing, pitches violently to one side at high speed. Its rear wheels instantly lose grip, rubbing violently against the cracked, rough surface. With a piercing tire screeching, a thick, nearly opaque cloud of white smoke billows out, instantly engulfing the rear of the car. A cloud of yellow dust and white smoke swirls beneath the glaring sunlight, which reflects off the car's smooth, curved surface in flowing patterns. The camera pans with the car's violent slide, capturing the car's graceful yet perilous sideways glide through the curve on the edge of control and loss of control. The entire scene is filled with raw mechanical power and adrenaline-pumping tension.", + "seed": 20846, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_1/index_4591_seed_20846_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.013395748008042574, + "vlm_logprob": -0.013395748008042574, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.013395748008042574, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic cherry-red race car with a widebody kit and a massive carbon fiber rear wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car performing an extreme drift on an asphalt track that appears scorched by the afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pitch violently to one side while its rear wheels lose grip and rub against the cracked, rough surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, nearly opaque cloud of white smoke billowing from the rear wheels, mixed with yellow dust swirling beneath the glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld view that pans with the car's violent slide?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016279619187116623 + ], + "opposite_logprob_tokens": [ + -4.766279697418213 + ], + "logprob_reward": -0.016279619187116623, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.016279619187116623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15074852108955383, + 0.08253376185894012, + 0.0847960114479065, + 0.083150714635849, + 0.09483673423528671, + 0.0781790241599083, + 0.10742360353469849, + 0.10294225066900253, + 0.0873454362154007, + 0.12804394960403442 + ], + "credit_map": [ + 0.016987113282084465, + 0.01768755353987217, + 0.02038021758198738, + 0.03216191381216049, + 0.015823228284716606, + 0.013715933077037334, + 0.01831391453742981, + 0.015678653493523598, + 0.012182284146547318, + 0.008790705353021622, + 0.011212162673473358, + 0.007679828442633152, + 0.013971945270895958, + 0.008255630731582642, + 0.008772934786975384, + 0.011668271385133266, + 0.005701437592506409, + 0.004686135333031416, + 0.006282490212470293, + 0.005556080956012011, + 0.006336400285363197, + 0.022415107116103172, + 0.0059878164902329445, + 0.027830541133880615, + 0.0085338419303298, + 0.008833645842969418, + 0.008391369134187698, + 0.014927292242646217, + 0.011370344087481499, + 0.007889186032116413, + 0.011325129307806492, + 0.011879903264343739, + 0.014843350276350975, + 0.00956796482205391, + 0.015916015952825546, + 0.011496060527861118, + 0.01523197628557682, + 0.006542818155139685, + 0.011258433572947979, + 0.009980118833482265, + 0.007267508655786514, + 0.006594883743673563, + 0.013873272575438023, + 0.012010914273560047, + 0.008672398515045643, + 0.010072363540530205, + 0.01215326227247715, + 0.007534414995461702, + 0.01633218303322792, + 0.017389265820384026, + 0.02020169049501419, + 0.011926453560590744, + 0.00938449427485466, + 0.013056553900241852, + 0.008687827736139297, + 0.010445136576890945, + 0.011284447275102139, + 0.01116707269102335, + 0.021038753911852837, + 0.016826177015900612, + 0.011269011534750462, + 0.012043769471347332, + 0.009095847606658936, + 0.010217166505753994, + 0.006546500604599714, + 0.013250171206891537, + 0.008699260652065277, + 0.007257232442498207, + 0.013530879281461239, + 0.013806631788611412, + 0.014962879940867424, + 0.009291879832744598, + 0.017858628183603287, + 0.030717261135578156, + 0.011766216717660427, + 0.017694178968667984, + 0.021751364693045616, + 0.010346855968236923, + 0.00745270075276494, + 0.010456745512783527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.016279619187116623 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic cherry-red race car with a widebody kit and a massive carbon fiber rear wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car performing an extreme drift on an asphalt track that appears scorched by the afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pitch violently to one side while its rear wheels lose grip and rub against the cracked, rough surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, nearly opaque cloud of white smoke billowing from the rear wheels, mixed with yellow dust swirling beneath the glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld view that pans with the car's violent slide?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014203879982233047 + ], + "opposite_logprob_tokens": [ + -4.264204025268555 + ], + "logprob_reward": -0.014203879982233047, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014203879982233047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17444071173667908, + 0.12596142292022705, + 0.0214664489030838, + 0.049278922379016876, + 0.01816871576011181, + 0.22121945023536682, + 0.06659608334302902, + 0.12135045975446701, + 0.1387881487607956, + 0.06272964924573898 + ], + "credit_map": [ + 0.021041622385382652, + 0.025791723281145096, + 0.021871743723750114, + 0.032372698187828064, + 0.01496074628084898, + 0.01731129363179207, + 0.013845685869455338, + 0.02724519371986389, + 0.01891312375664711, + 0.014956101775169373, + 0.01760476641356945, + 0.011053672060370445, + 0.016994718462228775, + 0.01659543067216873, + 0.011383023113012314, + 0.018460582941770554, + 0.0016855334397405386, + 0.0024522156454622746, + 0.0030937015544623137, + 0.0023557867389172316, + 0.0027865679003298283, + 0.002633088966831565, + 0.002418023534119129, + 0.004041530657559633, + 0.004490796942263842, + 0.004640650004148483, + 0.0038038580678403378, + 0.0046076965518295765, + 0.007000101264566183, + 0.005970621947199106, + 0.005796625278890133, + 0.012968569062650204, + 0.0035263460595160723, + 0.0015365456929430366, + 0.002323061926290393, + 0.001979722175747156, + 0.0026940342504531145, + 0.0013508066767826676, + 0.0023774458095431328, + 0.002380751771852374, + 0.05255170166492462, + 0.010269170626997948, + 0.09739378094673157, + 0.01804880052804947, + 0.0060676406137645245, + 0.008204263634979725, + 0.01606537587940693, + 0.012618707492947578, + 0.008677835576236248, + 0.009866766631603241, + 0.012936905026435852, + 0.009863353334367275, + 0.005518788937479258, + 0.006523520685732365, + 0.006767647806555033, + 0.006441262550652027, + 0.018384821712970734, + 0.014915594831109047, + 0.012625920586287975, + 0.014254827983677387, + 0.02173655293881893, + 0.01245203334838152, + 0.013272149488329887, + 0.0137085672467947, + 0.012880356051027775, + 0.020440753549337387, + 0.016835203394293785, + 0.012563728727400303, + 0.022528033703565598, + 0.0306645967066288, + 0.010447482578456402, + 0.012427995912730694, + 0.010522288270294666, + 0.008194575086236, + 0.007995924912393093, + 0.011244178749620914, + 0.006366148125380278, + 0.007768227718770504, + 0.004080208484083414, + 0.006558098830282688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.014203879982233047 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic cherry-red race car with a widebody kit and a massive carbon fiber rear wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car performing an extreme drift on an asphalt track that appears scorched by the afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pitch violently to one side while its rear wheels lose grip and rub against the cracked, rough surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, nearly opaque cloud of white smoke billowing from the rear wheels, mixed with yellow dust swirling beneath the glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld view that pans with the car's violent slide?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059316821396351 + ], + "opposite_logprob_tokens": [ + -3.8955931663513184 + ], + "logprob_reward": -0.02059316821396351, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059316821396351, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13903655111789703, + 0.1478084921836853, + 0.048471417278051376, + 0.050371237099170685, + 0.041134659200906754, + 0.11740211397409439, + 0.14941442012786865, + 0.11772754788398743, + 0.09315259009599686, + 0.09548100084066391 + ], + "credit_map": [ + 0.013843649998307228, + 0.020584847778081894, + 0.020415527746081352, + 0.024424845352768898, + 0.013790789991617203, + 0.014118485152721405, + 0.01290698442608118, + 0.018951416015625, + 0.031003110110759735, + 0.015152839943766594, + 0.02413269504904747, + 0.01148150209337473, + 0.015406222082674503, + 0.016295958310365677, + 0.012018656358122826, + 0.02231750078499317, + 0.003987798001617193, + 0.005404098425060511, + 0.004089532885700464, + 0.004792045336216688, + 0.003914623986929655, + 0.010222941637039185, + 0.003069329774007201, + 0.012991046532988548, + 0.005844165571033955, + 0.0043098763562738895, + 0.00503925746306777, + 0.0061943307518959045, + 0.006075030192732811, + 0.0054433452896773815, + 0.007252309937030077, + 0.010212919674813747, + 0.005829653237015009, + 0.004620606545358896, + 0.007268837187439203, + 0.004290428012609482, + 0.005782952532172203, + 0.0033142445608973503, + 0.005496112164109945, + 0.004531826823949814, + 0.0277712345123291, + 0.007276951801031828, + 0.045527614653110504, + 0.0092764962464571, + 0.005853778682649136, + 0.005282345227897167, + 0.009990094229578972, + 0.006423598621040583, + 0.021115748211741447, + 0.02087063156068325, + 0.02747262641787529, + 0.019431456923484802, + 0.01182161457836628, + 0.015031191520392895, + 0.017690764740109444, + 0.015980392694473267, + 0.019432242959737778, + 0.013825231231749058, + 0.012183480896055698, + 0.010749957524240017, + 0.019682185724377632, + 0.013515726663172245, + 0.012331534177064896, + 0.016007186844944954, + 0.007956943474709988, + 0.01384203601628542, + 0.010771221481263638, + 0.013828037306666374, + 0.011184086091816425, + 0.02022845856845379, + 0.008488218300044537, + 0.006853584200143814, + 0.012921074405312538, + 0.012538114562630653, + 0.009597563184797764, + 0.016595568507909775, + 0.013167406432330608, + 0.013476911932229996, + 0.007175578735768795, + 0.01000878307968378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.02059316821396351 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic cherry-red race car with a widebody kit and a massive carbon fiber rear wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car performing an extreme drift on an asphalt track that appears scorched by the afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pitch violently to one side while its rear wheels lose grip and rub against the cracked, rough surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, nearly opaque cloud of white smoke billowing from the rear wheels, mixed with yellow dust swirling beneath the glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld view that pans with the car's violent slide?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017131189815700054 + ], + "opposite_logprob_tokens": [ + -6.376713275909424 + ], + "logprob_reward": -0.0017131189815700054, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017131189815700054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04664906486868858, + 0.041896142065525055, + 0.08502556383609772, + 0.023504631593823433, + 0.052216991782188416, + 0.07701954245567322, + 0.11414685100317001, + 0.1449892520904541, + 0.19271916151046753, + 0.2218327522277832 + ], + "credit_map": [ + 0.003963956143707037, + 0.005505929235368967, + 0.005303786601871252, + 0.005988070275634527, + 0.006512596271932125, + 0.0067094978876411915, + 0.0040590520948171616, + 0.008606184273958206, + 0.008994946256279945, + 0.004065524321049452, + 0.006416651885956526, + 0.003599273506551981, + 0.003726549446582794, + 0.004011604003608227, + 0.004774609114974737, + 0.00630698399618268, + 0.013035918585956097, + 0.009908784180879593, + 0.008368835784494877, + 0.012686180882155895, + 0.011718026362359524, + 0.00979954656213522, + 0.006766179576516151, + 0.012742103077471256, + 0.003743545850738883, + 0.002683684229850769, + 0.0024532731622457504, + 0.002960572252050042, + 0.0023319192696362734, + 0.002322447719052434, + 0.002223810413852334, + 0.004785382654517889, + 0.010909488424658775, + 0.005357188172638416, + 0.007068927865475416, + 0.005954842083156109, + 0.006294550374150276, + 0.00442254263907671, + 0.0065864878706634045, + 0.005622969474643469, + 0.01678452081978321, + 0.004255818668752909, + 0.031778957694768906, + 0.006308977957814932, + 0.003069436876103282, + 0.004345286171883345, + 0.005785711109638214, + 0.004690841306000948, + 0.01737588457763195, + 0.01818392425775528, + 0.021266145631670952, + 0.013896279968321323, + 0.009721320122480392, + 0.013166791759431362, + 0.01008028257638216, + 0.010456236079335213, + 0.02623409777879715, + 0.019270721822977066, + 0.01596430502831936, + 0.015030919574201107, + 0.022295217961072922, + 0.014989457093179226, + 0.015388857573270798, + 0.015815693885087967, + 0.012848235666751862, + 0.031803783029317856, + 0.01893218420445919, + 0.027913378551602364, + 0.03354291617870331, + 0.03972671926021576, + 0.015097955241799355, + 0.01285401452332735, + 0.021421050652861595, + 0.042550455778837204, + 0.015593732707202435, + 0.04130769148468971, + 0.03314872831106186, + 0.02968944050371647, + 0.01446316484361887, + 0.023658515885472298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.0017131189815700054 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic cherry-red race car with a widebody kit and a massive carbon fiber rear wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car performing an extreme drift on an asphalt track that appears scorched by the afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pitch violently to one side while its rear wheels lose grip and rub against the cracked, rough surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, nearly opaque cloud of white smoke billowing from the rear wheels, mixed with yellow dust swirling beneath the glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld view that pans with the car's violent slide?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014188953675329685 + ], + "opposite_logprob_tokens": [ + -4.264188766479492 + ], + "logprob_reward": -0.014188953675329685, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014188953675329685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05393099784851074, + 0.05418790876865387, + 0.22749443352222443, + 0.029482562094926834, + 0.05900115147233009, + 0.15676483511924744, + 0.14368480443954468, + 0.11509695649147034, + 0.10233771055936813, + 0.058018606156110764 + ], + "credit_map": [ + 0.007743282709270716, + 0.006863832473754883, + 0.006792346481233835, + 0.005783070344477892, + 0.007189650554209948, + 0.005754657555371523, + 0.0052059609442949295, + 0.008598199114203453, + 0.01079112570732832, + 0.00541058462113142, + 0.007400485221296549, + 0.004671970847994089, + 0.007646874058991671, + 0.006713868584483862, + 0.0041136350482702255, + 0.007439367938786745, + 0.02997531369328499, + 0.020528849214315414, + 0.019729087129235268, + 0.025856584310531616, + 0.021849431097507477, + 0.041756510734558105, + 0.015579476952552795, + 0.052219197154045105, + 0.003370686899870634, + 0.0032597242388874292, + 0.003641413291916251, + 0.00439370796084404, + 0.004000173881649971, + 0.0029294034466147423, + 0.002790472935885191, + 0.005096982233226299, + 0.00567869795486331, + 0.004902153741568327, + 0.00958850048482418, + 0.009903524070978165, + 0.00757026020437479, + 0.003479474689811468, + 0.01155032217502594, + 0.006328223738819361, + 0.027912849560379982, + 0.008505982346832752, + 0.0469386950135231, + 0.018944526091217995, + 0.00994354858994484, + 0.01703934371471405, + 0.014501873403787613, + 0.012978026643395424, + 0.019601574167609215, + 0.021953891962766647, + 0.023025033995509148, + 0.01691330596804619, + 0.013245667330920696, + 0.018393436446785927, + 0.014825385063886642, + 0.015726517885923386, + 0.011811476200819016, + 0.011682256124913692, + 0.015217077918350697, + 0.01342250220477581, + 0.020360464230179787, + 0.014616093598306179, + 0.011537923477590084, + 0.01644917018711567, + 0.007639516144990921, + 0.01804739236831665, + 0.007987831719219685, + 0.018963200971484184, + 0.017921045422554016, + 0.015918821096420288, + 0.009768368676304817, + 0.006091548595577478, + 0.0057053882628679276, + 0.011645244434475899, + 0.004758518189191818, + 0.009714850224554539, + 0.00959627516567707, + 0.00586320273578167, + 0.0035749110393226147, + 0.007160218898206949 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 6 + ], + "selected_reward": -0.014188953675329685 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle, handheld shot closely follows a heavily modified, metallic cherry-red race car as it performs an extreme drift on an asphalt track scorched by the afternoon sun. The car, adorned with a widebody kit and a massive carbon fiber rear wing, pitches violently to one side at high speed. Its rear wheels instantly lose grip, rubbing violently against the cracked, rough surface. With a piercing tire screeching, a thick, nearly opaque cloud of white smoke billows out, instantly engulfing the rear of the car. A cloud of yellow dust and white smoke swirls beneath the glaring sunlight, which reflects off the car's smooth, curved surface in flowing patterns. The camera pans with the car's violent slide, capturing the car's graceful yet perilous sideways glide through the curve on the edge of control and loss of control. The entire scene is filled with raw mechanical power and adrenaline-pumping tension.", + "seed": 20847, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_1/index_4591_seed_20847_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.011455832654610277, + "vlm_logprob": -0.011455832654610277, + "vlm_yes_no_margin": 4.9, + "vlm_token_credit": -0.011455832654610277, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic cherry-red race car with a widebody kit and a massive carbon fiber rear wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car performing an extreme drift on an asphalt track that appears scorched by the afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pitch violently to one side while its rear wheels lose grip and rub against the cracked, rough surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, nearly opaque cloud of white smoke billowing from the rear wheels, mixed with yellow dust swirling beneath the glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld view that pans with the car's violent slide?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011903680860996246 + ], + "opposite_logprob_tokens": [ + -5.136903762817383 + ], + "logprob_reward": -0.011903680860996246, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.011903680860996246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03485285863280296, + 0.7113321423530579, + 0.02012237347662449, + 0.022942421957850456, + 0.01483864150941372, + 0.010582578368484974, + 0.043849777430295944, + 0.044397998601198196, + 0.041357625275850296, + 0.055723633617162704 + ], + "credit_map": [ + 0.004139587748795748, + 0.0025699471589177847, + 0.005184144247323275, + 0.004644704516977072, + 0.003541691694408655, + 0.0037713460624217987, + 0.005451574921607971, + 0.005549859255552292, + 0.04439438879489899, + 0.0902525931596756, + 0.1955406665802002, + 0.07329355180263519, + 0.07092801481485367, + 0.03882754594087601, + 0.0652494803071022, + 0.13284584879875183, + 0.0021238226909190416, + 0.001863318495452404, + 0.0023254614789038897, + 0.0019277320243418217, + 0.0018090116791427135, + 0.004459360148757696, + 0.0031609234865754843, + 0.0024527430068701506, + 0.0013060311321169138, + 0.003402682486921549, + 0.0036540816072374582, + 0.002648076042532921, + 0.0032144736032932997, + 0.0030471112113445997, + 0.0032411012798547745, + 0.0024288624990731478, + 0.0010578783694654703, + 0.001039607566781342, + 0.0014504072023555636, + 0.001828471664339304, + 0.001908216392621398, + 0.0012795887887477875, + 0.004005845636129379, + 0.002268624259158969, + 0.0010941626969724894, + 0.0010779984295368195, + 0.0009012772934511304, + 0.0013494860613718629, + 0.001345091382972896, + 0.001487372093833983, + 0.001998303225263953, + 0.0013288863701745868, + 0.0064316559582948685, + 0.004141690209507942, + 0.003652542596682906, + 0.0030756150372326374, + 0.005799705162644386, + 0.007911003194749355, + 0.005729759577661753, + 0.007107803598046303, + 0.003928073216229677, + 0.003076828783378005, + 0.004834073130041361, + 0.007030435837805271, + 0.008447385393083096, + 0.006918305065482855, + 0.005072529427707195, + 0.0050903670489788055, + 0.00600234791636467, + 0.004386641550809145, + 0.007637727539986372, + 0.004411980044096708, + 0.005238237325102091, + 0.006219799630343914, + 0.003672599559649825, + 0.0037882919423282146, + 0.0055783092975616455, + 0.0067025646567344666, + 0.005566725041717291, + 0.004198321141302586, + 0.011430171318352222, + 0.006778744515031576, + 0.010675146244466305, + 0.004793649539351463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 7 + ], + "selected_reward": -0.011903680860996246 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic cherry-red race car with a widebody kit and a massive carbon fiber rear wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car performing an extreme drift on an asphalt track that appears scorched by the afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pitch violently to one side while its rear wheels lose grip and rub against the cracked, rough surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, nearly opaque cloud of white smoke billowing from the rear wheels, mixed with yellow dust swirling beneath the glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld view that pans with the car's violent slide?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009787782095372677 + ], + "opposite_logprob_tokens": [ + -4.634787559509277 + ], + "logprob_reward": -0.009787782095372677, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009787782095372677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03792206570506096, + 0.610758900642395, + 0.012715116143226624, + 0.02219630591571331, + 0.013692089356482029, + 0.013307783752679825, + 0.012884024530649185, + 0.03582581877708435, + 0.16777126491069794, + 0.07292664051055908 + ], + "credit_map": [ + 0.005677466280758381, + 0.002645892323926091, + 0.009079103358089924, + 0.00449299206957221, + 0.0024383587297052145, + 0.0017287764931097627, + 0.005762120708823204, + 0.006097354926168919, + 0.039925333112478256, + 0.08370883017778397, + 0.15649382770061493, + 0.0611196793615818, + 0.06566765159368515, + 0.03494001924991608, + 0.05871083587408066, + 0.1101926788687706, + 0.0016197594814002514, + 0.0010965804103761911, + 0.0012995764845982194, + 0.0015935273841023445, + 0.0010469163535162807, + 0.0018191479612141848, + 0.0026710988022387028, + 0.0015685094986110926, + 0.0015755785861983895, + 0.0024776856880635023, + 0.0025163704995065928, + 0.0026779405307024717, + 0.0038184295408427715, + 0.002952005248516798, + 0.0035307954531162977, + 0.002647499553859234, + 0.0007729334174655378, + 0.0011314251460134983, + 0.0012304375413805246, + 0.0016010269755497575, + 0.0017418299103155732, + 0.0011702119372785091, + 0.003760592546314001, + 0.00228363205678761, + 0.001465570298023522, + 0.001559610478579998, + 0.0012971301330253482, + 0.0019024513894692063, + 0.0016074466984719038, + 0.0014555209781974554, + 0.0020744777284562588, + 0.001945575582794845, + 0.0015367664163932204, + 0.0011459034867584705, + 0.0016741483705118299, + 0.0019985183607786894, + 0.0013123666867613792, + 0.0017894326010718942, + 0.0016875426517799497, + 0.0017393451416864991, + 0.0029324425850063562, + 0.002304556779563427, + 0.00350221898406744, + 0.005263120401650667, + 0.005547936540096998, + 0.00584478210657835, + 0.006545905955135822, + 0.003884852398186922, + 0.012982388027012348, + 0.020412415266036987, + 0.025493083521723747, + 0.020036106929183006, + 0.024042747914791107, + 0.02532614953815937, + 0.020558154210448265, + 0.01892021857202053, + 0.006029173266142607, + 0.007205389905720949, + 0.008514663204550743, + 0.0069993338547647, + 0.013014265336096287, + 0.01013360358774662, + 0.01209003571420908, + 0.008940177038311958 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 9 + ], + "selected_reward": -0.009787782095372677 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic cherry-red race car with a widebody kit and a massive carbon fiber rear wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car performing an extreme drift on an asphalt track that appears scorched by the afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pitch violently to one side while its rear wheels lose grip and rub against the cracked, rough surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, nearly opaque cloud of white smoke billowing from the rear wheels, mixed with yellow dust swirling beneath the glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld view that pans with the car's violent slide?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018193509429693222 + ], + "opposite_logprob_tokens": [ + -4.01819372177124 + ], + "logprob_reward": -0.018193509429693222, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018193509429693222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049085184931755066, + 0.7427486181259155, + 0.0152380196377635, + 0.02585950680077076, + 0.03546638786792755, + 0.010256639681756496, + 0.02329023741185665, + 0.025308093056082726, + 0.018355464562773705, + 0.054391901940107346 + ], + "credit_map": [ + 0.008561349473893642, + 0.0023710359819233418, + 0.011933940462768078, + 0.004972724709659815, + 0.0031136153265833855, + 0.0029430496506392956, + 0.006105717271566391, + 0.009083747863769531, + 0.04804876446723938, + 0.09862486273050308, + 0.19615602493286133, + 0.07507938891649246, + 0.07774924486875534, + 0.04186864569783211, + 0.06959454715251923, + 0.1356271207332611, + 0.0018896351102739573, + 0.0013582282699644566, + 0.0018615834414958954, + 0.0014734120341017842, + 0.0015638445038348436, + 0.003054504981264472, + 0.0020827408879995346, + 0.001954068895429373, + 0.0015042219310998917, + 0.0032924688421189785, + 0.0037256567738950253, + 0.0031034715939313173, + 0.003661962691694498, + 0.0037145379465073347, + 0.003822849364951253, + 0.003034335095435381, + 0.0025460978504270315, + 0.0038695905823260546, + 0.004629114642739296, + 0.003281103679910302, + 0.005369088612496853, + 0.003911946434527636, + 0.007427960634231567, + 0.00443148473277688, + 0.0011275401338934898, + 0.0009423561859875917, + 0.0009537663427181542, + 0.0016876839799806476, + 0.0014556106179952621, + 0.0013460282934829593, + 0.001364562544040382, + 0.001379090710543096, + 0.003000600729137659, + 0.0029738997109234333, + 0.0038728127256035805, + 0.0021184771321713924, + 0.002551896497607231, + 0.0037942160852253437, + 0.0018866591854020953, + 0.003091671969741583, + 0.003131527453660965, + 0.0019037615275010467, + 0.002802066272124648, + 0.0034688326995819807, + 0.0035705880727618933, + 0.0038953714538365602, + 0.0038543713744729757, + 0.0026815738528966904, + 0.0025064717046916485, + 0.0019088328117504716, + 0.0030204730574041605, + 0.0017596138641238213, + 0.001994009595364332, + 0.002415480324998498, + 0.002172463806346059, + 0.0025781195145100355, + 0.005369246006011963, + 0.005726467818021774, + 0.006357688922435045, + 0.004930428229272366, + 0.008714604191482067, + 0.0065277572721242905, + 0.010685297660529613, + 0.006080406717956066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.018193509429693222 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic cherry-red race car with a widebody kit and a massive carbon fiber rear wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car performing an extreme drift on an asphalt track that appears scorched by the afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pitch violently to one side while its rear wheels lose grip and rub against the cracked, rough surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, nearly opaque cloud of white smoke billowing from the rear wheels, mixed with yellow dust swirling beneath the glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld view that pans with the car's violent slide?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001335324952378869 + ], + "opposite_logprob_tokens": [ + -6.626335144042969 + ], + "logprob_reward": -0.001335324952378869, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001335324952378869, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08469072729349136, + 0.2844878137111664, + 0.03180320933461189, + 0.029944781213998795, + 0.03774794563651085, + 0.023281440138816833, + 0.05377456545829773, + 0.1083606630563736, + 0.2325429767370224, + 0.11336591094732285 + ], + "credit_map": [ + 0.013862096704542637, + 0.004398612305521965, + 0.02012822777032852, + 0.009220787324011326, + 0.005711629055440426, + 0.005259441677480936, + 0.01270463690161705, + 0.013405285775661469, + 0.020696304738521576, + 0.0454380102455616, + 0.05933878570795059, + 0.02705094777047634, + 0.03599384054541588, + 0.018996063619852066, + 0.030725594609975815, + 0.04624827206134796, + 0.00404112646356225, + 0.0025276425294578075, + 0.0034942745696753263, + 0.0032037575729191303, + 0.0028856853023171425, + 0.006364203058183193, + 0.005213033873587847, + 0.00407348433509469, + 0.0017266483046114445, + 0.0036718142218887806, + 0.004058583173900843, + 0.0036222029011696577, + 0.005372584797441959, + 0.004123202059417963, + 0.0039668008685112, + 0.0034029448870569468, + 0.0022371727973222733, + 0.0036271039862185717, + 0.004185188096016645, + 0.0038946839049458504, + 0.005248669534921646, + 0.003340590512380004, + 0.009394335560500622, + 0.005820197984576225, + 0.0023543373681604862, + 0.002576980274170637, + 0.0017620138823986053, + 0.0038807974196970463, + 0.003774677636101842, + 0.002743361983448267, + 0.0035109694581478834, + 0.0026783014182001352, + 0.00791796576231718, + 0.005055172834545374, + 0.005077432841062546, + 0.0063409036956727505, + 0.006038065068423748, + 0.007493171375244856, + 0.006813127547502518, + 0.009038724936544895, + 0.009652881883084774, + 0.008092258125543594, + 0.009187632240355015, + 0.018760619685053825, + 0.019557129591703415, + 0.018260523676872253, + 0.013835537247359753, + 0.011014079675078392, + 0.019586026668548584, + 0.028511246666312218, + 0.03347820043563843, + 0.030013244599103928, + 0.0334252305328846, + 0.03297103941440582, + 0.02952621690928936, + 0.025031767785549164, + 0.012561955489218235, + 0.011464635841548443, + 0.016102856025099754, + 0.00949228648096323, + 0.019852053374052048, + 0.014952285215258598, + 0.014109988696873188, + 0.014829853549599648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 9 + ], + "selected_reward": -0.001335324952378869 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic cherry-red race car with a widebody kit and a massive carbon fiber rear wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car performing an extreme drift on an asphalt track that appears scorched by the afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pitch violently to one side while its rear wheels lose grip and rub against the cracked, rough surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, nearly opaque cloud of white smoke billowing from the rear wheels, mixed with yellow dust swirling beneath the glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld view that pans with the car's violent slide?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016058865934610367 + ], + "opposite_logprob_tokens": [ + -4.141058921813965 + ], + "logprob_reward": -0.016058865934610367, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016058865934610367, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0897025540471077, + 0.6056160926818848, + 0.03402978554368019, + 0.018090812489390373, + 0.043929848819971085, + 0.010526676662266254, + 0.023514682427048683, + 0.03850630298256874, + 0.05393611267209053, + 0.08214706927537918 + ], + "credit_map": [ + 0.020310157909989357, + 0.005145510658621788, + 0.030514702200889587, + 0.007023505866527557, + 0.0038478178903460503, + 0.0057538351975381374, + 0.006700298748910427, + 0.010406731627881527, + 0.04042879864573479, + 0.08948023617267609, + 0.1443021297454834, + 0.0593884103000164, + 0.06862053275108337, + 0.03560738265514374, + 0.061528246849775314, + 0.10626034438610077, + 0.003855203976854682, + 0.003980291076004505, + 0.003739949082955718, + 0.005083103664219379, + 0.0035693079698830843, + 0.004160984419286251, + 0.005569043103605509, + 0.0040719034150242805, + 0.001027058926410973, + 0.002565260510891676, + 0.0021726805716753006, + 0.0017019135411828756, + 0.004252478014677763, + 0.0024906930048018694, + 0.0020361696369946003, + 0.0018445579335093498, + 0.0023547580931335688, + 0.003903165925294161, + 0.002809898927807808, + 0.005072769708931446, + 0.006705069448798895, + 0.003870841581374407, + 0.01173916645348072, + 0.007474182639271021, + 0.0012639492051675916, + 0.0011616847477853298, + 0.0009180399356409907, + 0.0012334076454862952, + 0.0014267575461417437, + 0.0013763562310487032, + 0.0016351257218047976, + 0.0015113562112674117, + 0.003018446732312441, + 0.0023948512971401215, + 0.002935930388048291, + 0.0030142536852508783, + 0.0024793767370283604, + 0.0037143148947507143, + 0.0027183773927390575, + 0.00323913199827075, + 0.003276935312896967, + 0.0028034152928739786, + 0.003822598373517394, + 0.005749037954956293, + 0.006970567628741264, + 0.006349426694214344, + 0.004893573932349682, + 0.004640747793018818, + 0.005732203368097544, + 0.006670266389846802, + 0.005928783677518368, + 0.006719171069562435, + 0.0075053079053759575, + 0.00744415819644928, + 0.007104791235178709, + 0.006831432692706585, + 0.009974698536098003, + 0.007500388193875551, + 0.01486952044069767, + 0.009917962364852428, + 0.009319884702563286, + 0.010403003543615341, + 0.00591182429343462, + 0.014249793253839016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.016058865934610367 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_10/index_tensor([2016, 2016])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_10/index_tensor([2016, 2016])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..575fb43106703d756b754c31547188a31d16e5b1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_10/index_tensor([2016, 2016])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a slowly zooming close-up, a delicate memory stick hovers in a pure, profound black void. The world is dominated by a high-contrast black and white palette, the sole source of light coming from a sharp spotlight emanating from outside the frame. The stick's dark substrate is covered in a labyrinth of intricate silver circuitry, which gleams with a cold metallic sheen in the light. The square integrated circuit chip has a matte finish, while the neatly arranged gold-plated fingers at one end capture and reflect a dazzling, almost pure white light as they rotate. It performs an elegant and deliberate gyroscopic motion, steadily rotating around its long axis while simultaneously tumbling slowly end-to-end. Each rotation causes light and shadow to dance fluidly across its surface, outlining every minute structural detail, creating a surreal, hypnotic, cinematic quality that is a chilling celebration of technological beauty.", + "seed": 27720, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_10/index_2016_seed_27720_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.20424212850630283, + "vlm_logprob": -0.20424212850630283, + "vlm_yes_no_margin": 2.0249999880790712, + "vlm_token_credit": -0.20424212850630283, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a memory stick with visible silver circuitry and a gold-plated connector?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure black void with a high-contrast black and white palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a sharp spotlight originating from outside the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the memory stick perform a combined rotation around its long axis and a slow end-to-end tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow zoom-in movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03394080698490143 + ], + "opposite_logprob_tokens": [ + -3.5339407920837402 + ], + "logprob_reward": -0.03394080698490143, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03394080698490143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10239453613758087, + 0.0585363544523716, + 0.5731402635574341, + 0.04425657540559769, + 0.022178683429956436, + 0.031752899289131165, + 0.026931792497634888, + 0.023253675550222397, + 0.02684628590941429, + 0.0907088965177536 + ], + "credit_map": [ + 0.012506753206253052, + 0.008210474625229836, + 0.011624980717897415, + 0.013514513149857521, + 0.011062052100896835, + 0.019363122060894966, + 0.010229531675577164, + 0.015883123502135277, + 0.009255069307982922, + 0.009358037263154984, + 0.007063419558107853, + 0.01634037122130394, + 0.00294071389362216, + 0.004689653869718313, + 0.0031071952544152737, + 0.005781901068985462, + 0.01837346702814102, + 0.011445247568190098, + 0.008251466788351536, + 0.009923020377755165, + 0.010672030039131641, + 0.2606869637966156, + 0.009919383563101292, + 0.24386873841285706, + 0.007421240210533142, + 0.004902138840407133, + 0.007155652157962322, + 0.004072172101587057, + 0.003139237407594919, + 0.005753723904490471, + 0.00467191357165575, + 0.007140504661947489, + 0.00269153225235641, + 0.0031348569318652153, + 0.0039410968311131, + 0.0018294386100023985, + 0.0016412381082773209, + 0.0032436323817819357, + 0.002509733196347952, + 0.003187159774824977, + 0.0036358924116939306, + 0.002878637984395027, + 0.004689718130975962, + 0.002969572087749839, + 0.002511815167963505, + 0.004675031639635563, + 0.003672172548249364, + 0.006720064207911491, + 0.004076537676155567, + 0.0031753284856677055, + 0.004624425899237394, + 0.0020164402667433023, + 0.001859012758359313, + 0.00400901772081852, + 0.0028208508156239986, + 0.004350183065980673, + 0.0033779439982026815, + 0.00307375262491405, + 0.003361764596775174, + 0.0018520931480452418, + 0.0017930959584191442, + 0.0035308923106640577, + 0.0023669907823204994, + 0.003897146787494421, + 0.004162450786679983, + 0.00284654856659472, + 0.004048499278724194, + 0.0022615515626966953, + 0.0025224059354513884, + 0.0044403960928320885, + 0.0031210167799144983, + 0.003443419700488448, + 0.012483082711696625, + 0.008823494426906109, + 0.013895378448069096, + 0.01213555783033371, + 0.007454978302121162, + 0.01348846685141325, + 0.012826267629861832, + 0.009601681493222713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.03394080698490143 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a memory stick with visible silver circuitry and a gold-plated connector?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure black void with a high-contrast black and white palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a sharp spotlight originating from outside the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the memory stick perform a combined rotation around its long axis and a slow end-to-end tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow zoom-in movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486289978027344 + ], + "opposite_logprob_tokens": [ + -1.2236289978027344 + ], + "logprob_reward": -0.3486289978027344, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486289978027344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06063781678676605, + 0.034817617386579514, + 0.42728304862976074, + 0.0216640904545784, + 0.013408483006060123, + 0.012334893457591534, + 0.01581357605755329, + 0.014029481448233128, + 0.10984132438898087, + 0.2901696562767029 + ], + "credit_map": [ + 0.010966084897518158, + 0.0038637069519609213, + 0.007846382446587086, + 0.0076826331205666065, + 0.003755716374143958, + 0.0092213936150074, + 0.004612497519701719, + 0.012689399532973766, + 0.00546829542145133, + 0.004568910226225853, + 0.003382667200639844, + 0.009998803958296776, + 0.0022159749642014503, + 0.0028017940931022167, + 0.0024137706495821476, + 0.003967401571571827, + 0.015239516273140907, + 0.009066274389624596, + 0.009289414621889591, + 0.00826358050107956, + 0.007846140302717686, + 0.19102656841278076, + 0.007693869061768055, + 0.178857684135437, + 0.005349180661141872, + 0.0020324888173490763, + 0.00490872748196125, + 0.001564012374728918, + 0.0009819469414651394, + 0.0031579334754496813, + 0.001144004869274795, + 0.002525797812268138, + 0.0017778718611225486, + 0.0010991788003593683, + 0.0033213989809155464, + 0.0006421909783966839, + 0.001065694261342287, + 0.00355824688449502, + 0.0009143720380961895, + 0.0010295292595401406, + 0.0011491329642012715, + 0.0008872259058989584, + 0.002310636453330517, + 0.0010342702735215425, + 0.0007590712048113346, + 0.0023421593941748142, + 0.001492135925218463, + 0.002360261045396328, + 0.0017455779016017914, + 0.0010948512936010957, + 0.0019789806101471186, + 0.0010037998436018825, + 0.0009982322808355093, + 0.002826387295499444, + 0.0016197243239730597, + 0.004546023439615965, + 0.001677813590504229, + 0.0013661427656188607, + 0.0020588848274201155, + 0.0010081997606903315, + 0.0006878594867885113, + 0.002580022905021906, + 0.0012416793033480644, + 0.003408879041671753, + 0.015151097439229488, + 0.00989307276904583, + 0.016651898622512817, + 0.00753896776586771, + 0.008188361302018166, + 0.019175883382558823, + 0.012792202644050121, + 0.020449845120310783, + 0.04078036919236183, + 0.022076072171330452, + 0.05635146051645279, + 0.025924421846866608, + 0.023737270385026932, + 0.04696907848119736, + 0.03333220258355141, + 0.04099879041314125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 8 + ], + "selected_reward": -0.3486289978027344 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a memory stick with visible silver circuitry and a gold-plated connector?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure black void with a high-contrast black and white palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a sharp spotlight originating from outside the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the memory stick perform a combined rotation around its long axis and a slow end-to-end tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow zoom-in movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42893558740615845 + ], + "opposite_logprob_tokens": [ + -1.0539355278015137 + ], + "logprob_reward": -0.42893558740615845, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.42893558740615845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11375117301940918, + 0.04564701393246651, + 0.5685820579528809, + 0.059615157544612885, + 0.011098017916083336, + 0.03495927155017853, + 0.037978630512952805, + 0.022056061774492264, + 0.01974630542099476, + 0.08656623959541321 + ], + "credit_map": [ + 0.01474755723029375, + 0.00630134018138051, + 0.014277820475399494, + 0.01116915326565504, + 0.010438773781061172, + 0.023800304159522057, + 0.009008553810417652, + 0.024007681757211685, + 0.00673864409327507, + 0.0087434658780694, + 0.006240869406610727, + 0.00873362086713314, + 0.003424969967454672, + 0.003923902753740549, + 0.0027926156762987375, + 0.005048932041972876, + 0.022693894803524017, + 0.01177810039371252, + 0.010891781188547611, + 0.010462558828294277, + 0.010928718373179436, + 0.2560841739177704, + 0.009361892007291317, + 0.23638097941875458, + 0.016259238123893738, + 0.00569689879193902, + 0.013121556490659714, + 0.004607229493558407, + 0.0025737418327480555, + 0.009559005498886108, + 0.0024103906471282244, + 0.005387101788073778, + 0.001714997342787683, + 0.0010860186303034425, + 0.0020448213908821344, + 0.0009019074495881796, + 0.0006980281905271113, + 0.0017703870544210076, + 0.0010026291711255908, + 0.0018792295595631003, + 0.0032997436355799437, + 0.002247604075819254, + 0.005873267073184252, + 0.0026094738859683275, + 0.002415032358840108, + 0.005954965949058533, + 0.004534854553639889, + 0.00802433118224144, + 0.005247634369879961, + 0.003352832980453968, + 0.008650923147797585, + 0.0032595768570899963, + 0.0026233112439513206, + 0.0069953761994838715, + 0.003052055137231946, + 0.004796923138201237, + 0.0026458406355232, + 0.0023715882562100887, + 0.0036165097262710333, + 0.0014255426358431578, + 0.0010577442590147257, + 0.0035787068773061037, + 0.0022443875204771757, + 0.0051157427951693535, + 0.003189088311046362, + 0.0017933895578607917, + 0.0027680285274982452, + 0.0013421315234154463, + 0.0016961718210950494, + 0.0031704159919172525, + 0.002481688978150487, + 0.003305393271148205, + 0.011767907999455929, + 0.01024333480745554, + 0.014159628190100193, + 0.008750957436859608, + 0.008071934804320335, + 0.012434440664947033, + 0.010229607112705708, + 0.010908431373536587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.42893558740615845 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a memory stick with visible silver circuitry and a gold-plated connector?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure black void with a high-contrast black and white palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a sharp spotlight originating from outside the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the memory stick perform a combined rotation around its long axis and a slow end-to-end tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow zoom-in movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029810817912220955 + ], + "opposite_logprob_tokens": [ + -3.529810905456543 + ], + "logprob_reward": -0.029810817912220955, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029810817912220955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037204861640930176, + 0.052403487265110016, + 0.7557445168495178, + 0.02135850489139557, + 0.006886545103043318, + 0.009311719797551632, + 0.03159503638744354, + 0.025373097509145737, + 0.017934957519173622, + 0.04218733310699463 + ], + "credit_map": [ + 0.006029641255736351, + 0.0029010504949837923, + 0.0036483267322182655, + 0.004765032324939966, + 0.0031146900728344917, + 0.0066656372509896755, + 0.0029958540108054876, + 0.00708462530747056, + 0.009407729841768742, + 0.00869764480739832, + 0.006555250380188227, + 0.00802403874695301, + 0.0033950256183743477, + 0.005303647834807634, + 0.0035983030684292316, + 0.007421837653964758, + 0.024166101589798927, + 0.014842568896710873, + 0.012604795396327972, + 0.013555587269365788, + 0.014366299845278263, + 0.3421309292316437, + 0.013552581891417503, + 0.320525586605072, + 0.005127408541738987, + 0.0020536654628813267, + 0.00422406941652298, + 0.0017344166990369558, + 0.001049464917741716, + 0.002882143948227167, + 0.001267594750970602, + 0.0030197403393685818, + 0.0009008717024698853, + 0.0008785125683061779, + 0.0011865930864587426, + 0.0005792066804133356, + 0.0005671551334671676, + 0.001338838366791606, + 0.000605045526754111, + 0.000830321863759309, + 0.0011339082848280668, + 0.0009429000201635063, + 0.0016750876093283296, + 0.0010036787716671824, + 0.0005520299891941249, + 0.0015554315177723765, + 0.000984125304967165, + 0.001464557833969593, + 0.0036508950870484114, + 0.002435453701764345, + 0.003871388966217637, + 0.0016325920587405562, + 0.0018355731153860688, + 0.005796162877231836, + 0.0029650588985532522, + 0.009407908655703068, + 0.003784364089369774, + 0.0041058906354010105, + 0.0034188248682767153, + 0.002561141736805439, + 0.0017353732837364078, + 0.0035330092068761587, + 0.0028218405786901712, + 0.003412652062252164, + 0.002055075718089938, + 0.0014456629287451506, + 0.003407473908737302, + 0.0012956465361639857, + 0.0011999462731182575, + 0.003082713345065713, + 0.0016243568388745189, + 0.0038240801077336073, + 0.004891013260930777, + 0.004019512794911861, + 0.007660445757210255, + 0.004336253274232149, + 0.004067084286361933, + 0.007120165973901749, + 0.0038840589113533497, + 0.00620879465714097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 9 + ], + "selected_reward": -0.029810817912220955 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a memory stick with visible silver circuitry and a gold-plated connector?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure black void with a high-contrast black and white palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a sharp spotlight originating from outside the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the memory stick perform a combined rotation around its long axis and a slow end-to-end tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow zoom-in movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17989443242549896 + ], + "opposite_logprob_tokens": [ + -1.8048944473266602 + ], + "logprob_reward": -0.17989443242549896, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17989443242549896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02734972909092903, + 0.05989154055714607, + 0.6604452133178711, + 0.04775082692503929, + 0.007902529090642929, + 0.03282120078802109, + 0.030417554080486298, + 0.02761744149029255, + 0.01824619621038437, + 0.08755778521299362 + ], + "credit_map": [ + 0.003602866316214204, + 0.002837312640622258, + 0.003488104557618499, + 0.003837013151496649, + 0.0030680643394589424, + 0.00345263141207397, + 0.00283170398324728, + 0.004232033155858517, + 0.011738060042262077, + 0.00960472971200943, + 0.005689654033631086, + 0.014256191439926624, + 0.003110290039330721, + 0.004944654181599617, + 0.002907420974224806, + 0.007640542462468147, + 0.02246684581041336, + 0.01277694758027792, + 0.009752470999956131, + 0.011376356706023216, + 0.012644036673009396, + 0.3002568483352661, + 0.011304198764264584, + 0.2798675000667572, + 0.011405258439481258, + 0.004857603460550308, + 0.009085085242986679, + 0.0036237293388694525, + 0.0029208878986537457, + 0.0063734110444784164, + 0.003895670408383012, + 0.0055891782976686954, + 0.0008131833747029305, + 0.0009842937579378486, + 0.0015879177954047918, + 0.0006647029658779502, + 0.0006774540524929762, + 0.001463990076445043, + 0.0006650366703979671, + 0.001045950804837048, + 0.0028869477100670338, + 0.002038339152932167, + 0.005010898690670729, + 0.0023572873324155807, + 0.0021520231384783983, + 0.006388320121914148, + 0.004071284085512161, + 0.0079161012545228, + 0.0035582627169787884, + 0.0022562951780855656, + 0.007346871308982372, + 0.0019521884387359023, + 0.0015220569912344217, + 0.0059953550808131695, + 0.002754961373284459, + 0.00503156241029501, + 0.0029038020875304937, + 0.003547021420672536, + 0.0042808810248970985, + 0.0018130744574591517, + 0.002009555697441101, + 0.004974063485860825, + 0.003192072967067361, + 0.00489696953445673, + 0.0026288358494639397, + 0.0014074950013309717, + 0.0027124988846480846, + 0.0012726652203127742, + 0.001524436636827886, + 0.00388842960819602, + 0.0020757922902703285, + 0.0027360429521650076, + 0.010412695817649364, + 0.009965162724256516, + 0.015533562749624252, + 0.009125962853431702, + 0.008600626140832901, + 0.014419122599065304, + 0.008825893513858318, + 0.01067475788295269 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 1 + ], + "selected_reward": -0.17989443242549896 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a slowly zooming close-up, a delicate memory stick hovers in a pure, profound black void. The world is dominated by a high-contrast black and white palette, the sole source of light coming from a sharp spotlight emanating from outside the frame. The stick's dark substrate is covered in a labyrinth of intricate silver circuitry, which gleams with a cold metallic sheen in the light. The square integrated circuit chip has a matte finish, while the neatly arranged gold-plated fingers at one end capture and reflect a dazzling, almost pure white light as they rotate. It performs an elegant and deliberate gyroscopic motion, steadily rotating around its long axis while simultaneously tumbling slowly end-to-end. Each rotation causes light and shadow to dance fluidly across its surface, outlining every minute structural detail, creating a surreal, hypnotic, cinematic quality that is a chilling celebration of technological beauty.", + "seed": 27721, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_10/index_2016_seed_27721_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.15468772128224373, + "vlm_logprob": -0.15468772128224373, + "vlm_yes_no_margin": 2.475, + "vlm_token_credit": -0.15468772128224373, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a memory stick with visible silver circuitry and a gold-plated connector?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure black void with a high-contrast black and white palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a sharp spotlight originating from outside the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the memory stick perform a combined rotation around its long axis and a slow end-to-end tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow zoom-in movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03396039456129074 + ], + "opposite_logprob_tokens": [ + -3.5339603424072266 + ], + "logprob_reward": -0.03396039456129074, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03396039456129074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17272143065929413, + 0.09356255829334259, + 0.057824570685625076, + 0.11362629383802414, + 0.34308922290802, + 0.07112598419189453, + 0.08506287634372711, + 0.015036556869745255, + 0.020142197608947754, + 0.027808379381895065 + ], + "credit_map": [ + 0.012860158458352089, + 0.00888401735574007, + 0.013070082291960716, + 0.01238584890961647, + 0.010570530779659748, + 0.05703315883874893, + 0.010778349824249744, + 0.04713926464319229, + 0.006713161710649729, + 0.004065279848873615, + 0.006829314399510622, + 0.0060874721966683865, + 0.003639017231762409, + 0.02820487506687641, + 0.005470531992614269, + 0.03255290165543556, + 0.0073367273434996605, + 0.007250658702105284, + 0.006573044694960117, + 0.00844288058578968, + 0.0061529227532446384, + 0.006180151365697384, + 0.009922628290951252, + 0.005965550895780325, + 0.012263957411050797, + 0.01040102168917656, + 0.01161404699087143, + 0.013074666261672974, + 0.011326206848025322, + 0.022387297824025154, + 0.01613663323223591, + 0.016422457993030548, + 0.02601313218474388, + 0.03928125277161598, + 0.029725242406129837, + 0.02498839609324932, + 0.037499379366636276, + 0.09506969898939133, + 0.021492237225174904, + 0.0690198466181755, + 0.006938534788787365, + 0.007760004606097937, + 0.010549869388341904, + 0.009532609023153782, + 0.0064013246446847916, + 0.0095779774710536, + 0.004672630224376917, + 0.01569303311407566, + 0.005435185972601175, + 0.005146893206983805, + 0.01760430447757244, + 0.007731874939054251, + 0.006889506243169308, + 0.023614859208464622, + 0.010769635438919067, + 0.007870618253946304, + 0.001839615753851831, + 0.0015992613043636084, + 0.001877782167866826, + 0.0017046883003786206, + 0.001494606607593596, + 0.0024667426478117704, + 0.002026747912168503, + 0.0020271106623113155, + 0.00209453864954412, + 0.001971947029232979, + 0.0034237122163176537, + 0.0018387591699138284, + 0.0011825121473520994, + 0.0052758511155843735, + 0.0014889565063640475, + 0.0028659191448241472, + 0.002834145212545991, + 0.0028492736164480448, + 0.003261784091591835, + 0.0035435687750577927, + 0.0027340988162904978, + 0.006368177477270365, + 0.002714211819693446, + 0.0035031153820455074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.03396039456129074 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a memory stick with visible silver circuitry and a gold-plated connector?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure black void with a high-contrast black and white palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a sharp spotlight originating from outside the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the memory stick perform a combined rotation around its long axis and a slow end-to-end tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow zoom-in movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3871079683303833 + ], + "opposite_logprob_tokens": [ + -1.1371079683303833 + ], + "logprob_reward": -0.3871079683303833, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3871079683303833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06318821758031845, + 0.45903295278549194, + 0.02733483351767063, + 0.03299711272120476, + 0.31773313879966736, + 0.03170203045010567, + 0.01651424914598465, + 0.015009825117886066, + 0.01094507984817028, + 0.025542570278048515 + ], + "credit_map": [ + 0.014138603582978249, + 0.004566479474306107, + 0.012101679109036922, + 0.005168522708117962, + 0.005633003544062376, + 0.0085979625582695, + 0.004864657763391733, + 0.008117311634123325, + 0.04795834794640541, + 0.014304939657449722, + 0.033462826162576675, + 0.013362380675971508, + 0.010776451788842678, + 0.16014385223388672, + 0.015187128446996212, + 0.1638370156288147, + 0.0037010053638368845, + 0.00514309573918581, + 0.0029459898360073566, + 0.0034291131887584925, + 0.003114425577223301, + 0.00390056986361742, + 0.0021914816461503506, + 0.002909153001382947, + 0.0049580917693674564, + 0.003275076160207391, + 0.006302055437117815, + 0.0029101090040057898, + 0.004094678908586502, + 0.0038005358073860407, + 0.00401427922770381, + 0.0036422857083380222, + 0.06168556213378906, + 0.013392219319939613, + 0.03845615312457085, + 0.012767506763339043, + 0.012776265852153301, + 0.0891728326678276, + 0.010106384754180908, + 0.0793762356042862, + 0.003121535759419203, + 0.003866496030241251, + 0.005639944225549698, + 0.004776219837367535, + 0.00279653980396688, + 0.0034399242140352726, + 0.002472261432558298, + 0.005589108914136887, + 0.0022100545465946198, + 0.0017914028139784932, + 0.0018301352392882109, + 0.001256462070159614, + 0.0015579711180180311, + 0.0032984139397740364, + 0.00140297575853765, + 0.003166832262650132, + 0.0026719653978943825, + 0.0014191769296303391, + 0.002325214445590973, + 0.0013719662092626095, + 0.00129403464961797, + 0.0030617916490882635, + 0.0011003506369888783, + 0.001765324384905398, + 0.0011859347578138113, + 0.000987762468867004, + 0.00136892672162503, + 0.0011929234024137259, + 0.0006377134704962373, + 0.0029078915249556303, + 0.0008509868639521301, + 0.0018129396485164762, + 0.003747976152226329, + 0.0037948195822536945, + 0.002963425824418664, + 0.0030901136342436075, + 0.002816873136907816, + 0.003964719828218222, + 0.002295016311109066, + 0.002869626507163048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.3871079683303833 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a memory stick with visible silver circuitry and a gold-plated connector?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure black void with a high-contrast black and white palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a sharp spotlight originating from outside the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the memory stick perform a combined rotation around its long axis and a slow end-to-end tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow zoom-in movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054969530552625656 + ], + "opposite_logprob_tokens": [ + -2.929969549179077 + ], + "logprob_reward": -0.054969530552625656, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054969530552625656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1202479749917984, + 0.07935395836830139, + 0.07759296894073486, + 0.17218680679798126, + 0.0430346354842186, + 0.13917359709739685, + 0.05757557600736618, + 0.2595450282096863, + 0.020451584830880165, + 0.030837828293442726 + ], + "credit_map": [ + 0.01182914525270462, + 0.006817393936216831, + 0.010259250178933144, + 0.00889051053673029, + 0.00799973402172327, + 0.03435710072517395, + 0.004911834839731455, + 0.035183023661375046, + 0.01275090966373682, + 0.00830173771828413, + 0.011309354566037655, + 0.007478781975805759, + 0.0065310741774737835, + 0.01354964729398489, + 0.008163463324308395, + 0.011268993839621544, + 0.015267876908183098, + 0.010048546828329563, + 0.01126046758145094, + 0.008005097508430481, + 0.00467686029151082, + 0.013695696368813515, + 0.0048905229195952415, + 0.009747910313308239, + 0.02826603129506111, + 0.014438093639910221, + 0.024743441492319107, + 0.01800808310508728, + 0.019283806905150414, + 0.021851636469364166, + 0.02506154775619507, + 0.020534195005893707, + 0.0036812899634242058, + 0.004744274076074362, + 0.004473615437746048, + 0.005255046300590038, + 0.004275815095752478, + 0.009044009260833263, + 0.004269165452569723, + 0.007291425485163927, + 0.011136243119835854, + 0.009165765717625618, + 0.011740244925022125, + 0.010984561406075954, + 0.007177266292273998, + 0.03943541646003723, + 0.005495534744113684, + 0.044038575142621994, + 0.008055455051362514, + 0.007243437692523003, + 0.008283770643174648, + 0.006491983309388161, + 0.0056904940865933895, + 0.008253737352788448, + 0.007172005251049995, + 0.006384698208421469, + 0.07165762782096863, + 0.018870189785957336, + 0.043843768537044525, + 0.0191955603659153, + 0.020322386175394058, + 0.03744583949446678, + 0.015464418567717075, + 0.032745249569416046, + 0.0018679799977689981, + 0.0027148532681167126, + 0.003159149317070842, + 0.002447674050927162, + 0.002238773973658681, + 0.002948346547782421, + 0.0026350284460932016, + 0.0024397813249379396, + 0.0035948632284998894, + 0.0025043573696166277, + 0.003821346443146467, + 0.0022455027792602777, + 0.0021797732915729284, + 0.008820566348731518, + 0.002454883186146617, + 0.005216538440436125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 5 + ], + "selected_reward": -0.054969530552625656 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a memory stick with visible silver circuitry and a gold-plated connector?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure black void with a high-contrast black and white palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a sharp spotlight originating from outside the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the memory stick perform a combined rotation around its long axis and a slow end-to-end tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow zoom-in movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607927680015564 + ], + "opposite_logprob_tokens": [ + -4.141079425811768 + ], + "logprob_reward": -0.01607927680015564, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607927680015564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19816158711910248, + 0.19437584280967712, + 0.09800684452056885, + 0.11236023157835007, + 0.1494477391242981, + 0.07526581734418869, + 0.06892719864845276, + 0.04912533238530159, + 0.03185421973466873, + 0.02247515693306923 + ], + "credit_map": [ + 0.021056655794382095, + 0.00879763811826706, + 0.015806078910827637, + 0.011284014210104942, + 0.009977329522371292, + 0.0646594986319542, + 0.008710158988833427, + 0.05787023529410362, + 0.01796572469174862, + 0.005650324746966362, + 0.012163622304797173, + 0.00650313263759017, + 0.005251052789390087, + 0.06886954605579376, + 0.008073488250374794, + 0.06989897787570953, + 0.011401643976569176, + 0.01589183509349823, + 0.01050335168838501, + 0.012615841813385487, + 0.010293816216289997, + 0.01834462769329548, + 0.007689295336604118, + 0.011266438290476799, + 0.015482926741242409, + 0.008337338455021381, + 0.017264656722545624, + 0.006916950922459364, + 0.008524230681359768, + 0.03061372973024845, + 0.009902726858854294, + 0.015317682176828384, + 0.022458389401435852, + 0.007822838611900806, + 0.01612558402121067, + 0.007590609136968851, + 0.007684594951570034, + 0.04500037059187889, + 0.007056931033730507, + 0.03570842742919922, + 0.008742716163396835, + 0.007966175675392151, + 0.010753079317510128, + 0.008195361122488976, + 0.007760593667626381, + 0.010416748002171516, + 0.0065034120343625546, + 0.014927743934094906, + 0.007716979831457138, + 0.005507095251232386, + 0.0130820507183671, + 0.0066296411678195, + 0.006214217282831669, + 0.015577303245663643, + 0.007238146848976612, + 0.006961768493056297, + 0.0115999486297369, + 0.004630949813872576, + 0.007559300400316715, + 0.004510785918682814, + 0.004103698302060366, + 0.00785604864358902, + 0.0028202226385474205, + 0.006044382695108652, + 0.004049223847687244, + 0.0035471601877361536, + 0.004467683844268322, + 0.004174784291535616, + 0.002747337566688657, + 0.004750954918563366, + 0.0032799814362078905, + 0.00483709666877985, + 0.003707974450662732, + 0.003020995995029807, + 0.0034031139221042395, + 0.002482525072991848, + 0.002828444354236126, + 0.002764961449429393, + 0.002059274585917592, + 0.0022078678011894226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.01607927680015564 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a memory stick with visible silver circuitry and a gold-plated connector?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure black void with a high-contrast black and white palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a sharp spotlight originating from outside the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the memory stick perform a combined rotation around its long axis and a slow end-to-end tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow zoom-in movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813214361667633 + ], + "opposite_logprob_tokens": [ + -1.406321406364441 + ], + "logprob_reward": -0.2813214361667633, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813214361667633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11542721837759018, + 0.25378742814064026, + 0.0621299110352993, + 0.07306376099586487, + 0.10043557733297348, + 0.26240357756614685, + 0.030265994369983673, + 0.0517822727560997, + 0.023549236357212067, + 0.027155084535479546 + ], + "credit_map": [ + 0.009672409854829311, + 0.00943500455468893, + 0.010660255327820778, + 0.009170169942080975, + 0.010785956867039204, + 0.03078456036746502, + 0.007951573468744755, + 0.02696727029979229, + 0.016542358323931694, + 0.007834983989596367, + 0.01621354930102825, + 0.00911602284759283, + 0.005391092039644718, + 0.10047722607851028, + 0.007519579492509365, + 0.09069260209798813, + 0.009126945398747921, + 0.010166305117309093, + 0.008410283364355564, + 0.008745117112994194, + 0.007449902594089508, + 0.007718780543655157, + 0.005343295633792877, + 0.005169274751096964, + 0.013915922492742538, + 0.007158579770475626, + 0.011240607127547264, + 0.00969189964234829, + 0.007498220074921846, + 0.006790667772293091, + 0.00912687461823225, + 0.007640976458787918, + 0.008507072925567627, + 0.011191985569894314, + 0.015620680525898933, + 0.00802221242338419, + 0.014435277320444584, + 0.025404516607522964, + 0.007411131635308266, + 0.009842691011726856, + 0.020202338695526123, + 0.018306180834770203, + 0.02198575809597969, + 0.016504814848303795, + 0.015231730416417122, + 0.07433714717626572, + 0.010375021025538445, + 0.08546052128076553, + 0.002968659158796072, + 0.003344778437167406, + 0.0036158952862024307, + 0.002330856630578637, + 0.0031380208674818277, + 0.005519601982086897, + 0.003436991246417165, + 0.005911185406148434, + 0.00893185194581747, + 0.005535618867725134, + 0.0066443076357245445, + 0.004652577452361584, + 0.00508735841140151, + 0.010929900221526623, + 0.0033620025496929884, + 0.006638648919761181, + 0.002745266305282712, + 0.0015408785548061132, + 0.0021625543013215065, + 0.0015450420323759317, + 0.0011130443308502436, + 0.0070649064145982265, + 0.0014884218107908964, + 0.005889119114726782, + 0.0036602180916815996, + 0.0028848007787019014, + 0.0032883642707020044, + 0.003244356717914343, + 0.00340672698803246, + 0.0045190234668552876, + 0.002990779932588339, + 0.0031608110293745995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.2813214361667633 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a slowly zooming close-up, a delicate memory stick hovers in a pure, profound black void. The world is dominated by a high-contrast black and white palette, the sole source of light coming from a sharp spotlight emanating from outside the frame. The stick's dark substrate is covered in a labyrinth of intricate silver circuitry, which gleams with a cold metallic sheen in the light. The square integrated circuit chip has a matte finish, while the neatly arranged gold-plated fingers at one end capture and reflect a dazzling, almost pure white light as they rotate. It performs an elegant and deliberate gyroscopic motion, steadily rotating around its long axis while simultaneously tumbling slowly end-to-end. Each rotation causes light and shadow to dance fluidly across its surface, outlining every minute structural detail, creating a surreal, hypnotic, cinematic quality that is a chilling celebration of technological beauty.", + "seed": 27722, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_10/index_2016_seed_27722_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2306427348405123, + "vlm_logprob": -0.2306427348405123, + "vlm_yes_no_margin": 1.7999999880790711, + "vlm_token_credit": -0.2306427348405123, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a memory stick with visible silver circuitry and a gold-plated connector?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure black void with a high-contrast black and white palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a sharp spotlight originating from outside the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the memory stick perform a combined rotation around its long axis and a slow end-to-end tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow zoom-in movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13217712938785553 + ], + "opposite_logprob_tokens": [ + -2.1321771144866943 + ], + "logprob_reward": -0.13217712938785553, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.13217712938785553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16443215310573578, + 0.1853335052728653, + 0.10697333514690399, + 0.056211620569229126, + 0.13419143855571747, + 0.08434485644102097, + 0.04060559347271919, + 0.10542032867670059, + 0.08142520487308502, + 0.041061993688344955 + ], + "credit_map": [ + 0.01523579191416502, + 0.016326887533068657, + 0.05026480555534363, + 0.01167350821197033, + 0.01336340606212616, + 0.01877683773636818, + 0.012600899673998356, + 0.026190010830760002, + 0.03362658619880676, + 0.01923239603638649, + 0.018332036212086678, + 0.012523220852017403, + 0.026229452341794968, + 0.026077888906002045, + 0.022247212007641792, + 0.027064714580774307, + 0.013407181017100811, + 0.00907935295253992, + 0.016800442710518837, + 0.009456281550228596, + 0.017657125368714333, + 0.012130029499530792, + 0.011408300139009953, + 0.017034631222486496, + 0.005159652326256037, + 0.006436855066567659, + 0.007812161464244127, + 0.007410402875393629, + 0.008717512711882591, + 0.009212871082127094, + 0.003248566295951605, + 0.00821360107511282, + 0.015574093908071518, + 0.0167064368724823, + 0.017691342160105705, + 0.015170111320912838, + 0.024511020630598068, + 0.013751423917710781, + 0.01594652608036995, + 0.014840491116046906, + 0.008454268798232079, + 0.011172257363796234, + 0.013278872705996037, + 0.01289819460362196, + 0.006712097208946943, + 0.008469440042972565, + 0.01494544930756092, + 0.008414272218942642, + 0.005946464370936155, + 0.0059425486251711845, + 0.006490137428045273, + 0.00348763819783926, + 0.004642242565751076, + 0.004256275948137045, + 0.005050569772720337, + 0.0047897156327962875, + 0.01600629836320877, + 0.010457863099873066, + 0.011241395957767963, + 0.014034741558134556, + 0.012301979586482048, + 0.011111319996416569, + 0.018351774662733078, + 0.011914952658116817, + 0.013357963413000107, + 0.007386958692222834, + 0.010089714080095291, + 0.012430972419679165, + 0.008790924213826656, + 0.01210976392030716, + 0.0074668629094958305, + 0.009792046621441841, + 0.006168224848806858, + 0.003974346909672022, + 0.006172122899442911, + 0.0041429526172578335, + 0.005767022725194693, + 0.0058203572407364845, + 0.003303036792203784, + 0.005713930353522301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.13217712938785553 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a memory stick with visible silver circuitry and a gold-plated connector?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure black void with a high-contrast black and white palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a sharp spotlight originating from outside the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the memory stick perform a combined rotation around its long axis and a slow end-to-end tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow zoom-in movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17989861965179443 + ], + "opposite_logprob_tokens": [ + -1.8048986196517944 + ], + "logprob_reward": -0.17989861965179443, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17989861965179443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08796632289886475, + 0.2065889686346054, + 0.05294150859117508, + 0.2975251078605652, + 0.12268203496932983, + 0.04977802932262421, + 0.02846081741154194, + 0.05606643855571747, + 0.03420015051960945, + 0.0637906864285469 + ], + "credit_map": [ + 0.012084314599633217, + 0.006397818680852652, + 0.02502722106873989, + 0.008441891521215439, + 0.00546904606744647, + 0.011417088098824024, + 0.006332221440970898, + 0.012796725146472454, + 0.030151523649692535, + 0.027478652074933052, + 0.02635914646089077, + 0.02168017253279686, + 0.02099679596722126, + 0.031149331480264664, + 0.020651424303650856, + 0.02812192402780056, + 0.004881408531218767, + 0.003940050024539232, + 0.0054649850353598595, + 0.005601900164037943, + 0.0055589573457837105, + 0.004088190384209156, + 0.010587756521999836, + 0.012818262912333012, + 0.024312447756528854, + 0.02591904252767563, + 0.04095497354865074, + 0.04643382877111435, + 0.03649072349071503, + 0.03943954408168793, + 0.02104809880256653, + 0.06292644143104553, + 0.014733715914189816, + 0.017231615260243416, + 0.018742525950074196, + 0.014831893146038055, + 0.016236701980233192, + 0.012548115104436874, + 0.015297999605536461, + 0.013059461489319801, + 0.004473240580409765, + 0.00530471745878458, + 0.007970712147653103, + 0.007306003011763096, + 0.0035116386134177446, + 0.004536387510597706, + 0.011590485461056232, + 0.0050848438404500484, + 0.004215771798044443, + 0.004074889235198498, + 0.003747300710529089, + 0.0025596481282263994, + 0.0021380994003266096, + 0.003130187513306737, + 0.0038730979431420565, + 0.004721820820122957, + 0.007112658582627773, + 0.005958156194537878, + 0.006186370737850666, + 0.008869397453963757, + 0.00689690001308918, + 0.0048178439028561115, + 0.008373339660465717, + 0.007851770147681236, + 0.0041277045384049416, + 0.003451010910794139, + 0.0043316916562616825, + 0.00547934090718627, + 0.0043838853016495705, + 0.004601608961820602, + 0.0034224961418658495, + 0.0044024065136909485, + 0.01407997403293848, + 0.004532512277364731, + 0.0076725780963897705, + 0.007428720593452454, + 0.007062466349452734, + 0.006411247421056032, + 0.006153543945401907, + 0.010449642315506935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 4 + ], + "selected_reward": -0.17989861965179443 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a memory stick with visible silver circuitry and a gold-plated connector?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure black void with a high-contrast black and white palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a sharp spotlight originating from outside the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the memory stick perform a combined rotation around its long axis and a slow end-to-end tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow zoom-in movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742998480796814 + ], + "opposite_logprob_tokens": [ + -0.9742998480796814 + ], + "logprob_reward": -0.4742998480796814, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742998480796814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3652520179748535, + 0.11864683777093887, + 0.0410868339240551, + 0.04863471910357475, + 0.0975552350282669, + 0.02099989540874958, + 0.032031673938035965, + 0.12269937992095947, + 0.0972023755311966, + 0.055891018360853195 + ], + "credit_map": [ + 0.03482205793261528, + 0.019347168505191803, + 0.1308189481496811, + 0.021827178075909615, + 0.014827306382358074, + 0.03531460464000702, + 0.018549954518675804, + 0.08974483609199524, + 0.01743384450674057, + 0.012470223009586334, + 0.013887947425246239, + 0.011132831685245037, + 0.012841012328863144, + 0.0170818530023098, + 0.019823579117655754, + 0.013975542970001698, + 0.007073316257447004, + 0.005058614537119865, + 0.0051983012817800045, + 0.004746470134705305, + 0.0041489023715257645, + 0.0035080378875136375, + 0.005761430133134127, + 0.005591756664216518, + 0.0043779406696558, + 0.00511033833026886, + 0.006276892498135567, + 0.004146948456764221, + 0.007384186144918203, + 0.008178072981536388, + 0.004459348972886801, + 0.008700991980731487, + 0.005970493424683809, + 0.009255108423531055, + 0.013356816954910755, + 0.010373837314546108, + 0.02330087125301361, + 0.011569039896130562, + 0.011263170279562473, + 0.01246589981019497, + 0.0021894637029618025, + 0.0032804086804389954, + 0.0032914457842707634, + 0.002908067312091589, + 0.0023029933217912912, + 0.002556778024882078, + 0.002222099807113409, + 0.0022486383095383644, + 0.006229687016457319, + 0.0037146045360714197, + 0.004179663024842739, + 0.003313440131023526, + 0.0030489666387438774, + 0.0034449705854058266, + 0.0041762785986065865, + 0.003924062941223383, + 0.014920755289494991, + 0.013337703421711922, + 0.011667205952107906, + 0.01778711937367916, + 0.016304491087794304, + 0.01034854631870985, + 0.02154361642897129, + 0.01678994484245777, + 0.012277787551283836, + 0.009917914867401123, + 0.014010967686772346, + 0.022897249087691307, + 0.008712705224752426, + 0.007926098071038723, + 0.009523317217826843, + 0.011936335824429989, + 0.00940193235874176, + 0.005106587428599596, + 0.006507455371320248, + 0.006301886402070522, + 0.006720174103975296, + 0.007210026029497385, + 0.006419667974114418, + 0.008223292417824268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.4742998480796814 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a memory stick with visible silver circuitry and a gold-plated connector?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure black void with a high-contrast black and white palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a sharp spotlight originating from outside the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the memory stick perform a combined rotation around its long axis and a slow end-to-end tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow zoom-in movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018199244514107704 + ], + "opposite_logprob_tokens": [ + -4.018199443817139 + ], + "logprob_reward": -0.018199244514107704, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018199244514107704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4103628396987915, + 0.07242632657289505, + 0.05423184856772423, + 0.038488589227199554, + 0.11370544880628586, + 0.03467381373047829, + 0.025210892781615257, + 0.09844803065061569, + 0.09260313212871552, + 0.059849031269550323 + ], + "credit_map": [ + 0.04165387526154518, + 0.018601369112730026, + 0.1531575620174408, + 0.021309005096554756, + 0.016748584806919098, + 0.04374486580491066, + 0.016624482348561287, + 0.0985230952501297, + 0.013122611679136753, + 0.007041467819362879, + 0.006782811600714922, + 0.006096558179706335, + 0.009790661744773388, + 0.011122155003249645, + 0.00997124332934618, + 0.008498813025653362, + 0.008241277188062668, + 0.005100891459733248, + 0.007637877948582172, + 0.006848267279565334, + 0.006234355736523867, + 0.0040492401458323, + 0.00826957542449236, + 0.007850365713238716, + 0.004428567830473185, + 0.004147693049162626, + 0.005187124013900757, + 0.004324032459408045, + 0.008186387829482555, + 0.004433960188180208, + 0.003138994798064232, + 0.0046418276615440845, + 0.012360243126749992, + 0.014025584794580936, + 0.01548764668405056, + 0.014618266373872757, + 0.02359131909906864, + 0.009207354858517647, + 0.013554629869759083, + 0.010860408656299114, + 0.004043686669319868, + 0.004123907070606947, + 0.005630777683109045, + 0.005606596823781729, + 0.002403207588940859, + 0.004778943490236998, + 0.004629049450159073, + 0.003457646584138274, + 0.004569015000015497, + 0.0027046254836022854, + 0.0032968707382678986, + 0.0025677785743027925, + 0.0029122126288712025, + 0.002393450355157256, + 0.002817039377987385, + 0.00394990062341094, + 0.015229320153594017, + 0.009008119814097881, + 0.015422115102410316, + 0.014652006328105927, + 0.010808505117893219, + 0.009237274527549744, + 0.012385373003780842, + 0.011705311015248299, + 0.00936204381287098, + 0.00855949241667986, + 0.012747246772050858, + 0.02243422530591488, + 0.008278792724013329, + 0.009678494185209274, + 0.010786261409521103, + 0.010756569914519787, + 0.00988845806568861, + 0.00501437671482563, + 0.00859688501805067, + 0.007478987332433462, + 0.006498031318187714, + 0.007831929251551628, + 0.005699337925761938, + 0.008841026574373245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.018199244514107704 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a memory stick with visible silver circuitry and a gold-plated connector?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure black void with a high-contrast black and white palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a sharp spotlight originating from outside the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the memory stick perform a combined rotation around its long axis and a slow end-to-end tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow zoom-in movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486388325691223 + ], + "opposite_logprob_tokens": [ + -1.2236387729644775 + ], + "logprob_reward": -0.3486388325691223, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.3486388325691223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4825575649738312, + 0.048725999891757965, + 0.02286604419350624, + 0.016084754839539528, + 0.05686763674020767, + 0.020537471398711205, + 0.019565168768167496, + 0.11065371334552765, + 0.1940697729587555, + 0.028071889653801918 + ], + "credit_map": [ + 0.048064202070236206, + 0.021204480901360512, + 0.1836281716823578, + 0.02591472677886486, + 0.01707826741039753, + 0.051042746752500534, + 0.017470022663474083, + 0.11815494298934937, + 0.00606894213706255, + 0.006454769056290388, + 0.004556517116725445, + 0.004082204774022102, + 0.006821607705205679, + 0.006240827031433582, + 0.00785513874143362, + 0.006645992863923311, + 0.004050916526466608, + 0.002965663094073534, + 0.0030894256196916103, + 0.0032402321230620146, + 0.00208696024492383, + 0.0020459636580199003, + 0.0025899114552885294, + 0.0027969712391495705, + 0.0021271922159940004, + 0.002343276282772422, + 0.0020184568129479885, + 0.002116367220878601, + 0.0020726672373712063, + 0.0017818821361288428, + 0.0010036180028691888, + 0.002621295163407922, + 0.004034329205751419, + 0.005737075582146645, + 0.007751952391117811, + 0.00666973227635026, + 0.012175165116786957, + 0.00643960339948535, + 0.0076654329895973206, + 0.0063943457789719105, + 0.002343173837289214, + 0.00236925738863647, + 0.004054782446473837, + 0.002724507125094533, + 0.001610101549886167, + 0.00330735812895, + 0.0018645080272108316, + 0.0022637834772467613, + 0.0023138904944062233, + 0.0021906935144215822, + 0.002477820962667465, + 0.0014505828730762005, + 0.0021169104147702456, + 0.0018293389584869146, + 0.0040907966904342175, + 0.0030951332300901413, + 0.014117427170276642, + 0.011375656351447105, + 0.012296782806515694, + 0.016274483874440193, + 0.014256049878895283, + 0.009441045112907887, + 0.01719760149717331, + 0.015694672241806984, + 0.02165720798075199, + 0.01925038732588291, + 0.02863849140703678, + 0.04975969344377518, + 0.01846967823803425, + 0.01627488248050213, + 0.01817885972559452, + 0.02184057980775833, + 0.004534963984042406, + 0.0026824523229151964, + 0.003634108928963542, + 0.0032080740202218294, + 0.0027518155984580517, + 0.003234941978007555, + 0.0034400864969938993, + 0.004585447255522013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.3486388325691223 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a slowly zooming close-up, a delicate memory stick hovers in a pure, profound black void. The world is dominated by a high-contrast black and white palette, the sole source of light coming from a sharp spotlight emanating from outside the frame. The stick's dark substrate is covered in a labyrinth of intricate silver circuitry, which gleams with a cold metallic sheen in the light. The square integrated circuit chip has a matte finish, while the neatly arranged gold-plated fingers at one end capture and reflect a dazzling, almost pure white light as they rotate. It performs an elegant and deliberate gyroscopic motion, steadily rotating around its long axis while simultaneously tumbling slowly end-to-end. Each rotation causes light and shadow to dance fluidly across its surface, outlining every minute structural detail, creating a surreal, hypnotic, cinematic quality that is a chilling celebration of technological beauty.", + "seed": 27723, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_10/index_2016_seed_27723_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.16701310221105814, + "vlm_logprob": -0.16701310221105814, + "vlm_yes_no_margin": 2.675, + "vlm_token_credit": -0.16701310221105814, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a memory stick with visible silver circuitry and a gold-plated connector?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure black void with a high-contrast black and white palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a sharp spotlight originating from outside the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the memory stick perform a combined rotation around its long axis and a slow end-to-end tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow zoom-in movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011753357946872711 + ], + "opposite_logprob_tokens": [ + -4.886753559112549 + ], + "logprob_reward": -0.011753357946872711, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.011753357946872711, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18456193804740906, + 0.05293997377157211, + 0.03774419054389, + 0.17934173345565796, + 0.05486685410141945, + 0.11182954907417297, + 0.13100941479206085, + 0.04533550888299942, + 0.11964526027441025, + 0.08272554725408554 + ], + "credit_map": [ + 0.015556525439023972, + 0.016859082505106926, + 0.01885756105184555, + 0.02239340916275978, + 0.01924172416329384, + 0.027348140254616737, + 0.03374475613236427, + 0.03056073933839798, + 0.004351613111793995, + 0.006422816310077906, + 0.00758880702778697, + 0.0077669983729720116, + 0.005786364898085594, + 0.006990635301917791, + 0.0058685848489403725, + 0.00816415436565876, + 0.005558622069656849, + 0.003683476010337472, + 0.005930026527494192, + 0.002421679673716426, + 0.003517060074955225, + 0.005228477995842695, + 0.004140157252550125, + 0.007264688611030579, + 0.01749027706682682, + 0.02089526131749153, + 0.019976209849119186, + 0.020493807271122932, + 0.015882765874266624, + 0.02884943038225174, + 0.024025950580835342, + 0.03172804042696953, + 0.01023966446518898, + 0.004747361876070499, + 0.006701835431158543, + 0.0058693005703389645, + 0.007412001024931669, + 0.007163585629314184, + 0.006671675480902195, + 0.0060614305548369884, + 0.013322942890226841, + 0.011503864079713821, + 0.01086379587650299, + 0.013156890869140625, + 0.018288055434823036, + 0.014326373115181923, + 0.013423635624349117, + 0.01694399118423462, + 0.018372368067502975, + 0.012744072824716568, + 0.02085915207862854, + 0.016043923795223236, + 0.01150532253086567, + 0.017420750111341476, + 0.013459606096148491, + 0.020604217424988747, + 0.004869014490395784, + 0.00436873733997345, + 0.006915734149515629, + 0.0038993186317384243, + 0.00397479347884655, + 0.009112430736422539, + 0.003503536805510521, + 0.008691942319273949, + 0.012615491636097431, + 0.008653976023197174, + 0.016776371747255325, + 0.011303618550300598, + 0.00800275057554245, + 0.025447439402341843, + 0.01197703555226326, + 0.024868575856089592, + 0.005920581519603729, + 0.00989541970193386, + 0.008676943369209766, + 0.008278055116534233, + 0.016859842464327812, + 0.011671312153339386, + 0.010456516407430172, + 0.01096687838435173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.011753357946872711 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a memory stick with visible silver circuitry and a gold-plated connector?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure black void with a high-contrast black and white palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a sharp spotlight originating from outside the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the memory stick perform a combined rotation around its long axis and a slow end-to-end tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow zoom-in movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12707041203975677 + ], + "opposite_logprob_tokens": [ + -2.127070426940918 + ], + "logprob_reward": -0.12707041203975677, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12707041203975677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2658020853996277, + 0.04109824821352959, + 0.09625578671693802, + 0.1269877552986145, + 0.05812055245041847, + 0.19595932960510254, + 0.043162744492292404, + 0.05935428664088249, + 0.09145084768533707, + 0.021808331832289696 + ], + "credit_map": [ + 0.018740516155958176, + 0.0224055927246809, + 0.040715884417295456, + 0.030412469059228897, + 0.026568833738565445, + 0.03937854990363121, + 0.04062718525528908, + 0.04695306345820427, + 0.0043722596019506454, + 0.005067331716418266, + 0.005787387955933809, + 0.005009730812162161, + 0.004644110333174467, + 0.005416056606918573, + 0.004845416639000177, + 0.0059559522196650505, + 0.017859818413853645, + 0.007103501819074154, + 0.014785303734242916, + 0.005777751561254263, + 0.006777895148843527, + 0.015864145010709763, + 0.007487846538424492, + 0.020599521696567535, + 0.022336943075060844, + 0.010866222903132439, + 0.011653557419776917, + 0.010859214700758457, + 0.012189487926661968, + 0.02011510357260704, + 0.015183765441179276, + 0.023783452808856964, + 0.007406248711049557, + 0.005628161132335663, + 0.00998456310480833, + 0.006789275910705328, + 0.00765881547704339, + 0.007140430621802807, + 0.006507888436317444, + 0.007005168125033379, + 0.027131645008921623, + 0.01734485849738121, + 0.02620399184525013, + 0.01955084688961506, + 0.021403225138783455, + 0.03443336859345436, + 0.020071858540177345, + 0.0298195481300354, + 0.007393510080873966, + 0.004425602499395609, + 0.004591365810483694, + 0.00367420120164752, + 0.0037900437600910664, + 0.006906257476657629, + 0.004626566544175148, + 0.007755199447274208, + 0.009065416641533375, + 0.004424576181918383, + 0.00973901990801096, + 0.003161995206028223, + 0.003439664375036955, + 0.013676286675035954, + 0.0033285953104496002, + 0.012518730945885181, + 0.012048598378896713, + 0.005873082671314478, + 0.015558143146336079, + 0.007817773148417473, + 0.004490989726036787, + 0.01491617877036333, + 0.007651974447071552, + 0.023094112053513527, + 0.0026562742423266172, + 0.002265308052301407, + 0.0027908403426408768, + 0.002107902429997921, + 0.0037718513049185276, + 0.003393742488697171, + 0.0020187736954540014, + 0.0028036385774612427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.12707041203975677 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a memory stick with visible silver circuitry and a gold-plated connector?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure black void with a high-contrast black and white palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a sharp spotlight originating from outside the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the memory stick perform a combined rotation around its long axis and a slow end-to-end tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow zoom-in movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16037887334823608 + ], + "opposite_logprob_tokens": [ + -1.9103789329528809 + ], + "logprob_reward": -0.16037887334823608, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16037887334823608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14255158603191376, + 0.12192314863204956, + 0.16783060133457184, + 0.06651578098535538, + 0.05983695015311241, + 0.04473509266972542, + 0.08491040766239166, + 0.1339966207742691, + 0.15155397355556488, + 0.026145899668335915 + ], + "credit_map": [ + 0.013209691271185875, + 0.014064128510653973, + 0.017958875745534897, + 0.013524106703698635, + 0.014472569338977337, + 0.022332247346639633, + 0.024915633723139763, + 0.02207433059811592, + 0.014106218703091145, + 0.010632296092808247, + 0.011502185836434364, + 0.015298915095627308, + 0.01968672126531601, + 0.018586445599794388, + 0.015509990975260735, + 0.01660037226974964, + 0.021667424589395523, + 0.013043997809290886, + 0.021398693323135376, + 0.017800547182559967, + 0.01653839647769928, + 0.026717564091086388, + 0.018085358664393425, + 0.03257859870791435, + 0.006473049987107515, + 0.00827847234904766, + 0.008821132592856884, + 0.007928485982120037, + 0.008847925812005997, + 0.009447980672121048, + 0.00930994376540184, + 0.007408785168081522, + 0.011620075441896915, + 0.004135127179324627, + 0.00781021686270833, + 0.0036900171544402838, + 0.005556347779929638, + 0.009322304278612137, + 0.006539075635373592, + 0.011163778603076935, + 0.005177944898605347, + 0.005029661115258932, + 0.007582563906908035, + 0.003612768603488803, + 0.005595822352916002, + 0.007201847154647112, + 0.0045052170753479, + 0.006029264070093632, + 0.01348321232944727, + 0.009867759421467781, + 0.01701219752430916, + 0.008869851008057594, + 0.006285924464464188, + 0.01123077142983675, + 0.0077990577556192875, + 0.010361621156334877, + 0.02243465930223465, + 0.009689431637525558, + 0.02104787342250347, + 0.00872788205742836, + 0.0073527684435248375, + 0.02876877412199974, + 0.005404787138104439, + 0.03057043068110943, + 0.017006659880280495, + 0.010629035532474518, + 0.024371784180402756, + 0.0157863087952137, + 0.007650885730981827, + 0.025936506688594818, + 0.0169331356883049, + 0.033239640295505524, + 0.004066347610205412, + 0.0030582009349018335, + 0.0032708796206861734, + 0.0027970841620117426, + 0.002808590419590473, + 0.0044861407950520515, + 0.002736677648499608, + 0.0029219756834208965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.16037887334823608 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a memory stick with visible silver circuitry and a gold-plated connector?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure black void with a high-contrast black and white palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a sharp spotlight originating from outside the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the memory stick perform a combined rotation around its long axis and a slow end-to-end tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow zoom-in movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541239149868488 + ], + "opposite_logprob_tokens": [ + -4.3875412940979 + ], + "logprob_reward": -0.012541239149868488, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541239149868488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13633601367473602, + 0.06434889137744904, + 0.0783257931470871, + 0.16154621541500092, + 0.1079963743686676, + 0.10816238075494766, + 0.11854042857885361, + 0.07417117059230804, + 0.11118843406438828, + 0.03938434645533562 + ], + "credit_map": [ + 0.008585549890995026, + 0.013517133891582489, + 0.015759479254484177, + 0.017879093065857887, + 0.016052432358264923, + 0.020620081573724747, + 0.024144353345036507, + 0.01977788656949997, + 0.009327874518930912, + 0.005415237043052912, + 0.006705134641379118, + 0.006591108627617359, + 0.007572481408715248, + 0.011327634565532207, + 0.007081322837620974, + 0.010328104719519615, + 0.011060948483645916, + 0.006176456809043884, + 0.01271230448037386, + 0.0077751330099999905, + 0.007700292393565178, + 0.01129542849957943, + 0.007480233907699585, + 0.014124996028840542, + 0.02318519726395607, + 0.01587364636361599, + 0.016720877960324287, + 0.020623140037059784, + 0.018642956390976906, + 0.02312617190182209, + 0.019650321453809738, + 0.02372390776872635, + 0.02370453253388405, + 0.009115733206272125, + 0.015859823673963547, + 0.008960110135376453, + 0.010190549306571484, + 0.017127221450209618, + 0.008191794157028198, + 0.014846610836684704, + 0.013578071258962154, + 0.008481292985379696, + 0.014594635926187038, + 0.009185196831822395, + 0.01462467759847641, + 0.019457481801509857, + 0.012734423391520977, + 0.015506597235798836, + 0.018768934532999992, + 0.013752871192991734, + 0.02442825585603714, + 0.012520979158580303, + 0.010962176136672497, + 0.014457577839493752, + 0.013547638431191444, + 0.01010199636220932, + 0.011856885626912117, + 0.004719710908830166, + 0.012137122452259064, + 0.005000400822609663, + 0.0043084993958473206, + 0.016143780201673508, + 0.004614369478076696, + 0.015390396118164062, + 0.014636882580816746, + 0.003987621050328016, + 0.02797693759202957, + 0.005430852063000202, + 0.0052757877856493, + 0.02016105316579342, + 0.007174145430326462, + 0.02654515765607357, + 0.0052529629319906235, + 0.004512826446443796, + 0.004894394893199205, + 0.0031575439497828484, + 0.005069616716355085, + 0.00768020935356617, + 0.0037946230731904507, + 0.0050221700221300125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.012541239149868488 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a memory stick with visible silver circuitry and a gold-plated connector?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure black void with a high-contrast black and white palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a sharp spotlight originating from outside the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the memory stick perform a combined rotation around its long axis and a slow end-to-end tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow zoom-in movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233216285705566 + ], + "opposite_logprob_tokens": [ + -0.8983216285705566 + ], + "logprob_reward": -0.5233216285705566, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233216285705566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08561861515045166, + 0.05868827924132347, + 0.0633644163608551, + 0.03011377900838852, + 0.08297105878591537, + 0.04591994360089302, + 0.11157021671533585, + 0.30000507831573486, + 0.19048160314559937, + 0.03126703202724457 + ], + "credit_map": [ + 0.006874779239296913, + 0.008331820368766785, + 0.014772824011743069, + 0.010404287837445736, + 0.009014680981636047, + 0.010762696154415607, + 0.012363814748823643, + 0.013093709945678711, + 0.0054996623657643795, + 0.005670263897627592, + 0.005482563283294439, + 0.007013152819126844, + 0.007056110072880983, + 0.01199383195489645, + 0.005531699396669865, + 0.010440990328788757, + 0.007220170460641384, + 0.008664893917739391, + 0.00926969014108181, + 0.006449026521295309, + 0.006516312714666128, + 0.008192412555217743, + 0.00607694685459137, + 0.010974962264299393, + 0.003216363722458482, + 0.003494283417239785, + 0.002553352853283286, + 0.004572310484945774, + 0.004241867922246456, + 0.004718015436083078, + 0.003353983396664262, + 0.003963604103773832, + 0.008218483068048954, + 0.007592986803501844, + 0.012667731381952763, + 0.0121977673843503, + 0.009499464184045792, + 0.01098208874464035, + 0.008662923239171505, + 0.013149612583220005, + 0.00457656430080533, + 0.005895943380892277, + 0.00619219895452261, + 0.005803178530186415, + 0.004839506931602955, + 0.0067314812913537025, + 0.005094761960208416, + 0.006786306854337454, + 0.015862474218010902, + 0.0108236288651824, + 0.01552548911422491, + 0.01675603911280632, + 0.009485816583037376, + 0.013608923181891441, + 0.01010483130812645, + 0.019403008744120598, + 0.04805303364992142, + 0.020707523450255394, + 0.048712361603975296, + 0.01554575189948082, + 0.017474886029958725, + 0.06739586591720581, + 0.01266458723694086, + 0.06945105642080307, + 0.023612625896930695, + 0.012121039442718029, + 0.025484232231974602, + 0.018452521413564682, + 0.010513345710933208, + 0.03476769104599953, + 0.018231218680739403, + 0.047298941761255264, + 0.0031454742420464754, + 0.0033061921130865812, + 0.003710787510499358, + 0.0034273453056812286, + 0.0048607890494167805, + 0.0038561252877116203, + 0.004725980572402477, + 0.0042343405075371265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.5233216285705566 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_11/index_tensor([907, 907])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_11/index_tensor([907, 907])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2e397918c864f0cb0118e619e4a9bce472cdd620 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_11/index_tensor([907, 907])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Handheld, panning, wide-angle shots immerse the viewer directly in the raw energy of cattle ranching. The unsteady movement conveys a sense of immediate, improvisational action. Across a vast expanse of rolling, dry, golden-brown grassland, a handful of people on foot purposefully herd a large herd of reddish-brown cattle. The ranchers, a group of men and women, are dressed in practical workwear: worn denim jeans, plaid flannel shirts, and sturdy leather boots. Some wear wide-brimmed hats, casting their weathered faces in shadow. They move with a brisk yet determined gait, flanking the herd and occasionally waving their arms to guide stray animals back to the main group. The cattle move at a slow, powerful pace, their coats coated in a thin layer of dust, their hooves churning the dry earth. The entire scene is bathed in the warm, soft light of the afternoon sun, casting long, dramatic shadows and illuminating the thick clouds of dust kicked up by their activity, creating a hazy, golden atmosphere that is both chaotic and serenely beautiful.", + "seed": 22669, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_11/index_907_seed_22669_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1667378604412079, + "vlm_logprob": -0.1667378604412079, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.1667378604412079, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement handheld and panning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast expanse of rolling, dry, golden-brown grassland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are people dressed in worn denim jeans, plaid flannel shirts, and sturdy leather boots herding reddish-brown cattle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the cattle moving at a slow, powerful pace with coats coated in a thin layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft afternoon sunlight casting long shadows and a hazy, golden atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7767996191978455 + ], + "opposite_logprob_tokens": [ + -0.6517996191978455 + ], + "logprob_reward": -0.7767996191978455, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7767996191978455, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2063245177268982, + 0.0580049529671669, + 0.10096511244773865, + 0.06290869414806366, + 0.13038675487041473, + 0.07936428487300873, + 0.08790489286184311, + 0.10583244264125824, + 0.11768488585948944, + 0.050623439252376556 + ], + "credit_map": [ + 0.024043699726462364, + 0.018029551953077316, + 0.024303562939167023, + 0.02068191207945347, + 0.026662243530154228, + 0.033134981989860535, + 0.03590669110417366, + 0.023561863228678703, + 0.006746761035174131, + 0.007236574310809374, + 0.00511582987383008, + 0.008969857357442379, + 0.007848563604056835, + 0.006338095758110285, + 0.007732190191745758, + 0.008017078973352909, + 0.016718553379178047, + 0.01278343889862299, + 0.012603304348886013, + 0.01046496070921421, + 0.01743924245238304, + 0.009043139405548573, + 0.013455579057335854, + 0.008456896059215069, + 0.007040924392640591, + 0.009203489869832993, + 0.009162641130387783, + 0.011836179532110691, + 0.00550277903676033, + 0.006609583273530006, + 0.00830516591668129, + 0.0052479337900877, + 0.012029618956148624, + 0.01951756700873375, + 0.010143530555069447, + 0.013100314885377884, + 0.00896430853754282, + 0.03409431874752045, + 0.02055218815803528, + 0.011984918266534805, + 0.012810747139155865, + 0.008823854848742485, + 0.00489506870508194, + 0.006897640414535999, + 0.014648822136223316, + 0.01614568382501602, + 0.008483801037073135, + 0.006658663973212242, + 0.010801543481647968, + 0.01775669865310192, + 0.00881914608180523, + 0.012436896562576294, + 0.01039064023643732, + 0.008674862794578075, + 0.009007124230265617, + 0.010017973370850086, + 0.025389542803168297, + 0.006531612481921911, + 0.0060980310663580894, + 0.007309543434530497, + 0.01003239769488573, + 0.034673046320676804, + 0.007000428158789873, + 0.008797844871878624, + 0.015893882140517235, + 0.013584106229245663, + 0.010633991099894047, + 0.012356959283351898, + 0.02217574417591095, + 0.013722949661314487, + 0.014417294412851334, + 0.014899960719048977, + 0.007909880951046944, + 0.00871980655938387, + 0.007635566405951977, + 0.005580675322562456, + 0.004887124057859182, + 0.006987273693084717, + 0.0046799988485872746, + 0.004223113413900137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.7767996191978455 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement handheld and panning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast expanse of rolling, dry, golden-brown grassland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are people dressed in worn denim jeans, plaid flannel shirts, and sturdy leather boots herding reddish-brown cattle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the cattle moving at a slow, powerful pace with coats coated in a thin layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft afternoon sunlight casting long shadows and a hazy, golden atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004111528396606445 + ], + "opposite_logprob_tokens": [ + -5.504111289978027 + ], + "logprob_reward": -0.004111528396606445, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004111528396606445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09194085001945496, + 0.0948721319437027, + 0.04180785268545151, + 0.37861815094947815, + 0.03444829210639, + 0.025714924558997154, + 0.15313740074634552, + 0.06959685683250427, + 0.07287826389074326, + 0.03698529303073883 + ], + "credit_map": [ + 0.01372700184583664, + 0.008813598193228245, + 0.0087488554418087, + 0.008396225981414318, + 0.012416639365255833, + 0.013660861179232597, + 0.016942724585533142, + 0.009234936907887459, + 0.008470828644931316, + 0.010895862244069576, + 0.011715072207152843, + 0.011296436190605164, + 0.011977264657616615, + 0.00941446702927351, + 0.018146555870771408, + 0.012955635786056519, + 0.005786992143839598, + 0.007536035962402821, + 0.004928742069751024, + 0.004188913851976395, + 0.006564710289239883, + 0.004410748369991779, + 0.004850498866289854, + 0.0035412083379924297, + 0.05047270655632019, + 0.04930707439780235, + 0.06369318068027496, + 0.1077272966504097, + 0.022997954860329628, + 0.019446227699518204, + 0.04444669187068939, + 0.020527008920907974, + 0.0038135081995278597, + 0.00306348642334342, + 0.003173292148858309, + 0.003991343546658754, + 0.003716310951858759, + 0.005326622631400824, + 0.0058601368218660355, + 0.005503587890416384, + 0.003533158218488097, + 0.0038318298757076263, + 0.0022078147158026695, + 0.003042478347197175, + 0.00398986367508769, + 0.005507164169102907, + 0.001826677704229951, + 0.0017759372713044286, + 0.021708259359002113, + 0.03521088883280754, + 0.017308231443166733, + 0.018968850374221802, + 0.019238892942667007, + 0.01821126975119114, + 0.013371399603784084, + 0.009119604714214802, + 0.018352719023823738, + 0.0036804177798330784, + 0.0030294801108539104, + 0.0036584963090717793, + 0.006244015879929066, + 0.026440322399139404, + 0.003923957701772451, + 0.004267444834113121, + 0.00954762939363718, + 0.008041152730584145, + 0.009525156579911709, + 0.006993381306529045, + 0.015051435679197311, + 0.008432860486209393, + 0.006366354878991842, + 0.008920283988118172, + 0.004852572456002235, + 0.0045150043442845345, + 0.004647377412766218, + 0.003306905273348093, + 0.005428692325949669, + 0.007766455411911011, + 0.0035224941093474627, + 0.0029457861091941595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 1 + ], + "selected_reward": -0.004111528396606445 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement handheld and panning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast expanse of rolling, dry, golden-brown grassland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are people dressed in worn denim jeans, plaid flannel shirts, and sturdy leather boots herding reddish-brown cattle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the cattle moving at a slow, powerful pace with coats coated in a thin layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft afternoon sunlight casting long shadows and a hazy, golden atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107686385512352 + ], + "opposite_logprob_tokens": [ + -4.511076927185059 + ], + "logprob_reward": -0.01107686385512352, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107686385512352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22267036139965057, + 0.13679726421833038, + 0.1497775763273239, + 0.12329233437776566, + 0.09804283827543259, + 0.05005871132016182, + 0.06114216148853302, + 0.05648335814476013, + 0.07594559341669083, + 0.025789804756641388 + ], + "credit_map": [ + 0.01903199777007103, + 0.0242527574300766, + 0.02381279692053795, + 0.02486255019903183, + 0.024900425225496292, + 0.04413392022252083, + 0.028739426285028458, + 0.032936472445726395, + 0.013329041190445423, + 0.016517002135515213, + 0.020091744139790535, + 0.014345732517540455, + 0.01832391880452633, + 0.016926554962992668, + 0.020524535328149796, + 0.01673874258995056, + 0.01902211457490921, + 0.015073482878506184, + 0.021377352997660637, + 0.014890304766595364, + 0.02190953679382801, + 0.016411012038588524, + 0.0176456980407238, + 0.023448076099157333, + 0.021867722272872925, + 0.019007699564099312, + 0.01750827766954899, + 0.019891370087862015, + 0.010727421380579472, + 0.010389117524027824, + 0.013064012862741947, + 0.010836716741323471, + 0.007409454323351383, + 0.007953863590955734, + 0.011790076270699501, + 0.014163278974592686, + 0.009539482183754444, + 0.009751616977155209, + 0.02467198483645916, + 0.012763088569045067, + 0.006014553364366293, + 0.00555763253942132, + 0.006612866651266813, + 0.007889381609857082, + 0.006675923243165016, + 0.007745583541691303, + 0.005258206278085709, + 0.0043045636266469955, + 0.009823397733271122, + 0.012126403860747814, + 0.007106426637619734, + 0.006792611908167601, + 0.00824945978820324, + 0.007419733330607414, + 0.005424176342785358, + 0.00419995654374361, + 0.012034500949084759, + 0.003993383143097162, + 0.0035649954807013273, + 0.004736433736979961, + 0.006101294420659542, + 0.016166748479008675, + 0.005065063014626503, + 0.004820938687771559, + 0.009304473176598549, + 0.009050090797245502, + 0.009566452354192734, + 0.006248573772609234, + 0.01516055315732956, + 0.008390056900680065, + 0.006738670170307159, + 0.011486722156405449, + 0.00372688309289515, + 0.003740341402590275, + 0.0032323417253792286, + 0.0029152599163353443, + 0.0027626187074929476, + 0.003902630414813757, + 0.0031223027035593987, + 0.0023874258622527122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.01107686385512352 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement handheld and panning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast expanse of rolling, dry, golden-brown grassland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are people dressed in worn denim jeans, plaid flannel shirts, and sturdy leather boots herding reddish-brown cattle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the cattle moving at a slow, powerful pace with coats coated in a thin layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft afternoon sunlight casting long shadows and a hazy, golden atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038090530782938004 + ], + "opposite_logprob_tokens": [ + -3.288090467453003 + ], + "logprob_reward": -0.038090530782938004, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038090530782938004, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17114292085170746, + 0.08290529996156693, + 0.10141749680042267, + 0.22194983065128326, + 0.060441818088293076, + 0.0355534553527832, + 0.18572765588760376, + 0.054897554218769073, + 0.05259184166789055, + 0.03337208554148674 + ], + "credit_map": [ + 0.02185378223657608, + 0.015884872525930405, + 0.015755239874124527, + 0.01685093529522419, + 0.020353244617581367, + 0.03259347379207611, + 0.02760659158229828, + 0.0202447772026062, + 0.00708678038790822, + 0.008139913901686668, + 0.009249329566955566, + 0.009999840520322323, + 0.010271181352436543, + 0.010962883941829205, + 0.01800379529595375, + 0.009191582910716534, + 0.01596139930188656, + 0.008617258630692959, + 0.015382429584860802, + 0.010662153363227844, + 0.019164633005857468, + 0.00844875629991293, + 0.012659402564167976, + 0.010521456599235535, + 0.03023476153612137, + 0.03204633668065071, + 0.0384969636797905, + 0.05070577561855316, + 0.016647564247250557, + 0.015507643111050129, + 0.02269197441637516, + 0.015618818812072277, + 0.006280376575887203, + 0.011649333871901035, + 0.0034788756165653467, + 0.005645522382110357, + 0.004801966715604067, + 0.017818164080381393, + 0.006252163089811802, + 0.004515410866588354, + 0.004369120113551617, + 0.004112984519451857, + 0.0036494540981948376, + 0.004191575106233358, + 0.007810911163687706, + 0.005633050110191107, + 0.003265399020165205, + 0.002520962618291378, + 0.02355332113802433, + 0.03536368906497955, + 0.023423580452799797, + 0.027370130643248558, + 0.020455211400985718, + 0.031354501843452454, + 0.013163033872842789, + 0.01104420144110918, + 0.008986342698335648, + 0.005263390019536018, + 0.004243525676429272, + 0.004386487882584333, + 0.005570690613240004, + 0.01454543974250555, + 0.0046994746662676334, + 0.007202204316854477, + 0.0065264031291007996, + 0.007814669981598854, + 0.006774283945560455, + 0.005034507252275944, + 0.004736066795885563, + 0.006381225306540728, + 0.005050501320511103, + 0.010274183005094528, + 0.0044695064425468445, + 0.0036757641937583685, + 0.0050826952792704105, + 0.0034626824781298637, + 0.004000421147793531, + 0.006117828190326691, + 0.0030320549849420786, + 0.003531134920194745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.038090530782938004 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement handheld and panning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast expanse of rolling, dry, golden-brown grassland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are people dressed in worn denim jeans, plaid flannel shirts, and sturdy leather boots herding reddish-brown cattle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the cattle moving at a slow, powerful pace with coats coated in a thin layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft afternoon sunlight casting long shadows and a hazy, golden atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003610759973526001 + ], + "opposite_logprob_tokens": [ + -5.628610610961914 + ], + "logprob_reward": -0.003610759973526001, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003610759973526001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1980636715888977, + 0.04592335224151611, + 0.081879623234272, + 0.09619545936584473, + 0.03706924989819527, + 0.049327146261930466, + 0.32018741965293884, + 0.025127340108156204, + 0.12163697183132172, + 0.02458973601460457 + ], + "credit_map": [ + 0.020020533353090286, + 0.01937570981681347, + 0.01728181540966034, + 0.0181154515594244, + 0.023731745779514313, + 0.041737303137779236, + 0.030165284872055054, + 0.027635831385850906, + 0.0050655934028327465, + 0.0049521601758897305, + 0.005468237679451704, + 0.005220626015216112, + 0.006695366930216551, + 0.005643244367092848, + 0.008172901347279549, + 0.00470522278919816, + 0.014076200313866138, + 0.007243617437779903, + 0.01362892147153616, + 0.010242714546620846, + 0.013304772786796093, + 0.00560829509049654, + 0.011083600111305714, + 0.006691506132483482, + 0.010144832544028759, + 0.015800949186086655, + 0.016003649681806564, + 0.021032294258475304, + 0.007764315232634544, + 0.012139773927628994, + 0.007174709811806679, + 0.00613493612036109, + 0.004245041869580746, + 0.006015280727297068, + 0.0035450051072984934, + 0.0050089131109416485, + 0.0030055991373956203, + 0.009619608521461487, + 0.0026215799152851105, + 0.003008223371580243, + 0.004396168049424887, + 0.0056716063991189, + 0.0032256648410111666, + 0.005492499563843012, + 0.010649804957211018, + 0.012419232167303562, + 0.0035717086866497993, + 0.0039004620630294085, + 0.05902590602636337, + 0.05045269429683685, + 0.04322047904133797, + 0.05000276118516922, + 0.03304899111390114, + 0.038867104798555374, + 0.022909484803676605, + 0.02266000770032406, + 0.002969935070723295, + 0.00344080594368279, + 0.0022419802844524384, + 0.002817160915583372, + 0.005084298085421324, + 0.0026632219087332487, + 0.002261787187308073, + 0.003648150246590376, + 0.016236573457717896, + 0.017640283331274986, + 0.015593909658491611, + 0.013357710093259811, + 0.01276689488440752, + 0.01309215184301138, + 0.011581775732338428, + 0.021367670968174934, + 0.0035150377079844475, + 0.003947196062654257, + 0.0024118053261190653, + 0.00337023357860744, + 0.004158817231655121, + 0.003402257803827524, + 0.0020988713949918747, + 0.0016855158610269427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.003610759973526001 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Handheld, panning, wide-angle shots immerse the viewer directly in the raw energy of cattle ranching. The unsteady movement conveys a sense of immediate, improvisational action. Across a vast expanse of rolling, dry, golden-brown grassland, a handful of people on foot purposefully herd a large herd of reddish-brown cattle. The ranchers, a group of men and women, are dressed in practical workwear: worn denim jeans, plaid flannel shirts, and sturdy leather boots. Some wear wide-brimmed hats, casting their weathered faces in shadow. They move with a brisk yet determined gait, flanking the herd and occasionally waving their arms to guide stray animals back to the main group. The cattle move at a slow, powerful pace, their coats coated in a thin layer of dust, their hooves churning the dry earth. The entire scene is bathed in the warm, soft light of the afternoon sun, casting long, dramatic shadows and illuminating the thick clouds of dust kicked up by their activity, creating a hazy, golden atmosphere that is both chaotic and serenely beautiful.", + "seed": 22670, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_11/index_907_seed_22670_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.17834689966402947, + "vlm_logprob": -0.17834689966402947, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.17834689966402947, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement handheld and panning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast expanse of rolling, dry, golden-brown grassland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are people dressed in worn denim jeans, plaid flannel shirts, and sturdy leather boots herding reddish-brown cattle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the cattle moving at a slow, powerful pace with coats coated in a thin layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft afternoon sunlight casting long shadows and a hazy, golden atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8425662517547607 + ], + "opposite_logprob_tokens": [ + -0.5925662517547607 + ], + "logprob_reward": -0.8425662517547607, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8425662517547607, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.132266104221344, + 0.07430783659219742, + 0.07950174808502197, + 0.05687928944826126, + 0.08276193588972092, + 0.0923258364200592, + 0.14074823260307312, + 0.26307496428489685, + 0.05017269030213356, + 0.027961358428001404 + ], + "credit_map": [ + 0.01951916329562664, + 0.013883762992918491, + 0.016211960464715958, + 0.010569680482149124, + 0.024080533534288406, + 0.014288470149040222, + 0.02012459561228752, + 0.013587926514446735, + 0.011340524069964886, + 0.007881294935941696, + 0.01171391736716032, + 0.00824716780334711, + 0.008041102439165115, + 0.007131997030228376, + 0.006586038041859865, + 0.013365787453949451, + 0.01900259219110012, + 0.023612987250089645, + 0.0061795152723789215, + 0.0050798095762729645, + 0.007708083372563124, + 0.008122725412249565, + 0.004211206920444965, + 0.005584827624261379, + 0.0069680786691606045, + 0.005906460341066122, + 0.006348357070237398, + 0.008825021795928478, + 0.008010516874492168, + 0.00818355567753315, + 0.0055823312141001225, + 0.007054965477436781, + 0.015764346346259117, + 0.009626107290387154, + 0.007987700402736664, + 0.0079707195982337, + 0.00799707043915987, + 0.019915740936994553, + 0.005761026870459318, + 0.007739221677184105, + 0.010828680358827114, + 0.009866444393992424, + 0.010412778705358505, + 0.014570367522537708, + 0.012171045877039433, + 0.012593906372785568, + 0.012525354512035847, + 0.00935724750161171, + 0.026890866458415985, + 0.01603175885975361, + 0.011388387531042099, + 0.016562949866056442, + 0.02806210331618786, + 0.010200902819633484, + 0.01353426557034254, + 0.018076999112963676, + 0.08859707415103912, + 0.0292385034263134, + 0.00880824588239193, + 0.006252022925764322, + 0.10280958563089371, + 0.01087197382003069, + 0.008463475853204727, + 0.008034090511500835, + 0.006845617201179266, + 0.006909735035151243, + 0.0064215767197310925, + 0.0063007548451423645, + 0.008777675218880177, + 0.004571362864226103, + 0.005649152211844921, + 0.004696811083704233, + 0.0031947719398885965, + 0.003952431958168745, + 0.0031731915660202503, + 0.005374828353524208, + 0.0035140884574502707, + 0.002694611670449376, + 0.003149012802168727, + 0.00290841911919415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.8425662517547607 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement handheld and panning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast expanse of rolling, dry, golden-brown grassland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are people dressed in worn denim jeans, plaid flannel shirts, and sturdy leather boots herding reddish-brown cattle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the cattle moving at a slow, powerful pace with coats coated in a thin layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft afternoon sunlight casting long shadows and a hazy, golden atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032055205665528774 + ], + "opposite_logprob_tokens": [ + -5.753205299377441 + ], + "logprob_reward": -0.0032055205665528774, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032055205665528774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27412912249565125, + 0.03569599986076355, + 0.036837387830019, + 0.058575280010700226, + 0.058079566806554794, + 0.05615592002868652, + 0.0745958462357521, + 0.3355494737625122, + 0.024218790233135223, + 0.04616255313158035 + ], + "credit_map": [ + 0.04076703265309334, + 0.033363427966833115, + 0.05647257715463638, + 0.03600969538092613, + 0.01860664039850235, + 0.03244532644748688, + 0.02152176946401596, + 0.034942660480737686, + 0.004770784173160791, + 0.005168932490050793, + 0.00596237787976861, + 0.004630607552826405, + 0.0037240074016153812, + 0.003376798937097192, + 0.0032549784518778324, + 0.004807513207197189, + 0.005883242469280958, + 0.006981495302170515, + 0.006633519195020199, + 0.0043103802017867565, + 0.003760822117328644, + 0.003116453066468239, + 0.0034027756191790104, + 0.0027486977633088827, + 0.006274000275880098, + 0.006231175735592842, + 0.010606671683490276, + 0.007839588448405266, + 0.007809899747371674, + 0.0064129335805773735, + 0.006863749120384455, + 0.006537260953336954, + 0.007396340370178223, + 0.006709475535899401, + 0.006615860853344202, + 0.006101375911384821, + 0.006385013926774263, + 0.014507927000522614, + 0.004455924034118652, + 0.005907649639993906, + 0.006259768269956112, + 0.006448040716350079, + 0.005087840370833874, + 0.007015117444097996, + 0.006875092163681984, + 0.00861441157758236, + 0.010708018206059933, + 0.0051476298831403255, + 0.023141000419855118, + 0.008183176629245281, + 0.0031103137880563736, + 0.003969569690525532, + 0.027484282851219177, + 0.0029803235083818436, + 0.0028288576286286116, + 0.0028983261436223984, + 0.11298450827598572, + 0.037216491997241974, + 0.011468745768070221, + 0.009221444837749004, + 0.12952321767807007, + 0.013860335573554039, + 0.012071787379682064, + 0.009202963672578335, + 0.003779049264267087, + 0.0026367383543401957, + 0.003598451381549239, + 0.0040321205742657185, + 0.0029405364766716957, + 0.0017562094144523144, + 0.003409693483263254, + 0.002065991749987006, + 0.005302817095071077, + 0.007149633951485157, + 0.004175894428044558, + 0.003527882741764188, + 0.009373401291668415, + 0.007174902129918337, + 0.005063745658844709, + 0.004394277464598417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.0032055205665528774 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement handheld and panning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast expanse of rolling, dry, golden-brown grassland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are people dressed in worn denim jeans, plaid flannel shirts, and sturdy leather boots herding reddish-brown cattle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the cattle moving at a slow, powerful pace with coats coated in a thin layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft afternoon sunlight casting long shadows and a hazy, golden atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978388637304306 + ], + "opposite_logprob_tokens": [ + -4.634783744812012 + ], + "logprob_reward": -0.00978388637304306, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978388637304306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16732753813266754, + 0.05901128053665161, + 0.14753709733486176, + 0.09202740341424942, + 0.02784237265586853, + 0.03346351906657219, + 0.13864752650260925, + 0.2635255455970764, + 0.0353827178478241, + 0.03523499146103859 + ], + "credit_map": [ + 0.014842321164906025, + 0.02286115102469921, + 0.023623384535312653, + 0.02972310408949852, + 0.02232389524579048, + 0.015152085572481155, + 0.021722497418522835, + 0.017079107463359833, + 0.009854458272457123, + 0.005272984504699707, + 0.009377328678965569, + 0.0073405602015554905, + 0.006759138312190771, + 0.00862126424908638, + 0.006053215824067593, + 0.005732336081564426, + 0.04274178668856621, + 0.051080215722322464, + 0.012166526168584824, + 0.009008950553834438, + 0.013179919682443142, + 0.005680544301867485, + 0.0073256343603134155, + 0.006353521253913641, + 0.014389877207577229, + 0.010553043335676193, + 0.011139892973005772, + 0.007725671865046024, + 0.011844953522086143, + 0.013156057335436344, + 0.012455880641937256, + 0.010762023739516735, + 0.004616521764546633, + 0.0028952108696103096, + 0.0034434194676578045, + 0.003042105818167329, + 0.0033817316871136427, + 0.00477790180593729, + 0.0032715341076254845, + 0.0024139494635164738, + 0.0039006080478429794, + 0.003283608937636018, + 0.003964417614042759, + 0.007661202922463417, + 0.003656851826235652, + 0.002606061054393649, + 0.005272496026009321, + 0.0031182707753032446, + 0.042315635830163956, + 0.01736767217516899, + 0.004743027966469526, + 0.004983586259186268, + 0.0536644347012043, + 0.0063953399658203125, + 0.004582419525831938, + 0.004595406353473663, + 0.08997619152069092, + 0.03002149425446987, + 0.008172390051186085, + 0.005926887970417738, + 0.10410130769014359, + 0.010059555061161518, + 0.007605306804180145, + 0.007662403397262096, + 0.005879343952983618, + 0.004823496099561453, + 0.003631660947576165, + 0.003974871709942818, + 0.008191023021936417, + 0.0024323463439941406, + 0.003937938716262579, + 0.0025120344944298267, + 0.0036048695910722017, + 0.004145492799580097, + 0.005520184990018606, + 0.003903899108991027, + 0.005787974689155817, + 0.00482242600992322, + 0.003747966606169939, + 0.0037021785974502563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.00978388637304306 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement handheld and panning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast expanse of rolling, dry, golden-brown grassland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are people dressed in worn denim jeans, plaid flannel shirts, and sturdy leather boots herding reddish-brown cattle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the cattle moving at a slow, powerful pace with coats coated in a thin layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft afternoon sunlight casting long shadows and a hazy, golden atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033695101737976074 + ], + "opposite_logprob_tokens": [ + -3.4086952209472656 + ], + "logprob_reward": -0.033695101737976074, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033695101737976074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21332544088363647, + 0.09771790355443954, + 0.07328640669584274, + 0.03746607527136803, + 0.13495950400829315, + 0.05309981107711792, + 0.17552031576633453, + 0.12281954288482666, + 0.052692513912916183, + 0.03911247104406357 + ], + "credit_map": [ + 0.023610802367329597, + 0.03126884251832962, + 0.0292796827852726, + 0.027025066316127777, + 0.027788737788796425, + 0.02104293927550316, + 0.02697845920920372, + 0.02633092924952507, + 0.012844766490161419, + 0.008008690550923347, + 0.017729103565216064, + 0.00967942364513874, + 0.017772777006030083, + 0.011726376600563526, + 0.011290078982710838, + 0.008666682057082653, + 0.018037831410765648, + 0.023497557267546654, + 0.006277413573116064, + 0.0039521451108157635, + 0.0076391082257032394, + 0.0043032243847846985, + 0.00399888725951314, + 0.00558023527264595, + 0.00456342613324523, + 0.003969086799770594, + 0.0054282876662909985, + 0.005455853417515755, + 0.004507420118898153, + 0.004387355875223875, + 0.005700869485735893, + 0.0034537736792117357, + 0.019414961338043213, + 0.015032140538096428, + 0.0094954464584589, + 0.015395422466099262, + 0.017503203824162483, + 0.034082163125276566, + 0.01094950083643198, + 0.013086680322885513, + 0.006039225496351719, + 0.00542911933735013, + 0.006486124824732542, + 0.007823624648153782, + 0.007971270941197872, + 0.006339013576507568, + 0.00773391081020236, + 0.005277520976960659, + 0.0531311072409153, + 0.01979058049619198, + 0.00724917184561491, + 0.00830911472439766, + 0.06308112293481827, + 0.008786740712821484, + 0.006988111883401871, + 0.008184368722140789, + 0.04136744141578674, + 0.013490141369402409, + 0.004142263438552618, + 0.0030231380369514227, + 0.0474180169403553, + 0.005007817409932613, + 0.003975991625338793, + 0.004394739866256714, + 0.008122412487864494, + 0.009044538252055645, + 0.003914286848157644, + 0.004838054068386555, + 0.010877502150833607, + 0.004679024685174227, + 0.00620739720761776, + 0.005009299144148827, + 0.005289916414767504, + 0.0037367884069681168, + 0.004804011434316635, + 0.004930174443870783, + 0.004847663454711437, + 0.005460538901388645, + 0.005504217464476824, + 0.004539160523563623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.033695101737976074 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement handheld and panning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast expanse of rolling, dry, golden-brown grassland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are people dressed in worn denim jeans, plaid flannel shirts, and sturdy leather boots herding reddish-brown cattle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the cattle moving at a slow, powerful pace with coats coated in a thin layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft afternoon sunlight casting long shadows and a hazy, golden atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002483737887814641 + ], + "opposite_logprob_tokens": [ + -6.00248384475708 + ], + "logprob_reward": -0.002483737887814641, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002483737887814641, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1467004418373108, + 0.043258052319288254, + 0.03824704885482788, + 0.08752118796110153, + 0.319326788187027, + 0.10543892532587051, + 0.03684847056865692, + 0.03606078773736954, + 0.08127210289239883, + 0.10532619804143906 + ], + "credit_map": [ + 0.01713305339217186, + 0.015037518925964832, + 0.020697763189673424, + 0.01836315728724003, + 0.021910356357693672, + 0.01393078826367855, + 0.023104887455701828, + 0.016522908583283424, + 0.007152431644499302, + 0.0038415631279349327, + 0.007724209688603878, + 0.00482043856754899, + 0.00600073579698801, + 0.004098910838365555, + 0.0038689891807734966, + 0.005750774871557951, + 0.005907965358346701, + 0.0059019532054662704, + 0.006376476958394051, + 0.005309916101396084, + 0.003714169841259718, + 0.003484134329482913, + 0.0037084792274981737, + 0.0038439517375081778, + 0.015469840727746487, + 0.011160055175423622, + 0.01917305216193199, + 0.014300062321126461, + 0.00615104241296649, + 0.0055109052918851376, + 0.008824290707707405, + 0.0069319275207817554, + 0.04588431119918823, + 0.038653530180454254, + 0.020718194544315338, + 0.036164041608572006, + 0.039667584002017975, + 0.0851965844631195, + 0.020757578313350677, + 0.032284971326589584, + 0.012150854803621769, + 0.014562444761395454, + 0.009888650849461555, + 0.014345093630254269, + 0.014592255465686321, + 0.014239905402064323, + 0.014749808236956596, + 0.010909901931881905, + 0.00682626198977232, + 0.004293795675039291, + 0.004103638231754303, + 0.004960200749337673, + 0.00796579197049141, + 0.0028290986083447933, + 0.002685654442757368, + 0.0031840279698371887, + 0.005849574692547321, + 0.005038386210799217, + 0.003928839694708586, + 0.0036663743667304516, + 0.005153198726475239, + 0.002716187620535493, + 0.004442536737769842, + 0.005265685264021158, + 0.010718860663473606, + 0.015580687671899796, + 0.006917861290276051, + 0.008466383442282677, + 0.021344568580389023, + 0.005144952796399593, + 0.006875673774629831, + 0.006223107688128948, + 0.013748426921665668, + 0.01925460435450077, + 0.006550321355462074, + 0.009134616702795029, + 0.02313542366027832, + 0.014242930337786674, + 0.010713629424571991, + 0.008546235039830208 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.002483737887814641 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Handheld, panning, wide-angle shots immerse the viewer directly in the raw energy of cattle ranching. The unsteady movement conveys a sense of immediate, improvisational action. Across a vast expanse of rolling, dry, golden-brown grassland, a handful of people on foot purposefully herd a large herd of reddish-brown cattle. The ranchers, a group of men and women, are dressed in practical workwear: worn denim jeans, plaid flannel shirts, and sturdy leather boots. Some wear wide-brimmed hats, casting their weathered faces in shadow. They move with a brisk yet determined gait, flanking the herd and occasionally waving their arms to guide stray animals back to the main group. The cattle move at a slow, powerful pace, their coats coated in a thin layer of dust, their hooves churning the dry earth. The entire scene is bathed in the warm, soft light of the afternoon sun, casting long, dramatic shadows and illuminating the thick clouds of dust kicked up by their activity, creating a hazy, golden atmosphere that is both chaotic and serenely beautiful.", + "seed": 22671, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_11/index_907_seed_22671_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.16352490331046282, + "vlm_logprob": -0.16352490331046282, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.16352490331046282, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement handheld and panning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast expanse of rolling, dry, golden-brown grassland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are people dressed in worn denim jeans, plaid flannel shirts, and sturdy leather boots herding reddish-brown cattle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the cattle moving at a slow, powerful pace with coats coated in a thin layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft afternoon sunlight casting long shadows and a hazy, golden atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7753727436065674 + ], + "opposite_logprob_tokens": [ + -0.6503727436065674 + ], + "logprob_reward": -0.7753727436065674, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7753727436065674, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06111554056406021, + 0.23603853583335876, + 0.04365939646959305, + 0.09370023012161255, + 0.02007233165204525, + 0.029782628640532494, + 0.19747571647167206, + 0.04714157432317734, + 0.14181219041347504, + 0.12920181453227997 + ], + "credit_map": [ + 0.006654970347881317, + 0.008699254132807255, + 0.005504747852683067, + 0.009885198436677456, + 0.00827909354120493, + 0.0063391379080712795, + 0.008096566423773766, + 0.007656573783606291, + 0.013629259541630745, + 0.08963054418563843, + 0.007895519956946373, + 0.008345263078808784, + 0.0946696400642395, + 0.006453168112784624, + 0.008052637800574303, + 0.007362521253526211, + 0.0033390289172530174, + 0.009541195817291737, + 0.0033297999761998653, + 0.0038765145000070333, + 0.011379078961908817, + 0.002929957816377282, + 0.003889207262545824, + 0.00537461880594492, + 0.013105073943734169, + 0.016159391030669212, + 0.008414706215262413, + 0.0077256932854652405, + 0.010349037125706673, + 0.00880054198205471, + 0.012085207737982273, + 0.01706058531999588, + 0.002780931768938899, + 0.002565415808930993, + 0.003608918283134699, + 0.0018794086063280702, + 0.0022244625724852085, + 0.0022799561265856028, + 0.001895640161819756, + 0.002837598556652665, + 0.006413922645151615, + 0.004336269106715918, + 0.0034442588221281767, + 0.0022397604770958424, + 0.0020496195647865534, + 0.0031222179532051086, + 0.003287555417045951, + 0.0048890262842178345, + 0.023189334198832512, + 0.06625572592020035, + 0.013031385838985443, + 0.009418881498277187, + 0.00860396958887577, + 0.009332604706287384, + 0.009963692165911198, + 0.05768013373017311, + 0.007240783888846636, + 0.008468043059110641, + 0.005394484847784042, + 0.006457535084336996, + 0.005335710011422634, + 0.0040621086955070496, + 0.0056532914750278, + 0.004529622383415699, + 0.026998715475201607, + 0.03281691297888756, + 0.00905392225831747, + 0.00965017732232809, + 0.022791074588894844, + 0.011418668553233147, + 0.015383966267108917, + 0.013698766008019447, + 0.013184120878577232, + 0.02885616570711136, + 0.007989780977368355, + 0.006716043222695589, + 0.010118243284523487, + 0.00683154072612524, + 0.04107251390814781, + 0.014433415606617928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 8 + ], + "selected_reward": -0.7753727436065674 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement handheld and panning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast expanse of rolling, dry, golden-brown grassland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are people dressed in worn denim jeans, plaid flannel shirts, and sturdy leather boots herding reddish-brown cattle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the cattle moving at a slow, powerful pace with coats coated in a thin layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft afternoon sunlight casting long shadows and a hazy, golden atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002508352743461728 + ], + "opposite_logprob_tokens": [ + -6.002508163452148 + ], + "logprob_reward": -0.002508352743461728, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002508352743461728, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08112866431474686, + 0.16192786395549774, + 0.2137327790260315, + 0.06294453889131546, + 0.09211511164903641, + 0.022367320954799652, + 0.11228125542402267, + 0.061776068061590195, + 0.0477147214114666, + 0.1440117508172989 + ], + "credit_map": [ + 0.010965773835778236, + 0.011381695047020912, + 0.008918135426938534, + 0.01137601863592863, + 0.0117411557585001, + 0.008404841646552086, + 0.009294080547988415, + 0.00904695875942707, + 0.009228912182152271, + 0.06608409434556961, + 0.007563424296677113, + 0.009109017439186573, + 0.047775931656360626, + 0.004927559290081263, + 0.0073966714553534985, + 0.009842232801020145, + 0.01576993055641651, + 0.07723549008369446, + 0.007996573112905025, + 0.009861016646027565, + 0.0664612352848053, + 0.011253751814365387, + 0.009513272903859615, + 0.01564149744808674, + 0.010800992138683796, + 0.011776115745306015, + 0.0055321636609733105, + 0.003748884191736579, + 0.009110335260629654, + 0.005658689886331558, + 0.006331506185233593, + 0.009985842742025852, + 0.013712320476770401, + 0.012653745710849762, + 0.01567496731877327, + 0.012699639424681664, + 0.009378067217767239, + 0.006929159164428711, + 0.008482305333018303, + 0.012584898620843887, + 0.005390058271586895, + 0.00429266644641757, + 0.0020942373666912317, + 0.00231363857164979, + 0.00201884051784873, + 0.0017027383437380195, + 0.0022673332132399082, + 0.0022878062445670366, + 0.019470935687422752, + 0.04322294890880585, + 0.006141048390418291, + 0.004845608491450548, + 0.005653138272464275, + 0.004747410770505667, + 0.003254658542573452, + 0.024945495650172234, + 0.010281075723469257, + 0.009185999631881714, + 0.006000409834086895, + 0.009981672279536724, + 0.00569023285061121, + 0.004690551199018955, + 0.006060924846678972, + 0.009885200299322605, + 0.00911671482026577, + 0.011458850465714931, + 0.0036417637020349503, + 0.004909992218017578, + 0.00646745041012764, + 0.003271677065640688, + 0.003877911251038313, + 0.004970355425029993, + 0.011341460980474949, + 0.028946593403816223, + 0.007916068658232689, + 0.0063273487612605095, + 0.008520548231899738, + 0.0080790463835001, + 0.05371541902422905, + 0.019165245816111565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 9 + ], + "selected_reward": -0.002508352743461728 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement handheld and panning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast expanse of rolling, dry, golden-brown grassland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are people dressed in worn denim jeans, plaid flannel shirts, and sturdy leather boots herding reddish-brown cattle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the cattle moving at a slow, powerful pace with coats coated in a thin layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft afternoon sunlight casting long shadows and a hazy, golden atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009785185568034649 + ], + "opposite_logprob_tokens": [ + -4.634785175323486 + ], + "logprob_reward": -0.009785185568034649, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009785185568034649, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06203838065266609, + 0.31188637018203735, + 0.2734511196613312, + 0.04432426765561104, + 0.03167590871453285, + 0.05503571406006813, + 0.13743802905082703, + 0.031541358679533005, + 0.02417960949242115, + 0.028429316356778145 + ], + "credit_map": [ + 0.009809711016714573, + 0.0101422518491745, + 0.00491377804428339, + 0.006386717315763235, + 0.009259925223886967, + 0.006215438712388277, + 0.007164208684116602, + 0.008146347478032112, + 0.010505088604986668, + 0.13471728563308716, + 0.005881838966161013, + 0.007966211996972561, + 0.11735155433416367, + 0.008506173267960548, + 0.011991329491138458, + 0.0149668725207448, + 0.013804235495626926, + 0.11407775431871414, + 0.006468162871897221, + 0.009493758901953697, + 0.09893374145030975, + 0.009388483129441738, + 0.008468400686979294, + 0.012816589325666428, + 0.005151689983904362, + 0.0058002895675599575, + 0.004333621822297573, + 0.004108575638383627, + 0.008932765573263168, + 0.005811712238937616, + 0.003978793043643236, + 0.006206819787621498, + 0.004561739042401314, + 0.00391868781298399, + 0.00508083263412118, + 0.0043563577346503735, + 0.003300396027043462, + 0.002486196346580982, + 0.004070349968969822, + 0.003901349613443017, + 0.010326764546334743, + 0.006874925922602415, + 0.00513860210776329, + 0.004737500101327896, + 0.0056907725520431995, + 0.006312786601483822, + 0.006852517835795879, + 0.009101847186684608, + 0.015827205032110214, + 0.046696536242961884, + 0.009113525971770287, + 0.006779612973332405, + 0.005860663950443268, + 0.006124100647866726, + 0.006298443768173456, + 0.04073793813586235, + 0.0034410618245601654, + 0.004461843520402908, + 0.002524526556953788, + 0.0036562413442879915, + 0.0035671440418809652, + 0.003983381669968367, + 0.006286962423473597, + 0.0036201970651745796, + 0.003195017110556364, + 0.004097782075405121, + 0.0019252158235758543, + 0.0019523785449564457, + 0.003664234420284629, + 0.002450471743941307, + 0.0039221360348165035, + 0.0029723725747317076, + 0.0021883933804929256, + 0.0042060259729623795, + 0.0020507611334323883, + 0.0019956021569669247, + 0.0028288650792092085, + 0.0019534225575625896, + 0.009545385837554932, + 0.003660857677459717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 6 + ], + "selected_reward": -0.009785185568034649 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement handheld and panning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast expanse of rolling, dry, golden-brown grassland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are people dressed in worn denim jeans, plaid flannel shirts, and sturdy leather boots herding reddish-brown cattle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the cattle moving at a slow, powerful pace with coats coated in a thin layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft afternoon sunlight casting long shadows and a hazy, golden atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026347355917096138 + ], + "opposite_logprob_tokens": [ + -3.6513473987579346 + ], + "logprob_reward": -0.026347355917096138, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026347355917096138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06094488501548767, + 0.3607102632522583, + 0.2731151282787323, + 0.033438220620155334, + 0.03894149512052536, + 0.022136563435196877, + 0.12622112035751343, + 0.023284005001187325, + 0.013100497424602509, + 0.04810779169201851 + ], + "credit_map": [ + 0.007297006901353598, + 0.01154661551117897, + 0.005539599806070328, + 0.007829675450921059, + 0.00888124480843544, + 0.006339063867926598, + 0.0070028407499194145, + 0.006508835591375828, + 0.014587009325623512, + 0.16570927202701569, + 0.005880110431462526, + 0.008466058410704136, + 0.13452398777008057, + 0.008760469034314156, + 0.01120926532894373, + 0.011574091389775276, + 0.01602308638393879, + 0.10705447942018509, + 0.009526458568871021, + 0.012596972286701202, + 0.09494374692440033, + 0.010227041319012642, + 0.010776561684906483, + 0.011966797523200512, + 0.0037340903654694557, + 0.0037909243255853653, + 0.004274744540452957, + 0.004505236633121967, + 0.003199649276211858, + 0.004133197944611311, + 0.004853715188801289, + 0.004946662578731775, + 0.006067302078008652, + 0.005054296925663948, + 0.006363490130752325, + 0.004771074745804071, + 0.0034187976270914078, + 0.0038311651442199945, + 0.0041114287450909615, + 0.005323937628418207, + 0.004009178373962641, + 0.003302587429061532, + 0.0027107661589980125, + 0.0027646662201732397, + 0.002125222934409976, + 0.001941655995324254, + 0.0018967135110870004, + 0.0033857738599181175, + 0.01737573929131031, + 0.045247990638017654, + 0.008087389171123505, + 0.006060448475182056, + 0.005462766624987125, + 0.005181996151804924, + 0.004502146039158106, + 0.034302644431591034, + 0.0034149405546486378, + 0.0041221734136343, + 0.0026391262654215097, + 0.0037496520671993494, + 0.0021667410619556904, + 0.001936226966790855, + 0.0021493600215762854, + 0.003105785930529237, + 0.0020965079311281443, + 0.0028594178147614002, + 0.001115479040890932, + 0.0011886623688042164, + 0.0014982593711465597, + 0.0010564645053818822, + 0.0017479460220783949, + 0.0015377592062577605, + 0.0037144499365240335, + 0.00882144644856453, + 0.0027415992226451635, + 0.002464441116899252, + 0.002895087469369173, + 0.002836509607732296, + 0.017979737371206284, + 0.006654522847384214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 6 + ], + "selected_reward": -0.026347355917096138 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement handheld and panning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast expanse of rolling, dry, golden-brown grassland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are people dressed in worn denim jeans, plaid flannel shirts, and sturdy leather boots herding reddish-brown cattle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the cattle moving at a slow, powerful pace with coats coated in a thin layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft afternoon sunlight casting long shadows and a hazy, golden atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036108787171542645 + ], + "opposite_logprob_tokens": [ + -5.628611087799072 + ], + "logprob_reward": -0.0036108787171542645, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036108787171542645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04573991522192955, + 0.038131024688482285, + 0.6826938986778259, + 0.03228415921330452, + 0.04764200374484062, + 0.009259536862373352, + 0.015874261036515236, + 0.04741876572370529, + 0.026656392961740494, + 0.05429999902844429 + ], + "credit_map": [ + 0.006542883813381195, + 0.007796032354235649, + 0.004541036207228899, + 0.006695770192891359, + 0.006736063864082098, + 0.0041180928237736225, + 0.004828397184610367, + 0.004481640178710222, + 0.006219981238245964, + 0.00922726932913065, + 0.003052622312679887, + 0.003865036414936185, + 0.0029004469979554415, + 0.0035741827450692654, + 0.003569017630070448, + 0.00572246452793479, + 0.03132062405347824, + 0.26337340474128723, + 0.020680101588368416, + 0.025201819837093353, + 0.25396105647087097, + 0.02607238106429577, + 0.024855848401784897, + 0.037228722125291824, + 0.004674107767641544, + 0.0050870506092906, + 0.0029671036172658205, + 0.0025656563229858875, + 0.005618824623525143, + 0.0037829666398465633, + 0.003038279013708234, + 0.004550169687718153, + 0.007807587739080191, + 0.0068346369080245495, + 0.008421177975833416, + 0.00646160589531064, + 0.00501559441909194, + 0.0031360911671072245, + 0.003689332166686654, + 0.006275977473706007, + 0.0011202743044123054, + 0.0009673958411440253, + 0.0011249436065554619, + 0.0013536588521674275, + 0.001179801532998681, + 0.0009216430480591953, + 0.0013339333236217499, + 0.0012578866444528103, + 0.0016565888654440641, + 0.003652099287137389, + 0.001403129892423749, + 0.0013745789183303714, + 0.001063191913999617, + 0.0010956815676763654, + 0.001621734001673758, + 0.004007257521152496, + 0.008055259473621845, + 0.005655546672642231, + 0.00517639797180891, + 0.005263980943709612, + 0.0039911880157887936, + 0.0032111883629113436, + 0.004412870854139328, + 0.011652330867946148, + 0.006042136810719967, + 0.00790859293192625, + 0.0016746236942708492, + 0.002297761617228389, + 0.0024373000487685204, + 0.0016339238500222564, + 0.002183155156672001, + 0.0024788971059024334, + 0.00392139283940196, + 0.009359590709209442, + 0.0034827624913305044, + 0.0034095619339495897, + 0.0037424855399876833, + 0.003206404158845544, + 0.019857538864016533, + 0.007320263888686895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 4 + ], + "selected_reward": -0.0036108787171542645 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Handheld, panning, wide-angle shots immerse the viewer directly in the raw energy of cattle ranching. The unsteady movement conveys a sense of immediate, improvisational action. Across a vast expanse of rolling, dry, golden-brown grassland, a handful of people on foot purposefully herd a large herd of reddish-brown cattle. The ranchers, a group of men and women, are dressed in practical workwear: worn denim jeans, plaid flannel shirts, and sturdy leather boots. Some wear wide-brimmed hats, casting their weathered faces in shadow. They move with a brisk yet determined gait, flanking the herd and occasionally waving their arms to guide stray animals back to the main group. The cattle move at a slow, powerful pace, their coats coated in a thin layer of dust, their hooves churning the dry earth. The entire scene is bathed in the warm, soft light of the afternoon sun, casting long, dramatic shadows and illuminating the thick clouds of dust kicked up by their activity, creating a hazy, golden atmosphere that is both chaotic and serenely beautiful.", + "seed": 22672, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_11/index_907_seed_22672_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.13825244856998326, + "vlm_logprob": -0.13825244856998326, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.13825244856998326, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement handheld and panning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast expanse of rolling, dry, golden-brown grassland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are people dressed in worn denim jeans, plaid flannel shirts, and sturdy leather boots herding reddish-brown cattle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the cattle moving at a slow, powerful pace with coats coated in a thin layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft afternoon sunlight casting long shadows and a hazy, golden atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6534406542778015 + ], + "opposite_logprob_tokens": [ + -0.7784406542778015 + ], + "logprob_reward": -0.6534406542778015, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6534406542778015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22506865859031677, + 0.18364858627319336, + 0.043624527752399445, + 0.12616382539272308, + 0.1517607420682907, + 0.04654408246278763, + 0.03530857339501381, + 0.10689160227775574, + 0.03386692330241203, + 0.047122519463300705 + ], + "credit_map": [ + 0.0214031171053648, + 0.015048437751829624, + 0.025817211717367172, + 0.05419979244470596, + 0.026652159169316292, + 0.02255224995315075, + 0.026159020140767097, + 0.033236656337976456, + 0.019779153168201447, + 0.015783796086907387, + 0.029281219467520714, + 0.03869874030351639, + 0.014970849268138409, + 0.021476970985531807, + 0.021027397364377975, + 0.022630447521805763, + 0.004341857973486185, + 0.004733858164399862, + 0.003872441127896309, + 0.0031343360897153616, + 0.005726808216422796, + 0.008828980848193169, + 0.009080605581402779, + 0.003905640682205558, + 0.01864069141447544, + 0.012844053097069263, + 0.014688055031001568, + 0.012797740288078785, + 0.010572276078164577, + 0.01552942581474781, + 0.014668194577097893, + 0.02642340399324894, + 0.008497689850628376, + 0.011439663358032703, + 0.013127927668392658, + 0.05772162228822708, + 0.009951991960406303, + 0.006591160781681538, + 0.01696583442389965, + 0.027464861050248146, + 0.004680618178099394, + 0.004401642829179764, + 0.007926946505904198, + 0.0067104012705385685, + 0.00401737866923213, + 0.0061591193079948425, + 0.0056267729960381985, + 0.007021204102784395, + 0.005486457608640194, + 0.003854443086311221, + 0.004386072978377342, + 0.0033956822007894516, + 0.005089438054710627, + 0.004672251176089048, + 0.003646621247753501, + 0.004777609836310148, + 0.01187037955969572, + 0.014444392174482346, + 0.015499883331358433, + 0.010665365494787693, + 0.013123179785907269, + 0.011754215694963932, + 0.014622597023844719, + 0.014911595731973648, + 0.0022040025796741247, + 0.006632140371948481, + 0.005154819693416357, + 0.00394831458106637, + 0.007518415339291096, + 0.0030977048445492983, + 0.0022990081924945116, + 0.0030125179328024387, + 0.004858911037445068, + 0.007488959934562445, + 0.00895550288259983, + 0.005105624906718731, + 0.0051144142635166645, + 0.007725235540419817, + 0.0035847239196300507, + 0.004289149772375822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.6534406542778015 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement handheld and panning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast expanse of rolling, dry, golden-brown grassland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are people dressed in worn denim jeans, plaid flannel shirts, and sturdy leather boots herding reddish-brown cattle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the cattle moving at a slow, powerful pace with coats coated in a thin layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft afternoon sunlight casting long shadows and a hazy, golden atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004651203285902739 + ], + "opposite_logprob_tokens": [ + -5.379651069641113 + ], + "logprob_reward": -0.004651203285902739, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004651203285902739, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1702180951833725, + 0.0591871552169323, + 0.0934772714972496, + 0.25792157649993896, + 0.07224678993225098, + 0.026328234001994133, + 0.08913899213075638, + 0.09315723180770874, + 0.031040353700518608, + 0.10728434473276138 + ], + "credit_map": [ + 0.01354005467146635, + 0.023903291672468185, + 0.016263017430901527, + 0.02686592936515808, + 0.012227663770318031, + 0.0320512019097805, + 0.01572650484740734, + 0.0296404417604208, + 0.0070128715597093105, + 0.004676154814660549, + 0.010241961106657982, + 0.008272002451121807, + 0.007345155812799931, + 0.007243850734084845, + 0.008113241754472256, + 0.0062819248996675014, + 0.008939670398831367, + 0.008047716692090034, + 0.008081077598035336, + 0.008057927712798119, + 0.015638891607522964, + 0.014886178076267242, + 0.020422514528036118, + 0.009403307922184467, + 0.034372299909591675, + 0.03185528144240379, + 0.029435230419039726, + 0.02773459628224373, + 0.027262523770332336, + 0.030004236847162247, + 0.02658098191022873, + 0.050676435232162476, + 0.008676989004015923, + 0.008004236035048962, + 0.006833372171968222, + 0.022535694763064384, + 0.007918229326605797, + 0.0037115798331797123, + 0.00869151670485735, + 0.00587517349049449, + 0.0024712001904845238, + 0.002947438508272171, + 0.0028091941494494677, + 0.0028501213528215885, + 0.0038180865813046694, + 0.0031180086079984903, + 0.0033974985126405954, + 0.004916686099022627, + 0.013044497929513454, + 0.01020777877420187, + 0.010903039015829563, + 0.007035234477370977, + 0.006152733229100704, + 0.015591713599860668, + 0.008147520013153553, + 0.018056487664580345, + 0.031121745705604553, + 0.030933069065213203, + 0.007925368845462799, + 0.004774201661348343, + 0.0057305339723825455, + 0.004153027664870024, + 0.003274146467447281, + 0.005245147738605738, + 0.003713223384693265, + 0.0050961533561348915, + 0.0029833426233381033, + 0.002887467620894313, + 0.01128039974719286, + 0.0016579861985519528, + 0.0017206474440172315, + 0.0017011339077726007, + 0.010889100842177868, + 0.01171611063182354, + 0.018316052854061127, + 0.011873957701027393, + 0.010738097131252289, + 0.019430069252848625, + 0.010686332359910011, + 0.013634636998176575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.004651203285902739 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement handheld and panning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast expanse of rolling, dry, golden-brown grassland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are people dressed in worn denim jeans, plaid flannel shirts, and sturdy leather boots herding reddish-brown cattle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the cattle moving at a slow, powerful pace with coats coated in a thin layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft afternoon sunlight casting long shadows and a hazy, golden atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004641829524189234 + ], + "opposite_logprob_tokens": [ + -5.379642009735107 + ], + "logprob_reward": -0.004641829524189234, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004641829524189234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12605392932891846, + 0.04696705937385559, + 0.04023438319563866, + 0.4832870662212372, + 0.0655200183391571, + 0.05921464040875435, + 0.04533696174621582, + 0.04868742451071739, + 0.03944133222103119, + 0.04525721073150635 + ], + "credit_map": [ + 0.009528011083602905, + 0.00837812852114439, + 0.011955749243497849, + 0.0316140241920948, + 0.011049935594201088, + 0.012085529044270515, + 0.010971718467772007, + 0.0304708369076252, + 0.003945773933082819, + 0.0057283188216388226, + 0.00735382130369544, + 0.0070459540002048016, + 0.004602690692991018, + 0.007506846450269222, + 0.005317286122590303, + 0.005466369446367025, + 0.004211890511214733, + 0.00426511513069272, + 0.003358340123668313, + 0.004618295002728701, + 0.0068314881063997746, + 0.005630542524158955, + 0.007946598343551159, + 0.0033721153158694506, + 0.07626824080944061, + 0.061458081007003784, + 0.057623155415058136, + 0.04972193017601967, + 0.04659837856888771, + 0.058590710163116455, + 0.05021609738469124, + 0.0828104317188263, + 0.0035791650880128145, + 0.0070939273573458195, + 0.005849290639162064, + 0.021062780171632767, + 0.006283637136220932, + 0.003672814927995205, + 0.006791502237319946, + 0.011186901479959488, + 0.004007220733910799, + 0.007227419409900904, + 0.006314793601632118, + 0.005829743109643459, + 0.011993234977126122, + 0.007718914654105902, + 0.007693855091929436, + 0.008429461158812046, + 0.005775907076895237, + 0.005412334576249123, + 0.006251126527786255, + 0.0031122618820518255, + 0.004171098582446575, + 0.006654052063822746, + 0.006873099133372307, + 0.007087081670761108, + 0.009513246826827526, + 0.010083272121846676, + 0.007640525698661804, + 0.00391504867002368, + 0.004293637350201607, + 0.004709147848188877, + 0.003497492056339979, + 0.0050350576639175415, + 0.0031818889547139406, + 0.004927434958517551, + 0.013045820407569408, + 0.0051726954989135265, + 0.005600455682724714, + 0.001982624875381589, + 0.0023522810079157352, + 0.0031781296711415052, + 0.0036723404191434383, + 0.008285118266940117, + 0.006475499831140041, + 0.005371226463466883, + 0.004963899031281471, + 0.007961072027683258, + 0.0041438499465584755, + 0.004384202416986227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.004641829524189234 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement handheld and panning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast expanse of rolling, dry, golden-brown grassland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are people dressed in worn denim jeans, plaid flannel shirts, and sturdy leather boots herding reddish-brown cattle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the cattle moving at a slow, powerful pace with coats coated in a thin layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft afternoon sunlight casting long shadows and a hazy, golden atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023283299058675766 + ], + "opposite_logprob_tokens": [ + -3.7732832431793213 + ], + "logprob_reward": -0.023283299058675766, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023283299058675766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08793151378631592, + 0.10834652185440063, + 0.035506412386894226, + 0.02974030189216137, + 0.25907111167907715, + 0.020221762359142303, + 0.0704844743013382, + 0.053963713347911835, + 0.16508063673973083, + 0.1696535348892212 + ], + "credit_map": [ + 0.007464413531124592, + 0.009523538872599602, + 0.00821385532617569, + 0.016115868464112282, + 0.009869945235550404, + 0.009401964023709297, + 0.009222869761288166, + 0.01811905764043331, + 0.008927639573812485, + 0.01275415625423193, + 0.017811140045523643, + 0.017207639291882515, + 0.009256124496459961, + 0.015337219461798668, + 0.01721079833805561, + 0.009841798804700375, + 0.003894136054441333, + 0.004601232707500458, + 0.0026094510685652494, + 0.0028902722988277674, + 0.007506920490413904, + 0.0046061198227107525, + 0.005963547620922327, + 0.0034347334876656532, + 0.004015863873064518, + 0.004179253242909908, + 0.003555959789082408, + 0.0024987729266285896, + 0.0031326860189437866, + 0.0029326167423278093, + 0.0031637661159038544, + 0.006261384580284357, + 0.018139731138944626, + 0.012561002746224403, + 0.02182639390230179, + 0.09149453043937683, + 0.015123260207474232, + 0.014459444209933281, + 0.02555019222199917, + 0.059916578233242035, + 0.0017947086598724127, + 0.0025638295337557793, + 0.002542483387514949, + 0.0032434333115816116, + 0.00340481661260128, + 0.0026993337087333202, + 0.0021830787882208824, + 0.0017900774255394936, + 0.015086112543940544, + 0.009875222109258175, + 0.00951339676976204, + 0.00405201455578208, + 0.005843952298164368, + 0.008686107583343983, + 0.005039702169597149, + 0.012387962080538273, + 0.013535335659980774, + 0.013288748450577259, + 0.005741179455071688, + 0.0040787686593830585, + 0.004412414506077766, + 0.004694940522313118, + 0.0034044014755636454, + 0.004807925317436457, + 0.01287410780787468, + 0.046871427446603775, + 0.0149317542091012, + 0.006390772759914398, + 0.07002457976341248, + 0.004764999728649855, + 0.004393448121845722, + 0.004829541314393282, + 0.015211755409836769, + 0.0371035672724247, + 0.053513478487730026, + 0.018220476806163788, + 0.011379718780517578, + 0.011490101926028728, + 0.014368918724358082, + 0.008365526795387268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.023283299058675766 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement handheld and panning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast expanse of rolling, dry, golden-brown grassland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are people dressed in worn denim jeans, plaid flannel shirts, and sturdy leather boots herding reddish-brown cattle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the cattle moving at a slow, powerful pace with coats coated in a thin layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft afternoon sunlight casting long shadows and a hazy, golden atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005245256703346968 + ], + "opposite_logprob_tokens": [ + -5.255245208740234 + ], + "logprob_reward": -0.005245256703346968, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005245256703346968, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19079555571079254, + 0.10105139017105103, + 0.10768496245145798, + 0.07361595332622528, + 0.20481480658054352, + 0.04581991210579872, + 0.05114481970667839, + 0.08801715075969696, + 0.03252256661653519, + 0.10453289747238159 + ], + "credit_map": [ + 0.01859268546104431, + 0.01644914597272873, + 0.02104763686656952, + 0.043505437672138214, + 0.019790468737483025, + 0.022598320618271828, + 0.020201006904244423, + 0.028610842302441597, + 0.008513777516782284, + 0.009832116775214672, + 0.016461791470646858, + 0.021331509575247765, + 0.008488980121910572, + 0.012258332222700119, + 0.013160798698663712, + 0.011004079133272171, + 0.009109419770538807, + 0.010373539291322231, + 0.009486858732998371, + 0.007239568512886763, + 0.024183284491300583, + 0.013215044513344765, + 0.02452601119875908, + 0.009551226161420345, + 0.008706629276275635, + 0.007618389092385769, + 0.007711938116699457, + 0.009503846988081932, + 0.007658586837351322, + 0.007665747310966253, + 0.008334885351359844, + 0.016415925696492195, + 0.011777788400650024, + 0.01697288453578949, + 0.017539454624056816, + 0.08176775276660919, + 0.008807390928268433, + 0.007788307964801788, + 0.023980701342225075, + 0.0361805222928524, + 0.0038034149911254644, + 0.005662580952048302, + 0.005377487745136023, + 0.005499769002199173, + 0.006374563090503216, + 0.006544788833707571, + 0.005636893212795258, + 0.006920413114130497, + 0.007571740075945854, + 0.007277477066963911, + 0.007809429429471493, + 0.005386377219110727, + 0.006234823260456324, + 0.006383575964719057, + 0.0038705384358763695, + 0.006610855460166931, + 0.02660072036087513, + 0.024143656715750694, + 0.010174195282161236, + 0.0065658860839903355, + 0.006208277307450771, + 0.0044701118022203445, + 0.004218584392219782, + 0.005635711830109358, + 0.00426831841468811, + 0.008540975861251354, + 0.00615283939987421, + 0.00375180272385478, + 0.0029792471323162317, + 0.001845389953814447, + 0.002312354976311326, + 0.0026716359425336123, + 0.008439564146101475, + 0.022920606657862663, + 0.034530725330114365, + 0.011789186857640743, + 0.007088668644428253, + 0.0058744922280311584, + 0.009112103842198849, + 0.004777536727488041 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.005245256703346968 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_12/index_tensor([1720, 1720])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_12/index_tensor([1720, 1720])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c8b39615d47dabb84f008416d83a987b8a1b6933 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_12/index_tensor([1720, 1720])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle close-up captures the determined ascent of a tiny, furry orange tabby, its curious green eyes wide as it cautiously places its tiny paws on the face of a larger, more complex cat. Designed to resemble a rugged mountain, the structure is covered with beige shisha ropes of various textures and a patch of brown carpet, making the kitten's ascent seem like a grand expedition. As the kitten pauses, its tail twitching slightly to maintain balance, the camera executes a slow, smooth arc to the right, gradually expanding the frame. This movement reveals the scratchboard's true setting in a spacious, sunny living room. Bright, warm sunlight pours in through the floor-to-ceiling windows, illuminating the dust motes dancing on the beams and casting soft, elongated shadows across the polished hardwood floor. The background unfolds to reveal a chic, minimalist space with a comfortable gray sofa, an understated wooden coffee table, and lush potted green plants, creating a tranquil and welcoming atmosphere, contrasting the kitten's miniature adventure with the vast, peaceful world of home it inhabits.", + "seed": 32324, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_12/index_1720_seed_32324_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.026008345186710358, + "vlm_logprob": -0.026008345186710358, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.026008345186710358, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, furry orange tabby kitten with wide green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kitten climbing the face of a larger cat that is designed to resemble a rugged mountain covered in beige shisha ropes and brown carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the kitten pause with its tail twitching slightly to maintain balance during the ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, smooth arc to the right that gradually expands the frame to reveal a spacious, sunny living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist living room featuring a gray sofa, a wooden coffee table, potted green plants, and bright sunlight illuminating dust motes and hardwood floors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0098367715254426 + ], + "opposite_logprob_tokens": [ + -4.759836673736572 + ], + "logprob_reward": -0.0098367715254426, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.0098367715254426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14229753613471985, + 0.07385797053575516, + 0.06956065446138382, + 0.053714144974946976, + 0.053033214062452316, + 0.07837913185358047, + 0.04391338303685188, + 0.06119992583990097, + 0.20173634588718414, + 0.22230768203735352 + ], + "credit_map": [ + 0.02188575081527233, + 0.0164070762693882, + 0.01551057118922472, + 0.03153878450393677, + 0.019960710778832436, + 0.012135930359363556, + 0.012366079725325108, + 0.012492652982473373, + 0.009463579393923283, + 0.015133568085730076, + 0.00750968512147665, + 0.012008782476186752, + 0.008542961440980434, + 0.0061973147094249725, + 0.007863898761570454, + 0.007138182874768972, + 0.006112922448664904, + 0.008310403674840927, + 0.0057549625635147095, + 0.014970585703849792, + 0.011624208651483059, + 0.007060430943965912, + 0.006725768558681011, + 0.009001378901302814, + 0.007817421108484268, + 0.005209034308791161, + 0.00599007960408926, + 0.007492575328797102, + 0.007306011859327555, + 0.007310235872864723, + 0.005271967500448227, + 0.007316822186112404, + 0.010014006868004799, + 0.007309819106012583, + 0.006793500855565071, + 0.008584212511777878, + 0.007134766783565283, + 0.00436465535312891, + 0.0041945455595850945, + 0.004637707956135273, + 0.010560673661530018, + 0.007522169500589371, + 0.009342863224446774, + 0.013075131922960281, + 0.017538463696837425, + 0.006659373641014099, + 0.006969868205487728, + 0.0067105903290212154, + 0.006329848896712065, + 0.008284252136945724, + 0.003992652520537376, + 0.004383784718811512, + 0.006005008704960346, + 0.005495975259691477, + 0.0043662055395543575, + 0.00505566131323576, + 0.006126836407929659, + 0.016632426530122757, + 0.005963505245745182, + 0.0042673698626458645, + 0.01313124317675829, + 0.00627320958301425, + 0.0038922273088246584, + 0.004913111682981253, + 0.02364702709019184, + 0.03188559412956238, + 0.025718659162521362, + 0.015981139615178108, + 0.029390938580036163, + 0.027732307091355324, + 0.01496575865894556, + 0.03241494297981262, + 0.02058073878288269, + 0.028662431985139847, + 0.017010463401675224, + 0.030434904620051384, + 0.025439364835619926, + 0.03950510546565056, + 0.03737347573041916, + 0.02330121025443077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.0098367715254426 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, furry orange tabby kitten with wide green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kitten climbing the face of a larger cat that is designed to resemble a rugged mountain covered in beige shisha ropes and brown carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the kitten pause with its tail twitching slightly to maintain balance during the ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, smooth arc to the right that gradually expands the frame to reveal a spacious, sunny living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist living room featuring a gray sofa, a wooden coffee table, potted green plants, and bright sunlight illuminating dust motes and hardwood floors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026363147422671318 + ], + "opposite_logprob_tokens": [ + -3.6513631343841553 + ], + "logprob_reward": -0.026363147422671318, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026363147422671318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24071051180362701, + 0.09281037747859955, + 0.07970566302537918, + 0.023721417412161827, + 0.03880651295185089, + 0.040594857186079025, + 0.05736462026834488, + 0.11675681918859482, + 0.14227212965488434, + 0.16725704073905945 + ], + "credit_map": [ + 0.027812091633677483, + 0.031049931421875954, + 0.020938940346240997, + 0.04039405286312103, + 0.03562581166625023, + 0.02273278869688511, + 0.03057766892015934, + 0.03157925605773926, + 0.009618599899113178, + 0.011092946864664555, + 0.009015888907015324, + 0.012828708626329899, + 0.02041534334421158, + 0.008903569541871548, + 0.010015726089477539, + 0.010919606313109398, + 0.011187741532921791, + 0.010319494642317295, + 0.007844829000532627, + 0.018469875678420067, + 0.00952194258570671, + 0.008071529679000378, + 0.007756110746413469, + 0.0065341489389538765, + 0.0035198936238884926, + 0.00279403617605567, + 0.0030712082516402006, + 0.004020339343696833, + 0.0039024907164275646, + 0.0028601132798939943, + 0.0018253056332468987, + 0.0017280327156186104, + 0.008139435201883316, + 0.007565715815871954, + 0.004863003734499216, + 0.004344194661825895, + 0.004478453192859888, + 0.0026341548655182123, + 0.0026573899667710066, + 0.004124168306589127, + 0.008610106073319912, + 0.0045915585942566395, + 0.004913486540317535, + 0.00460045924410224, + 0.007435054052621126, + 0.0034937572199851274, + 0.0033055713865906, + 0.003644866170361638, + 0.008469045162200928, + 0.008979042991995811, + 0.005956489592790604, + 0.0054888674058020115, + 0.008742169477045536, + 0.006272332742810249, + 0.005480934865772724, + 0.007975742220878601, + 0.011117406189441681, + 0.02863776870071888, + 0.011086962185800076, + 0.007729435805231333, + 0.02331268973648548, + 0.016841556876897812, + 0.005700429901480675, + 0.012330582365393639, + 0.011743622832000256, + 0.016384782269597054, + 0.020062219351530075, + 0.011987664736807346, + 0.02237194962799549, + 0.016786916181445122, + 0.014798828400671482, + 0.028136152774095535, + 0.015808800235390663, + 0.022455213591456413, + 0.012404846027493477, + 0.018150264397263527, + 0.022066717967391014, + 0.03463337942957878, + 0.024603836238384247, + 0.01713397353887558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.026363147422671318 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, furry orange tabby kitten with wide green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kitten climbing the face of a larger cat that is designed to resemble a rugged mountain covered in beige shisha ropes and brown carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the kitten pause with its tail twitching slightly to maintain balance during the ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, smooth arc to the right that gradually expands the frame to reveal a spacious, sunny living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist living room featuring a gray sofa, a wooden coffee table, potted green plants, and bright sunlight illuminating dust motes and hardwood floors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026355832815170288 + ], + "opposite_logprob_tokens": [ + -3.651355743408203 + ], + "logprob_reward": -0.026355832815170288, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026355832815170288, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20552100241184235, + 0.07923903316259384, + 0.06276349723339081, + 0.03555234149098396, + 0.04771457612514496, + 0.05792488157749176, + 0.05853952839970589, + 0.0609518364071846, + 0.1207462027668953, + 0.27104708552360535 + ], + "credit_map": [ + 0.027656497433781624, + 0.025165319442749023, + 0.02044086903333664, + 0.04527037963271141, + 0.029850317165255547, + 0.015242960304021835, + 0.01955561153590679, + 0.022339053452014923, + 0.009797994047403336, + 0.010323534719645977, + 0.006952659226953983, + 0.01197399664670229, + 0.011450480669736862, + 0.009699404239654541, + 0.008221455849707127, + 0.010819499380886555, + 0.012668004259467125, + 0.01027690339833498, + 0.0062211789190769196, + 0.014014366082847118, + 0.006155053619295359, + 0.004355823155492544, + 0.005232453811913729, + 0.0038397114258259535, + 0.006715518422424793, + 0.0045859068632125854, + 0.0039296261966228485, + 0.0047197239473462105, + 0.0056414855644106865, + 0.00402960367500782, + 0.0030953576788306236, + 0.0028351154178380966, + 0.007670135237276554, + 0.0064406320452690125, + 0.006304874550551176, + 0.007721726316958666, + 0.007700597867369652, + 0.004918508231639862, + 0.0036737245973199606, + 0.003284378442913294, + 0.010438576340675354, + 0.006218620575964451, + 0.006869451142847538, + 0.008180752396583557, + 0.012131701223552227, + 0.0057382178492844105, + 0.004442884586751461, + 0.003904674667865038, + 0.010786700062453747, + 0.008619788102805614, + 0.008758144453167915, + 0.0069579388946294785, + 0.007273960392922163, + 0.006855243816971779, + 0.00391777977347374, + 0.005369969177991152, + 0.00894018355756998, + 0.008912152610719204, + 0.0059684868901968, + 0.0049127088859677315, + 0.010781870223581791, + 0.009161616675555706, + 0.006926633417606354, + 0.0053481790237128735, + 0.010305012576282024, + 0.012436502613127232, + 0.01380744855850935, + 0.008540600538253784, + 0.014780577272176743, + 0.021706117317080498, + 0.013574794866144657, + 0.02559514157474041, + 0.026642676442861557, + 0.03296636790037155, + 0.01790791004896164, + 0.034629158675670624, + 0.028418241068720818, + 0.05670133978128433, + 0.04527590796351433, + 0.028505459427833557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.026355832815170288 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, furry orange tabby kitten with wide green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kitten climbing the face of a larger cat that is designed to resemble a rugged mountain covered in beige shisha ropes and brown carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the kitten pause with its tail twitching slightly to maintain balance during the ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, smooth arc to the right that gradually expands the frame to reveal a spacious, sunny living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist living room featuring a gray sofa, a wooden coffee table, potted green plants, and bright sunlight illuminating dust motes and hardwood floors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494944378733635 + ], + "opposite_logprob_tokens": [ + -2.9299495220184326 + ], + "logprob_reward": -0.05494944378733635, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494944378733635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13091301918029785, + 0.07315123826265335, + 0.023945482447743416, + 0.06269997358322144, + 0.05914557725191116, + 0.03520762920379639, + 0.10629632323980331, + 0.05544580891728401, + 0.08187727630138397, + 0.37131768465042114 + ], + "credit_map": [ + 0.013111061416566372, + 0.015930449590086937, + 0.012866387143731117, + 0.01822883076965809, + 0.015714051201939583, + 0.01498575322329998, + 0.02146567963063717, + 0.01861080713570118, + 0.011294702999293804, + 0.009223759174346924, + 0.007916165515780449, + 0.013294856995344162, + 0.009148070588707924, + 0.00818236917257309, + 0.005656435154378414, + 0.008434890769422054, + 0.005174591206014156, + 0.0031929470133036375, + 0.0024926832411438227, + 0.0036197900772094727, + 0.0027436078526079655, + 0.002371378242969513, + 0.0021249663550406694, + 0.002225519623607397, + 0.00968838855624199, + 0.008262122981250286, + 0.006534854881465435, + 0.010552754625678062, + 0.008750908076763153, + 0.006128587760031223, + 0.005518651567399502, + 0.007263713516294956, + 0.011823000386357307, + 0.011567341163754463, + 0.009402311407029629, + 0.00598803348839283, + 0.009904870763421059, + 0.003707483410835266, + 0.0031623998656868935, + 0.0035901456139981747, + 0.006976426113396883, + 0.005049943458288908, + 0.004135114140808582, + 0.0036947971675544977, + 0.004510780796408653, + 0.0033428524620831013, + 0.002592452336102724, + 0.004905263893306255, + 0.02575863152742386, + 0.0244939886033535, + 0.010764810256659985, + 0.009510341100394726, + 0.012174283154308796, + 0.007993931882083416, + 0.006767062935978174, + 0.008833277970552444, + 0.005959174130111933, + 0.010165464133024216, + 0.005509034264832735, + 0.003951133228838444, + 0.009276320226490498, + 0.010590204037725925, + 0.005040095653384924, + 0.004954386968165636, + 0.008533896878361702, + 0.008326190523803234, + 0.009226595051586628, + 0.006221314892172813, + 0.01093896385282278, + 0.011709675192832947, + 0.010740308091044426, + 0.016180334612727165, + 0.03903031349182129, + 0.04476476460695267, + 0.023734521120786667, + 0.04691329225897789, + 0.03853513300418854, + 0.07683952897787094, + 0.0630645677447319, + 0.038435593247413635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.05494944378733635 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, furry orange tabby kitten with wide green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kitten climbing the face of a larger cat that is designed to resemble a rugged mountain covered in beige shisha ropes and brown carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the kitten pause with its tail twitching slightly to maintain balance during the ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, smooth arc to the right that gradually expands the frame to reveal a spacious, sunny living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist living room featuring a gray sofa, a wooden coffee table, potted green plants, and bright sunlight illuminating dust motes and hardwood floors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012536530382931232 + ], + "opposite_logprob_tokens": [ + -4.387536525726318 + ], + "logprob_reward": -0.012536530382931232, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012536530382931232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050983671098947525, + 0.058665040880441666, + 0.028834538534283638, + 0.05065074563026428, + 0.03644373267889023, + 0.027431847527623177, + 0.0518670417368412, + 0.07401598989963531, + 0.04647873342037201, + 0.5746286511421204 + ], + "credit_map": [ + 0.0056451596319675446, + 0.0070792390033602715, + 0.006003790069371462, + 0.00845879316329956, + 0.007331959903240204, + 0.0058234017342329025, + 0.005209116265177727, + 0.0054322159849107265, + 0.007439407054334879, + 0.0074796415865421295, + 0.005492190830409527, + 0.012705590575933456, + 0.00832483358681202, + 0.005293586291372776, + 0.0052746739238500595, + 0.0066551207564771175, + 0.005407976917922497, + 0.0036896232049912214, + 0.003169008996337652, + 0.007311363238841295, + 0.0031491194386035204, + 0.0020767601672559977, + 0.002177141373977065, + 0.0018535470589995384, + 0.010934765450656414, + 0.007955277338624, + 0.005717556457966566, + 0.0089814318343997, + 0.005544423125684261, + 0.003852555062621832, + 0.0029660258442163467, + 0.00469871936365962, + 0.007381473667919636, + 0.006919887848198414, + 0.0059559764340519905, + 0.0037285806611180305, + 0.005877149291336536, + 0.0021760324016213417, + 0.0020702474284917116, + 0.0023343865759670734, + 0.006431357469409704, + 0.004203188698738813, + 0.0029432831797748804, + 0.0024692874867469072, + 0.003063380252569914, + 0.002541269874200225, + 0.002220583613961935, + 0.0035594990476965904, + 0.011654352769255638, + 0.010458732023835182, + 0.006278970744460821, + 0.005186592228710651, + 0.006386109162122011, + 0.005053448025137186, + 0.002953306771814823, + 0.003895536996424198, + 0.006757795810699463, + 0.01459814514964819, + 0.007662667427212, + 0.005190813448280096, + 0.012568091042339802, + 0.016562579199671745, + 0.003899955190718174, + 0.006775950081646442, + 0.00578027218580246, + 0.003453670535236597, + 0.005158659070730209, + 0.00413703266531229, + 0.004774312488734722, + 0.007433272898197174, + 0.005037775728851557, + 0.010703742504119873, + 0.05928250029683113, + 0.07128935307264328, + 0.03573426231741905, + 0.0779351145029068, + 0.059879668056964874, + 0.10636213421821594, + 0.10493088513612747, + 0.059214793145656586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 1 + ], + "selected_reward": -0.012536530382931232 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle close-up captures the determined ascent of a tiny, furry orange tabby, its curious green eyes wide as it cautiously places its tiny paws on the face of a larger, more complex cat. Designed to resemble a rugged mountain, the structure is covered with beige shisha ropes of various textures and a patch of brown carpet, making the kitten's ascent seem like a grand expedition. As the kitten pauses, its tail twitching slightly to maintain balance, the camera executes a slow, smooth arc to the right, gradually expanding the frame. This movement reveals the scratchboard's true setting in a spacious, sunny living room. Bright, warm sunlight pours in through the floor-to-ceiling windows, illuminating the dust motes dancing on the beams and casting soft, elongated shadows across the polished hardwood floor. The background unfolds to reveal a chic, minimalist space with a comfortable gray sofa, an understated wooden coffee table, and lush potted green plants, creating a tranquil and welcoming atmosphere, contrasting the kitten's miniature adventure with the vast, peaceful world of home it inhabits.", + "seed": 32325, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_12/index_1720_seed_32325_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.024929341580718755, + "vlm_logprob": -0.024929341580718755, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.024929341580718755, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, furry orange tabby kitten with wide green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kitten climbing the face of a larger cat that is designed to resemble a rugged mountain covered in beige shisha ropes and brown carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the kitten pause with its tail twitching slightly to maintain balance during the ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, smooth arc to the right that gradually expands the frame to reveal a spacious, sunny living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist living room featuring a gray sofa, a wooden coffee table, potted green plants, and bright sunlight illuminating dust motes and hardwood floors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07072257995605469 + ], + "opposite_logprob_tokens": [ + -2.6957225799560547 + ], + "logprob_reward": -0.07072257995605469, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07072257995605469, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2050984799861908, + 0.12459735572338104, + 0.13307592272758484, + 0.14674212038516998, + 0.20086728036403656, + 0.019691623747348785, + 0.021247100085020065, + 0.03200249373912811, + 0.04453373700380325, + 0.07214383035898209 + ], + "credit_map": [ + 0.027719618752598763, + 0.036126118153333664, + 0.03279531002044678, + 0.018350357189774513, + 0.025826888158917427, + 0.015566718764603138, + 0.01982743665575981, + 0.028886038810014725, + 0.013735332526266575, + 0.015188504941761494, + 0.012680516578257084, + 0.009607378393411636, + 0.022623663768172264, + 0.015779215842485428, + 0.01677793264389038, + 0.0182048212736845, + 0.029116379097104073, + 0.010916938073933125, + 0.017962878569960594, + 0.01848686672747135, + 0.01045202650129795, + 0.008062705397605896, + 0.013890497386455536, + 0.024187639355659485, + 0.01469662506133318, + 0.03720906376838684, + 0.026356175541877747, + 0.01125143188983202, + 0.014694267883896828, + 0.009661227464675903, + 0.017790304496884346, + 0.015083033591508865, + 0.05449088290333748, + 0.061630070209503174, + 0.03101087547838688, + 0.009526672773063183, + 0.011040902696549892, + 0.007400117348879576, + 0.017031116411089897, + 0.008736670017242432, + 0.003942654933780432, + 0.001821398618631065, + 0.0036298509221524, + 0.0018251098226755857, + 0.001878349925391376, + 0.0022729558404535055, + 0.0019279244588688016, + 0.0023933814372867346, + 0.0034017323050647974, + 0.0032425345852971077, + 0.002254222519695759, + 0.0024319051299244165, + 0.0028043317142874002, + 0.0024760819505900145, + 0.002543680602684617, + 0.002092614769935608, + 0.003734920872375369, + 0.003240534570068121, + 0.003665327560156584, + 0.0030267194379121065, + 0.003660377347841859, + 0.004468460567295551, + 0.004392481409013271, + 0.005813676863908768, + 0.004744127858430147, + 0.002788847079500556, + 0.010073008947074413, + 0.0038418893236666918, + 0.002923273481428623, + 0.002084923442453146, + 0.00634766323491931, + 0.011730008758604527, + 0.018151920288801193, + 0.007555729243904352, + 0.008702774532139301, + 0.006506586913019419, + 0.005550738424062729, + 0.004433967173099518, + 0.015528964810073376, + 0.005713155027478933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.07072257995605469 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, furry orange tabby kitten with wide green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kitten climbing the face of a larger cat that is designed to resemble a rugged mountain covered in beige shisha ropes and brown carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the kitten pause with its tail twitching slightly to maintain balance during the ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, smooth arc to the right that gradually expands the frame to reveal a spacious, sunny living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist living room featuring a gray sofa, a wooden coffee table, potted green plants, and bright sunlight illuminating dust motes and hardwood floors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012552070431411266 + ], + "opposite_logprob_tokens": [ + -4.387552261352539 + ], + "logprob_reward": -0.012552070431411266, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012552070431411266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11080863326787949, + 0.09090229868888855, + 0.06384319812059402, + 0.20712608098983765, + 0.09291234612464905, + 0.03506843000650406, + 0.04767404869198799, + 0.034219831228256226, + 0.22859002649784088, + 0.08885504305362701 + ], + "credit_map": [ + 0.02367161400616169, + 0.017926782369613647, + 0.018859509378671646, + 0.008019521832466125, + 0.010240557603538036, + 0.007300373632460833, + 0.00907889660447836, + 0.01571139320731163, + 0.009073853492736816, + 0.011254983022809029, + 0.011983354575932026, + 0.01228009257465601, + 0.011546841822564602, + 0.008725241757929325, + 0.014156944118440151, + 0.011880999431014061, + 0.013763920404016972, + 0.006199448369443417, + 0.006748800631612539, + 0.008577424101531506, + 0.005327756050974131, + 0.00383621989749372, + 0.0068609281443059444, + 0.012528708204627037, + 0.05121422931551933, + 0.08611530810594559, + 0.03213190287351608, + 0.005553238559514284, + 0.009193502366542816, + 0.005721746478229761, + 0.007723384536802769, + 0.009472791105508804, + 0.011257429607212543, + 0.009527149610221386, + 0.02336299605667591, + 0.006724256556481123, + 0.007124839350581169, + 0.005309763830155134, + 0.02198571339249611, + 0.007620206102728844, + 0.006378563586622477, + 0.0030572533141821623, + 0.0044282907620072365, + 0.003863438731059432, + 0.0032779062166810036, + 0.0028706705197691917, + 0.00669432058930397, + 0.004497992340475321, + 0.006841565947979689, + 0.0072568804025650024, + 0.0037704568821936846, + 0.005206929985433817, + 0.00665696244686842, + 0.0047547658905386925, + 0.006446178071200848, + 0.00674031674861908, + 0.0048087346367537975, + 0.004188592545688152, + 0.004832007922232151, + 0.0037593445740640163, + 0.004266422241926193, + 0.004372528288513422, + 0.004249978810548782, + 0.0037422263994812965, + 0.05900048837065697, + 0.01258399523794651, + 0.0561056062579155, + 0.018977615982294083, + 0.010612263344228268, + 0.009252794086933136, + 0.027995554730296135, + 0.0340617373585701, + 0.026110010221600533, + 0.01025544572621584, + 0.012589085847139359, + 0.008122427389025688, + 0.007501528132706881, + 0.004471273627132177, + 0.012836531735956669, + 0.006968757137656212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.012552070431411266 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, furry orange tabby kitten with wide green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kitten climbing the face of a larger cat that is designed to resemble a rugged mountain covered in beige shisha ropes and brown carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the kitten pause with its tail twitching slightly to maintain balance during the ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, smooth arc to the right that gradually expands the frame to reveal a spacious, sunny living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist living room featuring a gray sofa, a wooden coffee table, potted green plants, and bright sunlight illuminating dust motes and hardwood floors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634584717452526 + ], + "opposite_logprob_tokens": [ + -3.651345729827881 + ], + "logprob_reward": -0.02634584717452526, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634584717452526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15479743480682373, + 0.16737058758735657, + 0.0796009823679924, + 0.10776570439338684, + 0.09775751084089279, + 0.023348210379481316, + 0.03482173755764961, + 0.037354398518800735, + 0.1939748078584671, + 0.10320866852998734 + ], + "credit_map": [ + 0.018632257357239723, + 0.022031540051102638, + 0.029206329956650734, + 0.015296096913516521, + 0.016485420987010002, + 0.019212905317544937, + 0.014029488898813725, + 0.01990339532494545, + 0.02156292274594307, + 0.03229453042149544, + 0.018933380022644997, + 0.014385930262506008, + 0.025065530091524124, + 0.017655212432146072, + 0.017507873475551605, + 0.01996520347893238, + 0.020725982263684273, + 0.009099019691348076, + 0.006315488368272781, + 0.0051175681874156, + 0.007651752792298794, + 0.005352865904569626, + 0.005801434628665447, + 0.01953687146306038, + 0.011923152953386307, + 0.022534247487783432, + 0.012992233969271183, + 0.010339864529669285, + 0.012571646831929684, + 0.008782419376075268, + 0.01522863656282425, + 0.013393500819802284, + 0.01048735249787569, + 0.01360315177589655, + 0.011979199014604092, + 0.007266517262905836, + 0.01146379578858614, + 0.010618307627737522, + 0.02046840265393257, + 0.011870780028402805, + 0.003774587530642748, + 0.0032052372116595507, + 0.003028194885700941, + 0.0025281161069869995, + 0.0029419453348964453, + 0.0028150673024356365, + 0.0027410313487052917, + 0.0023140308912843466, + 0.004864871501922607, + 0.004177585709840059, + 0.004704155493527651, + 0.004542303737252951, + 0.0045667910017073154, + 0.003937994595617056, + 0.004061102867126465, + 0.003966931719332933, + 0.00413403520360589, + 0.004729045554995537, + 0.0047667743638157845, + 0.004434769973158836, + 0.004046709276735783, + 0.004012049175798893, + 0.005976506974548101, + 0.005254508927464485, + 0.04435603693127632, + 0.009565925225615501, + 0.07024239003658295, + 0.021468132734298706, + 0.008265754207968712, + 0.006354878190904856, + 0.024712756276130676, + 0.009008937515318394, + 0.0173978079110384, + 0.011444618925452232, + 0.008588938042521477, + 0.009261805564165115, + 0.018012765794992447, + 0.00823418702930212, + 0.02294309251010418, + 0.00732545368373394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.02634584717452526 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, furry orange tabby kitten with wide green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kitten climbing the face of a larger cat that is designed to resemble a rugged mountain covered in beige shisha ropes and brown carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the kitten pause with its tail twitching slightly to maintain balance during the ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, smooth arc to the right that gradually expands the frame to reveal a spacious, sunny living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist living room featuring a gray sofa, a wooden coffee table, potted green plants, and bright sunlight illuminating dust motes and hardwood floors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005249288398772478 + ], + "opposite_logprob_tokens": [ + -5.255249500274658 + ], + "logprob_reward": -0.005249288398772478, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005249288398772478, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08536576479673386, + 0.04492639750242233, + 0.09460300952196121, + 0.08662167936563492, + 0.09878840297460556, + 0.03554471209645271, + 0.07729469239711761, + 0.04314236342906952, + 0.35477587580680847, + 0.0789370909333229 + ], + "credit_map": [ + 0.01240584161132574, + 0.017665259540081024, + 0.019495604559779167, + 0.0062753427773714066, + 0.007616326212882996, + 0.006694614887237549, + 0.005287182983011007, + 0.009925593622028828, + 0.00609245803207159, + 0.004691306035965681, + 0.005240836646407843, + 0.0034503424540162086, + 0.006805654615163803, + 0.004678473342210054, + 0.007564574480056763, + 0.006402753293514252, + 0.018247265368700027, + 0.00905078835785389, + 0.0128088453784585, + 0.013813814148306847, + 0.007280545774847269, + 0.005608950275927782, + 0.007282687351107597, + 0.020510118454694748, + 0.025115547701716423, + 0.03366558253765106, + 0.007210139650851488, + 0.004454213660210371, + 0.00444177258759737, + 0.002959250472486019, + 0.0038893786258995533, + 0.004885807633399963, + 0.03383021801710129, + 0.02849615551531315, + 0.008135540410876274, + 0.0028890904504805803, + 0.004927958361804485, + 0.003648380981758237, + 0.01208446454256773, + 0.004776597488671541, + 0.00565726961940527, + 0.0036661624908447266, + 0.006047582253813744, + 0.004628208000212908, + 0.002949911868199706, + 0.0036917529068887234, + 0.005331811029464006, + 0.0035720141604542732, + 0.010193193331360817, + 0.012992029078304768, + 0.008449348621070385, + 0.009428928606212139, + 0.007406868040561676, + 0.011520477943122387, + 0.009386918507516384, + 0.007916929200291634, + 0.005817161872982979, + 0.004808942787349224, + 0.006635183934122324, + 0.0048687271773815155, + 0.005071074236184359, + 0.0053190868347883224, + 0.005338737275451422, + 0.005283454433083534, + 0.08465811610221863, + 0.017293130978941917, + 0.13142068684101105, + 0.039370693266391754, + 0.010953490622341633, + 0.007888306863605976, + 0.035463523119688034, + 0.027727937325835228, + 0.012173223309218884, + 0.009604367427527905, + 0.010130086913704872, + 0.00766198243945837, + 0.013767631724476814, + 0.006386630702763796, + 0.01200213935226202, + 0.007211028598248959 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 2 + ], + "selected_reward": -0.005249288398772478 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, furry orange tabby kitten with wide green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kitten climbing the face of a larger cat that is designed to resemble a rugged mountain covered in beige shisha ropes and brown carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the kitten pause with its tail twitching slightly to maintain balance during the ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, smooth arc to the right that gradually expands the frame to reveal a spacious, sunny living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist living room featuring a gray sofa, a wooden coffee table, potted green plants, and bright sunlight illuminating dust motes and hardwood floors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009776921942830086 + ], + "opposite_logprob_tokens": [ + -4.634777069091797 + ], + "logprob_reward": -0.009776921942830086, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009776921942830086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19281388819217682, + 0.12103911489248276, + 0.07750964909791946, + 0.15198099613189697, + 0.08806218951940536, + 0.06050308421254158, + 0.08771134167909622, + 0.04427442327141762, + 0.07498998939990997, + 0.10111535340547562 + ], + "credit_map": [ + 0.02549932524561882, + 0.029625261202454567, + 0.03145191818475723, + 0.016475170850753784, + 0.021210094913840294, + 0.022548126056790352, + 0.019448334351181984, + 0.026555655524134636, + 0.014067362993955612, + 0.011509331874549389, + 0.013888312503695488, + 0.008190722204744816, + 0.029492899775505066, + 0.01152314804494381, + 0.014079184271395206, + 0.018288152292370796, + 0.02456529624760151, + 0.0070983171463012695, + 0.005044781137257814, + 0.004277580417692661, + 0.005408802069723606, + 0.0035126740112900734, + 0.0048765260726213455, + 0.0227256678044796, + 0.031025422737002373, + 0.057435043156147, + 0.01962287165224552, + 0.006332010962069035, + 0.009941394440829754, + 0.008774636313319206, + 0.009353301487863064, + 0.009496312588453293, + 0.009030451066792011, + 0.013534430414438248, + 0.016308730468153954, + 0.006375852972269058, + 0.0070700678043067455, + 0.005978129804134369, + 0.021535256877541542, + 0.00822926964610815, + 0.012267369776964188, + 0.004598543979227543, + 0.01394782867282629, + 0.0050208792090415955, + 0.00589615385979414, + 0.006792263127863407, + 0.006260584108531475, + 0.005719454493373632, + 0.016234874725341797, + 0.016681216657161713, + 0.009414558298885822, + 0.009056582115590572, + 0.009185773320496082, + 0.007726849988102913, + 0.012155892327427864, + 0.007255596574395895, + 0.006032636389136314, + 0.006802864372730255, + 0.007943636737763882, + 0.0057596866972744465, + 0.0036613065749406815, + 0.004999763797968626, + 0.005292740650475025, + 0.0037817885167896748, + 0.010440127924084663, + 0.0047155325300991535, + 0.02048547752201557, + 0.007152701262384653, + 0.005157252307981253, + 0.004736206028610468, + 0.007326492574065924, + 0.014976197853684425, + 0.01659107767045498, + 0.015823453664779663, + 0.01473207212984562, + 0.01075948029756546, + 0.0093406792730093, + 0.006184084806591272, + 0.016628755256533623, + 0.011055756360292435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.009776921942830086 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle close-up captures the determined ascent of a tiny, furry orange tabby, its curious green eyes wide as it cautiously places its tiny paws on the face of a larger, more complex cat. Designed to resemble a rugged mountain, the structure is covered with beige shisha ropes of various textures and a patch of brown carpet, making the kitten's ascent seem like a grand expedition. As the kitten pauses, its tail twitching slightly to maintain balance, the camera executes a slow, smooth arc to the right, gradually expanding the frame. This movement reveals the scratchboard's true setting in a spacious, sunny living room. Bright, warm sunlight pours in through the floor-to-ceiling windows, illuminating the dust motes dancing on the beams and casting soft, elongated shadows across the polished hardwood floor. The background unfolds to reveal a chic, minimalist space with a comfortable gray sofa, an understated wooden coffee table, and lush potted green plants, creating a tranquil and welcoming atmosphere, contrasting the kitten's miniature adventure with the vast, peaceful world of home it inhabits.", + "seed": 32326, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_12/index_1720_seed_32326_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02918590260669589, + "vlm_logprob": -0.02918590260669589, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.02918590260669589, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, furry orange tabby kitten with wide green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kitten climbing the face of a larger cat that is designed to resemble a rugged mountain covered in beige shisha ropes and brown carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the kitten pause with its tail twitching slightly to maintain balance during the ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, smooth arc to the right that gradually expands the frame to reveal a spacious, sunny living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist living room featuring a gray sofa, a wooden coffee table, potted green plants, and bright sunlight illuminating dust motes and hardwood floors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.055366963148117065 + ], + "opposite_logprob_tokens": [ + -3.0553669929504395 + ], + "logprob_reward": -0.055366963148117065, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.055366963148117065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2818755507469177, + 0.04186137765645981, + 0.05076980218291283, + 0.08028145134449005, + 0.05368315428495407, + 0.23678548634052277, + 0.04912205785512924, + 0.1360371857881546, + 0.044134095311164856, + 0.025449901819229126 + ], + "credit_map": [ + 0.040566083043813705, + 0.029704110696911812, + 0.05267344415187836, + 0.01863170973956585, + 0.031379010528326035, + 0.06270387768745422, + 0.01874649152159691, + 0.027470793575048447, + 0.00479228887706995, + 0.006978979334235191, + 0.005958130583167076, + 0.004254640080034733, + 0.004725183360278606, + 0.006422830745577812, + 0.0038149061147123575, + 0.004914416465908289, + 0.007185583468526602, + 0.004035027232021093, + 0.006457265932112932, + 0.015785057097673416, + 0.0035485578700900078, + 0.0026899969670921564, + 0.007092426531016827, + 0.003975881263613701, + 0.008908644318580627, + 0.01737811602652073, + 0.010647905990481377, + 0.010342945344746113, + 0.01274706982076168, + 0.007703192997723818, + 0.006184806581586599, + 0.00636876467615366, + 0.006695938762277365, + 0.006095248740166426, + 0.009242182597517967, + 0.004795589484274387, + 0.005018637049943209, + 0.010392831638455391, + 0.0039968243800103664, + 0.007445896975696087, + 0.04292133077979088, + 0.03280409052968025, + 0.04105113074183464, + 0.024072730913758278, + 0.018529532477259636, + 0.018164748325943947, + 0.03552566468715668, + 0.023716235533356667, + 0.008850359357893467, + 0.006022999063134193, + 0.005108469165861607, + 0.007919970899820328, + 0.0048149023205041885, + 0.0032644090242683887, + 0.007243704982101917, + 0.005897235590964556, + 0.01634480245411396, + 0.01676708273589611, + 0.01948326639831066, + 0.017823223024606705, + 0.015025452710688114, + 0.019116755574941635, + 0.016461454331874847, + 0.015015139244496822, + 0.008300742134451866, + 0.005250285845249891, + 0.006878065876662731, + 0.0068848212249577045, + 0.003477109596133232, + 0.0028208435978740454, + 0.004721387755125761, + 0.0058008404448628426, + 0.001931273378431797, + 0.0021351282484829426, + 0.0028023547492921352, + 0.004230692982673645, + 0.00471659516915679, + 0.0026790339034050703, + 0.0038263334427028894, + 0.0031284880824387074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.055366963148117065 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, furry orange tabby kitten with wide green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kitten climbing the face of a larger cat that is designed to resemble a rugged mountain covered in beige shisha ropes and brown carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the kitten pause with its tail twitching slightly to maintain balance during the ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, smooth arc to the right that gradually expands the frame to reveal a spacious, sunny living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist living room featuring a gray sofa, a wooden coffee table, potted green plants, and bright sunlight illuminating dust motes and hardwood floors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023325342684984207 + ], + "opposite_logprob_tokens": [ + -3.7733254432678223 + ], + "logprob_reward": -0.023325342684984207, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023325342684984207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1301247775554657, + 0.04104239493608475, + 0.1563672423362732, + 0.0487697497010231, + 0.2124536782503128, + 0.10388562083244324, + 0.06273924559354782, + 0.12664714455604553, + 0.07302428036928177, + 0.04494595155119896 + ], + "credit_map": [ + 0.013918973505496979, + 0.015548733994364738, + 0.01933203637599945, + 0.008537502959370613, + 0.016418123617768288, + 0.02335124835371971, + 0.008857403881847858, + 0.024160750210285187, + 0.004410466179251671, + 0.008261567912995815, + 0.005271829664707184, + 0.004414235707372427, + 0.00402480736374855, + 0.006973244249820709, + 0.0035048960708081722, + 0.004181349650025368, + 0.017354724928736687, + 0.02230673097074032, + 0.02266087383031845, + 0.022307801991701126, + 0.017062323167920113, + 0.012297806330025196, + 0.025488443672657013, + 0.01688854768872261, + 0.005588545463979244, + 0.009114916436374187, + 0.006995008792728186, + 0.006539545487612486, + 0.00800827145576477, + 0.0031752155628055334, + 0.0045592207461595535, + 0.004789022263139486, + 0.04046523943543434, + 0.030586401000618935, + 0.04861684516072273, + 0.01623675599694252, + 0.023471727967262268, + 0.020558981224894524, + 0.01553747896105051, + 0.016980236396193504, + 0.01867680624127388, + 0.020433463156223297, + 0.01171896792948246, + 0.01287931576371193, + 0.0070716398768126965, + 0.007451216224581003, + 0.015492726117372513, + 0.010161480866372585, + 0.008732732385396957, + 0.006503966636955738, + 0.008545673452317715, + 0.009985299780964851, + 0.0074097951874136925, + 0.005107165314257145, + 0.008501182310283184, + 0.007953427731990814, + 0.01981472037732601, + 0.016355687752366066, + 0.015499728731811047, + 0.015348580665886402, + 0.012083162553608418, + 0.017718631774187088, + 0.016030211001634598, + 0.013796413317322731, + 0.012391213327646255, + 0.006817835848778486, + 0.013409863226115704, + 0.010831945575773716, + 0.005379101261496544, + 0.00477246381342411, + 0.009075623005628586, + 0.010346232913434505, + 0.003929217346012592, + 0.004483905620872974, + 0.005404610652476549, + 0.01005595177412033, + 0.0040772478096187115, + 0.005681673064827919, + 0.005834552925080061, + 0.005478793289512396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.023325342684984207 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, furry orange tabby kitten with wide green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kitten climbing the face of a larger cat that is designed to resemble a rugged mountain covered in beige shisha ropes and brown carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the kitten pause with its tail twitching slightly to maintain balance during the ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, smooth arc to the right that gradually expands the frame to reveal a spacious, sunny living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist living room featuring a gray sofa, a wooden coffee table, potted green plants, and bright sunlight illuminating dust motes and hardwood floors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029810817912220955 + ], + "opposite_logprob_tokens": [ + -3.529810905456543 + ], + "logprob_reward": -0.029810817912220955, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029810817912220955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25957080721855164, + 0.04676405340433121, + 0.11471915990114212, + 0.044552549719810486, + 0.08217645436525345, + 0.14354775846004486, + 0.06580698490142822, + 0.1508854329586029, + 0.05095601826906204, + 0.041020773351192474 + ], + "credit_map": [ + 0.025085629895329475, + 0.04049199819564819, + 0.028956275433301926, + 0.0186117272824049, + 0.04207184910774231, + 0.045981552451848984, + 0.02030155248939991, + 0.038070209324359894, + 0.004285749979317188, + 0.007721308618783951, + 0.00558166578412056, + 0.006080070976167917, + 0.00575270177796483, + 0.006099173799157143, + 0.0040262057445943356, + 0.007217176724225283, + 0.01315552182495594, + 0.01724329963326454, + 0.017959361895918846, + 0.01806863583624363, + 0.014912637881934643, + 0.010361488908529282, + 0.013635647483170033, + 0.00938256923109293, + 0.005235371645539999, + 0.006265558302402496, + 0.005318957380950451, + 0.0076033612713217735, + 0.004533801693469286, + 0.004970113281160593, + 0.006836671847850084, + 0.0037887159269303083, + 0.00940989051014185, + 0.012427120469510555, + 0.01566743105649948, + 0.009417125023901463, + 0.00875356886535883, + 0.007921701297163963, + 0.009641045704483986, + 0.008938571438193321, + 0.03196113929152489, + 0.023065632209181786, + 0.019286705181002617, + 0.013598749414086342, + 0.011652447283267975, + 0.010992750525474548, + 0.017653584480285645, + 0.015336748212575912, + 0.008631263859570026, + 0.006022259593009949, + 0.007662299554795027, + 0.01131535042077303, + 0.008946609683334827, + 0.005367065314203501, + 0.010428081266582012, + 0.007434057537466288, + 0.019700899720191956, + 0.019604912027716637, + 0.0216445904225111, + 0.015216005966067314, + 0.01561914011836052, + 0.022533215582370758, + 0.019641639664769173, + 0.016925031319260597, + 0.007904012687504292, + 0.009245835244655609, + 0.0071887909434735775, + 0.006175497081130743, + 0.005519067402929068, + 0.004820335656404495, + 0.005287436302751303, + 0.004815043415874243, + 0.003598140785470605, + 0.0032023137900978327, + 0.004864145070314407, + 0.006830580998212099, + 0.0074919150210917, + 0.0043089985847473145, + 0.006032360717654228, + 0.004692319314926863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.029810817912220955 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, furry orange tabby kitten with wide green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kitten climbing the face of a larger cat that is designed to resemble a rugged mountain covered in beige shisha ropes and brown carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the kitten pause with its tail twitching slightly to maintain balance during the ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, smooth arc to the right that gradually expands the frame to reveal a spacious, sunny living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist living room featuring a gray sofa, a wooden coffee table, potted green plants, and bright sunlight illuminating dust motes and hardwood floors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630960550159216 + ], + "opposite_logprob_tokens": [ + -4.882630825042725 + ], + "logprob_reward": -0.007630960550159216, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630960550159216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12995891273021698, + 0.0581740103662014, + 0.19499604403972626, + 0.05753786861896515, + 0.10295019298791885, + 0.08412054181098938, + 0.09218977391719818, + 0.11966505646705627, + 0.09527608752250671, + 0.0651315227150917 + ], + "credit_map": [ + 0.011836109682917595, + 0.012981681153178215, + 0.03456110879778862, + 0.006456427741795778, + 0.008836647495627403, + 0.011709308251738548, + 0.007178564555943012, + 0.03639906272292137, + 0.006155502516776323, + 0.0062067448161542416, + 0.007212736178189516, + 0.005638019181787968, + 0.004722973797470331, + 0.011383017525076866, + 0.005610923748463392, + 0.011244087480008602, + 0.02434142492711544, + 0.02456602267920971, + 0.03488156199455261, + 0.03606631979346275, + 0.02782553993165493, + 0.015786824747920036, + 0.01851847395300865, + 0.013009858317673206, + 0.006864771246910095, + 0.004651590250432491, + 0.00821837317198515, + 0.006016097497195005, + 0.004016044083982706, + 0.008943676017224789, + 0.012021861039102077, + 0.006805450655519962, + 0.009101074188947678, + 0.017350459471344948, + 0.01289413496851921, + 0.017764024436473846, + 0.013194565661251545, + 0.008138888515532017, + 0.013838281854987144, + 0.01066875271499157, + 0.009694629348814487, + 0.017488298937678337, + 0.008628112263977528, + 0.012433544732630253, + 0.006836018059402704, + 0.00694883381947875, + 0.015493307262659073, + 0.0065977913327515125, + 0.011802553199231625, + 0.00998731143772602, + 0.009633511304855347, + 0.017476774752140045, + 0.010518110357224941, + 0.007956422865390778, + 0.013738966546952724, + 0.011076115071773529, + 0.0161594208329916, + 0.018059665337204933, + 0.014040601439774036, + 0.01447417214512825, + 0.014839094132184982, + 0.018938371911644936, + 0.009401977993547916, + 0.013751747086644173, + 0.014314769767224789, + 0.010718372650444508, + 0.013971727341413498, + 0.011079441756010056, + 0.013013403862714767, + 0.008219162933528423, + 0.010722867213189602, + 0.01323633175343275, + 0.004707734566181898, + 0.004902354907244444, + 0.007704799994826317, + 0.01746920868754387, + 0.007643857039511204, + 0.0059781670570373535, + 0.008499393239617348, + 0.008225993253290653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.007630960550159216 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, furry orange tabby kitten with wide green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kitten climbing the face of a larger cat that is designed to resemble a rugged mountain covered in beige shisha ropes and brown carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the kitten pause with its tail twitching slightly to maintain balance during the ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, smooth arc to the right that gradually expands the frame to reveal a spacious, sunny living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist living room featuring a gray sofa, a wooden coffee table, potted green plants, and bright sunlight illuminating dust motes and hardwood floors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979542873799801 + ], + "opposite_logprob_tokens": [ + -3.5297954082489014 + ], + "logprob_reward": -0.02979542873799801, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979542873799801, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1400550752878189, + 0.044987715780735016, + 0.08303934335708618, + 0.03642956539988518, + 0.07731395214796066, + 0.03805605322122574, + 0.10982572287321091, + 0.12762810289859772, + 0.18407705426216125, + 0.15858744084835052 + ], + "credit_map": [ + 0.019117427989840508, + 0.018058914691209793, + 0.030709456652402878, + 0.006688325200229883, + 0.01248831395059824, + 0.02073591761291027, + 0.007605085149407387, + 0.024651633575558662, + 0.005289021879434586, + 0.004399995319545269, + 0.006593589670956135, + 0.004838858265429735, + 0.0043961708433926105, + 0.008630428463220596, + 0.004844848066568375, + 0.00599480327218771, + 0.009777512401342392, + 0.011224452406167984, + 0.010114988312125206, + 0.01829579286277294, + 0.00781304482370615, + 0.005784925073385239, + 0.01226852834224701, + 0.0077600907534360886, + 0.003942749463021755, + 0.007368003949522972, + 0.004005612339824438, + 0.004088942892849445, + 0.005697917193174362, + 0.004125913605093956, + 0.003864510217681527, + 0.0033359178341925144, + 0.008690825663506985, + 0.010336139239370823, + 0.014561144635081291, + 0.01052207313477993, + 0.0097171850502491, + 0.006050484254956245, + 0.008114613592624664, + 0.009321488440036774, + 0.005189318209886551, + 0.004301605746150017, + 0.0037609138526022434, + 0.0038526186253875494, + 0.004661668557673693, + 0.005366610363125801, + 0.004089116584509611, + 0.0068341996520757675, + 0.01784130185842514, + 0.0118608009070158, + 0.01094727125018835, + 0.01791173405945301, + 0.013393505476415157, + 0.009396667592227459, + 0.015411403961479664, + 0.013063041493296623, + 0.026652036234736443, + 0.013796710409224033, + 0.011041625402867794, + 0.008837168104946613, + 0.01746196299791336, + 0.021945461630821228, + 0.0134397242218256, + 0.014453405514359474, + 0.01722743548452854, + 0.020240237936377525, + 0.02874492108821869, + 0.017947452142834663, + 0.04736221954226494, + 0.01742997020483017, + 0.016723711043596268, + 0.018401101231575012, + 0.012299979105591774, + 0.012083254754543304, + 0.01998850703239441, + 0.031024765223264694, + 0.03181749954819679, + 0.01290186494588852, + 0.018835505470633507, + 0.019636064767837524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.02979542873799801 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle close-up captures the determined ascent of a tiny, furry orange tabby, its curious green eyes wide as it cautiously places its tiny paws on the face of a larger, more complex cat. Designed to resemble a rugged mountain, the structure is covered with beige shisha ropes of various textures and a patch of brown carpet, making the kitten's ascent seem like a grand expedition. As the kitten pauses, its tail twitching slightly to maintain balance, the camera executes a slow, smooth arc to the right, gradually expanding the frame. This movement reveals the scratchboard's true setting in a spacious, sunny living room. Bright, warm sunlight pours in through the floor-to-ceiling windows, illuminating the dust motes dancing on the beams and casting soft, elongated shadows across the polished hardwood floor. The background unfolds to reveal a chic, minimalist space with a comfortable gray sofa, an understated wooden coffee table, and lush potted green plants, creating a tranquil and welcoming atmosphere, contrasting the kitten's miniature adventure with the vast, peaceful world of home it inhabits.", + "seed": 32327, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_12/index_1720_seed_32327_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.015815768018364908, + "vlm_logprob": -0.015815768018364908, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.015815768018364908, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, furry orange tabby kitten with wide green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kitten climbing the face of a larger cat that is designed to resemble a rugged mountain covered in beige shisha ropes and brown carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the kitten pause with its tail twitching slightly to maintain balance during the ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, smooth arc to the right that gradually expands the frame to reveal a spacious, sunny living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist living room featuring a gray sofa, a wooden coffee table, potted green plants, and bright sunlight illuminating dust motes and hardwood floors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02013481967151165 + ], + "opposite_logprob_tokens": [ + -4.020134925842285 + ], + "logprob_reward": -0.02013481967151165, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02013481967151165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11937917023897171, + 0.1233150064945221, + 0.2106572538614273, + 0.05067899078130722, + 0.03492611646652222, + 0.28886598348617554, + 0.06225765496492386, + 0.03896882385015488, + 0.029958056285977364, + 0.040992893278598785 + ], + "credit_map": [ + 0.018503565341234207, + 0.009088773280382156, + 0.01564105600118637, + 0.01978326588869095, + 0.01169714517891407, + 0.01363565493375063, + 0.015056855045258999, + 0.015972865745425224, + 0.01602509804069996, + 0.026470445096492767, + 0.011299464851617813, + 0.012073752470314503, + 0.016787273809313774, + 0.01837502419948578, + 0.012443525716662407, + 0.009840433485805988, + 0.021692318841814995, + 0.02179255709052086, + 0.03668106347322464, + 0.02248331531882286, + 0.035021014511585236, + 0.03106055036187172, + 0.018633641302585602, + 0.023292822763323784, + 0.007366322446614504, + 0.006307318806648254, + 0.004577595740556717, + 0.005060573574155569, + 0.0072659398429095745, + 0.0058213165029883385, + 0.007848454639315605, + 0.006431471556425095, + 0.004197494592517614, + 0.0034651444293558598, + 0.003699640044942498, + 0.005427992437034845, + 0.006988126784563065, + 0.0034851296804845333, + 0.004291225224733353, + 0.0033713674638420343, + 0.021347349509596825, + 0.026592858135700226, + 0.018610823899507523, + 0.029985416680574417, + 0.03856759890913963, + 0.08062849938869476, + 0.030702292919158936, + 0.04243115335702896, + 0.00510542094707489, + 0.007341603748500347, + 0.0032962874975055456, + 0.007943115197122097, + 0.013174165040254593, + 0.010668070986866951, + 0.007826381362974644, + 0.0069026099517941475, + 0.0073272548615932465, + 0.004589464049786329, + 0.00402239290997386, + 0.002948312321677804, + 0.003939629998058081, + 0.004926348105072975, + 0.004452231340110302, + 0.0067631942220032215, + 0.003916936926543713, + 0.002777090761810541, + 0.002526872791349888, + 0.004552926868200302, + 0.005463018082082272, + 0.0025518182665109634, + 0.00397625332698226, + 0.004193141590803862, + 0.005132194142788649, + 0.0042235408909618855, + 0.00448257802054286, + 0.005969013553112745, + 0.00396464578807354, + 0.004428654909133911, + 0.006914668716490269, + 0.005877599120140076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.02013481967151165 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, furry orange tabby kitten with wide green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kitten climbing the face of a larger cat that is designed to resemble a rugged mountain covered in beige shisha ropes and brown carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the kitten pause with its tail twitching slightly to maintain balance during the ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, smooth arc to the right that gradually expands the frame to reveal a spacious, sunny living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist living room featuring a gray sofa, a wooden coffee table, potted green plants, and bright sunlight illuminating dust motes and hardwood floors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060309424996376 + ], + "opposite_logprob_tokens": [ + -3.8956031799316406 + ], + "logprob_reward": -0.02060309424996376, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060309424996376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18003277480602264, + 0.15889431536197662, + 0.06724023818969727, + 0.08224153518676758, + 0.0866115540266037, + 0.14777493476867676, + 0.04343843460083008, + 0.08675628155469894, + 0.08371735364198685, + 0.06329259276390076 + ], + "credit_map": [ + 0.026971913874149323, + 0.022643061354756355, + 0.023500783368945122, + 0.023607371374964714, + 0.02444509230554104, + 0.023569051176309586, + 0.021779198199510574, + 0.013516291044652462, + 0.028183404356241226, + 0.02354702353477478, + 0.01603221520781517, + 0.018001515418291092, + 0.019245600327849388, + 0.02171185426414013, + 0.01882966421544552, + 0.013343020342290401, + 0.007007039617747068, + 0.009468588046729565, + 0.009904558770358562, + 0.006447880063205957, + 0.012871567159891129, + 0.00930926576256752, + 0.005953557789325714, + 0.006277774926275015, + 0.009432769380509853, + 0.011211546137928963, + 0.007075792644172907, + 0.008636676706373692, + 0.016776643693447113, + 0.00968015193939209, + 0.013224327005445957, + 0.006203618366271257, + 0.008925815112888813, + 0.008913343772292137, + 0.006843249779194593, + 0.01242354977875948, + 0.015162555500864983, + 0.009943629615008831, + 0.016482926905155182, + 0.007916475646197796, + 0.011731898412108421, + 0.012415740638971329, + 0.010899732820689678, + 0.010939320549368858, + 0.019956745207309723, + 0.04463959485292435, + 0.017200319096446037, + 0.01999158039689064, + 0.004229639191180468, + 0.004466752056032419, + 0.0044123511761426926, + 0.004210616461932659, + 0.00581025006249547, + 0.008314243517816067, + 0.0060216342099010944, + 0.005972946062684059, + 0.013258863240480423, + 0.015122995711863041, + 0.004729471635073423, + 0.007862811908125877, + 0.016943523660302162, + 0.009874631650745869, + 0.010366765782237053, + 0.008597202599048615, + 0.009523599408566952, + 0.007539685349911451, + 0.006027617491781712, + 0.012889698147773743, + 0.01676885038614273, + 0.00814198236912489, + 0.010898726992309093, + 0.01192717906087637, + 0.007602257654070854, + 0.005409616976976395, + 0.007316580042243004, + 0.01361177209764719, + 0.0071933395229279995, + 0.005686124786734581, + 0.010092035867273808, + 0.00638086162507534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.02060309424996376 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, furry orange tabby kitten with wide green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kitten climbing the face of a larger cat that is designed to resemble a rugged mountain covered in beige shisha ropes and brown carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the kitten pause with its tail twitching slightly to maintain balance during the ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, smooth arc to the right that gradually expands the frame to reveal a spacious, sunny living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist living room featuring a gray sofa, a wooden coffee table, potted green plants, and bright sunlight illuminating dust motes and hardwood floors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018189996480941772 + ], + "opposite_logprob_tokens": [ + -4.018189907073975 + ], + "logprob_reward": -0.018189996480941772, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018189996480941772, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13363489508628845, + 0.16011086106300354, + 0.14720642566680908, + 0.08381550759077072, + 0.04643594101071358, + 0.22719287872314453, + 0.059802647680044174, + 0.0623357929289341, + 0.04382074996829033, + 0.035644277930259705 + ], + "credit_map": [ + 0.01778784580528736, + 0.009829008020460606, + 0.020791245624423027, + 0.016613351181149483, + 0.01631171815097332, + 0.018693404272198677, + 0.01827825792133808, + 0.015330053865909576, + 0.019172105938196182, + 0.023444045335054398, + 0.016134068369865417, + 0.01533288974314928, + 0.02881881780922413, + 0.028612487018108368, + 0.015774685889482498, + 0.012821758165955544, + 0.017377588897943497, + 0.012974467128515244, + 0.019233224913477898, + 0.016713574528694153, + 0.016970036551356316, + 0.022741790860891342, + 0.01556794811040163, + 0.02562779188156128, + 0.008236807771027088, + 0.007803656160831451, + 0.005988330114632845, + 0.008621708489954472, + 0.011051801964640617, + 0.013672902248799801, + 0.018666133284568787, + 0.009774171747267246, + 0.007616790942847729, + 0.006301079411059618, + 0.00380407995544374, + 0.007029898464679718, + 0.005010564811527729, + 0.005159569438546896, + 0.006788117811083794, + 0.004725837614387274, + 0.017282970249652863, + 0.02223970927298069, + 0.013558244332671165, + 0.02322143316268921, + 0.031860023736953735, + 0.058798499405384064, + 0.02507276087999344, + 0.03515922650694847, + 0.004931436851620674, + 0.006877702195197344, + 0.004477617330849171, + 0.006030200514942408, + 0.01247471570968628, + 0.010332493111491203, + 0.00807403028011322, + 0.0066044507548213005, + 0.011362818069756031, + 0.011050905101001263, + 0.0035973037593066692, + 0.006029150448739529, + 0.010367097333073616, + 0.007075869012624025, + 0.006792039610445499, + 0.00606060354039073, + 0.00569141935557127, + 0.00412873737514019, + 0.004204787779599428, + 0.0066510154865682125, + 0.0059739057905972, + 0.0049133882857859135, + 0.005626052618026733, + 0.0066314456053078175, + 0.004025560338050127, + 0.004068118520081043, + 0.0036429213359951973, + 0.00573026854544878, + 0.004102534148842096, + 0.003655715612694621, + 0.005467066541314125, + 0.004952090792357922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 2 + ], + "selected_reward": -0.018189996480941772 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, furry orange tabby kitten with wide green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kitten climbing the face of a larger cat that is designed to resemble a rugged mountain covered in beige shisha ropes and brown carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the kitten pause with its tail twitching slightly to maintain balance during the ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, smooth arc to the right that gradually expands the frame to reveal a spacious, sunny living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist living room featuring a gray sofa, a wooden coffee table, potted green plants, and bright sunlight illuminating dust motes and hardwood floors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004092533141374588 + ], + "opposite_logprob_tokens": [ + -5.504092693328857 + ], + "logprob_reward": -0.004092533141374588, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004092533141374588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15118439495563507, + 0.06331642717123032, + 0.06249873340129852, + 0.08449248224496841, + 0.06873877346515656, + 0.07706010341644287, + 0.09754839539527893, + 0.06970515102148056, + 0.12281090021133423, + 0.20264461636543274 + ], + "credit_map": [ + 0.015607924200594425, + 0.012138963676989079, + 0.016099754720926285, + 0.022647589445114136, + 0.02028762921690941, + 0.026218468323349953, + 0.016527730971574783, + 0.0216563418507576, + 0.007241150364279747, + 0.010246970690786839, + 0.007682155352085829, + 0.006027382332831621, + 0.008900079876184464, + 0.009836106561124325, + 0.00706855533644557, + 0.006314028054475784, + 0.010780069045722485, + 0.011108756065368652, + 0.01228959858417511, + 0.007199278101325035, + 0.004429756198078394, + 0.005935986526310444, + 0.004591578152030706, + 0.006163710728287697, + 0.0114577142521739, + 0.01871335506439209, + 0.007245361339300871, + 0.006832213141024113, + 0.012342886067926884, + 0.009104151278734207, + 0.00935834925621748, + 0.009438447654247284, + 0.009417872875928879, + 0.006615362595766783, + 0.005895019508898258, + 0.009088072925806046, + 0.010705465450882912, + 0.007161438465118408, + 0.012277581728994846, + 0.007577960845082998, + 0.006882363930344582, + 0.0060700923204422, + 0.004552189260721207, + 0.007373647298663855, + 0.013984550721943378, + 0.01613518036901951, + 0.009078540839254856, + 0.012983540073037148, + 0.005691102240234613, + 0.015709374099969864, + 0.006193908862769604, + 0.010506105609238148, + 0.022948212921619415, + 0.011397496797144413, + 0.010370713658630848, + 0.014731476083397865, + 0.011372881010174751, + 0.009917198680341244, + 0.006074362900108099, + 0.006656798068434, + 0.006184385623782873, + 0.00862659141421318, + 0.0077730645425617695, + 0.013099867850542068, + 0.0205118078738451, + 0.015242119319736958, + 0.00891430489718914, + 0.014985749498009682, + 0.017082681879401207, + 0.012970607727766037, + 0.014739337377250195, + 0.018364304676651955, + 0.03142504394054413, + 0.018989793956279755, + 0.02389279566705227, + 0.03194178268313408, + 0.017760708928108215, + 0.020612463355064392, + 0.03426084294915199, + 0.023761186748743057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.004092533141374588 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, furry orange tabby kitten with wide green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the kitten climbing the face of a larger cat that is designed to resemble a rugged mountain covered in beige shisha ropes and brown carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the kitten pause with its tail twitching slightly to maintain balance during the ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, smooth arc to the right that gradually expands the frame to reveal a spacious, sunny living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist living room featuring a gray sofa, a wooden coffee table, potted green plants, and bright sunlight illuminating dust motes and hardwood floors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01605839654803276 + ], + "opposite_logprob_tokens": [ + -4.141058444976807 + ], + "logprob_reward": -0.01605839654803276, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01605839654803276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15013012290000916, + 0.06971357017755508, + 0.1143454909324646, + 0.10521208494901657, + 0.06402050703763962, + 0.1246321052312851, + 0.07823598384857178, + 0.06000731512904167, + 0.1395016610622406, + 0.09420116245746613 + ], + "credit_map": [ + 0.02295437641441822, + 0.02020186558365822, + 0.012557984329760075, + 0.019768262282013893, + 0.015231121331453323, + 0.020430590957403183, + 0.016471628099679947, + 0.022514287382364273, + 0.008652525022625923, + 0.010065726935863495, + 0.006881135981529951, + 0.0075548673048615456, + 0.009738470427691936, + 0.012493020854890347, + 0.007213904522359371, + 0.0071139223873615265, + 0.01838347315788269, + 0.013198514468967915, + 0.015564106404781342, + 0.011153783649206161, + 0.015352538786828518, + 0.020152036100625992, + 0.009758568368852139, + 0.010782471857964993, + 0.014210246503353119, + 0.016346698626875877, + 0.009634352289140224, + 0.009287065826356411, + 0.018465163186192513, + 0.011320197023451328, + 0.01269499771296978, + 0.013253367505967617, + 0.008857312612235546, + 0.005696374457329512, + 0.005910062231123447, + 0.007935944013297558, + 0.011169868521392345, + 0.006436803378164768, + 0.011145975440740585, + 0.006868163123726845, + 0.011080355383455753, + 0.012955107726156712, + 0.008765079081058502, + 0.013779421336948872, + 0.017466271296143532, + 0.03228786960244179, + 0.012071887031197548, + 0.016226110979914665, + 0.010674775578081608, + 0.01284733135253191, + 0.0060774413868784904, + 0.013041123747825623, + 0.01014751847833395, + 0.010808867402374744, + 0.008289837278425694, + 0.00634909188374877, + 0.011954523622989655, + 0.008844178169965744, + 0.004289078526198864, + 0.007403241004794836, + 0.008108745329082012, + 0.007185253314673901, + 0.006832289043813944, + 0.005390002857893705, + 0.02170969918370247, + 0.014545458368957043, + 0.011213289573788643, + 0.010193049907684326, + 0.02908826433122158, + 0.013977478258311749, + 0.014188656583428383, + 0.02458575740456581, + 0.01652620919048786, + 0.012601854279637337, + 0.00799095444381237, + 0.008138468489050865, + 0.012325647287070751, + 0.011344078928232193, + 0.012216813862323761, + 0.013057135045528412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.01605839654803276 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_13/index_tensor([589, 589])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_13/index_tensor([589, 589])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..76145e303b4b9954c939d011a284dbd77d69bdeb --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_13/index_tensor([589, 589])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dramatic low-angle wide shot anchors the massive form of a modern wind turbine, silhouetted against a clear, deep blue sky. Suddenly, with a catastrophic shudder, one of its massive, gracefully curved blades violently breaks off from its central hub at the base. The camera quickly tilts downward, maintaining focus and tracking the detached blade as it begins its terrifying plunge toward the earth. Made of a smooth, off-white composite material, the blade tumbles from head to toe in a slow, eerie, graceful descent. As it spins through the air, its vast, long surface catches the sharp, directed rays of light from the low, afternoon sun, creating a brilliant, blinding flash of light that travels along its length. In the background, a vast wind farm of identical turbines continues its rhythmic rotation against the rolling green hills. Their tranquil motion contrasts starkly and eerily with the catastrophe in the foreground, creating an atmosphere of sudden, awe-inspiring mechanical failure and impending impact.", + "seed": 17210, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_13/index_589_seed_17210_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06436058208346367, + "vlm_logprob": -0.06436058208346367, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.06436058208346367, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive modern wind turbine with gracefully curved blades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does one of the turbine blades violently break off from the central hub at the base?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the detached blade made of a smooth, off-white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to track the detached blade as it plummets toward the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast wind farm of identical turbines rotating against rolling green hills?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043748728930950165 + ], + "opposite_logprob_tokens": [ + -3.793748617172241 + ], + "logprob_reward": -0.043748728930950165, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.043748728930950165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18039347231388092, + 0.07128627598285675, + 0.18929147720336914, + 0.07847675681114197, + 0.1815105527639389, + 0.06081061065196991, + 0.08063176274299622, + 0.05375293269753456, + 0.0601688027381897, + 0.043677281588315964 + ], + "credit_map": [ + 0.02236444316804409, + 0.02585049904882908, + 0.02466273494064808, + 0.014205379411578178, + 0.020484138280153275, + 0.0169084295630455, + 0.031364236027002335, + 0.024553632363677025, + 0.009670658968389034, + 0.007041384465992451, + 0.012781066820025444, + 0.008032113313674927, + 0.007677282672375441, + 0.006944885011762381, + 0.009779853746294975, + 0.00935903750360012, + 0.016296034678816795, + 0.0076371761970222, + 0.027349617332220078, + 0.014210372231900692, + 0.010717377066612244, + 0.04761038348078728, + 0.005490190349519253, + 0.05998033657670021, + 0.007492674048990011, + 0.012326888740062714, + 0.01619383692741394, + 0.009530792012810707, + 0.009495713748037815, + 0.008234790526330471, + 0.006091678980737925, + 0.009110390208661556, + 0.020588325336575508, + 0.02231779508292675, + 0.033390890806913376, + 0.030666859820485115, + 0.01707560196518898, + 0.028782915323972702, + 0.008775955066084862, + 0.019912226125597954, + 0.00872400589287281, + 0.009754443541169167, + 0.007247175555676222, + 0.008589399978518486, + 0.0035575467627495527, + 0.008464783430099487, + 0.002417186973616481, + 0.012056073173880577, + 0.008025211282074451, + 0.009910102933645248, + 0.020040635019540787, + 0.01576659269630909, + 0.004545851610600948, + 0.007911458611488342, + 0.005568881519138813, + 0.008863032795488834, + 0.004667392000555992, + 0.005071212071925402, + 0.007033540867269039, + 0.00721386494114995, + 0.006882373709231615, + 0.006127089727669954, + 0.008247756399214268, + 0.00850970670580864, + 0.0054749357514083385, + 0.004019454587250948, + 0.009038204327225685, + 0.009581740945577621, + 0.007083616219460964, + 0.007000036537647247, + 0.007650298066437244, + 0.010320519097149372, + 0.006735547445714474, + 0.004243810661137104, + 0.0066301547922194, + 0.006685100961476564, + 0.002468834398314357, + 0.006054014433175325, + 0.003301067976281047, + 0.007558752316981554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.043748728930950165 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive modern wind turbine with gracefully curved blades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does one of the turbine blades violently break off from the central hub at the base?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the detached blade made of a smooth, off-white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to track the detached blade as it plummets toward the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast wind farm of identical turbines rotating against rolling green hills?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2016259729862213 + ], + "opposite_logprob_tokens": [ + -1.701625943183899 + ], + "logprob_reward": -0.2016259729862213, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2016259729862213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11727253347635269, + 0.04528055340051651, + 0.1575435847043991, + 0.14474913477897644, + 0.060899220407009125, + 0.17941106855869293, + 0.09218888729810715, + 0.08002161979675293, + 0.08004169166088104, + 0.04259173199534416 + ], + "credit_map": [ + 0.010157372802495956, + 0.012592638842761517, + 0.03200903534889221, + 0.00893271341919899, + 0.013709211722016335, + 0.012001012451946735, + 0.010459483601152897, + 0.017411060631275177, + 0.0099453991279006, + 0.004498154390603304, + 0.005960945971310139, + 0.003565879538655281, + 0.005744697526097298, + 0.007120612543076277, + 0.003987229894846678, + 0.00445763673633337, + 0.011683949269354343, + 0.009301934391260147, + 0.022241102531552315, + 0.01352290716022253, + 0.013923360966145992, + 0.03326015546917915, + 0.009130481630563736, + 0.04447968676686287, + 0.015210926532745361, + 0.016749383881688118, + 0.03671780228614807, + 0.02682894468307495, + 0.011244275607168674, + 0.01221653912216425, + 0.008151301182806492, + 0.017629964277148247, + 0.0071638040244579315, + 0.008846328593790531, + 0.009463105350732803, + 0.00920216552913189, + 0.0076631177216768265, + 0.008486364968121052, + 0.0036249703262001276, + 0.006449362263083458, + 0.022546373307704926, + 0.034149978309869766, + 0.017595162615180016, + 0.024150308221578598, + 0.02035820484161377, + 0.02073308452963829, + 0.009507237002253532, + 0.03037071041762829, + 0.009342587552964687, + 0.010313370265066624, + 0.025817083194851875, + 0.019417675212025642, + 0.0023636724799871445, + 0.010818811133503914, + 0.003191509284079075, + 0.010924178175628185, + 0.007022135891020298, + 0.006986311636865139, + 0.00849505327641964, + 0.007127953693270683, + 0.016780145466327667, + 0.009939783252775669, + 0.008804700337350368, + 0.01486553531140089, + 0.007022165227681398, + 0.005848743487149477, + 0.014675823971629143, + 0.013045979663729668, + 0.0063322470523417, + 0.007614434231072664, + 0.008874173276126385, + 0.016628123819828033, + 0.004411541856825352, + 0.003432654310017824, + 0.006119545549154282, + 0.008826942183077335, + 0.002703130478039384, + 0.007199691608548164, + 0.0027105773333460093, + 0.007187650538980961 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 3 + ], + "selected_reward": -0.2016259729862213 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive modern wind turbine with gracefully curved blades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does one of the turbine blades violently break off from the central hub at the base?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the detached blade made of a smooth, off-white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to track the detached blade as it plummets toward the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast wind farm of identical turbines rotating against rolling green hills?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007653911132365465 + ], + "opposite_logprob_tokens": [ + -4.882653713226318 + ], + "logprob_reward": -0.007653911132365465, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007653911132365465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24315862357616425, + 0.15083639323711395, + 0.22229143977165222, + 0.06334082782268524, + 0.030885137617588043, + 0.07027354836463928, + 0.10445969551801682, + 0.0661742240190506, + 0.019101016223430634, + 0.029479071497917175 + ], + "credit_map": [ + 0.022539937868714333, + 0.03456341102719307, + 0.038329340517520905, + 0.016579166054725647, + 0.027305223047733307, + 0.0221280287951231, + 0.049709729850292206, + 0.03200377896428108, + 0.01980605721473694, + 0.012723584659397602, + 0.037704091519117355, + 0.012684930115938187, + 0.012774513103067875, + 0.017773689702153206, + 0.01382056437432766, + 0.023548977449536324, + 0.020111052319407463, + 0.010316216386854649, + 0.03324507921934128, + 0.020091723650693893, + 0.016551168635487556, + 0.048955097794532776, + 0.01282268576323986, + 0.06019841134548187, + 0.006931878626346588, + 0.006112401373684406, + 0.0177882332354784, + 0.010368719696998596, + 0.0064713372848927975, + 0.005357157904654741, + 0.0028284257277846336, + 0.007482674438506365, + 0.00418101018294692, + 0.0035061012022197247, + 0.004972038324922323, + 0.0038435773458331823, + 0.0028910550754517317, + 0.004319807980209589, + 0.001794222043827176, + 0.005377326160669327, + 0.007658035960048437, + 0.013552362099289894, + 0.006716763600707054, + 0.009745890274643898, + 0.00858626514673233, + 0.007953420281410217, + 0.0032701236195862293, + 0.012790686450898647, + 0.010271268896758556, + 0.00867516454309225, + 0.030436063185334206, + 0.025590017437934875, + 0.0023791231215000153, + 0.009286800399422646, + 0.004466685000807047, + 0.013354572467505932, + 0.005164578091353178, + 0.0037959457840770483, + 0.007620917167514563, + 0.0065031410194933414, + 0.01624687947332859, + 0.006641021464020014, + 0.005014319438487291, + 0.015187425538897514, + 0.001943696872331202, + 0.0012398065300658345, + 0.003453055862337351, + 0.004649772774428129, + 0.0010324963368475437, + 0.0018753459444269538, + 0.0018685634713619947, + 0.0030382778495550156, + 0.004993454087525606, + 0.003729399526491761, + 0.0044153607450425625, + 0.004947420675307512, + 0.0018204784719273448, + 0.003492511110380292, + 0.0015911608934402466, + 0.004489284940063953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.007653911132365465 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive modern wind turbine with gracefully curved blades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does one of the turbine blades violently break off from the central hub at the base?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the detached blade made of a smooth, off-white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to track the detached blade as it plummets toward the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast wind farm of identical turbines rotating against rolling green hills?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067406888119876385 + ], + "opposite_logprob_tokens": [ + -5.006740570068359 + ], + "logprob_reward": -0.0067406888119876385, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067406888119876385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1756969392299652, + 0.06066368147730827, + 0.14601927995681763, + 0.10960114002227783, + 0.05847259238362312, + 0.10324379056692123, + 0.08732539415359497, + 0.09118109196424484, + 0.09735704958438873, + 0.07043905556201935 + ], + "credit_map": [ + 0.0182377640157938, + 0.02383880503475666, + 0.025815308094024658, + 0.014700190164148808, + 0.019682414829730988, + 0.016926808282732964, + 0.03559543937444687, + 0.020900201052427292, + 0.009923793375492096, + 0.005336762871593237, + 0.011387521401047707, + 0.005421897396445274, + 0.005882524885237217, + 0.007179449312388897, + 0.009199108928442001, + 0.0063326237723231316, + 0.01181305106729269, + 0.005898334551602602, + 0.021801002323627472, + 0.011542027816176414, + 0.006684147287160158, + 0.036341309547424316, + 0.004441959783434868, + 0.047497425228357315, + 0.011026391759514809, + 0.01604587770998478, + 0.01881093531847, + 0.013993678614497185, + 0.0203902255743742, + 0.011706872843205929, + 0.01028014812618494, + 0.007346996571868658, + 0.006379618309438229, + 0.005789639428257942, + 0.01165601797401905, + 0.009216364473104477, + 0.005415323656052351, + 0.008589965291321278, + 0.003178267739713192, + 0.008247392252087593, + 0.010270137339830399, + 0.02276313304901123, + 0.012802334502339363, + 0.013762890361249447, + 0.009180250577628613, + 0.014904320240020752, + 0.006853342521935701, + 0.012707367539405823, + 0.01019103080034256, + 0.012141554616391659, + 0.01938006840646267, + 0.01139722391963005, + 0.00450926972553134, + 0.010937647894024849, + 0.004838715307414532, + 0.01392988208681345, + 0.010762226767838001, + 0.006342568434774876, + 0.010106307454407215, + 0.00929489079862833, + 0.01743810437619686, + 0.011368230916559696, + 0.012012521736323833, + 0.013856235891580582, + 0.007780064363032579, + 0.005928861442953348, + 0.022660650312900543, + 0.010324801318347454, + 0.00384580553509295, + 0.011191025376319885, + 0.007358334958553314, + 0.028267504647374153, + 0.008206172846257687, + 0.007030415814369917, + 0.011492098681628704, + 0.015160751529037952, + 0.004516012966632843, + 0.010299989953637123, + 0.002865537768229842, + 0.010868075303733349 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.0067406888119876385 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive modern wind turbine with gracefully curved blades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does one of the turbine blades violently break off from the central hub at the base?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the detached blade made of a smooth, off-white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to track the detached blade as it plummets toward the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast wind farm of identical turbines rotating against rolling green hills?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203360855579376 + ], + "opposite_logprob_tokens": [ + -2.8120336532592773 + ], + "logprob_reward": -0.06203360855579376, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203360855579376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03858982026576996, + 0.018075447529554367, + 0.08816990256309509, + 0.025670643895864487, + 0.04261000454425812, + 0.15303638577461243, + 0.08664389699697495, + 0.10577892512083054, + 0.4060879647731781, + 0.035337019711732864 + ], + "credit_map": [ + 0.004096743185073137, + 0.0058287205174565315, + 0.005076550878584385, + 0.003287374274805188, + 0.005650791339576244, + 0.0050927745178341866, + 0.004964082967489958, + 0.0045927842147648335, + 0.002425813116133213, + 0.001412104582414031, + 0.00492390152066946, + 0.0012645811075344682, + 0.0016894558211788535, + 0.0017987830797210336, + 0.0014934198698028922, + 0.0030673877336084843, + 0.007565283682197332, + 0.0038604275323450565, + 0.01675383374094963, + 0.007286465726792812, + 0.005609012674540281, + 0.016936922445893288, + 0.0053580813109874725, + 0.02479987032711506, + 0.003052891232073307, + 0.004680250771343708, + 0.005076555069535971, + 0.0037894221022725105, + 0.0023454504553228617, + 0.002981662517413497, + 0.0011949323816224933, + 0.0025494771543890238, + 0.005776930600404739, + 0.005652822088450193, + 0.00616674916818738, + 0.005798495374619961, + 0.004001418594270945, + 0.005891813896596432, + 0.002916628960520029, + 0.006405147723853588, + 0.02040139213204384, + 0.02384338714182377, + 0.016789818182587624, + 0.024900823831558228, + 0.016437657177448273, + 0.01695171743631363, + 0.007194305304437876, + 0.02651727758347988, + 0.007951172068715096, + 0.005678276065737009, + 0.024579478427767754, + 0.021546700969338417, + 0.002670696470886469, + 0.007040934171527624, + 0.003207230009138584, + 0.01396941114217043, + 0.009828331880271435, + 0.00856765452772379, + 0.013030603528022766, + 0.010560121387243271, + 0.022175200283527374, + 0.014985651709139347, + 0.008215552195906639, + 0.018415816128253937, + 0.03383340314030647, + 0.021898292005062103, + 0.08340953290462494, + 0.044495172798633575, + 0.03592654690146446, + 0.04859290271997452, + 0.036565277725458145, + 0.10136682540178299, + 0.0037772534415125847, + 0.0034364801831543446, + 0.005095836706459522, + 0.007088371552526951, + 0.0018870499916374683, + 0.006373930722475052, + 0.0024841304402798414, + 0.005193968303501606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.06203360855579376 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic low-angle wide shot anchors the massive form of a modern wind turbine, silhouetted against a clear, deep blue sky. Suddenly, with a catastrophic shudder, one of its massive, gracefully curved blades violently breaks off from its central hub at the base. The camera quickly tilts downward, maintaining focus and tracking the detached blade as it begins its terrifying plunge toward the earth. Made of a smooth, off-white composite material, the blade tumbles from head to toe in a slow, eerie, graceful descent. As it spins through the air, its vast, long surface catches the sharp, directed rays of light from the low, afternoon sun, creating a brilliant, blinding flash of light that travels along its length. In the background, a vast wind farm of identical turbines continues its rhythmic rotation against the rolling green hills. Their tranquil motion contrasts starkly and eerily with the catastrophe in the foreground, creating an atmosphere of sudden, awe-inspiring mechanical failure and impending impact.", + "seed": 17211, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_13/index_589_seed_17211_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06505048051476478, + "vlm_logprob": -0.06505048051476478, + "vlm_yes_no_margin": 3.349999952316284, + "vlm_token_credit": -0.06505048051476478, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive modern wind turbine with gracefully curved blades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does one of the turbine blades violently break off from the central hub at the base?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the detached blade made of a smooth, off-white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to track the detached blade as it plummets toward the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast wind farm of identical turbines rotating against rolling green hills?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021206889301538467 + ], + "opposite_logprob_tokens": [ + -4.396206855773926 + ], + "logprob_reward": -0.021206889301538467, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.021206889301538467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12550680339336395, + 0.0590628944337368, + 0.08762536942958832, + 0.04929869994521141, + 0.10174072533845901, + 0.17012372612953186, + 0.2606285512447357, + 0.04444894567131996, + 0.08395764231681824, + 0.017606625333428383 + ], + "credit_map": [ + 0.012581425718963146, + 0.005745040252804756, + 0.008843299001455307, + 0.007765532471239567, + 0.035165540874004364, + 0.02033446915447712, + 0.011617710813879967, + 0.02345380187034607, + 0.005862267222255468, + 0.0027607562951743603, + 0.003803614526987076, + 0.0016803944017738104, + 0.019986402243375778, + 0.011357088573276997, + 0.0037795936223119497, + 0.009832785464823246, + 0.006515412125736475, + 0.001813822309486568, + 0.0356852263212204, + 0.0037211687304079533, + 0.009280942380428314, + 0.005383417475968599, + 0.004403539001941681, + 0.020821847021579742, + 0.0050209322944283485, + 0.002822872716933489, + 0.008076749742031097, + 0.004791571758687496, + 0.006375767290592194, + 0.006866480223834515, + 0.008318597450852394, + 0.007025732193142176, + 0.009522555395960808, + 0.013898412697017193, + 0.015034012496471405, + 0.01272421982139349, + 0.00964641198515892, + 0.010713580995798111, + 0.01471463218331337, + 0.01548691000789404, + 0.017894305288791656, + 0.01990647427737713, + 0.01989222690463066, + 0.024295026436448097, + 0.01978384517133236, + 0.02375442162156105, + 0.023361017927527428, + 0.021236425265669823, + 0.026990899816155434, + 0.02755923941731453, + 0.03849135711789131, + 0.04599776864051819, + 0.02561820112168789, + 0.041962310671806335, + 0.022635752335190773, + 0.031373031437397, + 0.0044139474630355835, + 0.005015384405851364, + 0.005900534801185131, + 0.007986676879227161, + 0.004166466649621725, + 0.007186146918684244, + 0.0027877981774508953, + 0.006991998758167028, + 0.007299595978111029, + 0.009318264201283455, + 0.0248141847550869, + 0.02063072845339775, + 0.005573425907641649, + 0.004664193838834763, + 0.0048309266567230225, + 0.006826338358223438, + 0.0019518849439918995, + 0.0029613159131258726, + 0.003789599286392331, + 0.0032248960342258215, + 0.001347952987998724, + 0.0013505812967196107, + 0.001378617831505835, + 0.0016017785528674722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.021206889301538467 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive modern wind turbine with gracefully curved blades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does one of the turbine blades violently break off from the central hub at the base?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the detached blade made of a smooth, off-white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to track the detached blade as it plummets toward the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast wind farm of identical turbines rotating against rolling green hills?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2016168087720871 + ], + "opposite_logprob_tokens": [ + -1.7016167640686035 + ], + "logprob_reward": -0.2016168087720871, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2016168087720871, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05536050722002983, + 0.051044244319200516, + 0.038481198251247406, + 0.06375515460968018, + 0.08775954693555832, + 0.06575625389814377, + 0.178217813372612, + 0.08789782971143723, + 0.16979779303073883, + 0.20192968845367432 + ], + "credit_map": [ + 0.006705228239297867, + 0.0036216597072780132, + 0.007013834081590176, + 0.0045706722885370255, + 0.013309262692928314, + 0.007886076346039772, + 0.005419289227575064, + 0.006834483705461025, + 0.004694217350333929, + 0.0014077892992645502, + 0.0021429385524243116, + 0.0017028700094670057, + 0.018307043239474297, + 0.010175188072025776, + 0.0036416228394955397, + 0.008972578682005405, + 0.0031633474864065647, + 0.0010003087809309363, + 0.014335091225802898, + 0.0020509161986410618, + 0.004180915653705597, + 0.0037140659987926483, + 0.002612992888316512, + 0.007423560135066509, + 0.005647154990583658, + 0.004239980131387711, + 0.009324358776211739, + 0.006169814150780439, + 0.008855306543409824, + 0.007693242281675339, + 0.012016450986266136, + 0.009808851405978203, + 0.009365594014525414, + 0.008182594552636147, + 0.015295471996068954, + 0.00971770565956831, + 0.005853527691215277, + 0.008112048730254173, + 0.014577512629330158, + 0.01665508560836315, + 0.006764209363609552, + 0.007596340961754322, + 0.008209612220525742, + 0.009364682249724865, + 0.007326226215809584, + 0.009116229601204395, + 0.008881477639079094, + 0.008497473783791065, + 0.021091852337121964, + 0.02186080813407898, + 0.023996496573090553, + 0.026867976412177086, + 0.018269972875714302, + 0.03431091457605362, + 0.014400270767509937, + 0.017419513314962387, + 0.012931806966662407, + 0.012083153240382671, + 0.01609688065946102, + 0.013411305844783783, + 0.0090594831854105, + 0.008927365764975548, + 0.005636520683765411, + 0.009751313365995884, + 0.016191020607948303, + 0.019292965531349182, + 0.02719241939485073, + 0.027018645778298378, + 0.026451369747519493, + 0.016204219311475754, + 0.019382832571864128, + 0.018064314499497414, + 0.03069200925529003, + 0.05548177659511566, + 0.03773481398820877, + 0.028782682493329048, + 0.015715934336185455, + 0.013981824740767479, + 0.008245929144322872, + 0.011294731870293617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.2016168087720871 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive modern wind turbine with gracefully curved blades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does one of the turbine blades violently break off from the central hub at the base?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the detached blade made of a smooth, off-white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to track the detached blade as it plummets toward the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast wind farm of identical turbines rotating against rolling green hills?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01422715000808239 + ], + "opposite_logprob_tokens": [ + -4.264226913452148 + ], + "logprob_reward": -0.01422715000808239, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01422715000808239, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058625783771276474, + 0.044036902487277985, + 0.13831306993961334, + 0.0666276142001152, + 0.16284671425819397, + 0.12125928699970245, + 0.0717443972826004, + 0.04880847409367561, + 0.09419775754213333, + 0.19353994727134705 + ], + "credit_map": [ + 0.006994929630309343, + 0.004707846324890852, + 0.008806788362562656, + 0.006860954686999321, + 0.008276830427348614, + 0.0061735957860946655, + 0.007610292173922062, + 0.009194545447826385, + 0.004403580445796251, + 0.0037286817096173763, + 0.004398633725941181, + 0.004047911148518324, + 0.010662502609193325, + 0.006793394684791565, + 0.004317543935030699, + 0.005684652365744114, + 0.010721084661781788, + 0.002176560927182436, + 0.05715922638773918, + 0.0053660511039197445, + 0.01485719159245491, + 0.008716891519725323, + 0.005217161029577255, + 0.03409890457987785, + 0.006203237920999527, + 0.0040533398278057575, + 0.011812935583293438, + 0.007277446798980236, + 0.007018711417913437, + 0.008029489777982235, + 0.01245527621358633, + 0.009777176193892956, + 0.014599684625864029, + 0.015292413532733917, + 0.03419027104973793, + 0.02082577720284462, + 0.013208926655352116, + 0.015383992344141006, + 0.017225757241249084, + 0.03211989626288414, + 0.013325423002243042, + 0.0160836149007082, + 0.013007118366658688, + 0.018016885966062546, + 0.013834628276526928, + 0.018143828958272934, + 0.014547758735716343, + 0.014300025068223476, + 0.007272150367498398, + 0.006581319496035576, + 0.012270809151232243, + 0.014417995698750019, + 0.005757386330515146, + 0.010253374464809895, + 0.0054871090687811375, + 0.009704258292913437, + 0.004664329811930656, + 0.0045352717861533165, + 0.009129935875535011, + 0.009601152502000332, + 0.003795117838308215, + 0.007548709865659475, + 0.0030505529139190912, + 0.006483403034508228, + 0.008949819952249527, + 0.008510026149451733, + 0.031093623489141464, + 0.024477282539010048, + 0.005781805608421564, + 0.004668160807341337, + 0.0035227600019425154, + 0.007194277364760637, + 0.025616072118282318, + 0.048685185611248016, + 0.038244131952524185, + 0.02982342429459095, + 0.013999616727232933, + 0.013278571888804436, + 0.011653805151581764, + 0.012239140458405018 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 2 + ], + "selected_reward": -0.01422715000808239 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive modern wind turbine with gracefully curved blades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does one of the turbine blades violently break off from the central hub at the base?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the detached blade made of a smooth, off-white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to track the detached blade as it plummets toward the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast wind farm of identical turbines rotating against rolling green hills?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018191518262028694 + ], + "opposite_logprob_tokens": [ + -4.018191337585449 + ], + "logprob_reward": -0.018191518262028694, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018191518262028694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12665845453739166, + 0.03133757784962654, + 0.15132786333560944, + 0.08046654611825943, + 0.13230949640274048, + 0.0868837758898735, + 0.09445345401763916, + 0.03062095120549202, + 0.04841870069503784, + 0.21752318739891052 + ], + "credit_map": [ + 0.011591638438403606, + 0.004484998062252998, + 0.011761548928916454, + 0.00750510161742568, + 0.03970838710665703, + 0.03379695490002632, + 0.009195687249302864, + 0.008614144288003445, + 0.0038101959507912397, + 0.0017560424748808146, + 0.004395581781864166, + 0.0026359972544014454, + 0.006207793019711971, + 0.003823547624051571, + 0.0038680494762957096, + 0.004840370267629623, + 0.011401085183024406, + 0.0021853900980204344, + 0.06451334059238434, + 0.005605593789368868, + 0.01547072920948267, + 0.009975889697670937, + 0.005105965305119753, + 0.03706987202167511, + 0.004946145229041576, + 0.003754250705242157, + 0.012447896413505077, + 0.009489261545240879, + 0.01299450546503067, + 0.008917327970266342, + 0.01629173755645752, + 0.011625424958765507, + 0.01130425464361906, + 0.01372568029910326, + 0.023836858570575714, + 0.016734199598431587, + 0.010176872834563255, + 0.012983818538486958, + 0.017960814759135246, + 0.025586996227502823, + 0.008654571138322353, + 0.010360968299210072, + 0.009981989860534668, + 0.010963237844407558, + 0.01105713564902544, + 0.012311032973229885, + 0.011378197930753231, + 0.01217664685100317, + 0.016738122329115868, + 0.016337808221578598, + 0.007536748424172401, + 0.006715577095746994, + 0.0050586010329425335, + 0.03353249654173851, + 0.0032296031713485718, + 0.005304502323269844, + 0.002972949296236038, + 0.003257433185353875, + 0.005251043941825628, + 0.005255616270005703, + 0.003221208695322275, + 0.0035900643561035395, + 0.0025145967956632376, + 0.004558040294796228, + 0.004701830446720123, + 0.005891763139516115, + 0.00915980339050293, + 0.007660169620066881, + 0.003791585098952055, + 0.00625979108735919, + 0.004767835605889559, + 0.006185928825289011, + 0.02989974059164524, + 0.05573824420571327, + 0.042027153074741364, + 0.03262575343251228, + 0.016316009685397148, + 0.015313549898564816, + 0.01228562742471695, + 0.013317125849425793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 4 + ], + "selected_reward": -0.018191518262028694 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive modern wind turbine with gracefully curved blades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does one of the turbine blades violently break off from the central hub at the base?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the detached blade made of a smooth, off-white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to track the detached blade as it plummets toward the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast wind farm of identical turbines rotating against rolling green hills?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001003623008728 + ], + "opposite_logprob_tokens": [ + -2.69500994682312 + ], + "logprob_reward": -0.07001003623008728, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001003623008728, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0643884614109993, + 0.04975859075784683, + 0.06717011332511902, + 0.050337012857198715, + 0.03855198621749878, + 0.12106599658727646, + 0.09730355441570282, + 0.05343789979815483, + 0.3527282178401947, + 0.10525818914175034 + ], + "credit_map": [ + 0.005539617035537958, + 0.0034332536160945892, + 0.007449022959917784, + 0.005502927582710981, + 0.018052073195576668, + 0.01067038718611002, + 0.006170518230646849, + 0.007570663932710886, + 0.004958339501172304, + 0.0017536997329443693, + 0.0027328557334840298, + 0.0016209405148401856, + 0.01798795536160469, + 0.010059692896902561, + 0.002737394766882062, + 0.007907711900770664, + 0.005087666213512421, + 0.0010901144705712795, + 0.027316778898239136, + 0.0028424630872905254, + 0.005914162844419479, + 0.004439704120159149, + 0.0031055465806275606, + 0.017373673617839813, + 0.0028562319930642843, + 0.0034907746594399214, + 0.008019081316888332, + 0.006771706510335207, + 0.004478114191442728, + 0.005063909571617842, + 0.011600849218666553, + 0.008056343533098698, + 0.0048258802853524685, + 0.003677741391584277, + 0.005294557195156813, + 0.004078366328030825, + 0.004358806647360325, + 0.007034669630229473, + 0.004210091661661863, + 0.005071874242275953, + 0.014879439026117325, + 0.010330725461244583, + 0.021249687299132347, + 0.019247202202677727, + 0.009265709668397903, + 0.015188377350568771, + 0.010301253758370876, + 0.0206036064773798, + 0.011783115565776825, + 0.012057261541485786, + 0.011668666265904903, + 0.013362940400838852, + 0.008623220957815647, + 0.021340738981962204, + 0.00743908854201436, + 0.01102852076292038, + 0.006612979341298342, + 0.004141718149185181, + 0.009167256765067577, + 0.010104276239871979, + 0.0033712356816977262, + 0.009222897700965405, + 0.0021079417783766985, + 0.008709595538675785, + 0.03095628321170807, + 0.037393972277641296, + 0.12055237591266632, + 0.09723140299320221, + 0.015994004905223846, + 0.016514191403985023, + 0.010739645920693874, + 0.023346351459622383, + 0.012313039973378181, + 0.024508610367774963, + 0.021288493648171425, + 0.01958238147199154, + 0.007907435297966003, + 0.006801459472626448, + 0.0050200498662889, + 0.007836715318262577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -0.07001003623008728 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic low-angle wide shot anchors the massive form of a modern wind turbine, silhouetted against a clear, deep blue sky. Suddenly, with a catastrophic shudder, one of its massive, gracefully curved blades violently breaks off from its central hub at the base. The camera quickly tilts downward, maintaining focus and tracking the detached blade as it begins its terrifying plunge toward the earth. Made of a smooth, off-white composite material, the blade tumbles from head to toe in a slow, eerie, graceful descent. As it spins through the air, its vast, long surface catches the sharp, directed rays of light from the low, afternoon sun, creating a brilliant, blinding flash of light that travels along its length. In the background, a vast wind farm of identical turbines continues its rhythmic rotation against the rolling green hills. Their tranquil motion contrasts starkly and eerily with the catastrophe in the foreground, creating an atmosphere of sudden, awe-inspiring mechanical failure and impending impact.", + "seed": 17212, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_13/index_589_seed_17212_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.11716844392940402, + "vlm_logprob": -0.11716844392940402, + "vlm_yes_no_margin": 3.199999976158142, + "vlm_token_credit": -0.11716844392940402, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive modern wind turbine with gracefully curved blades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does one of the turbine blades violently break off from the central hub at the base?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the detached blade made of a smooth, off-white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to track the detached blade as it plummets toward the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast wind farm of identical turbines rotating against rolling green hills?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3954061269760132 + ], + "opposite_logprob_tokens": [ + -1.1454061269760132 + ], + "logprob_reward": -0.3954061269760132, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3954061269760132, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10423983633518219, + 0.297983318567276, + 0.043496809899806976, + 0.05550777167081833, + 0.13482660055160522, + 0.03189794346690178, + 0.1745152473449707, + 0.07637061178684235, + 0.034341201186180115, + 0.046820707619190216 + ], + "credit_map": [ + 0.012046952731907368, + 0.015339452773332596, + 0.016115451231598854, + 0.02242857590317726, + 0.010004020296037197, + 0.009770719334483147, + 0.007999389432370663, + 0.010535268113017082, + 0.02441554330289364, + 0.11522238701581955, + 0.04874882847070694, + 0.02028653956949711, + 0.006605697795748711, + 0.07141105830669403, + 0.004640617873519659, + 0.006652609445154667, + 0.005451499950140715, + 0.011555494740605354, + 0.00354529800824821, + 0.0036834264174103737, + 0.004240675829350948, + 0.007422067690640688, + 0.00332860485650599, + 0.004269735887646675, + 0.006028106901794672, + 0.007128922268748283, + 0.00842121709138155, + 0.00712971156463027, + 0.008121708407998085, + 0.007305440027266741, + 0.004529866855591536, + 0.006842793431133032, + 0.015667865052819252, + 0.02155090682208538, + 0.014159333892166615, + 0.014774296432733536, + 0.01458643190562725, + 0.02272169105708599, + 0.012034821324050426, + 0.019331254065036774, + 0.0034237857908010483, + 0.0054891761392354965, + 0.0033599629532545805, + 0.0016704746522009373, + 0.005309558007866144, + 0.00546989543363452, + 0.0035459951031953096, + 0.0036290932912379503, + 0.020241642370820045, + 0.05557896941900253, + 0.007256796117872, + 0.00980218593031168, + 0.020114397630095482, + 0.04509124904870987, + 0.009248088113963604, + 0.007181913126260042, + 0.0062095001339912415, + 0.01918499357998371, + 0.0062422603368759155, + 0.006112367380410433, + 0.009796497412025928, + 0.01769072562456131, + 0.005742894951254129, + 0.005391368176788092, + 0.003380491863936186, + 0.004488660953938961, + 0.003546122694388032, + 0.0028366667684167624, + 0.007026939652860165, + 0.00476432777941227, + 0.003959208261221647, + 0.004338779021054506, + 0.0054825483821332455, + 0.006248333025723696, + 0.005760607775300741, + 0.003965878859162331, + 0.011309558525681496, + 0.00812410656362772, + 0.003415683750063181, + 0.0025139874778687954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 4 + ], + "selected_reward": -0.3954061269760132 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive modern wind turbine with gracefully curved blades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does one of the turbine blades violently break off from the central hub at the base?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the detached blade made of a smooth, off-white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to track the detached blade as it plummets toward the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast wind farm of identical turbines rotating against rolling green hills?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0337274894118309 + ], + "opposite_logprob_tokens": [ + -3.4087274074554443 + ], + "logprob_reward": -0.0337274894118309, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0337274894118309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07064913213253021, + 0.06673680990934372, + 0.03511909767985344, + 0.036829933524131775, + 0.04196697100996971, + 0.045699991285800934, + 0.20852556824684143, + 0.14145527780056, + 0.029246516525745392, + 0.3237707316875458 + ], + "credit_map": [ + 0.0072334990836679935, + 0.008915713988244534, + 0.009063471108675003, + 0.011585108004510403, + 0.013992579653859138, + 0.006518582347780466, + 0.0060035912320017815, + 0.007336581125855446, + 0.004977824166417122, + 0.01220595370978117, + 0.010567247867584229, + 0.007690295111387968, + 0.005809451453387737, + 0.012417268007993698, + 0.005724640563130379, + 0.007344122510403395, + 0.0031705305445939302, + 0.004904980771243572, + 0.0061815581284463406, + 0.00599663844332099, + 0.003335135756060481, + 0.004846196621656418, + 0.003388709155842662, + 0.003295348957180977, + 0.003504745662212372, + 0.0056526511907577515, + 0.004751121159642935, + 0.004433751571923494, + 0.007212928030639887, + 0.005389499943703413, + 0.0027391209732741117, + 0.0031461145263165236, + 0.004537195898592472, + 0.006143420003354549, + 0.004399527329951525, + 0.004835294093936682, + 0.004413892049342394, + 0.00763653265312314, + 0.004103173967450857, + 0.005897929426282644, + 0.004731959663331509, + 0.006486446596682072, + 0.005723869893699884, + 0.0039574624970555305, + 0.008210881613194942, + 0.007004649844020605, + 0.005421432200819254, + 0.00416328851133585, + 0.017238596454262733, + 0.07728774100542068, + 0.01043422520160675, + 0.006093354430049658, + 0.016311563551425934, + 0.06420139223337173, + 0.009904750622808933, + 0.007053928915411234, + 0.015141142532229424, + 0.049397267401218414, + 0.006846205797046423, + 0.0048258136957883835, + 0.010172196663916111, + 0.03754604607820511, + 0.008879548870027065, + 0.008647050708532333, + 0.002640625461935997, + 0.003358863526955247, + 0.003322046948596835, + 0.0022633220069110394, + 0.0060616047121584415, + 0.003462371416389942, + 0.0046411254443228245, + 0.0034965570084750652, + 0.03531709313392639, + 0.038486432284116745, + 0.031492285430431366, + 0.024835260584950447, + 0.08156059682369232, + 0.05718403309583664, + 0.035359252244234085, + 0.019535772502422333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.0337274894118309 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive modern wind turbine with gracefully curved blades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does one of the turbine blades violently break off from the central hub at the base?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the detached blade made of a smooth, off-white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to track the detached blade as it plummets toward the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast wind farm of identical turbines rotating against rolling green hills?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00527229392901063 + ], + "opposite_logprob_tokens": [ + -5.255272388458252 + ], + "logprob_reward": -0.00527229392901063, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00527229392901063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05819609761238098, + 0.719427764415741, + 0.01841348223388195, + 0.030404245480895042, + 0.04843614622950554, + 0.02888023667037487, + 0.029658813029527664, + 0.02226732298731804, + 0.0100243603810668, + 0.0342915803194046 + ], + "credit_map": [ + 0.006577702704817057, + 0.006961822509765625, + 0.009598735719919205, + 0.011014191433787346, + 0.004501791670918465, + 0.006324762012809515, + 0.0047894730232656, + 0.008427614346146584, + 0.06500623375177383, + 0.29574865102767944, + 0.0950322151184082, + 0.035829681903123856, + 0.013709669932723045, + 0.19257627427577972, + 0.008966760709881783, + 0.012558289803564548, + 0.002247535390779376, + 0.0046695745550096035, + 0.00182630296330899, + 0.0018348790472373366, + 0.0019675293006002903, + 0.0028695673681795597, + 0.001440625637769699, + 0.001557468785904348, + 0.0035908198915421963, + 0.005375599954277277, + 0.0037071295082569122, + 0.0022119039203971624, + 0.0039306222461164, + 0.0047514792531728745, + 0.003051197621971369, + 0.003785491921007633, + 0.005701096728444099, + 0.008254957385361195, + 0.004955857060849667, + 0.004958111792802811, + 0.005185310263186693, + 0.008487177081406116, + 0.004275586921721697, + 0.006618049927055836, + 0.0027766902931034565, + 0.006041930988430977, + 0.0028215055353939533, + 0.0019536451436579227, + 0.004815157037228346, + 0.0057626450434327126, + 0.001694485079497099, + 0.0030141782481223345, + 0.0034808588679879904, + 0.0037643839605152607, + 0.0034875504206866026, + 0.0021864748559892178, + 0.0077835009433329105, + 0.0034063791390508413, + 0.0032093357294797897, + 0.0023403309751302004, + 0.002085095038637519, + 0.004322962835431099, + 0.0025993993040174246, + 0.0032618329860270023, + 0.0027725351974368095, + 0.003310574684292078, + 0.0018310068408027291, + 0.002083913190290332, + 0.0009348503081128001, + 0.0011890149908140302, + 0.001596492831595242, + 0.0007558249053545296, + 0.0016908150864765048, + 0.001378132146783173, + 0.0011781810317188501, + 0.0013010482070967555, + 0.003648197976872325, + 0.004134349524974823, + 0.007461633067578077, + 0.003942952491343021, + 0.005453765857964754, + 0.004158183466643095, + 0.002786344848573208, + 0.0027061516884714365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.00527229392901063 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive modern wind turbine with gracefully curved blades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does one of the turbine blades violently break off from the central hub at the base?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the detached blade made of a smooth, off-white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to track the detached blade as it plummets toward the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast wind farm of identical turbines rotating against rolling green hills?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638749830424786 + ], + "opposite_logprob_tokens": [ + -4.758638858795166 + ], + "logprob_reward": -0.008638749830424786, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638749830424786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05547820031642914, + 0.6107489466667175, + 0.036614418029785156, + 0.03667164593935013, + 0.015373093076050282, + 0.04215363413095474, + 0.05627899616956711, + 0.07123161107301712, + 0.027200784534215927, + 0.04824867844581604 + ], + "credit_map": [ + 0.005780787207186222, + 0.006045603659003973, + 0.00958959385752678, + 0.01378022413700819, + 0.006347260437905788, + 0.004339933395385742, + 0.004176810383796692, + 0.005417985841631889, + 0.05262784659862518, + 0.24093998968601227, + 0.08802765607833862, + 0.03547961637377739, + 0.012994149699807167, + 0.1553468108177185, + 0.009385112673044205, + 0.015947770327329636, + 0.005325788166373968, + 0.008504786528646946, + 0.005980048794299364, + 0.004122158512473106, + 0.0019034274155274034, + 0.005234081763774157, + 0.0018924630712717772, + 0.0036516645923256874, + 0.003559424541890621, + 0.006204548291862011, + 0.0051431190222501755, + 0.005430706311017275, + 0.004397773649543524, + 0.005679191555827856, + 0.0028346774633973837, + 0.0034222037065774202, + 0.0018702975939959288, + 0.0018207652028650045, + 0.002978654345497489, + 0.0018592624692246318, + 0.0010654067154973745, + 0.0024819467216730118, + 0.001336402609013021, + 0.001960358116775751, + 0.004959952551871538, + 0.010681860148906708, + 0.004571296274662018, + 0.0027726555708795786, + 0.003550667315721512, + 0.009315150789916515, + 0.0020073717460036278, + 0.004294678568840027, + 0.00526272039860487, + 0.008422830142080784, + 0.009056090377271175, + 0.00821434985846281, + 0.0067206197418272495, + 0.010180140845477581, + 0.003949804697185755, + 0.004472437314689159, + 0.010335821658372879, + 0.015156633220613003, + 0.004823606926947832, + 0.005830507725477219, + 0.011822951957583427, + 0.01203501969575882, + 0.005221220664680004, + 0.006005849223583937, + 0.0022732566576451063, + 0.003093366278335452, + 0.0027288859710097313, + 0.0024107808712869883, + 0.006192454602569342, + 0.0032969459425657988, + 0.002975900424644351, + 0.004229194484651089, + 0.005504051223397255, + 0.006398355588316917, + 0.00734386732801795, + 0.004713712725788355, + 0.009951550513505936, + 0.006953277159482241, + 0.0031450772657990456, + 0.004238787107169628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 6 + ], + "selected_reward": -0.008638749830424786 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive modern wind turbine with gracefully curved blades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does one of the turbine blades violently break off from the central hub at the base?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the detached blade made of a smooth, off-white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to track the detached blade as it plummets toward the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast wind farm of identical turbines rotating against rolling green hills?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1427975594997406 + ], + "opposite_logprob_tokens": [ + -2.0177974700927734 + ], + "logprob_reward": -0.1427975594997406, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.1427975594997406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07837370038032532, + 0.523898184299469, + 0.029182802885770798, + 0.05077497288584709, + 0.03757335618138313, + 0.06537014991044998, + 0.034933093935251236, + 0.07386308163404465, + 0.029466819018125534, + 0.07656382769346237 + ], + "credit_map": [ + 0.006657253485172987, + 0.005341826938092709, + 0.01958145573735237, + 0.022935543209314346, + 0.008087174035608768, + 0.004268214572221041, + 0.004922584164887667, + 0.0065796393901109695, + 0.046498171985149384, + 0.21576760709285736, + 0.06652279943227768, + 0.026843028143048286, + 0.009962952695786953, + 0.1420799344778061, + 0.007447447162121534, + 0.0087762251496315, + 0.0035921509843319654, + 0.005114004481583834, + 0.005624923389405012, + 0.005108626093715429, + 0.0018163160420954227, + 0.0032856452744454145, + 0.0018691045697778463, + 0.0027720334473997355, + 0.005707028787583113, + 0.006943383254110813, + 0.008660916239023209, + 0.008800294250249863, + 0.005269573535770178, + 0.006569393444806337, + 0.004202579613775015, + 0.004621801897883415, + 0.004957502707839012, + 0.006357725244015455, + 0.005898757837712765, + 0.005775421857833862, + 0.0033449509646743536, + 0.004782964941114187, + 0.0033353150356560946, + 0.003120716428384185, + 0.006837611086666584, + 0.010126887820661068, + 0.01081823743879795, + 0.008880667388439178, + 0.005606391467154026, + 0.00757951196283102, + 0.004521689843386412, + 0.010999159887433052, + 0.00438128923997283, + 0.0038641446735709906, + 0.005846105515956879, + 0.005594633985310793, + 0.005391203798353672, + 0.0038278778083622456, + 0.002713691908866167, + 0.003314146539196372, + 0.007677010260522366, + 0.009362204931676388, + 0.011861685663461685, + 0.011135576292872429, + 0.0076155816204845905, + 0.00918482057750225, + 0.007001529447734356, + 0.010024670511484146, + 0.0034726683516055346, + 0.003480812767520547, + 0.004894740879535675, + 0.006427636835724115, + 0.002955025527626276, + 0.0030925243627279997, + 0.0022657415829598904, + 0.002877669408917427, + 0.008278269320726395, + 0.008241206407546997, + 0.018192829564213753, + 0.015018008649349213, + 0.008334679529070854, + 0.00818417314440012, + 0.004096534103155136, + 0.006218125578016043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.1427975594997406 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic low-angle wide shot anchors the massive form of a modern wind turbine, silhouetted against a clear, deep blue sky. Suddenly, with a catastrophic shudder, one of its massive, gracefully curved blades violently breaks off from its central hub at the base. The camera quickly tilts downward, maintaining focus and tracking the detached blade as it begins its terrifying plunge toward the earth. Made of a smooth, off-white composite material, the blade tumbles from head to toe in a slow, eerie, graceful descent. As it spins through the air, its vast, long surface catches the sharp, directed rays of light from the low, afternoon sun, creating a brilliant, blinding flash of light that travels along its length. In the background, a vast wind farm of identical turbines continues its rhythmic rotation against the rolling green hills. Their tranquil motion contrasts starkly and eerily with the catastrophe in the foreground, creating an atmosphere of sudden, awe-inspiring mechanical failure and impending impact.", + "seed": 17213, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_13/index_589_seed_17213_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03294882569462061, + "vlm_logprob": -0.03294882569462061, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.03294882569462061, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive modern wind turbine with gracefully curved blades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does one of the turbine blades violently break off from the central hub at the base?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the detached blade made of a smooth, off-white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to track the detached blade as it plummets toward the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast wind farm of identical turbines rotating against rolling green hills?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05553646758198738 + ], + "opposite_logprob_tokens": [ + -3.1805365085601807 + ], + "logprob_reward": -0.05553646758198738, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.05553646758198738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23617678880691528, + 0.04493613541126251, + 0.11451411992311478, + 0.16487377882003784, + 0.10338404029607773, + 0.06483899801969528, + 0.05882009491324425, + 0.08144974708557129, + 0.05790063366293907, + 0.07310572266578674 + ], + "credit_map": [ + 0.017047060653567314, + 0.019065888598561287, + 0.04858550801873207, + 0.043540019541978836, + 0.027693821117281914, + 0.032981615513563156, + 0.024459103122353554, + 0.02280375361442566, + 0.005755867343395948, + 0.006753535009920597, + 0.006836786400526762, + 0.007252518553286791, + 0.0043659089133143425, + 0.0061646089889109135, + 0.00367256929166615, + 0.004134336952120066, + 0.010696884244680405, + 0.014426562003791332, + 0.028550898656249046, + 0.035508397966623306, + 0.0037360601127147675, + 0.003382429014891386, + 0.00864437036216259, + 0.00956849567592144, + 0.01256046537309885, + 0.01847865991294384, + 0.04212687164545059, + 0.03226582333445549, + 0.010438424535095692, + 0.011810383759438992, + 0.024554621428251266, + 0.012638518586754799, + 0.010060814209282398, + 0.012044697999954224, + 0.016204124316573143, + 0.016695600003004074, + 0.009614965878427029, + 0.012487523257732391, + 0.011002376675605774, + 0.015273927710950375, + 0.005805373657494783, + 0.007871674373745918, + 0.00891862902790308, + 0.008808443322777748, + 0.0067751738242805, + 0.01071838941425085, + 0.008523532189428806, + 0.007417773827910423, + 0.007380925118923187, + 0.009746784344315529, + 0.005207158625125885, + 0.0068235755898058414, + 0.0051233526319265366, + 0.01566980965435505, + 0.0029746373184025288, + 0.005893845111131668, + 0.007946415804326534, + 0.009632456116378307, + 0.008205577731132507, + 0.010073159821331501, + 0.008564223535358906, + 0.023667024448513985, + 0.004250378347933292, + 0.00911050383001566, + 0.008956573903560638, + 0.007669470272958279, + 0.007782028056681156, + 0.00832255557179451, + 0.003556670621037483, + 0.009141407907009125, + 0.004310754127800465, + 0.008161166682839394, + 0.010224657133221626, + 0.010087601840496063, + 0.00964770931750536, + 0.00913073681294918, + 0.005967081058770418, + 0.0078095621429383755, + 0.009257575497031212, + 0.010980797000229359 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.05553646758198738 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive modern wind turbine with gracefully curved blades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does one of the turbine blades violently break off from the central hub at the base?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the detached blade made of a smooth, off-white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to track the detached blade as it plummets toward the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast wind farm of identical turbines rotating against rolling green hills?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07007016241550446 + ], + "opposite_logprob_tokens": [ + -2.695070266723633 + ], + "logprob_reward": -0.07007016241550446, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07007016241550446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0842680111527443, + 0.04200253263115883, + 0.1656341403722763, + 0.04092789813876152, + 0.04272089898586273, + 0.13670890033245087, + 0.16071826219558716, + 0.10682151466608047, + 0.05836980417370796, + 0.16182799637317657 + ], + "credit_map": [ + 0.013749279081821442, + 0.019159091636538506, + 0.011527417227625847, + 0.011655411683022976, + 0.007027797866612673, + 0.00863924901932478, + 0.004317623563110828, + 0.008192135952413082, + 0.0034213706385344267, + 0.0041809710673987865, + 0.007029686588793993, + 0.006169636268168688, + 0.004715966992080212, + 0.007902495563030243, + 0.0039921049028635025, + 0.00459029758349061, + 0.0403887964785099, + 0.045789822936058044, + 0.029322274029254913, + 0.022376004606485367, + 0.005599325988441706, + 0.004962137434631586, + 0.007249907590448856, + 0.009945886209607124, + 0.004252068232744932, + 0.004607572220265865, + 0.0077524385415017605, + 0.0067919776774942875, + 0.004104618914425373, + 0.0029190974310040474, + 0.007451995275914669, + 0.0030481310095638037, + 0.00404141191393137, + 0.004709088243544102, + 0.007888996973633766, + 0.00537925073876977, + 0.0038577511440962553, + 0.004150622524321079, + 0.005502668209373951, + 0.007191111333668232, + 0.01881563290953636, + 0.02104981429874897, + 0.01708434894680977, + 0.013804267160594463, + 0.013077391311526299, + 0.03165167570114136, + 0.008853760547935963, + 0.01237200852483511, + 0.020799599587917328, + 0.025583509355783463, + 0.013360671699047089, + 0.01515183225274086, + 0.022456489503383636, + 0.042298562824726105, + 0.007135148625820875, + 0.013932449743151665, + 0.01005261018872261, + 0.013866259716451168, + 0.012274243868887424, + 0.018558381125330925, + 0.009770010598003864, + 0.022253289818763733, + 0.006729417480528355, + 0.01331730280071497, + 0.006929589435458183, + 0.006431919522583485, + 0.008833794854581356, + 0.007402547635138035, + 0.003600458148866892, + 0.011053895577788353, + 0.0046843355521559715, + 0.009433265775442123, + 0.021097145974636078, + 0.01742289774119854, + 0.024655845016241074, + 0.022451523691415787, + 0.013202435337007046, + 0.022727131843566895, + 0.016216201707720757, + 0.024054819718003273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 6 + ], + "selected_reward": -0.07007016241550446 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive modern wind turbine with gracefully curved blades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does one of the turbine blades violently break off from the central hub at the base?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the detached blade made of a smooth, off-white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to track the detached blade as it plummets toward the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast wind farm of identical turbines rotating against rolling green hills?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01821223832666874 + ], + "opposite_logprob_tokens": [ + -4.01821231842041 + ], + "logprob_reward": -0.01821223832666874, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01821223832666874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14506830275058746, + 0.04587121680378914, + 0.13689468801021576, + 0.15135633945465088, + 0.08511640876531601, + 0.0514727383852005, + 0.07366414368152618, + 0.1112242192029953, + 0.08282912522554398, + 0.11650282889604568 + ], + "credit_map": [ + 0.012073269113898277, + 0.015812119469046593, + 0.02060704492032528, + 0.021242860704660416, + 0.018340371549129486, + 0.02404944971203804, + 0.01665065623819828, + 0.016292531043291092, + 0.0036780585069209337, + 0.0047266120091080666, + 0.0064780875109136105, + 0.005733075551688671, + 0.0054398286156356335, + 0.009338825941085815, + 0.004609816242009401, + 0.005866911727935076, + 0.02716263197362423, + 0.03306848555803299, + 0.026204286143183708, + 0.02892448566854, + 0.0038541913963854313, + 0.0036839551758021116, + 0.004876772873103619, + 0.009119885042309761, + 0.011461974121630192, + 0.01734085939824581, + 0.03188280761241913, + 0.026638060808181763, + 0.012339775450527668, + 0.01374004315584898, + 0.022554323077201843, + 0.015398501418530941, + 0.008300242014229298, + 0.009675128385424614, + 0.012419710867106915, + 0.014104770496487617, + 0.00716137932613492, + 0.009730582125484943, + 0.008734446950256824, + 0.014990153722465038, + 0.005829809233546257, + 0.007207794114947319, + 0.008273333311080933, + 0.006213187240064144, + 0.004984128754585981, + 0.009380603209137917, + 0.00492828106507659, + 0.004655599128454924, + 0.008728621527552605, + 0.008684104308485985, + 0.013717245310544968, + 0.01098496001213789, + 0.005294688045978546, + 0.013434505090117455, + 0.0034651414025574923, + 0.009354882873594761, + 0.008258084766566753, + 0.014952347613871098, + 0.011099250055849552, + 0.02047199197113514, + 0.012598445639014244, + 0.025428906083106995, + 0.006060766987502575, + 0.01235442515462637, + 0.009416581131517887, + 0.009774267673492432, + 0.011616012081503868, + 0.013006795197725296, + 0.0061837430112063885, + 0.012705404311418533, + 0.008217104710638523, + 0.01190921850502491, + 0.013324729166924953, + 0.017140207812190056, + 0.014993301592767239, + 0.011898296885192394, + 0.012011243961751461, + 0.016033291816711426, + 0.015491821803152561, + 0.015609941445291042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.01821223832666874 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive modern wind turbine with gracefully curved blades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does one of the turbine blades violently break off from the central hub at the base?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the detached blade made of a smooth, off-white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to track the detached blade as it plummets toward the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast wind farm of identical turbines rotating against rolling green hills?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006736189126968384 + ], + "opposite_logprob_tokens": [ + -5.0067362785339355 + ], + "logprob_reward": -0.006736189126968384, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006736189126968384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09982828050851822, + 0.03927025571465492, + 0.28326502442359924, + 0.06169562786817551, + 0.030255157500505447, + 0.15964825451374054, + 0.059579502791166306, + 0.06192191317677498, + 0.04333089664578438, + 0.16120506823062897 + ], + "credit_map": [ + 0.009858864359557629, + 0.010155306197702885, + 0.013920395635068417, + 0.015401863493025303, + 0.014945483766496181, + 0.016331113874912262, + 0.009178252890706062, + 0.010036998428404331, + 0.0031527031678706408, + 0.003657506313174963, + 0.006379192229360342, + 0.005082646384835243, + 0.004710120148956776, + 0.007945351302623749, + 0.003611162770539522, + 0.004731574095785618, + 0.044988151639699936, + 0.05921269208192825, + 0.06063920259475708, + 0.08139437437057495, + 0.007376589346677065, + 0.004721402656286955, + 0.01229910645633936, + 0.012633498758077621, + 0.005490467417985201, + 0.006738727912306786, + 0.01240469329059124, + 0.011185099370777607, + 0.00505682872608304, + 0.006187355145812035, + 0.005757877137511969, + 0.008874577470123768, + 0.0038101773243397474, + 0.004121049307286739, + 0.00416357908397913, + 0.00422616908326745, + 0.0033269990235567093, + 0.0026916528586298227, + 0.003763278014957905, + 0.004152250476181507, + 0.02582714892923832, + 0.024140559136867523, + 0.023317674174904823, + 0.013789174146950245, + 0.015404507517814636, + 0.03123653680086136, + 0.011602828279137611, + 0.014329825527966022, + 0.00825970433652401, + 0.0107059795409441, + 0.005482889246195555, + 0.005718611646443605, + 0.006952696479856968, + 0.01490144245326519, + 0.0024277400225400925, + 0.005130438134074211, + 0.007383819669485092, + 0.008609792217612267, + 0.0077091408893466, + 0.008560611866414547, + 0.0056457617320120335, + 0.010926200076937675, + 0.005311092361807823, + 0.007775497157126665, + 0.005337791983038187, + 0.006073552183806896, + 0.005875115282833576, + 0.00561996316537261, + 0.003730383235961199, + 0.0060346173122525215, + 0.004685896448791027, + 0.005973581690341234, + 0.022368988022208214, + 0.02420179173350334, + 0.016483953222632408, + 0.027459124103188515, + 0.01339834462851286, + 0.02399945631623268, + 0.012599224224686623, + 0.020694194361567497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 5 + ], + "selected_reward": -0.006736189126968384 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive modern wind turbine with gracefully curved blades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does one of the turbine blades violently break off from the central hub at the base?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the detached blade made of a smooth, off-white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to track the detached blade as it plummets toward the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast wind farm of identical turbines rotating against rolling green hills?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014189071021974087 + ], + "opposite_logprob_tokens": [ + -4.26418924331665 + ], + "logprob_reward": -0.014189071021974087, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014189071021974087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10811522603034973, + 0.05937608703970909, + 0.1632545292377472, + 0.11117623746395111, + 0.07157523930072784, + 0.12022505700588226, + 0.11922571808099747, + 0.07955639809370041, + 0.08742348849773407, + 0.08007197827100754 + ], + "credit_map": [ + 0.011422174982726574, + 0.012176856398582458, + 0.01876389980316162, + 0.020774925127625465, + 0.012601221911609173, + 0.015300441533327103, + 0.0078060077503323555, + 0.009269695729017258, + 0.005033707246184349, + 0.006563263945281506, + 0.008923768997192383, + 0.0072445981204509735, + 0.006176264490932226, + 0.013621724210679531, + 0.0048599690198898315, + 0.0069527896121144295, + 0.01832551695406437, + 0.025152746587991714, + 0.04041387513279915, + 0.0507856085896492, + 0.0031730076298117638, + 0.003920766990631819, + 0.007340207230299711, + 0.014142811298370361, + 0.008512604050338268, + 0.01185927540063858, + 0.027002407237887383, + 0.021685970947146416, + 0.00594857195392251, + 0.007657772861421108, + 0.017613045871257782, + 0.010896585881710052, + 0.007892745546996593, + 0.00774060795083642, + 0.016131706535816193, + 0.01628751866519451, + 0.0032063822727650404, + 0.0053800493478775024, + 0.006582803092896938, + 0.008353425189852715, + 0.014623413793742657, + 0.016117751598358154, + 0.024061080068349838, + 0.02156839706003666, + 0.009460545144975185, + 0.012642478570342064, + 0.01012913417071104, + 0.01162225566804409, + 0.014733640477061272, + 0.012485131621360779, + 0.02894357219338417, + 0.018027454614639282, + 0.008924415335059166, + 0.014310038648545742, + 0.005700128618627787, + 0.01610133796930313, + 0.010339610278606415, + 0.011543232016265392, + 0.010210627689957619, + 0.011723523028194904, + 0.006740425247699022, + 0.013612356968224049, + 0.004319121595472097, + 0.01106750126928091, + 0.01045464351773262, + 0.016108427196741104, + 0.014188602566719055, + 0.016079455614089966, + 0.004305700771510601, + 0.007962668314576149, + 0.005334526300430298, + 0.012989459559321404, + 0.014060590416193008, + 0.009186558425426483, + 0.006748090498149395, + 0.0171998031437397, + 0.005371264647692442, + 0.01088251918554306, + 0.00609003659337759, + 0.010533117689192295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 6 + ], + "selected_reward": -0.014189071021974087 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_14/index_tensor([2368, 2368])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_14/index_tensor([2368, 2368])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..aa12a2016cb8c2676dcd3f8fe8b79c5ed99c3579 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_14/index_tensor([2368, 2368])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, right-facing, panoramic shot reveals the tranquil scene of a spacious, semi-open wooden workshop. Soft, golden morning sunlight slants through the open side, casting sharp beams of light through the air, shimmering with fine sawdust and dust. At the center of the frame, a massive, unfinished dragon boat rests upon its wooden frame like a sleeping beast. Its hull, crafted from dark solid wood, its intricately carved dragon scales gleaming in the morning light. Several craftsmen, dressed in simple cotton work clothes, are scattered around the boat, their weathered faces etched with focused composure. An elderly craftsman kneels at the bow, meticulously sanding the ferocious, majestic details of the dragon's head with fine sandpaper. Another stands amidships, a small chisel in hand, steadily and deliberately adding the finishing touches to a carved dragon scale. Still another uses a wide brush to evenly apply a layer of transparent tung oil to the side of the boat, imbuing the wood with a rich, vibrant color wherever it passes. The scene is silent, with only the gentle clink of tools on wood. An almost sacred feeling of dedication and reverence for traditional craftsmanship permeates the scene, while the overall tone is warm and earthy.", + "seed": 28327, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_14/index_2368_seed_28327_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03850570283830166, + "vlm_logprob": -0.03850570283830166, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.03850570283830166, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a spacious, semi-open wooden workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, unfinished dragon boat resting on a wooden frame at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are several craftsmen wearing simple cotton work clothes visible around the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is soft, golden morning sunlight slanting through the open side, casting sharp beams through sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slow, right-facing, and panoramic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043674442917108536 + ], + "opposite_logprob_tokens": [ + -3.9186744689941406 + ], + "logprob_reward": -0.043674442917108536, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.043674442917108536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5407306551933289, + 0.04838985204696655, + 0.05224365368485451, + 0.029349112883210182, + 0.05349278822541237, + 0.0733741894364357, + 0.030153371393680573, + 0.03734029084444046, + 0.07583293318748474, + 0.059093158692121506 + ], + "credit_map": [ + 0.05873452499508858, + 0.044934775680303574, + 0.03123491443693638, + 0.18498264253139496, + 0.03061891347169876, + 0.12921717762947083, + 0.028053484857082367, + 0.03295418620109558, + 0.0065614827908575535, + 0.005485034082084894, + 0.003953775390982628, + 0.0050088814459741116, + 0.007080342620611191, + 0.010176247917115688, + 0.005194004159420729, + 0.004930075723677874, + 0.00885110255330801, + 0.005006070714443922, + 0.005041266791522503, + 0.003449898911640048, + 0.0030621341429650784, + 0.005869888700544834, + 0.015132415108382702, + 0.005830874666571617, + 0.003928921185433865, + 0.003967545926570892, + 0.0031235027126967907, + 0.0032861880026757717, + 0.0028276259545236826, + 0.004835156723856926, + 0.004429530818015337, + 0.0029506387654691935, + 0.01052164752036333, + 0.012210343033075333, + 0.005518857855349779, + 0.004796555265784264, + 0.004799878224730492, + 0.006752669811248779, + 0.005415969993919134, + 0.00347686093300581, + 0.005806200206279755, + 0.009801721200346947, + 0.007747261319309473, + 0.004622036125510931, + 0.01976640149950981, + 0.004936713259667158, + 0.01425278652459383, + 0.006441064178943634, + 0.003305762307718396, + 0.0036809761077165604, + 0.0020171459764242172, + 0.0034810903016477823, + 0.0053350538946688175, + 0.005180739797651768, + 0.003239953424781561, + 0.003912647720426321, + 0.003674246370792389, + 0.0033953001257032156, + 0.002582517685368657, + 0.008490214124321938, + 0.004857048392295837, + 0.00583191541954875, + 0.004959282465279102, + 0.0035497620701789856, + 0.0180208757519722, + 0.01564495451748371, + 0.0033541403245180845, + 0.009795181453227997, + 0.011944391764700413, + 0.006648698356002569, + 0.005047036800533533, + 0.005377644672989845, + 0.006521792616695166, + 0.005297227296978235, + 0.005011626519262791, + 0.009516296908259392, + 0.006732095032930374, + 0.00762783782556653, + 0.0076345582492649555, + 0.010751724243164062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.043674442917108536 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a spacious, semi-open wooden workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, unfinished dragon boat resting on a wooden frame at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are several craftsmen wearing simple cotton work clothes visible around the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is soft, golden morning sunlight slanting through the open side, casting sharp beams through sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slow, right-facing, and panoramic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009803364053368568 + ], + "opposite_logprob_tokens": [ + -4.634803295135498 + ], + "logprob_reward": -0.009803364053368568, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009803364053368568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37107089161872864, + 0.050263579934835434, + 0.05731969699263573, + 0.052356258034706116, + 0.06650448590517044, + 0.06523146480321884, + 0.05581265687942505, + 0.16314002871513367, + 0.042494431138038635, + 0.07580656558275223 + ], + "credit_map": [ + 0.0372200608253479, + 0.02806982956826687, + 0.021039700135588646, + 0.12226781249046326, + 0.02033369056880474, + 0.1049705222249031, + 0.016643431037664413, + 0.020525803789496422, + 0.008500219322741032, + 0.007326738443225622, + 0.008102724328637123, + 0.004129183944314718, + 0.005906928796321154, + 0.006472295615822077, + 0.005039988551288843, + 0.004785494413226843, + 0.008019130676984787, + 0.006764277350157499, + 0.004118665121495724, + 0.0033245780505239964, + 0.003452188102528453, + 0.010383917950093746, + 0.015814684331417084, + 0.005442251451313496, + 0.004506838973611593, + 0.003996764775365591, + 0.004233493935316801, + 0.01098574697971344, + 0.0036819414235651493, + 0.015869000926613808, + 0.005231994669884443, + 0.0038504733238369226, + 0.009817344136536121, + 0.020068807527422905, + 0.007735802326351404, + 0.005039242561906576, + 0.006524602882564068, + 0.006502062547951937, + 0.005497741512954235, + 0.005318875890225172, + 0.0052864705212414265, + 0.014499529264867306, + 0.003591950051486492, + 0.003297596238553524, + 0.02073066122829914, + 0.006307102274149656, + 0.00684654526412487, + 0.004671601112931967, + 0.007262921892106533, + 0.004890509881079197, + 0.004481437616050243, + 0.009182737208902836, + 0.007439132779836655, + 0.006816677749156952, + 0.008013752289116383, + 0.0077254800125956535, + 0.02009708434343338, + 0.013257323764264584, + 0.008179188705980778, + 0.03320075199007988, + 0.020858917385339737, + 0.034044891595840454, + 0.018175367265939713, + 0.015326496213674545, + 0.01048193033784628, + 0.004265408031642437, + 0.0021129532251507044, + 0.005154075101017952, + 0.004660012200474739, + 0.006120485719293356, + 0.003634475637227297, + 0.00606508506461978, + 0.005980181973427534, + 0.005151303019374609, + 0.004836518783122301, + 0.015636924654245377, + 0.008183457888662815, + 0.007802903186529875, + 0.007475115358829498, + 0.02074015513062477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.009803364053368568 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a spacious, semi-open wooden workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, unfinished dragon boat resting on a wooden frame at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are several craftsmen wearing simple cotton work clothes visible around the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is soft, golden morning sunlight slanting through the open side, casting sharp beams through sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slow, right-facing, and panoramic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026360128074884415 + ], + "opposite_logprob_tokens": [ + -3.651360034942627 + ], + "logprob_reward": -0.026360128074884415, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026360128074884415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35639992356300354, + 0.10460255295038223, + 0.0476970411837101, + 0.08705746382474899, + 0.23308786749839783, + 0.029303422197699547, + 0.03389566391706467, + 0.03481759876012802, + 0.03303639218211174, + 0.04010207578539848 + ], + "credit_map": [ + 0.03895982354879379, + 0.0303628072142601, + 0.019924765452742577, + 0.1139996349811554, + 0.01942155323922634, + 0.0928235799074173, + 0.016427667811512947, + 0.024480056017637253, + 0.016766177490353584, + 0.019981907680630684, + 0.011565955355763435, + 0.01005688775330782, + 0.008995573967695236, + 0.016614830121397972, + 0.009101992473006248, + 0.011519222520291805, + 0.008506259880959988, + 0.0078081972897052765, + 0.0040010991506278515, + 0.003570618573576212, + 0.003902935655787587, + 0.009921661578118801, + 0.0068335277028381824, + 0.0031527371611446142, + 0.009335172362625599, + 0.006949776783585548, + 0.004468634724617004, + 0.02142404206097126, + 0.0036937519907951355, + 0.03253182768821716, + 0.004843642469495535, + 0.003810607362538576, + 0.027977164834737778, + 0.07760234922170639, + 0.037981871515512466, + 0.015666723251342773, + 0.023197760805487633, + 0.02094333805143833, + 0.018282774835824966, + 0.011435877531766891, + 0.005448338109999895, + 0.004557025618851185, + 0.0028670914471149445, + 0.0028663359116762877, + 0.0050043147057294846, + 0.0037187424022704363, + 0.0028085745871067047, + 0.002032996853813529, + 0.004844650626182556, + 0.0036995168775320053, + 0.0027056019753217697, + 0.006054506171494722, + 0.0051241363398730755, + 0.003802475519478321, + 0.0040056295692920685, + 0.003659144276753068, + 0.0038995358627289534, + 0.0029522283002734184, + 0.0022690878249704838, + 0.004771035630255938, + 0.005652963183820248, + 0.0073266709223389626, + 0.0042562903836369514, + 0.0036897845566272736, + 0.0064736101776361465, + 0.005027368199080229, + 0.002147240098565817, + 0.0027808677405118942, + 0.005179394036531448, + 0.003718375228345394, + 0.003327307989820838, + 0.00438222661614418, + 0.004127703607082367, + 0.0040113311260938644, + 0.003957677632570267, + 0.007467356510460377, + 0.0036976435221731663, + 0.0047029429115355015, + 0.00688853207975626, + 0.005248886998742819 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.026360128074884415 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a spacious, semi-open wooden workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, unfinished dragon boat resting on a wooden frame at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are several craftsmen wearing simple cotton work clothes visible around the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is soft, golden morning sunlight slanting through the open side, casting sharp beams through sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slow, right-facing, and panoramic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033721957355737686 + ], + "opposite_logprob_tokens": [ + -3.408721923828125 + ], + "logprob_reward": -0.033721957355737686, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033721957355737686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5551098585128784, + 0.05389692261815071, + 0.05777842923998833, + 0.06592324376106262, + 0.05086348578333855, + 0.07610058039426804, + 0.012925826013088226, + 0.03656815364956856, + 0.05650842934846878, + 0.03432510048151016 + ], + "credit_map": [ + 0.05650288239121437, + 0.04288215935230255, + 0.03065970167517662, + 0.19349049031734467, + 0.026283638551831245, + 0.13874869048595428, + 0.026345688849687576, + 0.04019654169678688, + 0.006126849912106991, + 0.005986866541206837, + 0.004278563894331455, + 0.0068520475178956985, + 0.007196096703410149, + 0.013310138136148453, + 0.005141337402164936, + 0.005005019251257181, + 0.014515742659568787, + 0.005944764241576195, + 0.004764479584991932, + 0.004013729281723499, + 0.004305717069655657, + 0.007692019920796156, + 0.01248132809996605, + 0.004060644656419754, + 0.006652156822383404, + 0.004273165483027697, + 0.004648817703127861, + 0.015344366431236267, + 0.004555804654955864, + 0.021038377657532692, + 0.006093472708016634, + 0.0033170776441693306, + 0.007705568801611662, + 0.011379574425518513, + 0.011172939091920853, + 0.003503199201077223, + 0.004205218516290188, + 0.004975213203579187, + 0.004762283526360989, + 0.0031594857573509216, + 0.009037250652909279, + 0.007862239144742489, + 0.010842649266123772, + 0.011552107520401478, + 0.011160655878484249, + 0.008816910907626152, + 0.011137175373733044, + 0.00569158373400569, + 0.0015858730766922235, + 0.001723777735605836, + 0.0012760190293192863, + 0.002019779523834586, + 0.0020731682889163494, + 0.0017088738968595862, + 0.00116306624840945, + 0.0013752675149589777, + 0.003760323394089937, + 0.0033593361731618643, + 0.0031086993403732777, + 0.005076362285763025, + 0.005068627651780844, + 0.00786847248673439, + 0.004798388108611107, + 0.0035279442090541124, + 0.012893235310912132, + 0.009325433522462845, + 0.0028883488848805428, + 0.008906913921236992, + 0.008309757336974144, + 0.004817745182663202, + 0.004749781917780638, + 0.004617209546267986, + 0.0041567487642169, + 0.0036975222174078226, + 0.002442855155095458, + 0.006760319694876671, + 0.003482433967292309, + 0.0028896848671138287, + 0.004310579039156437, + 0.0065849544480443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.033721957355737686 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a spacious, semi-open wooden workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, unfinished dragon boat resting on a wooden frame at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are several craftsmen wearing simple cotton work clothes visible around the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is soft, golden morning sunlight slanting through the open side, casting sharp beams through sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slow, right-facing, and panoramic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896862179040909 + ], + "opposite_logprob_tokens": [ + -2.5789685249328613 + ], + "logprob_reward": -0.07896862179040909, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896862179040909, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09652850776910782, + 0.04664259031414986, + 0.10032907128334045, + 0.08202251046895981, + 0.042172037065029144, + 0.08340376615524292, + 0.07769740372896194, + 0.0655413419008255, + 0.27789321541786194, + 0.12776954472064972 + ], + "credit_map": [ + 0.01338442973792553, + 0.008275406435132027, + 0.006988781038671732, + 0.021960467100143433, + 0.00899803638458252, + 0.025213418528437614, + 0.005436201114207506, + 0.006271769292652607, + 0.0062705716118216515, + 0.007374603301286697, + 0.004410783760249615, + 0.00506417453289032, + 0.004681739956140518, + 0.008786234073340893, + 0.006107127759605646, + 0.003947355784475803, + 0.02592516504228115, + 0.009804656729102135, + 0.010473429225385189, + 0.008198131807148457, + 0.006780652794986963, + 0.013610047288239002, + 0.017876731231808662, + 0.007660267874598503, + 0.009290060959756374, + 0.005375120788812637, + 0.00511199003085494, + 0.02160048857331276, + 0.0030643539503216743, + 0.026864156126976013, + 0.005623575299978256, + 0.005092771258205175, + 0.004581833723932505, + 0.010446375235915184, + 0.007294372655451298, + 0.0027246512472629547, + 0.0061995722353458405, + 0.004473924171179533, + 0.003970972262322903, + 0.002480340888723731, + 0.007693177554756403, + 0.015663355588912964, + 0.010883480310440063, + 0.00961283314973116, + 0.017593003809452057, + 0.005320496391505003, + 0.008853532373905182, + 0.007783886510878801, + 0.010737505741417408, + 0.00969003327190876, + 0.006203535478562117, + 0.008151774294674397, + 0.011185201816260815, + 0.00928949099034071, + 0.011753777973353863, + 0.010686087422072887, + 0.0072953347116708755, + 0.006172260269522667, + 0.004536577500402927, + 0.008852163329720497, + 0.009700734168291092, + 0.013769055716693401, + 0.008588641881942749, + 0.006626582238823175, + 0.05805585905909538, + 0.029253793880343437, + 0.013913379050791264, + 0.07522391527891159, + 0.033027853816747665, + 0.01799018122255802, + 0.020932717248797417, + 0.029495542868971825, + 0.016780462116003036, + 0.011000184342265129, + 0.0096949003636837, + 0.029271965846419334, + 0.010154606774449348, + 0.009698762558400631, + 0.014019741676747799, + 0.027148926630616188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 2 + ], + "selected_reward": -0.07896862179040909 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, right-facing, panoramic shot reveals the tranquil scene of a spacious, semi-open wooden workshop. Soft, golden morning sunlight slants through the open side, casting sharp beams of light through the air, shimmering with fine sawdust and dust. At the center of the frame, a massive, unfinished dragon boat rests upon its wooden frame like a sleeping beast. Its hull, crafted from dark solid wood, its intricately carved dragon scales gleaming in the morning light. Several craftsmen, dressed in simple cotton work clothes, are scattered around the boat, their weathered faces etched with focused composure. An elderly craftsman kneels at the bow, meticulously sanding the ferocious, majestic details of the dragon's head with fine sandpaper. Another stands amidships, a small chisel in hand, steadily and deliberately adding the finishing touches to a carved dragon scale. Still another uses a wide brush to evenly apply a layer of transparent tung oil to the side of the boat, imbuing the wood with a rich, vibrant color wherever it passes. The scene is silent, with only the gentle clink of tools on wood. An almost sacred feeling of dedication and reverence for traditional craftsmanship permeates the scene, while the overall tone is warm and earthy.", + "seed": 28328, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_14/index_2368_seed_28328_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.044568336196243764, + "vlm_logprob": -0.044568336196243764, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.044568336196243764, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a spacious, semi-open wooden workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, unfinished dragon boat resting on a wooden frame at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are several craftsmen wearing simple cotton work clothes visible around the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is soft, golden morning sunlight slanting through the open side, casting sharp beams through sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slow, right-facing, and panoramic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02015538327395916 + ], + "opposite_logprob_tokens": [ + -4.895155429840088 + ], + "logprob_reward": -0.02015538327395916, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.02015538327395916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07555466890335083, + 0.026581546291708946, + 0.019911957904696465, + 0.07609443366527557, + 0.06994147598743439, + 0.0627821609377861, + 0.20872974395751953, + 0.2153952270746231, + 0.20315814018249512, + 0.04185071960091591 + ], + "credit_map": [ + 0.011368113569915295, + 0.007499732542783022, + 0.005495030898600817, + 0.005642246920615435, + 0.020371215417981148, + 0.008231784217059612, + 0.005432691425085068, + 0.011513848789036274, + 0.0036324160173535347, + 0.004305660724639893, + 0.002149044768884778, + 0.00277679949067533, + 0.0036425774451345205, + 0.0023976375814527273, + 0.0039452072232961655, + 0.0037321976851671934, + 0.0031129512935876846, + 0.002396397525444627, + 0.0021563260816037655, + 0.001692955382168293, + 0.0019449194660410285, + 0.0017663835315033793, + 0.003381055546924472, + 0.003460968378931284, + 0.008935139514505863, + 0.013327404856681824, + 0.004007673356682062, + 0.007341062184423208, + 0.007547563407570124, + 0.005565057508647442, + 0.0153562156483531, + 0.014014308340847492, + 0.00932155828922987, + 0.008674930781126022, + 0.006469335872679949, + 0.0054938155226409435, + 0.005465674679726362, + 0.0094193359836936, + 0.011587473563849926, + 0.013509341515600681, + 0.008646691218018532, + 0.006691552698612213, + 0.006555110216140747, + 0.007029017433524132, + 0.006902200635522604, + 0.006884091533720493, + 0.010204420424997807, + 0.009869074448943138, + 0.013477394357323647, + 0.031049078330397606, + 0.005209935829043388, + 0.011645362712442875, + 0.009897003881633282, + 0.009327175095677376, + 0.07534724473953247, + 0.05277653783559799, + 0.01878662407398224, + 0.016726631671190262, + 0.010571549646556377, + 0.012308900244534016, + 0.009013255126774311, + 0.015863999724388123, + 0.06586231291294098, + 0.06626194715499878, + 0.02244739793241024, + 0.009502431377768517, + 0.03138657659292221, + 0.022863132879137993, + 0.018467653542757034, + 0.011513066478073597, + 0.06074479967355728, + 0.02623305656015873, + 0.005394366104155779, + 0.004226652905344963, + 0.0025137451011687517, + 0.006937377154827118, + 0.00413980707526207, + 0.0051065185107290745, + 0.007270219270139933, + 0.006262032315135002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.02015538327395916 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a spacious, semi-open wooden workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, unfinished dragon boat resting on a wooden frame at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are several craftsmen wearing simple cotton work clothes visible around the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is soft, golden morning sunlight slanting through the open side, casting sharp beams through sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slow, right-facing, and panoramic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008652104064822197 + ], + "opposite_logprob_tokens": [ + -4.758652210235596 + ], + "logprob_reward": -0.008652104064822197, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008652104064822197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0571112185716629, + 0.047460515052080154, + 0.020054738968610764, + 0.07763656228780746, + 0.1205662339925766, + 0.08060326427221298, + 0.07473929971456528, + 0.23295888304710388, + 0.16314341127872467, + 0.1257258802652359 + ], + "credit_map": [ + 0.008382013067603111, + 0.005104333162307739, + 0.006005673669278622, + 0.004968633875250816, + 0.009668475016951561, + 0.007377726957201958, + 0.006355323828756809, + 0.009249038994312286, + 0.008385349065065384, + 0.004150466062128544, + 0.004189407918602228, + 0.006012337747961283, + 0.00724109448492527, + 0.006134999915957451, + 0.003938628826290369, + 0.00740823382511735, + 0.002245200565084815, + 0.002892018062993884, + 0.0021492044907063246, + 0.0017106040613725781, + 0.0025263044517487288, + 0.0018655332969501615, + 0.003713322337716818, + 0.0029525517020374537, + 0.008326002396643162, + 0.014205081388354301, + 0.005584309343248606, + 0.008138702251017094, + 0.008833909407258034, + 0.0059148939326405525, + 0.014265337027609348, + 0.012368329800665379, + 0.01502738706767559, + 0.016760211437940598, + 0.013340839184820652, + 0.01081730518490076, + 0.013933577574789524, + 0.013928787782788277, + 0.01855534501373768, + 0.018202777951955795, + 0.009260344319045544, + 0.0116443345323205, + 0.005923353601247072, + 0.0068803648464381695, + 0.011742762289941311, + 0.008695898577570915, + 0.014632218517363071, + 0.011823986656963825, + 0.005503661930561066, + 0.00981920026242733, + 0.004463484976440668, + 0.005338233429938555, + 0.007957523688673973, + 0.007072212640196085, + 0.020278053358197212, + 0.0143069326877594, + 0.017729677259922028, + 0.016823511570692062, + 0.01333906315267086, + 0.015347514301538467, + 0.009958307258784771, + 0.018066078424453735, + 0.07075518369674683, + 0.0709395557641983, + 0.010058307088911533, + 0.006892262026667595, + 0.011650236323475838, + 0.02407902665436268, + 0.014340926893055439, + 0.00865807756781578, + 0.06269695609807968, + 0.024767613038420677, + 0.007870713248848915, + 0.009313182905316353, + 0.008512335829436779, + 0.02790192887187004, + 0.016061943024396896, + 0.016085008159279823, + 0.019387900829315186, + 0.02059287577867508 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.008652104064822197 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a spacious, semi-open wooden workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, unfinished dragon boat resting on a wooden frame at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are several craftsmen wearing simple cotton work clothes visible around the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is soft, golden morning sunlight slanting through the open side, casting sharp beams through sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slow, right-facing, and panoramic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294639855623245 + ], + "opposite_logprob_tokens": [ + -2.2379465103149414 + ], + "logprob_reward": -0.11294639855623245, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294639855623245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07491128891706467, + 0.053309954702854156, + 0.03671742230653763, + 0.050804369151592255, + 0.03428150340914726, + 0.03806021064519882, + 0.09101899713277817, + 0.07052024453878403, + 0.34415602684020996, + 0.20622001588344574 + ], + "credit_map": [ + 0.007640328258275986, + 0.014126242138445377, + 0.007018653675913811, + 0.0076951803639531136, + 0.009048622101545334, + 0.011522870510816574, + 0.00689062150195241, + 0.0109687689691782, + 0.0046679796651005745, + 0.008187818340957165, + 0.004038582090288401, + 0.004236011765897274, + 0.011371501721441746, + 0.007991146296262741, + 0.004596616607159376, + 0.008220297284424305, + 0.0038491631858050823, + 0.004892919212579727, + 0.0020797515753656626, + 0.002981898607686162, + 0.004864020738750696, + 0.0034054205752909184, + 0.008231365121901035, + 0.006412877701222897, + 0.004650832619518042, + 0.009362964890897274, + 0.0027254687156528234, + 0.005407399497926235, + 0.005248354747891426, + 0.003673101309686899, + 0.010597283020615578, + 0.009138960391283035, + 0.0038296375423669815, + 0.002894757315516472, + 0.003677941160276532, + 0.004183790180832148, + 0.0026029956061393023, + 0.008791914209723473, + 0.003965700976550579, + 0.0043347617611289024, + 0.0037192420568317175, + 0.004504384007304907, + 0.003927426412701607, + 0.003340394701808691, + 0.006209284067153931, + 0.005080254748463631, + 0.005377637222409248, + 0.005901583470404148, + 0.0061844270676374435, + 0.015697188675403595, + 0.002993290778249502, + 0.00465073436498642, + 0.005085863638669252, + 0.0051970211789011955, + 0.030504195019602776, + 0.020706279203295708, + 0.006085079163312912, + 0.0062645236030220985, + 0.007649267092347145, + 0.008381095714867115, + 0.005032849032431841, + 0.010102549567818642, + 0.015217612497508526, + 0.01178726926445961, + 0.03481869027018547, + 0.01315684337168932, + 0.04950443655252457, + 0.04061450436711311, + 0.03156784921884537, + 0.009691144339740276, + 0.11809377372264862, + 0.04670877382159233, + 0.012151454575359821, + 0.017183268442749977, + 0.012444334104657173, + 0.04229887202382088, + 0.0307887252420187, + 0.027325792238116264, + 0.03214528039097786, + 0.03188228979706764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.11294639855623245 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a spacious, semi-open wooden workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, unfinished dragon boat resting on a wooden frame at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are several craftsmen wearing simple cotton work clothes visible around the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is soft, golden morning sunlight slanting through the open side, casting sharp beams through sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slow, right-facing, and panoramic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011082758195698261 + ], + "opposite_logprob_tokens": [ + -4.511082649230957 + ], + "logprob_reward": -0.011082758195698261, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011082758195698261, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10314126312732697, + 0.0894746482372284, + 0.16002938151359558, + 0.06184905767440796, + 0.03935978189110756, + 0.036994460970163345, + 0.18151551485061646, + 0.10101746767759323, + 0.14100302755832672, + 0.08561540395021439 + ], + "credit_map": [ + 0.01993437670171261, + 0.009545442648231983, + 0.009347205050289631, + 0.009091398678719997, + 0.020775925368070602, + 0.013626707717776299, + 0.008471347391605377, + 0.012348857708275318, + 0.011148623190820217, + 0.014541730284690857, + 0.008756694383919239, + 0.007351174484938383, + 0.01415884681046009, + 0.008340596221387386, + 0.010023476555943489, + 0.015153505839407444, + 0.022282971069216728, + 0.021757764741778374, + 0.01632782444357872, + 0.014619169756770134, + 0.018612846732139587, + 0.012482671067118645, + 0.027069909498095512, + 0.026876237243413925, + 0.007864964194595814, + 0.009341387078166008, + 0.00461761187762022, + 0.007110999897122383, + 0.0070856367237865925, + 0.005488026421517134, + 0.010672204196453094, + 0.009668226353824139, + 0.006697668228298426, + 0.0053045847453176975, + 0.003511607414111495, + 0.003600157331675291, + 0.003938994836062193, + 0.005711589474231005, + 0.005786719731986523, + 0.004808463156223297, + 0.0033527310006320477, + 0.0044079190120100975, + 0.003620396601036191, + 0.004266299773007631, + 0.0043372586369514465, + 0.005135235842317343, + 0.005869950167834759, + 0.00600466737523675, + 0.012059863656759262, + 0.025944005697965622, + 0.006611277349293232, + 0.00863688439130783, + 0.011145158670842648, + 0.009857768192887306, + 0.06293080747127533, + 0.04432976245880127, + 0.005152804311364889, + 0.006803946569561958, + 0.0036624588537961245, + 0.005712445825338364, + 0.004533612169325352, + 0.012368481606245041, + 0.03365128114819527, + 0.02913244068622589, + 0.015488681383430958, + 0.00758626451715827, + 0.006215857807546854, + 0.02210988849401474, + 0.012376208789646626, + 0.010396038182079792, + 0.04791770875453949, + 0.018912386149168015, + 0.006123041268438101, + 0.010313100181519985, + 0.005602099467068911, + 0.01626468263566494, + 0.008186355233192444, + 0.01321351993829012, + 0.012363684363663197, + 0.01354891899973154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 8 + ], + "selected_reward": -0.011082758195698261 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a spacious, semi-open wooden workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, unfinished dragon boat resting on a wooden frame at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are several craftsmen wearing simple cotton work clothes visible around the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is soft, golden morning sunlight slanting through the open side, casting sharp beams through sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slow, right-facing, and panoramic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000503689050674 + ], + "opposite_logprob_tokens": [ + -2.695004940032959 + ], + "logprob_reward": -0.07000503689050674, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000503689050674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07466559112071991, + 0.023120854049921036, + 0.02247537299990654, + 0.07461170107126236, + 0.07395510375499725, + 0.06262719631195068, + 0.07201169431209564, + 0.2859192490577698, + 0.2585325539112091, + 0.052080657333135605 + ], + "credit_map": [ + 0.009850031696259975, + 0.012571865692734718, + 0.004701542668044567, + 0.006519767455756664, + 0.014601749368011951, + 0.008324647322297096, + 0.006843128707259893, + 0.011252861469984055, + 0.0025560331996530294, + 0.002550347475335002, + 0.0018256818875670433, + 0.0028319109696894884, + 0.0033989385701715946, + 0.0019807692151516676, + 0.00400256784632802, + 0.003974605817347765, + 0.004415622912347317, + 0.0028663103003054857, + 0.0017891336465254426, + 0.0018383195856586099, + 0.0025888807140290737, + 0.0022617788054049015, + 0.003227082546800375, + 0.0034882433246821165, + 0.00837809219956398, + 0.013047209940850735, + 0.00466956989839673, + 0.008783386088907719, + 0.008156470954418182, + 0.005779385566711426, + 0.01394839771091938, + 0.011849189177155495, + 0.013608873821794987, + 0.011297944001853466, + 0.006282479967921972, + 0.005459760781377554, + 0.010672977194190025, + 0.008435508236289024, + 0.009923807345330715, + 0.008273757062852383, + 0.006681917235255241, + 0.005702642258256674, + 0.006869290955364704, + 0.0062018875032663345, + 0.006123565603047609, + 0.007778450846672058, + 0.010410821996629238, + 0.0128586171194911, + 0.007053027860820293, + 0.010746260173618793, + 0.004804233089089394, + 0.006214871536940336, + 0.007081473246216774, + 0.006743457168340683, + 0.016559667885303497, + 0.012808708474040031, + 0.016993606463074684, + 0.014077296480536461, + 0.008752623572945595, + 0.01521360781043768, + 0.011874646879732609, + 0.02316812053322792, + 0.09746742248535156, + 0.09837191551923752, + 0.03236094117164612, + 0.015455946326255798, + 0.0398576483130455, + 0.02859395183622837, + 0.021434077993035316, + 0.009811951778829098, + 0.0773700550198555, + 0.033647987991571426, + 0.00395340658724308, + 0.004287755116820335, + 0.0038075640331953764, + 0.009454336017370224, + 0.004946044646203518, + 0.008029203861951828, + 0.009218195453286171, + 0.008384153246879578 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.07000503689050674 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, right-facing, panoramic shot reveals the tranquil scene of a spacious, semi-open wooden workshop. Soft, golden morning sunlight slants through the open side, casting sharp beams of light through the air, shimmering with fine sawdust and dust. At the center of the frame, a massive, unfinished dragon boat rests upon its wooden frame like a sleeping beast. Its hull, crafted from dark solid wood, its intricately carved dragon scales gleaming in the morning light. Several craftsmen, dressed in simple cotton work clothes, are scattered around the boat, their weathered faces etched with focused composure. An elderly craftsman kneels at the bow, meticulously sanding the ferocious, majestic details of the dragon's head with fine sandpaper. Another stands amidships, a small chisel in hand, steadily and deliberately adding the finishing touches to a carved dragon scale. Still another uses a wide brush to evenly apply a layer of transparent tung oil to the side of the boat, imbuing the wood with a rich, vibrant color wherever it passes. The scene is silent, with only the gentle clink of tools on wood. An almost sacred feeling of dedication and reverence for traditional craftsmanship permeates the scene, while the overall tone is warm and earthy.", + "seed": 28329, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_14/index_2368_seed_28329_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.044520852994173764, + "vlm_logprob": -0.044520852994173764, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.044520852994173764, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a spacious, semi-open wooden workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, unfinished dragon boat resting on a wooden frame at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are several craftsmen wearing simple cotton work clothes visible around the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is soft, golden morning sunlight slanting through the open side, casting sharp beams through sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slow, right-facing, and panoramic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020134001970291138 + ], + "opposite_logprob_tokens": [ + -5.145133972167969 + ], + "logprob_reward": -0.020134001970291138, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.020134001970291138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12561552226543427, + 0.08070315420627594, + 0.033876821398735046, + 0.18367208540439606, + 0.022310059517621994, + 0.029274050146341324, + 0.09360034018754959, + 0.04163992777466774, + 0.33695316314697266, + 0.0523548386991024 + ], + "credit_map": [ + 0.013773598708212376, + 0.01980026811361313, + 0.013287921436131, + 0.013155476190149784, + 0.01885478012263775, + 0.020201077684760094, + 0.013259399682283401, + 0.013283012434840202, + 0.008249932900071144, + 0.015125405043363571, + 0.013156708329916, + 0.008857264183461666, + 0.007569551933556795, + 0.011618453077971935, + 0.008576074615120888, + 0.00754977110773325, + 0.003869416657835245, + 0.005220450460910797, + 0.0037979246117174625, + 0.003991809207946062, + 0.0042602564208209515, + 0.005512417294085026, + 0.0030370382592082024, + 0.0041875094175338745, + 0.021471520885825157, + 0.020418697968125343, + 0.025847947224974632, + 0.02861151285469532, + 0.02265912853181362, + 0.02281622402369976, + 0.024359039962291718, + 0.017488017678260803, + 0.002684428123757243, + 0.0020040934905409813, + 0.002632289892062545, + 0.0027823359705507755, + 0.004063121508806944, + 0.003469842718914151, + 0.002520870417356491, + 0.0021530792582780123, + 0.003583868034183979, + 0.004768977407366037, + 0.0038363090716302395, + 0.0030414306093007326, + 0.004146391060203314, + 0.0032358733005821705, + 0.003976948093622923, + 0.0026842548977583647, + 0.010030646808445454, + 0.0042952438816428185, + 0.02631280943751335, + 0.008309673517942429, + 0.006075944751501083, + 0.02754765748977661, + 0.005071302410215139, + 0.005957063287496567, + 0.005749281495809555, + 0.005693993531167507, + 0.008204598911106586, + 0.004037005361169577, + 0.004644600208848715, + 0.004695984534919262, + 0.004490779712796211, + 0.0041236900724470615, + 0.08214348554611206, + 0.025498244911432266, + 0.01200850959867239, + 0.021045874804258347, + 0.13763922452926636, + 0.026675522327423096, + 0.01962330751121044, + 0.012319035828113556, + 0.007395618595182896, + 0.012030217796564102, + 0.005927581340074539, + 0.005123895592987537, + 0.008404691703617573, + 0.005292586516588926, + 0.003281872719526291, + 0.004898376297205687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.020134001970291138 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a spacious, semi-open wooden workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, unfinished dragon boat resting on a wooden frame at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are several craftsmen wearing simple cotton work clothes visible around the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is soft, golden morning sunlight slanting through the open side, casting sharp beams through sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slow, right-facing, and panoramic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046483553014695644 + ], + "opposite_logprob_tokens": [ + -5.379648208618164 + ], + "logprob_reward": -0.0046483553014695644, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046483553014695644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17761169373989105, + 0.09462333470582962, + 0.09967810660600662, + 0.025557931512594223, + 0.030459437519311905, + 0.02933364175260067, + 0.15382587909698486, + 0.07831728458404541, + 0.175561785697937, + 0.13503088057041168 + ], + "credit_map": [ + 0.024616839364171028, + 0.025617238134145737, + 0.03156175836920738, + 0.016715707257390022, + 0.02224235236644745, + 0.02235572226345539, + 0.016648536548018456, + 0.01785355620086193, + 0.010401356033980846, + 0.014408658258616924, + 0.013792573474347591, + 0.012183290906250477, + 0.009519579820334911, + 0.01020536758005619, + 0.013074098154902458, + 0.011038405820727348, + 0.0081716263666749, + 0.020426413044333458, + 0.012362908571958542, + 0.009728691540658474, + 0.00997913721948862, + 0.013667654246091843, + 0.013419294729828835, + 0.011922379024326801, + 0.0021372849587351084, + 0.0032369440887123346, + 0.0033541289158165455, + 0.003468662267550826, + 0.004726325161755085, + 0.003204319393262267, + 0.002728288061916828, + 0.0027019765693694353, + 0.0032549931202083826, + 0.0021833819337189198, + 0.004009608644992113, + 0.0033886516466736794, + 0.0066685485653579235, + 0.003544674487784505, + 0.0035659978166222572, + 0.0038435794413089752, + 0.006553344428539276, + 0.003243915271013975, + 0.0034157978370785713, + 0.00295299687422812, + 0.0033668389078229666, + 0.002309487434104085, + 0.004290667362511158, + 0.0032005927059799433, + 0.029216911643743515, + 0.007558288984000683, + 0.017941752448678017, + 0.009541952982544899, + 0.009690689854323864, + 0.05925248563289642, + 0.00877052079886198, + 0.011853273957967758, + 0.011963127180933952, + 0.007644449360668659, + 0.015804721042513847, + 0.007458824198693037, + 0.00954041164368391, + 0.012011867947876453, + 0.005745085421949625, + 0.008148795925080776, + 0.03948870673775673, + 0.018257366493344307, + 0.010686753317713737, + 0.011045028455555439, + 0.062491126358509064, + 0.013939963653683662, + 0.01015833392739296, + 0.009494509547948837, + 0.014452679082751274, + 0.02169612981379032, + 0.013315247371792793, + 0.0197320394217968, + 0.01821977086365223, + 0.012305723503232002, + 0.015315831638872623, + 0.019993450492620468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.0046483553014695644 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a spacious, semi-open wooden workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, unfinished dragon boat resting on a wooden frame at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are several craftsmen wearing simple cotton work clothes visible around the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is soft, golden morning sunlight slanting through the open side, casting sharp beams through sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slow, right-facing, and panoramic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014198355376720428 + ], + "opposite_logprob_tokens": [ + -4.264198303222656 + ], + "logprob_reward": -0.014198355376720428, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014198355376720428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11958838999271393, + 0.08149446547031403, + 0.03749830648303032, + 0.14777186512947083, + 0.02221127413213253, + 0.12956666946411133, + 0.10475796461105347, + 0.029848886653780937, + 0.2642271816730499, + 0.06303500384092331 + ], + "credit_map": [ + 0.0156256090849638, + 0.022299304604530334, + 0.017465928569436073, + 0.010668163187801838, + 0.017668887972831726, + 0.02014971897006035, + 0.00823940895497799, + 0.007471373304724693, + 0.009617445059120655, + 0.013686947524547577, + 0.014797125943005085, + 0.009135032072663307, + 0.007911410182714462, + 0.009951728396117687, + 0.008032695390284061, + 0.008362087421119213, + 0.0027007474564015865, + 0.007857423275709152, + 0.004995493683964014, + 0.005979443900287151, + 0.0031157969497144222, + 0.003701719455420971, + 0.005201869178563356, + 0.0039458139799535275, + 0.01724686473608017, + 0.01603672094643116, + 0.020741496235132217, + 0.023910503834486008, + 0.018978018313646317, + 0.01770758256316185, + 0.019502965733408928, + 0.013647724874317646, + 0.00243955641053617, + 0.0017501016845926642, + 0.0031402939930558205, + 0.0016741511644795537, + 0.005296811461448669, + 0.0032768198288977146, + 0.0016076525207608938, + 0.0030258879996836185, + 0.01935710944235325, + 0.024913467466831207, + 0.015272908844053745, + 0.014156244695186615, + 0.01770157739520073, + 0.016516322270035744, + 0.01082462165504694, + 0.010824423283338547, + 0.026962628588080406, + 0.005514285992830992, + 0.005739374551922083, + 0.005974954925477505, + 0.0060525620356202126, + 0.04037966579198837, + 0.006141049321740866, + 0.007993442937731743, + 0.0038422008510679007, + 0.002437060233205557, + 0.006402033381164074, + 0.0036824671551585197, + 0.003628387348726392, + 0.003790866583585739, + 0.0031237369403243065, + 0.0029421329963952303, + 0.0667482390999794, + 0.02040713280439377, + 0.010031812824308872, + 0.0172882042825222, + 0.10276566445827484, + 0.02194349281489849, + 0.015504385344684124, + 0.009538233280181885, + 0.007663878612220287, + 0.0110702533274889, + 0.004975516814738512, + 0.008023926056921482, + 0.010259988717734814, + 0.005986422766000032, + 0.006709728855639696, + 0.008345287293195724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 5 + ], + "selected_reward": -0.014198355376720428 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a spacious, semi-open wooden workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, unfinished dragon boat resting on a wooden frame at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are several craftsmen wearing simple cotton work clothes visible around the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is soft, golden morning sunlight slanting through the open side, casting sharp beams through sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slow, right-facing, and panoramic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329261600971222 + ], + "opposite_logprob_tokens": [ + -3.7732925415039062 + ], + "logprob_reward": -0.02329261600971222, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329261600971222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049211371690034866, + 0.0804145485162735, + 0.023271741345524788, + 0.15879380702972412, + 0.007833858951926231, + 0.03266623243689537, + 0.05395837873220444, + 0.015545892529189587, + 0.5634999871253967, + 0.014804182574152946 + ], + "credit_map": [ + 0.006108920089900494, + 0.01011303998529911, + 0.005979739595204592, + 0.004104251973330975, + 0.006418450269848108, + 0.007332551293075085, + 0.004232700914144516, + 0.004921718966215849, + 0.0086772246286273, + 0.014598154462873936, + 0.015042409300804138, + 0.00970218051224947, + 0.006800011731684208, + 0.009282206185162067, + 0.008312738500535488, + 0.00799962505698204, + 0.0016386541537940502, + 0.004920835141092539, + 0.0031265162397176027, + 0.0037227661814540625, + 0.00206832611002028, + 0.0019338004058226943, + 0.003391704522073269, + 0.0024691412691026926, + 0.018069157376885414, + 0.018466848880052567, + 0.020145447924733162, + 0.02508142590522766, + 0.020065201446413994, + 0.019578006118535995, + 0.021909505128860474, + 0.01547821145504713, + 0.0009743456030264497, + 0.0009130890830419958, + 0.0007657010573893785, + 0.0006742965779267251, + 0.0015165430959314108, + 0.0010258378461003304, + 0.0007028294494375587, + 0.0012612162390723825, + 0.004377175122499466, + 0.005558455362915993, + 0.0035238959826529026, + 0.0033510602079331875, + 0.005013554356992245, + 0.005052532069385052, + 0.002592111472040415, + 0.0031974490266293287, + 0.012199794873595238, + 0.002860577777028084, + 0.004551305435597897, + 0.003357881912961602, + 0.003433002158999443, + 0.02072790078818798, + 0.0028589663561433554, + 0.003968954086303711, + 0.00199255533516407, + 0.001560199074447155, + 0.002428200328722596, + 0.002215223852545023, + 0.0019933769945055246, + 0.001458832761272788, + 0.00234327488578856, + 0.0015542299952358007, + 0.13587313890457153, + 0.04299505800008774, + 0.020209942013025284, + 0.036136917769908905, + 0.22717499732971191, + 0.04958166182041168, + 0.033648546785116196, + 0.017879800871014595, + 0.0015720159281045198, + 0.0023681847378611565, + 0.002107155043631792, + 0.0018239484634250402, + 0.0019415895221754909, + 0.0015399340773001313, + 0.0015155663713812828, + 0.0019357884302735329 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 1 + ], + "selected_reward": -0.02329261600971222 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a spacious, semi-open wooden workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, unfinished dragon boat resting on a wooden frame at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are several craftsmen wearing simple cotton work clothes visible around the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is soft, golden morning sunlight slanting through the open side, casting sharp beams through sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slow, right-facing, and panoramic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16033093631267548 + ], + "opposite_logprob_tokens": [ + -1.910330891609192 + ], + "logprob_reward": -0.16033093631267548, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16033093631267548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1295887976884842, + 0.07535336911678314, + 0.07033468037843704, + 0.08314613997936249, + 0.028309009969234467, + 0.038684237748384476, + 0.25425636768341064, + 0.06348402798175812, + 0.2120533436536789, + 0.044790033251047134 + ], + "credit_map": [ + 0.013567137531936169, + 0.020381759852170944, + 0.016226887702941895, + 0.011558052152395248, + 0.019531283527612686, + 0.02405993826687336, + 0.011954234912991524, + 0.012309501878917217, + 0.007024993654340506, + 0.012387270107865334, + 0.010627562180161476, + 0.008406422100961208, + 0.010920108295977116, + 0.009811627678573132, + 0.00794221181422472, + 0.00823317002505064, + 0.005661951843649149, + 0.016665233299136162, + 0.008917199447751045, + 0.00835258886218071, + 0.00613351259380579, + 0.006226514931768179, + 0.010288095101714134, + 0.00808958150446415, + 0.009803364984691143, + 0.009027540683746338, + 0.010689835995435715, + 0.013628211803734303, + 0.009959418326616287, + 0.010030836798250675, + 0.011749467812478542, + 0.008257458917796612, + 0.002769608749076724, + 0.0029789539985358715, + 0.0032941263634711504, + 0.002969504566863179, + 0.00524444505572319, + 0.003964889328926802, + 0.003988543525338173, + 0.0030989404767751694, + 0.0046864040195941925, + 0.004742084536701441, + 0.0052271015010774136, + 0.0035634441301226616, + 0.006586593575775623, + 0.005702874157577753, + 0.003826826112344861, + 0.004348907619714737, + 0.05399750918149948, + 0.013196142390370369, + 0.02336912788450718, + 0.014723715372383595, + 0.01552136242389679, + 0.10004319995641708, + 0.014261286705732346, + 0.01914404146373272, + 0.008958244696259499, + 0.008485512807965279, + 0.009761159308254719, + 0.00902587827295065, + 0.008747017942368984, + 0.005045879166573286, + 0.006507810205221176, + 0.00695252139121294, + 0.039740875363349915, + 0.021060965955257416, + 0.013235457241535187, + 0.014847565442323685, + 0.0785604789853096, + 0.020059365779161453, + 0.014435166493058205, + 0.010113469325006008, + 0.004784350283443928, + 0.01015053503215313, + 0.003951533231884241, + 0.005509725771844387, + 0.006052418611943722, + 0.0038498295471072197, + 0.004171785898506641, + 0.006319855805486441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.16033093631267548 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, right-facing, panoramic shot reveals the tranquil scene of a spacious, semi-open wooden workshop. Soft, golden morning sunlight slants through the open side, casting sharp beams of light through the air, shimmering with fine sawdust and dust. At the center of the frame, a massive, unfinished dragon boat rests upon its wooden frame like a sleeping beast. Its hull, crafted from dark solid wood, its intricately carved dragon scales gleaming in the morning light. Several craftsmen, dressed in simple cotton work clothes, are scattered around the boat, their weathered faces etched with focused composure. An elderly craftsman kneels at the bow, meticulously sanding the ferocious, majestic details of the dragon's head with fine sandpaper. Another stands amidships, a small chisel in hand, steadily and deliberately adding the finishing touches to a carved dragon scale. Still another uses a wide brush to evenly apply a layer of transparent tung oil to the side of the boat, imbuing the wood with a rich, vibrant color wherever it passes. The scene is silent, with only the gentle clink of tools on wood. An almost sacred feeling of dedication and reverence for traditional craftsmanship permeates the scene, while the overall tone is warm and earthy.", + "seed": 28330, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_14/index_2368_seed_28330_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.041498757898807526, + "vlm_logprob": -0.041498757898807526, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.041498757898807526, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a spacious, semi-open wooden workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, unfinished dragon boat resting on a wooden frame at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are several craftsmen wearing simple cotton work clothes visible around the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is soft, golden morning sunlight slanting through the open side, casting sharp beams through sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slow, right-facing, and panoramic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022289352491497993 + ], + "opposite_logprob_tokens": [ + -4.397289276123047 + ], + "logprob_reward": -0.022289352491497993, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.022289352491497993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1461409330368042, + 0.1616157591342926, + 0.03181212395429611, + 0.22098809480667114, + 0.1688898503780365, + 0.054385460913181305, + 0.05680153891444206, + 0.06423166394233704, + 0.054070550948381424, + 0.04106400907039642 + ], + "credit_map": [ + 0.0129323098808527, + 0.00988512858748436, + 0.018866123631596565, + 0.028355466201901436, + 0.016288230195641518, + 0.016071442514657974, + 0.020044321194291115, + 0.02369791641831398, + 0.0190422423183918, + 0.01983778364956379, + 0.02042020857334137, + 0.0209635179489851, + 0.02996777556836605, + 0.014005209319293499, + 0.019374821335077286, + 0.018004203215241432, + 0.002555253216996789, + 0.004911618772894144, + 0.0029156231321394444, + 0.005368475802242756, + 0.003332854248583317, + 0.0038882389198988676, + 0.005200664047151804, + 0.003639396047219634, + 0.023201894015073776, + 0.0411701425909996, + 0.013232848607003689, + 0.02682148665189743, + 0.03170523792505264, + 0.026469361037015915, + 0.024310510605573654, + 0.03407661244273186, + 0.017506126314401627, + 0.01732484996318817, + 0.02048339508473873, + 0.03239944577217102, + 0.02441760152578354, + 0.016507379710674286, + 0.018981633707880974, + 0.021269408985972404, + 0.004646045621484518, + 0.004052553325891495, + 0.010112859308719635, + 0.006527018267661333, + 0.005344774108380079, + 0.004783016629517078, + 0.010854664258658886, + 0.008064529858529568, + 0.004865843802690506, + 0.004430423025041819, + 0.007725448813289404, + 0.008192582055926323, + 0.006011457648128271, + 0.005842116661369801, + 0.011590754613280296, + 0.008142915554344654, + 0.004593287594616413, + 0.0121222622692585, + 0.013035736046731472, + 0.008534066379070282, + 0.008205436170101166, + 0.004063427913933992, + 0.006360244937241077, + 0.007317208219319582, + 0.006099621299654245, + 0.00616710539907217, + 0.007315648719668388, + 0.008582169190049171, + 0.00849594920873642, + 0.0036855018697679043, + 0.008136214688420296, + 0.005588338244706392, + 0.004899375606328249, + 0.004747665952891111, + 0.009882723912596703, + 0.0051510510966181755, + 0.00372177897952497, + 0.002644107909873128, + 0.003389630001038313, + 0.006627676077187061 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -0.022289352491497993 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a spacious, semi-open wooden workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, unfinished dragon boat resting on a wooden frame at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are several craftsmen wearing simple cotton work clothes visible around the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is soft, golden morning sunlight slanting through the open side, casting sharp beams through sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slow, right-facing, and panoramic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008649386465549469 + ], + "opposite_logprob_tokens": [ + -4.7586493492126465 + ], + "logprob_reward": -0.008649386465549469, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008649386465549469, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14345510303974152, + 0.2571296691894531, + 0.03236370533704758, + 0.1307278722524643, + 0.1319337636232376, + 0.049316518008708954, + 0.037233076989650726, + 0.031551096588373184, + 0.06234438344836235, + 0.12394477427005768 + ], + "credit_map": [ + 0.01150850486010313, + 0.011619308032095432, + 0.023349899798631668, + 0.03287886828184128, + 0.014696203172206879, + 0.014277021400630474, + 0.016872581094503403, + 0.018252737820148468, + 0.024133609607815742, + 0.029603663831949234, + 0.03407856822013855, + 0.030988996848464012, + 0.04724598675966263, + 0.032404545694589615, + 0.0307643860578537, + 0.027909934520721436, + 0.002503094496205449, + 0.0035913665778934956, + 0.0032418379560112953, + 0.003729374147951603, + 0.0035909994039684534, + 0.005362831987440586, + 0.006156939547508955, + 0.004187264945358038, + 0.013797374442219734, + 0.025025280192494392, + 0.0075195468962192535, + 0.01227915845811367, + 0.019142724573612213, + 0.01681162789463997, + 0.015411464497447014, + 0.020740721374750137, + 0.013667205348610878, + 0.011202855966985226, + 0.015331094153225422, + 0.028956152498722076, + 0.02000066637992859, + 0.01194742415100336, + 0.01700608804821968, + 0.013822303153574467, + 0.004953871481120586, + 0.004568046424537897, + 0.007660639472305775, + 0.005999912042170763, + 0.009277653880417347, + 0.004269166383892298, + 0.006130601279437542, + 0.0064566307701170444, + 0.0031548377592116594, + 0.00382027099840343, + 0.007140154950320721, + 0.0036250432021915913, + 0.003613485721871257, + 0.003985375165939331, + 0.006189626641571522, + 0.0057042865082621574, + 0.0029956274665892124, + 0.003329686587676406, + 0.0040711467154324055, + 0.004441235680133104, + 0.006839750800281763, + 0.0023285895586013794, + 0.004478725604712963, + 0.0030663381330668926, + 0.005654930602759123, + 0.009663399308919907, + 0.008330519311130047, + 0.010347963310778141, + 0.008218030445277691, + 0.0039715152233839035, + 0.009180025197565556, + 0.006978002842515707, + 0.012167124077677727, + 0.009685621596872807, + 0.034551091492176056, + 0.015653960406780243, + 0.00994639378041029, + 0.005344804376363754, + 0.007332450710237026, + 0.029263341799378395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.008649386465549469 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a spacious, semi-open wooden workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, unfinished dragon boat resting on a wooden frame at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are several craftsmen wearing simple cotton work clothes visible around the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is soft, golden morning sunlight slanting through the open side, casting sharp beams through sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slow, right-facing, and panoramic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062056127935647964 + ], + "opposite_logprob_tokens": [ + -2.812056064605713 + ], + "logprob_reward": -0.062056127935647964, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062056127935647964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1158832460641861, + 0.08437824994325638, + 0.0503055639564991, + 0.1433972716331482, + 0.07944254577159882, + 0.05271810665726662, + 0.09787744283676147, + 0.1169571653008461, + 0.08499094843864441, + 0.17404943704605103 + ], + "credit_map": [ + 0.011418540962040424, + 0.009497587569057941, + 0.013332647271454334, + 0.01727088913321495, + 0.015087229199707508, + 0.015693984925746918, + 0.014202626422047615, + 0.019379742443561554, + 0.007947123609483242, + 0.009931782260537148, + 0.011972804553806782, + 0.009686415083706379, + 0.012474567629396915, + 0.012767665088176727, + 0.010210473090410233, + 0.009387422353029251, + 0.005439619068056345, + 0.007791491225361824, + 0.004874982871115208, + 0.005380807910114527, + 0.007719128392636776, + 0.00469549186527729, + 0.008303109556436539, + 0.006100932136178017, + 0.012436664663255215, + 0.020727895200252533, + 0.012948784045875072, + 0.01812717504799366, + 0.017243489623069763, + 0.024858197197318077, + 0.017579084262251854, + 0.019475974142551422, + 0.007907942868769169, + 0.006611818913370371, + 0.010865343734622002, + 0.01026124320924282, + 0.009748033247888088, + 0.008561787195503712, + 0.008127731271088123, + 0.017358645796775818, + 0.0066013759933412075, + 0.004828394390642643, + 0.00885460153222084, + 0.004873140249401331, + 0.004278318490833044, + 0.004582815803587437, + 0.009928977116942406, + 0.008770481683313847, + 0.011353944428265095, + 0.018875526264309883, + 0.01120771374553442, + 0.011348897591233253, + 0.00988644827157259, + 0.009575380943715572, + 0.018020298331975937, + 0.007609230000525713, + 0.010034078732132912, + 0.010009251534938812, + 0.012731525115668774, + 0.017563985660672188, + 0.02941056340932846, + 0.010747106745839119, + 0.013808020390570164, + 0.012652634643018246, + 0.008729759603738785, + 0.009389687329530716, + 0.010738629847764969, + 0.014404538087546825, + 0.011392960324883461, + 0.00463657034561038, + 0.01603960245847702, + 0.009659204632043839, + 0.01501445472240448, + 0.026689711958169937, + 0.052601270377635956, + 0.0256663765758276, + 0.020824484527111053, + 0.0068507413379848, + 0.013479347340762615, + 0.012923061847686768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 7 + ], + "selected_reward": -0.062056127935647964 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a spacious, semi-open wooden workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, unfinished dragon boat resting on a wooden frame at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are several craftsmen wearing simple cotton work clothes visible around the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is soft, golden morning sunlight slanting through the open side, casting sharp beams through sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slow, right-facing, and panoramic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014202939346432686 + ], + "opposite_logprob_tokens": [ + -4.264203071594238 + ], + "logprob_reward": -0.014202939346432686, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202939346432686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25030115246772766, + 0.11818322539329529, + 0.021068301051855087, + 0.15833202004432678, + 0.21072691679000854, + 0.05237031728029251, + 0.04531991481781006, + 0.02438216842710972, + 0.06319694221019745, + 0.05611900985240936 + ], + "credit_map": [ + 0.023451613262295723, + 0.02985326759517193, + 0.04094603285193443, + 0.054407838732004166, + 0.028138620778918266, + 0.020063595846295357, + 0.025482404977083206, + 0.027957763522863388, + 0.016915304586291313, + 0.01490895263850689, + 0.014188618399202824, + 0.01209414005279541, + 0.019538696855306625, + 0.013948537409305573, + 0.013629882596433163, + 0.012959085404872894, + 0.002698437310755253, + 0.002739001763984561, + 0.0022405022755265236, + 0.0025728626642376184, + 0.0022752394434064627, + 0.002817396307364106, + 0.003166039939969778, + 0.0025588201824575663, + 0.01782885752618313, + 0.029787158593535423, + 0.010109137743711472, + 0.01896197721362114, + 0.02233795076608658, + 0.019352229312062263, + 0.01710372418165207, + 0.022850986570119858, + 0.026437897235155106, + 0.022529294714331627, + 0.02983047254383564, + 0.03494098037481308, + 0.025016162544488907, + 0.025095259770751, + 0.02420908212661743, + 0.022667765617370605, + 0.006970301736146212, + 0.005851145833730698, + 0.007695294450968504, + 0.00866242777556181, + 0.0054017286747694016, + 0.005052474793046713, + 0.0066103278659284115, + 0.006126614287495613, + 0.004501848015934229, + 0.011293614283204079, + 0.007045859005302191, + 0.00624194648116827, + 0.006005672737956047, + 0.0034721677657216787, + 0.003673106897622347, + 0.0030856956727802753, + 0.002021563006564975, + 0.002620961982756853, + 0.003576200921088457, + 0.003817963181063533, + 0.0036720747593790293, + 0.001906590536236763, + 0.003964788280427456, + 0.0028020229656249285, + 0.005918871611356735, + 0.00987511221319437, + 0.007514165714383125, + 0.009671746753156185, + 0.009222309105098248, + 0.0038666273467242718, + 0.00992859248071909, + 0.007199516054242849, + 0.006160808261483908, + 0.007175622042268515, + 0.01640586368739605, + 0.008314485661685467, + 0.007255654316395521, + 0.002529214369133115, + 0.0034013648983091116, + 0.004875991027802229 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.014202939346432686 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a spacious, semi-open wooden workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, unfinished dragon boat resting on a wooden frame at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are several craftsmen wearing simple cotton work clothes visible around the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is soft, golden morning sunlight slanting through the open side, casting sharp beams through sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slow, right-facing, and panoramic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10029598325490952 + ], + "opposite_logprob_tokens": [ + -2.3502960205078125 + ], + "logprob_reward": -0.10029598325490952, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10029598325490952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07470977306365967, + 0.0624316968023777, + 0.09504756331443787, + 0.10738974809646606, + 0.3812785744667053, + 0.02236858382821083, + 0.0798661857843399, + 0.04911627620458603, + 0.07029229402542114, + 0.05749928206205368 + ], + "credit_map": [ + 0.0064002289436757565, + 0.006051820702850819, + 0.008647037670016289, + 0.007684561423957348, + 0.0060945614241063595, + 0.0062578702345490456, + 0.02069523185491562, + 0.012878459878265858, + 0.006925028283149004, + 0.0055594490841031075, + 0.008239350281655788, + 0.006935134995728731, + 0.009596453979611397, + 0.010805126279592514, + 0.007525989785790443, + 0.006845163647085428, + 0.011092099361121655, + 0.013860824517905712, + 0.009099499322474003, + 0.012387117370963097, + 0.008948620408773422, + 0.01059043686836958, + 0.015432950109243393, + 0.01363601628690958, + 0.008865977637469769, + 0.024287516251206398, + 0.006671382579952478, + 0.01315338909626007, + 0.012047570198774338, + 0.01585203781723976, + 0.01262346375733614, + 0.013888409361243248, + 0.02849096618592739, + 0.01695997454226017, + 0.12684465944766998, + 0.04039132222533226, + 0.0677160769701004, + 0.02842300571501255, + 0.04108726233243942, + 0.03136530891060829, + 0.0025136389303952456, + 0.0025108589325100183, + 0.0032227938063442707, + 0.00286966934800148, + 0.003023403463885188, + 0.0025805532932281494, + 0.0034203522372990847, + 0.002227312419563532, + 0.0053367093205451965, + 0.009848619811236858, + 0.019642986357212067, + 0.0070580653846263885, + 0.006348425522446632, + 0.0060222032479941845, + 0.01860804110765457, + 0.007001133169978857, + 0.004251827485859394, + 0.00934018287807703, + 0.009713619016110897, + 0.008226139470934868, + 0.0044084470719099045, + 0.0028582557570189238, + 0.005811648443341255, + 0.004506152588874102, + 0.005728613119572401, + 0.012065806426107883, + 0.009768816642463207, + 0.011676264926791191, + 0.009350761771202087, + 0.0032049280125647783, + 0.01183602586388588, + 0.006661080755293369, + 0.005295476410537958, + 0.006252232473343611, + 0.019153781235218048, + 0.008486900478601456, + 0.004887841641902924, + 0.002452776301652193, + 0.0029843421652913094, + 0.007985933683812618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.10029598325490952 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_15/index_tensor([5393, 5393])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_15/index_tensor([5393, 5393])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1f36721af058dc5f5007f1c5bdcb62f1a19b139f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_15/index_tensor([5393, 5393])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from a horizontal angle, the camera held still, its focus fixed squarely on a stoneware casserole set against a stark black void. A strong overhead light, precisely directed at the casserole's rounded, heavy, curved top, outlines its ancient contours and casts deep shadows across its rough, dark brown and terracotta-colored surface, highlighting the fine pores and grain left by firing. Suddenly, a hairline white crack silently shatters the casserole's tranquility. Then, as if triggering a chain reaction, countless cracks spread outward from this point with astonishing speed, instantly forming a dense spiderweb that shimmers with a pale white glow in the light. In the instant the cracks spread, the entire casserole's structure reaches a breaking point. Rather than bursting outward, it violently and dramatically sags inward, collapsing inward. The solid clay instantly disintegrated under the enormous inward pressure, turning into a cloud of dust mixed with tiny fragments and a large amount of dark powder. This chaotic cloud illuminated by the top light rolled and expanded briefly in place, then slowly drifted downwards due to the loss of all support, and finally completely disappeared in the endless darkness below, leaving only nothingness.", + "seed": 351, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_15/index_5393_seed_351_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01974322597961873, + "vlm_logprob": -0.01974322597961873, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.01974322597961873, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a stoneware casserole against a stark black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the casserole's surface dark brown and terracotta-colored with visible fine pores and grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline white crack appear and instantly spread into a dense, shimmering spiderweb pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casserole violently sag inward and collapse rather than bursting outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the collapsing structure turn into a cloud of dust and dark powder that drifts downwards and disappears into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.035029031336307526 + ], + "opposite_logprob_tokens": [ + -3.5350289344787598 + ], + "logprob_reward": -0.035029031336307526, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.035029031336307526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04191756993532181, + 0.012613177299499512, + 0.008466529659926891, + 0.00640099635347724, + 0.011942725628614426, + 0.06920412927865982, + 0.01650940254330635, + 0.7602450251579285, + 0.05140819773077965, + 0.021292274817824364 + ], + "credit_map": [ + 0.00888005644083023, + 0.0020373365841805935, + 0.005038327071815729, + 0.004309770651161671, + 0.0024448104668408632, + 0.00944144744426012, + 0.004116268828511238, + 0.005649556405842304, + 0.0019930077251046896, + 0.0014455419732257724, + 0.0011357237817719579, + 0.0013366970233619213, + 0.0018259246135130525, + 0.0014057758962735534, + 0.0017977359239012003, + 0.0016727701295167208, + 0.0016016921726986766, + 0.0008121742866933346, + 0.0007202618871815503, + 0.0009868022752925754, + 0.0009425858152098954, + 0.0012021561851724982, + 0.0012603801442310214, + 0.0009404771262779832, + 0.0013220557011663914, + 0.0004858076572418213, + 0.0007384428754448891, + 0.0005986732430756092, + 0.0005238166195340455, + 0.0013072217116132379, + 0.0006564069190062582, + 0.0007685713353566825, + 0.0017871285090222955, + 0.0017221842426806688, + 0.0008891359902918339, + 0.0017359532648697495, + 0.0013517794432118535, + 0.0018776989309117198, + 0.0014365004608407617, + 0.0011423449032008648, + 0.01525849662721157, + 0.029544441029429436, + 0.0019215561915189028, + 0.0034388049971312284, + 0.013973942026495934, + 0.0027674268931150436, + 0.001266009989194572, + 0.0010334481485188007, + 0.0013466012896969914, + 0.0021907463669776917, + 0.0019360374426469207, + 0.00250904425047338, + 0.0030451370403170586, + 0.002121845493093133, + 0.002096800599247217, + 0.0012631904100999236, + 0.0364517979323864, + 0.17558354139328003, + 0.014022786170244217, + 0.020548196509480476, + 0.29364341497421265, + 0.18080396950244904, + 0.019615422934293747, + 0.019575906917452812, + 0.007050611078739166, + 0.008197180926799774, + 0.004798973444849253, + 0.005375223234295845, + 0.006297969724982977, + 0.006913037970662117, + 0.0059442464262247086, + 0.006830957718193531, + 0.0018483921885490417, + 0.003083297284319997, + 0.0033250388223677874, + 0.003335090819746256, + 0.002652163850143552, + 0.0022615049965679646, + 0.0029026258271187544, + 0.0018841610290110111 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 8 + ], + "selected_reward": -0.035029031336307526 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a stoneware casserole against a stark black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the casserole's surface dark brown and terracotta-colored with visible fine pores and grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline white crack appear and instantly spread into a dense, shimmering spiderweb pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casserole violently sag inward and collapse rather than bursting outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the collapsing structure turn into a cloud of dust and dark powder that drifts downwards and disappears into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012559015303850174 + ], + "opposite_logprob_tokens": [ + -4.387558937072754 + ], + "logprob_reward": -0.012559015303850174, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012559015303850174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.038366593420505524, + 0.03452595695853233, + 0.01034989021718502, + 0.01001163013279438, + 0.013946624472737312, + 0.16731245815753937, + 0.012258144095540047, + 0.6644448041915894, + 0.02363925613462925, + 0.02514456957578659 + ], + "credit_map": [ + 0.007968040183186531, + 0.003424772759899497, + 0.0037566705141216516, + 0.0037999916821718216, + 0.003599917981773615, + 0.004650058690458536, + 0.004248861689120531, + 0.006918282248079777, + 0.005348094739019871, + 0.0037638640496879816, + 0.0031779934652149677, + 0.0029732121620327234, + 0.0039495197124779224, + 0.006388286128640175, + 0.0036889812909066677, + 0.005236009135842323, + 0.0014614732936024666, + 0.0012052858946844935, + 0.0011949222534894943, + 0.0010444559156894684, + 0.001216250704601407, + 0.001468723639845848, + 0.0012479395372793078, + 0.0015108393272385001, + 0.0014687527436763048, + 0.0009740911773405969, + 0.0011086879530921578, + 0.0011161902220919728, + 0.0011972037609666586, + 0.0012469226494431496, + 0.0013712227810174227, + 0.0015285596018657088, + 0.0022086361423134804, + 0.0013871339615434408, + 0.0022349136415868998, + 0.0018605377990752459, + 0.0011485365685075521, + 0.0022070209961384535, + 0.0014448586152866483, + 0.0014549883781000972, + 0.03764937445521355, + 0.07201078534126282, + 0.0042121619917452335, + 0.007595682051032782, + 0.033772461116313934, + 0.006496245041489601, + 0.0026686496566981077, + 0.0029071036260575056, + 0.0013496673200279474, + 0.001699118292890489, + 0.002463477896526456, + 0.001286985701881349, + 0.001923837000504136, + 0.0013571679592132568, + 0.0011391821317374706, + 0.0010387086076661944, + 0.03132201358675957, + 0.15302857756614685, + 0.01151650957763195, + 0.018308304250240326, + 0.25618696212768555, + 0.15987125039100647, + 0.01711631938815117, + 0.01709495671093464, + 0.002176632173359394, + 0.003975439816713333, + 0.001763132051564753, + 0.0028687689919024706, + 0.0039544422179460526, + 0.0027748385909944773, + 0.003348858328536153, + 0.0027771461755037308, + 0.0020303663332015276, + 0.003497904632240534, + 0.0030824802815914154, + 0.004223529249429703, + 0.0033146666828542948, + 0.0020906433928757906, + 0.003959053661674261, + 0.002945928368717432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.012559015303850174 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a stoneware casserole against a stark black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the casserole's surface dark brown and terracotta-colored with visible fine pores and grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline white crack appear and instantly spread into a dense, shimmering spiderweb pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casserole violently sag inward and collapse rather than bursting outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the collapsing structure turn into a cloud of dust and dark powder that drifts downwards and disappears into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738912779837847 + ], + "opposite_logprob_tokens": [ + -5.006739139556885 + ], + "logprob_reward": -0.006738912779837847, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738912779837847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03847089409828186, + 0.025479938834905624, + 0.011162432841956615, + 0.012499971315264702, + 0.028896529227495193, + 0.19236423075199127, + 0.058414604514837265, + 0.5155875086784363, + 0.05052131041884422, + 0.06660260260105133 + ], + "credit_map": [ + 0.006744309328496456, + 0.003329623956233263, + 0.0055617173202335835, + 0.004420605953782797, + 0.003325986210256815, + 0.005295583046972752, + 0.00328173884190619, + 0.006511322688311338, + 0.00495639955624938, + 0.002077965997159481, + 0.0029125972650945187, + 0.0025757302064448595, + 0.0024509290233254433, + 0.004820206668227911, + 0.0027400886174291372, + 0.0029460189398378134, + 0.0016490293201059103, + 0.0008824840770103037, + 0.0014272778062149882, + 0.001697659958153963, + 0.0009152418351732194, + 0.001151164178736508, + 0.0018018416594713926, + 0.0016377323772758245, + 0.002148672007024288, + 0.001364012248814106, + 0.0014362474903464317, + 0.0019547338597476482, + 0.0010996442288160324, + 0.0013250010088086128, + 0.0013909968547523022, + 0.0017806612886488438, + 0.00413612462580204, + 0.003555099479854107, + 0.0032065222039818764, + 0.002731718122959137, + 0.0033062708098441362, + 0.005275650881230831, + 0.003925010561943054, + 0.002760130213573575, + 0.04099389910697937, + 0.0787351205945015, + 0.006195226218551397, + 0.010193840600550175, + 0.03798440843820572, + 0.007932442240417004, + 0.004748290404677391, + 0.005580986849963665, + 0.005865844432264566, + 0.008205816149711609, + 0.008523719385266304, + 0.0073131960816681385, + 0.009289349429309368, + 0.006441676057875156, + 0.006998552940785885, + 0.005776442587375641, + 0.035038240253925323, + 0.11475169658660889, + 0.02179277502000332, + 0.020362360402941704, + 0.1871250569820404, + 0.09179545938968658, + 0.025154128670692444, + 0.01956772804260254, + 0.006721193436533213, + 0.0094684399664402, + 0.003758534789085388, + 0.003664295421913266, + 0.007745569106191397, + 0.0070183672942221165, + 0.005764807108789682, + 0.006380103528499603, + 0.005735939834266901, + 0.012340974994003773, + 0.00772837083786726, + 0.00841271597892046, + 0.011918428353965282, + 0.010583743453025818, + 0.005581009667366743, + 0.004301419015973806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 9 + ], + "selected_reward": -0.006738912779837847 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a stoneware casserole against a stark black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the casserole's surface dark brown and terracotta-colored with visible fine pores and grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline white crack appear and instantly spread into a dense, shimmering spiderweb pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casserole violently sag inward and collapse rather than bursting outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the collapsing structure turn into a cloud of dust and dark powder that drifts downwards and disappears into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305289313197136 + ], + "opposite_logprob_tokens": [ + -3.168052911758423 + ], + "logprob_reward": -0.04305289313197136, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305289313197136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09639550745487213, + 0.05133863911032677, + 0.021111546084284782, + 0.03257995471358299, + 0.028004705905914307, + 0.11875675618648529, + 0.04949541017413139, + 0.3676815927028656, + 0.1321779191493988, + 0.1024579405784607 + ], + "credit_map": [ + 0.015106646344065666, + 0.01087123155593872, + 0.009046011604368687, + 0.011068245396018028, + 0.01320459134876728, + 0.014099357649683952, + 0.011146425269544125, + 0.011853001080453396, + 0.008583440445363522, + 0.005765056703239679, + 0.00364744383841753, + 0.004456312861293554, + 0.006625750567764044, + 0.010802468284964561, + 0.005794144235551357, + 0.005664023105055094, + 0.0033320938237011433, + 0.0020746144000440836, + 0.001983600202947855, + 0.0019346189219504595, + 0.0022015946451574564, + 0.004204307217150927, + 0.002754835644736886, + 0.0026258823927491903, + 0.005063952878117561, + 0.0033063716255128384, + 0.0034344024024903774, + 0.003571774112060666, + 0.004352023359388113, + 0.003679053857922554, + 0.004770380444824696, + 0.00440199626609683, + 0.004395551513880491, + 0.0031337577383965254, + 0.003876164322718978, + 0.003231416456401348, + 0.003289245069026947, + 0.0037263156846165657, + 0.003208140144124627, + 0.003144115675240755, + 0.02262692339718342, + 0.043984122574329376, + 0.005276466254144907, + 0.006142650730907917, + 0.022710580378770828, + 0.005914739333093166, + 0.005480261519551277, + 0.006621010601520538, + 0.005432082340121269, + 0.008967257104814053, + 0.004549375269562006, + 0.005432353354990482, + 0.009979831986129284, + 0.006319611333310604, + 0.005258118733763695, + 0.0035567772574722767, + 0.017331287264823914, + 0.08167486637830734, + 0.010480036959052086, + 0.013844220899045467, + 0.13582484424114227, + 0.08442480862140656, + 0.012788851745426655, + 0.011312694288790226, + 0.010972447693347931, + 0.017313247546553612, + 0.015373862348496914, + 0.017337094992399216, + 0.02038521133363247, + 0.01595877669751644, + 0.020337391644716263, + 0.014499886892735958, + 0.009211973287165165, + 0.014855578541755676, + 0.014313065446913242, + 0.009261767379939556, + 0.015758313238620758, + 0.014341828413307667, + 0.01397015992552042, + 0.010745260864496231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.04305289313197136 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a stoneware casserole against a stark black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the casserole's surface dark brown and terracotta-colored with visible fine pores and grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline white crack appear and instantly spread into a dense, shimmering spiderweb pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casserole violently sag inward and collapse rather than bursting outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the collapsing structure turn into a cloud of dust and dark powder that drifts downwards and disappears into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013362773461267352 + ], + "opposite_logprob_tokens": [ + -6.626336097717285 + ], + "logprob_reward": -0.0013362773461267352, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013362773461267352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04021572321653366, + 0.02864384651184082, + 0.009381846524775028, + 0.00894204992800951, + 0.007856465876102448, + 0.051364652812480927, + 0.015813395380973816, + 0.6531607508659363, + 0.05170022323727608, + 0.13292105495929718 + ], + "credit_map": [ + 0.006086716894060373, + 0.0040442910976707935, + 0.0038244971074163914, + 0.004113662056624889, + 0.004698556382209063, + 0.007299062795937061, + 0.004970422014594078, + 0.005178514868021011, + 0.00478676613420248, + 0.0030543124303221703, + 0.0020269579254090786, + 0.0024925388861447573, + 0.003800734644755721, + 0.005852579139173031, + 0.0030768071301281452, + 0.0035531523171812296, + 0.0014631287194788456, + 0.0010391096584498882, + 0.0007541136001236737, + 0.0008547387551516294, + 0.0011230689706280828, + 0.0018617844907566905, + 0.001133162877522409, + 0.0011527398601174355, + 0.0012393095530569553, + 0.0010245214216411114, + 0.0007881140918470919, + 0.0007765808841213584, + 0.0011930186301469803, + 0.0015693396562710404, + 0.0011218752479180694, + 0.0012292901519685984, + 0.0010926943505182862, + 0.000823794980533421, + 0.0010500047355890274, + 0.0011128083569929004, + 0.0009094998240470886, + 0.0010542431846261024, + 0.0009579078177921474, + 0.0008555120439268649, + 0.01062200777232647, + 0.021036198362708092, + 0.0018379620742052794, + 0.0025839039590209723, + 0.010104614309966564, + 0.0022729462943971157, + 0.0018746127607300878, + 0.0010324075119569898, + 0.0017777547473087907, + 0.002566699171438813, + 0.0013330969959497452, + 0.0014100789558142424, + 0.0035133706405758858, + 0.001985615584999323, + 0.0019227531738579273, + 0.001304027740843594, + 0.03091721422970295, + 0.15010739862918854, + 0.010462752543389797, + 0.0176848117262125, + 0.25172245502471924, + 0.15856732428073883, + 0.01648845709860325, + 0.017210332676768303, + 0.005378641188144684, + 0.007284137420356274, + 0.007024367805570364, + 0.004466151352971792, + 0.007998738437891006, + 0.007604615762829781, + 0.006172709632664919, + 0.005770865362137556, + 0.011048216372728348, + 0.01638682559132576, + 0.02002178505063057, + 0.022978030145168304, + 0.01900780200958252, + 0.012585161253809929, + 0.018441399559378624, + 0.012451833114027977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.0013362773461267352 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a horizontal angle, the camera held still, its focus fixed squarely on a stoneware casserole set against a stark black void. A strong overhead light, precisely directed at the casserole's rounded, heavy, curved top, outlines its ancient contours and casts deep shadows across its rough, dark brown and terracotta-colored surface, highlighting the fine pores and grain left by firing. Suddenly, a hairline white crack silently shatters the casserole's tranquility. Then, as if triggering a chain reaction, countless cracks spread outward from this point with astonishing speed, instantly forming a dense spiderweb that shimmers with a pale white glow in the light. In the instant the cracks spread, the entire casserole's structure reaches a breaking point. Rather than bursting outward, it violently and dramatically sags inward, collapsing inward. The solid clay instantly disintegrated under the enormous inward pressure, turning into a cloud of dust mixed with tiny fragments and a large amount of dark powder. This chaotic cloud illuminated by the top light rolled and expanded briefly in place, then slowly drifted downwards due to the loss of all support, and finally completely disappeared in the endless darkness below, leaving only nothingness.", + "seed": 352, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_15/index_5393_seed_352_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.012093017459847033, + "vlm_logprob": -0.012093017459847033, + "vlm_yes_no_margin": 5.025, + "vlm_token_credit": -0.012093017459847033, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a stoneware casserole against a stark black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the casserole's surface dark brown and terracotta-colored with visible fine pores and grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline white crack appear and instantly spread into a dense, shimmering spiderweb pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casserole violently sag inward and collapse rather than bursting outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the collapsing structure turn into a cloud of dust and dark powder that drifts downwards and disappears into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017824934795498848 + ], + "opposite_logprob_tokens": [ + -4.267825126647949 + ], + "logprob_reward": -0.017824934795498848, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.017824934795498848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0950004979968071, + 0.025685183703899384, + 0.01299827266484499, + 0.015901068225502968, + 0.021189261227846146, + 0.007266798056662083, + 0.019325682893395424, + 0.0294049009680748, + 0.0805739164352417, + 0.6926543712615967 + ], + "credit_map": [ + 0.008157557807862759, + 0.0075331381522119045, + 0.01702776923775673, + 0.016533175483345985, + 0.006274331361055374, + 0.006070148199796677, + 0.018564455211162567, + 0.014839934185147285, + 0.002132577123120427, + 0.0019323882879689336, + 0.004155165050178766, + 0.004476424306631088, + 0.0022336975671350956, + 0.0014791616704314947, + 0.0057618883438408375, + 0.0035138807725161314, + 0.0012964813504368067, + 0.000987514154985547, + 0.002384546445682645, + 0.002172617707401514, + 0.0009676789049990475, + 0.0007235673256218433, + 0.0027824912685900927, + 0.0016833757981657982, + 0.0021610958501696587, + 0.0014205867191776633, + 0.002760054310783744, + 0.0025748051702976227, + 0.0012098668375983834, + 0.001107734628021717, + 0.0022532185539603233, + 0.0024137054570019245, + 0.002358986996114254, + 0.0023639544378966093, + 0.0029671918600797653, + 0.004097800701856613, + 0.0015515557024627924, + 0.0017188662895932794, + 0.0032321172766387463, + 0.0028987890109419823, + 0.0005685880896635354, + 0.0007893547299318016, + 0.0009761331020854414, + 0.0011180954752489924, + 0.0008121744613163173, + 0.0008930659969337285, + 0.0013495052698999643, + 0.0007598812808282673, + 0.0019090116256847978, + 0.002430328633636236, + 0.0019351382507011294, + 0.002837737323716283, + 0.0024897633120417595, + 0.0023424639366567135, + 0.0027851068880409002, + 0.0025961329229176044, + 0.003732292912900448, + 0.0038742227479815483, + 0.003047442762181163, + 0.004436384420841932, + 0.004916238598525524, + 0.0035374413710087538, + 0.003774489276111126, + 0.0020863888785243034, + 0.009634871035814285, + 0.01560088898986578, + 0.0044745649211108685, + 0.006401417311280966, + 0.019458698108792305, + 0.0151687515899539, + 0.0046180603094398975, + 0.0052166590467095375, + 0.07255052775144577, + 0.16864077746868134, + 0.03871490806341171, + 0.029157400131225586, + 0.16722746193408966, + 0.13813187181949615, + 0.044050026684999466, + 0.03418141230940819 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.017824934795498848 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a stoneware casserole against a stark black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the casserole's surface dark brown and terracotta-colored with visible fine pores and grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline white crack appear and instantly spread into a dense, shimmering spiderweb pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casserole violently sag inward and collapse rather than bursting outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the collapsing structure turn into a cloud of dust and dark powder that drifts downwards and disappears into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012560075148940086 + ], + "opposite_logprob_tokens": [ + -4.38755989074707 + ], + "logprob_reward": -0.012560075148940086, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012560075148940086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15026290714740753, + 0.1338900774717331, + 0.0831516906619072, + 0.020358379930257797, + 0.08024773746728897, + 0.032930731773376465, + 0.05096658319234848, + 0.03208267688751221, + 0.06617996096611023, + 0.3499293029308319 + ], + "credit_map": [ + 0.016385026276111603, + 0.016220735386013985, + 0.019329925999045372, + 0.027433501556515694, + 0.011696644127368927, + 0.014310548081994057, + 0.025958076119422913, + 0.01892843283712864, + 0.013534588739275932, + 0.011000236496329308, + 0.018053404986858368, + 0.026167986914515495, + 0.01393927913159132, + 0.009945127181708813, + 0.027441857382655144, + 0.013807592913508415, + 0.00871655810624361, + 0.008183496072888374, + 0.010906580835580826, + 0.014102054759860039, + 0.00574121717363596, + 0.007827216759324074, + 0.014820476062595844, + 0.012854093685746193, + 0.002044394612312317, + 0.001825972693040967, + 0.0030418303795158863, + 0.0036057692486792803, + 0.0020039656665176153, + 0.001693342113867402, + 0.0030074515379965305, + 0.0031356532126665115, + 0.008735030889511108, + 0.009199162013828754, + 0.011628423817455769, + 0.014358531683683395, + 0.007796970661729574, + 0.0060361395590007305, + 0.012124293483793736, + 0.010369189083576202, + 0.002484111348167062, + 0.003988140262663364, + 0.003530869260430336, + 0.00537072354927659, + 0.0036774128675460815, + 0.004380884114652872, + 0.005950333550572395, + 0.003548254957422614, + 0.0042879716493189335, + 0.006161569617688656, + 0.0055455779656767845, + 0.00806270632892847, + 0.00556383840739727, + 0.006955236196517944, + 0.00808772724121809, + 0.006301958579570055, + 0.005429902579635382, + 0.004128948785364628, + 0.0038408299442380667, + 0.0035315544810146093, + 0.004085378255695105, + 0.0037411744706332684, + 0.004431743640452623, + 0.0028931472916156054, + 0.009162581525743008, + 0.012753941118717194, + 0.004014924634248018, + 0.004104590509086847, + 0.015437926165759563, + 0.012072457000613213, + 0.0045250109396874905, + 0.004108530469238758, + 0.03236410766839981, + 0.07889227569103241, + 0.020704176276922226, + 0.015460343100130558, + 0.08745882660150528, + 0.0705290213227272, + 0.025534937158226967, + 0.018985606729984283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.012560075148940086 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a stoneware casserole against a stark black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the casserole's surface dark brown and terracotta-colored with visible fine pores and grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline white crack appear and instantly spread into a dense, shimmering spiderweb pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casserole violently sag inward and collapse rather than bursting outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the collapsing structure turn into a cloud of dust and dark powder that drifts downwards and disappears into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028228219598531723 + ], + "opposite_logprob_tokens": [ + -5.8778228759765625 + ], + "logprob_reward": -0.0028228219598531723, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028228219598531723, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04535326734185219, + 0.025752121582627296, + 0.0216396301984787, + 0.011620893143117428, + 0.0172961987555027, + 0.0257946215569973, + 0.10094040632247925, + 0.1249031275510788, + 0.04788511246442795, + 0.5788146257400513 + ], + "credit_map": [ + 0.004317866638302803, + 0.0055800629779696465, + 0.006457820069044828, + 0.008873446844518185, + 0.0027597288135439157, + 0.003004069672897458, + 0.009620222263038158, + 0.0047400458715856075, + 0.0019812064711004496, + 0.0034661919344216585, + 0.004158869385719299, + 0.005638801958411932, + 0.0015274330507963896, + 0.001438832376152277, + 0.004819969180971384, + 0.002720816759392619, + 0.0018103758338838816, + 0.0017604591557756066, + 0.002675315598025918, + 0.004721696022897959, + 0.0016901231138035655, + 0.0010163349797949195, + 0.0055878134444355965, + 0.0023775091394782066, + 0.0011679534800350666, + 0.00088585598859936, + 0.0018735153134912252, + 0.0024754663463681936, + 0.0008164024911820889, + 0.0006916375714354217, + 0.0018373323837295175, + 0.0018727285787463188, + 0.0016165210399776697, + 0.0012122089974582195, + 0.002378293313086033, + 0.003163783811032772, + 0.001235947827808559, + 0.001670227968133986, + 0.003251780988648534, + 0.002767433412373066, + 0.0018270608270540833, + 0.0031155389733612537, + 0.0029758610762655735, + 0.003670530393719673, + 0.0026783954817801714, + 0.002699656179174781, + 0.00587486894801259, + 0.0029527076985687017, + 0.008655172772705555, + 0.01013979036360979, + 0.010924132540822029, + 0.0169219933450222, + 0.009827637113630772, + 0.01327707339078188, + 0.01960803009569645, + 0.011586573906242847, + 0.023552728816866875, + 0.01888715662062168, + 0.012748615816235542, + 0.011552833020687103, + 0.01771279238164425, + 0.01847407966852188, + 0.0130161689594388, + 0.008958745747804642, + 0.006043031346052885, + 0.008921860717236996, + 0.0027316755149513483, + 0.001991269178688526, + 0.013470646925270557, + 0.009322014637291431, + 0.0029100188985466957, + 0.0024945882614701986, + 0.059627801179885864, + 0.1353147327899933, + 0.03159620612859726, + 0.02455000951886177, + 0.1423545777797699, + 0.11547241359949112, + 0.039198726415634155, + 0.030700111761689186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.0028228219598531723 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a stoneware casserole against a stark black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the casserole's surface dark brown and terracotta-colored with visible fine pores and grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline white crack appear and instantly spread into a dense, shimmering spiderweb pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casserole violently sag inward and collapse rather than bursting outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the collapsing structure turn into a cloud of dust and dark powder that drifts downwards and disappears into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026336906477808952 + ], + "opposite_logprob_tokens": [ + -3.651336908340454 + ], + "logprob_reward": -0.026336906477808952, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026336906477808952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15250615775585175, + 0.03309934586286545, + 0.08317691832780838, + 0.04452744498848915, + 0.04028831049799919, + 0.06784050911664963, + 0.15972791612148285, + 0.08137165755033493, + 0.11853106319904327, + 0.21893064677715302 + ], + "credit_map": [ + 0.010203634388744831, + 0.012980418279767036, + 0.02702554501593113, + 0.029220014810562134, + 0.010005262680351734, + 0.008307605981826782, + 0.033797573298215866, + 0.020966103300452232, + 0.003895427566021681, + 0.0032007202971726656, + 0.004385237116366625, + 0.0054429140873253345, + 0.003427868476137519, + 0.002554873703047633, + 0.005738822277635336, + 0.004453479778021574, + 0.009577752090990543, + 0.006292025558650494, + 0.012142694555222988, + 0.015113716945052147, + 0.0061523765325546265, + 0.0069283475168049335, + 0.0145352678373456, + 0.012434744276106358, + 0.005106991622596979, + 0.0032994847279042006, + 0.006895698141306639, + 0.00821259617805481, + 0.0037308302707970142, + 0.0029498343355953693, + 0.008225071243941784, + 0.006106935907155275, + 0.003900327021256089, + 0.0036645217332988977, + 0.0057592010125517845, + 0.009668475016951561, + 0.002182565862312913, + 0.0031728281173855066, + 0.0072868093848228455, + 0.00465357955545187, + 0.006031437776982784, + 0.0075378622859716415, + 0.007128921337425709, + 0.010935616679489613, + 0.0070060845464468, + 0.006034145597368479, + 0.016441507264971733, + 0.00672494200989604, + 0.013091117143630981, + 0.020123973488807678, + 0.01741814613342285, + 0.026254955679178238, + 0.018331311643123627, + 0.023082930594682693, + 0.0206135343760252, + 0.020811960101127625, + 0.012136202305555344, + 0.010605786927044392, + 0.0077029066160321236, + 0.006628823932260275, + 0.015133386477828026, + 0.011044558137655258, + 0.011045259423553944, + 0.007074734661728144, + 0.014247799292206764, + 0.020065447315573692, + 0.008665060624480247, + 0.007272578310221434, + 0.03057701140642166, + 0.019480332732200623, + 0.009431697428226471, + 0.008791133761405945, + 0.014510023407638073, + 0.02630872279405594, + 0.012131161987781525, + 0.015100355260074139, + 0.06368261575698853, + 0.05917949974536896, + 0.016476986929774284, + 0.01154129859060049 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.026336906477808952 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a stoneware casserole against a stark black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the casserole's surface dark brown and terracotta-colored with visible fine pores and grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline white crack appear and instantly spread into a dense, shimmering spiderweb pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casserole violently sag inward and collapse rather than bursting outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the collapsing structure turn into a cloud of dust and dark powder that drifts downwards and disappears into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009203489171341062 + ], + "opposite_logprob_tokens": [ + -7.000920295715332 + ], + "logprob_reward": -0.0009203489171341062, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009203489171341062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01607232540845871, + 0.014347388409078121, + 0.005994539242237806, + 0.0031123459339141846, + 0.00508136535063386, + 0.007380617316812277, + 0.012501291930675507, + 0.018742350861430168, + 0.032576292753219604, + 0.8841914534568787 + ], + "credit_map": [ + 0.001615465385839343, + 0.0019119424978271127, + 0.0018271717708557844, + 0.0033551999367773533, + 0.0010735132964327931, + 0.0008950211922638118, + 0.003153266618028283, + 0.002240746282041073, + 0.0010956914629787207, + 0.0016180529491975904, + 0.002444131998345256, + 0.0025422084145247936, + 0.0013732146471738815, + 0.0007454887381754816, + 0.0031158961355686188, + 0.0014127054018899798, + 0.0007159068481996655, + 0.0005653719417750835, + 0.0007982399547472596, + 0.0010650986805558205, + 0.0005502774147316813, + 0.000542733003385365, + 0.0009968371596187353, + 0.0007600749959237874, + 0.0002963616861961782, + 0.0001926047698361799, + 0.000529570272192359, + 0.000633878109510988, + 0.0002603994507808238, + 0.00022319859999697655, + 0.0005947314784862101, + 0.00038160168332979083, + 0.0005212928517721593, + 0.0004439039621502161, + 0.0007238705875352025, + 0.0009617929463274777, + 0.0003025089390575886, + 0.0004237426910549402, + 0.0009031373774632812, + 0.0008011168101802468, + 0.0005003496771678329, + 0.0007044609519653022, + 0.0008799513452686369, + 0.0010636469814926386, + 0.0006654521566815674, + 0.0008244551718235016, + 0.0014412997988983989, + 0.0013010018737986684, + 0.0015467305202037096, + 0.0014115386875346303, + 0.0012556861620396376, + 0.0019199289381504059, + 0.0011328819673508406, + 0.0015314517077058554, + 0.002052412601187825, + 0.0016506628599017859, + 0.0020217986311763525, + 0.0035949794109910727, + 0.00171414646320045, + 0.002065104665234685, + 0.0035830442793667316, + 0.002397130709141493, + 0.0022843130864202976, + 0.0010818365262821317, + 0.004995971918106079, + 0.00476620951667428, + 0.0026764085050672293, + 0.003543270519003272, + 0.005533704999834299, + 0.005879789590835571, + 0.0028861311729997396, + 0.0022948132827878, + 0.09040796756744385, + 0.2081415206193924, + 0.04659409075975418, + 0.03710492327809334, + 0.21904270350933075, + 0.17668698728084564, + 0.060200706124305725, + 0.046012599021196365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.0009203489171341062 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a horizontal angle, the camera held still, its focus fixed squarely on a stoneware casserole set against a stark black void. A strong overhead light, precisely directed at the casserole's rounded, heavy, curved top, outlines its ancient contours and casts deep shadows across its rough, dark brown and terracotta-colored surface, highlighting the fine pores and grain left by firing. Suddenly, a hairline white crack silently shatters the casserole's tranquility. Then, as if triggering a chain reaction, countless cracks spread outward from this point with astonishing speed, instantly forming a dense spiderweb that shimmers with a pale white glow in the light. In the instant the cracks spread, the entire casserole's structure reaches a breaking point. Rather than bursting outward, it violently and dramatically sags inward, collapsing inward. The solid clay instantly disintegrated under the enormous inward pressure, turning into a cloud of dust mixed with tiny fragments and a large amount of dark powder. This chaotic cloud illuminated by the top light rolled and expanded briefly in place, then slowly drifted downwards due to the loss of all support, and finally completely disappeared in the endless darkness below, leaving only nothingness.", + "seed": 353, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_15/index_5393_seed_353_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01467558927834034, + "vlm_logprob": -0.01467558927834034, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.01467558927834034, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a stoneware casserole against a stark black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the casserole's surface dark brown and terracotta-colored with visible fine pores and grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline white crack appear and instantly spread into a dense, shimmering spiderweb pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casserole violently sag inward and collapse rather than bursting outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the collapsing structure turn into a cloud of dust and dark powder that drifts downwards and disappears into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02229156717658043 + ], + "opposite_logprob_tokens": [ + -4.022291660308838 + ], + "logprob_reward": -0.02229156717658043, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02229156717658043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16586938500404358, + 0.03860635310411453, + 0.11800571531057358, + 0.19386155903339386, + 0.21225804090499878, + 0.05507481470704079, + 0.050061918795108795, + 0.04146832227706909, + 0.02919585444033146, + 0.095598004758358 + ], + "credit_map": [ + 0.01821240410208702, + 0.025538966059684753, + 0.017239317297935486, + 0.007871881127357483, + 0.03230994939804077, + 0.03050517663359642, + 0.009947966784238815, + 0.024243729189038277, + 0.004554819781333208, + 0.006016560830175877, + 0.004312219098210335, + 0.0018931523663923144, + 0.0071654184721410275, + 0.006937694735825062, + 0.0019015774596482515, + 0.0058249132707715034, + 0.017895422875881195, + 0.014226512052118778, + 0.009912673383951187, + 0.006920886691659689, + 0.02343122474849224, + 0.027400385588407516, + 0.0069758836179971695, + 0.011242733336985111, + 0.012612877413630486, + 0.029421431943774223, + 0.025151081383228302, + 0.011580009013414383, + 0.02912156470119953, + 0.02901025488972664, + 0.012327052652835846, + 0.04463730752468109, + 0.01048266515135765, + 0.01418117992579937, + 0.0336013063788414, + 0.04867815971374512, + 0.01671282947063446, + 0.021809672936797142, + 0.04412813112139702, + 0.022664101794362068, + 0.003990177530795336, + 0.005241701379418373, + 0.007538262754678726, + 0.007735996972769499, + 0.005192551761865616, + 0.008516517467796803, + 0.010723482817411423, + 0.0061361296102404594, + 0.003692605998367071, + 0.003946099895983934, + 0.0069907354190945625, + 0.009227896109223366, + 0.004488623235374689, + 0.004938179161399603, + 0.009339636191725731, + 0.007438146974891424, + 0.004810071550309658, + 0.002969995839521289, + 0.006210428196936846, + 0.006489179562777281, + 0.0032481730449944735, + 0.0044483221136033535, + 0.007362715434283018, + 0.005929435137659311, + 0.00333545939065516, + 0.0025298430118709803, + 0.0039558312855660915, + 0.003771427320316434, + 0.002473606960847974, + 0.004008539021015167, + 0.005437039770185947, + 0.003684109775349498, + 0.02665529027581215, + 0.005393179599195719, + 0.005355951376259327, + 0.005413728766143322, + 0.006628529168665409, + 0.00847618654370308, + 0.004988720174878836, + 0.03268643096089363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.02229156717658043 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a stoneware casserole against a stark black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the casserole's surface dark brown and terracotta-colored with visible fine pores and grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline white crack appear and instantly spread into a dense, shimmering spiderweb pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casserole violently sag inward and collapse rather than bursting outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the collapsing structure turn into a cloud of dust and dark powder that drifts downwards and disappears into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008658486418426037 + ], + "opposite_logprob_tokens": [ + -4.758658409118652 + ], + "logprob_reward": -0.008658486418426037, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008658486418426037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04990337789058685, + 0.032767556607723236, + 0.008315582759678364, + 0.06692828983068466, + 0.40302231907844543, + 0.057512685656547546, + 0.04011894017457962, + 0.037358060479164124, + 0.1204676479101181, + 0.18360552191734314 + ], + "credit_map": [ + 0.005557339638471603, + 0.011258346028625965, + 0.003602436510846019, + 0.0027089440263807774, + 0.010280060581862926, + 0.007013032678514719, + 0.0033612868282943964, + 0.006121932994574308, + 0.0043363203294575214, + 0.0036182450130581856, + 0.002863492351025343, + 0.0017771213315427303, + 0.0059568677097558975, + 0.007717466447502375, + 0.0019188300939276814, + 0.004579213447868824, + 0.0011825195979326963, + 0.00099984894040972, + 0.0009236266487278044, + 0.0007536067860201001, + 0.001634008134715259, + 0.0010465156519785523, + 0.0006659124628640711, + 0.001109544187784195, + 0.005982512142509222, + 0.008281641639769077, + 0.007807181682437658, + 0.004799268674105406, + 0.011812337674200535, + 0.01356691587716341, + 0.005986915901303291, + 0.008691519498825073, + 0.026053745299577713, + 0.03793597221374512, + 0.06377393752336502, + 0.0724620372056961, + 0.04011369124054909, + 0.03955429792404175, + 0.06757025420665741, + 0.05555839091539383, + 0.004204488825052977, + 0.0041610742919147015, + 0.008054257370531559, + 0.008381150662899017, + 0.005557816941291094, + 0.00889719557017088, + 0.012153275310993195, + 0.006103426218032837, + 0.004126216750591993, + 0.0030024389270693064, + 0.006008790340274572, + 0.008176947012543678, + 0.0037410184741020203, + 0.0035061996895819902, + 0.00619305856525898, + 0.005364268086850643, + 0.004303228575736284, + 0.0022742641158401966, + 0.00536869652569294, + 0.005090675316751003, + 0.00344482297077775, + 0.003121789079159498, + 0.010442228056490421, + 0.0033123590983450413, + 0.011309643276035786, + 0.006017743144184351, + 0.025226591154932976, + 0.011127816513180733, + 0.00570336077362299, + 0.006298533640801907, + 0.04165664687752724, + 0.013127317652106285, + 0.04626348614692688, + 0.012233713641762733, + 0.012829147279262543, + 0.017982520163059235, + 0.01354961283504963, + 0.014685418456792831, + 0.013624086044728756, + 0.05243752896785736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.008658486418426037 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a stoneware casserole against a stark black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the casserole's surface dark brown and terracotta-colored with visible fine pores and grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline white crack appear and instantly spread into a dense, shimmering spiderweb pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casserole violently sag inward and collapse rather than bursting outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the collapsing structure turn into a cloud of dust and dark powder that drifts downwards and disappears into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00409882515668869 + ], + "opposite_logprob_tokens": [ + -5.504098892211914 + ], + "logprob_reward": -0.00409882515668869, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00409882515668869, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046299345791339874, + 0.01819123700261116, + 0.023712672293186188, + 0.06529932469129562, + 0.07594051957130432, + 0.057499248534440994, + 0.04712900519371033, + 0.015532359480857849, + 0.039939023554325104, + 0.6104573011398315 + ], + "credit_map": [ + 0.005965523887425661, + 0.007647466380149126, + 0.004391197580844164, + 0.002331761410459876, + 0.010828694328665733, + 0.007941990159451962, + 0.0024857514072209597, + 0.004706955514848232, + 0.0017260601744055748, + 0.00208633323200047, + 0.001751625444740057, + 0.0010430279653519392, + 0.0037063229829072952, + 0.0040534124709665775, + 0.0009313134360127151, + 0.002893139375373721, + 0.0025091427378356457, + 0.002960948972031474, + 0.002201490104198456, + 0.0014572624349966645, + 0.004965931177139282, + 0.0047838217578828335, + 0.0013607292203232646, + 0.0034733451902866364, + 0.003278065240010619, + 0.012265139259397984, + 0.00794871523976326, + 0.003864435711875558, + 0.006905042566359043, + 0.005809485446661711, + 0.0046832747757434845, + 0.020545165985822678, + 0.005582550074905157, + 0.006324043031781912, + 0.012349735014140606, + 0.014178793877363205, + 0.006991054397076368, + 0.007513353135436773, + 0.009545888751745224, + 0.01345509011298418, + 0.004322628024965525, + 0.004570408258587122, + 0.008395213633775711, + 0.008582312613725662, + 0.005742120090872049, + 0.012196202762424946, + 0.007657462265342474, + 0.006032893434166908, + 0.0031895549036562443, + 0.003966624848544598, + 0.007797862868756056, + 0.010260612703859806, + 0.0030888859182596207, + 0.0046773385256528854, + 0.007099330425262451, + 0.007048796396702528, + 0.002172634471207857, + 0.0015664976090192795, + 0.001896157395094633, + 0.0017253689002245665, + 0.001493352116085589, + 0.0025692894123494625, + 0.0024475441314280033, + 0.0016615132335573435, + 0.004293893929570913, + 0.0035066946875303984, + 0.005012781359255314, + 0.005997675471007824, + 0.00360151007771492, + 0.005002238322049379, + 0.007691324222832918, + 0.004832903388887644, + 0.16517223417758942, + 0.03912803530693054, + 0.03587290272116661, + 0.05206277593970299, + 0.04013209417462349, + 0.050727106630802155, + 0.04237039387226105, + 0.18499170243740082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 3 + ], + "selected_reward": -0.00409882515668869 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a stoneware casserole against a stark black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the casserole's surface dark brown and terracotta-colored with visible fine pores and grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline white crack appear and instantly spread into a dense, shimmering spiderweb pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casserole violently sag inward and collapse rather than bursting outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the collapsing structure turn into a cloud of dust and dark powder that drifts downwards and disappears into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369222208857536 + ], + "opposite_logprob_tokens": [ + -3.4086921215057373 + ], + "logprob_reward": -0.03369222208857536, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369222208857536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04312821105122566, + 0.019120199605822563, + 0.013299886137247086, + 0.08606904000043869, + 0.07191584259271622, + 0.16993039846420288, + 0.05590590462088585, + 0.03724643588066101, + 0.07767236977815628, + 0.4257117211818695 + ], + "credit_map": [ + 0.006079456303268671, + 0.009085504338145256, + 0.004127697553485632, + 0.002366321161389351, + 0.006568397860974073, + 0.00535275274887681, + 0.0027300254441797733, + 0.006818055175244808, + 0.0022527778055518866, + 0.003423872170969844, + 0.0015543403569608927, + 0.001093145925551653, + 0.002807850018143654, + 0.004504095762968063, + 0.0012929170625284314, + 0.0021912003867328167, + 0.0018954882398247719, + 0.001967299496755004, + 0.0016902906354516745, + 0.0008537540561519563, + 0.0028753147926181555, + 0.0018177866004407406, + 0.0007033821893855929, + 0.001496570068411529, + 0.0052472795359790325, + 0.013391070999205112, + 0.011826600879430771, + 0.004770481493324041, + 0.014135281555354595, + 0.010955632664263248, + 0.005110705271363258, + 0.020631983876228333, + 0.004917356185615063, + 0.006477166898548603, + 0.010958843864500523, + 0.011233551427721977, + 0.009958839043974876, + 0.008563772775232792, + 0.007760350126773119, + 0.012045959942042828, + 0.014988751150667667, + 0.017136575654149055, + 0.027010375633835793, + 0.022853780537843704, + 0.016474897041916847, + 0.02764483541250229, + 0.023138053715229034, + 0.020683113485574722, + 0.004056787583976984, + 0.005778450984507799, + 0.007576826494187117, + 0.009553463198244572, + 0.003730839816853404, + 0.005166545044630766, + 0.011354146525263786, + 0.008688845671713352, + 0.00385540840215981, + 0.0025810073129832745, + 0.00590365007519722, + 0.008443308994174004, + 0.003163986373692751, + 0.002665025880560279, + 0.006914125289767981, + 0.003719923784956336, + 0.007390647195279598, + 0.0037552714347839355, + 0.01503179594874382, + 0.005740817170590162, + 0.004351824056357145, + 0.0051150843501091, + 0.028369605541229248, + 0.00791732594370842, + 0.11572620272636414, + 0.028899632394313812, + 0.023461293429136276, + 0.029116738587617874, + 0.03822314739227295, + 0.038236103951931, + 0.02713853120803833, + 0.12491007894277573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 3 + ], + "selected_reward": -0.03369222208857536 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a stoneware casserole against a stark black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the casserole's surface dark brown and terracotta-colored with visible fine pores and grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline white crack appear and instantly spread into a dense, shimmering spiderweb pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casserole violently sag inward and collapse rather than bursting outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the collapsing structure turn into a cloud of dust and dark powder that drifts downwards and disappears into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004636845551431179 + ], + "opposite_logprob_tokens": [ + -5.379636764526367 + ], + "logprob_reward": -0.004636845551431179, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004636845551431179, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.023788917809724808, + 0.0037184134125709534, + 0.006031996104866266, + 0.010535604320466518, + 0.014355136081576347, + 0.051241181790828705, + 0.016282886266708374, + 0.018666749820113182, + 0.036809228360652924, + 0.8185698986053467 + ], + "credit_map": [ + 0.0026685644406825304, + 0.004380150232464075, + 0.0025517211761325598, + 0.001018824870698154, + 0.003920466173440218, + 0.0037421409506350756, + 0.0011576327960938215, + 0.0043494184501469135, + 0.00045924793812446296, + 0.00048753770533949137, + 0.0003092865808866918, + 0.00022628802980761975, + 0.0007209376199170947, + 0.0007121497765183449, + 0.00020867549756076187, + 0.0005942902062088251, + 0.001020575175061822, + 0.0007970505976118147, + 0.0006489746738225222, + 0.00034707842860370874, + 0.0013713858788833022, + 0.001005623140372336, + 0.0003592459543142468, + 0.0004820619651582092, + 0.0008250561077147722, + 0.0012988258386030793, + 0.0010520362993702292, + 0.0012885379837825894, + 0.0017392822774127126, + 0.0015309936134144664, + 0.001627960940822959, + 0.0011729116085916758, + 0.000692314759362489, + 0.0008685496868565679, + 0.0029437928460538387, + 0.0032887046691030264, + 0.0007663661381229758, + 0.001670646364800632, + 0.002627006731927395, + 0.0014977544778957963, + 0.0030918808188289404, + 0.00435365317389369, + 0.007814656011760235, + 0.007457322906702757, + 0.005126393865793943, + 0.010540514253079891, + 0.0076212771236896515, + 0.0052354843355715275, + 0.0011786362156271935, + 0.0017503724666312337, + 0.0028283726423978806, + 0.002694380935281515, + 0.0011945303995162249, + 0.00149908394087106, + 0.0025856022257357836, + 0.002551908604800701, + 0.0017383272061124444, + 0.0009500581654720008, + 0.0027582834009081125, + 0.00279421778395772, + 0.0016335322288796306, + 0.0015740891685709357, + 0.005544544663280249, + 0.0016736967954784632, + 0.003508545458316803, + 0.002558021107688546, + 0.00731210270896554, + 0.005841004196554422, + 0.0023936196230351925, + 0.002586780348792672, + 0.008324896916747093, + 0.00428425520658493, + 0.22749356925487518, + 0.05284048989415169, + 0.04668556898832321, + 0.06086032837629318, + 0.05729911848902702, + 0.06777626276016235, + 0.051414936780929565, + 0.2541996240615845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.004636845551431179 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a horizontal angle, the camera held still, its focus fixed squarely on a stoneware casserole set against a stark black void. A strong overhead light, precisely directed at the casserole's rounded, heavy, curved top, outlines its ancient contours and casts deep shadows across its rough, dark brown and terracotta-colored surface, highlighting the fine pores and grain left by firing. Suddenly, a hairline white crack silently shatters the casserole's tranquility. Then, as if triggering a chain reaction, countless cracks spread outward from this point with astonishing speed, instantly forming a dense spiderweb that shimmers with a pale white glow in the light. In the instant the cracks spread, the entire casserole's structure reaches a breaking point. Rather than bursting outward, it violently and dramatically sags inward, collapsing inward. The solid clay instantly disintegrated under the enormous inward pressure, turning into a cloud of dust mixed with tiny fragments and a large amount of dark powder. This chaotic cloud illuminated by the top light rolled and expanded briefly in place, then slowly drifted downwards due to the loss of all support, and finally completely disappeared in the endless darkness below, leaving only nothingness.", + "seed": 354, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_15/index_5393_seed_354_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.041416293708607554, + "vlm_logprob": -0.041416293708607554, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.041416293708607554, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a stoneware casserole against a stark black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the casserole's surface dark brown and terracotta-colored with visible fine pores and grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline white crack appear and instantly spread into a dense, shimmering spiderweb pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casserole violently sag inward and collapse rather than bursting outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the collapsing structure turn into a cloud of dust and dark powder that drifts downwards and disappears into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016276217997074127 + ], + "opposite_logprob_tokens": [ + -4.7662763595581055 + ], + "logprob_reward": -0.016276217997074127, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.016276217997074127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11629626899957657, + 0.03159331530332565, + 0.06340108066797256, + 0.054018981754779816, + 0.1884801983833313, + 0.06464361399412155, + 0.13318690657615662, + 0.06604944169521332, + 0.09789405763149261, + 0.1844361275434494 + ], + "credit_map": [ + 0.008350239135324955, + 0.011617281474173069, + 0.012219729833304882, + 0.012428640387952328, + 0.022333607077598572, + 0.013135646469891071, + 0.018550416454672813, + 0.01766071282327175, + 0.002463198732584715, + 0.0023688108194619417, + 0.0033863456919789314, + 0.004165996797382832, + 0.006804468110203743, + 0.0030634752474725246, + 0.00485618319362402, + 0.004484834615141153, + 0.005667275749146938, + 0.004521490074694157, + 0.006682403385639191, + 0.006981569807976484, + 0.012563020922243595, + 0.006971198599785566, + 0.01152842678129673, + 0.008485691621899605, + 0.0034064489882439375, + 0.0034901881590485573, + 0.00511551508679986, + 0.0064645977690815926, + 0.010157596319913864, + 0.005807471927255392, + 0.01240735873579979, + 0.007169800810515881, + 0.01671505905687809, + 0.01666523888707161, + 0.01912042684853077, + 0.028139308094978333, + 0.02753961645066738, + 0.029756231233477592, + 0.02625170163810253, + 0.024292604997754097, + 0.006403992883861065, + 0.004362786188721657, + 0.007251168135553598, + 0.009898195043206215, + 0.007561857812106609, + 0.008746684528887272, + 0.008001659065485, + 0.012417273595929146, + 0.011218184605240822, + 0.008303608745336533, + 0.01994916796684265, + 0.028701161965727806, + 0.015189913101494312, + 0.015021927654743195, + 0.019724013283848763, + 0.015078919939696789, + 0.003969611134380102, + 0.004995432682335377, + 0.003560972400009632, + 0.017020899802446365, + 0.007020765449851751, + 0.004138265270739794, + 0.02040521427989006, + 0.00493828346952796, + 0.008431293070316315, + 0.014275174587965012, + 0.013637486845254898, + 0.013304369524121284, + 0.007115209940820932, + 0.005885779391974211, + 0.016784479841589928, + 0.018460258841514587, + 0.015850534662604332, + 0.01825064979493618, + 0.013256552629172802, + 0.019417982548475266, + 0.025418048724532127, + 0.023928090929985046, + 0.04028242826461792, + 0.0280318520963192 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 6 + ], + "selected_reward": -0.016276217997074127 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a stoneware casserole against a stark black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the casserole's surface dark brown and terracotta-colored with visible fine pores and grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline white crack appear and instantly spread into a dense, shimmering spiderweb pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casserole violently sag inward and collapse rather than bursting outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the collapsing structure turn into a cloud of dust and dark powder that drifts downwards and disappears into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009800177067518234 + ], + "opposite_logprob_tokens": [ + -4.634799957275391 + ], + "logprob_reward": -0.009800177067518234, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009800177067518234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1211843341588974, + 0.06177697330713272, + 0.0746283233165741, + 0.21142150461673737, + 0.08482044190168381, + 0.04528544098138809, + 0.14748111367225647, + 0.08726592361927032, + 0.054133765399456024, + 0.11200214922428131 + ], + "credit_map": [ + 0.009421009570360184, + 0.009640991687774658, + 0.011019867844879627, + 0.01557267177850008, + 0.02404707483947277, + 0.012869172729551792, + 0.020618755370378494, + 0.01799478940665722, + 0.004153663758188486, + 0.005838502664119005, + 0.005051033105701208, + 0.006510801147669554, + 0.014334185980260372, + 0.006539055146276951, + 0.01053102221339941, + 0.00881870836019516, + 0.006947302725166082, + 0.005855280440300703, + 0.005099103786051273, + 0.006377215962857008, + 0.02170480787754059, + 0.009294677525758743, + 0.01014450378715992, + 0.009205431677401066, + 0.011931550689041615, + 0.01416576188057661, + 0.019272442907094955, + 0.026040514931082726, + 0.04592069983482361, + 0.02243495173752308, + 0.040182456374168396, + 0.031473129987716675, + 0.005921395029872656, + 0.006945576053112745, + 0.012457883916795254, + 0.014328636229038239, + 0.012954101897776127, + 0.010847892612218857, + 0.010379388928413391, + 0.010985570028424263, + 0.0034583345986902714, + 0.002895446727052331, + 0.0028673692140728235, + 0.014758522622287273, + 0.00425289710983634, + 0.005160266533493996, + 0.004842932801693678, + 0.007049669045954943, + 0.012132628820836544, + 0.01120476983487606, + 0.022823283448815346, + 0.025162020698189735, + 0.021655328571796417, + 0.013923835940659046, + 0.018291272222995758, + 0.022287972271442413, + 0.003797386772930622, + 0.00617796927690506, + 0.005534821189939976, + 0.023773323744535446, + 0.007936895824968815, + 0.005174554418772459, + 0.02721649594604969, + 0.007654479239135981, + 0.0037739863619208336, + 0.007499136496335268, + 0.007629032712429762, + 0.008880352601408958, + 0.005249627400189638, + 0.0038584894500672817, + 0.008835632354021072, + 0.00840750802308321, + 0.00994786899536848, + 0.012213644571602345, + 0.008414383977651596, + 0.011956749483942986, + 0.01575847901403904, + 0.017784131690859795, + 0.022811705246567726, + 0.013115180656313896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.009800177067518234 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a stoneware casserole against a stark black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the casserole's surface dark brown and terracotta-colored with visible fine pores and grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline white crack appear and instantly spread into a dense, shimmering spiderweb pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casserole violently sag inward and collapse rather than bursting outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the collapsing structure turn into a cloud of dust and dark powder that drifts downwards and disappears into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024925374891608953 + ], + "opposite_logprob_tokens": [ + -6.002492427825928 + ], + "logprob_reward": -0.0024925374891608953, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024925374891608953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05093028396368027, + 0.020393799990415573, + 0.02669687196612358, + 0.03274988755583763, + 0.08402547240257263, + 0.06937021762132645, + 0.5488661527633667, + 0.020330214872956276, + 0.05884210020303726, + 0.0877949669957161 + ], + "credit_map": [ + 0.004125666338950396, + 0.0048238663002848625, + 0.004792352672666311, + 0.00958783645182848, + 0.008353822864592075, + 0.005956633947789669, + 0.0070555671118199825, + 0.0062345401383936405, + 0.0017610390204936266, + 0.0017899578670039773, + 0.0018458198755979538, + 0.003398559056222439, + 0.004056368488818407, + 0.002642593812197447, + 0.002740711672231555, + 0.0021587524097412825, + 0.0023434071335941553, + 0.0019636854995042086, + 0.001982584595680237, + 0.002800989430397749, + 0.007421145215630531, + 0.0031641998793929815, + 0.0036517397966235876, + 0.0033691246062517166, + 0.0025548296980559826, + 0.0020134213846176863, + 0.0039012981578707695, + 0.003430658485740423, + 0.006062228232622147, + 0.0034847997594624758, + 0.00687999976798892, + 0.004422657657414675, + 0.007659071125090122, + 0.007083425298333168, + 0.006434283219277859, + 0.011122324503958225, + 0.010698909871280193, + 0.012609644792973995, + 0.01644478738307953, + 0.011973037384450436, + 0.004896387457847595, + 0.0036334716714918613, + 0.004850700963288546, + 0.021652596071362495, + 0.01033333595842123, + 0.006861282512545586, + 0.008006486110389233, + 0.009135959669947624, + 0.046671271324157715, + 0.03919421508908272, + 0.10270313918590546, + 0.09222850203514099, + 0.0735725536942482, + 0.05391979590058327, + 0.06900880485773087, + 0.07156795263290405, + 0.0013503443915396929, + 0.002222644165158272, + 0.0012736598728224635, + 0.0037506369408220053, + 0.003494426142424345, + 0.0015998802846297622, + 0.004859117791056633, + 0.0017795065650716424, + 0.005951055325567722, + 0.008739671669900417, + 0.006290340330451727, + 0.00810539722442627, + 0.007152985315769911, + 0.006708567962050438, + 0.007377321831882, + 0.008516762405633926, + 0.008036833256483078, + 0.009607866406440735, + 0.008246850222349167, + 0.010721671395003796, + 0.011487052775919437, + 0.010732141323387623, + 0.01844022423028946, + 0.010522338561713696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 4 + ], + "selected_reward": -0.0024925374891608953 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a stoneware casserole against a stark black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the casserole's surface dark brown and terracotta-colored with visible fine pores and grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline white crack appear and instantly spread into a dense, shimmering spiderweb pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casserole violently sag inward and collapse rather than bursting outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the collapsing structure turn into a cloud of dust and dark powder that drifts downwards and disappears into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16032616794109344 + ], + "opposite_logprob_tokens": [ + -1.9103261232376099 + ], + "logprob_reward": -0.16032616794109344, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16032616794109344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05858376994729042, + 0.03318633511662483, + 0.03047746606171131, + 0.013973359018564224, + 0.03711773082613945, + 0.02410210855305195, + 0.4856586158275604, + 0.11390300840139389, + 0.07816796004772186, + 0.12482969462871552 + ], + "credit_map": [ + 0.004981016740202904, + 0.005445179063826799, + 0.005495315417647362, + 0.010936164297163486, + 0.009811827912926674, + 0.007031654939055443, + 0.007339243311434984, + 0.0075433701276779175, + 0.0026074424386024475, + 0.002629384398460388, + 0.003151661716401577, + 0.00445570657029748, + 0.007365631405264139, + 0.003990960773080587, + 0.004607629030942917, + 0.004377919249236584, + 0.0029280357994139194, + 0.0027747589629143476, + 0.003121262416243553, + 0.002789194229990244, + 0.007589444052428007, + 0.003579007461667061, + 0.0038194013759493828, + 0.003876361297443509, + 0.0010489958804100752, + 0.0010156508069485426, + 0.0018966618226841092, + 0.002271488541737199, + 0.0021000339183956385, + 0.001153225195594132, + 0.002493356354534626, + 0.0019939462654292583, + 0.0027400744147598743, + 0.0023010806180536747, + 0.002535819076001644, + 0.005574191454797983, + 0.006134496070444584, + 0.0038304158952087164, + 0.010324571281671524, + 0.003677083645015955, + 0.0019137965282425284, + 0.001914888620376587, + 0.001939298352226615, + 0.0032626816537231207, + 0.0034216013737022877, + 0.002956962212920189, + 0.0033791246823966503, + 0.005313753616064787, + 0.04229399189352989, + 0.03386114537715912, + 0.0922263041138649, + 0.08351776003837585, + 0.06334327906370163, + 0.049007292836904526, + 0.06033291295170784, + 0.06107591837644577, + 0.008942293003201485, + 0.013432360254228115, + 0.007545486558228731, + 0.016998151317238808, + 0.019450627267360687, + 0.00897511187940836, + 0.022078244015574455, + 0.01648072898387909, + 0.006839938927441835, + 0.008934112265706062, + 0.010268640704452991, + 0.0168895423412323, + 0.006237098481506109, + 0.004636525176465511, + 0.01188929658383131, + 0.012472808361053467, + 0.01209668442606926, + 0.012780404649674892, + 0.0102433692663908, + 0.014667083509266376, + 0.016279127448797226, + 0.013836843892931938, + 0.028599735349416733, + 0.016326455399394035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.16032616794109344 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a stoneware casserole against a stark black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the casserole's surface dark brown and terracotta-colored with visible fine pores and grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline white crack appear and instantly spread into a dense, shimmering spiderweb pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casserole violently sag inward and collapse rather than bursting outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the collapsing structure turn into a cloud of dust and dark powder that drifts downwards and disappears into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818636804819107 + ], + "opposite_logprob_tokens": [ + -4.018186569213867 + ], + "logprob_reward": -0.01818636804819107, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818636804819107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12215178459882736, + 0.0426640622317791, + 0.043353550136089325, + 0.04695191606879234, + 0.03666584938764572, + 0.07343349605798721, + 0.08323848992586136, + 0.0911150574684143, + 0.2439729869365692, + 0.21645280718803406 + ], + "credit_map": [ + 0.00839921273291111, + 0.013479921966791153, + 0.010276294313371181, + 0.019757255911827087, + 0.02128337323665619, + 0.015704117715358734, + 0.015656601637601852, + 0.01759500242769718, + 0.0031578480266034603, + 0.003421713365241885, + 0.004587747622281313, + 0.006352643482387066, + 0.007511680945754051, + 0.005672759842127562, + 0.0060996138490736485, + 0.0058600506745278835, + 0.004125853534787893, + 0.0037369898054748774, + 0.0035792242269963026, + 0.0040757195092737675, + 0.011565811932086945, + 0.004965223837643862, + 0.005336342845112085, + 0.005968383979052305, + 0.003543640486896038, + 0.0036482117138803005, + 0.004854322876781225, + 0.005195004399865866, + 0.009674920700490475, + 0.004625643137842417, + 0.008141960948705673, + 0.0072682141326367855, + 0.0033125088084489107, + 0.002928736386820674, + 0.005526637192815542, + 0.005235784687101841, + 0.00495095644146204, + 0.0045026750303804874, + 0.006014508195221424, + 0.004194043111056089, + 0.006356801837682724, + 0.0035389389377087355, + 0.005027005448937416, + 0.02160029485821724, + 0.009009531699120998, + 0.011196971870958805, + 0.007346468977630138, + 0.009357484988868237, + 0.006471524480730295, + 0.006071140058338642, + 0.0059870462864637375, + 0.024069136008620262, + 0.008470128290355206, + 0.008447813801467419, + 0.012115619145333767, + 0.011606083251535892, + 0.007263438776135445, + 0.011515467427670956, + 0.008061254397034645, + 0.016756393015384674, + 0.013765291310846806, + 0.008329748176038265, + 0.015534508973360062, + 0.009888955391943455, + 0.021644342690706253, + 0.024701349437236786, + 0.041396722197532654, + 0.042704541236162186, + 0.019665950909256935, + 0.013505673967301846, + 0.043811455368995667, + 0.03654297813773155, + 0.024551549926400185, + 0.02713032066822052, + 0.019277341663837433, + 0.02710418403148651, + 0.02887757308781147, + 0.023912737146019936, + 0.04275239259004593, + 0.022846711799502373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.01818636804819107 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_2/index_tensor([3703, 3703])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_2/index_tensor([3703, 3703])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3b2f6b29f8c343c406b5ced7e34699861b3e0576 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_2/index_tensor([3703, 3703])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The film begins with a close-up of a wall clock, its hour and minute hands spinning at a visible speed, creating a series of blurry afterimages. The camera then smoothly pulls back to reveal a dimly lit children's study. At the center of the frame, a boy, approximately ten years old, wearing a blue and white striped T-shirt, hunched over a wooden desk piled high with textbooks and workbooks. He desperately tugs his hands into his tousled black hair, his fingers tugging at the roots, his body trembling with silent frustration. An old-fashioned desk lamp provides the sole source of light, its warm yellow glow illuminating the mess on the desk while casting long, deep shadows across the rest of the room. Behind him, a mother in a gray cardigan stands quietly at the edge of the shadows, her brow furrowed, her lips pressed into a serious line, her gaze darting back and forth intently at the thin report card in her hand. The scene creates a stark contrast between the rapid passage of time and the characters' frozen anxiety, creating a suffocating, stressful domestic atmosphere.", + "seed": 15961, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_2/index_3703_seed_15961_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.039717947691679, + "vlm_logprob": -0.039717947691679, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.039717947691679, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up of a wall clock with spinning hands creating blurry afterimages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back to reveal a dimly lit children's study?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy approximately ten years old wearing a blue and white striped T-shirt visible hunched over a wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy shown tugging at his tousled black hair while his body trembles with frustration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a mother in a gray cardigan visible standing in the shadows holding a thin report card?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020911574363708496 + ], + "opposite_logprob_tokens": [ + -4.520911693572998 + ], + "logprob_reward": -0.020911574363708496, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.020911574363708496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09154132753610611, + 0.19146892428398132, + 0.19927309453487396, + 0.07646284997463226, + 0.04870514199137688, + 0.0671609416604042, + 0.09986136853694916, + 0.06267732381820679, + 0.07600641995668411, + 0.08684255927801132 + ], + "credit_map": [ + 0.010394011624157429, + 0.008091898635029793, + 0.01226435974240303, + 0.01864100806415081, + 0.012675223872065544, + 0.00828314758837223, + 0.010659720748662949, + 0.010531971231102943, + 0.021406782791018486, + 0.035839471966028214, + 0.013798472471535206, + 0.029642339795827866, + 0.024520697072148323, + 0.02195928432047367, + 0.030170105397701263, + 0.01413178164511919, + 0.021484604105353355, + 0.029748626053333282, + 0.032383181154727936, + 0.025379346683621407, + 0.02675098553299904, + 0.02799166738986969, + 0.01857774704694748, + 0.016956964507699013, + 0.006478157360106707, + 0.008803059346973896, + 0.006256727501749992, + 0.00532917445525527, + 0.02207377552986145, + 0.016383403912186623, + 0.00737800681963563, + 0.0037605531979352236, + 0.006135435774922371, + 0.00455354992300272, + 0.00705352146178484, + 0.0029492138419300318, + 0.008291417732834816, + 0.007569962181150913, + 0.007738032378256321, + 0.004414013586938381, + 0.011586378328502178, + 0.0073997583240270615, + 0.01332476083189249, + 0.004741974174976349, + 0.009070293046534061, + 0.010949800722301006, + 0.005178049672394991, + 0.004909931216388941, + 0.011492275632917881, + 0.016936225816607475, + 0.014783008024096489, + 0.00820935145020485, + 0.013566665351390839, + 0.017993947491049767, + 0.00975565705448389, + 0.007124242838472128, + 0.006399292033165693, + 0.009334198199212551, + 0.010560188442468643, + 0.0058400994166731834, + 0.010037267580628395, + 0.008950464427471161, + 0.00597362918779254, + 0.005582189187407494, + 0.00821485836058855, + 0.014028575271368027, + 0.004948827903717756, + 0.005152133293449879, + 0.01589915342628956, + 0.012507469393312931, + 0.010265943594276905, + 0.004989461973309517, + 0.005862353835254908, + 0.020156770944595337, + 0.004591996315866709, + 0.005648827645927668, + 0.01710735447704792, + 0.015436840243637562, + 0.012121262028813362, + 0.005917153786867857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 6 + ], + "selected_reward": -0.020911574363708496 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up of a wall clock with spinning hands creating blurry afterimages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back to reveal a dimly lit children's study?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy approximately ten years old wearing a blue and white striped T-shirt visible hunched over a wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy shown tugging at his tousled black hair while his body trembles with frustration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a mother in a gray cardigan visible standing in the shadows holding a thin report card?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016107548028230667 + ], + "opposite_logprob_tokens": [ + -4.141107559204102 + ], + "logprob_reward": -0.016107548028230667, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016107548028230667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08152272552251816, + 0.22053204476833344, + 0.2879795730113983, + 0.09488426148891449, + 0.03392665088176727, + 0.11173185706138611, + 0.05940062552690506, + 0.0315905436873436, + 0.03153584152460098, + 0.04689587280154228 + ], + "credit_map": [ + 0.007567893713712692, + 0.0069880178198218346, + 0.009540191851556301, + 0.01514007430523634, + 0.009744731709361076, + 0.00770438089966774, + 0.013885672204196453, + 0.01095175463706255, + 0.013257535174489021, + 0.01782398670911789, + 0.04555042088031769, + 0.024253038689494133, + 0.02200051210820675, + 0.024209244176745415, + 0.052446071058511734, + 0.020991232246160507, + 0.03406400978565216, + 0.049805041402578354, + 0.04251663014292717, + 0.032477013766765594, + 0.0337938591837883, + 0.04899997636675835, + 0.025308890268206596, + 0.02101415954530239, + 0.010947732254862785, + 0.010272536426782608, + 0.01047657523304224, + 0.007045441307127476, + 0.022981692105531693, + 0.01940218172967434, + 0.008603258058428764, + 0.005154841113835573, + 0.004314318764954805, + 0.0031260650139302015, + 0.006151864770799875, + 0.0026140406262129545, + 0.0054578776471316814, + 0.0052479770965874195, + 0.0037856576964259148, + 0.003228851594030857, + 0.019978521391749382, + 0.010130681097507477, + 0.02640160545706749, + 0.007652278523892164, + 0.013176387175917625, + 0.019539255648851395, + 0.008311914280056953, + 0.006541216280311346, + 0.008406145498156548, + 0.006090413313359022, + 0.010774594731628895, + 0.0043759713880717754, + 0.009279359132051468, + 0.009780548512935638, + 0.007088659331202507, + 0.003604931989684701, + 0.004277629777789116, + 0.005032106768339872, + 0.0032260261941701174, + 0.0025790606159716845, + 0.006084210239350796, + 0.004702345933765173, + 0.003815615549683571, + 0.001873549772426486, + 0.004486836027354002, + 0.007194077596068382, + 0.002797281602397561, + 0.0026912372559309006, + 0.003865753998979926, + 0.0046675982885062695, + 0.0026740727480500937, + 0.003158981679007411, + 0.0037026123609393835, + 0.012086844071745872, + 0.0027833578642457724, + 0.0028462314512580633, + 0.008750810287892818, + 0.008266481570899487, + 0.005361592397093773, + 0.003097943728789687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 5 + ], + "selected_reward": -0.016107548028230667 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up of a wall clock with spinning hands creating blurry afterimages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back to reveal a dimly lit children's study?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy approximately ten years old wearing a blue and white striped T-shirt visible hunched over a wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy shown tugging at his tousled black hair while his body trembles with frustration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a mother in a gray cardigan visible standing in the shadows holding a thin report card?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018195733428001404 + ], + "opposite_logprob_tokens": [ + -4.018195629119873 + ], + "logprob_reward": -0.018195733428001404, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018195733428001404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05693887174129486, + 0.0951932743191719, + 0.16422317922115326, + 0.06848561018705368, + 0.03821415454149246, + 0.08077152073383331, + 0.11504817754030228, + 0.24048830568790436, + 0.06555648893117905, + 0.07508045434951782 + ], + "credit_map": [ + 0.005238620098680258, + 0.004204554483294487, + 0.008553605526685715, + 0.010668061673641205, + 0.00644217524677515, + 0.006250421050935984, + 0.007763155270367861, + 0.00781827699393034, + 0.009706754237413406, + 0.014802120625972748, + 0.012572170235216618, + 0.011816318146884441, + 0.015026136301457882, + 0.01174449548125267, + 0.012459756806492805, + 0.007065523415803909, + 0.01802026852965355, + 0.026000140234827995, + 0.028922997415065765, + 0.01607382670044899, + 0.022845668718218803, + 0.023989759385585785, + 0.014317217282950878, + 0.014053299091756344, + 0.006997036747634411, + 0.007409470621496439, + 0.007614519447088242, + 0.005767002701759338, + 0.014587977901101112, + 0.012928228825330734, + 0.007610175292938948, + 0.005571200046688318, + 0.0038936331402510405, + 0.005334608256816864, + 0.0029041774105280638, + 0.002530379220843315, + 0.007687215227633715, + 0.006087441463023424, + 0.0063129449263215065, + 0.0034637567587196827, + 0.008596415631473064, + 0.009010080248117447, + 0.007536835502833128, + 0.00849402416497469, + 0.014210174791514874, + 0.012971844524145126, + 0.010621295310556889, + 0.009330845437943935, + 0.016207071021199226, + 0.013186285272240639, + 0.020670583471655846, + 0.008441811427474022, + 0.018910309299826622, + 0.018259838223457336, + 0.01289183646440506, + 0.006480430718511343, + 0.02795134112238884, + 0.03514128550887108, + 0.03423633053898811, + 0.019302217289805412, + 0.028762508183717728, + 0.03489295020699501, + 0.040124811232089996, + 0.02007685787975788, + 0.005783964414149523, + 0.009333056397736073, + 0.008752136491239071, + 0.0058934190310537815, + 0.012828183360397816, + 0.010155601426959038, + 0.008225971832871437, + 0.004584155976772308, + 0.006382010877132416, + 0.016921866685152054, + 0.004630180075764656, + 0.00502815842628479, + 0.014994166791439056, + 0.012643737718462944, + 0.009343009442090988, + 0.005137322936207056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 6 + ], + "selected_reward": -0.018195733428001404 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up of a wall clock with spinning hands creating blurry afterimages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back to reveal a dimly lit children's study?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy approximately ten years old wearing a blue and white striped T-shirt visible hunched over a wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy shown tugging at his tousled black hair while his body trembles with frustration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a mother in a gray cardigan visible standing in the shadows holding a thin report card?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430726483464241 + ], + "opposite_logprob_tokens": [ + -3.1680727005004883 + ], + "logprob_reward": -0.0430726483464241, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430726483464241, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04266699030995369, + 0.17572592198848724, + 0.09311820566654205, + 0.2285141497850418, + 0.1726582646369934, + 0.08722423762083054, + 0.05513632670044899, + 0.03639699146151543, + 0.06874479353427887, + 0.039814114570617676 + ], + "credit_map": [ + 0.0034652354661375284, + 0.0032034520991146564, + 0.008045952767133713, + 0.0077455476857721806, + 0.0052612703293561935, + 0.004771058913320303, + 0.004323032218962908, + 0.005851440597325563, + 0.013489587232470512, + 0.015509435907006264, + 0.014728331938385963, + 0.024545155465602875, + 0.012281566858291626, + 0.02006715163588524, + 0.06134212017059326, + 0.01376256998628378, + 0.00971885398030281, + 0.015139620751142502, + 0.011926781386137009, + 0.009467964991927147, + 0.012082788161933422, + 0.012699547223746777, + 0.013365037739276886, + 0.008717621676623821, + 0.01914646290242672, + 0.02171640284359455, + 0.02387670800089836, + 0.016822446137666702, + 0.06102587282657623, + 0.042881421744823456, + 0.027471942827105522, + 0.015572880394756794, + 0.018326185643672943, + 0.023741962388157845, + 0.01412076223641634, + 0.0126944063231349, + 0.029327228665351868, + 0.032293789088726044, + 0.025886187329888344, + 0.016267739236354828, + 0.012553585693240166, + 0.007926332764327526, + 0.01619698666036129, + 0.00772499106824398, + 0.012476561591029167, + 0.01466638408601284, + 0.0083309980109334, + 0.007348398678004742, + 0.00841591227799654, + 0.005571029614657164, + 0.010386879555881023, + 0.005017932970076799, + 0.006407223641872406, + 0.009253819473087788, + 0.006263226270675659, + 0.0038203035946935415, + 0.0038532838225364685, + 0.007795647252351046, + 0.003456560894846916, + 0.003264250699430704, + 0.006239489186555147, + 0.005221433937549591, + 0.003656799905002117, + 0.0029095252975821495, + 0.0064994352869689465, + 0.009801767766475677, + 0.006871937308460474, + 0.005297672003507614, + 0.016163449734449387, + 0.011085819453001022, + 0.0077764010056853294, + 0.0052483160980045795, + 0.0036636122968047857, + 0.012738056480884552, + 0.002007837640121579, + 0.002151944674551487, + 0.004903682041913271, + 0.007907910272479057, + 0.0035900981165468693, + 0.0028509744442999363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 4 + ], + "selected_reward": -0.0430726483464241 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up of a wall clock with spinning hands creating blurry afterimages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back to reveal a dimly lit children's study?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy approximately ten years old wearing a blue and white striped T-shirt visible hunched over a wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy shown tugging at his tousled black hair while his body trembles with frustration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a mother in a gray cardigan visible standing in the shadows holding a thin report card?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030223429203033 + ], + "opposite_logprob_tokens": [ + -2.350302219390869 + ], + "logprob_reward": -0.10030223429203033, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030223429203033, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10469311475753784, + 0.14205127954483032, + 0.09203844517469406, + 0.058537889271974564, + 0.08376636356115341, + 0.08992891013622284, + 0.04964187368750572, + 0.10116010159254074, + 0.041422344744205475, + 0.23675967752933502 + ], + "credit_map": [ + 0.009122160263359547, + 0.008956460282206535, + 0.010827001184225082, + 0.023289814591407776, + 0.012369999662041664, + 0.009150818921625614, + 0.016212647780776024, + 0.014764208346605301, + 0.010751956142485142, + 0.008760902099311352, + 0.007661924231797457, + 0.021590979769825935, + 0.010287243872880936, + 0.016558131203055382, + 0.053280968219041824, + 0.013159177266061306, + 0.012403634376823902, + 0.01251106709241867, + 0.01614702120423317, + 0.010600846260786057, + 0.013615282252430916, + 0.010286795906722546, + 0.00895433034747839, + 0.007519471924751997, + 0.004658516496419907, + 0.007115357089787722, + 0.0048727961257100105, + 0.0033746687695384026, + 0.017514919862151146, + 0.011122922413051128, + 0.006141304038465023, + 0.0037374086678028107, + 0.008142402395606041, + 0.01210103090852499, + 0.0072395955212414265, + 0.004156747832894325, + 0.018631763756275177, + 0.01618705317378044, + 0.009563576430082321, + 0.007744194939732552, + 0.01231434103101492, + 0.007788648828864098, + 0.0137983039021492, + 0.00777962664142251, + 0.013645755127072334, + 0.01395219936966896, + 0.011782840825617313, + 0.008867189288139343, + 0.008957020938396454, + 0.008001725189387798, + 0.005380901508033276, + 0.004450206179171801, + 0.006586852017790079, + 0.008037347346544266, + 0.004724439233541489, + 0.0035033803433179855, + 0.011436937376856804, + 0.012958257459104061, + 0.009369372390210629, + 0.0073525863699615, + 0.020142508670687675, + 0.014525231905281544, + 0.01712857000529766, + 0.008246643468737602, + 0.0049947467632591724, + 0.00883184652775526, + 0.0038921746890991926, + 0.0037213803734630346, + 0.005171843804419041, + 0.005931363441050053, + 0.004245971795171499, + 0.004633018281310797, + 0.018208717927336693, + 0.06298427283763885, + 0.013041301630437374, + 0.015389608219265938, + 0.039900973439216614, + 0.042196448892354965, + 0.0294339582324028, + 0.015604403801262379 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.10030223429203033 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a close-up of a wall clock, its hour and minute hands spinning at a visible speed, creating a series of blurry afterimages. The camera then smoothly pulls back to reveal a dimly lit children's study. At the center of the frame, a boy, approximately ten years old, wearing a blue and white striped T-shirt, hunched over a wooden desk piled high with textbooks and workbooks. He desperately tugs his hands into his tousled black hair, his fingers tugging at the roots, his body trembling with silent frustration. An old-fashioned desk lamp provides the sole source of light, its warm yellow glow illuminating the mess on the desk while casting long, deep shadows across the rest of the room. Behind him, a mother in a gray cardigan stands quietly at the edge of the shadows, her brow furrowed, her lips pressed into a serious line, her gaze darting back and forth intently at the thin report card in her hand. The scene creates a stark contrast between the rapid passage of time and the characters' frozen anxiety, creating a suffocating, stressful domestic atmosphere.", + "seed": 15962, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_2/index_3703_seed_15962_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09589370805770159, + "vlm_logprob": -0.09589370805770159, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.09589370805770159, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up of a wall clock with spinning hands creating blurry afterimages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back to reveal a dimly lit children's study?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy approximately ten years old wearing a blue and white striped T-shirt visible hunched over a wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy shown tugging at his tousled black hair while his body trembles with frustration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a mother in a gray cardigan visible standing in the shadows holding a thin report card?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.040992457419633865 + ], + "opposite_logprob_tokens": [ + -3.540992498397827 + ], + "logprob_reward": -0.040992457419633865, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.040992457419633865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19599105417728424, + 0.3198843002319336, + 0.09559059888124466, + 0.07247473299503326, + 0.06504853069782257, + 0.02930017001926899, + 0.021287355571985245, + 0.031729914247989655, + 0.12280862778425217, + 0.04588474705815315 + ], + "credit_map": [ + 0.015114415436983109, + 0.02799147553741932, + 0.022104909643530846, + 0.033603888005018234, + 0.01632453128695488, + 0.018435383215546608, + 0.03330289199948311, + 0.029113557189702988, + 0.028631113469600677, + 0.03887258842587471, + 0.034895963966846466, + 0.025713486596941948, + 0.07364966720342636, + 0.07584583014249802, + 0.018533091992139816, + 0.023742571473121643, + 0.00803790707141161, + 0.008558650501072407, + 0.00876086950302124, + 0.019843274727463722, + 0.012029917910695076, + 0.014805484563112259, + 0.011110789142549038, + 0.012443705461919308, + 0.008212815970182419, + 0.010338862426578999, + 0.011182151734828949, + 0.012931815348565578, + 0.007935321889817715, + 0.005671790335327387, + 0.008280672132968903, + 0.007921301759779453, + 0.006382169667631388, + 0.009436462074518204, + 0.007854578085243702, + 0.0045297653414309025, + 0.003281746990978718, + 0.0173425804823637, + 0.004598802886903286, + 0.011622429825365543, + 0.0031915956642478704, + 0.0037936531007289886, + 0.004142064601182938, + 0.00254993187263608, + 0.004511286970227957, + 0.005024294834583998, + 0.0025195141788572073, + 0.0035678273998200893, + 0.004362763371318579, + 0.0024450861383229494, + 0.0030284321401268244, + 0.001663273898884654, + 0.005153184290975332, + 0.001679993118159473, + 0.0014131420757621527, + 0.0015414828667417169, + 0.0040516117587685585, + 0.005209442228078842, + 0.0045950328931212425, + 0.004124502185732126, + 0.0042474051006138325, + 0.0018714185571298003, + 0.005439270753413439, + 0.0021912348456680775, + 0.029984785243868828, + 0.009561830200254917, + 0.008410222828388214, + 0.0036401350516825914, + 0.04714629426598549, + 0.012717191129922867, + 0.005459542386233807, + 0.005888625513762236, + 0.007473784498870373, + 0.007055291440337896, + 0.008585627190768719, + 0.00877569429576397, + 0.0061076609417796135, + 0.0028960444033145905, + 0.002687331521883607, + 0.0023033132310956717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.040992457419633865 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up of a wall clock with spinning hands creating blurry afterimages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back to reveal a dimly lit children's study?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy approximately ten years old wearing a blue and white striped T-shirt visible hunched over a wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy shown tugging at his tousled black hair while his body trembles with frustration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a mother in a gray cardigan visible standing in the shadows holding a thin report card?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014250419102609158 + ], + "opposite_logprob_tokens": [ + -4.2642502784729 + ], + "logprob_reward": -0.014250419102609158, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014250419102609158, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09414432942867279, + 0.19392330944538116, + 0.12317045032978058, + 0.05227449908852577, + 0.0812932699918747, + 0.05804049223661423, + 0.04165080189704895, + 0.05752720683813095, + 0.2301163673400879, + 0.06785929203033447 + ], + "credit_map": [ + 0.011070407927036285, + 0.011149040423333645, + 0.011194671504199505, + 0.011777898296713829, + 0.00907097477465868, + 0.009801509790122509, + 0.018937386572360992, + 0.011142447590827942, + 0.010991011746227741, + 0.010630247183144093, + 0.01228866633027792, + 0.007907063700258732, + 0.06140996143221855, + 0.06935164332389832, + 0.007977988570928574, + 0.013366740196943283, + 0.012812475673854351, + 0.00999109260737896, + 0.01189518067985773, + 0.02752009406685829, + 0.010702770203351974, + 0.01804669015109539, + 0.014983481727540493, + 0.017218660563230515, + 0.005806917790323496, + 0.005309244617819786, + 0.00902062188833952, + 0.005961105227470398, + 0.00932299718260765, + 0.007546931505203247, + 0.005030250642448664, + 0.004276427906006575, + 0.010869679972529411, + 0.013025753200054169, + 0.011740047484636307, + 0.007467524614185095, + 0.008692911826074123, + 0.012625167146325111, + 0.006271135061979294, + 0.010601051151752472, + 0.006939948070794344, + 0.00719319935888052, + 0.007485189475119114, + 0.006390833295881748, + 0.010850080288946629, + 0.007688097655773163, + 0.005533705465495586, + 0.005959439091384411, + 0.007817975245416164, + 0.006972345989197493, + 0.005483240820467472, + 0.0034008193761110306, + 0.007036783266812563, + 0.004227675963193178, + 0.002962521044537425, + 0.0037494415882974863, + 0.008927209302783012, + 0.006054013967514038, + 0.009517712518572807, + 0.007321292068809271, + 0.0056675332598388195, + 0.0031801422592252493, + 0.01247527077794075, + 0.004384031519293785, + 0.051979124546051025, + 0.018895426765084267, + 0.02020956203341484, + 0.01723667047917843, + 0.07318177819252014, + 0.018612807616591454, + 0.012764796614646912, + 0.017236201092600822, + 0.011490856297314167, + 0.012456750497221947, + 0.01008226815611124, + 0.011633096262812614, + 0.008549444377422333, + 0.005163866560906172, + 0.004604124464094639, + 0.003878885181620717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 2 + ], + "selected_reward": -0.014250419102609158 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up of a wall clock with spinning hands creating blurry afterimages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back to reveal a dimly lit children's study?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy approximately ten years old wearing a blue and white striped T-shirt visible hunched over a wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy shown tugging at his tousled black hair while his body trembles with frustration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a mother in a gray cardigan visible standing in the shadows holding a thin report card?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029832107946276665 + ], + "opposite_logprob_tokens": [ + -3.529832124710083 + ], + "logprob_reward": -0.029832107946276665, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029832107946276665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0840751975774765, + 0.10816101729869843, + 0.028346173465251923, + 0.0173228457570076, + 0.056363727897405624, + 0.03914475440979004, + 0.04586532339453697, + 0.06765619665384293, + 0.38423019647598267, + 0.16883456707000732 + ], + "credit_map": [ + 0.007296680938452482, + 0.014289294369518757, + 0.00940120778977871, + 0.011135824955999851, + 0.010353359393775463, + 0.009494854137301445, + 0.013549811206758022, + 0.008554169908165932, + 0.004307095892727375, + 0.0049759200774133205, + 0.006469859275966883, + 0.005435247905552387, + 0.03688836097717285, + 0.040333040058612823, + 0.004403205122798681, + 0.005348296836018562, + 0.0030990219675004482, + 0.0033124692272394896, + 0.002442071447148919, + 0.0064078085124492645, + 0.0024891996290534735, + 0.003899046452715993, + 0.0036248695105314255, + 0.0030716850887984037, + 0.002119323005899787, + 0.002881720196455717, + 0.0020545904990285635, + 0.0035604650620371103, + 0.0019874051213264465, + 0.0012286377605050802, + 0.0015318652149289846, + 0.0019588391296565533, + 0.008491400629281998, + 0.010464475490152836, + 0.013179905712604523, + 0.004705602768808603, + 0.002785436110571027, + 0.007396461442112923, + 0.003041172167286277, + 0.006299270782619715, + 0.004516467917710543, + 0.004736160393804312, + 0.0046033672988414764, + 0.003215090837329626, + 0.008627520874142647, + 0.005532401148229837, + 0.004879102576524019, + 0.003034643130376935, + 0.006895002443343401, + 0.006296724081039429, + 0.009787379764020443, + 0.0047630141489207745, + 0.005984383635222912, + 0.0049553303979337215, + 0.003547350876033306, + 0.0036361373495310545, + 0.008778885938227177, + 0.011020662263035774, + 0.013807601295411587, + 0.010635207407176495, + 0.00840620044618845, + 0.0040595983155071735, + 0.007657433394342661, + 0.003290606662631035, + 0.09300708025693893, + 0.022420885041356087, + 0.02562941052019596, + 0.018170291557908058, + 0.14950895309448242, + 0.042475830763578415, + 0.01623433083295822, + 0.01678340695798397, + 0.02719282917678356, + 0.03952164202928543, + 0.019629288464784622, + 0.026288067921996117, + 0.02521049603819847, + 0.014593275263905525, + 0.00979752466082573, + 0.00660145515576005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 1 + ], + "selected_reward": -0.029832107946276665 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up of a wall clock with spinning hands creating blurry afterimages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back to reveal a dimly lit children's study?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy approximately ten years old wearing a blue and white striped T-shirt visible hunched over a wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy shown tugging at his tousled black hair while his body trembles with frustration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a mother in a gray cardigan visible standing in the shadows holding a thin report card?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11304169148206711 + ], + "opposite_logprob_tokens": [ + -2.238041639328003 + ], + "logprob_reward": -0.11304169148206711, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11304169148206711, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04740385338664055, + 0.13623180985450745, + 0.022379431873559952, + 0.023671669885516167, + 0.11508025974035263, + 0.05587682127952576, + 0.11280917376279831, + 0.07624530792236328, + 0.32265210151672363, + 0.08764958381652832 + ], + "credit_map": [ + 0.005007477011531591, + 0.006369059905409813, + 0.00576446857303381, + 0.005820582155138254, + 0.007243851665407419, + 0.0048960004933178425, + 0.007196336518973112, + 0.0051060752011835575, + 0.00516363512724638, + 0.006820983719080687, + 0.005443308502435684, + 0.004343442153185606, + 0.0512828603386879, + 0.052972424775362015, + 0.00445133401080966, + 0.005753832403570414, + 0.0021012125071138144, + 0.003090623999014497, + 0.002559301443397999, + 0.002875480568036437, + 0.0037221480160951614, + 0.0033577121794223785, + 0.0023916547652333975, + 0.002281299326568842, + 0.0025313543155789375, + 0.0028829346410930157, + 0.0031826093327254057, + 0.004202349577099085, + 0.0033313832245767117, + 0.00235914159566164, + 0.0024621596094220877, + 0.00271973735652864, + 0.01686420850455761, + 0.02230728417634964, + 0.021499045193195343, + 0.009176496416330338, + 0.005953630898147821, + 0.019507523626089096, + 0.006428182125091553, + 0.013343894854187965, + 0.0075136348605155945, + 0.007145612966269255, + 0.011064407415688038, + 0.006361696869134903, + 0.007079894654452801, + 0.007671060040593147, + 0.00457832682877779, + 0.004462187644094229, + 0.023974567651748657, + 0.016025295481085777, + 0.024046091362833977, + 0.011611922644078732, + 0.01873948983848095, + 0.005305218510329723, + 0.0064172265119850636, + 0.006689363159239292, + 0.011424162425100803, + 0.016381144523620605, + 0.013352559879422188, + 0.009774796664714813, + 0.010380761697888374, + 0.004385505802929401, + 0.006799959111958742, + 0.003746418282389641, + 0.06672582030296326, + 0.02354605495929718, + 0.022376833483576775, + 0.012414481490850449, + 0.12133482843637466, + 0.04602901637554169, + 0.013578091748058796, + 0.016646962612867355, + 0.013365127146244049, + 0.0162236075848341, + 0.014978347346186638, + 0.014581390656530857, + 0.01157284714281559, + 0.007565599866211414, + 0.005256596487015486, + 0.004106072708964348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 4 + ], + "selected_reward": -0.11304169148206711 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up of a wall clock with spinning hands creating blurry afterimages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back to reveal a dimly lit children's study?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy approximately ten years old wearing a blue and white striped T-shirt visible hunched over a wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy shown tugging at his tousled black hair while his body trembles with frustration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a mother in a gray cardigan visible standing in the shadows holding a thin report card?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28135186433792114 + ], + "opposite_logprob_tokens": [ + -1.4063518047332764 + ], + "logprob_reward": -0.28135186433792114, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28135186433792114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05617309361696243, + 0.20526841282844543, + 0.044548265635967255, + 0.0185075756162405, + 0.07043936848640442, + 0.07145335525274277, + 0.1538679599761963, + 0.1328335702419281, + 0.1021362766623497, + 0.14477218687534332 + ], + "credit_map": [ + 0.007960854098200798, + 0.007761203218251467, + 0.008205545134842396, + 0.0051016490906476974, + 0.006198665592819452, + 0.006320297252386808, + 0.006958057172596455, + 0.007666824851185083, + 0.007084400858730078, + 0.008938075974583626, + 0.007355124689638615, + 0.007578318007290363, + 0.07889703661203384, + 0.08041936904191971, + 0.006738867145031691, + 0.008257205598056316, + 0.0030966487247496843, + 0.005099352449178696, + 0.003459085011854768, + 0.009366015903651714, + 0.006296979263424873, + 0.006129922810941935, + 0.005643561016768217, + 0.0054566990584135056, + 0.002316669560968876, + 0.002443248638883233, + 0.002913211239501834, + 0.003340994706377387, + 0.0018944741459563375, + 0.0019117100164294243, + 0.001982436515390873, + 0.0017048291629180312, + 0.008566901087760925, + 0.011860311031341553, + 0.009617573581635952, + 0.005726885050535202, + 0.004051693715155125, + 0.015843208879232407, + 0.00430920859798789, + 0.01046358048915863, + 0.008485653437674046, + 0.008899278938770294, + 0.014612263068556786, + 0.008416169323027134, + 0.00829708855599165, + 0.008520293049514294, + 0.007478063460439444, + 0.006744538899511099, + 0.0256070327013731, + 0.030268976464867592, + 0.029000641778111458, + 0.015643035992980003, + 0.02937876060605049, + 0.008073119446635246, + 0.007556496653705835, + 0.008339894004166126, + 0.019579734653234482, + 0.020096231251955032, + 0.021904127672314644, + 0.020191846415400505, + 0.019674796611070633, + 0.008331061340868473, + 0.011731011793017387, + 0.011324767023324966, + 0.026124775409698486, + 0.007845123298466206, + 0.008377382531762123, + 0.004106443841010332, + 0.03733472898602486, + 0.006743215024471283, + 0.0057074264623224735, + 0.005897182505577803, + 0.02289162576198578, + 0.028603829443454742, + 0.025620531290769577, + 0.02389351651072502, + 0.018005499616265297, + 0.01087507326155901, + 0.007753224112093449, + 0.0071288831532001495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 9 + ], + "selected_reward": -0.28135186433792114 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a close-up of a wall clock, its hour and minute hands spinning at a visible speed, creating a series of blurry afterimages. The camera then smoothly pulls back to reveal a dimly lit children's study. At the center of the frame, a boy, approximately ten years old, wearing a blue and white striped T-shirt, hunched over a wooden desk piled high with textbooks and workbooks. He desperately tugs his hands into his tousled black hair, his fingers tugging at the roots, his body trembling with silent frustration. An old-fashioned desk lamp provides the sole source of light, its warm yellow glow illuminating the mess on the desk while casting long, deep shadows across the rest of the room. Behind him, a mother in a gray cardigan stands quietly at the edge of the shadows, her brow furrowed, her lips pressed into a serious line, her gaze darting back and forth intently at the thin report card in her hand. The scene creates a stark contrast between the rapid passage of time and the characters' frozen anxiety, creating a suffocating, stressful domestic atmosphere.", + "seed": 15963, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_2/index_3703_seed_15963_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.063995311409235, + "vlm_logprob": -0.063995311409235, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.063995311409235, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up of a wall clock with spinning hands creating blurry afterimages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back to reveal a dimly lit children's study?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy approximately ten years old wearing a blue and white striped T-shirt visible hunched over a wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy shown tugging at his tousled black hair while his body trembles with frustration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a mother in a gray cardigan visible standing in the shadows holding a thin report card?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07312346249818802 + ], + "opposite_logprob_tokens": [ + -2.8231234550476074 + ], + "logprob_reward": -0.07312346249818802, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07312346249818802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4856627285480499, + 0.1070902943611145, + 0.05091994255781174, + 0.06461687386035919, + 0.05928570404648781, + 0.08033571392297745, + 0.04260696843266487, + 0.03377259150147438, + 0.020401403307914734, + 0.05530770123004913 + ], + "credit_map": [ + 0.03177965432405472, + 0.05855352431535721, + 0.05313296988606453, + 0.0613541305065155, + 0.08567574620246887, + 0.08666743338108063, + 0.05957683175802231, + 0.048922501504421234, + 0.009588266722857952, + 0.0077565754763782024, + 0.008867434225976467, + 0.006354155484586954, + 0.00882294587790966, + 0.030101556330919266, + 0.005371698644012213, + 0.030227670446038246, + 0.006788092665374279, + 0.005238347686827183, + 0.00403035432100296, + 0.004266017582267523, + 0.005952834151685238, + 0.008084815926849842, + 0.009500655345618725, + 0.0070588244125247, + 0.009636806324124336, + 0.0084086824208498, + 0.008604470640420914, + 0.0077591934241354465, + 0.008239730261266232, + 0.0073892343789339066, + 0.01005308423191309, + 0.004525674041360617, + 0.005765171721577644, + 0.010981759987771511, + 0.005842773709446192, + 0.0072717247530817986, + 0.009755355305969715, + 0.00819714367389679, + 0.005181982181966305, + 0.006289795972406864, + 0.007242027670145035, + 0.011609244160354137, + 0.008280600421130657, + 0.010878457687795162, + 0.011259539984166622, + 0.011089284904301167, + 0.010651842691004276, + 0.009324721060693264, + 0.005532394628971815, + 0.006906076800078154, + 0.004972249269485474, + 0.003889587242156267, + 0.0064924550242722034, + 0.00571805564686656, + 0.004158613737672567, + 0.004937541671097279, + 0.005204567685723305, + 0.004682695027440786, + 0.0024365184362977743, + 0.004345064051449299, + 0.005582805722951889, + 0.005162152461707592, + 0.0035549732856452465, + 0.002803817158564925, + 0.002493822481483221, + 0.0031357721891254187, + 0.00195641303434968, + 0.002066704910248518, + 0.0030378466472029686, + 0.0032357987947762012, + 0.0020961856935173273, + 0.002378862351179123, + 0.004517490975558758, + 0.015961216762661934, + 0.003754179459065199, + 0.0025767735205590725, + 0.00872514583170414, + 0.011717524379491806, + 0.0037077944725751877, + 0.004347580019384623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.07312346249818802 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up of a wall clock with spinning hands creating blurry afterimages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back to reveal a dimly lit children's study?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy approximately ten years old wearing a blue and white striped T-shirt visible hunched over a wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy shown tugging at his tousled black hair while his body trembles with frustration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a mother in a gray cardigan visible standing in the shadows holding a thin report card?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03376644477248192 + ], + "opposite_logprob_tokens": [ + -3.408766508102417 + ], + "logprob_reward": -0.03376644477248192, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03376644477248192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07118800282478333, + 0.13356664776802063, + 0.1443546861410141, + 0.10003301501274109, + 0.11921712011098862, + 0.16854169964790344, + 0.0927354022860527, + 0.0774221271276474, + 0.0518278107047081, + 0.04111352562904358 + ], + "credit_map": [ + 0.010307238437235355, + 0.010174761526286602, + 0.008268305100500584, + 0.010026260279119015, + 0.009834710508584976, + 0.00755705451592803, + 0.008474627509713173, + 0.0065450468100607395, + 0.01172050554305315, + 0.008995606563985348, + 0.010769878514111042, + 0.0076971277594566345, + 0.011408654972910881, + 0.03559369966387749, + 0.01016892772167921, + 0.03721224516630173, + 0.01160260196775198, + 0.012936152517795563, + 0.010661059059202671, + 0.011576118879020214, + 0.0205976665019989, + 0.031275298446416855, + 0.029747264459729195, + 0.01595851220190525, + 0.011797656305134296, + 0.012635366059839725, + 0.013609667308628559, + 0.01388294342905283, + 0.011104254983365536, + 0.013467027805745602, + 0.012971852906048298, + 0.010564243420958519, + 0.016478387638926506, + 0.01699959859251976, + 0.010179828852415085, + 0.015764739364385605, + 0.01594533398747444, + 0.015574521385133266, + 0.014019712805747986, + 0.014255001209676266, + 0.014893424697220325, + 0.02235843986272812, + 0.014839887619018555, + 0.021559525281190872, + 0.027806434780359268, + 0.026357151567935944, + 0.023347802460193634, + 0.017379028722643852, + 0.013027263805270195, + 0.015269226394593716, + 0.010734125040471554, + 0.009351447224617004, + 0.014656213112175465, + 0.011695083230733871, + 0.0086516123265028, + 0.0093504274263978, + 0.008840618655085564, + 0.016136329621076584, + 0.005121967755258083, + 0.005691297817975283, + 0.011204327456653118, + 0.012800357304513454, + 0.008679570630192757, + 0.008947659283876419, + 0.004729378502815962, + 0.009279762394726276, + 0.0037834388203918934, + 0.006282885558903217, + 0.010310902260243893, + 0.006824494805186987, + 0.004508431535214186, + 0.00610851775854826, + 0.004333756398409605, + 0.005823282990604639, + 0.003592547494918108, + 0.0026944370474666357, + 0.008199545554816723, + 0.007222780957818031, + 0.005412208382040262, + 0.003834965405985713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.03376644477248192 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up of a wall clock with spinning hands creating blurry afterimages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back to reveal a dimly lit children's study?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy approximately ten years old wearing a blue and white striped T-shirt visible hunched over a wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy shown tugging at his tousled black hair while his body trembles with frustration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a mother in a gray cardigan visible standing in the shadows holding a thin report card?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033733829855918884 + ], + "opposite_logprob_tokens": [ + -3.40873384475708 + ], + "logprob_reward": -0.033733829855918884, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033733829855918884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1470371037721634, + 0.03517424687743187, + 0.039299387484788895, + 0.059502337127923965, + 0.03541167825460434, + 0.2653953433036804, + 0.08453144878149033, + 0.23032400012016296, + 0.049074120819568634, + 0.054250337183475494 + ], + "credit_map": [ + 0.013289352878928185, + 0.013318764977157116, + 0.023227287456393242, + 0.019176889210939407, + 0.022489648312330246, + 0.023094477131962776, + 0.017332732677459717, + 0.015107967890799046, + 0.0051386174745857716, + 0.0038610505871474743, + 0.004399529658257961, + 0.0034243902191519737, + 0.003262612968683243, + 0.006939989048987627, + 0.002993641421198845, + 0.005154419224709272, + 0.004377967678010464, + 0.005643094424158335, + 0.003765597240999341, + 0.003172939410433173, + 0.005308130290359259, + 0.007040668278932571, + 0.005285046529024839, + 0.004705945961177349, + 0.012795843183994293, + 0.007703156676143408, + 0.0077699837274849415, + 0.007547205314040184, + 0.004694910254329443, + 0.00846356526017189, + 0.006104248110204935, + 0.004423429258167744, + 0.004610623233020306, + 0.003871746128425002, + 0.0032912546303123236, + 0.00562847638502717, + 0.005254244897514582, + 0.004362385254353285, + 0.0045629204250872135, + 0.0038300275336951017, + 0.02341279201209545, + 0.03239559382200241, + 0.02937454916536808, + 0.0538751445710659, + 0.03867353871464729, + 0.02429378405213356, + 0.034305546432733536, + 0.02906443364918232, + 0.012310781516134739, + 0.011984151788055897, + 0.010794921778142452, + 0.010788911022245884, + 0.010694161057472229, + 0.010285316035151482, + 0.006530088372528553, + 0.011143120937049389, + 0.029676631093025208, + 0.03907572478055954, + 0.018034107983112335, + 0.03397059440612793, + 0.032415375113487244, + 0.0323956236243248, + 0.02457147277891636, + 0.02018449455499649, + 0.006291494704782963, + 0.006431166082620621, + 0.005428266245871782, + 0.003523136954754591, + 0.007837380282580853, + 0.007534714415669441, + 0.004974395502358675, + 0.00705356989055872, + 0.004925562534481287, + 0.01239690650254488, + 0.003282784251496196, + 0.002650979906320572, + 0.010695096105337143, + 0.008900098502635956, + 0.006149211898446083, + 0.005249699577689171 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.033733829855918884 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up of a wall clock with spinning hands creating blurry afterimages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back to reveal a dimly lit children's study?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy approximately ten years old wearing a blue and white striped T-shirt visible hunched over a wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy shown tugging at his tousled black hair while his body trembles with frustration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a mother in a gray cardigan visible standing in the shadows holding a thin report card?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07903537899255753 + ], + "opposite_logprob_tokens": [ + -2.5790352821350098 + ], + "logprob_reward": -0.07903537899255753, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07903537899255753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14831404387950897, + 0.04317836835980415, + 0.08378924429416656, + 0.05679049342870712, + 0.038191065192222595, + 0.2055722177028656, + 0.08403114974498749, + 0.12282807379961014, + 0.09624549001455307, + 0.12105987966060638 + ], + "credit_map": [ + 0.015912992879748344, + 0.0115693099796772, + 0.030920052900910378, + 0.022729594260454178, + 0.018781008198857307, + 0.01814442314207554, + 0.015600528568029404, + 0.01465613953769207, + 0.007127103861421347, + 0.004235154017806053, + 0.00546807236969471, + 0.004338418133556843, + 0.006303871050477028, + 0.006257749628275633, + 0.004712801426649094, + 0.004735198803246021, + 0.0072145829908549786, + 0.008774305693805218, + 0.005722066853195429, + 0.0067184134386479855, + 0.01374926045536995, + 0.017382333055138588, + 0.011961220763623714, + 0.012267066165804863, + 0.012123764492571354, + 0.009788244031369686, + 0.008041024208068848, + 0.006354033015668392, + 0.0049035451374948025, + 0.006009949371218681, + 0.00551465293392539, + 0.004055282101035118, + 0.004927941597998142, + 0.006547571159899235, + 0.0032769758254289627, + 0.005458255298435688, + 0.005428877659142017, + 0.004933062940835953, + 0.0038108539301902056, + 0.0038075302727520466, + 0.021101107820868492, + 0.02591407671570778, + 0.015076740644872189, + 0.025113390758633614, + 0.03242801874876022, + 0.034691307693719864, + 0.025380996987223625, + 0.02586658112704754, + 0.011040710844099522, + 0.009759529493749142, + 0.010353103280067444, + 0.00795974861830473, + 0.00995269138365984, + 0.010632560588419437, + 0.008420787751674652, + 0.0159120112657547, + 0.013489646837115288, + 0.020041555166244507, + 0.009696842171251774, + 0.015095935203135014, + 0.019720491021871567, + 0.01894453912973404, + 0.012626297771930695, + 0.013212759047746658, + 0.007280611898750067, + 0.013723169453442097, + 0.005989915691316128, + 0.01199687086045742, + 0.023473141714930534, + 0.012662563472986221, + 0.012015028856694698, + 0.009104191325604916, + 0.01138340588659048, + 0.027453405782580376, + 0.006179837044328451, + 0.005434638820588589, + 0.026182973757386208, + 0.0201472919434309, + 0.008563913404941559, + 0.015714410692453384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.07903537899255753 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up of a wall clock with spinning hands creating blurry afterimages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back to reveal a dimly lit children's study?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy approximately ten years old wearing a blue and white striped T-shirt visible hunched over a wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy shown tugging at his tousled black hair while his body trembles with frustration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a mother in a gray cardigan visible standing in the shadows holding a thin report card?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031744092702866 + ], + "opposite_logprob_tokens": [ + -2.3503174781799316 + ], + "logprob_reward": -0.10031744092702866, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031744092702866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08032415062189102, + 0.044261932373046875, + 0.08686347305774689, + 0.05850183591246605, + 0.039357028901576996, + 0.07164592295885086, + 0.21121740341186523, + 0.13290835916996002, + 0.20481060445308685, + 0.0701092854142189 + ], + "credit_map": [ + 0.009876775555312634, + 0.006821073591709137, + 0.01568063162267208, + 0.013338468037545681, + 0.009179789572954178, + 0.009618840180337429, + 0.008053966797888279, + 0.007754601072520018, + 0.006164809688925743, + 0.006407920736819506, + 0.0051016309298574924, + 0.005358328577131033, + 0.0047833966091275215, + 0.007655668538063765, + 0.0031506195664405823, + 0.005639558658003807, + 0.009766608476638794, + 0.011116724461317062, + 0.01092528086155653, + 0.011340538039803505, + 0.012272217310965061, + 0.01150131318718195, + 0.011952845379710197, + 0.007987947203218937, + 0.010757176205515862, + 0.009147840552031994, + 0.00731614138931036, + 0.006767259910702705, + 0.0071649132296442986, + 0.006178915034979582, + 0.007207883056253195, + 0.0039617083966732025, + 0.007089735008776188, + 0.005783736705780029, + 0.004403751343488693, + 0.005404962692409754, + 0.005964674986898899, + 0.0037183493841439486, + 0.0036528923083096743, + 0.003338924143463373, + 0.007481131702661514, + 0.009846167638897896, + 0.007946391589939594, + 0.009390005841851234, + 0.009881778620183468, + 0.008510930463671684, + 0.010243111290037632, + 0.008346406742930412, + 0.031050244346261024, + 0.0275385994464159, + 0.01979673095047474, + 0.01886039972305298, + 0.02478688955307007, + 0.03480127453804016, + 0.019984226673841476, + 0.03439903259277344, + 0.013525020331144333, + 0.028121711686253548, + 0.011379102244973183, + 0.017413856461644173, + 0.016262248158454895, + 0.018831636756658554, + 0.013940204866230488, + 0.013434589840471745, + 0.015440738759934902, + 0.028376169502735138, + 0.013001575134694576, + 0.015762658789753914, + 0.05485985055565834, + 0.027729403227567673, + 0.025460682809352875, + 0.024179518222808838, + 0.004533175379037857, + 0.016435297206044197, + 0.0048876297660171986, + 0.0038034876342862844, + 0.012339198961853981, + 0.015657352283596992, + 0.007785568479448557, + 0.00466757221147418 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.10031744092702866 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a close-up of a wall clock, its hour and minute hands spinning at a visible speed, creating a series of blurry afterimages. The camera then smoothly pulls back to reveal a dimly lit children's study. At the center of the frame, a boy, approximately ten years old, wearing a blue and white striped T-shirt, hunched over a wooden desk piled high with textbooks and workbooks. He desperately tugs his hands into his tousled black hair, his fingers tugging at the roots, his body trembling with silent frustration. An old-fashioned desk lamp provides the sole source of light, its warm yellow glow illuminating the mess on the desk while casting long, deep shadows across the rest of the room. Behind him, a mother in a gray cardigan stands quietly at the edge of the shadows, her brow furrowed, her lips pressed into a serious line, her gaze darting back and forth intently at the thin report card in her hand. The scene creates a stark contrast between the rapid passage of time and the characters' frozen anxiety, creating a suffocating, stressful domestic atmosphere.", + "seed": 15964, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_2/index_3703_seed_15964_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05331600196659565, + "vlm_logprob": -0.05331600196659565, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.05331600196659565, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up of a wall clock with spinning hands creating blurry afterimages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back to reveal a dimly lit children's study?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy approximately ten years old wearing a blue and white striped T-shirt visible hunched over a wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy shown tugging at his tousled black hair while his body trembles with frustration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a mother in a gray cardigan visible standing in the shadows holding a thin report card?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017327826470136642 + ], + "opposite_logprob_tokens": [ + -4.767327785491943 + ], + "logprob_reward": -0.017327826470136642, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.017327826470136642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2012680619955063, + 0.28803157806396484, + 0.19337902963161469, + 0.05998537316918373, + 0.027993813157081604, + 0.02327836863696575, + 0.03443841636180878, + 0.05050788074731827, + 0.05747775360941887, + 0.06363976001739502 + ], + "credit_map": [ + 0.024683089926838875, + 0.01913316920399666, + 0.0329168327152729, + 0.027973340824246407, + 0.01915949396789074, + 0.02793216146528721, + 0.028790459036827087, + 0.020679498091340065, + 0.01547516230493784, + 0.01048759650439024, + 0.018151719123125076, + 0.01185598038136959, + 0.009819014929234982, + 0.11029726266860962, + 0.013723178766667843, + 0.09822163730859756, + 0.007259568199515343, + 0.009210115298628807, + 0.01511297281831503, + 0.008067813701927662, + 0.007146754767745733, + 0.0708712711930275, + 0.0062120468355715275, + 0.0694984719157219, + 0.0076771508902311325, + 0.01074204221367836, + 0.005154477432370186, + 0.0037721246480941772, + 0.012985431589186192, + 0.009268330410122871, + 0.004182885400950909, + 0.0062029240652918816, + 0.0029266502242535353, + 0.005068748723715544, + 0.003535299329087138, + 0.0030035818926990032, + 0.002539806766435504, + 0.004657880403101444, + 0.002905054483562708, + 0.00335679086856544, + 0.0023973537608981133, + 0.0029021187219768763, + 0.0030207980889827013, + 0.0024938422720879316, + 0.0017306925728917122, + 0.004740199539810419, + 0.004226702265441418, + 0.0017666583880782127, + 0.005559550132602453, + 0.0069817896001040936, + 0.006037733517587185, + 0.0035049261059612036, + 0.001913827029056847, + 0.003614033106714487, + 0.0038244230672717094, + 0.0030021313577890396, + 0.008216295391321182, + 0.013223065994679928, + 0.005844080355018377, + 0.005216765217483044, + 0.0045096189714968204, + 0.0056402781046926975, + 0.0037744208239018917, + 0.004083349369466305, + 0.010163256898522377, + 0.014827288687229156, + 0.006243299692869186, + 0.004666235763579607, + 0.005370155442506075, + 0.006843256764113903, + 0.0034362913575023413, + 0.005927966441959143, + 0.007997579872608185, + 0.013608009554445744, + 0.006285055074840784, + 0.004534432198852301, + 0.012297358363866806, + 0.007302115671336651, + 0.006935871206223965, + 0.00467932503670454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.017327826470136642 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up of a wall clock with spinning hands creating blurry afterimages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back to reveal a dimly lit children's study?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy approximately ten years old wearing a blue and white striped T-shirt visible hunched over a wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy shown tugging at his tousled black hair while his body trembles with frustration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a mother in a gray cardigan visible standing in the shadows holding a thin report card?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008668649941682816 + ], + "opposite_logprob_tokens": [ + -4.758668422698975 + ], + "logprob_reward": -0.008668649941682816, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008668649941682816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11654053628444672, + 0.10759224742650986, + 0.5276063680648804, + 0.09220509976148605, + 0.026784930378198624, + 0.029933765530586243, + 0.015477280132472515, + 0.02568342350423336, + 0.03297925740480423, + 0.025197025388479233 + ], + "credit_map": [ + 0.013421834446489811, + 0.011782757937908173, + 0.011644489131867886, + 0.017038071528077126, + 0.010122117586433887, + 0.018452219665050507, + 0.020087385550141335, + 0.013991676270961761, + 0.007611863315105438, + 0.007262838538736105, + 0.010646275244653225, + 0.008063159883022308, + 0.008140130899846554, + 0.0309615321457386, + 0.00931108370423317, + 0.025595376268029213, + 0.017293918877840042, + 0.020505255088210106, + 0.027995338663458824, + 0.01533054281026125, + 0.03340119868516922, + 0.19703003764152527, + 0.02060001716017723, + 0.1954500824213028, + 0.008670157752931118, + 0.015211939811706543, + 0.0070442152209579945, + 0.005913027562201023, + 0.023318709805607796, + 0.014371211640536785, + 0.006286020390689373, + 0.011389827355742455, + 0.0026397169567644596, + 0.0056856609880924225, + 0.0030685195233672857, + 0.0028348241467028856, + 0.0019787438213825226, + 0.003537439974024892, + 0.003713476238772273, + 0.0033265529200434685, + 0.003695901483297348, + 0.0031382094603031874, + 0.00347668188624084, + 0.0038825436495244503, + 0.0030593378469347954, + 0.0036215430591255426, + 0.006818099413067102, + 0.002241453854367137, + 0.0021219453774392605, + 0.002432272769510746, + 0.0019799317233264446, + 0.0015281253727152944, + 0.00130566512234509, + 0.0016350812511518598, + 0.001618054579012096, + 0.002856205217540264, + 0.003684616880491376, + 0.0031685729045420885, + 0.004753204062581062, + 0.0018267774721607566, + 0.0031894848216325045, + 0.003468639450147748, + 0.001875434652902186, + 0.003716696985065937, + 0.005343767814338207, + 0.007125857751816511, + 0.0034145587123930454, + 0.002995393704622984, + 0.003265778534114361, + 0.003735155100002885, + 0.003292187349870801, + 0.003806565422564745, + 0.0028639142401516438, + 0.004789281636476517, + 0.0034733645152300596, + 0.002572518540546298, + 0.004335589706897736, + 0.0025972523726522923, + 0.002199422800913453, + 0.002365685300901532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.008668649941682816 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up of a wall clock with spinning hands creating blurry afterimages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back to reveal a dimly lit children's study?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy approximately ten years old wearing a blue and white striped T-shirt visible hunched over a wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy shown tugging at his tousled black hair while his body trembles with frustration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a mother in a gray cardigan visible standing in the shadows holding a thin report card?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018195031210780144 + ], + "opposite_logprob_tokens": [ + -4.018195152282715 + ], + "logprob_reward": -0.018195031210780144, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018195031210780144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17887331545352936, + 0.14353536069393158, + 0.23941642045974731, + 0.03487041965126991, + 0.06178225949406624, + 0.08100815117359161, + 0.02195735089480877, + 0.12863598763942719, + 0.035628996789455414, + 0.07429175078868866 + ], + "credit_map": [ + 0.02459818683564663, + 0.014440583996474743, + 0.03774731233716011, + 0.03503557667136192, + 0.01237367745488882, + 0.017734717577695847, + 0.02202446013689041, + 0.014918798580765724, + 0.005571325309574604, + 0.006218415684998035, + 0.006139818113297224, + 0.005999499931931496, + 0.007113736122846603, + 0.053255464881658554, + 0.008825942873954773, + 0.050411157310009, + 0.009506968781352043, + 0.007717778440564871, + 0.020113619044423103, + 0.009871884249150753, + 0.008291158825159073, + 0.08789082616567612, + 0.008935286663472652, + 0.08708888292312622, + 0.003879006952047348, + 0.007206284906715155, + 0.003077079076319933, + 0.004191359505057335, + 0.0057161725126206875, + 0.00444731255993247, + 0.002504476811736822, + 0.003848727559670806, + 0.0068595982156693935, + 0.010023972950875759, + 0.006461989600211382, + 0.009381736628711224, + 0.005920643452554941, + 0.008130630478262901, + 0.008258799090981483, + 0.006744890008121729, + 0.010930764488875866, + 0.013682185672223568, + 0.01169291976839304, + 0.008649979718029499, + 0.008117184974253178, + 0.009932315908372402, + 0.010907269082963467, + 0.007095537148416042, + 0.002729114843532443, + 0.0049927495419979095, + 0.0028154272586107254, + 0.0018767979927361012, + 0.0019470422994345427, + 0.002187883947044611, + 0.0025020733010023832, + 0.002906262641772628, + 0.01691124401986599, + 0.026076098904013634, + 0.013726468198001385, + 0.010305837728083134, + 0.02459261380136013, + 0.01606949418783188, + 0.012456351891160011, + 0.008497872389853, + 0.005067004822194576, + 0.01101456768810749, + 0.003988162614405155, + 0.0029235759284347296, + 0.00396035285666585, + 0.003972187638282776, + 0.002405948005616665, + 0.0022971979342401028, + 0.006780555006116629, + 0.013846692629158497, + 0.006612632889300585, + 0.005360960029065609, + 0.016850147396326065, + 0.00890444777905941, + 0.010037826374173164, + 0.005898497998714447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.018195031210780144 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up of a wall clock with spinning hands creating blurry afterimages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back to reveal a dimly lit children's study?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy approximately ten years old wearing a blue and white striped T-shirt visible hunched over a wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy shown tugging at his tousled black hair while his body trembles with frustration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a mother in a gray cardigan visible standing in the shadows holding a thin report card?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205623969435692 + ], + "opposite_logprob_tokens": [ + -2.812056303024292 + ], + "logprob_reward": -0.06205623969435692, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205623969435692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10445128381252289, + 0.11831904947757721, + 0.5336690545082092, + 0.05085020512342453, + 0.03401048481464386, + 0.030348848551511765, + 0.032033778727054596, + 0.048400767147541046, + 0.02223818376660347, + 0.025678325444459915 + ], + "credit_map": [ + 0.014670898206532001, + 0.007994339801371098, + 0.023520387709140778, + 0.015659775584936142, + 0.008271170780062675, + 0.012012070044875145, + 0.013279801234602928, + 0.009042836725711823, + 0.006270991638302803, + 0.0040984139777719975, + 0.005108638666570187, + 0.00406577717512846, + 0.004183404613286257, + 0.04687030240893364, + 0.005754565820097923, + 0.041966959834098816, + 0.010017536580562592, + 0.011678855866193771, + 0.02394718863070011, + 0.007797224447131157, + 0.008653335273265839, + 0.2301284819841385, + 0.012163904495537281, + 0.22928248345851898, + 0.008822955191135406, + 0.012080388143658638, + 0.004866601433604956, + 0.0031783345621079206, + 0.008480587974190712, + 0.0064130681566894054, + 0.002330909948796034, + 0.004677357617765665, + 0.0032442871015518904, + 0.006760905031114817, + 0.00325938337482512, + 0.004792397376149893, + 0.003222040832042694, + 0.0037784127052873373, + 0.0047761076129972935, + 0.0041769505478441715, + 0.0035559104289859533, + 0.004166681319475174, + 0.004825159907341003, + 0.0034306661691516638, + 0.003018667222931981, + 0.004165582824498415, + 0.005052210297435522, + 0.0021339678205549717, + 0.005018474999815226, + 0.0066873012110590935, + 0.005730042699724436, + 0.0029598840046674013, + 0.0021172580309212208, + 0.0031927861273288727, + 0.0033339294604957104, + 0.0029941045213490725, + 0.007265438325703144, + 0.006703691557049751, + 0.00706613389775157, + 0.00453357957303524, + 0.0062866718508303165, + 0.006322478875517845, + 0.00496448390185833, + 0.005258288234472275, + 0.0037245643325150013, + 0.005654074717313051, + 0.0022879913449287415, + 0.00159572868142277, + 0.00235809781588614, + 0.0028372074011713266, + 0.001690115430392325, + 0.002090404275804758, + 0.003156110877171159, + 0.005503408145159483, + 0.003540815319865942, + 0.0025066384114325047, + 0.0036688067484647036, + 0.0031096641905605793, + 0.002038555918261409, + 0.0021543256007134914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.06205623969435692 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up of a wall clock with spinning hands creating blurry afterimages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back to reveal a dimly lit children's study?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy approximately ten years old wearing a blue and white striped T-shirt visible hunched over a wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy shown tugging at his tousled black hair while his body trembles with frustration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a mother in a gray cardigan visible standing in the shadows holding a thin report card?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16033226251602173 + ], + "opposite_logprob_tokens": [ + -1.910332202911377 + ], + "logprob_reward": -0.16033226251602173, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16033226251602173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08014021813869476, + 0.05102561041712761, + 0.42813748121261597, + 0.06779439747333527, + 0.029200492426753044, + 0.033728040754795074, + 0.04029180854558945, + 0.07548866420984268, + 0.11723926663398743, + 0.07695407420396805 + ], + "credit_map": [ + 0.011801986023783684, + 0.0066862162202596664, + 0.008897177875041962, + 0.009644913487136364, + 0.007399295456707478, + 0.01319647766649723, + 0.013737987726926804, + 0.008776155300438404, + 0.003398413071408868, + 0.002516848733648658, + 0.004354660399258137, + 0.004217472393065691, + 0.003083562944084406, + 0.01559034176170826, + 0.006372319534420967, + 0.0114919887855649, + 0.007998333312571049, + 0.009775621816515923, + 0.018617616966366768, + 0.005183403380215168, + 0.008205139078199863, + 0.1845327615737915, + 0.009446055628359318, + 0.184378519654274, + 0.005944174248725176, + 0.019613943994045258, + 0.003451369237154722, + 0.0037999313790351152, + 0.01615666225552559, + 0.010465145111083984, + 0.001990963239222765, + 0.006372201256453991, + 0.0035622708965092897, + 0.004034753423184156, + 0.003420796012505889, + 0.004028817173093557, + 0.0025311452336609364, + 0.004996666684746742, + 0.0034425552003085613, + 0.003183485707268119, + 0.0037262453697621822, + 0.004704438149929047, + 0.0044638472609221935, + 0.004250444937497377, + 0.0028100477065891027, + 0.005642636679112911, + 0.00517914118245244, + 0.0029512345790863037, + 0.006456848233938217, + 0.011668909341096878, + 0.006835783366113901, + 0.0034356890246272087, + 0.0023382287472486496, + 0.0036861631087958813, + 0.002279793843626976, + 0.003590388223528862, + 0.010470747947692871, + 0.020235111936926842, + 0.008382480591535568, + 0.005546572618186474, + 0.010359887965023518, + 0.008151257410645485, + 0.005394641775637865, + 0.006947961635887623, + 0.019237633794546127, + 0.02963903173804283, + 0.013959317468106747, + 0.009836089797317982, + 0.012465252541005611, + 0.012983976863324642, + 0.00816639419645071, + 0.010951557196676731, + 0.00948952604085207, + 0.02231835573911667, + 0.008739897049963474, + 0.005291204899549484, + 0.012100670486688614, + 0.008414637297391891, + 0.006435814779251814, + 0.004163962323218584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.16033226251602173 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_3/index_tensor([1459, 1459])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_3/index_tensor([1459, 1459])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..26b39ccdd5f6e102a2eeaa4969d69b80dc354903 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_3/index_tensor([1459, 1459])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, wide-angle shot follows the side of a sailboat, moving forward in a smooth tracking motion, evoking an immersive sense of sailing. A classic sloop, with a sleek white fiberglass hull and a deep blue antifouling paint on the lower half, plows through the deep azure waters. The massive mainsail and foresail, fully filled by the strong wind, create a full, powerful arc, the visible texture and stitching details of the sails striking in the sunlight. The boat rises and falls rhythmically with the waves, and the deep blue waves crash powerfully against the side, stirring up layers of white spray and a fine mist. Some of the water even splashes onto the camera, momentarily blurring the image before quickly sliding back down. The bright afternoon sun casts countless sparkling diamond-like points of light on the shimmering sea, illuminating the boat's wet teak deck and gleaming metal railings, creating a visual effect of ultra-high-definition realism. The horizon is clearly visible in the distance, dividing the sky, dotted with a few fluffy white clouds, from the vast ocean. The entire picture is full of a magnificent sense of adventure and freedom, while also showing an awe-inspiring force of nature.", + "seed": 28582, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_3/index_1459_seed_28582_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08392277434468269, + "vlm_logprob": -0.08392277434468269, + "vlm_yes_no_margin": 3.199999988079071, + "vlm_token_credit": -0.08392277434468269, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic sloop sailboat with a white fiberglass hull and deep blue antifouling paint on the lower half?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the massive mainsail and foresail fully filled with wind, showing visible texture and stitching details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the side of the boat in a smooth tracking motion from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep azure ocean with white spray and mist crashing against the boat's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show water splashing onto the camera lens, momentarily blurring the image before clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020178981125354767 + ], + "opposite_logprob_tokens": [ + -4.14517879486084 + ], + "logprob_reward": -0.020178981125354767, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.020178981125354767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11616096645593643, + 0.036247022449970245, + 0.052834998816251755, + 0.032971058040857315, + 0.026813719421625137, + 0.02513572759926319, + 0.26631855964660645, + 0.06321084499359131, + 0.3024140000343323, + 0.07789307832717896 + ], + "credit_map": [ + 0.012216885574162006, + 0.01167063694447279, + 0.018974367529153824, + 0.016640042886137962, + 0.017868736758828163, + 0.014599746093153954, + 0.011979465372860432, + 0.012211091816425323, + 0.0035535357892513275, + 0.004885405767709017, + 0.004796018358319998, + 0.007337323389947414, + 0.005448770243674517, + 0.004757164511829615, + 0.002899722894653678, + 0.0025690828915685415, + 0.005142009351402521, + 0.006296535953879356, + 0.00504181208088994, + 0.010253742337226868, + 0.010397817008197308, + 0.006288169417530298, + 0.004920069128274918, + 0.004494850989431143, + 0.0038044911343604326, + 0.004341143649071455, + 0.0025562478695064783, + 0.005680664908140898, + 0.005722492467612028, + 0.004269023425877094, + 0.0033099299762398005, + 0.0032870660070329905, + 0.003063230775296688, + 0.004205514211207628, + 0.00232210592366755, + 0.002596296137198806, + 0.004247622098773718, + 0.0037813216913491488, + 0.0030285080429166555, + 0.0035691228695213795, + 0.004528414458036423, + 0.0026756427250802517, + 0.0028686688747256994, + 0.0024586643557995558, + 0.002188662067055702, + 0.004844678565859795, + 0.002470701467245817, + 0.0031002964824438095, + 0.0851249024271965, + 0.022564833983778954, + 0.008200624026358128, + 0.0064637986943125725, + 0.08877332508563995, + 0.037824131548404694, + 0.010707549750804901, + 0.006659420672804117, + 0.013105064630508423, + 0.005740860477089882, + 0.0050805117934942245, + 0.005341492593288422, + 0.012654717080295086, + 0.009393587708473206, + 0.00531342625617981, + 0.0065811919048428535, + 0.09717532247304916, + 0.024157805368304253, + 0.015731874853372574, + 0.013776781037449837, + 0.08908256888389587, + 0.03422565758228302, + 0.014188752509653568, + 0.014075277373194695, + 0.02488388679921627, + 0.00666567450389266, + 0.0051794592291116714, + 0.00307856616564095, + 0.020894581452012062, + 0.00900330301374197, + 0.004222217481583357, + 0.003965390846133232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.020178981125354767 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic sloop sailboat with a white fiberglass hull and deep blue antifouling paint on the lower half?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the massive mainsail and foresail fully filled with wind, showing visible texture and stitching details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the side of the boat in a smooth tracking motion from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep azure ocean with white spray and mist crashing against the boat's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show water splashing onto the camera lens, momentarily blurring the image before clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812874108552933 + ], + "opposite_logprob_tokens": [ + -3.2881288528442383 + ], + "logprob_reward": -0.03812874108552933, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812874108552933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09842769801616669, + 0.041242413222789764, + 0.04883201792836189, + 0.016481883823871613, + 0.014653376303613186, + 0.13244713842868805, + 0.08480188250541687, + 0.05710592865943909, + 0.3496989607810974, + 0.15630872547626495 + ], + "credit_map": [ + 0.01086785551160574, + 0.010287551209330559, + 0.014850229024887085, + 0.012376559898257256, + 0.015606596134603024, + 0.012883053161203861, + 0.010726072825491428, + 0.01082978118211031, + 0.003560979850590229, + 0.004964872729033232, + 0.004582373425364494, + 0.01055449154227972, + 0.005763212218880653, + 0.004868697840720415, + 0.003679516026750207, + 0.00326826935634017, + 0.0039159078150987625, + 0.005726478528231382, + 0.00496691232547164, + 0.008951984345912933, + 0.00973508507013321, + 0.00604195287451148, + 0.005381507333368063, + 0.004112193360924721, + 0.00227552093565464, + 0.0019193313783034682, + 0.002156459493562579, + 0.00230291741900146, + 0.0022321583237499, + 0.0017651250818744302, + 0.001885981415398419, + 0.0019443908240646124, + 0.0014615259133279324, + 0.0017315931618213654, + 0.0026862206868827343, + 0.0015736962668597698, + 0.0020743543282151222, + 0.001973007572814822, + 0.0013568807626143098, + 0.001796097494661808, + 0.037744052708148956, + 0.012271366082131863, + 0.009020316414535046, + 0.009371891617774963, + 0.014183803461492062, + 0.03241591900587082, + 0.009124846197664738, + 0.008314944803714752, + 0.024659881368279457, + 0.007252516224980354, + 0.0035600210539996624, + 0.0041544134728610516, + 0.0246135164052248, + 0.012537880800664425, + 0.0037660903763026, + 0.00425756024196744, + 0.017011981457471848, + 0.005410292185842991, + 0.0032380851916968822, + 0.002840493805706501, + 0.015441467985510826, + 0.006434425245970488, + 0.0032348823733627796, + 0.0034942973870784044, + 0.10232029855251312, + 0.029965659603476524, + 0.024048207327723503, + 0.020724065601825714, + 0.09370651841163635, + 0.0404345728456974, + 0.018852876499295235, + 0.019646774977445602, + 0.051080524921417236, + 0.013629114255309105, + 0.008991682901978493, + 0.0050725447945296764, + 0.047879479825496674, + 0.017442572861909866, + 0.0070873587392270565, + 0.005125441588461399 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.03812874108552933 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic sloop sailboat with a white fiberglass hull and deep blue antifouling paint on the lower half?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the massive mainsail and foresail fully filled with wind, showing visible texture and stitching details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the side of the boat in a smooth tracking motion from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep azure ocean with white spray and mist crashing against the boat's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show water splashing onto the camera lens, momentarily blurring the image before clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033695101737976074 + ], + "opposite_logprob_tokens": [ + -3.4086952209472656 + ], + "logprob_reward": -0.033695101737976074, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033695101737976074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06481041759252548, + 0.019562745466828346, + 0.016409190371632576, + 0.012389116920530796, + 0.025336310267448425, + 0.054699674248695374, + 0.18261674046516418, + 0.15711301565170288, + 0.45064035058021545, + 0.016422446817159653 + ], + "credit_map": [ + 0.008442326448857784, + 0.007014335133135319, + 0.013052382506430149, + 0.014064504764974117, + 0.007893702946603298, + 0.004865326453000307, + 0.004622541833668947, + 0.004855296574532986, + 0.0018989542732015252, + 0.0021864459849894047, + 0.0022695008665323257, + 0.004704735241830349, + 0.0030185107607394457, + 0.0020923104602843523, + 0.0019108421402052045, + 0.0014814439928159118, + 0.001685819705016911, + 0.002203107811510563, + 0.0018508653156459332, + 0.0026748734526336193, + 0.0028516948223114014, + 0.0019563562236726284, + 0.0016625047428533435, + 0.001523967133834958, + 0.0011369789717718959, + 0.0013947883853688836, + 0.0019602742977440357, + 0.0022726883180439472, + 0.0015221242792904377, + 0.0014550695195794106, + 0.001230014255270362, + 0.001417179242707789, + 0.002913371194154024, + 0.003376665525138378, + 0.0027199971955269575, + 0.002454105531796813, + 0.003925576340407133, + 0.003573744324967265, + 0.002614508382976055, + 0.00375834247097373, + 0.01038968563079834, + 0.005564927123486996, + 0.0066245854832232, + 0.003509672125801444, + 0.005844534374773502, + 0.013538621366024017, + 0.003578425385057926, + 0.005649226251989603, + 0.055468056350946426, + 0.015605784952640533, + 0.006405026651918888, + 0.006972269620746374, + 0.054884664714336395, + 0.026581576094031334, + 0.008084685541689396, + 0.008614675141870975, + 0.05018356069922447, + 0.013931121677160263, + 0.006524760741740465, + 0.0060134828090667725, + 0.0475405678153038, + 0.017353612929582596, + 0.007451761048287153, + 0.008114141412079334, + 0.13589128851890564, + 0.03809119760990143, + 0.032557256519794464, + 0.025340726599097252, + 0.12465079873800278, + 0.051461126655340195, + 0.020117612555623055, + 0.02253035083413124, + 0.0039060998242348433, + 0.0018249849090352654, + 0.001677559339441359, + 0.0014310914557427168, + 0.003193368436768651, + 0.001992961624637246, + 0.0010946746915578842, + 0.001301706419326365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.033695101737976074 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic sloop sailboat with a white fiberglass hull and deep blue antifouling paint on the lower half?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the massive mainsail and foresail fully filled with wind, showing visible texture and stitching details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the side of the boat in a smooth tracking motion from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep azure ocean with white spray and mist crashing against the boat's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show water splashing onto the camera lens, momentarily blurring the image before clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192361384630203 + ], + "opposite_logprob_tokens": [ + -4.264192581176758 + ], + "logprob_reward": -0.014192361384630203, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192361384630203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19605474174022675, + 0.2072104662656784, + 0.04587189108133316, + 0.04341574385762215, + 0.04321630299091339, + 0.05830909684300423, + 0.04963777959346771, + 0.09240130335092545, + 0.1681065708398819, + 0.09577606618404388 + ], + "credit_map": [ + 0.02426724135875702, + 0.022070754319429398, + 0.03021051175892353, + 0.033635515719652176, + 0.024758189916610718, + 0.021487118676304817, + 0.019085785374045372, + 0.02053963392972946, + 0.0169915072619915, + 0.02308368869125843, + 0.03763284534215927, + 0.05846641585230827, + 0.017518343403935432, + 0.01983872801065445, + 0.018217353150248528, + 0.015461585484445095, + 0.004142319317907095, + 0.0054311202839016914, + 0.00495197344571352, + 0.00983438827097416, + 0.008098132908344269, + 0.005289824679493904, + 0.0042692371644079685, + 0.003854897804558277, + 0.003953664097934961, + 0.005161725915968418, + 0.003984822891652584, + 0.008652505464851856, + 0.00663342559710145, + 0.005182493012398481, + 0.004921689163893461, + 0.004925417248159647, + 0.005038522183895111, + 0.007331459317356348, + 0.009067949838936329, + 0.004919609520584345, + 0.004586505237966776, + 0.005183799657970667, + 0.0037366885226219893, + 0.003351770807057619, + 0.00820267852395773, + 0.006274649873375893, + 0.007387056015431881, + 0.0049119023606181145, + 0.007709656376391649, + 0.011975475586950779, + 0.004549971781671047, + 0.007297703996300697, + 0.011250467039644718, + 0.004420877899974585, + 0.0039062073919922113, + 0.0038084599655121565, + 0.011021632701158524, + 0.008247706107795238, + 0.003374797524884343, + 0.0036076286341995, + 0.02884063869714737, + 0.007544511463493109, + 0.007279221899807453, + 0.003510781330987811, + 0.02689165621995926, + 0.009714924730360508, + 0.003943697549402714, + 0.004675865639001131, + 0.03511343523859978, + 0.01979830488562584, + 0.017005838453769684, + 0.01908244378864765, + 0.02951524592936039, + 0.021343473345041275, + 0.012197865173220634, + 0.014049971476197243, + 0.03148423507809639, + 0.008487062528729439, + 0.006081454921513796, + 0.004080884158611298, + 0.028112011030316353, + 0.010159728117287159, + 0.00463392399251461, + 0.0027367672882974148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.014192361384630203 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic sloop sailboat with a white fiberglass hull and deep blue antifouling paint on the lower half?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the massive mainsail and foresail fully filled with wind, showing visible texture and stitching details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the side of the boat in a smooth tracking motion from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep azure ocean with white spray and mist crashing against the boat's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show water splashing onto the camera lens, momentarily blurring the image before clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134186863899231 + ], + "opposite_logprob_tokens": [ + -1.3134186267852783 + ], + "logprob_reward": -0.3134186863899231, + "yes_no_margin_reward": 0.9999999403953552, + "token_credit_reward": -0.3134186863899231, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13271595537662506, + 0.10643085837364197, + 0.03939763829112053, + 0.017452271655201912, + 0.018685191869735718, + 0.04672930762171745, + 0.035513266921043396, + 0.12247779965400696, + 0.43184420466423035, + 0.04875344783067703 + ], + "credit_map": [ + 0.014688801020383835, + 0.014057625085115433, + 0.01651587523519993, + 0.017409713938832283, + 0.02340610884130001, + 0.018153594806790352, + 0.013600843958556652, + 0.014883396215736866, + 0.00947825238108635, + 0.011975960806012154, + 0.013856260105967522, + 0.025720587000250816, + 0.015857933089137077, + 0.010909614153206348, + 0.009954874403774738, + 0.008677381090819836, + 0.003890181425958872, + 0.004595290403813124, + 0.004341694992035627, + 0.008636553771793842, + 0.007684433367103338, + 0.004181452561169863, + 0.0030936768744140863, + 0.002974354662001133, + 0.0016193235060200095, + 0.0021579188760370016, + 0.0016398304142057896, + 0.003163181710988283, + 0.0032031964510679245, + 0.0018622552743181586, + 0.0019892689306288958, + 0.0018172963755205274, + 0.002061274601146579, + 0.0024037023540586233, + 0.00236597191542387, + 0.002176842885091901, + 0.002452680142596364, + 0.0025500759948045015, + 0.0018482899758964777, + 0.0028263533022254705, + 0.008233233354985714, + 0.00603550486266613, + 0.004419397562742233, + 0.00279525900259614, + 0.005998022388666868, + 0.011731578968465328, + 0.0038038217462599277, + 0.003712490200996399, + 0.00603064801543951, + 0.003814341500401497, + 0.003605286590754986, + 0.0033061217982321978, + 0.004987773951143026, + 0.008089717477560043, + 0.0029129101894795895, + 0.0027664704248309135, + 0.041093967854976654, + 0.011251583695411682, + 0.004669204354286194, + 0.004114015493541956, + 0.03773396462202072, + 0.013849605806171894, + 0.0050306785851716995, + 0.0047347769141197205, + 0.07141883671283722, + 0.06497733294963837, + 0.039502114057540894, + 0.05532792955636978, + 0.06267371773719788, + 0.07849905639886856, + 0.027453463524580002, + 0.03199175000190735, + 0.01648595929145813, + 0.0036803691182285547, + 0.0027980434242635965, + 0.001774566131643951, + 0.015075525268912315, + 0.005242303945124149, + 0.002046304289251566, + 0.0016503764782100916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.3134186863899231 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide-angle shot follows the side of a sailboat, moving forward in a smooth tracking motion, evoking an immersive sense of sailing. A classic sloop, with a sleek white fiberglass hull and a deep blue antifouling paint on the lower half, plows through the deep azure waters. The massive mainsail and foresail, fully filled by the strong wind, create a full, powerful arc, the visible texture and stitching details of the sails striking in the sunlight. The boat rises and falls rhythmically with the waves, and the deep blue waves crash powerfully against the side, stirring up layers of white spray and a fine mist. Some of the water even splashes onto the camera, momentarily blurring the image before quickly sliding back down. The bright afternoon sun casts countless sparkling diamond-like points of light on the shimmering sea, illuminating the boat's wet teak deck and gleaming metal railings, creating a visual effect of ultra-high-definition realism. The horizon is clearly visible in the distance, dividing the sky, dotted with a few fluffy white clouds, from the vast ocean. The entire picture is full of a magnificent sense of adventure and freedom, while also showing an awe-inspiring force of nature.", + "seed": 28583, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_3/index_1459_seed_28583_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03831769358366728, + "vlm_logprob": -0.03831769358366728, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.03831769358366728, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic sloop sailboat with a white fiberglass hull and deep blue antifouling paint on the lower half?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the massive mainsail and foresail fully filled with wind, showing visible texture and stitching details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the side of the boat in a smooth tracking motion from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep azure ocean with white spray and mist crashing against the boat's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show water splashing onto the camera lens, momentarily blurring the image before clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018907787278294563 + ], + "opposite_logprob_tokens": [ + -4.1439080238342285 + ], + "logprob_reward": -0.018907787278294563, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.018907787278294563, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1603003591299057, + 0.0436072051525116, + 0.036005087196826935, + 0.0672546923160553, + 0.03547750040888786, + 0.052201855927705765, + 0.29071351885795593, + 0.0705619528889656, + 0.2065063714981079, + 0.0373714342713356 + ], + "credit_map": [ + 0.024448039010167122, + 0.0322234146296978, + 0.015812788158655167, + 0.020367510616779327, + 0.013226734474301338, + 0.02008926123380661, + 0.01581578329205513, + 0.018316835165023804, + 0.006948564201593399, + 0.009282761253416538, + 0.00530085526406765, + 0.0038469848223030567, + 0.00707966135814786, + 0.0034707176964730024, + 0.004059175960719585, + 0.003618483664467931, + 0.004511349834501743, + 0.004012123215943575, + 0.00591921666637063, + 0.004354177974164486, + 0.0038759566377848387, + 0.004179587122052908, + 0.0050365193746984005, + 0.00411615613847971, + 0.009593289345502853, + 0.016077525913715363, + 0.008517355658113956, + 0.009460869245231152, + 0.008321666158735752, + 0.004888160154223442, + 0.005615636240690947, + 0.004780187737196684, + 0.00473021762445569, + 0.0044993688352406025, + 0.006584427785128355, + 0.008068648166954517, + 0.0024025877937674522, + 0.002688030246645212, + 0.00305589335039258, + 0.003448325674980879, + 0.005522833205759525, + 0.005911774933338165, + 0.009478839114308357, + 0.007251982111483812, + 0.00558695662766695, + 0.00634438032284379, + 0.005864758975803852, + 0.006240328308194876, + 0.04533105343580246, + 0.05731019377708435, + 0.04525010660290718, + 0.0316203236579895, + 0.0368775948882103, + 0.03484836220741272, + 0.016003992408514023, + 0.02347187139093876, + 0.014847001060843468, + 0.013805774971842766, + 0.00786785688251257, + 0.005402630195021629, + 0.007798083126544952, + 0.007793753873556852, + 0.005789350252598524, + 0.007257495541125536, + 0.07480781525373459, + 0.0683031678199768, + 0.013718405738472939, + 0.00866624340415001, + 0.01567017287015915, + 0.008736003190279007, + 0.008939997293055058, + 0.007664551958441734, + 0.004705200437456369, + 0.004544275812804699, + 0.004853862337768078, + 0.0054650139063596725, + 0.00531970988959074, + 0.005508956499397755, + 0.0036635848227888346, + 0.0033108291681855917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.018907787278294563 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic sloop sailboat with a white fiberglass hull and deep blue antifouling paint on the lower half?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the massive mainsail and foresail fully filled with wind, showing visible texture and stitching details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the side of the boat in a smooth tracking motion from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep azure ocean with white spray and mist crashing against the boat's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show water splashing onto the camera lens, momentarily blurring the image before clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07008550316095352 + ], + "opposite_logprob_tokens": [ + -2.6950855255126953 + ], + "logprob_reward": -0.07008550316095352, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07008550316095352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24083735048770905, + 0.050479043275117874, + 0.0348212793469429, + 0.07083030045032501, + 0.02726100943982601, + 0.042908333241939545, + 0.3878052532672882, + 0.04089541733264923, + 0.07263682782649994, + 0.031525205820798874 + ], + "credit_map": [ + 0.035675134509801865, + 0.04409649595618248, + 0.02664722129702568, + 0.02883380837738514, + 0.021617349237203598, + 0.027739671990275383, + 0.025403931736946106, + 0.030823739245533943, + 0.00801717210561037, + 0.010998811572790146, + 0.006531132850795984, + 0.004424344748258591, + 0.006136326584964991, + 0.004289865959435701, + 0.005329510662704706, + 0.004751876927912235, + 0.00463559664785862, + 0.005478456616401672, + 0.004587030503898859, + 0.004633634816855192, + 0.004609329160302877, + 0.004431804642081261, + 0.0035082974936813116, + 0.0029371276032179594, + 0.010008077137172222, + 0.015090584754943848, + 0.008352089673280716, + 0.009093801490962505, + 0.008073537610471249, + 0.00727660721167922, + 0.00709172897040844, + 0.0058438824489712715, + 0.003916698042303324, + 0.0042121876031160355, + 0.004346082918345928, + 0.00517065217718482, + 0.0032843968365341425, + 0.0020985796581953764, + 0.0018717783968895674, + 0.0023606345057487488, + 0.005276297684758902, + 0.00684292521327734, + 0.005725543946027756, + 0.00605781888589263, + 0.005177601240575314, + 0.004657758865505457, + 0.004639645107090473, + 0.004530745558440685, + 0.06061497703194618, + 0.08105622231960297, + 0.05832481384277344, + 0.042542144656181335, + 0.04999786615371704, + 0.04550115764141083, + 0.020915500819683075, + 0.028852542862296104, + 0.009586753323674202, + 0.00844517070800066, + 0.007241964805871248, + 0.004382411949336529, + 0.0032224198803305626, + 0.0031063349451869726, + 0.0023797110188752413, + 0.0025306520983576775, + 0.028136691078543663, + 0.02605370618402958, + 0.0039346907287836075, + 0.004019583575427532, + 0.003766390960663557, + 0.0023918503429740667, + 0.0026901718229055405, + 0.0016437452286481857, + 0.003929831087589264, + 0.004279272630810738, + 0.004083913750946522, + 0.00413103261962533, + 0.00494197616353631, + 0.003842402482405305, + 0.0033860744442790747, + 0.0029307017102837563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.07008550316095352 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic sloop sailboat with a white fiberglass hull and deep blue antifouling paint on the lower half?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the massive mainsail and foresail fully filled with wind, showing visible texture and stitching details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the side of the boat in a smooth tracking motion from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep azure ocean with white spray and mist crashing against the boat's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show water splashing onto the camera lens, momentarily blurring the image before clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005949999671429396 + ], + "opposite_logprob_tokens": [ + -5.130949974060059 + ], + "logprob_reward": -0.005949999671429396, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005949999671429396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10935519635677338, + 0.10238310694694519, + 0.0517389215528965, + 0.09629900008440018, + 0.060672577470541, + 0.16752742230892181, + 0.23022416234016418, + 0.10259151458740234, + 0.031495437026023865, + 0.04771265760064125 + ], + "credit_map": [ + 0.019565140828490257, + 0.033423908054828644, + 0.00966582354158163, + 0.011205989867448807, + 0.01272405032068491, + 0.008224203251302242, + 0.006394991651177406, + 0.008151085115969181, + 0.013776440173387527, + 0.015679026022553444, + 0.013506217859685421, + 0.013657161965966225, + 0.014566377736628056, + 0.011738738045096397, + 0.010630269534885883, + 0.008828877471387386, + 0.006979836151003838, + 0.006837034597992897, + 0.007395181339234114, + 0.00571622047573328, + 0.006508559919893742, + 0.006623669993132353, + 0.006009574513882399, + 0.005668841768056154, + 0.013310428708791733, + 0.019063441082835197, + 0.015532522462308407, + 0.016555536538362503, + 0.009479580447077751, + 0.006827982142567635, + 0.006852170452475548, + 0.008677328936755657, + 0.009865481406450272, + 0.009558864869177341, + 0.0068828873336315155, + 0.008753849193453789, + 0.007311898283660412, + 0.00640069367364049, + 0.004301734268665314, + 0.007597166579216719, + 0.020471811294555664, + 0.019960546866059303, + 0.027373148128390312, + 0.02760225348174572, + 0.016581129282712936, + 0.01795779913663864, + 0.017369110137224197, + 0.02021162584424019, + 0.030935389921069145, + 0.03780146315693855, + 0.0386529378592968, + 0.03387461230158806, + 0.023475145921111107, + 0.027564436197280884, + 0.017379119992256165, + 0.020541047677397728, + 0.03307762369513512, + 0.03050900064408779, + 0.00956562627106905, + 0.007037869654595852, + 0.006154948845505714, + 0.005262239836156368, + 0.005028198007494211, + 0.005955999717116356, + 0.00436601834371686, + 0.003523664316162467, + 0.004166864790022373, + 0.003926909528672695, + 0.003940877970308065, + 0.004889264702796936, + 0.0029408105183392763, + 0.003741026623174548, + 0.006876906380057335, + 0.0059666759334504604, + 0.0056588356383144855, + 0.00582079216837883, + 0.007146590854972601, + 0.005646866746246815, + 0.00657419441267848, + 0.004021796863526106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.005949999671429396 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic sloop sailboat with a white fiberglass hull and deep blue antifouling paint on the lower half?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the massive mainsail and foresail fully filled with wind, showing visible texture and stitching details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the side of the boat in a smooth tracking motion from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep azure ocean with white spray and mist crashing against the boat's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show water splashing onto the camera lens, momentarily blurring the image before clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076261102221906185 + ], + "opposite_logprob_tokens": [ + -4.882626056671143 + ], + "logprob_reward": -0.0076261102221906185, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076261102221906185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0770268365740776, + 0.2294055074453354, + 0.10642380267381668, + 0.029067806899547577, + 0.10912439227104187, + 0.06586948782205582, + 0.16621531546115875, + 0.0944351851940155, + 0.060538046061992645, + 0.061893660575151443 + ], + "credit_map": [ + 0.01401835959404707, + 0.022391771897673607, + 0.00741356797516346, + 0.008710017427802086, + 0.008916037157177925, + 0.005672764033079147, + 0.004433969035744667, + 0.005470341071486473, + 0.03500426933169365, + 0.05309813842177391, + 0.02731531299650669, + 0.025748737156391144, + 0.0257229283452034, + 0.030145861208438873, + 0.015668291598558426, + 0.01670197583734989, + 0.014285160228610039, + 0.016276279464364052, + 0.020762281492352486, + 0.017434582114219666, + 0.012936090119183064, + 0.007882784120738506, + 0.009022084064781666, + 0.007824547588825226, + 0.004500539507716894, + 0.004919072147458792, + 0.003957022912800312, + 0.00497957831248641, + 0.002752230502665043, + 0.002083267318084836, + 0.003550463356077671, + 0.0023256319109350443, + 0.015789300203323364, + 0.013743264600634575, + 0.01757180131971836, + 0.023238655179739, + 0.011262455955147743, + 0.008320623077452183, + 0.007590286433696747, + 0.011607999913394451, + 0.008334795013070107, + 0.008960348553955555, + 0.010051549412310123, + 0.009241292253136635, + 0.007196690421551466, + 0.009808268398046494, + 0.00553180743008852, + 0.006744740530848503, + 0.022113267332315445, + 0.02419266290962696, + 0.02772602066397667, + 0.02261318452656269, + 0.018409855663776398, + 0.01921551115810871, + 0.012352053076028824, + 0.01959274709224701, + 0.018586255609989166, + 0.013052825815975666, + 0.015398764982819557, + 0.01377271767705679, + 0.009556856006383896, + 0.008131665177643299, + 0.008466353639960289, + 0.007469750940799713, + 0.0160260871052742, + 0.014831052161753178, + 0.007714980281889439, + 0.006779973860830069, + 0.00435200659558177, + 0.0032237903214991093, + 0.004372273571789265, + 0.003237883560359478, + 0.011449619196355343, + 0.007694958709180355, + 0.009422344155609608, + 0.007911065593361855, + 0.004480539821088314, + 0.0075351521372795105, + 0.008027423173189163, + 0.005372558254748583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 4 + ], + "selected_reward": -0.0076261102221906185 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic sloop sailboat with a white fiberglass hull and deep blue antifouling paint on the lower half?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the massive mainsail and foresail fully filled with wind, showing visible texture and stitching details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the side of the boat in a smooth tracking motion from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep azure ocean with white spray and mist crashing against the boat's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show water splashing onto the camera lens, momentarily blurring the image before clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08901906758546829 + ], + "opposite_logprob_tokens": [ + -2.4640190601348877 + ], + "logprob_reward": -0.08901906758546829, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901906758546829, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41915497183799744, + 0.11318507045507431, + 0.05558663234114647, + 0.03144589066505432, + 0.06825815886259079, + 0.11207428574562073, + 0.06300077587366104, + 0.07986532896757126, + 0.03209354728460312, + 0.025335337966680527 + ], + "credit_map": [ + 0.06371515244245529, + 0.06541991978883743, + 0.036114633083343506, + 0.05371188372373581, + 0.04125401750206947, + 0.08095856755971909, + 0.038405802100896835, + 0.03957498073577881, + 0.014931121841073036, + 0.016552694141864777, + 0.015221552923321724, + 0.014560585841536522, + 0.02015940099954605, + 0.011020970530807972, + 0.012185722589492798, + 0.008553015999495983, + 0.00899161584675312, + 0.009287087246775627, + 0.011195865459740162, + 0.00869075208902359, + 0.004977653734385967, + 0.003654536558315158, + 0.00521031254902482, + 0.0035788060631603003, + 0.004497213754802942, + 0.004873996134847403, + 0.005425583571195602, + 0.004864589311182499, + 0.003291130531579256, + 0.0024108984507620335, + 0.0026121705304831266, + 0.003470306983217597, + 0.010693748481571674, + 0.011384915560483932, + 0.010054074227809906, + 0.010591208003461361, + 0.005941340699791908, + 0.006923660635948181, + 0.005457520019263029, + 0.007211693096905947, + 0.01316042523831129, + 0.013530001044273376, + 0.018912527710199356, + 0.01938735879957676, + 0.01165042258799076, + 0.010432071052491665, + 0.013213306665420532, + 0.01178817544132471, + 0.009127562865614891, + 0.008687064051628113, + 0.009122471325099468, + 0.008813817985355854, + 0.007680987939238548, + 0.006999657489359379, + 0.004787443205714226, + 0.0077817696146667, + 0.017380541190505028, + 0.012567320838570595, + 0.01111531537026167, + 0.009522606618702412, + 0.0102969566360116, + 0.008046932518482208, + 0.00639607198536396, + 0.004539585672318935, + 0.006684935186058283, + 0.00610975781455636, + 0.002675889525562525, + 0.004084830638021231, + 0.003074049251154065, + 0.002839003223925829, + 0.004224100615829229, + 0.002400981495156884, + 0.002473620930686593, + 0.002797899069264531, + 0.0038239601999521255, + 0.003878210671246052, + 0.003376281587406993, + 0.0032493227627128363, + 0.003159509738907218, + 0.002576532308012247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.08901906758546829 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide-angle shot follows the side of a sailboat, moving forward in a smooth tracking motion, evoking an immersive sense of sailing. A classic sloop, with a sleek white fiberglass hull and a deep blue antifouling paint on the lower half, plows through the deep azure waters. The massive mainsail and foresail, fully filled by the strong wind, create a full, powerful arc, the visible texture and stitching details of the sails striking in the sunlight. The boat rises and falls rhythmically with the waves, and the deep blue waves crash powerfully against the side, stirring up layers of white spray and a fine mist. Some of the water even splashes onto the camera, momentarily blurring the image before quickly sliding back down. The bright afternoon sun casts countless sparkling diamond-like points of light on the shimmering sea, illuminating the boat's wet teak deck and gleaming metal railings, creating a visual effect of ultra-high-definition realism. The horizon is clearly visible in the distance, dividing the sky, dotted with a few fluffy white clouds, from the vast ocean. The entire picture is full of a magnificent sense of adventure and freedom, while also showing an awe-inspiring force of nature.", + "seed": 28584, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_3/index_1459_seed_28584_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01234517318662256, + "vlm_logprob": -0.01234517318662256, + "vlm_yes_no_margin": 4.774999952316284, + "vlm_token_credit": -0.01234517318662256, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic sloop sailboat with a white fiberglass hull and deep blue antifouling paint on the lower half?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the massive mainsail and foresail fully filled with wind, showing visible texture and stitching details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the side of the boat in a smooth tracking motion from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep azure ocean with white spray and mist crashing against the boat's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show water splashing onto the camera lens, momentarily blurring the image before clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01298319362103939 + ], + "opposite_logprob_tokens": [ + -4.637983322143555 + ], + "logprob_reward": -0.01298319362103939, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.01298319362103939, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22199147939682007, + 0.06731315702199936, + 0.095122791826725, + 0.07007860392332077, + 0.06771349161863327, + 0.10232695937156677, + 0.06479769945144653, + 0.07320850342512131, + 0.052332084625959396, + 0.1851152628660202 + ], + "credit_map": [ + 0.02229641191661358, + 0.03796206787228584, + 0.025871045887470245, + 0.024468861520290375, + 0.04095837473869324, + 0.03346254676580429, + 0.019920406863093376, + 0.017051750794053078, + 0.009724395349621773, + 0.016016429290175438, + 0.00787253025919199, + 0.005591091699898243, + 0.008505483157932758, + 0.008324254304170609, + 0.00696330051869154, + 0.004315674304962158, + 0.013917206786572933, + 0.02866773121058941, + 0.00851211417466402, + 0.006209516432136297, + 0.008077092468738556, + 0.017106616869568825, + 0.005897556431591511, + 0.00673495139926672, + 0.006935741286724806, + 0.014959254302084446, + 0.008361171931028366, + 0.0048507219180464745, + 0.008289014920592308, + 0.01505383849143982, + 0.005633675958961248, + 0.0059951855801045895, + 0.0069728572852909565, + 0.009393233805894852, + 0.007560256868600845, + 0.005752785596996546, + 0.013305125758051872, + 0.012420125305652618, + 0.0059062521904706955, + 0.006402858067303896, + 0.012496273033320904, + 0.017387233674526215, + 0.010187646374106407, + 0.006981677375733852, + 0.023708125576376915, + 0.017029905691742897, + 0.006743245758116245, + 0.007792849093675613, + 0.008211299777030945, + 0.0090093445032835, + 0.005228853784501553, + 0.0064882622100412846, + 0.012742104940116405, + 0.011675151064991951, + 0.005138751119375229, + 0.0063039339147508144, + 0.008410777896642685, + 0.011169779114425182, + 0.00843481533229351, + 0.011220190674066544, + 0.010015035979449749, + 0.00698214303702116, + 0.009119245223701, + 0.007856515236198902, + 0.007048697676509619, + 0.01205996610224247, + 0.003407456446439028, + 0.004776204004883766, + 0.00754130631685257, + 0.007560238242149353, + 0.004462094511836767, + 0.005476125981658697, + 0.03937925025820732, + 0.014452394098043442, + 0.010224374942481518, + 0.05815918743610382, + 0.01460382528603077, + 0.024908384308218956, + 0.010932407341897488, + 0.012455424293875694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.01298319362103939 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic sloop sailboat with a white fiberglass hull and deep blue antifouling paint on the lower half?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the massive mainsail and foresail fully filled with wind, showing visible texture and stitching details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the side of the boat in a smooth tracking motion from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep azure ocean with white spray and mist crashing against the boat's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show water splashing onto the camera lens, momentarily blurring the image before clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018200064077973366 + ], + "opposite_logprob_tokens": [ + -4.018199920654297 + ], + "logprob_reward": -0.018200064077973366, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018200064077973366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1700943410396576, + 0.08399561792612076, + 0.0806654542684555, + 0.08476177603006363, + 0.0613902285695076, + 0.07997595518827438, + 0.056250471621751785, + 0.16392068564891815, + 0.05007891729474068, + 0.16886651515960693 + ], + "credit_map": [ + 0.015339606441557407, + 0.02631114050745964, + 0.023662403225898743, + 0.02280656062066555, + 0.028904441744089127, + 0.022243542596697807, + 0.016571007668972015, + 0.014255649410188198, + 0.010061811655759811, + 0.012993056327104568, + 0.010998317040503025, + 0.007947537116706371, + 0.01300997193902731, + 0.010842092335224152, + 0.008096077479422092, + 0.010046754963696003, + 0.006633174605667591, + 0.011316956020891666, + 0.0124715156853199, + 0.008752317167818546, + 0.0129743292927742, + 0.009899215772747993, + 0.00891130231320858, + 0.009706644341349602, + 0.013811898417770863, + 0.014989198185503483, + 0.009824078530073166, + 0.005132707767188549, + 0.011110147461295128, + 0.016176925972104073, + 0.006567121017724276, + 0.007149701472371817, + 0.006936469580978155, + 0.01280985213816166, + 0.005226136650890112, + 0.003659249749034643, + 0.012867567129433155, + 0.008761736564338207, + 0.005918173119425774, + 0.005211040377616882, + 0.00848664902150631, + 0.01679356023669243, + 0.008505906909704208, + 0.006956316065043211, + 0.016066938638687134, + 0.01050924975425005, + 0.0065996996127069, + 0.0060576340183615685, + 0.007022765465080738, + 0.00768172275274992, + 0.007051732856780291, + 0.0056205028668046, + 0.00893442053347826, + 0.006583182606846094, + 0.004647678229957819, + 0.008708471432328224, + 0.020016910508275032, + 0.024856403470039368, + 0.02222721464931965, + 0.022591818124055862, + 0.0195813849568367, + 0.019947508350014687, + 0.0166497603058815, + 0.018049683421850204, + 0.008455891162157059, + 0.006850070785731077, + 0.003406706964597106, + 0.003931041806936264, + 0.007950348779559135, + 0.006798597984015942, + 0.00439047534018755, + 0.008295786567032337, + 0.03546033799648285, + 0.013488655909895897, + 0.009663586504757404, + 0.054533518850803375, + 0.0126519575715065, + 0.018977968022227287, + 0.011789092794060707, + 0.012301390990614891 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.018200064077973366 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic sloop sailboat with a white fiberglass hull and deep blue antifouling paint on the lower half?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the massive mainsail and foresail fully filled with wind, showing visible texture and stitching details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the side of the boat in a smooth tracking motion from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep azure ocean with white spray and mist crashing against the boat's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show water splashing onto the camera lens, momentarily blurring the image before clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008633431047201157 + ], + "opposite_logprob_tokens": [ + -4.758633613586426 + ], + "logprob_reward": -0.008633431047201157, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008633431047201157, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07986035943031311, + 0.09076163917779922, + 0.061075661331415176, + 0.025605259463191032, + 0.12989448010921478, + 0.038864705711603165, + 0.06338507682085037, + 0.029304755851626396, + 0.31654348969459534, + 0.1647045910358429 + ], + "credit_map": [ + 0.016128594055771828, + 0.017366662621498108, + 0.009864652529358864, + 0.007684189360588789, + 0.008174236863851547, + 0.008441681042313576, + 0.0067932032980024815, + 0.0054071382619440556, + 0.007337009534239769, + 0.012458380311727524, + 0.010550374165177345, + 0.00987282581627369, + 0.012647912837564945, + 0.01620418205857277, + 0.008182388730347157, + 0.013508557341992855, + 0.014704247936606407, + 0.018326634541153908, + 0.006230277940630913, + 0.0042276810854673386, + 0.0036102093290537596, + 0.006537178996950388, + 0.0032586948946118355, + 0.0041807363741099834, + 0.003134818747639656, + 0.004135920200496912, + 0.003833470866084099, + 0.002741719363257289, + 0.003610698040574789, + 0.0032589815091341734, + 0.002715272596105933, + 0.002174378139898181, + 0.016624223440885544, + 0.025461766868829727, + 0.008521738462150097, + 0.0057973978109657764, + 0.019380822777748108, + 0.036793917417526245, + 0.009195671416819096, + 0.008118939585983753, + 0.005630614701658487, + 0.0063788434490561485, + 0.0031970590353012085, + 0.003571157343685627, + 0.005782669875770807, + 0.00665102107450366, + 0.0035412481520324945, + 0.004112090449780226, + 0.007982635870575905, + 0.00989528652280569, + 0.006673402618616819, + 0.006064563523977995, + 0.011979100294411182, + 0.007580720819532871, + 0.0054890746250748634, + 0.007720290683209896, + 0.0037629634607583284, + 0.002946807537227869, + 0.0036382596008479595, + 0.0033477982506155968, + 0.004704494960606098, + 0.0035497182980179787, + 0.0038666657637804747, + 0.0034880470484495163, + 0.044228166341781616, + 0.06194775924086571, + 0.021092703565955162, + 0.031033329665660858, + 0.05369297042489052, + 0.04429907724261284, + 0.026432260870933533, + 0.033817220479249954, + 0.02939864806830883, + 0.014868663623929024, + 0.0145340571179986, + 0.03575187921524048, + 0.018507277593016624, + 0.024453435093164444, + 0.011330134235322475, + 0.015860475599765778 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.008633431047201157 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic sloop sailboat with a white fiberglass hull and deep blue antifouling paint on the lower half?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the massive mainsail and foresail fully filled with wind, showing visible texture and stitching details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the side of the boat in a smooth tracking motion from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep azure ocean with white spray and mist crashing against the boat's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show water splashing onto the camera lens, momentarily blurring the image before clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013328249333426356 + ], + "opposite_logprob_tokens": [ + -6.626332759857178 + ], + "logprob_reward": -0.0013328249333426356, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013328249333426356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1838289350271225, + 0.18711528182029724, + 0.11149019747972488, + 0.05120816081762314, + 0.03392024338245392, + 0.05715158209204674, + 0.04920434206724167, + 0.06574598699808121, + 0.16730479896068573, + 0.09303052723407745 + ], + "credit_map": [ + 0.016439294442534447, + 0.02161051332950592, + 0.0287602711468935, + 0.032030537724494934, + 0.027611855417490005, + 0.02368236891925335, + 0.01727062277495861, + 0.016423439607024193, + 0.02177165448665619, + 0.027140114456415176, + 0.03399529308080673, + 0.02115747705101967, + 0.022303052246570587, + 0.02500137686729431, + 0.016260219737887383, + 0.019486084580421448, + 0.023349326103925705, + 0.018328379839658737, + 0.013891729526221752, + 0.009540481492877007, + 0.012226571328938007, + 0.015060852281749249, + 0.009053360670804977, + 0.010039485059678555, + 0.009722193703055382, + 0.011022873222827911, + 0.005803694948554039, + 0.0042854901403188705, + 0.004493551794439554, + 0.006574389059096575, + 0.0040219551883637905, + 0.0052840085700154305, + 0.004041224252432585, + 0.0062932102009654045, + 0.002706453436985612, + 0.0025142719969153404, + 0.005457304883748293, + 0.007776034530252218, + 0.002889689290896058, + 0.00224204920232296, + 0.008364714682102203, + 0.008778521791100502, + 0.0050570047460496426, + 0.003935749642550945, + 0.012562832795083523, + 0.01127425767481327, + 0.0038058708887547255, + 0.003372622886672616, + 0.006491606589406729, + 0.007361168507486582, + 0.00585532933473587, + 0.0056283604353666306, + 0.007627686019986868, + 0.005294365808367729, + 0.004936668090522289, + 0.0060091521590948105, + 0.007078253664076328, + 0.009939303621649742, + 0.006257149390876293, + 0.010878629982471466, + 0.00819653645157814, + 0.008190847933292389, + 0.008021526969969273, + 0.007183727342635393, + 0.02496860921382904, + 0.037089381366968155, + 0.009707932360470295, + 0.015709882602095604, + 0.02696801908314228, + 0.024455124512314796, + 0.012316394597291946, + 0.016089435666799545, + 0.012702557258307934, + 0.008666636422276497, + 0.00863881316035986, + 0.020431146025657654, + 0.009634294547140598, + 0.013225584290921688, + 0.008867782540619373, + 0.01086369901895523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.0013328249333426356 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic sloop sailboat with a white fiberglass hull and deep blue antifouling paint on the lower half?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the massive mainsail and foresail fully filled with wind, showing visible texture and stitching details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the side of the boat in a smooth tracking motion from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep azure ocean with white spray and mist crashing against the boat's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show water splashing onto the camera lens, momentarily blurring the image before clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057635225355625 + ], + "opposite_logprob_tokens": [ + -3.895576238632202 + ], + "logprob_reward": -0.02057635225355625, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057635225355625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03868412226438522, + 0.09760487079620361, + 0.031350694596767426, + 0.04929956793785095, + 0.018404508009552956, + 0.04604824259877205, + 0.03286275640130043, + 0.03474356606602669, + 0.1426856815814972, + 0.5083159804344177 + ], + "credit_map": [ + 0.0034658859949558973, + 0.005551881156861782, + 0.006355905439704657, + 0.005345137324184179, + 0.005631024483591318, + 0.0048706065863370895, + 0.003737234277650714, + 0.0037264430429786444, + 0.00880179088562727, + 0.01633288338780403, + 0.013843055814504623, + 0.00859836209565401, + 0.015545625239610672, + 0.016917996108531952, + 0.00936208013445139, + 0.008203069679439068, + 0.0044613489881157875, + 0.006227669306099415, + 0.004650932736694813, + 0.0023700662422925234, + 0.0034849978983402252, + 0.003620192175731063, + 0.003064554650336504, + 0.0034709321334958076, + 0.006978539749979973, + 0.008158249780535698, + 0.005893502850085497, + 0.004663869738578796, + 0.006610619369894266, + 0.008522220887243748, + 0.004339138045907021, + 0.0041334256529808044, + 0.0024344241246581078, + 0.003550612833350897, + 0.0026314109563827515, + 0.001558526884764433, + 0.0021266168914735317, + 0.0030645779334008694, + 0.001535896910354495, + 0.001502439146861434, + 0.004949800670146942, + 0.007164447568356991, + 0.004956889431923628, + 0.0034607150591909885, + 0.010768519714474678, + 0.007284054532647133, + 0.0035590839106589556, + 0.0039047319442033768, + 0.004297609906643629, + 0.004339627921581268, + 0.0026333427522331476, + 0.0039275274612009525, + 0.005520964041352272, + 0.0054826149716973305, + 0.0024732486344873905, + 0.004187816753983498, + 0.005248616449534893, + 0.005137813277542591, + 0.0030942843295633793, + 0.004855389706790447, + 0.004682387690991163, + 0.003976203501224518, + 0.003914854489266872, + 0.003834012895822525, + 0.019561169669032097, + 0.025352369993925095, + 0.010291971266269684, + 0.014150922186672688, + 0.02828853204846382, + 0.018757469952106476, + 0.011355656199157238, + 0.014927580021321774, + 0.11211377382278442, + 0.042232364416122437, + 0.03242950141429901, + 0.14847679436206818, + 0.03415537253022194, + 0.07851231098175049, + 0.02650638110935688, + 0.03388948738574982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.02057635225355625 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide-angle shot follows the side of a sailboat, moving forward in a smooth tracking motion, evoking an immersive sense of sailing. A classic sloop, with a sleek white fiberglass hull and a deep blue antifouling paint on the lower half, plows through the deep azure waters. The massive mainsail and foresail, fully filled by the strong wind, create a full, powerful arc, the visible texture and stitching details of the sails striking in the sunlight. The boat rises and falls rhythmically with the waves, and the deep blue waves crash powerfully against the side, stirring up layers of white spray and a fine mist. Some of the water even splashes onto the camera, momentarily blurring the image before quickly sliding back down. The bright afternoon sun casts countless sparkling diamond-like points of light on the shimmering sea, illuminating the boat's wet teak deck and gleaming metal railings, creating a visual effect of ultra-high-definition realism. The horizon is clearly visible in the distance, dividing the sky, dotted with a few fluffy white clouds, from the vast ocean. The entire picture is full of a magnificent sense of adventure and freedom, while also showing an awe-inspiring force of nature.", + "seed": 28585, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_3/index_1459_seed_28585_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05621835868805647, + "vlm_logprob": -0.05621835868805647, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.05621835868805647, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic sloop sailboat with a white fiberglass hull and deep blue antifouling paint on the lower half?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the massive mainsail and foresail fully filled with wind, showing visible texture and stitching details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the side of the boat in a smooth tracking motion from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep azure ocean with white spray and mist crashing against the boat's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show water splashing onto the camera lens, momentarily blurring the image before clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030982138589024544 + ], + "opposite_logprob_tokens": [ + -3.655982255935669 + ], + "logprob_reward": -0.030982138589024544, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.030982138589024544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12015121430158615, + 0.02738591656088829, + 0.025153057649731636, + 0.057202305644750595, + 0.049895092844963074, + 0.45442143082618713, + 0.08947273343801498, + 0.09714867174625397, + 0.03335459157824516, + 0.04581490531563759 + ], + "credit_map": [ + 0.018865082412958145, + 0.019305991008877754, + 0.01577392965555191, + 0.017904771491885185, + 0.017197351902723312, + 0.015483378432691097, + 0.01132898312062025, + 0.004291737452149391, + 0.003367126453667879, + 0.004030170384794474, + 0.003192904172465205, + 0.003022725461050868, + 0.004594804253429174, + 0.0026196446269750595, + 0.004294686950743198, + 0.0022638558875769377, + 0.0026500055100768805, + 0.0038322852924466133, + 0.003012016648426652, + 0.0035384809598326683, + 0.005116773769259453, + 0.001682749716565013, + 0.002975794253870845, + 0.0023449526634067297, + 0.004544793162494898, + 0.007247708737850189, + 0.007879475131630898, + 0.00811917893588543, + 0.005010089837014675, + 0.005157144274562597, + 0.014255649410188198, + 0.004988271743059158, + 0.0071155303157866, + 0.006948531139642, + 0.005246628541499376, + 0.007231695111840963, + 0.005221387837082148, + 0.00723864184692502, + 0.007294509094208479, + 0.0035981680266559124, + 0.037685610353946686, + 0.18236862123012543, + 0.010360857471823692, + 0.011989384889602661, + 0.027923699468374252, + 0.15445910394191742, + 0.020082540810108185, + 0.00955167319625616, + 0.011786254122853279, + 0.025792239233851433, + 0.007071575149893761, + 0.011302674189209938, + 0.006990241352468729, + 0.017641551792621613, + 0.005165656562894583, + 0.003722554538398981, + 0.014243200421333313, + 0.017936406657099724, + 0.013323270715773106, + 0.015341329388320446, + 0.013546394184231758, + 0.008902178145945072, + 0.006270233541727066, + 0.0075856707990169525, + 0.004078044556081295, + 0.0054523758590221405, + 0.0034417626447975636, + 0.003572290064767003, + 0.006424075458198786, + 0.00423889746889472, + 0.0038902999367564917, + 0.0022568493150174618, + 0.00679281959310174, + 0.007152051664888859, + 0.007263883948326111, + 0.004471277352422476, + 0.005855409428477287, + 0.005974921397864819, + 0.004786206874996424, + 0.003518336685374379 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.030982138589024544 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic sloop sailboat with a white fiberglass hull and deep blue antifouling paint on the lower half?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the massive mainsail and foresail fully filled with wind, showing visible texture and stitching details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the side of the boat in a smooth tracking motion from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep azure ocean with white spray and mist crashing against the boat's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show water splashing onto the camera lens, momentarily blurring the image before clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054995253682136536 + ], + "opposite_logprob_tokens": [ + -2.92999529838562 + ], + "logprob_reward": -0.054995253682136536, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054995253682136536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11151497066020966, + 0.03936915472149849, + 0.03275848180055618, + 0.06242585927248001, + 0.04807940125465393, + 0.24710406363010406, + 0.1322048157453537, + 0.2548503279685974, + 0.03814402595162392, + 0.03354889526963234 + ], + "credit_map": [ + 0.014962255954742432, + 0.018061712384223938, + 0.012227233499288559, + 0.016705133020877838, + 0.01992172747850418, + 0.015169116668403149, + 0.010625299997627735, + 0.0038424867670983076, + 0.004901192616671324, + 0.00663817347958684, + 0.0028220931999385357, + 0.006386639084666967, + 0.006505335681140423, + 0.003129086224362254, + 0.006095185410231352, + 0.0028914452996104956, + 0.0027634254656732082, + 0.005355825647711754, + 0.0031564997043460608, + 0.00419995654374361, + 0.006380448117852211, + 0.0023712676484137774, + 0.005564815364778042, + 0.0029662437736988068, + 0.005593921057879925, + 0.00849853828549385, + 0.0069131930358707905, + 0.008252829313278198, + 0.006525222212076187, + 0.007660362403839827, + 0.012349613010883331, + 0.006632176227867603, + 0.005320058669894934, + 0.007233561482280493, + 0.004274336155503988, + 0.005019767209887505, + 0.008689409121870995, + 0.007285390514880419, + 0.005437158048152924, + 0.0048197172582149506, + 0.021515214815735817, + 0.10138500481843948, + 0.005622208584100008, + 0.006479480769485235, + 0.013865163549780846, + 0.08501259982585907, + 0.008885912597179413, + 0.004338468424975872, + 0.016115659847855568, + 0.04546819627285004, + 0.005725562106817961, + 0.005673071835190058, + 0.01318264752626419, + 0.03514911234378815, + 0.006522585172206163, + 0.00436796247959137, + 0.04387637600302696, + 0.05738077312707901, + 0.028382007032632828, + 0.026933779940009117, + 0.03790019452571869, + 0.020761411637067795, + 0.023035245016217232, + 0.016580522060394287, + 0.004689814522862434, + 0.005493614822626114, + 0.004788585472851992, + 0.003711660159751773, + 0.0052177757024765015, + 0.006977966520935297, + 0.004003592766821384, + 0.0032610155176371336, + 0.004469714593142271, + 0.003108662785962224, + 0.005416829604655504, + 0.003319599898532033, + 0.0050477199256420135, + 0.005610888823866844, + 0.003391037695109844, + 0.003184441477060318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.054995253682136536 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic sloop sailboat with a white fiberglass hull and deep blue antifouling paint on the lower half?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the massive mainsail and foresail fully filled with wind, showing visible texture and stitching details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the side of the boat in a smooth tracking motion from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep azure ocean with white spray and mist crashing against the boat's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show water splashing onto the camera lens, momentarily blurring the image before clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076268198899924755 + ], + "opposite_logprob_tokens": [ + -4.882627010345459 + ], + "logprob_reward": -0.0076268198899924755, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076268198899924755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.033676113933324814, + 0.03705061972141266, + 0.03399476036429405, + 0.03751571476459503, + 0.0354793407022953, + 0.6339389681816101, + 0.03176504373550415, + 0.1145816519856453, + 0.02662232145667076, + 0.015375480055809021 + ], + "credit_map": [ + 0.004818398971110582, + 0.006211443804204464, + 0.004381740931421518, + 0.004006362985819578, + 0.005507381167262793, + 0.003919178154319525, + 0.0030668145045638084, + 0.0017647914355620742, + 0.0032833353616297245, + 0.00854333583265543, + 0.002678411779925227, + 0.004368265625089407, + 0.00831441767513752, + 0.002806868404150009, + 0.004684823099523783, + 0.0023711591493338346, + 0.002399224089458585, + 0.005430083256214857, + 0.0035053680185228586, + 0.004436615854501724, + 0.006307140924036503, + 0.0020254971459507942, + 0.005567371379584074, + 0.004323455039411783, + 0.0034236698411405087, + 0.007419114466756582, + 0.004160808864980936, + 0.004730277694761753, + 0.004993082489818335, + 0.004071247763931751, + 0.004704639781266451, + 0.0040128701366484165, + 0.0050278459675610065, + 0.005314616486430168, + 0.004295022692531347, + 0.006084968335926533, + 0.0037642770912498236, + 0.003942176699638367, + 0.004496158100664616, + 0.0025542753282934427, + 0.05557301640510559, + 0.27690237760543823, + 0.011575651355087757, + 0.009420076385140419, + 0.02825704589486122, + 0.2323119044303894, + 0.010422151535749435, + 0.009476752951741219, + 0.00445901183411479, + 0.007096938788890839, + 0.0022749428171664476, + 0.0037709646858274937, + 0.003749724943190813, + 0.005683676805347204, + 0.0028506761882454157, + 0.0018791074398905039, + 0.018986253067851067, + 0.024698413908481598, + 0.013388109393417835, + 0.012041050009429455, + 0.017977416515350342, + 0.009264442138373852, + 0.011155686341226101, + 0.007070265244692564, + 0.0033110063523054123, + 0.00490466458722949, + 0.003724006237462163, + 0.003387792268767953, + 0.0039119599387049675, + 0.003142193192616105, + 0.002876427723094821, + 0.0013642696430906653, + 0.0021597743034362793, + 0.0018947615753859282, + 0.002414594404399395, + 0.0015767986187711358, + 0.0019267144380137324, + 0.001899435301311314, + 0.0019076659809798002, + 0.0015957339201122522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 3 + ], + "selected_reward": -0.0076268198899924755 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic sloop sailboat with a white fiberglass hull and deep blue antifouling paint on the lower half?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the massive mainsail and foresail fully filled with wind, showing visible texture and stitching details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the side of the boat in a smooth tracking motion from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep azure ocean with white spray and mist crashing against the boat's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show water splashing onto the camera lens, momentarily blurring the image before clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076261102221906185 + ], + "opposite_logprob_tokens": [ + -4.882626056671143 + ], + "logprob_reward": -0.0076261102221906185, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076261102221906185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1214306429028511, + 0.1389056146144867, + 0.10886584222316742, + 0.06294355541467667, + 0.05456690117716789, + 0.2798246145248413, + 0.1370820254087448, + 0.04981577396392822, + 0.03518866375088692, + 0.011376382783055305 + ], + "credit_map": [ + 0.015348711051046848, + 0.021441642194986343, + 0.011299815028905869, + 0.020712660625576973, + 0.02158990316092968, + 0.012825281359255314, + 0.013545749709010124, + 0.004666885826736689, + 0.012484611943364143, + 0.027450069785118103, + 0.00738421780988574, + 0.02472088672220707, + 0.02919924445450306, + 0.011224967427551746, + 0.019321629777550697, + 0.007119991350919008, + 0.009650185704231262, + 0.01935158297419548, + 0.013131853193044662, + 0.01557092647999525, + 0.019841846078634262, + 0.007021427154541016, + 0.011951862834393978, + 0.012346158735454082, + 0.006112312898039818, + 0.009985653683543205, + 0.007003449369221926, + 0.0096025625243783, + 0.0056960429064929485, + 0.0058435373939573765, + 0.013612720184028149, + 0.0050872815772891045, + 0.009131565690040588, + 0.00942708645015955, + 0.005438929423689842, + 0.008210325613617897, + 0.005491775460541248, + 0.00617041252553463, + 0.00750949140638113, + 0.0031873155385255814, + 0.025218326598405838, + 0.12207482010126114, + 0.004988321103155613, + 0.005024252459406853, + 0.012404688633978367, + 0.10098515450954437, + 0.004977018106728792, + 0.004152046982198954, + 0.020468764007091522, + 0.027516309171915054, + 0.017299318686127663, + 0.03210864216089249, + 0.009273873642086983, + 0.013063582591712475, + 0.011775369755923748, + 0.005576167721301317, + 0.008370415307581425, + 0.010771939530968666, + 0.0048563238233327866, + 0.0052762883715331554, + 0.008205489255487919, + 0.004573916085064411, + 0.004792922642081976, + 0.0029684784822165966, + 0.004250062629580498, + 0.007424029987305403, + 0.004892511758953333, + 0.004202972631901503, + 0.0049148378893733025, + 0.003672081744298339, + 0.003875770838931203, + 0.0019563937094062567, + 0.001476043718867004, + 0.0016761374427005649, + 0.0023937670048326254, + 0.001759663806296885, + 0.0011330803390592337, + 0.0010693228105083108, + 0.0011353215668350458, + 0.0007330456865020096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 6 + ], + "selected_reward": -0.0076261102221906185 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic sloop sailboat with a white fiberglass hull and deep blue antifouling paint on the lower half?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the massive mainsail and foresail fully filled with wind, showing visible texture and stitching details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the side of the boat in a smooth tracking motion from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep azure ocean with white spray and mist crashing against the boat's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show water splashing onto the camera lens, momentarily blurring the image before clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17986147105693817 + ], + "opposite_logprob_tokens": [ + -1.8048614263534546 + ], + "logprob_reward": -0.17986147105693817, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17986147105693817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0964721292257309, + 0.3229418098926544, + 0.07400820404291153, + 0.0854455828666687, + 0.0955844298005104, + 0.1027931496500969, + 0.04417581483721733, + 0.05774106830358505, + 0.08766322582960129, + 0.03317461907863617 + ], + "credit_map": [ + 0.015330505557358265, + 0.017372917383909225, + 0.011956169269979, + 0.015655897557735443, + 0.015270366333425045, + 0.007720829453319311, + 0.009670023806393147, + 0.003495418466627598, + 0.0369754359126091, + 0.07110701501369476, + 0.01878640614449978, + 0.03669567033648491, + 0.07839293032884598, + 0.028434917330741882, + 0.0356069877743721, + 0.016942420974373817, + 0.0075280447490513325, + 0.014219817705452442, + 0.006305876653641462, + 0.009143330156803131, + 0.013018875382840633, + 0.006247387267649174, + 0.008121727965772152, + 0.009423145093023777, + 0.005314708687365055, + 0.013747344724833965, + 0.010055653750896454, + 0.012522762641310692, + 0.012803436256945133, + 0.011037563905119896, + 0.012803872115910053, + 0.007160237058997154, + 0.010206760838627815, + 0.012517433613538742, + 0.013524500653147697, + 0.015367679297924042, + 0.010256211273372173, + 0.007860755547881126, + 0.017164383083581924, + 0.008686700835824013, + 0.008343884721398354, + 0.017923733219504356, + 0.011967265978455544, + 0.011622265912592411, + 0.012956812046468258, + 0.014056923799216747, + 0.017918966710567474, + 0.008003291673958302, + 0.004695321898907423, + 0.006855292711406946, + 0.003910232800990343, + 0.008694416843354702, + 0.00605397205799818, + 0.006175910588353872, + 0.005238007754087448, + 0.0025526583194732666, + 0.009108892641961575, + 0.011992005631327629, + 0.0058468715287745, + 0.008616393432021141, + 0.009177202358841896, + 0.005199146922677755, + 0.0045895688235759735, + 0.0032109813764691353, + 0.01128897350281477, + 0.014220994897186756, + 0.009896844625473022, + 0.010916361585259438, + 0.017800411209464073, + 0.009424036368727684, + 0.009151691570878029, + 0.0049639116041362286, + 0.004849802237004042, + 0.004025872331112623, + 0.004467362072318792, + 0.004384047817438841, + 0.0043502734042704105, + 0.003133950522169471, + 0.0040125371888279915, + 0.003950771875679493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.17986147105693817 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_4/index_tensor([8850, 8850])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_4/index_tensor([8850, 8850])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..03fc04a575ed386466003b58e96fbd80dc9effda --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_4/index_tensor([8850, 8850])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up static shot frames a simple, clear glass filled with water against a deep, featureless black background. A round, white, talc tablet is dropped into the water, and as it immerses itself, it explodes into a violent chemical reaction. The scene is rendered in dramatic, high-contrast black and white, with a sharp key light from the side sculpting every detail. A chaotic hiss of bright white microbubbles erupts violently from the tablet's surface, forming a turbulent, upward-rushing column that stirs the dark water. The tablet itself rapidly tumbles and disintegrates, releasing larger, shimmering bubbles that rush to the surface and burst. The visual is a powerful, contained storm, the stark white of the flaky chaos contrasting sharply with the inky black water, emphasizing the primal, kinetic energy of the dissolution.", + "seed": 14245, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_4/index_8850_seed_14245_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03877815278246999, + "vlm_logprob": -0.03877815278246999, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.03877815278246999, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glass filled with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deep, featureless black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round, white tablet dropped into the water and shown disintegrating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style high-contrast black and white with a sharp key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09491543471813202 + ], + "opposite_logprob_tokens": [ + -2.4699153900146484 + ], + "logprob_reward": -0.09491543471813202, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09491543471813202, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25204163789749146, + 0.194295272231102, + 0.06172893941402435, + 0.12847334146499634, + 0.09336000680923462, + 0.062478844076395035, + 0.07234036177396774, + 0.061418939381837845, + 0.039514604955911636, + 0.03434803709387779 + ], + "credit_map": [ + 0.05013815313577652, + 0.022757647559046745, + 0.03122258000075817, + 0.03125980496406555, + 0.020181721076369286, + 0.04033356159925461, + 0.02632366120815277, + 0.02982451394200325, + 0.03490342199802399, + 0.017471401020884514, + 0.026338323950767517, + 0.01973910629749298, + 0.017937542870640755, + 0.02221352979540825, + 0.03335026651620865, + 0.022341685369610786, + 0.007385365664958954, + 0.00552718248218298, + 0.008142682723701, + 0.007348146755248308, + 0.006642578635364771, + 0.0060913232155144215, + 0.013365617953240871, + 0.007226042915135622, + 0.019061502069234848, + 0.013831284828484058, + 0.018821923062205315, + 0.010681047104299068, + 0.010448489338159561, + 0.012199204415082932, + 0.025516660884022713, + 0.017913229763507843, + 0.016039956361055374, + 0.012989037670195103, + 0.013354688882827759, + 0.008952971547842026, + 0.011873104609549046, + 0.00637776218354702, + 0.01632203720510006, + 0.007450447883456945, + 0.009760767221450806, + 0.007945718243718147, + 0.010261774063110352, + 0.005880030337721109, + 0.005531188566237688, + 0.005847943481057882, + 0.011454029008746147, + 0.005797390826046467, + 0.011970684863626957, + 0.010737881995737553, + 0.010648141615092754, + 0.007265910040587187, + 0.005380984861403704, + 0.0062650153413414955, + 0.008421968668699265, + 0.011649779975414276, + 0.006784264463931322, + 0.0063951401971280575, + 0.007804026361554861, + 0.009869491681456566, + 0.006147800944745541, + 0.008870061486959457, + 0.007963367737829685, + 0.00758478743955493, + 0.0054644010961055756, + 0.004879686050117016, + 0.006169202271848917, + 0.0049036177806556225, + 0.0024458817206323147, + 0.003202419728040695, + 0.0035493592731654644, + 0.008900036104023457, + 0.004024224821478128, + 0.003648698329925537, + 0.004613899625837803, + 0.003908423241227865, + 0.0032329356763511896, + 0.003483742242679, + 0.003842115169391036, + 0.007593997288495302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.09491543471813202 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glass filled with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deep, featureless black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round, white tablet dropped into the water and shown disintegrating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style high-contrast black and white with a sharp key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014207405038177967 + ], + "opposite_logprob_tokens": [ + -4.264207363128662 + ], + "logprob_reward": -0.014207405038177967, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014207405038177967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27959033846855164, + 0.26141607761383057, + 0.04351874068379402, + 0.07887492328882217, + 0.05483686923980713, + 0.04829120263457298, + 0.09141476452350616, + 0.04735275357961655, + 0.024189939722418785, + 0.07051443308591843 + ], + "credit_map": [ + 0.053499363362789154, + 0.041652169078588486, + 0.03776824101805687, + 0.03533978387713432, + 0.015004893764853477, + 0.04008539393544197, + 0.02180366776883602, + 0.034436799585819244, + 0.07922904938459396, + 0.01430327445268631, + 0.04258487746119499, + 0.015490975230932236, + 0.015808049589395523, + 0.04451153427362442, + 0.014175117947161198, + 0.03531317785382271, + 0.009776520542800426, + 0.00440251873806119, + 0.006980775855481625, + 0.004586069379001856, + 0.00345168961212039, + 0.004936055280268192, + 0.004011821001768112, + 0.00537328515201807, + 0.018195202574133873, + 0.008918310515582561, + 0.015275785699486732, + 0.006070063915103674, + 0.007769162766635418, + 0.0046296147629618645, + 0.009662835858762264, + 0.008353942073881626, + 0.015321841463446617, + 0.006209115497767925, + 0.011148969642817974, + 0.005700622219592333, + 0.004271309357136488, + 0.0034994117449969053, + 0.004044439177960157, + 0.0046411543153226376, + 0.009182292968034744, + 0.007657301612198353, + 0.008736207149922848, + 0.004112702328711748, + 0.0038944086991250515, + 0.005073921289294958, + 0.005691361613571644, + 0.003943005111068487, + 0.022476904094219208, + 0.00944841280579567, + 0.016565604135394096, + 0.008045255206525326, + 0.007319652009755373, + 0.006106140092015266, + 0.013843805529177189, + 0.007608989719301462, + 0.006328223273158073, + 0.005280464421957731, + 0.0057226745411753654, + 0.008513635955750942, + 0.003869426902383566, + 0.006336198654025793, + 0.005473388358950615, + 0.005828738212585449, + 0.0033713066950440407, + 0.0031464812345802784, + 0.00310590211302042, + 0.0052615562453866005, + 0.001363174756988883, + 0.0021355641074478626, + 0.0026075325440615416, + 0.003198418766260147, + 0.008097142912447453, + 0.007886058650910854, + 0.00861452892422676, + 0.016973892226815224, + 0.005776829551905394, + 0.005149951204657555, + 0.009676885791122913, + 0.008339142426848412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.014207405038177967 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glass filled with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deep, featureless black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round, white tablet dropped into the water and shown disintegrating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style high-contrast black and white with a sharp key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005667987279593945 + ], + "opposite_logprob_tokens": [ + -7.5005669593811035 + ], + "logprob_reward": -0.0005667987279593945, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005667987279593945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5054575204849243, + 0.16619911789894104, + 0.05296695604920387, + 0.03113737516105175, + 0.020383957773447037, + 0.023508137091994286, + 0.04409262537956238, + 0.04197030887007713, + 0.035110779106616974, + 0.07917314022779465 + ], + "credit_map": [ + 0.10158946365118027, + 0.03789571672677994, + 0.07039958983659744, + 0.05608950927853584, + 0.03324129059910774, + 0.08747927099466324, + 0.051759134978055954, + 0.06700357794761658, + 0.048417940735816956, + 0.009827672503888607, + 0.02556217648088932, + 0.012258995324373245, + 0.009797166101634502, + 0.029796723276376724, + 0.008887844160199165, + 0.021650608628988266, + 0.007184403948485851, + 0.0053395857103168964, + 0.0065856026485562325, + 0.006565477233380079, + 0.005378257017582655, + 0.007517625577747822, + 0.009229952469468117, + 0.0051660495810210705, + 0.004486422520130873, + 0.003449848620221019, + 0.004562607500702143, + 0.003313652239739895, + 0.0028474905993789434, + 0.002947877859696746, + 0.005430127494037151, + 0.004099349025636911, + 0.0026168087497353554, + 0.0019875906873494387, + 0.0029288020450621843, + 0.00265288888476789, + 0.002911201911047101, + 0.0020364720840007067, + 0.0026320714969187975, + 0.0026181212160736322, + 0.003948986530303955, + 0.0029927045106887817, + 0.0039792838506400585, + 0.0020557965617626905, + 0.0018996584694832563, + 0.0024381752591580153, + 0.0034984133671969175, + 0.002695118309929967, + 0.004967826418578625, + 0.005719166249036789, + 0.005489241797477007, + 0.004914549179375172, + 0.005672483239322901, + 0.006575603969395161, + 0.005978791508823633, + 0.004774962551891804, + 0.004910336807370186, + 0.0046297842636704445, + 0.005149506498128176, + 0.006734281312674284, + 0.003995212726294994, + 0.005927995312958956, + 0.005044977180659771, + 0.005578217562288046, + 0.004492627922445536, + 0.005204974673688412, + 0.004865096881985664, + 0.007312989328056574, + 0.0020004822872579098, + 0.003659809473901987, + 0.00362168881110847, + 0.00395311089232564, + 0.009521129541099072, + 0.010449426248669624, + 0.0093574533239007, + 0.016559487208724022, + 0.006281164474785328, + 0.006129220128059387, + 0.00873893778771162, + 0.01213632058352232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0005667987279593945 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glass filled with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deep, featureless black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round, white tablet dropped into the water and shown disintegrating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style high-contrast black and white with a sharp key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01420164667069912 + ], + "opposite_logprob_tokens": [ + -4.264201641082764 + ], + "logprob_reward": -0.01420164667069912, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01420164667069912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3868974447250366, + 0.18046419322490692, + 0.08347746729850769, + 0.047390636056661606, + 0.04064406454563141, + 0.04614173620939255, + 0.03475017100572586, + 0.03270738944411278, + 0.030079906806349754, + 0.11744701862335205 + ], + "credit_map": [ + 0.07343477010726929, + 0.02999262884259224, + 0.04986506327986717, + 0.05749054625630379, + 0.025465913116931915, + 0.06314069032669067, + 0.03833555430173874, + 0.04917224124073982, + 0.05271318927407265, + 0.013739993795752525, + 0.02951914072036743, + 0.01190009992569685, + 0.011800093576312065, + 0.025815945118665695, + 0.01055612787604332, + 0.02441958710551262, + 0.010370361618697643, + 0.007452747318893671, + 0.011003650724887848, + 0.012990296818315983, + 0.009413075633347034, + 0.012663769535720348, + 0.012677567079663277, + 0.0069059934467077255, + 0.007804410997778177, + 0.005328997038304806, + 0.007292496971786022, + 0.004507986828684807, + 0.004448527470231056, + 0.002995549002662301, + 0.009162040427327156, + 0.005850623827427626, + 0.007923581637442112, + 0.00388971995562315, + 0.0063933408819139, + 0.004469577223062515, + 0.003904229262843728, + 0.003550884546712041, + 0.004585434217005968, + 0.00592729402706027, + 0.007861842401325703, + 0.005132749676704407, + 0.007019089534878731, + 0.004262813366949558, + 0.0032996907830238342, + 0.0043404595926404, + 0.0089761633425951, + 0.005248925648629665, + 0.006161992438137531, + 0.003638897091150284, + 0.005151485092937946, + 0.0032649561762809753, + 0.0031414490658789873, + 0.0039062099531292915, + 0.006617885082960129, + 0.0028672930784523487, + 0.004798876121640205, + 0.004088063724339008, + 0.0043113273568451405, + 0.004731474444270134, + 0.0032829931005835533, + 0.003098395885899663, + 0.00434234319254756, + 0.004053914453834295, + 0.003588138148188591, + 0.003924299031496048, + 0.004006804898381233, + 0.0068960366770625114, + 0.0017245732014998794, + 0.0025221090763807297, + 0.002913590520620346, + 0.004504353739321232, + 0.009522837586700916, + 0.011713218875229359, + 0.015529907308518887, + 0.02426164411008358, + 0.011035752482712269, + 0.008952044881880283, + 0.021036578342318535, + 0.0153950285166502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.01420164667069912 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glass filled with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deep, featureless black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round, white tablet dropped into the water and shown disintegrating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style high-contrast black and white with a sharp key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999947875738144 + ], + "opposite_logprob_tokens": [ + -2.6949994564056396 + ], + "logprob_reward": -0.06999947875738144, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999947875738144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2889399528503418, + 0.155568927526474, + 0.08268198370933533, + 0.08467007428407669, + 0.03217068687081337, + 0.04491424933075905, + 0.04685310646891594, + 0.09614057093858719, + 0.03952241316437721, + 0.12853804230690002 + ], + "credit_map": [ + 0.06499986350536346, + 0.030134664848446846, + 0.03662960231304169, + 0.03775949031114578, + 0.014456531964242458, + 0.042394112795591354, + 0.023352473974227905, + 0.039213210344314575, + 0.026690484955906868, + 0.008768640458583832, + 0.025928806513547897, + 0.015472210012376308, + 0.011537499725818634, + 0.026512671262025833, + 0.021046863868832588, + 0.01961173303425312, + 0.011258040554821491, + 0.00877160020172596, + 0.010200965218245983, + 0.012727033346891403, + 0.007766663562506437, + 0.010880090296268463, + 0.012009491212666035, + 0.00906810350716114, + 0.016586454585194588, + 0.009752701967954636, + 0.015881454572081566, + 0.007806271780282259, + 0.01021726243197918, + 0.005054463632404804, + 0.010944195091724396, + 0.008427268825471401, + 0.008066359907388687, + 0.003536039963364601, + 0.006294922903180122, + 0.0024678544141352177, + 0.0032898983918130398, + 0.002392232185229659, + 0.0026169673074036837, + 0.00350641505792737, + 0.008138911798596382, + 0.004960200749337673, + 0.008127891458570957, + 0.00445530517026782, + 0.0031032422557473183, + 0.004853169433772564, + 0.007307747378945351, + 0.003967783413827419, + 0.00536328786984086, + 0.006875184830278158, + 0.005629728548228741, + 0.007150150369852781, + 0.005623729433864355, + 0.004889798816293478, + 0.006831496022641659, + 0.004489729180932045, + 0.012022965587675571, + 0.010840735398232937, + 0.013070688582956791, + 0.014674852602183819, + 0.00960534531623125, + 0.013243700377643108, + 0.01244595181196928, + 0.010236325673758984, + 0.005485051311552525, + 0.0043925633653998375, + 0.005065485369414091, + 0.005173819605261087, + 0.00441243639215827, + 0.005642272531986237, + 0.004483258351683617, + 0.0048675257712602615, + 0.013254405930638313, + 0.016325296834111214, + 0.014939555898308754, + 0.025057272985577583, + 0.010846794582903385, + 0.011205299757421017, + 0.0177422147244215, + 0.019167209044098854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.06999947875738144 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot frames a simple, clear glass filled with water against a deep, featureless black background. A round, white, talc tablet is dropped into the water, and as it immerses itself, it explodes into a violent chemical reaction. The scene is rendered in dramatic, high-contrast black and white, with a sharp key light from the side sculpting every detail. A chaotic hiss of bright white microbubbles erupts violently from the tablet's surface, forming a turbulent, upward-rushing column that stirs the dark water. The tablet itself rapidly tumbles and disintegrates, releasing larger, shimmering bubbles that rush to the surface and burst. The visual is a powerful, contained storm, the stark white of the flaky chaos contrasting sharply with the inky black water, emphasizing the primal, kinetic energy of the dissolution.", + "seed": 14246, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_4/index_8850_seed_14246_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05985032762400806, + "vlm_logprob": -0.05985032762400806, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.05985032762400806, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glass filled with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deep, featureless black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round, white tablet dropped into the water and shown disintegrating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style high-contrast black and white with a sharp key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022328289225697517 + ], + "opposite_logprob_tokens": [ + -3.8973283767700195 + ], + "logprob_reward": -0.022328289225697517, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.022328289225697517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1893993616104126, + 0.14412926137447357, + 0.05294979363679886, + 0.060384366661310196, + 0.058685366064310074, + 0.22243501245975494, + 0.03762407228350639, + 0.15416467189788818, + 0.03396783769130707, + 0.04626024514436722 + ], + "credit_map": [ + 0.024927347898483276, + 0.014845774509012699, + 0.029205814003944397, + 0.022122599184513092, + 0.014268431812524796, + 0.0382855199277401, + 0.021543540060520172, + 0.024200333282351494, + 0.020130624994635582, + 0.011394885368645191, + 0.02409801259636879, + 0.02395353838801384, + 0.01155123021453619, + 0.014185157604515553, + 0.019482187926769257, + 0.019333630800247192, + 0.00953582115471363, + 0.005027646664530039, + 0.005828219000250101, + 0.006386413238942623, + 0.006831514183431864, + 0.008326096460223198, + 0.005943706259131432, + 0.0050703734159469604, + 0.011915436014533043, + 0.006093417294323444, + 0.007201024796813726, + 0.011526697315275669, + 0.005371874198317528, + 0.005596832372248173, + 0.008066576905548573, + 0.0046125068329274654, + 0.008254023268818855, + 0.006013273727148771, + 0.008941331878304482, + 0.0050428868271410465, + 0.005226627457886934, + 0.006277905777096748, + 0.007526192348450422, + 0.01140312198549509, + 0.02909182198345661, + 0.03268177807331085, + 0.027739498764276505, + 0.030115121975541115, + 0.01894730143249035, + 0.02477097138762474, + 0.034747619181871414, + 0.024340899661183357, + 0.00564555311575532, + 0.006199483759701252, + 0.003280685283243656, + 0.004485104698687792, + 0.005255469121038914, + 0.004642285406589508, + 0.0029738580342382193, + 0.0051416330970823765, + 0.023911943659186363, + 0.03009009175002575, + 0.014725659042596817, + 0.014154509641230106, + 0.011476434767246246, + 0.0279577374458313, + 0.016171816736459732, + 0.0156764704734087, + 0.010069725103676319, + 0.005217236001044512, + 0.0034624736290425062, + 0.002198396250605583, + 0.0026034973561763763, + 0.0033080419525504112, + 0.0019784104079008102, + 0.005130057223141193, + 0.004691643174737692, + 0.011147923767566681, + 0.004845934920012951, + 0.004642253275960684, + 0.0037650030571967363, + 0.007756621111184359, + 0.0045281886123120785, + 0.004882676061242819 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.022328289225697517 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glass filled with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deep, featureless black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round, white tablet dropped into the water and shown disintegrating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style high-contrast black and white with a sharp key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307573288679123 + ], + "opposite_logprob_tokens": [ + -3.1680757999420166 + ], + "logprob_reward": -0.04307573288679123, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307573288679123, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3142598271369934, + 0.05277334153652191, + 0.04119434580206871, + 0.04261329770088196, + 0.16338922083377838, + 0.04202551022171974, + 0.032944679260253906, + 0.11716825515031815, + 0.04720903933048248, + 0.14642243087291718 + ], + "credit_map": [ + 0.04333728179335594, + 0.03342634439468384, + 0.04705176129937172, + 0.03217789903283119, + 0.02014509215950966, + 0.04846705496311188, + 0.04096592590212822, + 0.04868851229548454, + 0.007150784134864807, + 0.0057319398038089275, + 0.008181118406355381, + 0.008583830669522285, + 0.005563751328736544, + 0.004664688836783171, + 0.00569929787889123, + 0.0071979365311563015, + 0.006448127795010805, + 0.004406656604260206, + 0.004267649259418249, + 0.004470926709473133, + 0.0063934712670743465, + 0.006887936498969793, + 0.0035348329693078995, + 0.004784747026860714, + 0.007295691408216953, + 0.005439981818199158, + 0.00604380713775754, + 0.0036263323854655027, + 0.004183618351817131, + 0.005276297219097614, + 0.005047972779721022, + 0.00569960055872798, + 0.026763301342725754, + 0.012451007030904293, + 0.024332275614142418, + 0.029136905446648598, + 0.01035313867032528, + 0.03222764655947685, + 0.00998399406671524, + 0.018140971660614014, + 0.006750460714101791, + 0.0045915935188531876, + 0.007559291087090969, + 0.004488085862249136, + 0.002595417434349656, + 0.00569656677544117, + 0.0056852069683372974, + 0.004658891353756189, + 0.004647336434572935, + 0.00324994046241045, + 0.002656666561961174, + 0.005135516170412302, + 0.003211390692740679, + 0.005215214565396309, + 0.0027098103892058134, + 0.00611881073564291, + 0.018571477383375168, + 0.02790086716413498, + 0.013876235112547874, + 0.011440487578511238, + 0.009061654098331928, + 0.017029665410518646, + 0.009884888306260109, + 0.009402994066476822, + 0.011727344244718552, + 0.011114676482975483, + 0.004425661638379097, + 0.004500887356698513, + 0.0025619789958000183, + 0.004513331688940525, + 0.002815528539940715, + 0.005549636669456959, + 0.01420958898961544, + 0.03371507301926613, + 0.01715792715549469, + 0.01645750179886818, + 0.012124140746891499, + 0.02291932702064514, + 0.013119488023221493, + 0.0167193915694952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.04307573288679123 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glass filled with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deep, featureless black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round, white tablet dropped into the water and shown disintegrating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style high-contrast black and white with a sharp key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002493250882253051 + ], + "opposite_logprob_tokens": [ + -6.002493381500244 + ], + "logprob_reward": -0.002493250882253051, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002493250882253051, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32350555062294006, + 0.056426774710416794, + 0.06707631051540375, + 0.12762591242790222, + 0.1636086404323578, + 0.07788146287202835, + 0.06163527071475983, + 0.06004248559474945, + 0.02492363005876541, + 0.03727398440241814 + ], + "credit_map": [ + 0.04250374436378479, + 0.029355712234973907, + 0.04391808435320854, + 0.024684159085154533, + 0.02113622985780239, + 0.04268405959010124, + 0.045059993863105774, + 0.0741635411977768, + 0.008613967336714268, + 0.004945988766849041, + 0.008115043863654137, + 0.009163838811218739, + 0.005644035059958696, + 0.0051741283386945724, + 0.007457011379301548, + 0.007312760688364506, + 0.009682981297373772, + 0.006539489608258009, + 0.010593848302960396, + 0.01040035579353571, + 0.008764058351516724, + 0.007397868204861879, + 0.0069790193811059, + 0.006718689575791359, + 0.029029469937086105, + 0.015769848600029945, + 0.017870506271719933, + 0.01697535626590252, + 0.010234540328383446, + 0.013842263258993626, + 0.01321655698120594, + 0.01068737544119358, + 0.025977646932005882, + 0.013520247302949429, + 0.01502967905253172, + 0.011411421932280064, + 0.016591735184192657, + 0.04782300069928169, + 0.011935792863368988, + 0.02131912112236023, + 0.011274885386228561, + 0.008178968913853168, + 0.011186493560671806, + 0.008648496121168137, + 0.008629759773612022, + 0.009321277029812336, + 0.01241473201662302, + 0.008226852864027023, + 0.008749354630708694, + 0.00979580357670784, + 0.005010091699659824, + 0.011122790165245533, + 0.007876989431679249, + 0.005842284299433231, + 0.005152150988578796, + 0.008085806854069233, + 0.009452334605157375, + 0.007417693734169006, + 0.00709543377161026, + 0.005044862162321806, + 0.005088130943477154, + 0.013452616520226002, + 0.006701447069644928, + 0.005789970047771931, + 0.004130688030272722, + 0.003119942033663392, + 0.0026815414894372225, + 0.004266391973942518, + 0.0014918133383616805, + 0.003216603770852089, + 0.0020616059191524982, + 0.003955043386667967, + 0.005293741822242737, + 0.007980093359947205, + 0.004307389724999666, + 0.004860718734562397, + 0.003394553204998374, + 0.003765373257920146, + 0.0029228634666651487, + 0.00474925059825182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.002493250882253051 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glass filled with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deep, featureless black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round, white tablet dropped into the water and shown disintegrating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style high-contrast black and white with a sharp key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980908192694187 + ], + "opposite_logprob_tokens": [ + -3.52980899810791 + ], + "logprob_reward": -0.02980908192694187, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980908192694187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2562227249145508, + 0.18283908069133759, + 0.04464390128850937, + 0.07632599025964737, + 0.1472642421722412, + 0.049054861068725586, + 0.09660463035106659, + 0.07225725054740906, + 0.04467129707336426, + 0.030116068199276924 + ], + "credit_map": [ + 0.036193475127220154, + 0.022314105182886124, + 0.03663504496216774, + 0.02201058529317379, + 0.015429312363266945, + 0.03771265223622322, + 0.02561580017209053, + 0.060311753302812576, + 0.02644415572285652, + 0.010921383276581764, + 0.03695543110370636, + 0.022727331146597862, + 0.017932940274477005, + 0.02047852799296379, + 0.021574363112449646, + 0.025804942473769188, + 0.007334153167903423, + 0.0037025234196335077, + 0.0066336654126644135, + 0.007049955427646637, + 0.006324100308120251, + 0.004713412374258041, + 0.004602008033543825, + 0.004284081049263477, + 0.011891532689332962, + 0.014802646823227406, + 0.009305243380367756, + 0.006735453847795725, + 0.010082575492560863, + 0.007055397145450115, + 0.009703277610242367, + 0.006749862339347601, + 0.020716479048132896, + 0.009868227876722813, + 0.012263430282473564, + 0.02318364940583706, + 0.013951306231319904, + 0.04010212421417236, + 0.011941430158913136, + 0.015237588435411453, + 0.0069149867631495, + 0.007463615387678146, + 0.005561903119087219, + 0.0056680114939808846, + 0.004235929809510708, + 0.005730781238526106, + 0.006295547354966402, + 0.007184087298810482, + 0.0156451016664505, + 0.01083292718976736, + 0.01059132069349289, + 0.01607944816350937, + 0.009866411797702312, + 0.013528717681765556, + 0.00803296361118555, + 0.0120277414098382, + 0.011939208954572678, + 0.007842440158128738, + 0.005480663385242224, + 0.006196331232786179, + 0.004998588468879461, + 0.021384358406066895, + 0.005589035339653492, + 0.008826625533401966, + 0.01280602440237999, + 0.005415856372565031, + 0.0036001300904899836, + 0.0037805605679750443, + 0.0021311244927346706, + 0.004082432948052883, + 0.0023678727447986603, + 0.010487295687198639, + 0.004079955630004406, + 0.005776145029813051, + 0.0036956327967345715, + 0.004372918512672186, + 0.0025578979402780533, + 0.00417590094730258, + 0.0023645341861993074, + 0.0030930829234421253 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.02980908192694187 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glass filled with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deep, featureless black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round, white tablet dropped into the water and shown disintegrating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style high-contrast black and white with a sharp key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20154528319835663 + ], + "opposite_logprob_tokens": [ + -1.701545238494873 + ], + "logprob_reward": -0.20154528319835663, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20154528319835663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2879403233528137, + 0.11662887036800385, + 0.0591779500246048, + 0.11436649411916733, + 0.06120651960372925, + 0.08067813515663147, + 0.053098298609256744, + 0.05502632260322571, + 0.1475401222705841, + 0.024337032809853554 + ], + "credit_map": [ + 0.03952999413013458, + 0.02903604507446289, + 0.04140776768326759, + 0.030036984011530876, + 0.017708411440253258, + 0.0399923101067543, + 0.039538536220788956, + 0.05069024860858917, + 0.016361773014068604, + 0.00912260077893734, + 0.017937738448381424, + 0.022664792835712433, + 0.008841503411531448, + 0.009713045321404934, + 0.017901379615068436, + 0.014086039736866951, + 0.00855331216007471, + 0.005183557979762554, + 0.008265560492873192, + 0.012130928225815296, + 0.00820479542016983, + 0.005429337732493877, + 0.006496152840554714, + 0.004914307966828346, + 0.02659023180603981, + 0.015394484624266624, + 0.01681048981845379, + 0.011181123554706573, + 0.009121529757976532, + 0.013672919943928719, + 0.010982080362737179, + 0.010613632388412952, + 0.00849309004843235, + 0.007097747642546892, + 0.013260613195598125, + 0.004975572694092989, + 0.005969583056867123, + 0.004533696454018354, + 0.00559353269636631, + 0.01128268614411354, + 0.01534171774983406, + 0.012133401818573475, + 0.013910443522036076, + 0.009748372249305248, + 0.005420043598860502, + 0.00922660157084465, + 0.006947041023522615, + 0.007950508035719395, + 0.009680553339421749, + 0.005168510600924492, + 0.004637221805751324, + 0.008518200367689133, + 0.00540146604180336, + 0.008871541358530521, + 0.005035405978560448, + 0.005785400979220867, + 0.00898387935012579, + 0.012610998935997486, + 0.005341701675206423, + 0.004751978442072868, + 0.004080347251147032, + 0.00825593899935484, + 0.00627409340813756, + 0.004727385472506285, + 0.04736988991498947, + 0.013983129523694515, + 0.012632445432245731, + 0.00835871696472168, + 0.004424782004207373, + 0.013349990360438824, + 0.007164081558585167, + 0.04025707393884659, + 0.003584921592846513, + 0.004349824972450733, + 0.003402853850275278, + 0.0024586771614849567, + 0.0018526931526139379, + 0.0029628626070916653, + 0.002713366411626339, + 0.003011832246556878 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.20154528319835663 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot frames a simple, clear glass filled with water against a deep, featureless black background. A round, white, talc tablet is dropped into the water, and as it immerses itself, it explodes into a violent chemical reaction. The scene is rendered in dramatic, high-contrast black and white, with a sharp key light from the side sculpting every detail. A chaotic hiss of bright white microbubbles erupts violently from the tablet's surface, forming a turbulent, upward-rushing column that stirs the dark water. The tablet itself rapidly tumbles and disintegrates, releasing larger, shimmering bubbles that rush to the surface and burst. The visual is a powerful, contained storm, the stark white of the flaky chaos contrasting sharply with the inky black water, emphasizing the primal, kinetic energy of the dissolution.", + "seed": 14247, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_4/index_8850_seed_14247_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05902878739871085, + "vlm_logprob": -0.05902878739871085, + "vlm_yes_no_margin": 3.4749999761581423, + "vlm_token_credit": -0.05902878739871085, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glass filled with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deep, featureless black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round, white tablet dropped into the water and shown disintegrating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style high-contrast black and white with a sharp key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04222596064209938 + ], + "opposite_logprob_tokens": [ + -3.2922260761260986 + ], + "logprob_reward": -0.04222596064209938, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04222596064209938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34499475359916687, + 0.06369617581367493, + 0.11991412937641144, + 0.02709464728832245, + 0.0732693076133728, + 0.1325644701719284, + 0.04622301086783409, + 0.05723428726196289, + 0.09952712059020996, + 0.035482052713632584 + ], + "credit_map": [ + 0.05251038819551468, + 0.04002779722213745, + 0.044384561479091644, + 0.03222425654530525, + 0.026629790663719177, + 0.06148827075958252, + 0.04849134013056755, + 0.0392383448779583, + 0.01416495069861412, + 0.008000090718269348, + 0.010181250981986523, + 0.006487179547548294, + 0.0056343767791986465, + 0.006634205114096403, + 0.007372197695076466, + 0.005221927538514137, + 0.026739874854683876, + 0.010370886884629726, + 0.02011202462017536, + 0.011160134337842464, + 0.007001521065831184, + 0.008656558580696583, + 0.025609852746129036, + 0.01026328094303608, + 0.004585140850394964, + 0.0025593191385269165, + 0.004761811345815659, + 0.0028151152655482292, + 0.002513066167011857, + 0.0020496267825365067, + 0.005106956698000431, + 0.0027036115061491728, + 0.016762694343924522, + 0.006706740707159042, + 0.012598760426044464, + 0.005939006339758635, + 0.006443735212087631, + 0.004454961512237787, + 0.013787634670734406, + 0.006575775798410177, + 0.017843840643763542, + 0.018082929775118828, + 0.018474366515874863, + 0.02349746599793434, + 0.013507655821740627, + 0.01133600901812315, + 0.017290659248828888, + 0.01253154594451189, + 0.005942432675510645, + 0.009594559669494629, + 0.007080717943608761, + 0.004937973339110613, + 0.0037419379223138094, + 0.003675950923934579, + 0.006727742962539196, + 0.00452169356867671, + 0.009968946687877178, + 0.006467883475124836, + 0.010010276921093464, + 0.006907428149133921, + 0.004388509318232536, + 0.005461815744638443, + 0.007398385554552078, + 0.006631043739616871, + 0.014686591923236847, + 0.007375923451036215, + 0.0145036606118083, + 0.017665673047304153, + 0.009926803410053253, + 0.008504514582455158, + 0.014780727215111256, + 0.012083224952220917, + 0.00503813149407506, + 0.0034306426532566547, + 0.005745427217334509, + 0.0046177757903933525, + 0.002489052014425397, + 0.003748337971046567, + 0.005538823548704386, + 0.004873865284025669 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.04222596064209938 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glass filled with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deep, featureless black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round, white tablet dropped into the water and shown disintegrating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style high-contrast black and white with a sharp key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003594920039177 + ], + "opposite_logprob_tokens": [ + -2.3503594398498535 + ], + "logprob_reward": -0.1003594920039177, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003594920039177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2050529420375824, + 0.11610584706068039, + 0.1469356268644333, + 0.11669708788394928, + 0.0827157273888588, + 0.05980416387319565, + 0.04777897894382477, + 0.0687510147690773, + 0.08505786210298538, + 0.07110074162483215 + ], + "credit_map": [ + 0.03141387179493904, + 0.018407728523015976, + 0.02711699903011322, + 0.01693727821111679, + 0.013545091263949871, + 0.04589337483048439, + 0.026987213641405106, + 0.024751396849751472, + 0.030026497319340706, + 0.013340653851628304, + 0.019824570044875145, + 0.010347314178943634, + 0.008305632509291172, + 0.008691458031535149, + 0.014281099662184715, + 0.011288629844784737, + 0.026387695223093033, + 0.01758650317788124, + 0.02209257334470749, + 0.013392734341323376, + 0.012277839705348015, + 0.017942676320672035, + 0.022425226867198944, + 0.014830389060080051, + 0.040774814784526825, + 0.008274313993752003, + 0.02889746055006981, + 0.006497835740447044, + 0.006143661681562662, + 0.00579209765419364, + 0.014229469001293182, + 0.006087443791329861, + 0.014579505659639835, + 0.011400795541703701, + 0.013461781665682793, + 0.009276381693780422, + 0.007355729583650827, + 0.006371243856847286, + 0.00856485590338707, + 0.011705446988344193, + 0.009065361693501472, + 0.008869594894349575, + 0.009622863493859768, + 0.011019855737686157, + 0.005189350340515375, + 0.005036815069615841, + 0.004959477577358484, + 0.006040852051228285, + 0.007432040758430958, + 0.00856331828981638, + 0.00787273421883583, + 0.006772774737328291, + 0.004342608619481325, + 0.00316657149232924, + 0.005548845510929823, + 0.004080088809132576, + 0.009522228501737118, + 0.011288884095847607, + 0.010621710680425167, + 0.016576481983065605, + 0.00365620874799788, + 0.006532756146043539, + 0.0061142174527049065, + 0.004438531119376421, + 0.013742814771831036, + 0.006301390007138252, + 0.012450676411390305, + 0.01305277831852436, + 0.006885248702019453, + 0.00765733327716589, + 0.009339646436274052, + 0.01562797836959362, + 0.014364342205226421, + 0.007442566100507975, + 0.012004776857793331, + 0.0079342657700181, + 0.005402239039540291, + 0.006179932039231062, + 0.008565451949834824, + 0.009207173250615597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.1003594920039177 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glass filled with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deep, featureless black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round, white tablet dropped into the water and shown disintegrating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style high-contrast black and white with a sharp key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022052035201340914 + ], + "opposite_logprob_tokens": [ + -6.1272053718566895 + ], + "logprob_reward": -0.0022052035201340914, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022052035201340914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28900855779647827, + 0.09886138886213303, + 0.11306167393922806, + 0.044333964586257935, + 0.08853762596845627, + 0.09623494744300842, + 0.06963693350553513, + 0.06234893202781677, + 0.06166744604706764, + 0.07630859315395355 + ], + "credit_map": [ + 0.03903720900416374, + 0.021507004275918007, + 0.0346900150179863, + 0.03475514054298401, + 0.02812069095671177, + 0.03634217008948326, + 0.06150740757584572, + 0.033048905432224274, + 0.017271583899855614, + 0.00969563890248537, + 0.01375760417431593, + 0.01090957224369049, + 0.008579657413065434, + 0.007657782174646854, + 0.02237812802195549, + 0.008611424826085567, + 0.021349336951971054, + 0.01200601551681757, + 0.016268588602542877, + 0.012568848207592964, + 0.011070316657423973, + 0.01275221910327673, + 0.015337726101279259, + 0.01170862466096878, + 0.00783732533454895, + 0.004462209064513445, + 0.008141427300870419, + 0.005812736693769693, + 0.003914139233529568, + 0.004227886907756329, + 0.006439421325922012, + 0.0034988182596862316, + 0.01261636707931757, + 0.010091462172567844, + 0.012246399186551571, + 0.007399340160191059, + 0.009449336677789688, + 0.006486229132860899, + 0.018703406676650047, + 0.011545087210834026, + 0.012064244598150253, + 0.013968215323984623, + 0.013196106068789959, + 0.013273968361318111, + 0.010556093417108059, + 0.007173006422817707, + 0.01648966781795025, + 0.009513648226857185, + 0.009396785870194435, + 0.013171120546758175, + 0.011877411045134068, + 0.008372214622795582, + 0.00584186427295208, + 0.00525428494438529, + 0.008895191363990307, + 0.006828058045357466, + 0.01109743770211935, + 0.009176583960652351, + 0.010059175081551075, + 0.010614089667797089, + 0.0038348075468093157, + 0.00431813532486558, + 0.0056420788168907166, + 0.007606624625623226, + 0.008104376494884491, + 0.005276333075016737, + 0.009510213509202003, + 0.01152313593775034, + 0.00488741509616375, + 0.006446454674005508, + 0.007229242939502001, + 0.008690273389220238, + 0.012428770773112774, + 0.007601235993206501, + 0.013060661032795906, + 0.009518434293568134, + 0.005450292956084013, + 0.006529810838401318, + 0.008852987550199032, + 0.012866401113569736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0022052035201340914 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glass filled with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deep, featureless black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round, white tablet dropped into the water and shown disintegrating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style high-contrast black and white with a sharp key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0233065914362669 + ], + "opposite_logprob_tokens": [ + -3.7733066082000732 + ], + "logprob_reward": -0.0233065914362669, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0233065914362669, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16717983782291412, + 0.07915078103542328, + 0.06182493269443512, + 0.046995941549539566, + 0.12385831773281097, + 0.10204733163118362, + 0.10635004192590714, + 0.08811676502227783, + 0.08906899392604828, + 0.13540701568126678 + ], + "credit_map": [ + 0.02503071539103985, + 0.01791561394929886, + 0.02381097711622715, + 0.012739659287035465, + 0.017101654782891273, + 0.02854340896010399, + 0.020932545885443687, + 0.021105283871293068, + 0.013675106689333916, + 0.009641243144869804, + 0.011590240523219109, + 0.007693685591220856, + 0.006962875835597515, + 0.006691203452646732, + 0.013950963504612446, + 0.008945471607148647, + 0.010381050407886505, + 0.007923166267573833, + 0.009336515329778194, + 0.00797234009951353, + 0.0060679237358272076, + 0.0055452133528888226, + 0.00905455183237791, + 0.005544179584830999, + 0.008057004772126675, + 0.003764564171433449, + 0.007165994960814714, + 0.0061617037281394005, + 0.004868113435804844, + 0.0034719263203442097, + 0.009805758483707905, + 0.003700879169628024, + 0.017459532245993614, + 0.0092149144038558, + 0.018378620967268944, + 0.015813911333680153, + 0.009322598576545715, + 0.008530369959771633, + 0.024359410628676414, + 0.020778978243470192, + 0.013710458762943745, + 0.016298165544867516, + 0.013443023897707462, + 0.012557277455925941, + 0.009175479412078857, + 0.006238725036382675, + 0.021074289456009865, + 0.009549916721880436, + 0.015455407090485096, + 0.020968405529856682, + 0.0173640139400959, + 0.012173552066087723, + 0.009841530583798885, + 0.0071238926611840725, + 0.014992578886449337, + 0.008430675603449345, + 0.0147715387865901, + 0.012903863564133644, + 0.01341179572045803, + 0.013472761027514935, + 0.005736508872359991, + 0.010526571422815323, + 0.005883580539375544, + 0.01141014602035284, + 0.01418632548302412, + 0.006204719189554453, + 0.013703973963856697, + 0.01529793068766594, + 0.007499160245060921, + 0.008398163132369518, + 0.010344823822379112, + 0.013433909974992275, + 0.024880951270461082, + 0.015208104625344276, + 0.0223627258092165, + 0.017064319923520088, + 0.01015832182019949, + 0.009919815696775913, + 0.01370457373559475, + 0.02210821397602558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.0233065914362669 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glass filled with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deep, featureless black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round, white tablet dropped into the water and shown disintegrating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style high-contrast black and white with a sharp key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704668939113617 + ], + "opposite_logprob_tokens": [ + -2.127046585083008 + ], + "logprob_reward": -0.12704668939113617, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12704668939113617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1919070929288864, + 0.1074753925204277, + 0.11849702149629593, + 0.033532995730638504, + 0.05774514377117157, + 0.05985315889120102, + 0.07725002616643906, + 0.14826780557632446, + 0.1529221385717392, + 0.05254920572042465 + ], + "credit_map": [ + 0.02450479008257389, + 0.02260447107255459, + 0.02546207420527935, + 0.012735618278384209, + 0.01609564572572708, + 0.04009503126144409, + 0.021555360406637192, + 0.02885410562157631, + 0.014598090201616287, + 0.010436738841235638, + 0.013928335160017014, + 0.013161535374820232, + 0.010487254709005356, + 0.008302099071443081, + 0.026225563138723373, + 0.010335774160921574, + 0.01922811195254326, + 0.014783890917897224, + 0.01675890013575554, + 0.013391045853495598, + 0.011950645595788956, + 0.01153520680963993, + 0.018430080264806747, + 0.012419145554304123, + 0.006062252912670374, + 0.002719891956076026, + 0.0051047164015471935, + 0.0041256011463701725, + 0.0038174709770828485, + 0.00350197940133512, + 0.005740838591009378, + 0.0024602448102086782, + 0.009112734347581863, + 0.0064770132303237915, + 0.008550208993256092, + 0.007114847656339407, + 0.004431883338838816, + 0.005350784864276648, + 0.007080520037561655, + 0.009627153165638447, + 0.007464007940143347, + 0.009213627316057682, + 0.007981796748936176, + 0.008643049746751785, + 0.0067985886707901955, + 0.004284485708922148, + 0.009611532092094421, + 0.00585607485845685, + 0.011851491406559944, + 0.014472097158432007, + 0.015636438503861427, + 0.007141246926039457, + 0.006028412375599146, + 0.005141865462064743, + 0.00951319932937622, + 0.007465272210538387, + 0.02278600074350834, + 0.021663429215550423, + 0.023637384176254272, + 0.03542398288846016, + 0.00847258698195219, + 0.013724523596465588, + 0.011847632005810738, + 0.010712271556258202, + 0.023496773093938828, + 0.014588582329452038, + 0.02211066521704197, + 0.029190437868237495, + 0.012975751422345638, + 0.014844660647213459, + 0.015533854253590107, + 0.020181424915790558, + 0.010486386716365814, + 0.006489704363048077, + 0.007150860503315926, + 0.005687016993761063, + 0.005966337397694588, + 0.003450770629569888, + 0.007512860000133514, + 0.005805269815027714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.12704668939113617 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot frames a simple, clear glass filled with water against a deep, featureless black background. A round, white, talc tablet is dropped into the water, and as it immerses itself, it explodes into a violent chemical reaction. The scene is rendered in dramatic, high-contrast black and white, with a sharp key light from the side sculpting every detail. A chaotic hiss of bright white microbubbles erupts violently from the tablet's surface, forming a turbulent, upward-rushing column that stirs the dark water. The tablet itself rapidly tumbles and disintegrates, releasing larger, shimmering bubbles that rush to the surface and burst. The visual is a powerful, contained storm, the stark white of the flaky chaos contrasting sharply with the inky black water, emphasizing the primal, kinetic energy of the dissolution.", + "seed": 14248, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_4/index_8850_seed_14248_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07403625631704927, + "vlm_logprob": -0.07403625631704927, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.07403625631704927, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glass filled with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deep, featureless black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round, white tablet dropped into the water and shown disintegrating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style high-contrast black and white with a sharp key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03625105321407318 + ], + "opposite_logprob_tokens": [ + -3.4112510681152344 + ], + "logprob_reward": -0.03625105321407318, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03625105321407318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20409046113491058, + 0.06363383680582047, + 0.13806815445423126, + 0.04132193699479103, + 0.048691753298044205, + 0.06217677891254425, + 0.13131995499134064, + 0.07038095593452454, + 0.15445375442504883, + 0.0858624130487442 + ], + "credit_map": [ + 0.02879508212208748, + 0.01904844120144844, + 0.03496123477816582, + 0.02693243883550167, + 0.016399148851633072, + 0.022564420476555824, + 0.025295894593000412, + 0.030093781650066376, + 0.008460336364805698, + 0.0053121293894946575, + 0.0052147069945931435, + 0.007177577819675207, + 0.005238570272922516, + 0.008241072297096252, + 0.01065122801810503, + 0.013338212855160236, + 0.01130601018667221, + 0.012647919356822968, + 0.009922712109982967, + 0.02571513131260872, + 0.013420735485851765, + 0.028811264783143997, + 0.024066079407930374, + 0.012178287841379642, + 0.004184429068118334, + 0.0046455590054392815, + 0.006248714402318001, + 0.00561533635482192, + 0.005300484597682953, + 0.005332821514457464, + 0.004864688031375408, + 0.00512990215793252, + 0.004250285215675831, + 0.003856916446238756, + 0.004301086999475956, + 0.004462359473109245, + 0.004902887623757124, + 0.012529812753200531, + 0.004567788913846016, + 0.009820609353482723, + 0.006367740221321583, + 0.011182858608663082, + 0.004041412845253944, + 0.004657199140638113, + 0.01698405295610428, + 0.006372425705194473, + 0.006461434531956911, + 0.006109650246798992, + 0.025361740961670876, + 0.014253795146942139, + 0.010262240655720234, + 0.01118842326104641, + 0.019559061154723167, + 0.01516983937472105, + 0.020391667261719704, + 0.015133184380829334, + 0.008973393589258194, + 0.0038917188066989183, + 0.0066132149659097195, + 0.01239769533276558, + 0.00369674782268703, + 0.008865980431437492, + 0.006676646415144205, + 0.019265560433268547, + 0.017954999580979347, + 0.01201880443841219, + 0.016098883002996445, + 0.0325322300195694, + 0.014407938346266747, + 0.018335727974772453, + 0.025735732167959213, + 0.01736942119896412, + 0.009463171474635601, + 0.008707785978913307, + 0.01567923091351986, + 0.008735417388379574, + 0.007189392577856779, + 0.013744686730206013, + 0.00991814024746418, + 0.012424581684172153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.03625105321407318 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glass filled with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deep, featureless black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round, white tablet dropped into the water and shown disintegrating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style high-contrast black and white with a sharp key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06208738684654236 + ], + "opposite_logprob_tokens": [ + -2.812087297439575 + ], + "logprob_reward": -0.06208738684654236, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06208738684654236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08206681907176971, + 0.35592323541641235, + 0.13166403770446777, + 0.02901514247059822, + 0.0312734954059124, + 0.0815604031085968, + 0.05299077555537224, + 0.12067214399576187, + 0.06225551292300224, + 0.05257846787571907 + ], + "credit_map": [ + 0.008689514361321926, + 0.005364695563912392, + 0.016889484599232674, + 0.009841732680797577, + 0.007534915581345558, + 0.011944450438022614, + 0.011414945125579834, + 0.010387079790234566, + 0.05318764969706535, + 0.029534611850976944, + 0.03576890751719475, + 0.027602681890130043, + 0.021148599684238434, + 0.06300612539052963, + 0.03793570399284363, + 0.08773895353078842, + 0.014511686749756336, + 0.018495237454771996, + 0.00920132640749216, + 0.018965646624565125, + 0.012612075544893742, + 0.025673676282167435, + 0.012589657679200172, + 0.019614730030298233, + 0.0033612616825848818, + 0.0030351649038493633, + 0.003872910048812628, + 0.004974768031388521, + 0.002723757643252611, + 0.0031295691151171923, + 0.004912903066724539, + 0.0030048079788684845, + 0.0019642335828393698, + 0.0011006873100996017, + 0.0012394095538184047, + 0.0017304723151028156, + 0.0043226126581430435, + 0.009289953857660294, + 0.0032122163102030754, + 0.008413909003138542, + 0.009503318928182125, + 0.02331485226750374, + 0.002813922241330147, + 0.004468854516744614, + 0.021694110706448555, + 0.006714941933751106, + 0.008073673583567142, + 0.0049767266027629375, + 0.007404920645058155, + 0.008389952592551708, + 0.004286595620214939, + 0.00534512335434556, + 0.0064986650831997395, + 0.005705702118575573, + 0.009051455184817314, + 0.0063083590939641, + 0.012131672352552414, + 0.00922523532062769, + 0.006568137090653181, + 0.017453545704483986, + 0.008320523425936699, + 0.02535199001431465, + 0.006092149298638105, + 0.035528890788555145, + 0.007168573793023825, + 0.006046436261385679, + 0.008907336741685867, + 0.01140511129051447, + 0.004658435937017202, + 0.009099840186536312, + 0.008128265850245953, + 0.006841509137302637, + 0.005027088336646557, + 0.0036826017312705517, + 0.012769784778356552, + 0.00650021992623806, + 0.00323483906686306, + 0.005047043785452843, + 0.009506377391517162, + 0.006810514722019434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 7 + ], + "selected_reward": -0.06208738684654236 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glass filled with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deep, featureless black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round, white tablet dropped into the water and shown disintegrating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style high-contrast black and white with a sharp key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00319886626675725 + ], + "opposite_logprob_tokens": [ + -5.753199100494385 + ], + "logprob_reward": -0.00319886626675725, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00319886626675725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17023901641368866, + 0.14148183166980743, + 0.08454571664333344, + 0.018402215093374252, + 0.03406691923737526, + 0.14164921641349792, + 0.10338299721479416, + 0.076851025223732, + 0.05968248099088669, + 0.169698566198349 + ], + "credit_map": [ + 0.018038177862763405, + 0.01364502590149641, + 0.03272365406155586, + 0.02292906679213047, + 0.015119539573788643, + 0.018450874835252762, + 0.025127068161964417, + 0.024205615743994713, + 0.019874602556228638, + 0.012863370589911938, + 0.014477740041911602, + 0.012173282913863659, + 0.008409001864492893, + 0.02335388958454132, + 0.017206581309437752, + 0.033123359084129333, + 0.008850837126374245, + 0.011619100347161293, + 0.008619291707873344, + 0.011511706747114658, + 0.012389350682497025, + 0.010080150328576565, + 0.013620632700622082, + 0.007854645140469074, + 0.002358959522098303, + 0.0020042036194354296, + 0.003087342018261552, + 0.0021598441526293755, + 0.0022158471401780844, + 0.0022036470472812653, + 0.0019835527054965496, + 0.0023888200521469116, + 0.0025582294911146164, + 0.00201858370564878, + 0.0025038686580955982, + 0.0024424660950899124, + 0.004016433842480183, + 0.010013295337557793, + 0.0030610165558755398, + 0.00745302764698863, + 0.016245102509856224, + 0.03885910287499428, + 0.00427690614014864, + 0.009475795552134514, + 0.041481271386146545, + 0.00808083638548851, + 0.015157787129282951, + 0.00807242002338171, + 0.015783242881298065, + 0.009651816450059414, + 0.006172169931232929, + 0.010006595402956009, + 0.016007503494620323, + 0.01058940589427948, + 0.020939519628882408, + 0.014232741668820381, + 0.008172783069312572, + 0.004545923788100481, + 0.003610062412917614, + 0.013610300607979298, + 0.0046803453005850315, + 0.014537186361849308, + 0.0045425985008478165, + 0.023151831701397896, + 0.005957962945103645, + 0.006954624317586422, + 0.007636927533894777, + 0.012526883743703365, + 0.0038481131196022034, + 0.005995096173137426, + 0.011220668442547321, + 0.005542202852666378, + 0.01825893670320511, + 0.014315944164991379, + 0.021400390192866325, + 0.029291534796357155, + 0.01040397584438324, + 0.01836305297911167, + 0.03499307483434677, + 0.0226716548204422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.00319886626675725 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glass filled with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deep, featureless black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round, white tablet dropped into the water and shown disintegrating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style high-contrast black and white with a sharp key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308463633060455 + ], + "opposite_logprob_tokens": [ + -3.1680846214294434 + ], + "logprob_reward": -0.04308463633060455, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308463633060455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15606136620044708, + 0.11859104782342911, + 0.030149366706609726, + 0.016206860542297363, + 0.0771024078130722, + 0.06313861906528473, + 0.07199109345674515, + 0.2726300358772278, + 0.02303539216518402, + 0.1710938662290573 + ], + "credit_map": [ + 0.017342938110232353, + 0.010914149694144726, + 0.030572528019547462, + 0.018928710371255875, + 0.012088839896023273, + 0.021020451560616493, + 0.0259094201028347, + 0.019284315407276154, + 0.013661015778779984, + 0.013604989275336266, + 0.011884349398314953, + 0.012346026487648487, + 0.008553420193493366, + 0.017095210030674934, + 0.01861541159451008, + 0.022830618545413017, + 0.003168713301420212, + 0.00460249837487936, + 0.0025202948600053787, + 0.0036835691425949335, + 0.004101309459656477, + 0.004202880430966616, + 0.004663627129048109, + 0.003206470748409629, + 0.001675753970630467, + 0.0015969931846484542, + 0.0016511697322130203, + 0.0025680982507765293, + 0.001489434391260147, + 0.0020975174847990274, + 0.0030564654152840376, + 0.0020714281126856804, + 0.004591061733663082, + 0.002251607831567526, + 0.002149636158719659, + 0.002950243651866913, + 0.009327623061835766, + 0.02656658925116062, + 0.005168514791876078, + 0.024097127839922905, + 0.006196198984980583, + 0.01233513094484806, + 0.005724267102777958, + 0.0060228342190384865, + 0.006981716025620699, + 0.005525325424969196, + 0.0057107494212687016, + 0.01464239601045847, + 0.009870684705674648, + 0.007482511457055807, + 0.011698269285261631, + 0.010932930745184422, + 0.005080971401184797, + 0.006752071436494589, + 0.012813865207135677, + 0.007359786424785852, + 0.030064910650253296, + 0.016263049095869064, + 0.014271516352891922, + 0.04080832004547119, + 0.017977938055992126, + 0.05408444628119469, + 0.016962574794888496, + 0.08219727873802185, + 0.0032958120573312044, + 0.0017635647673159838, + 0.0031636205967515707, + 0.004596773535013199, + 0.0017113848589360714, + 0.0030356766656041145, + 0.00284024141728878, + 0.002628317568451166, + 0.017976824194192886, + 0.013611790724098682, + 0.022037150338292122, + 0.029614098370075226, + 0.00941457785665989, + 0.01568596437573433, + 0.03981349989771843, + 0.02293993905186653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.04308463633060455 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glass filled with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deep, featureless black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round, white tablet dropped into the water and shown disintegrating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style high-contrast black and white with a sharp key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22555933892726898 + ], + "opposite_logprob_tokens": [ + -1.6005593538284302 + ], + "logprob_reward": -0.22555933892726898, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22555933892726898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26278239488601685, + 0.13603374361991882, + 0.06513883918523788, + 0.02815084531903267, + 0.04518457502126694, + 0.20553816854953766, + 0.062324490398168564, + 0.08132259547710419, + 0.045488763600587845, + 0.06803558021783829 + ], + "credit_map": [ + 0.028269724920392036, + 0.017521319910883904, + 0.05751613900065422, + 0.0287953931838274, + 0.02719760127365589, + 0.035734377801418304, + 0.034737177193164825, + 0.033010657876729965, + 0.0164899080991745, + 0.014723972417414188, + 0.015279731713235378, + 0.013283809646964073, + 0.010273965075612068, + 0.021817315369844437, + 0.019691361114382744, + 0.024473683908581734, + 0.003854662412777543, + 0.006739548873156309, + 0.00417065154761076, + 0.015205956064164639, + 0.006311416160315275, + 0.011610216461122036, + 0.01124047115445137, + 0.006005911622196436, + 0.0034033150877803564, + 0.002953152870759368, + 0.002867605770006776, + 0.003304183017462492, + 0.002297792350873351, + 0.004993898794054985, + 0.0036801262758672237, + 0.004650771152228117, + 0.0034899732563644648, + 0.00220108381472528, + 0.0026445447001606226, + 0.0023718569427728653, + 0.005815317388623953, + 0.01314719207584858, + 0.004313682205975056, + 0.011200923472642899, + 0.024466216564178467, + 0.036556098610162735, + 0.0055528623051941395, + 0.018491683527827263, + 0.0681343823671341, + 0.011353439651429653, + 0.02495119720697403, + 0.01603228785097599, + 0.009962520562112331, + 0.0060192556120455265, + 0.005166211165487766, + 0.006123314145952463, + 0.008458409458398819, + 0.007144988980144262, + 0.009863835759460926, + 0.009585953317582607, + 0.008323093876242638, + 0.005758653860539198, + 0.007040148600935936, + 0.016655681654810905, + 0.005346936639398336, + 0.011817729100584984, + 0.007341418880969286, + 0.019038932397961617, + 0.004377116449177265, + 0.00368320569396019, + 0.004995526745915413, + 0.009595847688615322, + 0.0050187185406684875, + 0.005626247730106115, + 0.007462804671376944, + 0.004729297012090683, + 0.006724582053720951, + 0.005174665246158838, + 0.009672585874795914, + 0.009658949449658394, + 0.0054127867333590984, + 0.008434590883553028, + 0.014225807972252369, + 0.008731613866984844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.22555933892726898 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_5/index_tensor([8967, 8967])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_5/index_tensor([8967, 8967])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1ff987de6010c6362109bac672247b1d7afa4c99 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_5/index_tensor([8967, 8967])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "tight, almost motionless shot, shot from a slightly upward angle, locks the viewer's gaze firmly on a boundless, rolling sea of clouds. A resolute Eastern warrior stands barefoot on the tip of a slender, emerald green bamboo pole. His long, jet-black hair is tied into a high bun with a simple cloth belt. A few strands, blown by the sharp wind from high above, gently brush against his angular features. He wears a loose, washed-bleached linen warrior's uniform, its sleeves rustling in the wind. His core remains rock-solid, swaying only subtly, imperceptibly, against the force of the swaying bamboo pole. The camera meticulously captures his deep, unwavering eyes, revealing a calmness and focus beyond his years, as if at one with the universe. Beneath his feet, the tip of the bamboo pole, bearing his entire weight, vibrated visibly and at high frequency due to the combined forces of wind and gravity, creating a stark contrast to the absolute stability of his upper body. The first rays of golden morning sunlight pierced the clouds, casting a warm glow on his bronzed skin and capturing every detail of his tense toes gripping the slippery pole. The entire scene, shrouded in a vast sense of loneliness, was imbued with an ultimate, restrained, and powerful tension, creating an otherworldly martial arts realm.", + "seed": 25113, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_5/index_8967_seed_25113_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02376009160652757, + "vlm_logprob": -0.02376009160652757, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.02376009160652757, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed tightly with a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Eastern warrior standing barefoot on the tip of a slender emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior have long jet-black hair tied in a high bun with a simple cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior wearing a loose washed-bleached linen uniform while standing on a rolling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the first rays of golden morning sunlight piercing through the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029220648109912872 + ], + "opposite_logprob_tokens": [ + -4.7792205810546875 + ], + "logprob_reward": -0.029220648109912872, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.029220648109912872, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03247180953621864, + 0.019411388784646988, + 0.34302395582199097, + 0.021779565140604973, + 0.015530344098806381, + 0.045525163412094116, + 0.43687474727630615, + 0.017641335725784302, + 0.03288919851183891, + 0.034852467477321625 + ], + "credit_map": [ + 0.0051766871474683285, + 0.003469115123152733, + 0.004831248894333839, + 0.003297971561551094, + 0.00408692192286253, + 0.004887372255325317, + 0.0031275234650820494, + 0.0035949687007814646, + 0.00442899763584137, + 0.0019496552413329482, + 0.0037710873875766993, + 0.001369291334412992, + 0.0012160500045865774, + 0.0027686324901878834, + 0.0009966110810637474, + 0.002911064075306058, + 0.10979855060577393, + 0.04435106739401817, + 0.02749772183597088, + 0.008588564582169056, + 0.01402815617620945, + 0.03599923104047775, + 0.01686338149011135, + 0.08589724451303482, + 0.002976665971800685, + 0.0017532330239191651, + 0.004718025214970112, + 0.001974476035684347, + 0.0025333785451948643, + 0.0033587568905204535, + 0.001960220979526639, + 0.0025048081297427416, + 0.002609990304335952, + 0.0014601602451875806, + 0.003546475199982524, + 0.0008300283225253224, + 0.0011103543220087886, + 0.00411316379904747, + 0.0005169857759028673, + 0.0013431863626465201, + 0.012146220542490482, + 0.005288493819534779, + 0.012039241380989552, + 0.002216719090938568, + 0.0031284443102777004, + 0.004552476108074188, + 0.0025585382245481014, + 0.003595030168071389, + 0.13240131735801697, + 0.03825432434678078, + 0.15713860094547272, + 0.01597907952964306, + 0.011871587485074997, + 0.02123568020761013, + 0.01283322274684906, + 0.04716096818447113, + 0.004450482316315174, + 0.0021551158279180527, + 0.0019665765576064587, + 0.001435140147805214, + 0.001295286463573575, + 0.003937965724617243, + 0.0013116102200001478, + 0.0010891592828556895, + 0.0072141229175031185, + 0.00353516754694283, + 0.0031092537101358175, + 0.0016157639911398292, + 0.0013819485902786255, + 0.012070221826434135, + 0.001056450535543263, + 0.002906271256506443, + 0.00986982136964798, + 0.006150674074888229, + 0.0031010203529149294, + 0.0015212384751066566, + 0.0014180063735693693, + 0.0032518289517611265, + 0.0018376103835180402, + 0.007702267728745937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 5 + ], + "selected_reward": -0.029220648109912872 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed tightly with a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Eastern warrior standing barefoot on the tip of a slender emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior have long jet-black hair tied in a high bun with a simple cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior wearing a loose washed-bleached linen uniform while standing on a rolling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the first rays of golden morning sunlight piercing through the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430884063243866 + ], + "opposite_logprob_tokens": [ + -3.168088436126709 + ], + "logprob_reward": -0.0430884063243866, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430884063243866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0569508820772171, + 0.045460283756256104, + 0.11528323590755463, + 0.03556140884757042, + 0.02222561277449131, + 0.03496389836072922, + 0.4381701648235321, + 0.17525430023670197, + 0.03171470761299133, + 0.04441551864147186 + ], + "credit_map": [ + 0.010688702575862408, + 0.0050289807841181755, + 0.010740318335592747, + 0.007123768329620361, + 0.00438184617087245, + 0.007847603410482407, + 0.004362940788269043, + 0.006776722148060799, + 0.011206201277673244, + 0.004849096294492483, + 0.008032377809286118, + 0.003134107915684581, + 0.002590952441096306, + 0.0062974258325994015, + 0.0023404823150485754, + 0.0070096394047141075, + 0.03623981401324272, + 0.014435806311666965, + 0.011067445389926434, + 0.003448148025199771, + 0.005409935023635626, + 0.012132645584642887, + 0.005480228923261166, + 0.027069218456745148, + 0.003717961022630334, + 0.003935015294700861, + 0.008977781049907207, + 0.003181102219969034, + 0.0032245097681879997, + 0.00677529675886035, + 0.002270352328196168, + 0.0034793918021023273, + 0.006320162210613489, + 0.0023816123139113188, + 0.0024311142042279243, + 0.0011490086326375604, + 0.0011042006080970168, + 0.005546968430280685, + 0.001453476957976818, + 0.0018390691839158535, + 0.0028531611897051334, + 0.004244351293891668, + 0.008314015343785286, + 0.004213516600430012, + 0.0032782997004687786, + 0.005687564145773649, + 0.0028040858451277018, + 0.0035689063370227814, + 0.1287255883216858, + 0.039258379489183426, + 0.15650403499603271, + 0.01573040895164013, + 0.011661847122013569, + 0.02437436208128929, + 0.014972207136452198, + 0.04694332182407379, + 0.049361806362867355, + 0.015272427350282669, + 0.013807041570544243, + 0.0045324633829295635, + 0.008611762896180153, + 0.07079645991325378, + 0.006855132523924112, + 0.0060171931982040405, + 0.006433124653995037, + 0.004672711715102196, + 0.00495564891025424, + 0.0033112738747149706, + 0.002273038262501359, + 0.004177235532552004, + 0.002428918844088912, + 0.0034627593122422695, + 0.006027416791766882, + 0.009527323767542839, + 0.003797239391133189, + 0.002204712014645338, + 0.0033751896116882563, + 0.0032100132666528225, + 0.002441712422296405, + 0.01383191253989935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 2 + ], + "selected_reward": -0.0430884063243866 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed tightly with a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Eastern warrior standing barefoot on the tip of a slender emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior have long jet-black hair tied in a high bun with a simple cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior wearing a loose washed-bleached linen uniform while standing on a rolling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the first rays of golden morning sunlight piercing through the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026333538815379143 + ], + "opposite_logprob_tokens": [ + -3.6513335704803467 + ], + "logprob_reward": -0.026333538815379143, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026333538815379143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06966280937194824, + 0.032436612993478775, + 0.14077022671699524, + 0.02345053106546402, + 0.027285784482955933, + 0.0391254685819149, + 0.4530581831932068, + 0.08292253315448761, + 0.11147419363260269, + 0.019813677296042442 + ], + "credit_map": [ + 0.007199966348707676, + 0.006850041914731264, + 0.012603316456079483, + 0.0075796074233949184, + 0.00997776910662651, + 0.011283881962299347, + 0.005684337113052607, + 0.008483890444040298, + 0.008316473104059696, + 0.00327997375279665, + 0.005679013207554817, + 0.0027594072744250298, + 0.0021085762418806553, + 0.0040172007866203785, + 0.002372085116803646, + 0.0039038865361362696, + 0.04444512352347374, + 0.017939915880560875, + 0.011321119964122772, + 0.0035883388482034206, + 0.005813462659716606, + 0.015334157273173332, + 0.006919257342815399, + 0.035408858209848404, + 0.003130497643724084, + 0.002639691112563014, + 0.004142878111451864, + 0.0020841602236032486, + 0.0027784453704953194, + 0.0041650510393083096, + 0.0023870861623436213, + 0.002122720470651984, + 0.003057606052607298, + 0.0024268757551908493, + 0.006934361066669226, + 0.00377585063688457, + 0.002570416545495391, + 0.0033416030928492546, + 0.0026681579183787107, + 0.002510914346203208, + 0.008950240910053253, + 0.006459056865423918, + 0.005095077678561211, + 0.0030647108796983957, + 0.003289386397227645, + 0.006858577486127615, + 0.0025929389521479607, + 0.0028154789470136166, + 0.13938361406326294, + 0.04068824276328087, + 0.16445845365524292, + 0.015579981729388237, + 0.011819581501185894, + 0.019149990752339363, + 0.013168737292289734, + 0.04880959540605545, + 0.020648719742894173, + 0.008116901852190495, + 0.007341606076806784, + 0.003033042885363102, + 0.005545764230191708, + 0.030726520344614983, + 0.003660994116216898, + 0.003848977154120803, + 0.023619618266820908, + 0.01001435425132513, + 0.01206961926072836, + 0.005963941570371389, + 0.00657094269990921, + 0.04072418808937073, + 0.004498987924307585, + 0.008012545295059681, + 0.0021373152267187834, + 0.003594764741137624, + 0.0031852335669100285, + 0.0013895784504711628, + 0.0014485192950814962, + 0.0020958425011485815, + 0.0015203231014311314, + 0.0044420999474823475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 8 + ], + "selected_reward": -0.026333538815379143 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed tightly with a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Eastern warrior standing barefoot on the tip of a slender emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior have long jet-black hair tied in a high bun with a simple cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior wearing a loose washed-bleached linen uniform while standing on a rolling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the first rays of golden morning sunlight piercing through the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040937201119959354 + ], + "opposite_logprob_tokens": [ + -5.504093647003174 + ], + "logprob_reward": -0.0040937201119959354, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040937201119959354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07656912505626678, + 0.05960817635059357, + 0.23122285306453705, + 0.07838229089975357, + 0.03879053518176079, + 0.10277403891086578, + 0.17356158792972565, + 0.08778080344200134, + 0.08177962899208069, + 0.06953096389770508 + ], + "credit_map": [ + 0.019295217469334602, + 0.008025217801332474, + 0.015471856109797955, + 0.008155560120940208, + 0.004181259777396917, + 0.008909943513572216, + 0.004881481174379587, + 0.007648590486496687, + 0.017034059390425682, + 0.00634565157815814, + 0.010900205932557583, + 0.0038382690399885178, + 0.0027323393151164055, + 0.0072638350538909435, + 0.002912248717620969, + 0.008581569418311119, + 0.07318957149982452, + 0.029492482542991638, + 0.019483555108308792, + 0.00635369960218668, + 0.010552763938903809, + 0.02363933064043522, + 0.011314108036458492, + 0.05719734728336334, + 0.008585760369896889, + 0.008238100446760654, + 0.01759887859225273, + 0.005321064498275518, + 0.006451806053519249, + 0.017764508724212646, + 0.003721104934811592, + 0.010701068677008152, + 0.010426950640976429, + 0.0034516809973865747, + 0.004781925585120916, + 0.0018677138723433018, + 0.0017954822396859527, + 0.011711496859788895, + 0.0014743461506441236, + 0.003280939534306526, + 0.027297476306557655, + 0.010791147127747536, + 0.03365948051214218, + 0.003737723920494318, + 0.0035174244549125433, + 0.012007161974906921, + 0.002482678974047303, + 0.009280955418944359, + 0.050250258296728134, + 0.01508786715567112, + 0.062361348420381546, + 0.005765240639448166, + 0.005162164568901062, + 0.011489716358482838, + 0.005288104061037302, + 0.01815688982605934, + 0.024272730574011803, + 0.00852341577410698, + 0.005683625116944313, + 0.0026514511555433273, + 0.005865694023668766, + 0.03372829779982567, + 0.0036797355860471725, + 0.003375849686563015, + 0.016126325353980064, + 0.008511419408023357, + 0.010140281170606613, + 0.0032869214192032814, + 0.0028414896223694086, + 0.03199981153011322, + 0.0028327363543212414, + 0.006040643434971571, + 0.01661829650402069, + 0.009514800272881985, + 0.006877202074974775, + 0.002451114822179079, + 0.0023270361125469208, + 0.005200221203267574, + 0.0028951966669410467, + 0.02364709973335266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 5 + ], + "selected_reward": -0.0040937201119959354 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed tightly with a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Eastern warrior standing barefoot on the tip of a slender emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior have long jet-black hair tied in a high bun with a simple cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior wearing a loose washed-bleached linen uniform while standing on a rolling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the first rays of golden morning sunlight piercing through the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016064144670963287 + ], + "opposite_logprob_tokens": [ + -4.141064167022705 + ], + "logprob_reward": -0.016064144670963287, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016064144670963287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05195613577961922, + 0.01113045122474432, + 0.34375259280204773, + 0.02920328639447689, + 0.26694998145103455, + 0.03405841812491417, + 0.10789190977811813, + 0.0830957368016243, + 0.03799498826265335, + 0.03396639600396156 + ], + "credit_map": [ + 0.005154602229595184, + 0.004923313390463591, + 0.011189116165041924, + 0.0055261882953345776, + 0.005026804283261299, + 0.00968187302350998, + 0.004177335184067488, + 0.006276904139667749, + 0.001096126507036388, + 0.0014204184990376234, + 0.002222376875579357, + 0.0010906922398135066, + 0.0012572850100696087, + 0.0014491970650851727, + 0.0012937881983816624, + 0.0013005681103095412, + 0.10908056795597076, + 0.04468844085931778, + 0.028074495494365692, + 0.00850208941847086, + 0.01345616765320301, + 0.03616192564368248, + 0.016939710825681686, + 0.08684924244880676, + 0.0037377469707280397, + 0.002511108061298728, + 0.00751328794285655, + 0.0017798100598156452, + 0.0024819921236485243, + 0.00645506801083684, + 0.0016479676123708487, + 0.0030763098038733006, + 0.07525932788848877, + 0.023970330134034157, + 0.03510098159313202, + 0.00965642835944891, + 0.007781419437378645, + 0.08582081645727158, + 0.008011152036488056, + 0.021349547430872917, + 0.005334897432476282, + 0.002262476831674576, + 0.010706901550292969, + 0.0020735799334943295, + 0.0026299734599888325, + 0.0062499865889549255, + 0.0015948767540976405, + 0.0032057282514870167, + 0.03221307322382927, + 0.009046446532011032, + 0.03952523693442345, + 0.003778024110943079, + 0.002946029417216778, + 0.005191563628613949, + 0.003361346898600459, + 0.011830201372504234, + 0.021971317008137703, + 0.007575582712888718, + 0.008717525750398636, + 0.0026660531293600798, + 0.003300201613456011, + 0.03306103125214577, + 0.002723749727010727, + 0.003080283524468541, + 0.009143597446382046, + 0.005139497108757496, + 0.004116704221814871, + 0.0029808853287249804, + 0.002304157242178917, + 0.009667177684605122, + 0.002222007606178522, + 0.002420966047793627, + 0.010252944193780422, + 0.01252553891390562, + 0.002857656916603446, + 0.0011629888322204351, + 0.0015480566071346402, + 0.00288063264451921, + 0.0011661640601232648, + 0.0015724172117188573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 6 + ], + "selected_reward": -0.016064144670963287 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, almost motionless shot, shot from a slightly upward angle, locks the viewer's gaze firmly on a boundless, rolling sea of clouds. A resolute Eastern warrior stands barefoot on the tip of a slender, emerald green bamboo pole. His long, jet-black hair is tied into a high bun with a simple cloth belt. A few strands, blown by the sharp wind from high above, gently brush against his angular features. He wears a loose, washed-bleached linen warrior's uniform, its sleeves rustling in the wind. His core remains rock-solid, swaying only subtly, imperceptibly, against the force of the swaying bamboo pole. The camera meticulously captures his deep, unwavering eyes, revealing a calmness and focus beyond his years, as if at one with the universe. Beneath his feet, the tip of the bamboo pole, bearing his entire weight, vibrated visibly and at high frequency due to the combined forces of wind and gravity, creating a stark contrast to the absolute stability of his upper body. The first rays of golden morning sunlight pierced the clouds, casting a warm glow on his bronzed skin and capturing every detail of his tense toes gripping the slippery pole. The entire scene, shrouded in a vast sense of loneliness, was imbued with an ultimate, restrained, and powerful tension, creating an otherworldly martial arts realm.", + "seed": 25114, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_5/index_8967_seed_25114_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.025589977949857713, + "vlm_logprob": -0.025589977949857713, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.025589977949857713, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed tightly with a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Eastern warrior standing barefoot on the tip of a slender emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior have long jet-black hair tied in a high bun with a simple cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior wearing a loose washed-bleached linen uniform while standing on a rolling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the first rays of golden morning sunlight piercing through the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03472441807389259 + ], + "opposite_logprob_tokens": [ + -3.909724473953247 + ], + "logprob_reward": -0.03472441807389259, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.03472441807389259, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22092187404632568, + 0.045592136681079865, + 0.03278680890798569, + 0.04513603076338768, + 0.057793453335762024, + 0.20452164113521576, + 0.06140050292015076, + 0.12801752984523773, + 0.14442971348762512, + 0.0594002828001976 + ], + "credit_map": [ + 0.018288739025592804, + 0.04042830318212509, + 0.060558948665857315, + 0.030988002195954323, + 0.022400908172130585, + 0.012796509079635143, + 0.022160647436976433, + 0.013299833983182907, + 0.003638535039499402, + 0.004846398718655109, + 0.005588171072304249, + 0.00725287850946188, + 0.00488645676523447, + 0.005975113715976477, + 0.007090071216225624, + 0.006314517930150032, + 0.0026199761778116226, + 0.004029441624879837, + 0.003721794346347451, + 0.006127836182713509, + 0.003559341887012124, + 0.004365001805126667, + 0.0046050637029111385, + 0.0037583541125059128, + 0.004559640772640705, + 0.007528165355324745, + 0.0055679380893707275, + 0.0056253583170473576, + 0.004914100747555494, + 0.005149036645889282, + 0.006377138197422028, + 0.005414656363427639, + 0.0038922857493162155, + 0.0042152791284024715, + 0.00775441899895668, + 0.013452200219035149, + 0.004574927967041731, + 0.005379330832511187, + 0.011392625980079174, + 0.007132386788725853, + 0.009470164775848389, + 0.0073450650088489056, + 0.07532473653554916, + 0.08246845752000809, + 0.0074268514290452, + 0.002667035674676299, + 0.01674148254096508, + 0.0030778630170971155, + 0.004463630262762308, + 0.006947746034711599, + 0.007639732211828232, + 0.011513011530041695, + 0.006308949552476406, + 0.00439572986215353, + 0.013515966013073921, + 0.006615741644054651, + 0.01492227241396904, + 0.011177664622664452, + 0.03911464288830757, + 0.01561990287154913, + 0.009003297425806522, + 0.012784178368747234, + 0.015197950415313244, + 0.010197627358138561, + 0.01226238813251257, + 0.016159629449248314, + 0.03096412681043148, + 0.019518528133630753, + 0.013539742678403854, + 0.0132740568369627, + 0.02475447952747345, + 0.013956766575574875, + 0.006608116906136274, + 0.006072618067264557, + 0.010746839456260204, + 0.009967425838112831, + 0.005135491024702787, + 0.006279189605265856, + 0.007076950743794441, + 0.0075136576779186726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.03472441807389259 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed tightly with a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Eastern warrior standing barefoot on the tip of a slender emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior have long jet-black hair tied in a high bun with a simple cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior wearing a loose washed-bleached linen uniform while standing on a rolling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the first rays of golden morning sunlight piercing through the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026362799108028412 + ], + "opposite_logprob_tokens": [ + -3.651362895965576 + ], + "logprob_reward": -0.026362799108028412, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026362799108028412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13031107187271118, + 0.08823951333761215, + 0.04155593365430832, + 0.03324606642127037, + 0.025210976600646973, + 0.26534897089004517, + 0.050016727298498154, + 0.2092328816652298, + 0.11644470691680908, + 0.04039314389228821 + ], + "credit_map": [ + 0.010762101970613003, + 0.012078388594090939, + 0.03410382941365242, + 0.024290459230542183, + 0.017191562801599503, + 0.009729770943522453, + 0.013083740137517452, + 0.009071226231753826, + 0.009330446831882, + 0.014964153058826923, + 0.01359613984823227, + 0.013200719840824604, + 0.013687248341739178, + 0.007640617433935404, + 0.008703944273293018, + 0.007116246037185192, + 0.002539407927542925, + 0.0033841473050415516, + 0.0055245086550712585, + 0.00785752572119236, + 0.004194166045635939, + 0.0060987211763858795, + 0.006241519469767809, + 0.005715939216315746, + 0.0030760911758989096, + 0.006263607647269964, + 0.004709181841462851, + 0.005234804004430771, + 0.003991367295384407, + 0.0029116207733750343, + 0.004458257462829351, + 0.0026011355221271515, + 0.0012792866909876466, + 0.0019067812245339155, + 0.003996613901108503, + 0.003734714351594448, + 0.003156328806653619, + 0.004169694613665342, + 0.0036417751107364893, + 0.003325785044580698, + 0.011775420047342777, + 0.009435856714844704, + 0.09542994201183319, + 0.1041698157787323, + 0.009937215596437454, + 0.005996340420097113, + 0.020983658730983734, + 0.007620713207870722, + 0.003703588154166937, + 0.006415752228349447, + 0.006197797600179911, + 0.010196714662015438, + 0.005986722651869059, + 0.005682431627064943, + 0.006572002544999123, + 0.005261718761175871, + 0.024209938943386078, + 0.016662145033478737, + 0.06746238470077515, + 0.01478924322873354, + 0.014439348131418228, + 0.023748919367790222, + 0.02166905254125595, + 0.02625187672674656, + 0.01037560123950243, + 0.013320785015821457, + 0.028775189071893692, + 0.013714531436562538, + 0.009909335523843765, + 0.01025767344981432, + 0.01741795614361763, + 0.012673640623688698, + 0.003367079421877861, + 0.004017204511910677, + 0.010300382040441036, + 0.004286166746169329, + 0.0020078637171536684, + 0.006460690870881081, + 0.00387369142845273, + 0.006080069113522768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.026362799108028412 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed tightly with a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Eastern warrior standing barefoot on the tip of a slender emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior have long jet-black hair tied in a high bun with a simple cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior wearing a loose washed-bleached linen uniform while standing on a rolling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the first rays of golden morning sunlight piercing through the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026334352791309357 + ], + "opposite_logprob_tokens": [ + -3.651334285736084 + ], + "logprob_reward": -0.026334352791309357, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026334352791309357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3635512888431549, + 0.152215376496315, + 0.04932492971420288, + 0.060136280953884125, + 0.06527324765920639, + 0.11199190467596054, + 0.048696693032979965, + 0.04209680110216141, + 0.05562354251742363, + 0.05109000205993652 + ], + "credit_map": [ + 0.024905597791075706, + 0.05285843089222908, + 0.11914235353469849, + 0.06026395410299301, + 0.036371368914842606, + 0.017103534191846848, + 0.03778236731886864, + 0.015123657882213593, + 0.009665028192102909, + 0.014538981020450592, + 0.03344868868589401, + 0.027151383459568024, + 0.02479846403002739, + 0.01263357326388359, + 0.0195145346224308, + 0.010464699938893318, + 0.0030720033682882786, + 0.004554763436317444, + 0.005487483460456133, + 0.01218703854829073, + 0.004604042042046785, + 0.005515635944902897, + 0.009071692824363708, + 0.004832264501601458, + 0.005343882367014885, + 0.008766865357756615, + 0.007301175966858864, + 0.012373425997793674, + 0.005991584621369839, + 0.005187148228287697, + 0.010407545603811741, + 0.004764643497765064, + 0.0033495412208139896, + 0.0048118033446371555, + 0.00963587500154972, + 0.02271055243909359, + 0.004887707997113466, + 0.004354477860033512, + 0.011412391439080238, + 0.004110889974981546, + 0.005371879320591688, + 0.0044101146049797535, + 0.03920856863260269, + 0.042619578540325165, + 0.004852072801440954, + 0.003023143857717514, + 0.008494485169649124, + 0.004012053832411766, + 0.004531749989837408, + 0.0070279487408697605, + 0.007135036401450634, + 0.00805168878287077, + 0.004863664973527193, + 0.005865494254976511, + 0.0060676406137645245, + 0.005153460428118706, + 0.005043970886617899, + 0.005648871883749962, + 0.006745445542037487, + 0.0053457655012607574, + 0.004423946607857943, + 0.005013253539800644, + 0.004877857398241758, + 0.004997684620320797, + 0.003754219738766551, + 0.006170306820422411, + 0.011076051741838455, + 0.007696649059653282, + 0.00499174976721406, + 0.00437094597145915, + 0.011156353168189526, + 0.006407260894775391, + 0.0109532680362463, + 0.0042545367032289505, + 0.005646214820444584, + 0.009258382953703403, + 0.0037669260054826736, + 0.00469932658597827, + 0.006517387460917234, + 0.005993950180709362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.026334352791309357 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed tightly with a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Eastern warrior standing barefoot on the tip of a slender emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior have long jet-black hair tied in a high bun with a simple cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior wearing a loose washed-bleached linen uniform while standing on a rolling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the first rays of golden morning sunlight piercing through the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014190716668963432 + ], + "opposite_logprob_tokens": [ + -4.264190673828125 + ], + "logprob_reward": -0.014190716668963432, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014190716668963432, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27244916558265686, + 0.17783628404140472, + 0.07244493067264557, + 0.06933555006980896, + 0.12834228575229645, + 0.1291881948709488, + 0.03889762610197067, + 0.06039665266871452, + 0.029500974342226982, + 0.021608388051390648 + ], + "credit_map": [ + 0.021258467808365822, + 0.025022031739354134, + 0.04458512365818024, + 0.06940821558237076, + 0.03726469725370407, + 0.013438417576253414, + 0.04611790552735329, + 0.015354298986494541, + 0.018493812531232834, + 0.027944713830947876, + 0.031177090480923653, + 0.031409960240125656, + 0.026446552947163582, + 0.010132203809916973, + 0.020978761836886406, + 0.011253169737756252, + 0.003184537636116147, + 0.005785876885056496, + 0.009277601726353168, + 0.023598525673151016, + 0.00551564572378993, + 0.004141390323638916, + 0.016095275059342384, + 0.004846072290092707, + 0.0037286276929080486, + 0.0066098240204155445, + 0.008343725465238094, + 0.01595880836248398, + 0.007661717943847179, + 0.006620210595428944, + 0.014473812654614449, + 0.005938820540904999, + 0.005203592125326395, + 0.008397444151341915, + 0.017616691067814827, + 0.04843804985284805, + 0.01112202275544405, + 0.0037519382312893867, + 0.02649308182299137, + 0.007319448050111532, + 0.005976257845759392, + 0.005079145077615976, + 0.04771296679973602, + 0.05092844367027283, + 0.004863892216235399, + 0.0021351336035877466, + 0.009997562505304813, + 0.002494787098839879, + 0.0034850635565817356, + 0.006608390714973211, + 0.004078694619238377, + 0.00678597204387188, + 0.004273967817425728, + 0.0042864298447966576, + 0.005309741012752056, + 0.004069363698363304, + 0.006665852386504412, + 0.0067565301433205605, + 0.014888545498251915, + 0.005760285537689924, + 0.0053847720846533775, + 0.008150728419423103, + 0.004961312748491764, + 0.00782862026244402, + 0.003071187762543559, + 0.005343264900147915, + 0.004443277604877949, + 0.004856005776673555, + 0.003733585588634014, + 0.0024333714973181486, + 0.0029639494605362415, + 0.0026563280262053013, + 0.0036259687040001154, + 0.0021748696453869343, + 0.0025590884033590555, + 0.004210048820823431, + 0.001902652671560645, + 0.0018173076678067446, + 0.00331704830750823, + 0.00200140243396163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.014190716668963432 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed tightly with a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Eastern warrior standing barefoot on the tip of a slender emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior have long jet-black hair tied in a high bun with a simple cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior wearing a loose washed-bleached linen uniform while standing on a rolling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the first rays of golden morning sunlight piercing through the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026337603107094765 + ], + "opposite_logprob_tokens": [ + -3.6513376235961914 + ], + "logprob_reward": -0.026337603107094765, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026337603107094765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2798372507095337, + 0.06141030415892601, + 0.07863249629735947, + 0.043172240257263184, + 0.16976693272590637, + 0.19497352838516235, + 0.08821193128824234, + 0.03477991744875908, + 0.02201290801167488, + 0.027202391996979713 + ], + "credit_map": [ + 0.02099882997572422, + 0.036879971623420715, + 0.09164587408304214, + 0.047005705535411835, + 0.03028029017150402, + 0.010429125279188156, + 0.03056795336306095, + 0.012029546312987804, + 0.004022329114377499, + 0.00681390892714262, + 0.012002112343907356, + 0.011162184178829193, + 0.01299341768026352, + 0.0033574074041098356, + 0.008276667445898056, + 0.0027822840493172407, + 0.004986671730875969, + 0.008340729400515556, + 0.013678268529474735, + 0.020673833787441254, + 0.007585722021758556, + 0.0031458872836083174, + 0.016547325998544693, + 0.0036740717478096485, + 0.00435253232717514, + 0.00657578045502305, + 0.005281558260321617, + 0.0064263190142810345, + 0.005489595700055361, + 0.0031278461683541536, + 0.008035443723201752, + 0.003883166704326868, + 0.007614553440362215, + 0.010569386184215546, + 0.024778610095381737, + 0.06598802655935287, + 0.013880386017262936, + 0.004311522003263235, + 0.03405877202749252, + 0.008565709926187992, + 0.01267065666615963, + 0.015493117272853851, + 0.059263214468955994, + 0.05867292732000351, + 0.012802154757082462, + 0.005406116135418415, + 0.02216332033276558, + 0.00850205309689045, + 0.0072733014822006226, + 0.01230128388851881, + 0.010563910007476807, + 0.016382895410060883, + 0.007980595342814922, + 0.006955028045922518, + 0.016520986333489418, + 0.010233938694000244, + 0.004237662069499493, + 0.004942070692777634, + 0.0068765017203986645, + 0.005027293227612972, + 0.0038359391037374735, + 0.0034669251181185246, + 0.0030889350455254316, + 0.0033045909367501736, + 0.0015102863544598222, + 0.002086907159537077, + 0.0037034982815384865, + 0.0036410456523299217, + 0.0021008446346968412, + 0.002234171610325575, + 0.004761540796607733, + 0.001974615966901183, + 0.007675090339034796, + 0.0030107363127171993, + 0.003956236410886049, + 0.004083232954144478, + 0.002041940111666918, + 0.001997649436816573, + 0.002428207779303193, + 0.002009303541854024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.026337603107094765 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, almost motionless shot, shot from a slightly upward angle, locks the viewer's gaze firmly on a boundless, rolling sea of clouds. A resolute Eastern warrior stands barefoot on the tip of a slender, emerald green bamboo pole. His long, jet-black hair is tied into a high bun with a simple cloth belt. A few strands, blown by the sharp wind from high above, gently brush against his angular features. He wears a loose, washed-bleached linen warrior's uniform, its sleeves rustling in the wind. His core remains rock-solid, swaying only subtly, imperceptibly, against the force of the swaying bamboo pole. The camera meticulously captures his deep, unwavering eyes, revealing a calmness and focus beyond his years, as if at one with the universe. Beneath his feet, the tip of the bamboo pole, bearing his entire weight, vibrated visibly and at high frequency due to the combined forces of wind and gravity, creating a stark contrast to the absolute stability of his upper body. The first rays of golden morning sunlight pierced the clouds, casting a warm glow on his bronzed skin and capturing every detail of his tense toes gripping the slippery pole. The entire scene, shrouded in a vast sense of loneliness, was imbued with an ultimate, restrained, and powerful tension, creating an otherworldly martial arts realm.", + "seed": 25115, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_5/index_8967_seed_25115_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.056810795702040194, + "vlm_logprob": -0.056810795702040194, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.056810795702040194, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed tightly with a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Eastern warrior standing barefoot on the tip of a slender emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior have long jet-black hair tied in a high bun with a simple cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior wearing a loose washed-bleached linen uniform while standing on a rolling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the first rays of golden morning sunlight piercing through the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13036160171031952 + ], + "opposite_logprob_tokens": [ + -2.255361557006836 + ], + "logprob_reward": -0.13036160171031952, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.13036160171031952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1937149316072464, + 0.14343559741973877, + 0.07477989792823792, + 0.12747135758399963, + 0.026634708046913147, + 0.06546709686517715, + 0.07822906970977783, + 0.19106386601924896, + 0.037788935005664825, + 0.06141449138522148 + ], + "credit_map": [ + 0.02940911054611206, + 0.01605094224214554, + 0.03340943530201912, + 0.026535402983427048, + 0.02434019185602665, + 0.01945294253528118, + 0.01909973844885826, + 0.02541716955602169, + 0.01720605604350567, + 0.013595659285783768, + 0.012918026186525822, + 0.021961253136396408, + 0.021135399118065834, + 0.017149101942777634, + 0.022225696593523026, + 0.017244400456547737, + 0.007942995056509972, + 0.008749791420996189, + 0.008883507922291756, + 0.01296395342797041, + 0.012203920632600784, + 0.008164566941559315, + 0.007691430393606424, + 0.008179735392332077, + 0.016458824276924133, + 0.00936591625213623, + 0.02617417275905609, + 0.013190507888793945, + 0.009059591218829155, + 0.010964656248688698, + 0.016055617481470108, + 0.026202073320746422, + 0.0051535735838115215, + 0.002968485001474619, + 0.002534640021622181, + 0.0020336038433015347, + 0.004837051033973694, + 0.003530538873746991, + 0.002751661231741309, + 0.0028251546900719404, + 0.009431936778128147, + 0.006145661696791649, + 0.008111201226711273, + 0.006391420029103756, + 0.0058012851513922215, + 0.006820425856858492, + 0.006335501559078693, + 0.016429664567112923, + 0.008460449986159801, + 0.0090184910222888, + 0.0070524681359529495, + 0.01190265268087387, + 0.01064450852572918, + 0.008702118881046772, + 0.010911036282777786, + 0.011537346988916397, + 0.020667459815740585, + 0.02731279842555523, + 0.015601406805217266, + 0.03927362710237503, + 0.02265743911266327, + 0.020487433299422264, + 0.02812683954834938, + 0.016936855390667915, + 0.005773467011749744, + 0.003072975901886821, + 0.006815524771809578, + 0.002962552709504962, + 0.004576108418405056, + 0.003981288522481918, + 0.004039651248604059, + 0.00656737107783556, + 0.006936556193977594, + 0.004385924898087978, + 0.009722720831632614, + 0.006938112899661064, + 0.005928023252636194, + 0.007079575210809708, + 0.005278616212308407, + 0.01514496561139822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.13036160171031952 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed tightly with a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Eastern warrior standing barefoot on the tip of a slender emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior have long jet-black hair tied in a high bun with a simple cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior wearing a loose washed-bleached linen uniform while standing on a rolling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the first rays of golden morning sunlight piercing through the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900706678628922 + ], + "opposite_logprob_tokens": [ + -2.579007148742676 + ], + "logprob_reward": -0.07900706678628922, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900706678628922, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24353483319282532, + 0.07884512096643448, + 0.07691163569688797, + 0.05460739508271217, + 0.0518290139734745, + 0.028034187853336334, + 0.07697135210037231, + 0.2991738021373749, + 0.05670195817947388, + 0.03339061513543129 + ], + "credit_map": [ + 0.02948344312608242, + 0.021093087270855904, + 0.03357033431529999, + 0.036833181977272034, + 0.029231544584035873, + 0.02634112909436226, + 0.027674468234181404, + 0.039307672530412674, + 0.010049300268292427, + 0.0064018298871815205, + 0.007805893663316965, + 0.0128579530864954, + 0.011113460175693035, + 0.00816714484244585, + 0.011693255975842476, + 0.010756293311715126, + 0.0058220140635967255, + 0.00896411668509245, + 0.007644359488040209, + 0.014752931892871857, + 0.013236195780336857, + 0.009273676201701164, + 0.008557827211916447, + 0.008660520426928997, + 0.009381377138197422, + 0.004945410881191492, + 0.006259790621697903, + 0.005515332333743572, + 0.006774350069463253, + 0.006274663843214512, + 0.007711539510637522, + 0.007744935806840658, + 0.006191145163029432, + 0.0053734020330011845, + 0.00782765168696642, + 0.0043931277468800545, + 0.006776153575628996, + 0.005184630397707224, + 0.004208058584481478, + 0.011874848045408726, + 0.00390134216286242, + 0.0019919241312891245, + 0.00475877383723855, + 0.0022628523875027895, + 0.0031330573838204145, + 0.0029931352473795414, + 0.002474009757861495, + 0.0065190959721803665, + 0.015502980910241604, + 0.005104504060000181, + 0.011006728745996952, + 0.006485659163445234, + 0.006251205690205097, + 0.008578416891396046, + 0.00715159997344017, + 0.0168902650475502, + 0.030044473707675934, + 0.03917684033513069, + 0.02225235290825367, + 0.05647995322942734, + 0.04305918142199516, + 0.033495739102363586, + 0.04503355920314789, + 0.0296317245811224, + 0.005360987037420273, + 0.007241816725581884, + 0.005090879276394844, + 0.00800183042883873, + 0.009076629765331745, + 0.006353985983878374, + 0.00866649579256773, + 0.006909338757395744, + 0.00325140287168324, + 0.003737400518730283, + 0.004757114220410585, + 0.0038689160719513893, + 0.004803558811545372, + 0.0036966807674616575, + 0.003341731382533908, + 0.005933811888098717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.07900706678628922 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed tightly with a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Eastern warrior standing barefoot on the tip of a slender emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior have long jet-black hair tied in a high bun with a simple cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior wearing a loose washed-bleached linen uniform while standing on a rolling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the first rays of golden morning sunlight piercing through the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057156339287758 + ], + "opposite_logprob_tokens": [ + -3.89557147026062 + ], + "logprob_reward": -0.02057156339287758, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057156339287758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13560043275356293, + 0.1045290008187294, + 0.051101330667734146, + 0.17366090416908264, + 0.0462004691362381, + 0.050604548305273056, + 0.06314773112535477, + 0.3009651005268097, + 0.05846178159117699, + 0.015728751197457314 + ], + "credit_map": [ + 0.01937904767692089, + 0.01703243888914585, + 0.014193549752235413, + 0.018134059384465218, + 0.017827320843935013, + 0.014726022258400917, + 0.018605880439281464, + 0.015702107921242714, + 0.011474424041807652, + 0.010713381692767143, + 0.012842798605561256, + 0.01461808755993843, + 0.017580946907401085, + 0.011159366928040981, + 0.013848123140633106, + 0.012291868217289448, + 0.004435171838849783, + 0.004902387969195843, + 0.006078089587390423, + 0.008914164267480373, + 0.008134189993143082, + 0.005712559446692467, + 0.007959610782563686, + 0.004965154454112053, + 0.022051101550459862, + 0.013267355971038342, + 0.034755151718854904, + 0.020366691052913666, + 0.011281296610832214, + 0.014902034774422646, + 0.021260211244225502, + 0.035777051001787186, + 0.0055274637416005135, + 0.005115218460559845, + 0.0029908274300396442, + 0.00725009897723794, + 0.007224165368825197, + 0.004852441605180502, + 0.008264858275651932, + 0.004975392948836088, + 0.0068018753081560135, + 0.003185844048857689, + 0.006659104488790035, + 0.0044431607238948345, + 0.0043480233289301395, + 0.005674239248037338, + 0.005902864038944244, + 0.013589433394372463, + 0.010575381107628345, + 0.004619842395186424, + 0.009442608803510666, + 0.0062063029035925865, + 0.005333464127033949, + 0.006705380976200104, + 0.004962523002177477, + 0.015302221290767193, + 0.031866297125816345, + 0.04200756922364235, + 0.02707459218800068, + 0.05977468565106392, + 0.038326364010572433, + 0.03241119533777237, + 0.041046272963285446, + 0.028458114713430405, + 0.010740304365754128, + 0.004025436472147703, + 0.01188623160123825, + 0.004246899392455816, + 0.004866302944719791, + 0.006333878729492426, + 0.004672860261052847, + 0.011689863167703152, + 0.0015591472620144486, + 0.001838783617131412, + 0.0020727303344756365, + 0.002323579043149948, + 0.0025709320325404406, + 0.001868718070909381, + 0.0016289901686832309, + 0.0018658703193068504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.02057156339287758 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed tightly with a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Eastern warrior standing barefoot on the tip of a slender emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior have long jet-black hair tied in a high bun with a simple cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior wearing a loose washed-bleached linen uniform while standing on a rolling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the first rays of golden morning sunlight piercing through the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011068728752434254 + ], + "opposite_logprob_tokens": [ + -4.511068820953369 + ], + "logprob_reward": -0.011068728752434254, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011068728752434254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22646671533584595, + 0.05134844407439232, + 0.047431278973817825, + 0.12044854462146759, + 0.07518285512924194, + 0.07053341716527939, + 0.13055886328220367, + 0.10232793539762497, + 0.1465437412261963, + 0.0291582141071558 + ], + "credit_map": [ + 0.03697574883699417, + 0.018036553636193275, + 0.05348464846611023, + 0.02012069895863533, + 0.020257659256458282, + 0.026844093576073647, + 0.019561585038900375, + 0.031185736879706383, + 0.008750294335186481, + 0.00528918020427227, + 0.010029568336904049, + 0.004380715545266867, + 0.005064633674919605, + 0.005893061403185129, + 0.00626740837469697, + 0.005673579405993223, + 0.008603179827332497, + 0.003440724452957511, + 0.00804133340716362, + 0.004418815020471811, + 0.004038576036691666, + 0.005576899740844965, + 0.005469704046845436, + 0.007842046208679676, + 0.01553073525428772, + 0.007951663807034492, + 0.02102624624967575, + 0.013379407115280628, + 0.010822475887835026, + 0.013310684822499752, + 0.013185945339500904, + 0.025241388007998466, + 0.010696438141167164, + 0.00676838681101799, + 0.014580313116312027, + 0.007306476589292288, + 0.006620487198233604, + 0.007544063031673431, + 0.006994359660893679, + 0.01467233244329691, + 0.00926773902028799, + 0.007747935131192207, + 0.008737221360206604, + 0.005803469568490982, + 0.013172862119972706, + 0.009013413451611996, + 0.006183713208884001, + 0.010607069358229637, + 0.02907514199614525, + 0.01009418722242117, + 0.021250514313578606, + 0.007299755699932575, + 0.012330330908298492, + 0.014510542154312134, + 0.008878187276422977, + 0.027120200917124748, + 0.01119229570031166, + 0.012589678168296814, + 0.00991408433765173, + 0.016325777396559715, + 0.016934243962168694, + 0.011694765649735928, + 0.011900416575372219, + 0.011776674538850784, + 0.02487998642027378, + 0.010432794690132141, + 0.03112138994038105, + 0.009767432697117329, + 0.011888450011610985, + 0.016410842537879944, + 0.010117137804627419, + 0.03192570433020592, + 0.004064069129526615, + 0.0031114390585571527, + 0.0034589909482747316, + 0.004839668981730938, + 0.004186780191957951, + 0.003336532274261117, + 0.003126898780465126, + 0.003033833345398307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.011068728752434254 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed tightly with a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Eastern warrior standing barefoot on the tip of a slender emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior have long jet-black hair tied in a high bun with a simple cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior wearing a loose washed-bleached linen uniform while standing on a rolling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the first rays of golden morning sunlight piercing through the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304501786828041 + ], + "opposite_logprob_tokens": [ + -3.1680450439453125 + ], + "logprob_reward": -0.04304501786828041, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304501786828041, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17078141868114471, + 0.08434294164180756, + 0.06159121170639992, + 0.31731027364730835, + 0.03355443850159645, + 0.049523040652275085, + 0.034813057631254196, + 0.17211994528770447, + 0.04762967303395271, + 0.028333932161331177 + ], + "credit_map": [ + 0.034343887120485306, + 0.019448690116405487, + 0.035609904676675797, + 0.012774425558745861, + 0.01453297771513462, + 0.018784867599606514, + 0.01528963353484869, + 0.019997036084532738, + 0.012262959964573383, + 0.00921491626650095, + 0.011291698552668095, + 0.014149712398648262, + 0.01137822400778532, + 0.0098875742405653, + 0.009349732659757137, + 0.006808116100728512, + 0.010140171274542809, + 0.006421075668185949, + 0.010646882466971874, + 0.006458721589297056, + 0.009707356803119183, + 0.0073787434957921505, + 0.005833432078361511, + 0.005004833918064833, + 0.04336587339639664, + 0.021158404648303986, + 0.059812720865011215, + 0.03621397912502289, + 0.024041632190346718, + 0.031481724232435226, + 0.03786037862300873, + 0.06337558478116989, + 0.004318708553910255, + 0.0025624632835388184, + 0.004876295570284128, + 0.0037709625903517008, + 0.00275175622664392, + 0.0031557916663587093, + 0.00434652017429471, + 0.007771939504891634, + 0.006810755934566259, + 0.004877759143710136, + 0.00582424271851778, + 0.003982312977313995, + 0.008357084356248379, + 0.005667951423674822, + 0.004329662304371595, + 0.00967327505350113, + 0.005980841349810362, + 0.0032926532439887524, + 0.0036323866806924343, + 0.004996676929295063, + 0.0028945032972842455, + 0.003469018964096904, + 0.0032646367326378822, + 0.0072823395021259785, + 0.01836959645152092, + 0.025326671078801155, + 0.013763940893113613, + 0.03631305322051048, + 0.020243290811777115, + 0.018280191347002983, + 0.0248960442841053, + 0.014927150681614876, + 0.007615850772708654, + 0.0032300585880875587, + 0.010200674645602703, + 0.0031883884221315384, + 0.0034764630254358053, + 0.005382528994232416, + 0.003376963548362255, + 0.011158744804561138, + 0.0030484579037874937, + 0.0024626704398542643, + 0.004207865800708532, + 0.0037731199990957975, + 0.0032467704731971025, + 0.0029868725687265396, + 0.003100304864346981, + 0.005507869645953178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.04304501786828041 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, almost motionless shot, shot from a slightly upward angle, locks the viewer's gaze firmly on a boundless, rolling sea of clouds. A resolute Eastern warrior stands barefoot on the tip of a slender, emerald green bamboo pole. His long, jet-black hair is tied into a high bun with a simple cloth belt. A few strands, blown by the sharp wind from high above, gently brush against his angular features. He wears a loose, washed-bleached linen warrior's uniform, its sleeves rustling in the wind. His core remains rock-solid, swaying only subtly, imperceptibly, against the force of the swaying bamboo pole. The camera meticulously captures his deep, unwavering eyes, revealing a calmness and focus beyond his years, as if at one with the universe. Beneath his feet, the tip of the bamboo pole, bearing his entire weight, vibrated visibly and at high frequency due to the combined forces of wind and gravity, creating a stark contrast to the absolute stability of his upper body. The first rays of golden morning sunlight pierced the clouds, casting a warm glow on his bronzed skin and capturing every detail of his tense toes gripping the slippery pole. The entire scene, shrouded in a vast sense of loneliness, was imbued with an ultimate, restrained, and powerful tension, creating an otherworldly martial arts realm.", + "seed": 25116, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_5/index_8967_seed_25116_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03573183938860893, + "vlm_logprob": -0.03573183938860893, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.03573183938860893, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed tightly with a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Eastern warrior standing barefoot on the tip of a slender emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior have long jet-black hair tied in a high bun with a simple cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior wearing a loose washed-bleached linen uniform while standing on a rolling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the first rays of golden morning sunlight piercing through the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04945746809244156 + ], + "opposite_logprob_tokens": [ + -3.799457550048828 + ], + "logprob_reward": -0.04945746809244156, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.04945746809244156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08806151151657104, + 0.06547123193740845, + 0.04729929566383362, + 0.5227591395378113, + 0.03853493928909302, + 0.026667343452572823, + 0.06303076446056366, + 0.10136372596025467, + 0.028490306809544563, + 0.018321756273508072 + ], + "credit_map": [ + 0.010793210007250309, + 0.01338382437825203, + 0.014821280725300312, + 0.010034279897809029, + 0.005566230975091457, + 0.005791387055069208, + 0.005287869833409786, + 0.022383425384759903, + 0.009529506787657738, + 0.006919432897120714, + 0.009044735692441463, + 0.009809404611587524, + 0.005962439812719822, + 0.0071760620921850204, + 0.007972050458192825, + 0.009057598188519478, + 0.005623225588351488, + 0.004975916817784309, + 0.0061568268574774265, + 0.005388377234339714, + 0.006765126716345549, + 0.004270750563591719, + 0.0068052480928599834, + 0.007313819136470556, + 0.030905285850167274, + 0.011572650633752346, + 0.19963635504245758, + 0.1349681317806244, + 0.011796689592301846, + 0.10811043530702591, + 0.009875877760350704, + 0.01589367166161537, + 0.004103976301848888, + 0.0031176144257187843, + 0.00928330235183239, + 0.006517594680190086, + 0.0032381804194301367, + 0.006752234883606434, + 0.0032061324454844, + 0.0023158995900303125, + 0.004651627037674189, + 0.002260376000776887, + 0.0060044233687222, + 0.0032595465891063213, + 0.0024812493938952684, + 0.003196579637005925, + 0.0020854719914495945, + 0.002728069666773081, + 0.00781865231692791, + 0.006267154589295387, + 0.02029157616198063, + 0.006233553867787123, + 0.002893768483772874, + 0.013604896143078804, + 0.0031751426868140697, + 0.0027460181154310703, + 0.005379774142056704, + 0.004716155584901571, + 0.04053012281656265, + 0.008511732332408428, + 0.0025533423759043217, + 0.03128615394234657, + 0.0029280069284141064, + 0.005458441097289324, + 0.0037775104865431786, + 0.0027451051864773035, + 0.0038142141420394182, + 0.0054421210661530495, + 0.0023715223651379347, + 0.003393769497051835, + 0.0022747786715626717, + 0.004671287257224321, + 0.0016544954851269722, + 0.0017369389533996582, + 0.0040622069500386715, + 0.0018128009978681803, + 0.0019140007207170129, + 0.003160364693030715, + 0.0015839064726606011, + 0.0023970413021743298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.04945746809244156 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed tightly with a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Eastern warrior standing barefoot on the tip of a slender emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior have long jet-black hair tied in a high bun with a simple cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior wearing a loose washed-bleached linen uniform while standing on a rolling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the first rays of golden morning sunlight piercing through the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372403234243393 + ], + "opposite_logprob_tokens": [ + -3.408724069595337 + ], + "logprob_reward": -0.03372403234243393, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372403234243393, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.026449793949723244, + 0.020627764984965324, + 0.017540959641337395, + 0.8657954335212708, + 0.014332451857626438, + 0.017153911292552948, + 0.012184846214950085, + 0.010555781424045563, + 0.006124125327914953, + 0.009234903380274773 + ], + "credit_map": [ + 0.0035463678650557995, + 0.0038359989412128925, + 0.0041025904938578606, + 0.004444865044206381, + 0.0016889467369765043, + 0.0027462439611554146, + 0.00211490411311388, + 0.003969878423959017, + 0.0028019617311656475, + 0.0021190810948610306, + 0.003271507564932108, + 0.0034796663094311953, + 0.001953606493771076, + 0.0017546929884701967, + 0.0025275282096117735, + 0.002719721756875515, + 0.0017905060667544603, + 0.0013051442801952362, + 0.002537206746637821, + 0.002868308685719967, + 0.0018268806161358953, + 0.0021057557314634323, + 0.0022640663664788008, + 0.0028430926613509655, + 0.05222879350185394, + 0.018335776403546333, + 0.3321659564971924, + 0.22298088669776917, + 0.018555330112576485, + 0.18679291009902954, + 0.013982579112052917, + 0.020753251388669014, + 0.0016718379920348525, + 0.0012224785750731826, + 0.002903963904827833, + 0.0026341662742197514, + 0.0013914337614551187, + 0.0011497990926727653, + 0.0018210721900686622, + 0.0015377014642581344, + 0.002994849346578121, + 0.0011347370455041528, + 0.004576963372528553, + 0.002371863927692175, + 0.0008163430029526353, + 0.002672641770914197, + 0.001074943458661437, + 0.0015115711139515042, + 0.001864087302237749, + 0.0017116295639425516, + 0.0015125644858926535, + 0.0013933514710515738, + 0.0014485756400972605, + 0.0013999835355207324, + 0.0013075766619294882, + 0.0015470779035240412, + 0.0011722301132977009, + 0.0013560954248532653, + 0.002467390615493059, + 0.0019651823677122593, + 0.0008879370870999992, + 0.0014049465535208583, + 0.0006013806560076773, + 0.0007006203522905707, + 0.001205135602504015, + 0.0009518364095129073, + 0.0008619839791208506, + 0.0009235207107849419, + 0.00046937342267483473, + 0.0007768672658130527, + 0.00046136436867527664, + 0.0004740442964248359, + 0.0008522608550265431, + 0.0013950187712907791, + 0.0013246657326817513, + 0.0017085329163819551, + 0.0008007906144484878, + 0.00181394477840513, + 0.0006135375588200986, + 0.0007261530845426023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.03372403234243393 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed tightly with a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Eastern warrior standing barefoot on the tip of a slender emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior have long jet-black hair tied in a high bun with a simple cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior wearing a loose washed-bleached linen uniform while standing on a rolling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the first rays of golden morning sunlight piercing through the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026334119960665703 + ], + "opposite_logprob_tokens": [ + -3.651334047317505 + ], + "logprob_reward": -0.026334119960665703, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026334119960665703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17329032719135284, + 0.07434746623039246, + 0.023596249520778656, + 0.46755945682525635, + 0.1244787648320198, + 0.0219045989215374, + 0.05002405494451523, + 0.030303964391350746, + 0.014554225839674473, + 0.019940882921218872 + ], + "credit_map": [ + 0.018649576231837273, + 0.02614484541118145, + 0.03546402230858803, + 0.021257435902953148, + 0.010250918567180634, + 0.014240270480513573, + 0.009973453357815742, + 0.03730979561805725, + 0.007952009327709675, + 0.007280130870640278, + 0.010084518231451511, + 0.010822806507349014, + 0.00954658817499876, + 0.0081942705437541, + 0.01092213299125433, + 0.00954501237720251, + 0.0029527186416089535, + 0.002417465439066291, + 0.003793517593294382, + 0.00395236536860466, + 0.0024301439989358187, + 0.0024445108138024807, + 0.0020681675523519516, + 0.003537359181791544, + 0.02888849377632141, + 0.010346765629947186, + 0.17362995445728302, + 0.11586808413267136, + 0.010837296955287457, + 0.10594344139099121, + 0.008823358453810215, + 0.013222052715718746, + 0.009736267849802971, + 0.006034426856786013, + 0.04011877253651619, + 0.024495285004377365, + 0.006917682942003012, + 0.022232720628380775, + 0.008173322305083275, + 0.006770292297005653, + 0.0037820430006831884, + 0.0017202373128384352, + 0.00495826406404376, + 0.002750921295955777, + 0.002105516381561756, + 0.002649447415024042, + 0.002280857414007187, + 0.0016573121538385749, + 0.004265325143933296, + 0.0035166058223694563, + 0.017591018229722977, + 0.009421263821423054, + 0.0018243174999952316, + 0.008835330605506897, + 0.002006443915888667, + 0.0025637487415224314, + 0.003615909954532981, + 0.002812662860378623, + 0.007440236397087574, + 0.00488108629360795, + 0.0019393941620364785, + 0.0043052495457232, + 0.0020683107431977987, + 0.0032411147840321064, + 0.0012845090823248029, + 0.0012037717970088124, + 0.0020487289875745773, + 0.002365873195230961, + 0.0012848512269556522, + 0.002104984363541007, + 0.001390445395372808, + 0.002871060511097312, + 0.0014510019682347775, + 0.0021255468018352985, + 0.0032874520402401686, + 0.0023180812131613493, + 0.002033090917393565, + 0.0034106802195310593, + 0.0018034547101706266, + 0.003511573886498809 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.026334119960665703 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed tightly with a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Eastern warrior standing barefoot on the tip of a slender emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior have long jet-black hair tied in a high bun with a simple cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior wearing a loose washed-bleached linen uniform while standing on a rolling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the first rays of golden morning sunlight piercing through the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014197180047631264 + ], + "opposite_logprob_tokens": [ + -4.26419734954834 + ], + "logprob_reward": -0.014197180047631264, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014197180047631264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.013760371133685112, + 0.01320992223918438, + 0.006304811220616102, + 0.9034551978111267, + 0.014172607101500034, + 0.009690227918326855, + 0.015141436830163002, + 0.012571665458381176, + 0.006904866546392441, + 0.004788883961737156 + ], + "credit_map": [ + 0.0014353669248521328, + 0.0021608243696391582, + 0.0028927563689649105, + 0.0019256377127021551, + 0.0008964987355284393, + 0.0011663269251585007, + 0.0009226792608387768, + 0.0023602808360010386, + 0.0022297995164990425, + 0.0017644044710323215, + 0.002071266993880272, + 0.00221330882050097, + 0.0009818400721997023, + 0.00110380700789392, + 0.0015068229986354709, + 0.0013386734062805772, + 0.0009780809050425887, + 0.000516428321134299, + 0.0011885996209457517, + 0.0011678008595481515, + 0.0003955366264563054, + 0.0008969811606220901, + 0.00046823019511066377, + 0.0006931530660949647, + 0.054432716220617294, + 0.019320311024785042, + 0.3460034728050232, + 0.23227310180664062, + 0.01941782981157303, + 0.19580969214439392, + 0.014651555567979813, + 0.021546510979533195, + 0.0013368971413001418, + 0.0005255031283013523, + 0.0047808317467570305, + 0.0029083252884447575, + 0.0005061448900960386, + 0.0029541365802288055, + 0.0005978741100989282, + 0.0005628941580653191, + 0.002198807429522276, + 0.0010105789406225085, + 0.0022609515581279993, + 0.0014226045459508896, + 0.0004023685760330409, + 0.0011055042268708348, + 0.0005444038542918861, + 0.0007450085249729455, + 0.0021456500981003046, + 0.0020478046499192715, + 0.004386537708342075, + 0.0016488443361595273, + 0.0005594475078396499, + 0.0033654849976301193, + 0.0005597167764790356, + 0.0004279497661627829, + 0.0010611393954604864, + 0.0012084118789061904, + 0.0035379724577069283, + 0.0017979067051783204, + 0.0007003798382356763, + 0.002446399535983801, + 0.0006190022686496377, + 0.0012004532618448138, + 0.0014767649117857218, + 0.0010105798719450831, + 0.0008520967094227672, + 0.0011069910833612084, + 0.0005920901312492788, + 0.0008063918794505298, + 0.0004635895020328462, + 0.0005963628063909709, + 0.00044993613846600056, + 0.0007888968684710562, + 0.0008105949964374304, + 0.0010542948730289936, + 0.000220559086301364, + 0.0008682948537170887, + 0.0002393314935034141, + 0.0003569757391232997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 4 + ], + "selected_reward": -0.014197180047631264 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed tightly with a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Eastern warrior standing barefoot on the tip of a slender emerald green bamboo pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior have long jet-black hair tied in a high bun with a simple cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior wearing a loose washed-bleached linen uniform while standing on a rolling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the first rays of golden morning sunlight piercing through the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494639649987221 + ], + "opposite_logprob_tokens": [ + -2.9299464225769043 + ], + "logprob_reward": -0.05494639649987221, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494639649987221, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0711720660328865, + 0.041852425783872604, + 0.035222724080085754, + 0.16417603194713593, + 0.06879324465990067, + 0.08536731451749802, + 0.11603332310914993, + 0.24435976147651672, + 0.08288110047578812, + 0.09014201164245605 + ], + "credit_map": [ + 0.007880894467234612, + 0.012221111916005611, + 0.016198471188545227, + 0.007629002444446087, + 0.0053221965208649635, + 0.004846903961151838, + 0.004377584904432297, + 0.012695896439254284, + 0.007208125665783882, + 0.005334788467735052, + 0.005341924261301756, + 0.005286574363708496, + 0.004340391606092453, + 0.004588364623486996, + 0.005077242851257324, + 0.004675010684877634, + 0.004339329432696104, + 0.0029599564149975777, + 0.0053122080862522125, + 0.006993426941335201, + 0.002654607640579343, + 0.003361119655892253, + 0.00222432310692966, + 0.007377752102911472, + 0.013260748237371445, + 0.004288923926651478, + 0.05881498381495476, + 0.03954099863767624, + 0.0039785634726285934, + 0.0346178337931633, + 0.003121744142845273, + 0.0065522389486432076, + 0.006278046872466803, + 0.003042984986677766, + 0.021729299798607826, + 0.013073365204036236, + 0.003473181976005435, + 0.012701320461928844, + 0.003963572438806295, + 0.004531461745500565, + 0.019734906032681465, + 0.009100912138819695, + 0.018501752987504005, + 0.009288265369832516, + 0.004767140373587608, + 0.008640793152153492, + 0.004447717685252428, + 0.010885822586715221, + 0.014742259867489338, + 0.014946172945201397, + 0.036646220833063126, + 0.021007420495152473, + 0.004068049602210522, + 0.016351686790585518, + 0.004577303770929575, + 0.0036942046135663986, + 0.01770893856883049, + 0.01922752894461155, + 0.08275211602449417, + 0.027834584936499596, + 0.009643815457820892, + 0.06159355118870735, + 0.01354027260094881, + 0.01205893512815237, + 0.011365124955773354, + 0.007766458205878735, + 0.010927737690508366, + 0.014550921507179737, + 0.007658171933144331, + 0.016341056674718857, + 0.005526895634829998, + 0.008744723163545132, + 0.008606036193668842, + 0.012179824523627758, + 0.022301191464066505, + 0.011815222911536694, + 0.007164936978369951, + 0.017359700053930283, + 0.005136170424520969, + 0.005578917916864157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.05494639649987221 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_6/index_tensor([4279, 4279])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_6/index_tensor([4279, 4279])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..79d0ae4837dfd738dc1cb56b1fb6bc8a9c6390fe --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_6/index_tensor([4279, 4279])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning macro close-up, a slow and steady upward tilt, focuses on an intricate, circular spiderweb, dappled with shimmering dewdrops in the morning air. The first rays of soft, golden sunlight filter through the web, casting each plump dewdrop like a tiny magnifying glass, refracting a dazzling light. A tiny, glossy, jet-black spider, its velvety, round body contrasting with eight slender, hairy legs, delicately and rhythmically ascends a main thread. Its legs tap the shimmering silk with precise tips, each step causing the web to tremble subtly, sending the dewdrops swaying and shimmering like diamonds. The background is a completely blurred, out-of-focus deep green foliage. This shallow depth of field dramatically accentuates the astonishing details of the spider and web in the foreground. The entire image, framed by the cool background and the warm morning light, is imbued with the tranquility of early morning and the exquisite beauty of nature, creating an atmosphere of focused concentration and peace.", + "seed": 9973, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_6/index_4279_seed_9973_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02921478860080242, + "vlm_logprob": -0.02921478860080242, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.02921478860080242, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a circular spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spider jet-black with a velvety round body and eight slender hairy legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the spiderweb strands dappled with shimmering dewdrops that refract light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a completely blurred, out-of-focus deep green foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow and steady upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009822133928537369 + ], + "opposite_logprob_tokens": [ + -4.759822368621826 + ], + "logprob_reward": -0.009822133928537369, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.009822133928537369, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05629463121294975, + 0.012368720024824142, + 0.0344986729323864, + 0.013113527558743954, + 0.09597480297088623, + 0.03567807003855705, + 0.022571982815861702, + 0.04660949856042862, + 0.059629183262586594, + 0.6232608556747437 + ], + "credit_map": [ + 0.01759054698050022, + 0.0053843422792851925, + 0.0072627286426723, + 0.004377509467303753, + 0.004781135357916355, + 0.006606339477002621, + 0.004141696263104677, + 0.0061503322795033455, + 0.001244044047780335, + 0.002436296548694372, + 0.0013072856236249208, + 0.00163882109336555, + 0.0013249438488855958, + 0.0018502635648474097, + 0.0014385511167347431, + 0.001128513365983963, + 0.0037275368813425303, + 0.0031705847941339016, + 0.004477855283766985, + 0.0033600942697376013, + 0.004891548305749893, + 0.0049681938253343105, + 0.004658957477658987, + 0.005243901629000902, + 0.002132012275978923, + 0.0011400299845263362, + 0.001578712835907936, + 0.0011530447518453002, + 0.001935988082550466, + 0.0019683903083205223, + 0.0016920879716053605, + 0.0015132598346099257, + 0.008287626318633556, + 0.01233127061277628, + 0.00644807331264019, + 0.009383650496602058, + 0.014086300507187843, + 0.01286929938942194, + 0.022482452914118767, + 0.01008613407611847, + 0.004277936182916164, + 0.0035252226516604424, + 0.003643359523266554, + 0.004468430299311876, + 0.0031410474330186844, + 0.005509199108928442, + 0.00561333866789937, + 0.005499538499861956, + 0.0034553194418549538, + 0.001903928117826581, + 0.0022445092909038067, + 0.0018025552853941917, + 0.0016918402398005128, + 0.003938241396099329, + 0.003283134661614895, + 0.004252452868968248, + 0.0060651907697319984, + 0.004510512575507164, + 0.005430150311440229, + 0.005711485631763935, + 0.00434171874076128, + 0.006712646223604679, + 0.006179432850331068, + 0.007658359128981829, + 0.0024640141054987907, + 0.0017910490278154612, + 0.0024817718658596277, + 0.002405136590823531, + 0.003040845040231943, + 0.022519884631037712, + 0.0029861205257475376, + 0.021940361708402634, + 0.034926313906908035, + 0.025599272921681404, + 0.029190635308623314, + 0.016417386010289192, + 0.024505890905857086, + 0.2353169322013855, + 0.03961678221821785, + 0.21768765151500702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 8 + ], + "selected_reward": -0.009822133928537369 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a circular spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spider jet-black with a velvety round body and eight slender hairy legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the spiderweb strands dappled with shimmering dewdrops that refract light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a completely blurred, out-of-focus deep green foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow and steady upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029831182211637497 + ], + "opposite_logprob_tokens": [ + -3.5298311710357666 + ], + "logprob_reward": -0.029831182211637497, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029831182211637497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08435849845409393, + 0.07329528778791428, + 0.052239421755075455, + 0.04678839445114136, + 0.25488606095314026, + 0.047354742884635925, + 0.10027007758617401, + 0.0808229148387909, + 0.08077409118413925, + 0.17921048402786255 + ], + "credit_map": [ + 0.020485976710915565, + 0.007399005349725485, + 0.011036873795092106, + 0.008632016368210316, + 0.007917680777609348, + 0.010657768696546555, + 0.00808090902864933, + 0.010148273780941963, + 0.007689818739891052, + 0.011105349287390709, + 0.008449403569102287, + 0.011657142080366611, + 0.009306428954005241, + 0.009957724250853062, + 0.009259498678147793, + 0.005869929678738117, + 0.005114878993481398, + 0.007784229703247547, + 0.00529555743560195, + 0.00615704245865345, + 0.01003020815551281, + 0.008204932324588299, + 0.0055650267750024796, + 0.004087551962584257, + 0.007113116327673197, + 0.004451186861842871, + 0.00771971745416522, + 0.005039933603256941, + 0.004086512606590986, + 0.007448822725564241, + 0.007343633566051722, + 0.003585475031286478, + 0.017070066183805466, + 0.03177080675959587, + 0.01619560271501541, + 0.029860077425837517, + 0.030425800010561943, + 0.03742353618144989, + 0.0676039606332779, + 0.02453623153269291, + 0.005607572849839926, + 0.0043791658245027065, + 0.004518692847341299, + 0.006175158079713583, + 0.004074636846780777, + 0.006955527700483799, + 0.008903167210519314, + 0.006740827113389969, + 0.013901961967349052, + 0.008035714738070965, + 0.00897503923624754, + 0.009558379650115967, + 0.00915373396128416, + 0.01821945793926716, + 0.0141422338783741, + 0.018283570185303688, + 0.011941004544496536, + 0.00812322273850441, + 0.010009263642132282, + 0.008197013288736343, + 0.008252275176346302, + 0.011701377108693123, + 0.009279799647629261, + 0.013318966142833233, + 0.005297349765896797, + 0.004677475430071354, + 0.008123603649437428, + 0.006591302342712879, + 0.007136868312954903, + 0.018826259300112724, + 0.009350993670523167, + 0.020770249888300896, + 0.03127627819776535, + 0.009199854917824268, + 0.021224485710263252, + 0.008570599369704723, + 0.007849314250051975, + 0.04444132745265961, + 0.01290771272033453, + 0.043740928173065186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 6 + ], + "selected_reward": -0.029831182211637497 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a circular spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spider jet-black with a velvety round body and eight slender hairy legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the spiderweb strands dappled with shimmering dewdrops that refract light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a completely blurred, out-of-focus deep green foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow and steady upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628949359059334 + ], + "opposite_logprob_tokens": [ + -4.882628917694092 + ], + "logprob_reward": -0.007628949359059334, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628949359059334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04723656550049782, + 0.04927853122353554, + 0.033068299293518066, + 0.03848747909069061, + 0.018822219222784042, + 0.03722865879535675, + 0.10333327203989029, + 0.03473018482327461, + 0.24643319845199585, + 0.3913815915584564 + ], + "credit_map": [ + 0.012338261120021343, + 0.003847817424684763, + 0.005449921824038029, + 0.0046287584118545055, + 0.0044391583651304245, + 0.005834834650158882, + 0.005875836592167616, + 0.004821975249797106, + 0.005500324536114931, + 0.010083643719553947, + 0.005450777243822813, + 0.004809357691556215, + 0.0048294588923454285, + 0.007096685003489256, + 0.006396687589585781, + 0.005111596081405878, + 0.004370179492980242, + 0.0028583554085344076, + 0.005463423207402229, + 0.003398999571800232, + 0.00459736492484808, + 0.0048034656792879105, + 0.0029915415216237307, + 0.00458496855571866, + 0.004923171829432249, + 0.0033974593970924616, + 0.006153949070721865, + 0.0034646126441657543, + 0.006240165326744318, + 0.005875390022993088, + 0.004638245329260826, + 0.003794486401602626, + 0.002528159413486719, + 0.0020532887428998947, + 0.0017221341840922832, + 0.002980514895170927, + 0.001795244519598782, + 0.0026003411039710045, + 0.002862138208001852, + 0.002280399203300476, + 0.0043419417925179005, + 0.003393440740182996, + 0.0037999653723090887, + 0.0046305228024721146, + 0.0040686349384486675, + 0.005434527527540922, + 0.0076200952753424644, + 0.003939530812203884, + 0.01526585128158331, + 0.008316787891089916, + 0.010672607459127903, + 0.008802005089819431, + 0.008654911071062088, + 0.019492048770189285, + 0.012760668992996216, + 0.019368385896086693, + 0.0032252361997962, + 0.002681610407307744, + 0.0032247030176222324, + 0.0035976190119981766, + 0.0026877548079937696, + 0.006668818183243275, + 0.004724788945168257, + 0.00791965238749981, + 0.009925168938934803, + 0.006866969633847475, + 0.011755037121474743, + 0.009595025330781937, + 0.013085848651826382, + 0.09001746773719788, + 0.013002874329686165, + 0.09218478947877884, + 0.015599455684423447, + 0.014882341027259827, + 0.01384690310806036, + 0.009849722497165203, + 0.017107276245951653, + 0.15245865285396576, + 0.025550831109285355, + 0.14208641648292542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.007628949359059334 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a circular spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spider jet-black with a velvety round body and eight slender hairy legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the spiderweb strands dappled with shimmering dewdrops that refract light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a completely blurred, out-of-focus deep green foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow and steady upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780462831258774 + ], + "opposite_logprob_tokens": [ + -4.634780406951904 + ], + "logprob_reward": -0.009780462831258774, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780462831258774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23446208238601685, + 0.026488400995731354, + 0.06800786405801773, + 0.05026855319738388, + 0.0471046045422554, + 0.08430534601211548, + 0.06035394221544266, + 0.2696245014667511, + 0.07762477546930313, + 0.08175992965698242 + ], + "credit_map": [ + 0.025461547076702118, + 0.01961934007704258, + 0.02932991459965706, + 0.030461620539426804, + 0.025064198300242424, + 0.028589393943548203, + 0.04292331635951996, + 0.0330127514898777, + 0.003023892641067505, + 0.0036425015423446894, + 0.0032812440767884254, + 0.002554871141910553, + 0.0024968076031655073, + 0.004137052223086357, + 0.00433523440733552, + 0.0030167961958795786, + 0.007079578470438719, + 0.00641857273876667, + 0.008240271359682083, + 0.008775843307375908, + 0.010774235241115093, + 0.00941191241145134, + 0.006465358659625053, + 0.01084208581596613, + 0.008530127815902233, + 0.0047604842111468315, + 0.007456263527274132, + 0.005134821869432926, + 0.003732056822627783, + 0.007891632616519928, + 0.007448482792824507, + 0.005314679350703955, + 0.0056015849113464355, + 0.005627792328596115, + 0.004302704706788063, + 0.0066763595677912235, + 0.00519854249432683, + 0.006870912853628397, + 0.006767227314412594, + 0.006059478502720594, + 0.011763176880776882, + 0.007703178562223911, + 0.008740226738154888, + 0.010354643687605858, + 0.00701250322163105, + 0.013206890784204006, + 0.01350948866456747, + 0.012015242129564285, + 0.006808329839259386, + 0.005645470228046179, + 0.006144469138234854, + 0.006872240919619799, + 0.0064244987443089485, + 0.00888532493263483, + 0.0088996272534132, + 0.010673982091248035, + 0.02209923416376114, + 0.013956795446574688, + 0.017504271119832993, + 0.013822326436638832, + 0.016767261549830437, + 0.08004575967788696, + 0.01872396469116211, + 0.08670490235090256, + 0.0034383083693683147, + 0.0029505600687116385, + 0.00418366864323616, + 0.004151078872382641, + 0.0034304901491850615, + 0.02754816599190235, + 0.005154982674866915, + 0.02676752582192421, + 0.011452386155724525, + 0.006002750713378191, + 0.006871264427900314, + 0.006564280949532986, + 0.006192915141582489, + 0.017219508066773415, + 0.010374810546636581, + 0.017082013189792633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.009780462831258774 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a circular spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spider jet-black with a velvety round body and eight slender hairy legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the spiderweb strands dappled with shimmering dewdrops that refract light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a completely blurred, out-of-focus deep green foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow and steady upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08901121467351913 + ], + "opposite_logprob_tokens": [ + -2.4640111923217773 + ], + "logprob_reward": -0.08901121467351913, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901121467351913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06387145072221756, + 0.04035475477576256, + 0.02818111702799797, + 0.04944843426346779, + 0.1549222320318222, + 0.020650574937462807, + 0.10902636498212814, + 0.12814249098300934, + 0.24420632421970367, + 0.16119632124900818 + ], + "credit_map": [ + 0.01619499921798706, + 0.00519611407071352, + 0.009369430132210255, + 0.004082241095602512, + 0.0066540916450321674, + 0.008445760235190392, + 0.005936066620051861, + 0.007992736995220184, + 0.005268814507871866, + 0.006433906499296427, + 0.00561866257339716, + 0.004633671138435602, + 0.004339518025517464, + 0.005815759301185608, + 0.00459540169686079, + 0.003649016609415412, + 0.002987393643707037, + 0.0030208323150873184, + 0.0033073718659579754, + 0.0033044908195734024, + 0.004858419764786959, + 0.004114743787795305, + 0.003139419946819544, + 0.003448442555963993, + 0.005463009234517813, + 0.0036156647838652134, + 0.0054032825864851475, + 0.005689309909939766, + 0.004506549332290888, + 0.0070702205412089825, + 0.012432455085217953, + 0.005267934873700142, + 0.010020517744123936, + 0.019508644938468933, + 0.009636648930609226, + 0.019026881083846092, + 0.016158288344740868, + 0.024474384263157845, + 0.04281255975365639, + 0.013284292072057724, + 0.002926440443843603, + 0.001560401520691812, + 0.0021360241807997227, + 0.002673166571184993, + 0.0015019659185782075, + 0.0029877282213419676, + 0.0033755949698388577, + 0.003489249851554632, + 0.01447511836886406, + 0.009190988726913929, + 0.009998694062232971, + 0.012058878317475319, + 0.010543815791606903, + 0.01790790446102619, + 0.014773678965866566, + 0.020077278837561607, + 0.01302517019212246, + 0.007756839040666819, + 0.010741813108325005, + 0.009809620678424835, + 0.008031698875129223, + 0.03332878649234772, + 0.010141322389245033, + 0.03530723601579666, + 0.005302167497575283, + 0.004876057617366314, + 0.00600967463105917, + 0.004620156716555357, + 0.008079706691205502, + 0.10272690653800964, + 0.009893716312944889, + 0.1026979312300682, + 0.012327971868216991, + 0.0059038931503891945, + 0.006830346304923296, + 0.006644850596785545, + 0.006432840600609779, + 0.05521468445658684, + 0.010281464084982872, + 0.05756024271249771 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.08901121467351913 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning macro close-up, a slow and steady upward tilt, focuses on an intricate, circular spiderweb, dappled with shimmering dewdrops in the morning air. The first rays of soft, golden sunlight filter through the web, casting each plump dewdrop like a tiny magnifying glass, refracting a dazzling light. A tiny, glossy, jet-black spider, its velvety, round body contrasting with eight slender, hairy legs, delicately and rhythmically ascends a main thread. Its legs tap the shimmering silk with precise tips, each step causing the web to tremble subtly, sending the dewdrops swaying and shimmering like diamonds. The background is a completely blurred, out-of-focus deep green foliage. This shallow depth of field dramatically accentuates the astonishing details of the spider and web in the foreground. The entire image, framed by the cool background and the warm morning light, is imbued with the tranquility of early morning and the exquisite beauty of nature, creating an atmosphere of focused concentration and peace.", + "seed": 9974, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_6/index_4279_seed_9974_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04792668055742979, + "vlm_logprob": -0.04792668055742979, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.04792668055742979, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a circular spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spider jet-black with a velvety round body and eight slender hairy legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the spiderweb strands dappled with shimmering dewdrops that refract light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a completely blurred, out-of-focus deep green foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow and steady upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009237172082066536 + ], + "opposite_logprob_tokens": [ + -5.009237289428711 + ], + "logprob_reward": -0.009237172082066536, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.009237172082066536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3188895285129547, + 0.06290178745985031, + 0.07431841641664505, + 0.030649632215499878, + 0.05350888520479202, + 0.07613679021596909, + 0.02735537476837635, + 0.1849675327539444, + 0.09130337089300156, + 0.07996862381696701 + ], + "credit_map": [ + 0.021439658477902412, + 0.015739327296614647, + 0.037238504737615585, + 0.023567350581288338, + 0.05803433433175087, + 0.09542776644229889, + 0.03266046568751335, + 0.03478216007351875, + 0.006037180311977863, + 0.009303792379796505, + 0.006879584863781929, + 0.0077844406478106976, + 0.010097624734044075, + 0.007505157496780157, + 0.006934863980859518, + 0.008359148167073727, + 0.005965552758425474, + 0.004801563918590546, + 0.01775222085416317, + 0.0048965634778141975, + 0.0048875631764531136, + 0.008626594208180904, + 0.004038844257593155, + 0.023349525406956673, + 0.0031571858562529087, + 0.0044252462685108185, + 0.0030787293799221516, + 0.004323434084653854, + 0.0029284118209034204, + 0.0046046581119298935, + 0.004405577667057514, + 0.003726392751559615, + 0.007823779247701168, + 0.005058967042714357, + 0.005746070295572281, + 0.004114995710551739, + 0.006632709410041571, + 0.009048818610608578, + 0.008447127416729927, + 0.0066364239901304245, + 0.007177468854933977, + 0.008329108357429504, + 0.009364297613501549, + 0.007984064519405365, + 0.00872849766165018, + 0.012245168909430504, + 0.013224964961409569, + 0.009083225391805172, + 0.0029578227549791336, + 0.003992949612438679, + 0.004322470165789127, + 0.002793230814859271, + 0.003781647887080908, + 0.003293314017355442, + 0.003061868716031313, + 0.00315207545645535, + 0.013389364816248417, + 0.008434406481683254, + 0.010294950567185879, + 0.048035878688097, + 0.015912942588329315, + 0.07132289558649063, + 0.009063096717000008, + 0.00851400662213564, + 0.02678076922893524, + 0.005785150453448296, + 0.007236127741634846, + 0.005815166048705578, + 0.003428508061915636, + 0.030184639617800713, + 0.004287172574549913, + 0.00778584461659193, + 0.002770542399957776, + 0.005086466204375029, + 0.0038764760829508305, + 0.003886507824063301, + 0.003510952228680253, + 0.027986545115709305, + 0.0039383769035339355, + 0.02891276590526104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.009237172082066536 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a circular spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spider jet-black with a velvety round body and eight slender hairy legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the spiderweb strands dappled with shimmering dewdrops that refract light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a completely blurred, out-of-focus deep green foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow and steady upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08908919245004654 + ], + "opposite_logprob_tokens": [ + -2.4640891551971436 + ], + "logprob_reward": -0.08908919245004654, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08908919245004654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31395450234413147, + 0.07376682013273239, + 0.11760979145765305, + 0.056440237909555435, + 0.06070459634065628, + 0.06338532269001007, + 0.051460150629282, + 0.17202086746692657, + 0.05805116891860962, + 0.032606467604637146 + ], + "credit_map": [ + 0.02637936919927597, + 0.021137526258826256, + 0.046180952340364456, + 0.03597012534737587, + 0.030969558283686638, + 0.07770565897226334, + 0.03273514658212662, + 0.04287618026137352, + 0.007710892707109451, + 0.004542314447462559, + 0.013867981731891632, + 0.006032358855009079, + 0.007003724575042725, + 0.017997441813349724, + 0.0061792791821062565, + 0.010432826355099678, + 0.009810869581997395, + 0.009338272735476494, + 0.029340572655200958, + 0.008226205594837666, + 0.00741662597283721, + 0.02933632768690586, + 0.008581161499023438, + 0.015559753403067589, + 0.006052096839994192, + 0.01217422541230917, + 0.0055694920010864735, + 0.007610434200614691, + 0.006858406588435173, + 0.0055807214230299, + 0.005662177223712206, + 0.006932685151696205, + 0.010198886506259441, + 0.005646918900310993, + 0.010677317157387733, + 0.005320829804986715, + 0.004446717910468578, + 0.010149835608899593, + 0.006454538553953171, + 0.00780955096706748, + 0.00738147459924221, + 0.005600163713097572, + 0.005043529439717531, + 0.0063281552866101265, + 0.007410363759845495, + 0.009450317360460758, + 0.014591546729207039, + 0.007579772733151913, + 0.009259096346795559, + 0.004799686372280121, + 0.005929740611463785, + 0.008485584519803524, + 0.005991731770336628, + 0.006246140226721764, + 0.005255425814539194, + 0.005492744036018848, + 0.017575090751051903, + 0.009005224332213402, + 0.014122053049504757, + 0.03986683860421181, + 0.013791874051094055, + 0.05807522311806679, + 0.008999486453831196, + 0.010585085488855839, + 0.01709025725722313, + 0.0038000966887921095, + 0.004744740203022957, + 0.00367296626791358, + 0.002240144880488515, + 0.01899140141904354, + 0.0026209275238215923, + 0.004890633746981621, + 0.001489138463512063, + 0.0019101607613265514, + 0.0016142809763550758, + 0.0014124492881819606, + 0.001759658451192081, + 0.011171068996191025, + 0.002283745212480426, + 0.010965965688228607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.08908919245004654 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a circular spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spider jet-black with a velvety round body and eight slender hairy legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the spiderweb strands dappled with shimmering dewdrops that refract light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a completely blurred, out-of-focus deep green foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow and steady upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014203056693077087 + ], + "opposite_logprob_tokens": [ + -4.264203071594238 + ], + "logprob_reward": -0.014203056693077087, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014203056693077087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17860724031925201, + 0.035706181079149246, + 0.4275249242782593, + 0.015617440454661846, + 0.053508657962083817, + 0.1052372008562088, + 0.014248257502913475, + 0.08814869076013565, + 0.045380864292383194, + 0.03602055460214615 + ], + "credit_map": [ + 0.016292225569486618, + 0.012382248416543007, + 0.0276106595993042, + 0.025221988558769226, + 0.014803632162511349, + 0.035000912845134735, + 0.02129167504608631, + 0.026003902778029442, + 0.0033469514455646276, + 0.003316170535981655, + 0.005857890471816063, + 0.004174914676696062, + 0.004730584565550089, + 0.006793182343244553, + 0.0030930819921195507, + 0.004393406677991152, + 0.021515360102057457, + 0.011280789971351624, + 0.031266264617443085, + 0.014728652313351631, + 0.01881968230009079, + 0.1421031653881073, + 0.027467014268040657, + 0.16034400463104248, + 0.0014301674673333764, + 0.0018675141036510468, + 0.0015021348372101784, + 0.0017348239198327065, + 0.0016933594597503543, + 0.002960567595437169, + 0.0019734478555619717, + 0.002455424750223756, + 0.008906573988497257, + 0.005186923313885927, + 0.007855823263525963, + 0.0032315070275217295, + 0.0047073387540876865, + 0.010707177221775055, + 0.005170832853764296, + 0.007742483634501696, + 0.014279022812843323, + 0.014093611389398575, + 0.019780872389674187, + 0.008972707204520702, + 0.0071986704133450985, + 0.016897808760404587, + 0.009642169810831547, + 0.014372332952916622, + 0.0016657038358971477, + 0.00146950944326818, + 0.0016791543457657099, + 0.0017665036721155047, + 0.0023199901916086674, + 0.0017657916760072112, + 0.001699384767562151, + 0.0018822189886122942, + 0.007517586927860975, + 0.0051002479158341885, + 0.007512237876653671, + 0.01910385489463806, + 0.007250766269862652, + 0.029733942821621895, + 0.005434694699943066, + 0.0064953588880598545, + 0.0130744194611907, + 0.002520678099244833, + 0.0034194004256278276, + 0.0028148400597274303, + 0.0017571505159139633, + 0.015988117083907127, + 0.002084252890199423, + 0.003722007619217038, + 0.0016457964666187763, + 0.002463363343849778, + 0.002157286973670125, + 0.0017742383060976863, + 0.0014186317566782236, + 0.012524682097136974, + 0.001565056387335062, + 0.012471499852836132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.014203056693077087 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a circular spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spider jet-black with a velvety round body and eight slender hairy legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the spiderweb strands dappled with shimmering dewdrops that refract light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a completely blurred, out-of-focus deep green foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow and steady upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193889684975147 + ], + "opposite_logprob_tokens": [ + -4.264194011688232 + ], + "logprob_reward": -0.014193889684975147, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193889684975147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14976853132247925, + 0.13350912928581238, + 0.18406663835048676, + 0.10045037418603897, + 0.05765685811638832, + 0.056901317089796066, + 0.034150801599025726, + 0.1680983453989029, + 0.0792713314294815, + 0.03612663969397545 + ], + "credit_map": [ + 0.01205061562359333, + 0.010504013858735561, + 0.019943902269005775, + 0.01599378138780594, + 0.01566084660589695, + 0.03520563989877701, + 0.019635720178484917, + 0.02077402174472809, + 0.013503811322152615, + 0.018837597221136093, + 0.015141615644097328, + 0.015483085066080093, + 0.01838614046573639, + 0.01787159964442253, + 0.01599712297320366, + 0.01828817091882229, + 0.009726438671350479, + 0.005477190483361483, + 0.007057737093418837, + 0.0074078296311199665, + 0.01000255811959505, + 0.05832362547516823, + 0.01235482282936573, + 0.07371645420789719, + 0.012601845897734165, + 0.013507891446352005, + 0.015688659623265266, + 0.012444941326975822, + 0.010867174714803696, + 0.015828533098101616, + 0.009496096521615982, + 0.010015233419835567, + 0.008578252047300339, + 0.006705520674586296, + 0.00468482868745923, + 0.003895985195413232, + 0.0058422815054655075, + 0.012195233255624771, + 0.007012773770838976, + 0.008741985075175762, + 0.006293981336057186, + 0.0068176607601344585, + 0.01499547902494669, + 0.00486717326566577, + 0.002952926093712449, + 0.009425264783203602, + 0.0031920415349304676, + 0.008356792852282524, + 0.004870998673141003, + 0.0037080987822264433, + 0.0037001098971813917, + 0.00496930256485939, + 0.0048508490435779095, + 0.004701583180576563, + 0.0034740010742098093, + 0.0038758614100515842, + 0.013127438724040985, + 0.00852954015135765, + 0.011093324981629848, + 0.041406892240047455, + 0.012227155268192291, + 0.05962848663330078, + 0.012109220027923584, + 0.00997630599886179, + 0.024060063064098358, + 0.003922565840184689, + 0.006224996875971556, + 0.004916831385344267, + 0.002908024936914444, + 0.028441522270441055, + 0.0031765976455062628, + 0.005620729643851519, + 0.0018101914320141077, + 0.0028766219038516283, + 0.002313675358891487, + 0.002078537130728364, + 0.0017469633603468537, + 0.012239187955856323, + 0.0018421319546177983, + 0.011219334788620472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.014193889684975147 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a circular spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spider jet-black with a velvety round body and eight slender hairy legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the spiderweb strands dappled with shimmering dewdrops that refract light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a completely blurred, out-of-focus deep green foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow and steady upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291009187698364 + ], + "opposite_logprob_tokens": [ + -2.237910032272339 + ], + "logprob_reward": -0.11291009187698364, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291009187698364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08055989444255829, + 0.10581133514642715, + 0.30315741896629333, + 0.17906464636325836, + 0.1458839327096939, + 0.05835311859846115, + 0.015269756317138672, + 0.037075672298669815, + 0.02585981786251068, + 0.048964399844408035 + ], + "credit_map": [ + 0.006418032571673393, + 0.004721920005977154, + 0.00953978206962347, + 0.005823131650686264, + 0.009165726602077484, + 0.022259674966335297, + 0.00989590398967266, + 0.01273573562502861, + 0.010743532329797745, + 0.015558718703687191, + 0.013768344186246395, + 0.013013415969908237, + 0.014637225307524204, + 0.012335313484072685, + 0.013923742808401585, + 0.011831052601337433, + 0.01641121692955494, + 0.009432036429643631, + 0.03860555216670036, + 0.012098410166800022, + 0.014129070565104485, + 0.09739834815263748, + 0.019852127879858017, + 0.09523066133260727, + 0.01864938996732235, + 0.0356828048825264, + 0.017882665619254112, + 0.02227046713232994, + 0.020876212045550346, + 0.017600206658244133, + 0.01919550821185112, + 0.0269074197858572, + 0.021636778488755226, + 0.014807886444032192, + 0.011327317915856838, + 0.012543313205242157, + 0.017473427578806877, + 0.025587869808077812, + 0.021669046953320503, + 0.02083829790353775, + 0.005443965550512075, + 0.005388819612562656, + 0.010621105320751667, + 0.005458827130496502, + 0.005864021833986044, + 0.008460883051156998, + 0.00935325026512146, + 0.007762250956147909, + 0.0017465646378695965, + 0.001611000276170671, + 0.0021919363643974066, + 0.0015859826235100627, + 0.0013542117085307837, + 0.0019078270997852087, + 0.0018532476387917995, + 0.003018987365067005, + 0.0034410979133099318, + 0.002398956334218383, + 0.002732998225837946, + 0.0076004378497600555, + 0.0028610024601221085, + 0.010345114395022392, + 0.004090211819857359, + 0.003605854231864214, + 0.005331626161932945, + 0.0017160223796963692, + 0.002315546851605177, + 0.0023056960199028254, + 0.0019276151433587074, + 0.007409979589283466, + 0.0025403189938515425, + 0.002313014818355441, + 0.0018083011964336038, + 0.0032914213370531797, + 0.002269291551783681, + 0.0019206700380891562, + 0.0026256530545651913, + 0.017863860353827477, + 0.0029006691183894873, + 0.016284538432955742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 4 + ], + "selected_reward": -0.11291009187698364 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning macro close-up, a slow and steady upward tilt, focuses on an intricate, circular spiderweb, dappled with shimmering dewdrops in the morning air. The first rays of soft, golden sunlight filter through the web, casting each plump dewdrop like a tiny magnifying glass, refracting a dazzling light. A tiny, glossy, jet-black spider, its velvety, round body contrasting with eight slender, hairy legs, delicately and rhythmically ascends a main thread. Its legs tap the shimmering silk with precise tips, each step causing the web to tremble subtly, sending the dewdrops swaying and shimmering like diamonds. The background is a completely blurred, out-of-focus deep green foliage. This shallow depth of field dramatically accentuates the astonishing details of the spider and web in the foreground. The entire image, framed by the cool background and the warm morning light, is imbued with the tranquility of early morning and the exquisite beauty of nature, creating an atmosphere of focused concentration and peace.", + "seed": 9975, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_6/index_4279_seed_9975_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.039375022333115337, + "vlm_logprob": -0.039375022333115337, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.039375022333115337, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a circular spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spider jet-black with a velvety round body and eight slender hairy legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the spiderweb strands dappled with shimmering dewdrops that refract light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a completely blurred, out-of-focus deep green foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow and steady upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008813646622002125 + ], + "opposite_logprob_tokens": [ + -4.883813858032227 + ], + "logprob_reward": -0.008813646622002125, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.008813646622002125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1557939052581787, + 0.12471479922533035, + 0.029652539640665054, + 0.05291984975337982, + 0.26472094655036926, + 0.22196879982948303, + 0.05415145307779312, + 0.018845027312636375, + 0.043400462716817856, + 0.033832233399152756 + ], + "credit_map": [ + 0.02137843146920204, + 0.016139965504407883, + 0.014915123581886292, + 0.020656395703554153, + 0.0156935453414917, + 0.016307594254612923, + 0.029454857110977173, + 0.021247990429401398, + 0.018834276124835014, + 0.014627218246459961, + 0.01903228648006916, + 0.012255401350557804, + 0.013228417374193668, + 0.015696339309215546, + 0.015846800059080124, + 0.015194063074886799, + 0.0034594256430864334, + 0.0033806064166128635, + 0.005825246684253216, + 0.002457639668136835, + 0.0040198941715061665, + 0.003634244669228792, + 0.003462475258857012, + 0.0034130054991692305, + 0.007965495809912682, + 0.004767861682921648, + 0.005146597512066364, + 0.00753186596557498, + 0.00796187948435545, + 0.005693752318620682, + 0.008936187252402306, + 0.004916205536574125, + 0.04673779010772705, + 0.023568643257021904, + 0.051990192383527756, + 0.021336818113923073, + 0.025809060782194138, + 0.03807913884520531, + 0.02929428033530712, + 0.027905019000172615, + 0.038137443363666534, + 0.01859917864203453, + 0.023587780073285103, + 0.016574813053011894, + 0.026645390316843987, + 0.02668064460158348, + 0.03410382568836212, + 0.03763972595334053, + 0.005580106284469366, + 0.005564852152019739, + 0.013736546970903873, + 0.006835554726421833, + 0.004148220177739859, + 0.00600366061553359, + 0.005909096449613571, + 0.006373421289026737, + 0.0029942195396870375, + 0.002973536029458046, + 0.0026447304990142584, + 0.0018825778970494866, + 0.00263899564743042, + 0.0017226941417902708, + 0.0019931932911276817, + 0.0019950817804783583, + 0.007382875774055719, + 0.006026675458997488, + 0.004582197871059179, + 0.006414571776986122, + 0.004439090844243765, + 0.005498744547367096, + 0.0034736422821879387, + 0.005582662299275398, + 0.005374837201088667, + 0.004803400486707687, + 0.0041823177598416805, + 0.002861649263650179, + 0.00564833078533411, + 0.004004168789833784, + 0.0028164591640233994, + 0.004141069483011961 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.008813646622002125 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a circular spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spider jet-black with a velvety round body and eight slender hairy legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the spiderweb strands dappled with shimmering dewdrops that refract light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a completely blurred, out-of-focus deep green foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow and steady upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054985325783491135 + ], + "opposite_logprob_tokens": [ + -2.929985284805298 + ], + "logprob_reward": -0.054985325783491135, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054985325783491135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.229404017329216, + 0.16484184563159943, + 0.05411823093891144, + 0.09328927099704742, + 0.05346396565437317, + 0.13339833915233612, + 0.06152094528079033, + 0.08008827269077301, + 0.05955604836344719, + 0.07031910866498947 + ], + "credit_map": [ + 0.02084580808877945, + 0.02415549010038376, + 0.026538060978055, + 0.04181881621479988, + 0.027875935658812523, + 0.022831667214632034, + 0.035473428666591644, + 0.029864797368645668, + 0.023533187806606293, + 0.016700925305485725, + 0.02461542934179306, + 0.01802770420908928, + 0.017843419685959816, + 0.017902923747897148, + 0.024007169529795647, + 0.022211061790585518, + 0.006769968196749687, + 0.005336023401468992, + 0.013903540559113026, + 0.004134822636842728, + 0.006112386006861925, + 0.0053346785716712475, + 0.0066057532094419, + 0.005921054165810347, + 0.013344873674213886, + 0.010366294533014297, + 0.00915804598480463, + 0.00931671354919672, + 0.01223908644169569, + 0.014876202680170536, + 0.012836681678891182, + 0.01115136407315731, + 0.008401268161833286, + 0.004830094054341316, + 0.006219473201781511, + 0.0056423465721309185, + 0.006885170470923185, + 0.006161162164062262, + 0.009227392263710499, + 0.006097050383687019, + 0.024149445816874504, + 0.010552309453487396, + 0.014294111169874668, + 0.006775636225938797, + 0.015791311860084534, + 0.017439618706703186, + 0.020913677290081978, + 0.023482225835323334, + 0.0083604222163558, + 0.00958030205219984, + 0.007738191168755293, + 0.006593232974410057, + 0.006513554602861404, + 0.00883090402930975, + 0.007673858664929867, + 0.006230471655726433, + 0.009683768264949322, + 0.009934835135936737, + 0.012458451092243195, + 0.0060321856290102005, + 0.016243312507867813, + 0.007851368747651577, + 0.010982653126120567, + 0.006901693530380726, + 0.006402893923223019, + 0.005846536252647638, + 0.008137705735862255, + 0.005407274235039949, + 0.00654353154823184, + 0.012120503932237625, + 0.0060381232760846615, + 0.009059479460120201, + 0.008915534242987633, + 0.0075437165796756744, + 0.008822341449558735, + 0.00899426732212305, + 0.007826042361557484, + 0.00715978816151619, + 0.01226884126663208, + 0.008788572624325752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.054985325783491135 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a circular spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spider jet-black with a velvety round body and eight slender hairy legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the spiderweb strands dappled with shimmering dewdrops that refract light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a completely blurred, out-of-focus deep green foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow and steady upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541357427835464 + ], + "opposite_logprob_tokens": [ + -4.3875412940979 + ], + "logprob_reward": -0.012541357427835464, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541357427835464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12710529565811157, + 0.041285060346126556, + 0.11591722071170807, + 0.18277864158153534, + 0.21560296416282654, + 0.09664703160524368, + 0.10388431698083878, + 0.030469505116343498, + 0.05237016826868057, + 0.03393976017832756 + ], + "credit_map": [ + 0.015700718387961388, + 0.013554027304053307, + 0.014424238353967667, + 0.010554123669862747, + 0.014020128175616264, + 0.016640206798911095, + 0.02253379300236702, + 0.01967807114124298, + 0.004540771245956421, + 0.00463259918615222, + 0.006948440335690975, + 0.0038110273890197277, + 0.004672720562666655, + 0.005828788038343191, + 0.006222895812243223, + 0.0046278201043605804, + 0.015132270753383636, + 0.010644197463989258, + 0.018831241875886917, + 0.007280547637492418, + 0.015223351307213306, + 0.019246205687522888, + 0.013104455545544624, + 0.01645496115088463, + 0.02972857654094696, + 0.017015662044286728, + 0.020542358979582787, + 0.01917862892150879, + 0.025824766606092453, + 0.024389512836933136, + 0.024632466956973076, + 0.021466676145792007, + 0.03799695894122124, + 0.01883450709283352, + 0.04699121043086052, + 0.015571770258247852, + 0.021029865369200706, + 0.02983899600803852, + 0.022557944059371948, + 0.022781729698181152, + 0.015259007923305035, + 0.009314410388469696, + 0.010950951837003231, + 0.009376928210258484, + 0.011326239444315434, + 0.011804603971540928, + 0.01346494909375906, + 0.015149950981140137, + 0.009965694509446621, + 0.011235871352255344, + 0.02506842650473118, + 0.00967622920870781, + 0.01111588254570961, + 0.012668185867369175, + 0.01194800902158022, + 0.012206025421619415, + 0.004113577306270599, + 0.0039791204035282135, + 0.0043378183618187904, + 0.00344839901663363, + 0.004082726314663887, + 0.002827220130711794, + 0.00350252166390419, + 0.004178122617304325, + 0.010056082159280777, + 0.005845906678587198, + 0.006691817659884691, + 0.006071735639125109, + 0.006342713721096516, + 0.005787605885416269, + 0.005312067922204733, + 0.0062622432596981525, + 0.0046827360056340694, + 0.004967993125319481, + 0.004073192365467548, + 0.0037968445103615522, + 0.005968297366052866, + 0.003637169022113085, + 0.0034067805390805006, + 0.003406751435250044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.012541357427835464 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a circular spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spider jet-black with a velvety round body and eight slender hairy legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the spiderweb strands dappled with shimmering dewdrops that refract light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a completely blurred, out-of-focus deep green foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow and steady upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00762575538828969 + ], + "opposite_logprob_tokens": [ + -4.882625579833984 + ], + "logprob_reward": -0.00762575538828969, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00762575538828969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1496383249759674, + 0.13858762383460999, + 0.09098200500011444, + 0.24610145390033722, + 0.0652623400092125, + 0.09370658546686172, + 0.060872770845890045, + 0.05657827854156494, + 0.038697242736816406, + 0.05957338586449623 + ], + "credit_map": [ + 0.025862926617264748, + 0.017352210357785225, + 0.019633326679468155, + 0.020562317222356796, + 0.017252560704946518, + 0.012780561111867428, + 0.0187823623418808, + 0.017412059009075165, + 0.020878484472632408, + 0.015048250555992126, + 0.023364653810858727, + 0.012787509709596634, + 0.015302074141800404, + 0.013556133955717087, + 0.021048560738563538, + 0.016601957380771637, + 0.012248788960278034, + 0.008551585488021374, + 0.01159090269356966, + 0.008979110047221184, + 0.013408653438091278, + 0.012206793762743473, + 0.012787552550435066, + 0.011208617128431797, + 0.035825636237859726, + 0.02516336366534233, + 0.026886167004704475, + 0.026985082775354385, + 0.03649472817778587, + 0.03071637824177742, + 0.03665643930435181, + 0.027373651042580605, + 0.009310532361268997, + 0.0063756247982382774, + 0.01304255984723568, + 0.007525905501097441, + 0.006885285954922438, + 0.007017067633569241, + 0.009725157171487808, + 0.005380204413086176, + 0.011936108581721783, + 0.009638695977628231, + 0.010773912072181702, + 0.012408877722918987, + 0.010183272883296013, + 0.010863488540053368, + 0.01300498191267252, + 0.014897252433001995, + 0.00669129379093647, + 0.005659884307533503, + 0.015882812440395355, + 0.006379211787134409, + 0.005608043633401394, + 0.0071744839660823345, + 0.006337273865938187, + 0.007139769848436117, + 0.011603965424001217, + 0.004984561819583178, + 0.010612934827804565, + 0.004159335047006607, + 0.00816316343843937, + 0.004781367722898722, + 0.005334370769560337, + 0.006938580423593521, + 0.00633129570633173, + 0.005002032499760389, + 0.004530272446572781, + 0.005280701443552971, + 0.0043483455665409565, + 0.0041115968488156796, + 0.0041331578977406025, + 0.004959838464856148, + 0.009693924337625504, + 0.007770750671625137, + 0.009470457211136818, + 0.005591453053057194, + 0.008305607363581657, + 0.007184747606515884, + 0.005827993620187044, + 0.005728455726057291 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.00762575538828969 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a circular spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spider jet-black with a velvety round body and eight slender hairy legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the spiderweb strands dappled with shimmering dewdrops that refract light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a completely blurred, out-of-focus deep green foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow and steady upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11290902644395828 + ], + "opposite_logprob_tokens": [ + -2.2379090785980225 + ], + "logprob_reward": -0.11290902644395828, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11290902644395828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048551540821790695, + 0.05397672951221466, + 0.02434387058019638, + 0.03741764649748802, + 0.5387201309204102, + 0.08076402544975281, + 0.025211403146386147, + 0.038017019629478455, + 0.04948950931429863, + 0.10350820422172546 + ], + "credit_map": [ + 0.006025938782840967, + 0.005698331631720066, + 0.004982102196663618, + 0.005857623647898436, + 0.005481267347931862, + 0.0043735564686357975, + 0.009102950803935528, + 0.007029768079519272, + 0.0060676084831357, + 0.008819305337965488, + 0.005411762278527021, + 0.005234808661043644, + 0.006713551003485918, + 0.008818304166197777, + 0.005697239190340042, + 0.007214146200567484, + 0.0030582514591515064, + 0.002097652992233634, + 0.0040896981954574585, + 0.002336275763809681, + 0.004226870369166136, + 0.0025736086536198854, + 0.00334560708142817, + 0.0026159060653299093, + 0.006081098690629005, + 0.0034725030418485403, + 0.0038855678867548704, + 0.003720988519489765, + 0.005599683616310358, + 0.00522523932158947, + 0.005363741889595985, + 0.004068821202963591, + 0.09273117780685425, + 0.04650287702679634, + 0.11879315972328186, + 0.03824830800294876, + 0.05949036404490471, + 0.06748271733522415, + 0.05838025361299515, + 0.05709123983979225, + 0.012736665084958076, + 0.0069067287258803844, + 0.00910460390150547, + 0.006421427708119154, + 0.009008264169096947, + 0.009271577931940556, + 0.01293121837079525, + 0.014383536763489246, + 0.002514360472559929, + 0.003232383169233799, + 0.0040884679183363914, + 0.0022897059097886086, + 0.0036684356164187193, + 0.003926179371774197, + 0.0026304053608328104, + 0.0028614636976271868, + 0.004048231523483992, + 0.004833372309803963, + 0.004436183255165815, + 0.0025714323855936527, + 0.008256149478256702, + 0.0034609378781169653, + 0.0062447888776659966, + 0.004165921825915575, + 0.006803812924772501, + 0.0077933962456882, + 0.00507269287481904, + 0.008685334585607052, + 0.0070556532591581345, + 0.004422873258590698, + 0.005036668386310339, + 0.00461907684803009, + 0.010374401696026325, + 0.010929102078080177, + 0.023310011252760887, + 0.008439753204584122, + 0.013715899549424648, + 0.009629113599658012, + 0.00884498842060566, + 0.01826494187116623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 5 + ], + "selected_reward": -0.11290902644395828 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning macro close-up, a slow and steady upward tilt, focuses on an intricate, circular spiderweb, dappled with shimmering dewdrops in the morning air. The first rays of soft, golden sunlight filter through the web, casting each plump dewdrop like a tiny magnifying glass, refracting a dazzling light. A tiny, glossy, jet-black spider, its velvety, round body contrasting with eight slender, hairy legs, delicately and rhythmically ascends a main thread. Its legs tap the shimmering silk with precise tips, each step causing the web to tremble subtly, sending the dewdrops swaying and shimmering like diamonds. The background is a completely blurred, out-of-focus deep green foliage. This shallow depth of field dramatically accentuates the astonishing details of the spider and web in the foreground. The entire image, framed by the cool background and the warm morning light, is imbued with the tranquility of early morning and the exquisite beauty of nature, creating an atmosphere of focused concentration and peace.", + "seed": 9976, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_6/index_4279_seed_9976_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.037865282502025366, + "vlm_logprob": -0.037865282502025366, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.037865282502025366, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a circular spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spider jet-black with a velvety round body and eight slender hairy legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the spiderweb strands dappled with shimmering dewdrops that refract light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a completely blurred, out-of-focus deep green foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow and steady upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006433016154915094 + ], + "opposite_logprob_tokens": [ + -5.256433010101318 + ], + "logprob_reward": -0.006433016154915094, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.006433016154915094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10902929306030273, + 0.07807176560163498, + 0.0965873971581459, + 0.10789737850427628, + 0.11951139569282532, + 0.19006015360355377, + 0.10929200798273087, + 0.09982027113437653, + 0.04044860601425171, + 0.049281734973192215 + ], + "credit_map": [ + 0.010399334132671356, + 0.012422380968928337, + 0.019879966974258423, + 0.012631474994122982, + 0.01138118002563715, + 0.013157139532268047, + 0.0151745043694973, + 0.013983307406306267, + 0.01058973278850317, + 0.0070156618021428585, + 0.016905885189771652, + 0.008976208977401257, + 0.010280106216669083, + 0.006714218761771917, + 0.007323013618588448, + 0.010266930796205997, + 0.008482432924211025, + 0.009864808060228825, + 0.016787968575954437, + 0.008201166987419128, + 0.007478497922420502, + 0.010417775250971317, + 0.026823319494724274, + 0.008531421422958374, + 0.009239954873919487, + 0.013428892940282822, + 0.012531212531030178, + 0.012161044403910637, + 0.013248042203485966, + 0.020631086081266403, + 0.01629706472158432, + 0.010360072366893291, + 0.010459421202540398, + 0.014055443927645683, + 0.006259164772927761, + 0.012697972357273102, + 0.018703272566199303, + 0.01965579204261303, + 0.021545937284827232, + 0.016134390607476234, + 0.01171551737934351, + 0.02014775201678276, + 0.017280304804444313, + 0.01503865048289299, + 0.01982152834534645, + 0.043361350893974304, + 0.04882396385073662, + 0.013871059753000736, + 0.012449899688363075, + 0.00955423153936863, + 0.013014569878578186, + 0.010169214569032192, + 0.018801243975758553, + 0.018084360286593437, + 0.017324600368738174, + 0.009893875569105148, + 0.010016320273280144, + 0.00927454512566328, + 0.00982721522450447, + 0.021016964688897133, + 0.013157986104488373, + 0.013228224590420723, + 0.010195580311119556, + 0.013103431090712547, + 0.006247228477150202, + 0.0039473120123147964, + 0.005002880003303289, + 0.0039421506226062775, + 0.0028374039102345705, + 0.005541798658668995, + 0.005752841476351023, + 0.007176986429840326, + 0.01131987664848566, + 0.005063164513558149, + 0.0055671678856015205, + 0.004513108637183905, + 0.003593763103708625, + 0.00524681992828846, + 0.007543709594756365, + 0.006434121634811163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.006433016154915094 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a circular spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spider jet-black with a velvety round body and eight slender hairy legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the spiderweb strands dappled with shimmering dewdrops that refract light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a completely blurred, out-of-focus deep green foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow and steady upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867760092020035 + ], + "opposite_logprob_tokens": [ + -3.048677682876587 + ], + "logprob_reward": -0.04867760092020035, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867760092020035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10568007826805115, + 0.3570454716682434, + 0.030648421496152878, + 0.021003173664212227, + 0.04991944134235382, + 0.2124710977077484, + 0.06366093456745148, + 0.0626181960105896, + 0.07000194489955902, + 0.02695120871067047 + ], + "credit_map": [ + 0.012051970697939396, + 0.013616571202874184, + 0.010981299914419651, + 0.013126373291015625, + 0.009758573956787586, + 0.014306038618087769, + 0.014138860628008842, + 0.01770039089024067, + 0.024947764351963997, + 0.013580246828496456, + 0.14441904425621033, + 0.01704307645559311, + 0.10348501801490784, + 0.022224606946110725, + 0.019072452560067177, + 0.012273251079022884, + 0.0029109965544193983, + 0.003164856694638729, + 0.00737672159448266, + 0.002912336029112339, + 0.0026008046697825193, + 0.0035399366170167923, + 0.005485087167471647, + 0.0026576840318739414, + 0.0021705948747694492, + 0.003567690495401621, + 0.0022075623273849487, + 0.0025060256011784077, + 0.0023023057729005814, + 0.002680055797100067, + 0.003030588151887059, + 0.00253835110925138, + 0.005142465233802795, + 0.006101439706981182, + 0.00353831285610795, + 0.00503844628110528, + 0.009356830269098282, + 0.006628359202295542, + 0.00803126860409975, + 0.006082317791879177, + 0.012170481495559216, + 0.010328116826713085, + 0.017899103462696075, + 0.012922688387334347, + 0.014694921672344208, + 0.060137756168842316, + 0.06970704346895218, + 0.01461098063737154, + 0.005663271062076092, + 0.004667457193136215, + 0.008517500944435596, + 0.008082631044089794, + 0.005962314084172249, + 0.0102385887876153, + 0.01009688712656498, + 0.010432288981974125, + 0.006420542020350695, + 0.007907561957836151, + 0.005618820898234844, + 0.011847216635942459, + 0.007919680327177048, + 0.008927891962230206, + 0.0075647663325071335, + 0.00641171308234334, + 0.006461655721068382, + 0.014056483283638954, + 0.00754478620365262, + 0.006507616024464369, + 0.012120087631046772, + 0.008606676943600178, + 0.008803830482065678, + 0.005900808610022068, + 0.007690743077546358, + 0.0025277009699493647, + 0.003522450802847743, + 0.0019502427894622087, + 0.0019377870485186577, + 0.0027745382394641638, + 0.003524202387779951, + 0.003023544093593955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.04867760092020035 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a circular spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spider jet-black with a velvety round body and eight slender hairy legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the spiderweb strands dappled with shimmering dewdrops that refract light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a completely blurred, out-of-focus deep green foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow and steady upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251660011708736 + ], + "opposite_logprob_tokens": [ + -5.255251884460449 + ], + "logprob_reward": -0.005251660011708736, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251660011708736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10641036182641983, + 0.05931486561894417, + 0.04437090456485748, + 0.06184697523713112, + 0.1281658560037613, + 0.3887670934200287, + 0.03345750644803047, + 0.0695842057466507, + 0.06953812390565872, + 0.03854409605264664 + ], + "credit_map": [ + 0.010426538065075874, + 0.012470518238842487, + 0.015175375156104565, + 0.012165647931396961, + 0.010999314486980438, + 0.016650022938847542, + 0.011130179278552532, + 0.01739276759326458, + 0.005104546435177326, + 0.006398437544703484, + 0.014261704869568348, + 0.004937996622174978, + 0.009500527754426003, + 0.005105202551931143, + 0.0051076398231089115, + 0.00889880582690239, + 0.004285536706447601, + 0.0036613037809729576, + 0.007485703565180302, + 0.0038566607981920242, + 0.004030448850244284, + 0.005066183395683765, + 0.011036846786737442, + 0.0049482244066894054, + 0.004690511617809534, + 0.0096701979637146, + 0.007995834574103355, + 0.005005367565900087, + 0.006134512834250927, + 0.012778250500559807, + 0.008454406633973122, + 0.007117897272109985, + 0.010360795073211193, + 0.01028409879654646, + 0.010298938490450382, + 0.011808906681835651, + 0.011500272899866104, + 0.025860026478767395, + 0.027105679735541344, + 0.020947126671671867, + 0.020863529294729233, + 0.020630573853850365, + 0.02713158167898655, + 0.02314741164445877, + 0.028074398636817932, + 0.11294058710336685, + 0.12957218289375305, + 0.02640683203935623, + 0.0033439903054386377, + 0.002639530226588249, + 0.003922194708138704, + 0.004002649802714586, + 0.004391627851873636, + 0.004877029452472925, + 0.00501516368240118, + 0.005265322048217058, + 0.008412143215537071, + 0.006461914628744125, + 0.006787916179746389, + 0.013342093676328659, + 0.00903560221195221, + 0.009339291602373123, + 0.007380032446235418, + 0.00882520992308855, + 0.011262111365795135, + 0.010981961153447628, + 0.007803876884281635, + 0.007669089362025261, + 0.007543583400547504, + 0.009686515666544437, + 0.007789227180182934, + 0.006801762618124485, + 0.01050123292952776, + 0.003193662501871586, + 0.00457256892696023, + 0.003001068253070116, + 0.0034498225431889296, + 0.004626400303095579, + 0.004879065323621035, + 0.00432027829810977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.005251660011708736 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a circular spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spider jet-black with a velvety round body and eight slender hairy legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the spiderweb strands dappled with shimmering dewdrops that refract light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a completely blurred, out-of-focus deep green foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow and steady upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016065435484051704 + ], + "opposite_logprob_tokens": [ + -4.14106559753418 + ], + "logprob_reward": -0.016065435484051704, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016065435484051704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08047816902399063, + 0.4320186674594879, + 0.04743397235870361, + 0.04602830484509468, + 0.03064795769751072, + 0.10327933728694916, + 0.058653973042964935, + 0.14751753211021423, + 0.04131012409925461, + 0.01263197511434555 + ], + "credit_map": [ + 0.007329105399549007, + 0.009793149307370186, + 0.011443201452493668, + 0.010037313215434551, + 0.007422729395329952, + 0.014914022758603096, + 0.009776115417480469, + 0.009762524627149105, + 0.0326843187212944, + 0.01426326297223568, + 0.17130990326404572, + 0.020438771694898605, + 0.12332602590322495, + 0.02673698402941227, + 0.02394014783203602, + 0.019319240003824234, + 0.004196619614958763, + 0.008222388103604317, + 0.00778442807495594, + 0.004500250797718763, + 0.004437005612999201, + 0.007911029271781445, + 0.006263307761400938, + 0.00411894591525197, + 0.0038131342735141516, + 0.004444502294063568, + 0.004584876354783773, + 0.00393094215542078, + 0.005644555669277906, + 0.005877077579498291, + 0.008679797872900963, + 0.009053423069417477, + 0.002850170014426112, + 0.002744851168245077, + 0.0028837358113378286, + 0.0033533170353621244, + 0.0023418792989104986, + 0.004946904256939888, + 0.006111158058047295, + 0.005415941588580608, + 0.01059676893055439, + 0.007208716589957476, + 0.017276225611567497, + 0.009646538645029068, + 0.010737555101513863, + 0.017046581953763962, + 0.0183915626257658, + 0.012375381775200367, + 0.005843964405357838, + 0.004365659784525633, + 0.004864652641117573, + 0.006953978445380926, + 0.007561382371932268, + 0.011245517060160637, + 0.010576694272458553, + 0.007242124527692795, + 0.0146026024594903, + 0.013850856572389603, + 0.013674692250788212, + 0.0334576778113842, + 0.021828262135386467, + 0.016582904383540154, + 0.014630665071308613, + 0.01888987421989441, + 0.005964663345366716, + 0.004779726732522249, + 0.003920828457921743, + 0.0034615567419677973, + 0.0062063876539468765, + 0.005655235610902309, + 0.005825949367135763, + 0.005495778750628233, + 0.0020314485300332308, + 0.0014210592489689589, + 0.0014254320412874222, + 0.001443859189748764, + 0.0013124864781275392, + 0.001487645786255598, + 0.0015086286002770066, + 0.002001415938138962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 5 + ], + "selected_reward": -0.016065435484051704 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a circular spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spider jet-black with a velvety round body and eight slender hairy legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the spiderweb strands dappled with shimmering dewdrops that refract light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a completely blurred, out-of-focus deep green foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow and steady upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11289869993925095 + ], + "opposite_logprob_tokens": [ + -2.237898588180542 + ], + "logprob_reward": -0.11289869993925095, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11289869993925095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15078052878379822, + 0.1473408043384552, + 0.04491281509399414, + 0.1181139349937439, + 0.034326303750276566, + 0.20774437487125397, + 0.06456587463617325, + 0.08504632115364075, + 0.07490382343530655, + 0.07226517796516418 + ], + "credit_map": [ + 0.013638496398925781, + 0.012919646687805653, + 0.014226836152374744, + 0.01418473944067955, + 0.01621531881392002, + 0.04157114773988724, + 0.014452380128204823, + 0.023571979254484177, + 0.011073251254856586, + 0.007535818964242935, + 0.05360139533877373, + 0.00788912270218134, + 0.03878209739923477, + 0.009788976982235909, + 0.009861258789896965, + 0.00880888756364584, + 0.004858012776821852, + 0.0041098156943917274, + 0.007041286211460829, + 0.00443614274263382, + 0.005255978088825941, + 0.006007099058479071, + 0.008724928833544254, + 0.0044795554131269455, + 0.013673437759280205, + 0.017728956416249275, + 0.014840888790786266, + 0.011796525679528713, + 0.014314629137516022, + 0.01712578348815441, + 0.01874917931854725, + 0.00988454557955265, + 0.003163798013702035, + 0.003726170863956213, + 0.003150064731016755, + 0.0049026478081941605, + 0.0036228972021490335, + 0.004418551456183195, + 0.00579998129978776, + 0.005542194936424494, + 0.012636125087738037, + 0.016902221366763115, + 0.017028426751494408, + 0.014352534897625446, + 0.015527787618339062, + 0.05520561710000038, + 0.060927558690309525, + 0.015164121985435486, + 0.0073974221013486385, + 0.005274347960948944, + 0.008815258741378784, + 0.007951364852488041, + 0.008316019549965858, + 0.00924368854612112, + 0.009317701682448387, + 0.008250079117715359, + 0.009393676184117794, + 0.009407352656126022, + 0.008053121156990528, + 0.016677377745509148, + 0.012950560078024864, + 0.01168063934892416, + 0.009979873895645142, + 0.006903727073222399, + 0.008571242913603783, + 0.015172486193478107, + 0.008387000299990177, + 0.007287570275366306, + 0.008055792190134525, + 0.007604426238685846, + 0.012653131037950516, + 0.00717218266800046, + 0.01956121250987053, + 0.006296235602349043, + 0.009041358716785908, + 0.0057868813164532185, + 0.005674939602613449, + 0.009351015090942383, + 0.00869246106594801, + 0.007861083373427391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.11289869993925095 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_7/index_tensor([6114, 6114])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_7/index_tensor([6114, 6114])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0961c2036de2aac769ee7c71d0f1587696e72463 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_7/index_tensor([6114, 6114])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a documentary-style wide shot, the camera slowly and deliberately pans from left to right across a vast, cavernous metallurgical workshop, its high-beamed ceiling vanishing into industrial haze and shadows. The primary light source is a massive steel ingot, emitting an intense, pulsating orange-white heat that casts a fiery, dramatic glow and long, dancing shadows across the scene. This ingot slides relentlessly across a massive roller bed, gliding toward the center of the frame, where a towering industrial press awaits, constructed of heavy, dark, greased, and pitted steel. As the glowing metal enters the press's chute, two massive cylindrical rollers come into contact, exerting incredible pressure that visibly thickens and elongates the ingot. This action triggers a violent and spectacular eruption of vibrant yellow and white sparks, which blast outward in a chaotic shower, momentarily illuminating the pristine, functional details of the surrounding machinery. The entire process is captured in crisp, authentic detail, emphasising the sheer scale and raw power of the operation as once-thick steel is ruthlessly transformed into thinner, longer plates, creating a breathtaking and intense atmosphere of industrial might.", + "seed": 11110, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_7/index_6114_seed_11110_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.033212058385834096, + "vlm_logprob": -0.033212058385834096, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.033212058385834096, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous metallurgical workshop with a high-beamed ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and deliberately pan from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive steel ingot visible emitting an intense, pulsating orange-white heat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ingot slide across a roller bed toward a towering industrial press and get compressed by rollers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a violent eruption of vibrant yellow and white sparks when the ingot enters the press?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018473710864782333 + ], + "opposite_logprob_tokens": [ + -4.3934736251831055 + ], + "logprob_reward": -0.018473710864782333, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.018473710864782333, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23071695864200592, + 0.06369329243898392, + 0.3878169357776642, + 0.08384137600660324, + 0.04296368360519409, + 0.034620389342308044, + 0.0337798148393631, + 0.07653474062681198, + 0.011208509095013142, + 0.03482428565621376 + ], + "credit_map": [ + 0.020964745432138443, + 0.027149450033903122, + 0.01500717643648386, + 0.012451071292161942, + 0.030050983652472496, + 0.04837863892316818, + 0.027178751304745674, + 0.04953611642122269, + 0.006094174925237894, + 0.013779464177787304, + 0.005251019727438688, + 0.0042686560191214085, + 0.014723642729222775, + 0.008037648163735867, + 0.005323825404047966, + 0.006214862689375877, + 0.023865437135100365, + 0.04279983788728714, + 0.05695148929953575, + 0.03636275604367256, + 0.05691191926598549, + 0.05702934041619301, + 0.06685769557952881, + 0.04703846573829651, + 0.006614259444177151, + 0.010368501767516136, + 0.011141994036734104, + 0.007800602354109287, + 0.018505576997995377, + 0.01246216893196106, + 0.006772420834749937, + 0.010175853967666626, + 0.011561555787920952, + 0.006208108738064766, + 0.006514021661132574, + 0.005016801413148642, + 0.004337830934673548, + 0.0034861534368246794, + 0.0022226343862712383, + 0.0036165763158351183, + 0.005118147470057011, + 0.006371976807713509, + 0.0060888295993208885, + 0.003714144928380847, + 0.003883480792865157, + 0.002957886317744851, + 0.0028984909877181053, + 0.0035874343011528254, + 0.00479046581313014, + 0.006176695227622986, + 0.005037510767579079, + 0.005551357287913561, + 0.0036960819270461798, + 0.0018861599965021014, + 0.0011128545738756657, + 0.005528689362108707, + 0.011596553958952427, + 0.02709910273551941, + 0.010876131244003773, + 0.011954297311604023, + 0.004891381599009037, + 0.0029763805214315653, + 0.004292869009077549, + 0.002848027041181922, + 0.001193992211483419, + 0.0017621967708691955, + 0.0013040249468758702, + 0.0013481297064572573, + 0.0011903007980436087, + 0.001570741180330515, + 0.0014867569552734494, + 0.0013523666420951486, + 0.004702980630099773, + 0.00908228661864996, + 0.0018972895340994, + 0.002044256078079343, + 0.0071901618503034115, + 0.004624234978109598, + 0.002676185453310609, + 0.002606890397146344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.018473710864782333 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous metallurgical workshop with a high-beamed ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and deliberately pan from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive steel ingot visible emitting an intense, pulsating orange-white heat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ingot slide across a roller bed toward a towering industrial press and get compressed by rollers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a violent eruption of vibrant yellow and white sparks when the ingot enters the press?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899561524391174 + ], + "opposite_logprob_tokens": [ + -2.578995704650879 + ], + "logprob_reward": -0.07899561524391174, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899561524391174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0733451172709465, + 0.08358851820230484, + 0.17727479338645935, + 0.07462555915117264, + 0.10895633697509766, + 0.07010434567928314, + 0.06153951212763786, + 0.18925230205059052, + 0.09903518110513687, + 0.062278300523757935 + ], + "credit_map": [ + 0.006181454751640558, + 0.00933720264583826, + 0.005704506300389767, + 0.006036825943738222, + 0.010475428774952888, + 0.011898865923285484, + 0.008969517424702644, + 0.014741308987140656, + 0.0072393096052110195, + 0.013571435585618019, + 0.010198957286775112, + 0.00999197643250227, + 0.012709848582744598, + 0.01052210945636034, + 0.009494094178080559, + 0.009860791265964508, + 0.016955632716417313, + 0.018672626465559006, + 0.019186001271009445, + 0.01937612146139145, + 0.022459739819169044, + 0.024972965940833092, + 0.033813707530498505, + 0.021837998181581497, + 0.009897689335048199, + 0.00889587588608265, + 0.00916246883571148, + 0.007505467627197504, + 0.013073647394776344, + 0.011304624378681183, + 0.0070072319358587265, + 0.007778555620461702, + 0.020671386271715164, + 0.02058574929833412, + 0.015765510499477386, + 0.010504577308893204, + 0.014891361817717552, + 0.01080238539725542, + 0.006750533357262611, + 0.008984834887087345, + 0.010637558996677399, + 0.011799827218055725, + 0.015222539193928242, + 0.008470137603580952, + 0.006162545178085566, + 0.005706173367798328, + 0.0044936141930520535, + 0.0076119545847177505, + 0.008063730783760548, + 0.009653223678469658, + 0.011621701531112194, + 0.00880711805075407, + 0.006220838986337185, + 0.0036055443342775106, + 0.005369592923671007, + 0.008197763003408909, + 0.01663823612034321, + 0.05740031599998474, + 0.015926266089081764, + 0.052462901920080185, + 0.01148911565542221, + 0.010289464145898819, + 0.016274485737085342, + 0.00877151545137167, + 0.010680928826332092, + 0.020004697144031525, + 0.008710900321602821, + 0.007917022332549095, + 0.01177408266812563, + 0.018087029457092285, + 0.011303761973977089, + 0.01055675558745861, + 0.005638055503368378, + 0.011762765236198902, + 0.006976604461669922, + 0.005116356536746025, + 0.009156419895589352, + 0.008964264765381813, + 0.006797492504119873, + 0.007866340689361095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 4 + ], + "selected_reward": -0.07899561524391174 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous metallurgical workshop with a high-beamed ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and deliberately pan from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive steel ingot visible emitting an intense, pulsating orange-white heat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ingot slide across a roller bed toward a towering industrial press and get compressed by rollers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a violent eruption of vibrant yellow and white sparks when the ingot enters the press?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00118900160305202 + ], + "opposite_logprob_tokens": [ + -6.751189231872559 + ], + "logprob_reward": -0.00118900160305202, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.00118900160305202, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10181176662445068, + 0.059294912964105606, + 0.18934497237205505, + 0.13532036542892456, + 0.0648866817355156, + 0.08220717310905457, + 0.09377262741327286, + 0.18712981045246124, + 0.04942774027585983, + 0.03680403530597687 + ], + "credit_map": [ + 0.009415498934686184, + 0.016578152775764465, + 0.00708680460229516, + 0.007622789591550827, + 0.011696058325469494, + 0.01692827045917511, + 0.01070205308496952, + 0.02178211510181427, + 0.00646443385630846, + 0.007365908473730087, + 0.009332174435257912, + 0.005883923266083002, + 0.008497023954987526, + 0.007082488853484392, + 0.00686681130900979, + 0.007802134845405817, + 0.012432139366865158, + 0.017135122790932655, + 0.03367234766483307, + 0.01624627411365509, + 0.024682598188519478, + 0.02898629754781723, + 0.030821077525615692, + 0.025369077920913696, + 0.012848760932683945, + 0.019597655162215233, + 0.01292375661432743, + 0.01420010905712843, + 0.02987835742533207, + 0.020573848858475685, + 0.012507341802120209, + 0.012790516950190067, + 0.007943278178572655, + 0.008251393213868141, + 0.010546699166297913, + 0.008491486310958862, + 0.010547789745032787, + 0.005766021087765694, + 0.007836567237973213, + 0.005503436084836721, + 0.01128124725073576, + 0.016518937423825264, + 0.0138582494109869, + 0.008398025296628475, + 0.009151231497526169, + 0.007312331348657608, + 0.007082889787852764, + 0.008604246191680431, + 0.019135400652885437, + 0.016193661838769913, + 0.015730956569314003, + 0.013251572847366333, + 0.008954236283898354, + 0.005547450389713049, + 0.004521480295807123, + 0.01043785735964775, + 0.028813792392611504, + 0.0437302403151989, + 0.024851743131875992, + 0.02934129163622856, + 0.018593937158584595, + 0.013270915485918522, + 0.015107909217476845, + 0.01341994758695364, + 0.005794038064777851, + 0.00880123395472765, + 0.005637590307742357, + 0.003697149222716689, + 0.006546483840793371, + 0.011580238118767738, + 0.003263145685195923, + 0.004107849672436714, + 0.005794191733002663, + 0.009001011028885841, + 0.0023235527332872152, + 0.0019188887672498822, + 0.0064386045560240746, + 0.004077930469065905, + 0.0037309241015464067, + 0.0035189285408705473 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 3 + ], + "selected_reward": -0.00118900160305202 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous metallurgical workshop with a high-beamed ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and deliberately pan from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive steel ingot visible emitting an intense, pulsating orange-white heat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ingot slide across a roller bed toward a towering industrial press and get compressed by rollers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a violent eruption of vibrant yellow and white sparks when the ingot enters the press?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370409458875656 + ], + "opposite_logprob_tokens": [ + -3.4087040424346924 + ], + "logprob_reward": -0.03370409458875656, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370409458875656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07502994686365128, + 0.0673767477273941, + 0.19248545169830322, + 0.09707767516374588, + 0.08850444108247757, + 0.11075980216264725, + 0.06546001881361008, + 0.2116541564464569, + 0.05385429039597511, + 0.037797391414642334 + ], + "credit_map": [ + 0.007534993812441826, + 0.014827326871454716, + 0.005756077356636524, + 0.006599385756999254, + 0.01367940567433834, + 0.009376553818583488, + 0.009057731367647648, + 0.008198474533855915, + 0.007006141357123852, + 0.010074006393551826, + 0.0103001082316041, + 0.004832091275602579, + 0.010756555013358593, + 0.008512995205819607, + 0.006275547668337822, + 0.009619307704269886, + 0.01245214231312275, + 0.022916026413440704, + 0.02387392893433571, + 0.018213549628853798, + 0.03287578374147415, + 0.02947482280433178, + 0.03139792010188103, + 0.0212812889367342, + 0.00904908962547779, + 0.012080679647624493, + 0.009484115056693554, + 0.012580358423292637, + 0.01910313218832016, + 0.014631079509854317, + 0.010325386188924313, + 0.009823838248848915, + 0.01211280096322298, + 0.012775384820997715, + 0.012097946368157864, + 0.01340933796018362, + 0.012691429816186428, + 0.007916352711617947, + 0.009274625219404697, + 0.008226564154028893, + 0.011226913891732693, + 0.020692359656095505, + 0.01779940538108349, + 0.011637515388429165, + 0.013485914096236229, + 0.011278912425041199, + 0.010953603312373161, + 0.013685165904462337, + 0.006118743680417538, + 0.010836958885192871, + 0.010165686719119549, + 0.009156804531812668, + 0.010099302977323532, + 0.006043347995728254, + 0.004534150008112192, + 0.008505022153258324, + 0.03637602552771568, + 0.02692292630672455, + 0.033168718218803406, + 0.04455840215086937, + 0.021913062781095505, + 0.015063819475471973, + 0.01875741221010685, + 0.014893811196088791, + 0.005708498880267143, + 0.01309678889811039, + 0.004889520816504955, + 0.0053774104453623295, + 0.006660275626927614, + 0.00935744121670723, + 0.0044655222445726395, + 0.004298829939216375, + 0.005775706376880407, + 0.00705271540209651, + 0.003267238149419427, + 0.003324304474517703, + 0.005108972080051899, + 0.005551304202526808, + 0.0035708914510905743, + 0.004146257881075144 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 5 + ], + "selected_reward": -0.03370409458875656 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous metallurgical workshop with a high-beamed ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and deliberately pan from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive steel ingot visible emitting an intense, pulsating orange-white heat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ingot slide across a roller bed toward a towering industrial press and get compressed by rollers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a violent eruption of vibrant yellow and white sparks when the ingot enters the press?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369786962866783 + ], + "opposite_logprob_tokens": [ + -3.4086978435516357 + ], + "logprob_reward": -0.03369786962866783, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369786962866783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07368098944425583, + 0.03869134932756424, + 0.10845346003770828, + 0.041854918003082275, + 0.11514967679977417, + 0.048798929899930954, + 0.09120918810367584, + 0.2246926724910736, + 0.21368198096752167, + 0.04378677159547806 + ], + "credit_map": [ + 0.006359780672937632, + 0.005870587658137083, + 0.005356828682124615, + 0.005909178406000137, + 0.007553506642580032, + 0.016579607501626015, + 0.011655508540570736, + 0.01439599972218275, + 0.0052398135885596275, + 0.006814639084041119, + 0.0035892219748347998, + 0.003516846802085638, + 0.008251013234257698, + 0.0053763301111757755, + 0.002879417734220624, + 0.003024069359526038, + 0.006748152431100607, + 0.012941215187311172, + 0.015630904585123062, + 0.01116916723549366, + 0.014345642179250717, + 0.01652343198657036, + 0.01941560208797455, + 0.011679352261126041, + 0.003508171997964382, + 0.005027834791690111, + 0.005439125467091799, + 0.006152397487312555, + 0.006037133280187845, + 0.006194238550961018, + 0.004001494497060776, + 0.005494526121765375, + 0.01107901893556118, + 0.01173991709947586, + 0.02460452727973461, + 0.016475390642881393, + 0.019054561853408813, + 0.01016298495233059, + 0.012547175399959087, + 0.00948611181229353, + 0.009100589901208878, + 0.011023441329598427, + 0.00653636222705245, + 0.005741432309150696, + 0.0051969303749501705, + 0.00343124195933342, + 0.0034812490921467543, + 0.0042876857332885265, + 0.01333787851035595, + 0.01701485738158226, + 0.01596931740641594, + 0.012123558670282364, + 0.00997236743569374, + 0.005556702148169279, + 0.0038641218561679125, + 0.013370389118790627, + 0.025368398055434227, + 0.07351302355527878, + 0.026829563081264496, + 0.05018090456724167, + 0.015405282378196716, + 0.01155410148203373, + 0.012185794301331043, + 0.009655630216002464, + 0.02342446893453598, + 0.03301059827208519, + 0.024509932845830917, + 0.01670379377901554, + 0.028885550796985626, + 0.05493941158056259, + 0.014486690983176231, + 0.01772153750061989, + 0.005583042744547129, + 0.010898381471633911, + 0.0034822621382772923, + 0.002716870978474617, + 0.009342022240161896, + 0.0053797210566699505, + 0.0036836424842476845, + 0.0027008308097720146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 4 + ], + "selected_reward": -0.03369786962866783 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a documentary-style wide shot, the camera slowly and deliberately pans from left to right across a vast, cavernous metallurgical workshop, its high-beamed ceiling vanishing into industrial haze and shadows. The primary light source is a massive steel ingot, emitting an intense, pulsating orange-white heat that casts a fiery, dramatic glow and long, dancing shadows across the scene. This ingot slides relentlessly across a massive roller bed, gliding toward the center of the frame, where a towering industrial press awaits, constructed of heavy, dark, greased, and pitted steel. As the glowing metal enters the press's chute, two massive cylindrical rollers come into contact, exerting incredible pressure that visibly thickens and elongates the ingot. This action triggers a violent and spectacular eruption of vibrant yellow and white sparks, which blast outward in a chaotic shower, momentarily illuminating the pristine, functional details of the surrounding machinery. The entire process is captured in crisp, authentic detail, emphasising the sheer scale and raw power of the operation as once-thick steel is ruthlessly transformed into thinner, longer plates, creating a breathtaking and intense atmosphere of industrial might.", + "seed": 11111, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_7/index_6114_seed_11111_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04973130756989121, + "vlm_logprob": -0.04973130756989121, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.04973130756989121, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous metallurgical workshop with a high-beamed ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and deliberately pan from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive steel ingot visible emitting an intense, pulsating orange-white heat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ingot slide across a roller bed toward a towering industrial press and get compressed by rollers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a violent eruption of vibrant yellow and white sparks when the ingot enters the press?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05273779481649399 + ], + "opposite_logprob_tokens": [ + -3.1777377128601074 + ], + "logprob_reward": -0.05273779481649399, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.05273779481649399, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14554829895496368, + 0.27924448251724243, + 0.032320037484169006, + 0.12672659754753113, + 0.10089278966188431, + 0.0430019237101078, + 0.03667053207755089, + 0.03645014390349388, + 0.1610468327999115, + 0.03809833899140358 + ], + "credit_map": [ + 0.019795777276158333, + 0.017043596133589745, + 0.019028570502996445, + 0.012197134084999561, + 0.028357654809951782, + 0.020781299099326134, + 0.011671333573758602, + 0.016672927886247635, + 0.04044824838638306, + 0.036957670003175735, + 0.028670357540249825, + 0.03143933787941933, + 0.03806522116065025, + 0.04370734840631485, + 0.027601098641753197, + 0.032355185598134995, + 0.003427276387810707, + 0.004765704739838839, + 0.003196445992216468, + 0.0028536671306937933, + 0.003966310527175665, + 0.006393801886588335, + 0.0031932187266647816, + 0.004523609299212694, + 0.017766179516911507, + 0.0208443570882082, + 0.011060129851102829, + 0.009293296374380589, + 0.02469838596880436, + 0.017637671902775764, + 0.00938639510422945, + 0.016040189191699028, + 0.017385466024279594, + 0.01678265444934368, + 0.009461344219744205, + 0.008292417041957378, + 0.016154849901795387, + 0.011580951511859894, + 0.007402776274830103, + 0.013832329772412777, + 0.00802415981888771, + 0.006714036222547293, + 0.004032038152217865, + 0.0026349620893597603, + 0.008569145575165749, + 0.004722991492599249, + 0.0019636701326817274, + 0.006340921390801668, + 0.005584250669926405, + 0.005208837799727917, + 0.005059844348579645, + 0.003997844643890858, + 0.005966546013951302, + 0.004361982923001051, + 0.002653030678629875, + 0.0038381984923034906, + 0.004768504295498133, + 0.00758872926235199, + 0.006184936501085758, + 0.003794850781559944, + 0.0027994662523269653, + 0.003861525561660528, + 0.0026205675676465034, + 0.0048315636813640594, + 0.033745177090168, + 0.019025437533855438, + 0.007142062298953533, + 0.0054000611416995525, + 0.006612430792301893, + 0.031366512179374695, + 0.0058006541803479195, + 0.051954496651887894, + 0.0075598773546516895, + 0.007309579756110907, + 0.0037997437175363302, + 0.004132064059376717, + 0.004738445859402418, + 0.003031268948689103, + 0.003526445711031556, + 0.004000912886112928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.05273779481649399 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous metallurgical workshop with a high-beamed ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and deliberately pan from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive steel ingot visible emitting an intense, pulsating orange-white heat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ingot slide across a roller bed toward a towering industrial press and get compressed by rollers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a violent eruption of vibrant yellow and white sparks when the ingot enters the press?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033771395683289 + ], + "opposite_logprob_tokens": [ + -2.3503377437591553 + ], + "logprob_reward": -0.10033771395683289, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033771395683289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12347595393657684, + 0.17625249922275543, + 0.06859234720468521, + 0.05939777195453644, + 0.07915467023849487, + 0.11328476667404175, + 0.08722028881311417, + 0.10190553218126297, + 0.14906303584575653, + 0.04165305197238922 + ], + "credit_map": [ + 0.018182456493377686, + 0.016406569629907608, + 0.011574224568903446, + 0.007468029856681824, + 0.029280005022883415, + 0.018016744405031204, + 0.008958647958934307, + 0.01358927134424448, + 0.025280965492129326, + 0.02186412550508976, + 0.014309880323708057, + 0.02129816636443138, + 0.02209370769560337, + 0.02783927321434021, + 0.0185950119048357, + 0.024971365928649902, + 0.007399279158562422, + 0.008789931423962116, + 0.0062806447967886925, + 0.007058735471218824, + 0.012684273533523083, + 0.011165431700646877, + 0.006462180521339178, + 0.008751871064305305, + 0.010487259365618229, + 0.009844370186328888, + 0.005256138741970062, + 0.004604964051395655, + 0.008942991495132446, + 0.006428400985896587, + 0.00456945039331913, + 0.009264198131859303, + 0.016875898465514183, + 0.016204513609409332, + 0.006911390461027622, + 0.004910814110189676, + 0.012612796388566494, + 0.00880291499197483, + 0.0031801112927496433, + 0.009656229987740517, + 0.023756954818964005, + 0.01557314582169056, + 0.015859294682741165, + 0.007153053767979145, + 0.01589922606945038, + 0.014689473435282707, + 0.008027258329093456, + 0.012326357886195183, + 0.01212781947106123, + 0.013225606642663479, + 0.0109077924862504, + 0.008843719027936459, + 0.013781317509710789, + 0.008490177802741528, + 0.01022428274154663, + 0.009619572199881077, + 0.012551684863865376, + 0.011345366016030312, + 0.01872287504374981, + 0.010053517296910286, + 0.008647923357784748, + 0.010035138577222824, + 0.009519416838884354, + 0.021029604598879814, + 0.02592785656452179, + 0.02518978901207447, + 0.010035735554993153, + 0.01001326646655798, + 0.021138355135917664, + 0.018361685797572136, + 0.011523905210196972, + 0.026872439309954643, + 0.007079732604324818, + 0.007491475436836481, + 0.007450910285115242, + 0.00399785116314888, + 0.0045394813641905785, + 0.0035474724136292934, + 0.0036108025815337896, + 0.00393532682210207 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.10033771395683289 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous metallurgical workshop with a high-beamed ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and deliberately pan from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive steel ingot visible emitting an intense, pulsating orange-white heat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ingot slide across a roller bed toward a towering industrial press and get compressed by rollers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a violent eruption of vibrant yellow and white sparks when the ingot enters the press?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002501931507140398 + ], + "opposite_logprob_tokens": [ + -6.002501964569092 + ], + "logprob_reward": -0.002501931507140398, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002501931507140398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16879424452781677, + 0.19336308538913727, + 0.06788481026887894, + 0.03320109471678734, + 0.03547107055783272, + 0.030540093779563904, + 0.04486992955207825, + 0.062053095549345016, + 0.3378545045852661, + 0.025968024507164955 + ], + "credit_map": [ + 0.01558529119938612, + 0.016911007463932037, + 0.02197965979576111, + 0.017850743606686592, + 0.02655795030295849, + 0.026340946555137634, + 0.020868808031082153, + 0.02269984595477581, + 0.027056921273469925, + 0.02215392515063286, + 0.01781775802373886, + 0.02085217833518982, + 0.02772107720375061, + 0.03370120748877525, + 0.01839568093419075, + 0.02566433884203434, + 0.008253945969045162, + 0.008911343291401863, + 0.007000766694545746, + 0.008054292760789394, + 0.010348731651902199, + 0.008173633366823196, + 0.007343775127083063, + 0.009798327460885048, + 0.004864226561039686, + 0.003803447587415576, + 0.003173637669533491, + 0.003315263194963336, + 0.004234153311699629, + 0.00435596564784646, + 0.0030942705925554037, + 0.006360131781548262, + 0.0053921714425086975, + 0.0057540163397789, + 0.0030492821242660284, + 0.0030635239090770483, + 0.005732415243983269, + 0.004391905386000872, + 0.002918823156505823, + 0.0051689366810023785, + 0.005825613625347614, + 0.004347097594290972, + 0.0061815641820430756, + 0.002217488130554557, + 0.004257408436387777, + 0.0036518918350338936, + 0.0016175437485799193, + 0.002441487042233348, + 0.004961096681654453, + 0.007093122694641352, + 0.005701084621250629, + 0.004717948846518993, + 0.009681189432740211, + 0.005021616350859404, + 0.003148781368508935, + 0.004545089788734913, + 0.0057234372943639755, + 0.009339902549982071, + 0.01457736361771822, + 0.004550306126475334, + 0.0047247400507330894, + 0.006066015921533108, + 0.004954242147505283, + 0.01211709063500166, + 0.06759782880544662, + 0.037325385957956314, + 0.013331501744687557, + 0.010835469700396061, + 0.018482480198144913, + 0.06644636392593384, + 0.012237200513482094, + 0.11159828305244446, + 0.003606748767197132, + 0.00517865689471364, + 0.003981537651270628, + 0.0026984531432390213, + 0.0032759394962340593, + 0.0026833845768123865, + 0.0019086870597675443, + 0.002634617732837796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.002501931507140398 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous metallurgical workshop with a high-beamed ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and deliberately pan from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive steel ingot visible emitting an intense, pulsating orange-white heat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ingot slide across a roller bed toward a towering industrial press and get compressed by rollers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a violent eruption of vibrant yellow and white sparks when the ingot enters the press?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038109004497528076 + ], + "opposite_logprob_tokens": [ + -3.288109064102173 + ], + "logprob_reward": -0.038109004497528076, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038109004497528076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11282948404550552, + 0.15993885695934296, + 0.08249843120574951, + 0.050458285957574844, + 0.061485107988119125, + 0.05440093204379082, + 0.09599830210208893, + 0.11706064641475677, + 0.22266721725463867, + 0.042662687599658966 + ], + "credit_map": [ + 0.01307609025388956, + 0.009152889251708984, + 0.01547801960259676, + 0.01146390289068222, + 0.019257327541708946, + 0.01932010054588318, + 0.01124079991132021, + 0.013840350322425365, + 0.02101340889930725, + 0.0215165875852108, + 0.014302663505077362, + 0.015398591756820679, + 0.02347317337989807, + 0.02744980715215206, + 0.014847934246063232, + 0.021936696022748947, + 0.009881957434117794, + 0.0082856435328722, + 0.007616223301738501, + 0.009150004014372826, + 0.012197860516607761, + 0.012816362082958221, + 0.00720155518501997, + 0.0153488265350461, + 0.006737583316862583, + 0.009662393480539322, + 0.004413866903632879, + 0.004336218815296888, + 0.00761225214228034, + 0.006301512476056814, + 0.005215442273765802, + 0.006179016083478928, + 0.011817123740911484, + 0.013191653415560722, + 0.00612351531162858, + 0.005202633794397116, + 0.007414175197482109, + 0.0069996933452785015, + 0.0038301320746541023, + 0.006906184833496809, + 0.00870989914983511, + 0.008541940711438656, + 0.00414197938516736, + 0.0038334897253662348, + 0.008528579026460648, + 0.009276088327169418, + 0.0035581951960921288, + 0.007810757961124182, + 0.010218305513262749, + 0.013654910027980804, + 0.012122206389904022, + 0.010498451068997383, + 0.020976385101675987, + 0.010231082327663898, + 0.008726775646209717, + 0.009570186957716942, + 0.013242263346910477, + 0.023420341312885284, + 0.016818292438983917, + 0.010592211037874222, + 0.01002085767686367, + 0.013869037851691246, + 0.011150280013680458, + 0.017947372049093246, + 0.03795505315065384, + 0.02355523779988289, + 0.010953880846500397, + 0.009281504899263382, + 0.017849454656243324, + 0.04357367008924484, + 0.010110507719218731, + 0.06938791275024414, + 0.006073269061744213, + 0.00807288009673357, + 0.005336685571819544, + 0.004465620964765549, + 0.005619124975055456, + 0.005139137618243694, + 0.0037345418240875006, + 0.004221425857394934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 7 + ], + "selected_reward": -0.038109004497528076 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous metallurgical workshop with a high-beamed ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and deliberately pan from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive steel ingot visible emitting an intense, pulsating orange-white heat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ingot slide across a roller bed toward a towering industrial press and get compressed by rollers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a violent eruption of vibrant yellow and white sparks when the ingot enters the press?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054970093071460724 + ], + "opposite_logprob_tokens": [ + -2.9299700260162354 + ], + "logprob_reward": -0.054970093071460724, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054970093071460724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0713161826133728, + 0.06657645106315613, + 0.015724079683423042, + 0.014682498760521412, + 0.0280294269323349, + 0.02164582349359989, + 0.02342456579208374, + 0.04581523686647415, + 0.6971898674964905, + 0.015595797449350357 + ], + "credit_map": [ + 0.00707738334313035, + 0.0067243874073028564, + 0.010191385634243488, + 0.007844860665500164, + 0.010202503763139248, + 0.011996614746749401, + 0.007576031610369682, + 0.009703023359179497, + 0.009177942760288715, + 0.008209728635847569, + 0.006983027793467045, + 0.0069277649745345116, + 0.00948147103190422, + 0.011229225434362888, + 0.005962211638689041, + 0.008605077862739563, + 0.0020171841606497765, + 0.0022926684468984604, + 0.001670799683779478, + 0.0020119724795222282, + 0.002069635083898902, + 0.0024121305905282497, + 0.0011327954707667232, + 0.0021168929524719715, + 0.0018284603720530868, + 0.002728031948208809, + 0.0014958069659769535, + 0.0011670879321172833, + 0.0022306949831545353, + 0.0018459748243913054, + 0.001620911993086338, + 0.001765530207194388, + 0.005671811290085316, + 0.005303636193275452, + 0.002969356719404459, + 0.0019475453300401568, + 0.004162384197115898, + 0.003236577846109867, + 0.0019972200971096754, + 0.0027408956084400415, + 0.0037764140870422125, + 0.002836081897839904, + 0.0030289290007203817, + 0.0021688146516680717, + 0.0026425744872540236, + 0.0020313768181949854, + 0.002533701481297612, + 0.0026279310695827007, + 0.0036982386372983456, + 0.003946202341467142, + 0.0028564129024744034, + 0.0015815563965588808, + 0.002711599227041006, + 0.0036062123253941536, + 0.001690004370175302, + 0.0033343397080898285, + 0.00358933350071311, + 0.005680534057319164, + 0.01071719266474247, + 0.004079097416251898, + 0.004688822664320469, + 0.0041736881248652935, + 0.0031636878848075867, + 0.009722882881760597, + 0.1373225301504135, + 0.07701033353805542, + 0.02768387645483017, + 0.023447051644325256, + 0.03846267610788345, + 0.13723066449165344, + 0.026968887075781822, + 0.22906392812728882, + 0.0026030712760984898, + 0.0034652857575565577, + 0.0016024429351091385, + 0.0017020613886415958, + 0.0024717445485293865, + 0.0013698078691959381, + 0.0011161138536408544, + 0.0012652698205783963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.054970093071460724 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a documentary-style wide shot, the camera slowly and deliberately pans from left to right across a vast, cavernous metallurgical workshop, its high-beamed ceiling vanishing into industrial haze and shadows. The primary light source is a massive steel ingot, emitting an intense, pulsating orange-white heat that casts a fiery, dramatic glow and long, dancing shadows across the scene. This ingot slides relentlessly across a massive roller bed, gliding toward the center of the frame, where a towering industrial press awaits, constructed of heavy, dark, greased, and pitted steel. As the glowing metal enters the press's chute, two massive cylindrical rollers come into contact, exerting incredible pressure that visibly thickens and elongates the ingot. This action triggers a violent and spectacular eruption of vibrant yellow and white sparks, which blast outward in a chaotic shower, momentarily illuminating the pristine, functional details of the surrounding machinery. The entire process is captured in crisp, authentic detail, emphasising the sheer scale and raw power of the operation as once-thick steel is ruthlessly transformed into thinner, longer plates, creating a breathtaking and intense atmosphere of industrial might.", + "seed": 11112, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_7/index_6114_seed_11112_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03354518241249025, + "vlm_logprob": -0.03354518241249025, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.03354518241249025, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous metallurgical workshop with a high-beamed ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and deliberately pan from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive steel ingot visible emitting an intense, pulsating orange-white heat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ingot slide across a roller bed toward a towering industrial press and get compressed by rollers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a violent eruption of vibrant yellow and white sparks when the ingot enters the press?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038663215935230255 + ], + "opposite_logprob_tokens": [ + -3.913663148880005 + ], + "logprob_reward": -0.038663215935230255, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.038663215935230255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4291415214538574, + 0.04766308143734932, + 0.06900995969772339, + 0.05581197515130043, + 0.15021155774593353, + 0.05114445462822914, + 0.06596590578556061, + 0.032538749277591705, + 0.06603825837373734, + 0.03247448056936264 + ], + "credit_map": [ + 0.08458789438009262, + 0.07096502929925919, + 0.02476884424686432, + 0.033977001905441284, + 0.04126327112317085, + 0.06024088338017464, + 0.05282476916909218, + 0.06051382049918175, + 0.006627094466239214, + 0.007082601077854633, + 0.0031242358963936567, + 0.0027993570547550917, + 0.008075810968875885, + 0.008566571399569511, + 0.0051142084412276745, + 0.006273201201111078, + 0.010614190250635147, + 0.01443836372345686, + 0.0041235024109482765, + 0.005304687190800905, + 0.01081439945846796, + 0.012515967711806297, + 0.00411611795425415, + 0.007082736119627953, + 0.009358593262732029, + 0.008477624505758286, + 0.0063574961386621, + 0.004104729276150465, + 0.008932088501751423, + 0.010251802392303944, + 0.003439896274358034, + 0.0048897466622292995, + 0.027467496693134308, + 0.012432285584509373, + 0.013521860353648663, + 0.011713341809809208, + 0.04321359097957611, + 0.01588073931634426, + 0.013233925215899944, + 0.012748322449624538, + 0.0041476949118077755, + 0.006266709417104721, + 0.005865233018994331, + 0.012295770458877087, + 0.006412947550415993, + 0.003685283474624157, + 0.004597493447363377, + 0.007873321883380413, + 0.009815539233386517, + 0.006871993653476238, + 0.008718287572264671, + 0.009187296964228153, + 0.008214739151299, + 0.007230408955365419, + 0.005214276257902384, + 0.010713368654251099, + 0.0034663178957998753, + 0.0044280546717345715, + 0.0030058114789426327, + 0.005496379919350147, + 0.0031354911625385284, + 0.004313058685511351, + 0.004457550588995218, + 0.004236085340380669, + 0.024393700063228607, + 0.016612553969025612, + 0.005985880736261606, + 0.005067446269094944, + 0.005616610869765282, + 0.003543198574334383, + 0.0024524845648556948, + 0.0023663847241550684, + 0.005250286776572466, + 0.006141068879514933, + 0.005061191972345114, + 0.003474648343399167, + 0.003758186474442482, + 0.004380360245704651, + 0.0025850131642073393, + 0.0018237227341160178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.038663215935230255 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous metallurgical workshop with a high-beamed ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and deliberately pan from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive steel ingot visible emitting an intense, pulsating orange-white heat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ingot slide across a roller bed toward a towering industrial press and get compressed by rollers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a violent eruption of vibrant yellow and white sparks when the ingot enters the press?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032995045185089 + ], + "opposite_logprob_tokens": [ + -2.350329875946045 + ], + "logprob_reward": -0.10032995045185089, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032995045185089, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19890755414962769, + 0.058157943189144135, + 0.09609845280647278, + 0.09730011224746704, + 0.08741414546966553, + 0.20157141983509064, + 0.03038589097559452, + 0.054079845547676086, + 0.09092730283737183, + 0.08515724539756775 + ], + "credit_map": [ + 0.03516651317477226, + 0.03790416941046715, + 0.009202089160680771, + 0.013435170985758305, + 0.03146275877952576, + 0.02845866233110428, + 0.02173168770968914, + 0.02154652215540409, + 0.013077975250780582, + 0.010505530051887035, + 0.003911452367901802, + 0.003228933783248067, + 0.007901164703071117, + 0.009545671753585339, + 0.004540364723652601, + 0.005446859169751406, + 0.013948180712759495, + 0.019895877689123154, + 0.0060389419086277485, + 0.013186541385948658, + 0.014216416515409946, + 0.014886713586747646, + 0.0048249252140522, + 0.009100856259465218, + 0.021058032289147377, + 0.025941258296370506, + 0.010572150349617004, + 0.007689738180488348, + 0.013490765355527401, + 0.00697832927107811, + 0.004796330817043781, + 0.0067735156044363976, + 0.019465778023004532, + 0.010599755682051182, + 0.007995188236236572, + 0.00906427763402462, + 0.021340521052479744, + 0.009316557087004185, + 0.00492718955501914, + 0.004704881459474564, + 0.015217912383377552, + 0.020622476935386658, + 0.023268813267350197, + 0.03799520805478096, + 0.03533029556274414, + 0.024897079914808273, + 0.017216159030795097, + 0.02702348679304123, + 0.004278362262994051, + 0.005068741738796234, + 0.004849747754633427, + 0.0046406090259552, + 0.0035120330285280943, + 0.002754892921075225, + 0.0023547399323433638, + 0.0029267659410834312, + 0.005907729733735323, + 0.00882936641573906, + 0.00485317874699831, + 0.009843080304563046, + 0.006374040152877569, + 0.007244584150612354, + 0.006261767353862524, + 0.004766105208545923, + 0.01144754234701395, + 0.0119955288246274, + 0.011218622326850891, + 0.013084884732961655, + 0.014439604245126247, + 0.010041393339633942, + 0.010109199211001396, + 0.008590536192059517, + 0.010818655602633953, + 0.008145958185195923, + 0.018683459609746933, + 0.01282534934580326, + 0.008831306360661983, + 0.009940416552126408, + 0.009113281965255737, + 0.006798820570111275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.10032995045185089 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous metallurgical workshop with a high-beamed ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and deliberately pan from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive steel ingot visible emitting an intense, pulsating orange-white heat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ingot slide across a roller bed toward a towering industrial press and get compressed by rollers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a violent eruption of vibrant yellow and white sparks when the ingot enters the press?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013443727511912584 + ], + "opposite_logprob_tokens": [ + -6.626344203948975 + ], + "logprob_reward": -0.0013443727511912584, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013443727511912584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4941680431365967, + 0.05369539186358452, + 0.11068303883075714, + 0.08724027127027512, + 0.0341535322368145, + 0.031957220286130905, + 0.026713943108916283, + 0.053123537451028824, + 0.05444319173693657, + 0.053821854293346405 + ], + "credit_map": [ + 0.07217004895210266, + 0.07374700903892517, + 0.029476603493094444, + 0.03818565234541893, + 0.0473971851170063, + 0.09755413234233856, + 0.0647083967924118, + 0.07092901319265366, + 0.007897211238741875, + 0.007280698046088219, + 0.0038214013911783695, + 0.004558532498776913, + 0.006105037871748209, + 0.007018871605396271, + 0.008696162141859531, + 0.008317475207149982, + 0.012502530589699745, + 0.016261544078588486, + 0.007695687934756279, + 0.017103001475334167, + 0.014772424474358559, + 0.014477088116109371, + 0.010271209292113781, + 0.0175995584577322, + 0.019745251163840294, + 0.026210905984044075, + 0.006204543169587851, + 0.005386924836784601, + 0.0086581502109766, + 0.008409153670072556, + 0.006570523139089346, + 0.006054820027202368, + 0.004394111223518848, + 0.00439923070371151, + 0.004401226062327623, + 0.0029393304139375687, + 0.007864504121243954, + 0.004621609579771757, + 0.002938302932307124, + 0.002595218364149332, + 0.004133615177124739, + 0.0033690694253891706, + 0.004735224414616823, + 0.006067308597266674, + 0.0038161699194461107, + 0.0025491896085441113, + 0.0024567153304815292, + 0.0048299310728907585, + 0.0038802651688456535, + 0.0042502908036112785, + 0.0030137875583022833, + 0.003155986312776804, + 0.0032387857791036367, + 0.002751243533566594, + 0.0033307720441371202, + 0.0030928130727261305, + 0.004711115267127752, + 0.008935893885791302, + 0.005614737048745155, + 0.007653399836272001, + 0.005937074776738882, + 0.00811717752367258, + 0.006650194525718689, + 0.0055039445869624615, + 0.016634376719594002, + 0.012397700920701027, + 0.0047331214882433414, + 0.00390815082937479, + 0.007439600769430399, + 0.0033253447618335485, + 0.003526078537106514, + 0.002478816779330373, + 0.008058733306825161, + 0.005896994844079018, + 0.01036122441291809, + 0.0073967380449175835, + 0.004642499145120382, + 0.005436592269688845, + 0.006683420389890671, + 0.005345646757632494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.0013443727511912584 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous metallurgical workshop with a high-beamed ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and deliberately pan from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive steel ingot visible emitting an intense, pulsating orange-white heat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ingot slide across a roller bed toward a towering industrial press and get compressed by rollers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a violent eruption of vibrant yellow and white sparks when the ingot enters the press?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023291684687137604 + ], + "opposite_logprob_tokens": [ + -3.77329158782959 + ], + "logprob_reward": -0.023291684687137604, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023291684687137604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2571253180503845, + 0.07739637792110443, + 0.08330134302377701, + 0.056868091225624084, + 0.10565420240163803, + 0.07284160703420639, + 0.047199394553899765, + 0.15167735517024994, + 0.09911620616912842, + 0.048820048570632935 + ], + "credit_map": [ + 0.03542178496718407, + 0.040454450994729996, + 0.015156514942646027, + 0.01873028837144375, + 0.03067840449512005, + 0.05559791624546051, + 0.029365189373493195, + 0.03172079101204872, + 0.013871456496417522, + 0.009636983275413513, + 0.0046255020424723625, + 0.006955557037144899, + 0.010431138798594475, + 0.011284026317298412, + 0.01048303209245205, + 0.01010867953300476, + 0.014334878884255886, + 0.01063372939825058, + 0.005651337094604969, + 0.011989467777311802, + 0.011408788152039051, + 0.01068204827606678, + 0.006902946159243584, + 0.01169815007597208, + 0.006182002369314432, + 0.006728689651936293, + 0.007131831254810095, + 0.006541281472891569, + 0.007122151553630829, + 0.009777208790183067, + 0.006140041630715132, + 0.007244886830449104, + 0.013385321944952011, + 0.013226446695625782, + 0.012566863559186459, + 0.010151123628020287, + 0.025451723486185074, + 0.014547531493008137, + 0.010102339088916779, + 0.006222856231033802, + 0.006734390277415514, + 0.008964398875832558, + 0.009425305761396885, + 0.015568334609270096, + 0.00966024212539196, + 0.008176257833838463, + 0.005255409050732851, + 0.009057272225618362, + 0.004627902992069721, + 0.005863668862730265, + 0.004940933082252741, + 0.00620900047942996, + 0.0074828024953603745, + 0.005602387245744467, + 0.004664847627282143, + 0.007807855028659105, + 0.015022720210254192, + 0.01680060476064682, + 0.011442090384662151, + 0.0374075286090374, + 0.010688560083508492, + 0.026317045092582703, + 0.01968253403902054, + 0.014316272921860218, + 0.01973363757133484, + 0.012939768843352795, + 0.013143789954483509, + 0.01464476902037859, + 0.012552970089018345, + 0.010434887371957302, + 0.009109665639698505, + 0.006556717678904533, + 0.007294195704162121, + 0.005640849936753511, + 0.007077836897224188, + 0.004926701076328754, + 0.005948218051344156, + 0.008229618892073631, + 0.004771072883158922, + 0.004931556060910225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.023291684687137604 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous metallurgical workshop with a high-beamed ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and deliberately pan from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive steel ingot visible emitting an intense, pulsating orange-white heat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ingot slide across a roller bed toward a towering industrial press and get compressed by rollers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a violent eruption of vibrant yellow and white sparks when the ingot enters the press?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004096688237041235 + ], + "opposite_logprob_tokens": [ + -5.504096508026123 + ], + "logprob_reward": -0.004096688237041235, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004096688237041235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11919697374105453, + 0.0509001724421978, + 0.0642968937754631, + 0.09039109945297241, + 0.09032023698091507, + 0.04524172097444534, + 0.048520129173994064, + 0.06603474915027618, + 0.21393205225467682, + 0.21116593480110168 + ], + "credit_map": [ + 0.02306080050766468, + 0.01875331811606884, + 0.004187006037682295, + 0.0046315062791109085, + 0.03330733999609947, + 0.01729084551334381, + 0.00794728472828865, + 0.010018870234489441, + 0.00948378536850214, + 0.010487722232937813, + 0.004489230457693338, + 0.0037923266645520926, + 0.00723119406029582, + 0.007283306214958429, + 0.00415217038244009, + 0.00398044241592288, + 0.012236308306455612, + 0.014493920840322971, + 0.0027323642279952765, + 0.004572285804897547, + 0.011725103482604027, + 0.009728633798658848, + 0.0037751293275505304, + 0.005033145658671856, + 0.023719601333141327, + 0.028212042525410652, + 0.005573910195380449, + 0.004044262692332268, + 0.011278177611529827, + 0.007726467680186033, + 0.0049490174278616905, + 0.004887615796178579, + 0.015200184658169746, + 0.0076466212049126625, + 0.007992657832801342, + 0.006389012094587088, + 0.023895373567938805, + 0.012097897939383984, + 0.00903650838881731, + 0.008061985485255718, + 0.00489521399140358, + 0.0064710406586527824, + 0.005362458061426878, + 0.010058898478746414, + 0.005072727799415588, + 0.004647789057344198, + 0.004399869590997696, + 0.004333724267780781, + 0.006710071582347155, + 0.007000864017754793, + 0.004846577532589436, + 0.006387075409293175, + 0.006338847801089287, + 0.005183633416891098, + 0.004756859038025141, + 0.007296198979020119, + 0.005416130181401968, + 0.008901715278625488, + 0.005775768775492907, + 0.012445985339581966, + 0.006200053263455629, + 0.010448355227708817, + 0.010956862941384315, + 0.005889874882996082, + 0.04755210131406784, + 0.0433073453605175, + 0.015511313453316689, + 0.017828885465860367, + 0.03707963228225708, + 0.018120424821972847, + 0.02278822846710682, + 0.011744135059416294, + 0.03040112927556038, + 0.024955879896879196, + 0.03918870538473129, + 0.03020624630153179, + 0.01840939000248909, + 0.020622408017516136, + 0.027899201959371567, + 0.019482985138893127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.004096688237041235 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a documentary-style wide shot, the camera slowly and deliberately pans from left to right across a vast, cavernous metallurgical workshop, its high-beamed ceiling vanishing into industrial haze and shadows. The primary light source is a massive steel ingot, emitting an intense, pulsating orange-white heat that casts a fiery, dramatic glow and long, dancing shadows across the scene. This ingot slides relentlessly across a massive roller bed, gliding toward the center of the frame, where a towering industrial press awaits, constructed of heavy, dark, greased, and pitted steel. As the glowing metal enters the press's chute, two massive cylindrical rollers come into contact, exerting incredible pressure that visibly thickens and elongates the ingot. This action triggers a violent and spectacular eruption of vibrant yellow and white sparks, which blast outward in a chaotic shower, momentarily illuminating the pristine, functional details of the surrounding machinery. The entire process is captured in crisp, authentic detail, emphasising the sheer scale and raw power of the operation as once-thick steel is ruthlessly transformed into thinner, longer plates, creating a breathtaking and intense atmosphere of industrial might.", + "seed": 11113, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_7/index_6114_seed_11113_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08413122678175569, + "vlm_logprob": -0.08413122678175569, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.08413122678175569, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous metallurgical workshop with a high-beamed ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and deliberately pan from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive steel ingot visible emitting an intense, pulsating orange-white heat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ingot slide across a roller bed toward a towering industrial press and get compressed by rollers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a violent eruption of vibrant yellow and white sparks when the ingot enters the press?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.19893881678581238 + ], + "opposite_logprob_tokens": [ + -1.8239388465881348 + ], + "logprob_reward": -0.19893881678581238, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.19893881678581238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3355247974395752, + 0.26062360405921936, + 0.09199875593185425, + 0.05910476669669151, + 0.02816716767847538, + 0.04352891072630882, + 0.04547707363963127, + 0.05688440799713135, + 0.051825616508722305, + 0.026864908635616302 + ], + "credit_map": [ + 0.040997814387083054, + 0.038027770817279816, + 0.04591812938451767, + 0.07301631569862366, + 0.04091396555304527, + 0.03813039883971214, + 0.03610146790742874, + 0.02241896651685238, + 0.03443970903754234, + 0.026789924129843712, + 0.04339274764060974, + 0.06658413261175156, + 0.026567349210381508, + 0.024731092154979706, + 0.02193666435778141, + 0.016181958839297295, + 0.012152668088674545, + 0.015794193372130394, + 0.016620613634586334, + 0.01613379828631878, + 0.0077534448355436325, + 0.013031959533691406, + 0.0050592427141964436, + 0.0054528312757611275, + 0.006135910749435425, + 0.007432625629007816, + 0.011827563866972923, + 0.010737217962741852, + 0.006137290969491005, + 0.007516152691096067, + 0.0038045363035053015, + 0.005513471085578203, + 0.004577822983264923, + 0.0031114784069359303, + 0.004173746798187494, + 0.0023538873065263033, + 0.002421370707452297, + 0.005884155631065369, + 0.002407764084637165, + 0.0032369426917284727, + 0.004215002991259098, + 0.0024308310821652412, + 0.006978380028158426, + 0.0047388081438839436, + 0.002450748812407255, + 0.004758916795253754, + 0.004142774268984795, + 0.013813445344567299, + 0.009050666354596615, + 0.004228345118463039, + 0.008195924572646618, + 0.004877662286162376, + 0.005452445708215237, + 0.004381245002150536, + 0.002737463219091296, + 0.006553323939442635, + 0.0047119492664933205, + 0.0053671011701226234, + 0.0060872300527989864, + 0.008810724131762981, + 0.0059500327333807945, + 0.015523466281592846, + 0.0038925078697502613, + 0.006541394628584385, + 0.004570845980197191, + 0.007763260044157505, + 0.007329494692385197, + 0.0053144400008022785, + 0.008061627857387066, + 0.004458603449165821, + 0.006787165999412537, + 0.0075401803478598595, + 0.002766765421256423, + 0.00344588840380311, + 0.0038082993123680353, + 0.003483704524114728, + 0.0029026770498603582, + 0.002898011589422822, + 0.0036225803196430206, + 0.003936981316655874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.19893881678581238 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous metallurgical workshop with a high-beamed ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and deliberately pan from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive steel ingot visible emitting an intense, pulsating orange-white heat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ingot slide across a roller bed toward a towering industrial press and get compressed by rollers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a violent eruption of vibrant yellow and white sparks when the ingot enters the press?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08907075971364975 + ], + "opposite_logprob_tokens": [ + -2.4640707969665527 + ], + "logprob_reward": -0.08907075971364975, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08907075971364975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12022385746240616, + 0.08628890663385391, + 0.05467139557003975, + 0.0466206856071949, + 0.06244067847728729, + 0.1279054433107376, + 0.018488837406039238, + 0.35687610507011414, + 0.0714588537812233, + 0.05502521991729736 + ], + "credit_map": [ + 0.019797053188085556, + 0.011377975344657898, + 0.014463559724390507, + 0.027028609067201614, + 0.014915554784238338, + 0.015459924936294556, + 0.010118522681295872, + 0.007062646094709635, + 0.016457751393318176, + 0.00992648210376501, + 0.008021285757422447, + 0.011690077371895313, + 0.009289783425629139, + 0.012276698835194111, + 0.00842970609664917, + 0.010197112336754799, + 0.004681093152612448, + 0.007327208761125803, + 0.010033298283815384, + 0.010052810423076153, + 0.003079249057918787, + 0.007726316340267658, + 0.0044898297637701035, + 0.0072815860621631145, + 0.005727104842662811, + 0.007742187473922968, + 0.010410703718662262, + 0.0073234038427472115, + 0.0028117834590375423, + 0.005174871534109116, + 0.00281626358628273, + 0.004614365287125111, + 0.009600989520549774, + 0.006616339553147554, + 0.00925027672201395, + 0.006852739490568638, + 0.0059327902272343636, + 0.010428589768707752, + 0.007241342216730118, + 0.006517607253044844, + 0.009759023785591125, + 0.007802306674420834, + 0.021764308214187622, + 0.011035137809813023, + 0.006534479092806578, + 0.011513716541230679, + 0.011606142856180668, + 0.0478903166949749, + 0.0025941322091966867, + 0.0013514942256733775, + 0.002870549215003848, + 0.0024507686030119658, + 0.003093135543167591, + 0.0015504079638049006, + 0.001781472354196012, + 0.002796876709908247, + 0.01812502183020115, + 0.012034221552312374, + 0.030828597024083138, + 0.0899304673075676, + 0.019543632864952087, + 0.12593072652816772, + 0.018896659836173058, + 0.04158676043152809, + 0.006778920069336891, + 0.009387986734509468, + 0.008448145352303982, + 0.008341648615896702, + 0.010817267000675201, + 0.006101496517658234, + 0.009170238859951496, + 0.012413141317665577, + 0.006201351061463356, + 0.006506777368485928, + 0.0071655940264463425, + 0.007973119616508484, + 0.005852469243109226, + 0.006211256608366966, + 0.008667096495628357, + 0.006447550840675831 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.08907075971364975 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous metallurgical workshop with a high-beamed ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and deliberately pan from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive steel ingot visible emitting an intense, pulsating orange-white heat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ingot slide across a roller bed toward a towering industrial press and get compressed by rollers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a violent eruption of vibrant yellow and white sparks when the ingot enters the press?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007643737364560366 + ], + "opposite_logprob_tokens": [ + -4.882643699645996 + ], + "logprob_reward": -0.007643737364560366, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007643737364560366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2065889686346054, + 0.09815383702516556, + 0.06629189848899841, + 0.1259140521287918, + 0.06622372567653656, + 0.02053132839500904, + 0.04268031567335129, + 0.23796628415584564, + 0.06097646802663803, + 0.07467308640480042 + ], + "credit_map": [ + 0.02990490011870861, + 0.020561110228300095, + 0.018627144396305084, + 0.03646422550082207, + 0.024739794433116913, + 0.029535388574004173, + 0.02822807803750038, + 0.018528325483202934, + 0.012063098140060902, + 0.008921797387301922, + 0.011775307357311249, + 0.01657690294086933, + 0.010854625143110752, + 0.014219353906810284, + 0.013090117834508419, + 0.010652633383870125, + 0.0074023534543812275, + 0.009672300890088081, + 0.013199824839830399, + 0.012814890593290329, + 0.004888903349637985, + 0.008518120273947716, + 0.004250646568834782, + 0.005544860381633043, + 0.015434100292623043, + 0.021332886070013046, + 0.024068016558885574, + 0.0216983649879694, + 0.005966343451291323, + 0.011477390304207802, + 0.009520361199975014, + 0.016416577622294426, + 0.008397329598665237, + 0.009778144769370556, + 0.01036944892257452, + 0.007352237589657307, + 0.008588286116719246, + 0.00804835744202137, + 0.006746347527951002, + 0.006943578366190195, + 0.00313773937523365, + 0.0017831629374995828, + 0.003344411263242364, + 0.002336002653464675, + 0.001713066129013896, + 0.0037740038242191076, + 0.0015829183394089341, + 0.002860024804249406, + 0.005190667230635881, + 0.0032041040249168873, + 0.006228947080671787, + 0.005270404275506735, + 0.006233048625290394, + 0.005839341785758734, + 0.004200930707156658, + 0.006512871943414211, + 0.011665327474474907, + 0.009140432812273502, + 0.022567162290215492, + 0.06569404155015945, + 0.012865192256867886, + 0.07820288091897964, + 0.011909596621990204, + 0.025921644642949104, + 0.006726750638335943, + 0.008433838374912739, + 0.005227189976722002, + 0.008146542124450207, + 0.010067869909107685, + 0.00533886905759573, + 0.007120299618691206, + 0.009915109723806381, + 0.005933075211942196, + 0.010071396827697754, + 0.011217175051569939, + 0.01085048820823431, + 0.009590177796781063, + 0.008236727677285671, + 0.008510276675224304, + 0.010263761505484581 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.007643737364560366 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous metallurgical workshop with a high-beamed ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and deliberately pan from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive steel ingot visible emitting an intense, pulsating orange-white heat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ingot slide across a roller bed toward a towering industrial press and get compressed by rollers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a violent eruption of vibrant yellow and white sparks when the ingot enters the press?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497133359313011 + ], + "opposite_logprob_tokens": [ + -2.929971218109131 + ], + "logprob_reward": -0.05497133359313011, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497133359313011, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14027602970600128, + 0.13435283303260803, + 0.037708256393671036, + 0.07795316725969315, + 0.10145719349384308, + 0.04153969883918762, + 0.06410406529903412, + 0.28585031628608704, + 0.03814486414194107, + 0.0786135345697403 + ], + "credit_map": [ + 0.016332097351551056, + 0.01121149305254221, + 0.014740136452019215, + 0.02729533240199089, + 0.018196210265159607, + 0.020243201404809952, + 0.017107827588915825, + 0.015149733051657677, + 0.01928398199379444, + 0.01358388364315033, + 0.013778652995824814, + 0.018517790362238884, + 0.011381945572793484, + 0.022020911797881126, + 0.01648407056927681, + 0.01930159516632557, + 0.0034775673411786556, + 0.005520785227417946, + 0.006962952669709921, + 0.0069748833775520325, + 0.003555124159902334, + 0.005700557958334684, + 0.0024862620048224926, + 0.003030124818906188, + 0.008187683299183846, + 0.014499609358608723, + 0.011878828518092632, + 0.012359922751784325, + 0.007080880459398031, + 0.009690843522548676, + 0.005512117873877287, + 0.008743276819586754, + 0.01506838295608759, + 0.010903887450695038, + 0.014315774664282799, + 0.009773263707756996, + 0.011801855638623238, + 0.016645431518554688, + 0.012358054518699646, + 0.010590540245175362, + 0.005093827843666077, + 0.0025815065018832684, + 0.00638542277738452, + 0.004479661583900452, + 0.002687414176762104, + 0.004666835535317659, + 0.003129921155050397, + 0.012515108101069927, + 0.00635246466845274, + 0.004935422912240028, + 0.010468591004610062, + 0.00882479827851057, + 0.00994103029370308, + 0.007264731917530298, + 0.006816546432673931, + 0.009500477463006973, + 0.014017793349921703, + 0.010357915423810482, + 0.027163119986653328, + 0.08125792443752289, + 0.01640143431723118, + 0.0925426334142685, + 0.014061192981898785, + 0.030048277229070663, + 0.00431954488158226, + 0.0056204963475465775, + 0.004244400188326836, + 0.004445065278559923, + 0.005802256520837545, + 0.003731936914846301, + 0.005652023013681173, + 0.004329140763729811, + 0.006597194820642471, + 0.01277308538556099, + 0.008555748499929905, + 0.01106955949217081, + 0.009667831473052502, + 0.010629476048052311, + 0.011026718653738499, + 0.008293922990560532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.05497133359313011 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous metallurgical workshop with a high-beamed ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and deliberately pan from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive steel ingot visible emitting an intense, pulsating orange-white heat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ingot slide across a roller bed toward a towering industrial press and get compressed by rollers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a violent eruption of vibrant yellow and white sparks when the ingot enters the press?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003148645162582 + ], + "opposite_logprob_tokens": [ + -2.6950314044952393 + ], + "logprob_reward": -0.07003148645162582, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003148645162582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24332475662231445, + 0.09197079390287399, + 0.04414753615856171, + 0.06122790277004242, + 0.09881392866373062, + 0.02484798990190029, + 0.04208896681666374, + 0.2362709939479828, + 0.04124162718653679, + 0.11606548726558685 + ], + "credit_map": [ + 0.031021976843476295, + 0.026868410408496857, + 0.024863339960575104, + 0.04291510581970215, + 0.033336032181978226, + 0.032375335693359375, + 0.028843579813838005, + 0.023100977763533592, + 0.01710265688598156, + 0.009436681866645813, + 0.010596830397844315, + 0.01579960063099861, + 0.009043140336871147, + 0.010092143900692463, + 0.010120698250830173, + 0.009779039770364761, + 0.005085931625217199, + 0.004456644877791405, + 0.008051973767578602, + 0.008085616864264011, + 0.003850017674267292, + 0.006439841818064451, + 0.003399349283427, + 0.004778158850967884, + 0.007444308139383793, + 0.010403205640614033, + 0.009710265323519707, + 0.008129382506012917, + 0.005379804875701666, + 0.008285237476229668, + 0.003027935978025198, + 0.00884776096791029, + 0.013815265148878098, + 0.010170582681894302, + 0.013831321150064468, + 0.009635066613554955, + 0.01204681396484375, + 0.01672203652560711, + 0.011928290128707886, + 0.010664554312825203, + 0.002597827697172761, + 0.0020235679112374783, + 0.0032672041561454535, + 0.002571279415860772, + 0.0020795308519154787, + 0.003079465590417385, + 0.002126679988577962, + 0.007102435454726219, + 0.004740915261209011, + 0.0033319105859845877, + 0.0059707285836339, + 0.005649083759635687, + 0.005784227047115564, + 0.005831056274473667, + 0.0050504254177212715, + 0.0057306187227368355, + 0.011685333214700222, + 0.007641841657459736, + 0.02141737751662731, + 0.06183118373155594, + 0.014020553790032864, + 0.0812348797917366, + 0.011080706492066383, + 0.027359141036868095, + 0.004099024459719658, + 0.005489607807248831, + 0.005779458209872246, + 0.004478990565985441, + 0.006976118311285973, + 0.003965429496020079, + 0.0056107486598193645, + 0.004842248745262623, + 0.009730488993227482, + 0.014015053398907185, + 0.018513508141040802, + 0.01782550849020481, + 0.011233562603592873, + 0.014381402172148228, + 0.014117890037596226, + 0.016248077154159546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.07003148645162582 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_8/index_tensor([3102, 3102])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_8/index_tensor([3102, 3102])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4cbe9c7d6ff22addd6c27c67ad0616242de9266e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_8/index_tensor([3102, 3102])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "long, slightly upward shot from a fixed camera position captures the majestic scene atop the mountain with a documentary quality. At the center of the frame, a lone poet stands atop the highest point of a jagged, dark gray rock formation. He wears a loose, pure white robe, its thin cotton and linen material whipped violently in the fierce mountain breeze, its sleeves and hem fluttering and unfurling dramatically like a living banner. His long, dark hair, ruffled by the wind, contrasts sharply with the flowing white robe. Despite the distance, his upright figure is still recognizable, one hand clasped behind his back, gazing into the distance. The background is an endless expanse of mist-shrouded mountains, and the sky is a deep, almost inky blue, characteristic of high altitudes. The slanting afternoon sun penetrates the thin cloud cover, giving the white robe a sharp golden edge and casting long, sharply defined shadows on the rock. The color saturation of the entire picture is low, sharp and slightly grainy, creating a desolate, lonely, yet epic and tranquil atmosphere.", + "seed": 9858, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_8/index_3102_seed_9858_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.09214734244160354, + "vlm_logprob": -0.09214734244160354, + "vlm_yes_no_margin": 3.4749999761581423, + "vlm_token_credit": -0.09214734244160354, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a slightly upward angle from a fixed position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone figure wearing a loose, pure white robe standing atop a jagged, dark gray rock formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white robe's fabric visibly whipped violently by the wind, with sleeves and hem fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless expanse of mist-shrouded mountains under a deep, inky blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit low color saturation, sharp focus, and a slightly grainy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043450865894556046 + ], + "opposite_logprob_tokens": [ + -3.4184508323669434 + ], + "logprob_reward": -0.043450865894556046, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.043450865894556046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2998109459877014, + 0.08528964966535568, + 0.032112784683704376, + 0.06300470232963562, + 0.022848330438137054, + 0.055773474276065826, + 0.033268366008996964, + 0.06968206912279129, + 0.07171996682882309, + 0.2664896547794342 + ], + "credit_map": [ + 0.049990128725767136, + 0.03835393860936165, + 0.06943367421627045, + 0.03379478305578232, + 0.029259344562888145, + 0.03696981072425842, + 0.021659092977643013, + 0.02035018987953663, + 0.012999672442674637, + 0.011581304483115673, + 0.020167160779237747, + 0.008899541571736336, + 0.010625168681144714, + 0.009416015818715096, + 0.006032761186361313, + 0.005568023771047592, + 0.0029354507569223642, + 0.004422933328896761, + 0.006692002061754465, + 0.004779456183314323, + 0.0036140584852546453, + 0.003350307000800967, + 0.003937280736863613, + 0.0023812984582036734, + 0.00695406086742878, + 0.008937082253396511, + 0.012350575998425484, + 0.007360415067523718, + 0.005138437729328871, + 0.011426414363086224, + 0.005075383000075817, + 0.005762333516031504, + 0.0022917413152754307, + 0.0043637072667479515, + 0.0034125465899705887, + 0.0030343940015882254, + 0.002557500498369336, + 0.003278071526437998, + 0.00211463775485754, + 0.0017957318341359496, + 0.005342627875506878, + 0.0067794653587043285, + 0.007331607863306999, + 0.006295815575867891, + 0.011797034181654453, + 0.005393837578594685, + 0.007571529597043991, + 0.005261559039354324, + 0.005099940113723278, + 0.00598516408354044, + 0.006664170417934656, + 0.0035153250209987164, + 0.002883071545511484, + 0.004653653129935265, + 0.00267638242803514, + 0.0017906605498865247, + 0.005973567720502615, + 0.011823599226772785, + 0.008738870732486248, + 0.006348718889057636, + 0.012511927634477615, + 0.009031134657561779, + 0.009082334116101265, + 0.00617191893979907, + 0.007927591912448406, + 0.01057611033320427, + 0.011658664792776108, + 0.011176547966897488, + 0.00836391281336546, + 0.008961236104369164, + 0.006963014602661133, + 0.0060928924940526485, + 0.023671142756938934, + 0.06993268430233002, + 0.03376080095767975, + 0.02707112580537796, + 0.015733201056718826, + 0.071247398853302, + 0.013247959315776825, + 0.011825350113213062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.043450865894556046 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a slightly upward angle from a fixed position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone figure wearing a loose, pure white robe standing atop a jagged, dark gray rock formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white robe's fabric visibly whipped violently by the wind, with sleeves and hem fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless expanse of mist-shrouded mountains under a deep, inky blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit low color saturation, sharp focus, and a slightly grainy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036267952527850866 + ], + "opposite_logprob_tokens": [ + -5.628626823425293 + ], + "logprob_reward": -0.0036267952527850866, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036267952527850866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20112672448158264, + 0.07374880462884903, + 0.052740760147571564, + 0.14777924120426178, + 0.07815838605165482, + 0.06772810965776443, + 0.05804881080985069, + 0.10293764621019363, + 0.12323256582021713, + 0.09449897706508636 + ], + "credit_map": [ + 0.01746780052781105, + 0.025869913399219513, + 0.04682493954896927, + 0.030253710225224495, + 0.016857078298926353, + 0.024961698800325394, + 0.021646304056048393, + 0.017245305702090263, + 0.00861872173845768, + 0.009059586562216282, + 0.017189929261803627, + 0.012073054909706116, + 0.005603063851594925, + 0.008549013175070286, + 0.006383510772138834, + 0.006271927617490292, + 0.004813711158931255, + 0.0077347769401967525, + 0.00873958133161068, + 0.009350483305752277, + 0.004274560604244471, + 0.006191691849380732, + 0.005710936151444912, + 0.005925015080720186, + 0.013745120726525784, + 0.02370503544807434, + 0.02123131975531578, + 0.01003547478467226, + 0.01822558604180813, + 0.039753060787916183, + 0.010783195495605469, + 0.010300450026988983, + 0.006714201997965574, + 0.017488446086645126, + 0.009775192476809025, + 0.006858415435999632, + 0.010071060620248318, + 0.013866071589291096, + 0.007374846842139959, + 0.006010150071233511, + 0.006602949928492308, + 0.009039072319865227, + 0.007938997820019722, + 0.007686849217861891, + 0.014290578663349152, + 0.006690002977848053, + 0.008887715637683868, + 0.006591940764337778, + 0.004919450264424086, + 0.008976003155112267, + 0.011202165856957436, + 0.009089960716664791, + 0.004808023571968079, + 0.008596008643507957, + 0.005659161601215601, + 0.004798032343387604, + 0.008467370644211769, + 0.025505734607577324, + 0.008402494713664055, + 0.007592080160975456, + 0.016430091112852097, + 0.018344821408391, + 0.010292005725204945, + 0.00790304783731699, + 0.015580146573483944, + 0.034289952367544174, + 0.019668729975819588, + 0.010865762829780579, + 0.009234961122274399, + 0.02402617409825325, + 0.0051986780017614365, + 0.004368164576590061, + 0.00781469326466322, + 0.025404654443264008, + 0.009447605349123478, + 0.007930104620754719, + 0.0069126347079873085, + 0.026733649894595146, + 0.005593813490122557, + 0.004661827813833952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.0036267952527850866 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a slightly upward angle from a fixed position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone figure wearing a loose, pure white robe standing atop a jagged, dark gray rock formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white robe's fabric visibly whipped violently by the wind, with sleeves and hem fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless expanse of mist-shrouded mountains under a deep, inky blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit low color saturation, sharp focus, and a slightly grainy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813848853111267 + ], + "opposite_logprob_tokens": [ + -1.4063849449157715 + ], + "logprob_reward": -0.2813848853111267, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813848853111267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10016413778066635, + 0.07113073766231537, + 0.04657207056879997, + 0.15889640152454376, + 0.03982565551996231, + 0.018898725509643555, + 0.06758410483598709, + 0.03754661977291107, + 0.1331036388874054, + 0.32627785205841064 + ], + "credit_map": [ + 0.009457128122448921, + 0.013345411978662014, + 0.01683289371430874, + 0.015402640216052532, + 0.00982338935136795, + 0.015682512894272804, + 0.010151905007660389, + 0.009468263015151024, + 0.007318117655813694, + 0.008297860622406006, + 0.014024249278008938, + 0.010764723643660545, + 0.007879804819822311, + 0.00982046127319336, + 0.006374487187713385, + 0.006651029922068119, + 0.0052299220114946365, + 0.006379515863955021, + 0.007495924364775419, + 0.006300003733485937, + 0.0053854938596487045, + 0.005128443706780672, + 0.005535815842449665, + 0.00511695072054863, + 0.013364228419959545, + 0.025816885754466057, + 0.020156938582658768, + 0.010931484401226044, + 0.023279888555407524, + 0.04334670677781105, + 0.011301993392407894, + 0.01069826353341341, + 0.004825192969292402, + 0.006430026609450579, + 0.006785727571696043, + 0.00398642523214221, + 0.004820654634386301, + 0.005062750540673733, + 0.004007723182439804, + 0.003907156642526388, + 0.0021044176537543535, + 0.0033024949952960014, + 0.00316606811247766, + 0.002218186156824231, + 0.0025997599586844444, + 0.002132722642272711, + 0.0018965512281283736, + 0.0014785248786211014, + 0.0058342763222754, + 0.011724154464900494, + 0.007747658994048834, + 0.00629567913711071, + 0.01160120777785778, + 0.01034561451524496, + 0.008387633599340916, + 0.005647883750498295, + 0.0034118324983865023, + 0.004692289512604475, + 0.005031142849475145, + 0.004234168212860823, + 0.006331222597509623, + 0.00354732945561409, + 0.005991739220917225, + 0.004306896589696407, + 0.015172933228313923, + 0.029543615877628326, + 0.02066834457218647, + 0.011958155781030655, + 0.015418785624206066, + 0.021091006696224213, + 0.012609542347490788, + 0.006641250569373369, + 0.028768422082066536, + 0.08386906236410141, + 0.041583843529224396, + 0.034251801669597626, + 0.020008444786071777, + 0.08626803010702133, + 0.01655614748597145, + 0.014972115866839886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 8 + ], + "selected_reward": -0.2813848853111267 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a slightly upward angle from a fixed position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone figure wearing a loose, pure white robe standing atop a jagged, dark gray rock formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white robe's fabric visibly whipped violently by the wind, with sleeves and hem fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless expanse of mist-shrouded mountains under a deep, inky blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit low color saturation, sharp focus, and a slightly grainy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052458494901657104 + ], + "opposite_logprob_tokens": [ + -5.255245685577393 + ], + "logprob_reward": -0.0052458494901657104, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052458494901657104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11349552124738693, + 0.028905272483825684, + 0.04295135289430618, + 0.3625597059726715, + 0.03534296154975891, + 0.026181619614362717, + 0.025118432939052582, + 0.05917555093765259, + 0.1302628070116043, + 0.17600680887699127 + ], + "credit_map": [ + 0.024242030456662178, + 0.018474873155355453, + 0.022767554968595505, + 0.013373288325965405, + 0.008460086770355701, + 0.012412606738507748, + 0.00624435069039464, + 0.007520733401179314, + 0.0033566199708729982, + 0.004040149040520191, + 0.006911092437803745, + 0.0037099586334079504, + 0.003111958736553788, + 0.002776146400719881, + 0.0028823968023061752, + 0.0021169520914554596, + 0.005829087924212217, + 0.007264474406838417, + 0.007636555936187506, + 0.004512447863817215, + 0.004176511894911528, + 0.004592269659042358, + 0.004780285991728306, + 0.0041597201488912106, + 0.03576354309916496, + 0.059781089425086975, + 0.058646924793720245, + 0.03507912531495094, + 0.0339517667889595, + 0.08131427317857742, + 0.0317734032869339, + 0.026249565184116364, + 0.00339048495516181, + 0.006233118008822203, + 0.004451507702469826, + 0.0039339070208370686, + 0.005597260780632496, + 0.0051516154780983925, + 0.0036307796835899353, + 0.002954289084300399, + 0.0034415980335325003, + 0.00297786807641387, + 0.004816609434783459, + 0.0035988562740385532, + 0.0016590387094765902, + 0.003208970418199897, + 0.002560911001637578, + 0.00391776766628027, + 0.0018025467870756984, + 0.003511980175971985, + 0.004966358654201031, + 0.0036880625411868095, + 0.0027684615924954414, + 0.003267018822953105, + 0.0029237880371510983, + 0.002190216211602092, + 0.006452361587435007, + 0.006508020684123039, + 0.010007892735302448, + 0.006704118102788925, + 0.00885276310145855, + 0.006218637339770794, + 0.007746201474219561, + 0.006685556378215551, + 0.01291082613170147, + 0.026005670428276062, + 0.02324136719107628, + 0.017791366204619408, + 0.007708200253546238, + 0.024159623309969902, + 0.009882078506052494, + 0.008563673123717308, + 0.014109105803072453, + 0.04996345564723015, + 0.017962783575057983, + 0.014600592665374279, + 0.008307027630507946, + 0.054120711982250214, + 0.008044930174946785, + 0.008898184634745121 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 8 + ], + "selected_reward": -0.0052458494901657104 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a slightly upward angle from a fixed position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone figure wearing a loose, pure white robe standing atop a jagged, dark gray rock formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white robe's fabric visibly whipped violently by the wind, with sleeves and hem fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless expanse of mist-shrouded mountains under a deep, inky blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit low color saturation, sharp focus, and a slightly grainy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12702831625938416 + ], + "opposite_logprob_tokens": [ + -2.127028226852417 + ], + "logprob_reward": -0.12702831625938416, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12702831625938416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11496487259864807, + 0.02096305415034294, + 0.01795032061636448, + 0.1054053083062172, + 0.037748970091342926, + 0.10739308595657349, + 0.03164791688323021, + 0.028112217783927917, + 0.25105589628219604, + 0.2847583591938019 + ], + "credit_map": [ + 0.01367885246872902, + 0.012687421403825283, + 0.018380634486675262, + 0.015833452343940735, + 0.014799073338508606, + 0.015310453251004219, + 0.013754017651081085, + 0.010520966723561287, + 0.0021194054279476404, + 0.00275141978636384, + 0.0037956158630549908, + 0.0026697220746427774, + 0.0034834109246730804, + 0.002222271403297782, + 0.0023526186123490334, + 0.0015685888938605785, + 0.0020050136372447014, + 0.0028451092075556517, + 0.0033806264400482178, + 0.002620613668113947, + 0.0020480568055063486, + 0.0019534477032721043, + 0.0015015586977824569, + 0.0015958946896716952, + 0.010839501395821571, + 0.01524041872471571, + 0.01702253334224224, + 0.010704530403017998, + 0.012188361957669258, + 0.021233053877949715, + 0.009832581505179405, + 0.008344324305653572, + 0.002942409599199891, + 0.006596651393920183, + 0.0038376599550247192, + 0.0035769061651080847, + 0.005805252585560083, + 0.006685954984277487, + 0.004628411494195461, + 0.003675723448395729, + 0.010236541740596294, + 0.013937979005277157, + 0.012117079459130764, + 0.013657784089446068, + 0.022427035495638847, + 0.010234326124191284, + 0.014088260941207409, + 0.010694078169763088, + 0.0029380142223089933, + 0.005921473260968924, + 0.005762528162449598, + 0.0034782730508595705, + 0.0030952428933233023, + 0.003870166838169098, + 0.00428784778341651, + 0.0022943743970245123, + 0.003073540050536394, + 0.004058067686855793, + 0.004438706673681736, + 0.0022162669338285923, + 0.005241633392870426, + 0.0032621705904603004, + 0.0032401594799011946, + 0.0025816725101321936, + 0.032797567546367645, + 0.06543686240911484, + 0.04757584631443024, + 0.02349468134343624, + 0.0190462414175272, + 0.042213425040245056, + 0.010247877798974514, + 0.010243413969874382, + 0.024974724277853966, + 0.07098589092493057, + 0.03659426420927048, + 0.03162415325641632, + 0.01795930229127407, + 0.070560522377491, + 0.01643368974328041, + 0.015625793486833572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.12702831625938416 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, slightly upward shot from a fixed camera position captures the majestic scene atop the mountain with a documentary quality. At the center of the frame, a lone poet stands atop the highest point of a jagged, dark gray rock formation. He wears a loose, pure white robe, its thin cotton and linen material whipped violently in the fierce mountain breeze, its sleeves and hem fluttering and unfurling dramatically like a living banner. His long, dark hair, ruffled by the wind, contrasts sharply with the flowing white robe. Despite the distance, his upright figure is still recognizable, one hand clasped behind his back, gazing into the distance. The background is an endless expanse of mist-shrouded mountains, and the sky is a deep, almost inky blue, characteristic of high altitudes. The slanting afternoon sun penetrates the thin cloud cover, giving the white robe a sharp golden edge and casting long, sharply defined shadows on the rock. The color saturation of the entire picture is low, sharp and slightly grainy, creating a desolate, lonely, yet epic and tranquil atmosphere.", + "seed": 9859, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_8/index_3102_seed_9859_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07051145313307643, + "vlm_logprob": -0.07051145313307643, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.07051145313307643, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a slightly upward angle from a fixed position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone figure wearing a loose, pure white robe standing atop a jagged, dark gray rock formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white robe's fabric visibly whipped violently by the wind, with sleeves and hem fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless expanse of mist-shrouded mountains under a deep, inky blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit low color saturation, sharp focus, and a slightly grainy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14233705401420593 + ], + "opposite_logprob_tokens": [ + -2.1423370838165283 + ], + "logprob_reward": -0.14233705401420593, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.14233705401420593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11382076144218445, + 0.07368142902851105, + 0.03800463676452637, + 0.0561273917555809, + 0.1441020518541336, + 0.0669441968202591, + 0.26302334666252136, + 0.12373790889978409, + 0.08967716991901398, + 0.030881062150001526 + ], + "credit_map": [ + 0.011376850306987762, + 0.006897362414747477, + 0.0181317999958992, + 0.010449650697410107, + 0.016161097213625908, + 0.0202697291970253, + 0.014403785578906536, + 0.016130492091178894, + 0.010811123065650463, + 0.0043320562690496445, + 0.013148579746484756, + 0.007582575082778931, + 0.007714283186942339, + 0.01373258139938116, + 0.006572202313691378, + 0.009788035415112972, + 0.005093386396765709, + 0.0034608866553753614, + 0.005695204250514507, + 0.0035686541814357042, + 0.007127352990210056, + 0.00436751451343298, + 0.00463466253131628, + 0.004056977108120918, + 0.005162277724593878, + 0.0029066007118672132, + 0.006787599064409733, + 0.007865834049880505, + 0.005586488638073206, + 0.009928741492331028, + 0.009682157076895237, + 0.008207695558667183, + 0.017053311690688133, + 0.009626812301576138, + 0.02561042457818985, + 0.012943447567522526, + 0.012589411810040474, + 0.03936690464615822, + 0.011166981421411037, + 0.01574476808309555, + 0.008877758868038654, + 0.006649776827543974, + 0.007768880110234022, + 0.005293821915984154, + 0.007258734665811062, + 0.01600668393075466, + 0.007315982598811388, + 0.007772565353661776, + 0.03070293925702572, + 0.014614197425544262, + 0.043024174869060516, + 0.021399468183517456, + 0.024158967658877373, + 0.05731736496090889, + 0.036712225526571274, + 0.03509402647614479, + 0.014726389199495316, + 0.00934421643614769, + 0.018025165423750877, + 0.012017548084259033, + 0.015465263277292252, + 0.016466902568936348, + 0.0245071891695261, + 0.013185246847569942, + 0.008491897024214268, + 0.006323761772364378, + 0.01335381343960762, + 0.010083083994686604, + 0.00772476103156805, + 0.011668568477034569, + 0.012901868671178818, + 0.019129423424601555, + 0.004082514438778162, + 0.0016300504794344306, + 0.004687495529651642, + 0.0027487175539135933, + 0.0035760782193392515, + 0.006516634952276945, + 0.0026325909420847893, + 0.0050069838762283325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 7 + ], + "selected_reward": -0.14233705401420593 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a slightly upward angle from a fixed position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone figure wearing a loose, pure white robe standing atop a jagged, dark gray rock formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white robe's fabric visibly whipped violently by the wind, with sleeves and hem fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless expanse of mist-shrouded mountains under a deep, inky blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit low color saturation, sharp focus, and a slightly grainy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022055604495108128 + ], + "opposite_logprob_tokens": [ + -6.1272053718566895 + ], + "logprob_reward": -0.0022055604495108128, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022055604495108128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2134934961795807, + 0.08670744299888611, + 0.10549342632293701, + 0.04248714819550514, + 0.07268435508012772, + 0.2394275665283203, + 0.10118577629327774, + 0.04551153630018234, + 0.04922115430235863, + 0.043788012117147446 + ], + "credit_map": [ + 0.029482532292604446, + 0.011569427326321602, + 0.04795069992542267, + 0.020157143473625183, + 0.022425157949328423, + 0.03453296795487404, + 0.018507784232497215, + 0.028867775574326515, + 0.00717739574611187, + 0.006707107648253441, + 0.01538066752254963, + 0.008423738181591034, + 0.010127479210495949, + 0.015666397288441658, + 0.008293095044791698, + 0.01493156049400568, + 0.011735265143215656, + 0.009042195044457912, + 0.018791083246469498, + 0.015901761129498482, + 0.013419910334050655, + 0.010820920579135418, + 0.012233630754053593, + 0.013548660092055798, + 0.004035863094031811, + 0.004553795326501131, + 0.005522317718714476, + 0.006724832579493523, + 0.004713951610028744, + 0.005477360915392637, + 0.006248085293918848, + 0.005210944451391697, + 0.009369564242661, + 0.004493505693972111, + 0.010766115039587021, + 0.004989024251699448, + 0.01233790535479784, + 0.015371677465736866, + 0.005785746034234762, + 0.009570812806487083, + 0.026298290118575096, + 0.015980783849954605, + 0.03730498254299164, + 0.02278205193579197, + 0.02870739996433258, + 0.0552634634077549, + 0.026091627776622772, + 0.026998983696103096, + 0.014783608727157116, + 0.004745774902403355, + 0.011826335452497005, + 0.0065128253772854805, + 0.009795507416129112, + 0.0262334942817688, + 0.013424301519989967, + 0.01386394165456295, + 0.005329187493771315, + 0.0027172400150448084, + 0.007992170751094818, + 0.004506368190050125, + 0.006060007493942976, + 0.008414178155362606, + 0.005358658265322447, + 0.005133729428052902, + 0.006670984905213118, + 0.002742273034527898, + 0.004960448015481234, + 0.0043439255096018314, + 0.0052639273926615715, + 0.013593249022960663, + 0.003983514383435249, + 0.007662832736968994, + 0.006143063772469759, + 0.003717900952324271, + 0.006861498113721609, + 0.004730810411274433, + 0.004364852327853441, + 0.007229730486869812, + 0.005848302505910397, + 0.004891851916909218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.0022055604495108128 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a slightly upward angle from a fixed position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone figure wearing a loose, pure white robe standing atop a jagged, dark gray rock formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white robe's fabric visibly whipped violently by the wind, with sleeves and hem fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless expanse of mist-shrouded mountains under a deep, inky blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit low color saturation, sharp focus, and a slightly grainy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306282848119736 + ], + "opposite_logprob_tokens": [ + -3.168062925338745 + ], + "logprob_reward": -0.04306282848119736, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306282848119736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07867735624313354, + 0.055996403098106384, + 0.31691813468933105, + 0.03878132253885269, + 0.16571487486362457, + 0.05694554001092911, + 0.1149250715970993, + 0.1194692924618721, + 0.03243274614214897, + 0.020139237865805626 + ], + "credit_map": [ + 0.009202796965837479, + 0.009315629489719868, + 0.009058035910129547, + 0.006611540913581848, + 0.013121680356562138, + 0.011291462928056717, + 0.007183263543993235, + 0.012892959639430046, + 0.00650037731975317, + 0.0046215299516916275, + 0.006932002026587725, + 0.005710357800126076, + 0.004943630192428827, + 0.010239247232675552, + 0.006912125740200281, + 0.010137136094272137, + 0.03255888447165489, + 0.025707116350531578, + 0.04308939352631569, + 0.02784453146159649, + 0.05676127225160599, + 0.045477792620658875, + 0.05035639926791191, + 0.03512274846434593, + 0.004514704458415508, + 0.0030828260350972414, + 0.005711760371923447, + 0.005907660350203514, + 0.004084118641912937, + 0.006075556389987469, + 0.004403516184538603, + 0.005001181270927191, + 0.02193530462682247, + 0.009050232358276844, + 0.03284623473882675, + 0.01389814168214798, + 0.012868940830230713, + 0.04658458009362221, + 0.012964536435902119, + 0.015566903166472912, + 0.005208840128034353, + 0.005017600953578949, + 0.007612451445311308, + 0.006274342071264982, + 0.008487467654049397, + 0.009747899137437344, + 0.008314529433846474, + 0.006282413844019175, + 0.01166777778416872, + 0.006468248087912798, + 0.01745910383760929, + 0.009838650934398174, + 0.011165347881615162, + 0.02318708598613739, + 0.01761912740767002, + 0.017519744113087654, + 0.01346530206501484, + 0.00815882720053196, + 0.01769978366792202, + 0.011539840139448643, + 0.014882488176226616, + 0.016327517107129097, + 0.023455116897821426, + 0.013940426521003246, + 0.0029601329006254673, + 0.001909982063807547, + 0.0065138498321175575, + 0.00466161547228694, + 0.0034871120005846024, + 0.004087237641215324, + 0.004880605265498161, + 0.003932213876396418, + 0.002488781465217471, + 0.0014290270628407598, + 0.0034911015536636114, + 0.00242054695263505, + 0.0019462591735646129, + 0.0028119434136897326, + 0.0022408957593142986, + 0.0033106834162026644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 7 + ], + "selected_reward": -0.04306282848119736 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a slightly upward angle from a fixed position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone figure wearing a loose, pure white robe standing atop a jagged, dark gray rock formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white robe's fabric visibly whipped violently by the wind, with sleeves and hem fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless expanse of mist-shrouded mountains under a deep, inky blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit low color saturation, sharp focus, and a slightly grainy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046304380521178246 + ], + "opposite_logprob_tokens": [ + -5.3796305656433105 + ], + "logprob_reward": -0.0046304380521178246, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046304380521178246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05583825707435608, + 0.049145814031362534, + 0.01486114226281643, + 0.03781246021389961, + 0.3739049434661865, + 0.31602662801742554, + 0.056573737412691116, + 0.016730455681681633, + 0.0464780330657959, + 0.03262849524617195 + ], + "credit_map": [ + 0.006034230813384056, + 0.004036537371575832, + 0.014076491817831993, + 0.0052816541865468025, + 0.005081757437437773, + 0.009789709933102131, + 0.004450791981071234, + 0.007087085861712694, + 0.004747170489281416, + 0.0024616948794573545, + 0.0060922252014279366, + 0.003965908661484718, + 0.004382848273962736, + 0.012317911721765995, + 0.0046318392269313335, + 0.010546214878559113, + 0.0016674959333613515, + 0.0015363943530246615, + 0.0026514558121562004, + 0.0022255086805671453, + 0.0017572175711393356, + 0.0020474004559218884, + 0.0013332042144611478, + 0.0016424652421846986, + 0.0036681420169770718, + 0.001573077985085547, + 0.009329961612820625, + 0.004526125732809305, + 0.0022967462427914143, + 0.004492584615945816, + 0.006795292720198631, + 0.0051305294036865234, + 0.04822196066379547, + 0.023883357644081116, + 0.06710124760866165, + 0.03989817574620247, + 0.03031005896627903, + 0.09771022200584412, + 0.029637953266501427, + 0.03714194521307945, + 0.03447706624865532, + 0.024772996082901955, + 0.05151762068271637, + 0.03439398109912872, + 0.03663596138358116, + 0.06433483213186264, + 0.03733866289258003, + 0.03255552425980568, + 0.0069910031743347645, + 0.0037052335683256388, + 0.007770906202495098, + 0.005052383989095688, + 0.0059318142011761665, + 0.011559292674064636, + 0.007834609597921371, + 0.00772849703207612, + 0.0022872709669172764, + 0.0007936559268273413, + 0.002343920059502125, + 0.001234022551216185, + 0.0020137736573815346, + 0.0033554506953805685, + 0.0023792486172169447, + 0.0023231138475239277, + 0.0056204721331596375, + 0.0022507577668875456, + 0.006451476830989122, + 0.0024839274119585752, + 0.004740156698971987, + 0.010552634485065937, + 0.0039013156201690435, + 0.010477294214069843, + 0.0027774993795901537, + 0.0011801138753071427, + 0.0072932555340230465, + 0.001625047530978918, + 0.0030504418537020683, + 0.005764932371675968, + 0.0025686449371278286, + 0.008368561044335365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.0046304380521178246 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a slightly upward angle from a fixed position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone figure wearing a loose, pure white robe standing atop a jagged, dark gray rock formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white robe's fabric visibly whipped violently by the wind, with sleeves and hem fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless expanse of mist-shrouded mountains under a deep, inky blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit low color saturation, sharp focus, and a slightly grainy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16032138466835022 + ], + "opposite_logprob_tokens": [ + -1.9103213548660278 + ], + "logprob_reward": -0.16032138466835022, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16032138466835022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08167844265699387, + 0.0635988861322403, + 0.04509944096207619, + 0.03029971942305565, + 0.20304422080516815, + 0.4350385367870331, + 0.042272139340639114, + 0.02381531521677971, + 0.04511456936597824, + 0.03003864176571369 + ], + "credit_map": [ + 0.008639688603579998, + 0.008148516528308392, + 0.012098334729671478, + 0.0070341466926038265, + 0.009974947199225426, + 0.013024665415287018, + 0.0078027211129665375, + 0.014955430291593075, + 0.0050492314621806145, + 0.003045093035325408, + 0.007557637523859739, + 0.006580385845154524, + 0.005537827033549547, + 0.016208387911319733, + 0.005653060507029295, + 0.013967265374958515, + 0.005254415795207024, + 0.0043246569111943245, + 0.00781043479219079, + 0.0039509497582912445, + 0.004820163827389479, + 0.008729021064937115, + 0.0046973698772490025, + 0.005512434523552656, + 0.0029813293367624283, + 0.0018553712870925665, + 0.00399089464917779, + 0.004392764065414667, + 0.0034908049274235964, + 0.0035438609775155783, + 0.0055312891490757465, + 0.004513408523052931, + 0.019846739247441292, + 0.019383784383535385, + 0.03196144476532936, + 0.04003319889307022, + 0.02239077165722847, + 0.02249058708548546, + 0.023256294429302216, + 0.02368142455816269, + 0.06595519930124283, + 0.04523487389087677, + 0.05737033858895302, + 0.04459714889526367, + 0.054317720234394073, + 0.06979755312204361, + 0.05767698585987091, + 0.04008876904845238, + 0.004617398604750633, + 0.0030015413649380207, + 0.0060402024537324905, + 0.004401037935167551, + 0.00518924742937088, + 0.004932092502713203, + 0.007634390611201525, + 0.006456236820667982, + 0.002786208176985383, + 0.0015259503852576017, + 0.003874785266816616, + 0.0022722494322806597, + 0.0028968395199626684, + 0.0031448076479136944, + 0.00448420038446784, + 0.002830278128385544, + 0.004222742281854153, + 0.003604555269703269, + 0.0061639281921088696, + 0.005295838229358196, + 0.004694956820458174, + 0.007602469529956579, + 0.006112202070653439, + 0.007417881861329079, + 0.004327308386564255, + 0.00241973833180964, + 0.004607424605637789, + 0.002515666652470827, + 0.003163982182741165, + 0.005980511195957661, + 0.0027994676493108273, + 0.00422454671934247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.16032138466835022 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, slightly upward shot from a fixed camera position captures the majestic scene atop the mountain with a documentary quality. At the center of the frame, a lone poet stands atop the highest point of a jagged, dark gray rock formation. He wears a loose, pure white robe, its thin cotton and linen material whipped violently in the fierce mountain breeze, its sleeves and hem fluttering and unfurling dramatically like a living banner. His long, dark hair, ruffled by the wind, contrasts sharply with the flowing white robe. Despite the distance, his upright figure is still recognizable, one hand clasped behind his back, gazing into the distance. The background is an endless expanse of mist-shrouded mountains, and the sky is a deep, almost inky blue, characteristic of high altitudes. The slanting afternoon sun penetrates the thin cloud cover, giving the white robe a sharp golden edge and casting long, sharply defined shadows on the rock. The color saturation of the entire picture is low, sharp and slightly grainy, creating a desolate, lonely, yet epic and tranquil atmosphere.", + "seed": 9860, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_8/index_3102_seed_9860_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07617512429133058, + "vlm_logprob": -0.07617512429133058, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.07617512429133058, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a slightly upward angle from a fixed position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone figure wearing a loose, pure white robe standing atop a jagged, dark gray rock formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white robe's fabric visibly whipped violently by the wind, with sleeves and hem fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless expanse of mist-shrouded mountains under a deep, inky blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit low color saturation, sharp focus, and a slightly grainy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05276425555348396 + ], + "opposite_logprob_tokens": [ + -3.1777641773223877 + ], + "logprob_reward": -0.05276425555348396, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.05276425555348396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18553268909454346, + 0.29357290267944336, + 0.05437380075454712, + 0.05567228049039841, + 0.18102853000164032, + 0.06145566329360008, + 0.030121177434921265, + 0.06935954093933105, + 0.027925295755267143, + 0.04095808044075966 + ], + "credit_map": [ + 0.025299319997429848, + 0.02319902740418911, + 0.03886618837714195, + 0.014193854294717312, + 0.01646731235086918, + 0.03121541440486908, + 0.014255698770284653, + 0.022035904228687286, + 0.061184100806713104, + 0.01625707931816578, + 0.04103919863700867, + 0.01760798692703247, + 0.012067237868905067, + 0.09533844143152237, + 0.024516426026821136, + 0.02556246891617775, + 0.006073597818613052, + 0.0036067801993340254, + 0.009214579127728939, + 0.005465021822601557, + 0.006071790121495724, + 0.01154459360986948, + 0.005300300661474466, + 0.007097139954566956, + 0.005265537649393082, + 0.005380900111049414, + 0.009940609335899353, + 0.004696843679994345, + 0.007762485183775425, + 0.007725631818175316, + 0.006873322185128927, + 0.008026954717934132, + 0.029875651001930237, + 0.01662854291498661, + 0.027214795351028442, + 0.021571427583694458, + 0.014903596602380276, + 0.02440149337053299, + 0.019987870007753372, + 0.02644517458975315, + 0.01094595156610012, + 0.006232051178812981, + 0.011157737113535404, + 0.006493129301816225, + 0.00574621744453907, + 0.007413973566144705, + 0.006888320203870535, + 0.006578289438039064, + 0.004284767899662256, + 0.002682826016098261, + 0.004447444807738066, + 0.0042843385599553585, + 0.003447955474257469, + 0.003929843660444021, + 0.00401673186570406, + 0.003027272643521428, + 0.005430597346276045, + 0.016989341005682945, + 0.0051211523823440075, + 0.004640392493456602, + 0.004724473692476749, + 0.0199893731623888, + 0.006745434366166592, + 0.005718783941119909, + 0.0030442788265645504, + 0.0034519904293119907, + 0.004767449107021093, + 0.0028927880339324474, + 0.0031327232718467712, + 0.0036829423625022173, + 0.004250559490174055, + 0.002702568657696247, + 0.0050055780448019505, + 0.0064401887357234955, + 0.005572108086198568, + 0.004101223312318325, + 0.005707206670194864, + 0.005831076297909021, + 0.0037717174272984266, + 0.004528984893113375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.05276425555348396 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a slightly upward angle from a fixed position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone figure wearing a loose, pure white robe standing atop a jagged, dark gray rock formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white robe's fabric visibly whipped violently by the wind, with sleeves and hem fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless expanse of mist-shrouded mountains under a deep, inky blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit low color saturation, sharp focus, and a slightly grainy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004107373300939798 + ], + "opposite_logprob_tokens": [ + -5.504107475280762 + ], + "logprob_reward": -0.004107373300939798, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004107373300939798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07086948305368423, + 0.5182721614837646, + 0.08995826542377472, + 0.09326174855232239, + 0.06968869268894196, + 0.06009852886199951, + 0.024193473160266876, + 0.03354528173804283, + 0.017785051837563515, + 0.022327275946736336 + ], + "credit_map": [ + 0.012078126892447472, + 0.0057297260500490665, + 0.01029101200401783, + 0.005917171016335487, + 0.0040845079347491264, + 0.0216436218470335, + 0.004673262592405081, + 0.0064520565792918205, + 0.10579098016023636, + 0.026950592175126076, + 0.0786738395690918, + 0.034328531473875046, + 0.020332815125584602, + 0.16317114233970642, + 0.04362298548221588, + 0.04540126770734787, + 0.012305708602070808, + 0.006717512849718332, + 0.014255162328481674, + 0.008540610782802105, + 0.008612685836851597, + 0.018594691529870033, + 0.009307998232543468, + 0.011623896658420563, + 0.008024606853723526, + 0.008635500445961952, + 0.014327142387628555, + 0.012966102920472622, + 0.0152494627982378, + 0.008084457367658615, + 0.012656054459512234, + 0.013318419456481934, + 0.011320378631353378, + 0.006650310475379229, + 0.01027292013168335, + 0.008422525599598885, + 0.005427903030067682, + 0.010221289470791817, + 0.007989080622792244, + 0.00938428658992052, + 0.009956970810890198, + 0.006799886468797922, + 0.010224110446870327, + 0.00739414943382144, + 0.004999987781047821, + 0.007820721715688705, + 0.0052965618669986725, + 0.007606134749948978, + 0.004154028836637735, + 0.002431123051792383, + 0.003775399411097169, + 0.0031070627737790346, + 0.0021078772842884064, + 0.003321348922327161, + 0.0029726088978350163, + 0.0023240232840180397, + 0.002920767292380333, + 0.007474485784769058, + 0.002432588953524828, + 0.0030941436998546124, + 0.002269085729494691, + 0.008488845080137253, + 0.0042444984428584576, + 0.0026208662893623114, + 0.0014994734665378928, + 0.0020268121734261513, + 0.0034908282104879618, + 0.002303267363458872, + 0.00191026134416461, + 0.0020801196806132793, + 0.0023732767440378666, + 0.002101013669744134, + 0.0023636785335838795, + 0.003136703744530678, + 0.004179050680249929, + 0.00317192729562521, + 0.0015598866157233715, + 0.003154918784275651, + 0.0027100269217044115, + 0.0020510838367044926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.004107373300939798 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a slightly upward angle from a fixed position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone figure wearing a loose, pure white robe standing atop a jagged, dark gray rock formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white robe's fabric visibly whipped violently by the wind, with sleeves and hem fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless expanse of mist-shrouded mountains under a deep, inky blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit low color saturation, sharp focus, and a slightly grainy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866091161966324 + ], + "opposite_logprob_tokens": [ + -3.04866099357605 + ], + "logprob_reward": -0.04866091161966324, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866091161966324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08997581154108047, + 0.12468820810317993, + 0.16328009963035583, + 0.13951118290424347, + 0.06709793210029602, + 0.04723706468939781, + 0.02973390370607376, + 0.27137255668640137, + 0.029713565483689308, + 0.037389691919088364 + ], + "credit_map": [ + 0.01686245948076248, + 0.0060748085379600525, + 0.012555338442325592, + 0.005328229628503323, + 0.005343970376998186, + 0.03223240002989769, + 0.004549304023385048, + 0.007029297761619091, + 0.025362977758049965, + 0.006935392040759325, + 0.019684448838233948, + 0.009644600562751293, + 0.006026982795447111, + 0.03526206687092781, + 0.010255416855216026, + 0.011516328901052475, + 0.017501067370176315, + 0.013839389197528362, + 0.030665623024106026, + 0.01985097862780094, + 0.015618870966136456, + 0.02162269502878189, + 0.0194595605134964, + 0.02472190372645855, + 0.0140091422945261, + 0.011577868834137917, + 0.02216295711696148, + 0.02070648781955242, + 0.02042858488857746, + 0.01534125953912735, + 0.014914012514054775, + 0.020370883867144585, + 0.007601748686283827, + 0.005579626653343439, + 0.013437329791486263, + 0.007640082389116287, + 0.007154372986406088, + 0.009533561766147614, + 0.007449983153492212, + 0.008701225742697716, + 0.006836545653641224, + 0.0051859719678759575, + 0.008218824863433838, + 0.005767348222434521, + 0.003919912036508322, + 0.00677006458863616, + 0.004679563455283642, + 0.005858832038938999, + 0.003348597791045904, + 0.004974575247615576, + 0.003816381562501192, + 0.003208807436749339, + 0.001985428389161825, + 0.0062085348181426525, + 0.0031603844836354256, + 0.003031193045899272, + 0.01819293573498726, + 0.05615736171603203, + 0.021399173885583878, + 0.02622961439192295, + 0.021790549159049988, + 0.06340418756008148, + 0.04051637649536133, + 0.02368234097957611, + 0.002087309258058667, + 0.003255299525335431, + 0.0038117324002087116, + 0.0037784911692142487, + 0.004257390275597572, + 0.0036369774024933577, + 0.005279758479446173, + 0.0036066072061657906, + 0.003127399832010269, + 0.004908881615847349, + 0.0031531541608273983, + 0.005440493579953909, + 0.007842238061130047, + 0.005028435029089451, + 0.004446578677743673, + 0.0034425100311636925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 3 + ], + "selected_reward": -0.04866091161966324 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a slightly upward angle from a fixed position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone figure wearing a loose, pure white robe standing atop a jagged, dark gray rock formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white robe's fabric visibly whipped violently by the wind, with sleeves and hem fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless expanse of mist-shrouded mountains under a deep, inky blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit low color saturation, sharp focus, and a slightly grainy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023273399099707603 + ], + "opposite_logprob_tokens": [ + -3.773273468017578 + ], + "logprob_reward": -0.023273399099707603, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023273399099707603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08796954900026321, + 0.24780844151973724, + 0.13713541626930237, + 0.01676936447620392, + 0.16666178405284882, + 0.20919591188430786, + 0.09167986363172531, + 0.01693536527454853, + 0.015944698825478554, + 0.009899616241455078 + ], + "credit_map": [ + 0.01283853780478239, + 0.009257698431611061, + 0.0147218257188797, + 0.006636566016823053, + 0.005419598892331123, + 0.023823974654078484, + 0.005643381737172604, + 0.009627968072891235, + 0.051178060472011566, + 0.013230438344180584, + 0.03711759299039841, + 0.015762807801365852, + 0.009604043327271938, + 0.0784812644124031, + 0.021353185176849365, + 0.02108105644583702, + 0.023244716227054596, + 0.010604871436953545, + 0.020918944850564003, + 0.014428846538066864, + 0.013354579918086529, + 0.029068924486637115, + 0.011242141015827656, + 0.014272396452724934, + 0.0017297433223575354, + 0.001692502060905099, + 0.002802294911816716, + 0.0019503829535096884, + 0.0021194785367697477, + 0.0018553092377260327, + 0.002228871453553438, + 0.002390780486166477, + 0.027198396623134613, + 0.01465307641774416, + 0.029263528063893318, + 0.019695527851581573, + 0.013414200395345688, + 0.02299582213163376, + 0.017895139753818512, + 0.02154609188437462, + 0.037428829818964005, + 0.023264268413186073, + 0.03633998706936836, + 0.02461518719792366, + 0.015331810340285301, + 0.024946322664618492, + 0.019213847815990448, + 0.028055645525455475, + 0.01709545962512493, + 0.011961523443460464, + 0.013210266828536987, + 0.009288512170314789, + 0.005278297699987888, + 0.01823578029870987, + 0.008503668941557407, + 0.00810635183006525, + 0.0014186287298798561, + 0.0031955435406416655, + 0.002052928786724806, + 0.0017563884612172842, + 0.0012689445866271853, + 0.0035714837722480297, + 0.001903184107504785, + 0.0017682622419670224, + 0.002319306368008256, + 0.002353230956941843, + 0.0023946473374962807, + 0.0013691708445549011, + 0.0017627846682444215, + 0.0020169983617961407, + 0.001787374378181994, + 0.0019411870744079351, + 0.0009478201391175389, + 0.0017924968851730227, + 0.001228963490575552, + 0.0011072300840169191, + 0.000693555804900825, + 0.0023087032604962587, + 0.0009299339726567268, + 0.0008909129537642002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 4 + ], + "selected_reward": -0.023273399099707603 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a slightly upward angle from a fixed position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone figure wearing a loose, pure white robe standing atop a jagged, dark gray rock formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white robe's fabric visibly whipped violently by the wind, with sleeves and hem fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless expanse of mist-shrouded mountains under a deep, inky blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit low color saturation, sharp focus, and a slightly grainy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520696818828583 + ], + "opposite_logprob_tokens": [ + -1.5020697116851807 + ], + "logprob_reward": -0.2520696818828583, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520696818828583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21273303031921387, + 0.12688887119293213, + 0.06801949441432953, + 0.07305470108985901, + 0.15539810061454773, + 0.1538453996181488, + 0.08005393296480179, + 0.037136830389499664, + 0.06622530519962311, + 0.02664436399936676 + ], + "credit_map": [ + 0.03511509671807289, + 0.023279240354895592, + 0.037111248821020126, + 0.014206035062670708, + 0.012391428463160992, + 0.05099806562066078, + 0.01686566136777401, + 0.022766249254345894, + 0.02788838930428028, + 0.021122470498085022, + 0.015280584804713726, + 0.010650197975337505, + 0.008814850822091103, + 0.02575923129916191, + 0.0069777281023561954, + 0.010395403951406479, + 0.015535821206867695, + 0.004558328073471785, + 0.008365143090486526, + 0.004506342578679323, + 0.004257753025740385, + 0.0219992995262146, + 0.0037816164549440145, + 0.005015186034142971, + 0.007181537337601185, + 0.006187668070197105, + 0.009274063631892204, + 0.010291608050465584, + 0.013955029658973217, + 0.007864135317504406, + 0.009829714894294739, + 0.008470937609672546, + 0.0268913172185421, + 0.014852209948003292, + 0.022849926725029945, + 0.017061296850442886, + 0.012496540322899818, + 0.026100756600499153, + 0.01614714041352272, + 0.01899890787899494, + 0.024461401626467705, + 0.01966293714940548, + 0.026693396270275116, + 0.017658257856965065, + 0.011119391769170761, + 0.02196921966969967, + 0.013621095567941666, + 0.018659690394997597, + 0.013510003685951233, + 0.010571788065135479, + 0.012137466110289097, + 0.007999575696885586, + 0.005015619099140167, + 0.01436504814773798, + 0.008612261153757572, + 0.00784215983003378, + 0.004277897533029318, + 0.003579524578526616, + 0.005524123087525368, + 0.004962149076163769, + 0.0052471463568508625, + 0.003479883074760437, + 0.005651199258863926, + 0.004414905793964863, + 0.004799928516149521, + 0.007036156486719847, + 0.011056031100451946, + 0.008790171705186367, + 0.0072785476222634315, + 0.007548291701823473, + 0.010922725312411785, + 0.008793451823294163, + 0.0020587376784533262, + 0.003384183393791318, + 0.004987300839275122, + 0.003940073773264885, + 0.0019791878294199705, + 0.004016840364784002, + 0.0029523265548050404, + 0.0033257128670811653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.2520696818828583 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, slightly upward shot from a fixed camera position captures the majestic scene atop the mountain with a documentary quality. At the center of the frame, a lone poet stands atop the highest point of a jagged, dark gray rock formation. He wears a loose, pure white robe, its thin cotton and linen material whipped violently in the fierce mountain breeze, its sleeves and hem fluttering and unfurling dramatically like a living banner. His long, dark hair, ruffled by the wind, contrasts sharply with the flowing white robe. Despite the distance, his upright figure is still recognizable, one hand clasped behind his back, gazing into the distance. The background is an endless expanse of mist-shrouded mountains, and the sky is a deep, almost inky blue, characteristic of high altitudes. The slanting afternoon sun penetrates the thin cloud cover, giving the white robe a sharp golden edge and casting long, sharply defined shadows on the rock. The color saturation of the entire picture is low, sharp and slightly grainy, creating a desolate, lonely, yet epic and tranquil atmosphere.", + "seed": 9861, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_8/index_3102_seed_9861_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.048007619008421897, + "vlm_logprob": -0.048007619008421897, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.048007619008421897, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a slightly upward angle from a fixed position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone figure wearing a loose, pure white robe standing atop a jagged, dark gray rock formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white robe's fabric visibly whipped violently by the wind, with sleeves and hem fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless expanse of mist-shrouded mountains under a deep, inky blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit low color saturation, sharp focus, and a slightly grainy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03841809928417206 + ], + "opposite_logprob_tokens": [ + -3.5384180545806885 + ], + "logprob_reward": -0.03841809928417206, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03841809928417206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.264051616191864, + 0.07933760434389114, + 0.048203788697719574, + 0.03294609859585762, + 0.10778475552797318, + 0.046144552528858185, + 0.05577855184674263, + 0.02399236522614956, + 0.2738063633441925, + 0.06795434653759003 + ], + "credit_map": [ + 0.02764364518225193, + 0.06216869875788689, + 0.04647284001111984, + 0.029879573732614517, + 0.01627087965607643, + 0.03809770941734314, + 0.020679423585534096, + 0.022838840261101723, + 0.008943907916545868, + 0.017132097855210304, + 0.008015895262360573, + 0.011520965956151485, + 0.006414442323148251, + 0.0118513572961092, + 0.008053306490182877, + 0.0074056340381503105, + 0.006536422297358513, + 0.010175093077123165, + 0.0057534342631697655, + 0.004540855530649424, + 0.004925751592963934, + 0.00799576286226511, + 0.004064450040459633, + 0.004212021827697754, + 0.003715895814821124, + 0.006386458873748779, + 0.004097757861018181, + 0.002298277337104082, + 0.004620467312633991, + 0.006087671034038067, + 0.0032835439778864384, + 0.002456024056300521, + 0.007414109539240599, + 0.030249474570155144, + 0.007973913103342056, + 0.007219031453132629, + 0.007808556780219078, + 0.026311185210943222, + 0.012749891728162766, + 0.008058600127696991, + 0.00580691359937191, + 0.0068634808994829655, + 0.0063689099624753, + 0.004862606525421143, + 0.004703923128545284, + 0.00685862498357892, + 0.006374428980052471, + 0.004305662587285042, + 0.010350441560149193, + 0.007774385157972574, + 0.008677178993821144, + 0.0054015167988836765, + 0.005268734879791737, + 0.00868090707808733, + 0.005600302014499903, + 0.004025086760520935, + 0.003094561630859971, + 0.004836645442992449, + 0.0036232909187674522, + 0.00215251580812037, + 0.00231182761490345, + 0.004246286582201719, + 0.002065556589514017, + 0.0016616802895441651, + 0.017600176855921745, + 0.06015097722411156, + 0.03414762392640114, + 0.02199491113424301, + 0.018817514181137085, + 0.08615312725305557, + 0.01788693480193615, + 0.017055056989192963, + 0.010357549414038658, + 0.01254008337855339, + 0.010686112567782402, + 0.005608046427369118, + 0.0034465487115085125, + 0.016079753637313843, + 0.004567311145365238, + 0.004668947774916887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.03841809928417206 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a slightly upward angle from a fixed position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone figure wearing a loose, pure white robe standing atop a jagged, dark gray rock formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white robe's fabric visibly whipped violently by the wind, with sleeves and hem fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless expanse of mist-shrouded mountains under a deep, inky blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit low color saturation, sharp focus, and a slightly grainy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002499910071492195 + ], + "opposite_logprob_tokens": [ + -6.002500057220459 + ], + "logprob_reward": -0.002499910071492195, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002499910071492195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14563511312007904, + 0.047837335616350174, + 0.04495209455490112, + 0.08391421288251877, + 0.14339932799339294, + 0.030172400176525116, + 0.06234929710626602, + 0.03527366742491722, + 0.29269683361053467, + 0.11376973986625671 + ], + "credit_map": [ + 0.012145685032010078, + 0.019251329824328423, + 0.02829800173640251, + 0.016800006851553917, + 0.016205351799726486, + 0.020754076540470123, + 0.016693158075213432, + 0.015487506985664368, + 0.005276481620967388, + 0.008569341152906418, + 0.00546659342944622, + 0.007616003975272179, + 0.004485605750232935, + 0.006870659533888102, + 0.005007247440516949, + 0.0045453994534909725, + 0.005027559120208025, + 0.007107425015419722, + 0.005648525897413492, + 0.006014829967170954, + 0.004341326653957367, + 0.00798046588897705, + 0.00427756505087018, + 0.004554402083158493, + 0.009263584390282631, + 0.018749281764030457, + 0.009417983703315258, + 0.0054909419268369675, + 0.013471249490976334, + 0.015405152924358845, + 0.006995595991611481, + 0.005120421294122934, + 0.012301710434257984, + 0.04624000936746597, + 0.01009302120655775, + 0.01014342438429594, + 0.005102887284010649, + 0.039469778537750244, + 0.011923043988645077, + 0.008125446736812592, + 0.004427694715559483, + 0.004871707875281572, + 0.004250558093190193, + 0.0033561226446181536, + 0.0028195756021887064, + 0.0048276628367602825, + 0.003202317748218775, + 0.002416760427877307, + 0.011149371974170208, + 0.007763836067169905, + 0.012895419262349606, + 0.006817998364567757, + 0.0040727281011641026, + 0.009322105906903744, + 0.005163136869668961, + 0.00516469543799758, + 0.0028052586130797863, + 0.010683179832994938, + 0.00258138426579535, + 0.002018857514485717, + 0.0019288805779069662, + 0.009737890213727951, + 0.0035518165677785873, + 0.0019664019346237183, + 0.01878674328327179, + 0.06234045699238777, + 0.03525012731552124, + 0.019052891060709953, + 0.022467229515314102, + 0.09220200031995773, + 0.02432701364159584, + 0.018270350992679596, + 0.014318848960101604, + 0.023119954392313957, + 0.011755446903407574, + 0.008751037530601025, + 0.011861433275043964, + 0.022614434361457825, + 0.012624961324036121, + 0.008723624050617218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.002499910071492195 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a slightly upward angle from a fixed position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone figure wearing a loose, pure white robe standing atop a jagged, dark gray rock formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white robe's fabric visibly whipped violently by the wind, with sleeves and hem fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless expanse of mist-shrouded mountains under a deep, inky blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit low color saturation, sharp focus, and a slightly grainy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905690908432007 + ], + "opposite_logprob_tokens": [ + -2.464056968688965 + ], + "logprob_reward": -0.08905690908432007, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905690908432007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04546831548213959, + 0.02069542370736599, + 0.0154139818623662, + 0.02609432116150856, + 0.09336448460817337, + 0.021932311356067657, + 0.04352766275405884, + 0.01889420486986637, + 0.6622791290283203, + 0.052330173552036285 + ], + "credit_map": [ + 0.003856044262647629, + 0.009035279043018818, + 0.008813155815005302, + 0.005379387643188238, + 0.0034019509330391884, + 0.006742548663169146, + 0.003639903850853443, + 0.004600043408572674, + 0.0019685218576341867, + 0.0029666984919458628, + 0.002667392836883664, + 0.0035054213367402554, + 0.0030254602897912264, + 0.002668467117473483, + 0.00201767822727561, + 0.0018757825018838048, + 0.0019876547157764435, + 0.002365092048421502, + 0.001864815247245133, + 0.001492402865551412, + 0.0022340621799230576, + 0.0026890758890658617, + 0.001286385697312653, + 0.0014944933354854584, + 0.0019877320155501366, + 0.005069348029792309, + 0.0023587627802044153, + 0.001946312258951366, + 0.00393452774733305, + 0.006353373173624277, + 0.0025199996307492256, + 0.001924264244735241, + 0.005318040493875742, + 0.031235434114933014, + 0.004927629604935646, + 0.004333363845944405, + 0.006451849360018969, + 0.028255071491003036, + 0.008337902836501598, + 0.004505190532654524, + 0.0018961685709655285, + 0.003575388342142105, + 0.0026174369268119335, + 0.0019089507404714823, + 0.0028038490563631058, + 0.004650630056858063, + 0.0024799543898552656, + 0.001999931875616312, + 0.009648715145885944, + 0.004706273321062326, + 0.006835280917584896, + 0.005148338153958321, + 0.0030203047208487988, + 0.007606609724462032, + 0.0033164932392537594, + 0.0032456431072205305, + 0.0017814134480431676, + 0.004543560557067394, + 0.002055960241705179, + 0.0016747921472415328, + 0.0018488046480342746, + 0.003917987924069166, + 0.0017544060247018933, + 0.0013172795297577977, + 0.03745438531041145, + 0.15082301199436188, + 0.08695220202207565, + 0.03679967671632767, + 0.03614340350031853, + 0.22731991112232208, + 0.05119858682155609, + 0.03558797761797905, + 0.00268314266577363, + 0.01049787737429142, + 0.008972871117293835, + 0.0030289930291473866, + 0.0033080845605582, + 0.018153637647628784, + 0.003428659401834011, + 0.0022569047287106514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 9 + ], + "selected_reward": -0.08905690908432007 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a slightly upward angle from a fixed position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone figure wearing a loose, pure white robe standing atop a jagged, dark gray rock formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white robe's fabric visibly whipped violently by the wind, with sleeves and hem fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless expanse of mist-shrouded mountains under a deep, inky blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit low color saturation, sharp focus, and a slightly grainy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009772907942533493 + ], + "opposite_logprob_tokens": [ + -4.634772777557373 + ], + "logprob_reward": -0.009772907942533493, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009772907942533493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12770526111125946, + 0.15533055365085602, + 0.06201214715838432, + 0.035894349217414856, + 0.03557306155562401, + 0.021609297022223473, + 0.04525059461593628, + 0.0411171056330204, + 0.2364315688610077, + 0.23907609283924103 + ], + "credit_map": [ + 0.010969111695885658, + 0.013932731002569199, + 0.021296994760632515, + 0.017773237079381943, + 0.013814586214721203, + 0.014771775342524052, + 0.019620167091488838, + 0.01552665326744318, + 0.0154773173853755, + 0.02767655998468399, + 0.02398102730512619, + 0.022672416642308235, + 0.0077489945106208324, + 0.026724115014076233, + 0.015283606946468353, + 0.015766499564051628, + 0.008675818331539631, + 0.008812404237687588, + 0.008255082182586193, + 0.0074726250022649765, + 0.006168209481984377, + 0.008913898840546608, + 0.007073722779750824, + 0.00664038211107254, + 0.004320131149142981, + 0.00538821704685688, + 0.005580848548561335, + 0.003655198961496353, + 0.004008296877145767, + 0.006373502314090729, + 0.0038631078787148, + 0.0027050464414060116, + 0.005071834661066532, + 0.009174289181828499, + 0.003706736955791712, + 0.0028467602096498013, + 0.002206737408414483, + 0.008085539564490318, + 0.0020728546660393476, + 0.002408307045698166, + 0.0029514674097299576, + 0.006092050112783909, + 0.0019408641383051872, + 0.0011959077091887593, + 0.0011170549551025033, + 0.005510625895112753, + 0.0015570172108709812, + 0.0012443102896213531, + 0.004252218175679445, + 0.010621394962072372, + 0.006232401356101036, + 0.003191010793671012, + 0.003034282010048628, + 0.01070466823875904, + 0.0043344078585505486, + 0.0028802077285945415, + 0.009054811671376228, + 0.008062981069087982, + 0.005814694799482822, + 0.003967072814702988, + 0.0023529583122581244, + 0.006423584185540676, + 0.0021967412903904915, + 0.0032442559022456408, + 0.01871957816183567, + 0.054943084716796875, + 0.03486087918281555, + 0.010972526855766773, + 0.008168250322341919, + 0.08517926186323166, + 0.012713133357465267, + 0.010874824598431587, + 0.03589004650712013, + 0.041513144969940186, + 0.052788231521844864, + 0.016755113378167152, + 0.006760521326214075, + 0.06746716797351837, + 0.006373084615916014, + 0.011528784409165382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.009772907942533493 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a slightly upward angle from a fixed position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone figure wearing a loose, pure white robe standing atop a jagged, dark gray rock formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white robe's fabric visibly whipped violently by the wind, with sleeves and hem fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless expanse of mist-shrouded mountains under a deep, inky blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit low color saturation, sharp focus, and a slightly grainy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10029026865959167 + ], + "opposite_logprob_tokens": [ + -2.350290298461914 + ], + "logprob_reward": -0.10029026865959167, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10029026865959167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07693395763635635, + 0.05356639623641968, + 0.037049539387226105, + 0.031018352136015892, + 0.03894978016614914, + 0.03217588737607002, + 0.0799039676785469, + 0.09834811091423035, + 0.38779860734939575, + 0.16425539553165436 + ], + "credit_map": [ + 0.008081169798970222, + 0.011515561491250992, + 0.013600927777588367, + 0.010025019757449627, + 0.006293849553912878, + 0.00987861305475235, + 0.009404486045241356, + 0.008134323172271252, + 0.005945259239524603, + 0.00747717497870326, + 0.008068551309406757, + 0.007205496542155743, + 0.0071257660165429115, + 0.0065673599019646645, + 0.005419782362878323, + 0.005757002159953117, + 0.005990916863083839, + 0.0093811871483922, + 0.003572538960725069, + 0.002207410056143999, + 0.0040383050218224525, + 0.0074262432754039764, + 0.0022007785737514496, + 0.0022321545984596014, + 0.00514985341578722, + 0.006916165351867676, + 0.003852905472740531, + 0.0021962120663374662, + 0.0026508832816034555, + 0.00508101424202323, + 0.00246191187761724, + 0.0027094073593616486, + 0.0040173400193452835, + 0.00929155945777893, + 0.0035975235514342785, + 0.0044991024769842625, + 0.0027731028385460377, + 0.008078421466052532, + 0.003935968969017267, + 0.002756764180958271, + 0.002985285362228751, + 0.004171691834926605, + 0.004683738574385643, + 0.003456458682194352, + 0.0026853394228965044, + 0.006125225685536861, + 0.004917277488857508, + 0.0031508721876889467, + 0.012900114059448242, + 0.020818188786506653, + 0.007342081982642412, + 0.005580264609307051, + 0.004475174937397242, + 0.015349281020462513, + 0.008718530647456646, + 0.004720330238342285, + 0.017822809517383575, + 0.023025769740343094, + 0.011769965291023254, + 0.009396500885486603, + 0.004542183130979538, + 0.016544919461011887, + 0.009248284623026848, + 0.0059976717457175255, + 0.02222009003162384, + 0.08751317858695984, + 0.052492208778858185, + 0.021053466945886612, + 0.021164800971746445, + 0.13402259349822998, + 0.028178149834275246, + 0.021154116839170456, + 0.029498036950826645, + 0.02579907514154911, + 0.030695660039782524, + 0.019904300570487976, + 0.009469643235206604, + 0.025242595002055168, + 0.010464543476700783, + 0.013181540183722973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.10029026865959167 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_9/index_tensor([4115, 4115])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_9/index_tensor([4115, 4115])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..618152fa0a55d5e6398de9db174e52a582f2271a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_9/index_tensor([4115, 4115])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "steady, tight close-up locks the viewer's gaze on a shattered mirror, riddled with spiderweb-like cracks. Reflecting it is the fragmented image of a young woman. Her hair is neatly cut in a short orange bob, a few strands wet from the icy rain clinging to her smooth temples. Her emerald eyes radiate an unmistakable resolve. In the background, magenta and cyan neon lights cast hazy smears on the damp brick walls, while icy raindrops drift diagonally through the light, creating a decadent and psychedelic cyberpunk atmosphere. The girl's expression is remarkably focused. With an astonishingly steady hand, she grips an electronic eyeliner that emits a cold, blue light. Carefully, she traces the tip of the pencil along the contour of her upper lid, smoothly and decisively swiping it across, instantly creating a precise, sharp, cyborg-esque line. The cool and warm tones of neon are intertwined with the technological blue light of eyeliner, carving out distinct contours and deep shadows on her high nose bridge and tightly pursed lips. The whole picture is full of the ritual sense and resolute beauty of self-armor in the apocalyptic environment.", + "seed": 7684, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_9/index_4115_seed_7684_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01874165069311857, + "vlm_logprob": -0.01874165069311857, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.01874165069311857, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, tight close-up focused on a shattered mirror with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young woman with a short orange bob haircut and emerald eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding an electronic eyeliner that emits a cold blue light and applying it to her upper eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature magenta and cyan neon lights casting hazy smears on damp brick walls with diagonal icy raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's expression appear remarkably focused and resolute while applying the eyeliner?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.058170221745967865 + ], + "opposite_logprob_tokens": [ + -2.9331703186035156 + ], + "logprob_reward": -0.058170221745967865, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.058170221745967865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.145076721906662, + 0.1870303452014923, + 0.10121609270572662, + 0.07057086378335953, + 0.04188324138522148, + 0.02648480236530304, + 0.03315752372145653, + 0.15201498568058014, + 0.16699084639549255, + 0.07557457685470581 + ], + "credit_map": [ + 0.02368461713194847, + 0.017144162207841873, + 0.012651055119931698, + 0.010552880354225636, + 0.021826617419719696, + 0.028553782030940056, + 0.014173315837979317, + 0.01649029366672039, + 0.02642822265625, + 0.012103220447897911, + 0.04261976107954979, + 0.011862765066325665, + 0.015428869985044003, + 0.02572440542280674, + 0.020622415468096733, + 0.032240696251392365, + 0.011261934414505959, + 0.006214797031134367, + 0.021007560193538666, + 0.007868616841733456, + 0.0076210275292396545, + 0.02042997069656849, + 0.007546828128397465, + 0.01926535926759243, + 0.01462507899850607, + 0.009278037585318089, + 0.009490751661360264, + 0.00611757580190897, + 0.00811566412448883, + 0.006536670494824648, + 0.0066069685854017735, + 0.009800105355679989, + 0.005139325745403767, + 0.003385632997378707, + 0.006598429288715124, + 0.007899363525211811, + 0.00440560607239604, + 0.0046816118992865086, + 0.006474141497164965, + 0.003299129195511341, + 0.0031687384471297264, + 0.003423125483095646, + 0.0038134458009153605, + 0.002772774314507842, + 0.003739666659384966, + 0.0031214954797178507, + 0.0031663968693464994, + 0.003279158379882574, + 0.0038267013151198626, + 0.0035745007917284966, + 0.0036693804431706667, + 0.0032739611342549324, + 0.00564584182575345, + 0.004121856298297644, + 0.004006641451269388, + 0.005038643721491098, + 0.010603882372379303, + 0.00785225722938776, + 0.044833067804574966, + 0.011618150398135185, + 0.011514081619679928, + 0.01948343776166439, + 0.010763583704829216, + 0.03534654155373573, + 0.012392246164381504, + 0.004189892206341028, + 0.06818035989999771, + 0.013375538401305676, + 0.006284285802394152, + 0.017859002575278282, + 0.0068805706687271595, + 0.03782893344759941, + 0.0030654335860162973, + 0.0031522540375590324, + 0.023654207587242126, + 0.006924809888005257, + 0.003105208743363619, + 0.026120513677597046, + 0.0031852112151682377, + 0.006366933695971966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 7 + ], + "selected_reward": -0.058170221745967865 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, tight close-up focused on a shattered mirror with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young woman with a short orange bob haircut and emerald eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding an electronic eyeliner that emits a cold blue light and applying it to her upper eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature magenta and cyan neon lights casting hazy smears on damp brick walls with diagonal icy raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's expression appear remarkably focused and resolute while applying the eyeliner?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014205172657966614 + ], + "opposite_logprob_tokens": [ + -4.264204978942871 + ], + "logprob_reward": -0.014205172657966614, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014205172657966614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0976938009262085, + 0.06965924799442291, + 0.058819595724344254, + 0.18698449432849884, + 0.08767789602279663, + 0.054965995252132416, + 0.07755312323570251, + 0.17161786556243896, + 0.10192763805389404, + 0.09310032427310944 + ], + "credit_map": [ + 0.011359427124261856, + 0.016394125297665596, + 0.01009381003677845, + 0.012210114859044552, + 0.013486760668456554, + 0.01201622188091278, + 0.012549353763461113, + 0.009583992883563042, + 0.007620353251695633, + 0.00425513694062829, + 0.01767774298787117, + 0.005169210024178028, + 0.005399268586188555, + 0.013172725215554237, + 0.003402328584343195, + 0.012962491251528263, + 0.0045757051557302475, + 0.00651064096018672, + 0.009609599597752094, + 0.006869322154670954, + 0.008271263912320137, + 0.009839297272264957, + 0.004916381556540728, + 0.008227393962442875, + 0.0231570266187191, + 0.015986507758498192, + 0.02963867038488388, + 0.014710931107401848, + 0.024766946211457253, + 0.02112635225057602, + 0.03214671462774277, + 0.025451354682445526, + 0.008857444860041142, + 0.007392526138573885, + 0.011675148271024227, + 0.011181890964508057, + 0.012311802245676517, + 0.009517863392829895, + 0.01575743779540062, + 0.010983787477016449, + 0.004884404595941305, + 0.003918373491615057, + 0.005129203200340271, + 0.005774254444986582, + 0.007530883885920048, + 0.009808776900172234, + 0.007986556738615036, + 0.009933540597558022, + 0.006974643561989069, + 0.008722854778170586, + 0.008754913695156574, + 0.007637773174792528, + 0.012911487370729446, + 0.009522249922156334, + 0.014626095071434975, + 0.00840311124920845, + 0.014276707544922829, + 0.008826504461467266, + 0.04733448103070259, + 0.01144678145647049, + 0.01654699817299843, + 0.01971093751490116, + 0.016095401719212532, + 0.03738005831837654, + 0.008463617414236069, + 0.0036925827153027058, + 0.038867413997650146, + 0.007023702375590801, + 0.006019866093993187, + 0.010571940802037716, + 0.005675606429576874, + 0.021612918004393578, + 0.005521957762539387, + 0.003350525861606002, + 0.023879466578364372, + 0.007752920500934124, + 0.006262155249714851, + 0.03138580545783043, + 0.004007310606539249, + 0.01094018667936325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 8 + ], + "selected_reward": -0.014205172657966614 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, tight close-up focused on a shattered mirror with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young woman with a short orange bob haircut and emerald eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding an electronic eyeliner that emits a cold blue light and applying it to her upper eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature magenta and cyan neon lights casting hazy smears on damp brick walls with diagonal icy raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's expression appear remarkably focused and resolute while applying the eyeliner?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008645722642540932 + ], + "opposite_logprob_tokens": [ + -4.758645534515381 + ], + "logprob_reward": -0.008645722642540932, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008645722642540932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12338076531887054, + 0.05622149258852005, + 0.0981106087565422, + 0.054064005613327026, + 0.0604826882481575, + 0.06145979464054108, + 0.11595308035612106, + 0.13073159754276276, + 0.22199653089046478, + 0.0775994136929512 + ], + "credit_map": [ + 0.011413736268877983, + 0.016334963962435722, + 0.012339744716882706, + 0.012156300246715546, + 0.019031455740332603, + 0.0239199660718441, + 0.014182071201503277, + 0.014002524316310883, + 0.004634947050362825, + 0.004955499432981014, + 0.010836360976099968, + 0.006887477356940508, + 0.007592468056827784, + 0.007943850010633469, + 0.005011898931115866, + 0.008358988910913467, + 0.011905655264854431, + 0.0069611030630767345, + 0.012974354438483715, + 0.009790485724806786, + 0.009783283807337284, + 0.018448952585458755, + 0.00584390526637435, + 0.022402852773666382, + 0.01200060360133648, + 0.006951768882572651, + 0.010442589409649372, + 0.003533349372446537, + 0.0044348230585455894, + 0.005998179782181978, + 0.004497681278735399, + 0.006205009296536446, + 0.0062683080323040485, + 0.00889572687447071, + 0.00955407414585352, + 0.00828147679567337, + 0.006885158829391003, + 0.007438118103891611, + 0.007745180279016495, + 0.005414640065282583, + 0.007678857073187828, + 0.01110724825412035, + 0.007001435849815607, + 0.006654935423284769, + 0.007477600127458572, + 0.0057709128595888615, + 0.00991519819945097, + 0.005853604059666395, + 0.012306512333452702, + 0.01664493791759014, + 0.012391778640449047, + 0.013428944163024426, + 0.013301342725753784, + 0.011546280235052109, + 0.020687201991677284, + 0.01564606837928295, + 0.008685626089572906, + 0.007618635892868042, + 0.035721972584724426, + 0.009257311932742596, + 0.013805101625621319, + 0.016161933541297913, + 0.009403225965797901, + 0.03007778339087963, + 0.01613166183233261, + 0.006048033945262432, + 0.09042796492576599, + 0.015194270759820938, + 0.006170704960823059, + 0.025043414905667305, + 0.009731730446219444, + 0.05324874073266983, + 0.0036077862605452538, + 0.003181762993335724, + 0.024900805205106735, + 0.007429865188896656, + 0.004403980448842049, + 0.022984379902482033, + 0.004556702915579081, + 0.0065341307781636715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.008645722642540932 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, tight close-up focused on a shattered mirror with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young woman with a short orange bob haircut and emerald eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding an electronic eyeliner that emits a cold blue light and applying it to her upper eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature magenta and cyan neon lights casting hazy smears on damp brick walls with diagonal icy raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's expression appear remarkably focused and resolute while applying the eyeliner?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006736899726092815 + ], + "opposite_logprob_tokens": [ + -5.006736755371094 + ], + "logprob_reward": -0.006736899726092815, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006736899726092815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10682753473520279, + 0.09094485640525818, + 0.14916540682315826, + 0.06816690415143967, + 0.05316857621073723, + 0.029711302369832993, + 0.04812086373567581, + 0.1135203018784523, + 0.313666433095932, + 0.026707785204052925 + ], + "credit_map": [ + 0.00791043322533369, + 0.019771870225667953, + 0.011711963452398777, + 0.012095851823687553, + 0.01892838068306446, + 0.013149896636605263, + 0.013055280782282352, + 0.010203867219388485, + 0.007965631783008575, + 0.006635167170315981, + 0.02167322486639023, + 0.007302282378077507, + 0.008432034403085709, + 0.015425916761159897, + 0.007455973420292139, + 0.016054634004831314, + 0.014218445867300034, + 0.0086141899228096, + 0.03879866376519203, + 0.011644498445093632, + 0.011010640300810337, + 0.027305344119668007, + 0.00987215619534254, + 0.02770148403942585, + 0.008979607373476028, + 0.007414267398416996, + 0.011754142120480537, + 0.005121106281876564, + 0.009802418760955334, + 0.0086392005905509, + 0.009537163190543652, + 0.006919001694768667, + 0.008736963383853436, + 0.004223558120429516, + 0.007372868247330189, + 0.009645868092775345, + 0.003942144103348255, + 0.006696287076920271, + 0.004541217349469662, + 0.008009673096239567, + 0.004528519697487354, + 0.0030300926882773638, + 0.004880719818174839, + 0.0041797105222940445, + 0.0030212237033993006, + 0.0029795081354677677, + 0.00409690709784627, + 0.0029946232680231333, + 0.004306919872760773, + 0.006065449211746454, + 0.00623422721400857, + 0.004840326961129904, + 0.007668857462704182, + 0.0044451202265918255, + 0.008192643523216248, + 0.006367319729179144, + 0.007694579195231199, + 0.005974845960736275, + 0.033339668065309525, + 0.008236872032284737, + 0.01045319065451622, + 0.013852899894118309, + 0.008494576439261436, + 0.02547367475926876, + 0.02276116982102394, + 0.00695370277389884, + 0.13417527079582214, + 0.02579808048903942, + 0.007897241972386837, + 0.03524895757436752, + 0.011396588757634163, + 0.06943543255329132, + 0.0014063628623262048, + 0.002294749952852726, + 0.006967547815293074, + 0.003362280083820224, + 0.002388837281614542, + 0.006401794496923685, + 0.0018180216429755092, + 0.00206819293089211 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 7 + ], + "selected_reward": -0.006736899726092815 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, tight close-up focused on a shattered mirror with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young woman with a short orange bob haircut and emerald eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding an electronic eyeliner that emits a cold blue light and applying it to her upper eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature magenta and cyan neon lights casting hazy smears on damp brick walls with diagonal icy raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's expression appear remarkably focused and resolute while applying the eyeliner?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950236693024635 + ], + "opposite_logprob_tokens": [ + -5.130950450897217 + ], + "logprob_reward": -0.005950236693024635, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950236693024635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06278899312019348, + 0.05528508871793747, + 0.04036261513829231, + 0.10235244035720825, + 0.07444985955953598, + 0.07919494062662125, + 0.11007054895162582, + 0.0979406088590622, + 0.3382008671760559, + 0.03935406357049942 + ], + "credit_map": [ + 0.006005231756716967, + 0.006623232737183571, + 0.0060860030353069305, + 0.00682467594742775, + 0.015124642290174961, + 0.006936878897249699, + 0.00936515349894762, + 0.005823176354169846, + 0.009373310953378677, + 0.004997730255126953, + 0.01125982403755188, + 0.004950111731886864, + 0.004481807351112366, + 0.006800471339374781, + 0.005470139905810356, + 0.007951691746711731, + 0.003327935002744198, + 0.00437368918210268, + 0.007243753410875797, + 0.004725780338048935, + 0.0040191165171563625, + 0.005800460930913687, + 0.004429997410625219, + 0.006441881414502859, + 0.02057638019323349, + 0.01243353821337223, + 0.01557291578501463, + 0.007089758291840553, + 0.00934815313667059, + 0.012191906571388245, + 0.006705186795443296, + 0.01843460276722908, + 0.015949998050928116, + 0.010236548259854317, + 0.009136594831943512, + 0.010317281819880009, + 0.0068345870822668076, + 0.007578278426080942, + 0.007273796014487743, + 0.007122777868062258, + 0.009670862928032875, + 0.010652662254869938, + 0.011329399421811104, + 0.01279411930590868, + 0.007333486340939999, + 0.008916175924241543, + 0.010039176791906357, + 0.008459053933620453, + 0.01313429232686758, + 0.022298837080597878, + 0.012461617588996887, + 0.012685438618063927, + 0.01281864196062088, + 0.009984366595745087, + 0.013485605828464031, + 0.013201738707721233, + 0.006079091690480709, + 0.005759795196354389, + 0.027835000306367874, + 0.007584456354379654, + 0.010620974004268646, + 0.011252171359956264, + 0.008757255040109158, + 0.0200518686324358, + 0.023102179169654846, + 0.008904910646378994, + 0.14028681814670563, + 0.023723654448986053, + 0.011138399131596088, + 0.03989435359835625, + 0.014399191364645958, + 0.07675134390592575, + 0.0024392269551753998, + 0.0018229738343507051, + 0.011557463556528091, + 0.0042659868486225605, + 0.0028370229993015528, + 0.01020224392414093, + 0.00214579701423645, + 0.004083352163434029 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 3 + ], + "selected_reward": -0.005950236693024635 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, tight close-up locks the viewer's gaze on a shattered mirror, riddled with spiderweb-like cracks. Reflecting it is the fragmented image of a young woman. Her hair is neatly cut in a short orange bob, a few strands wet from the icy rain clinging to her smooth temples. Her emerald eyes radiate an unmistakable resolve. In the background, magenta and cyan neon lights cast hazy smears on the damp brick walls, while icy raindrops drift diagonally through the light, creating a decadent and psychedelic cyberpunk atmosphere. The girl's expression is remarkably focused. With an astonishingly steady hand, she grips an electronic eyeliner that emits a cold, blue light. Carefully, she traces the tip of the pencil along the contour of her upper lid, smoothly and decisively swiping it across, instantly creating a precise, sharp, cyborg-esque line. The cool and warm tones of neon are intertwined with the technological blue light of eyeliner, carving out distinct contours and deep shadows on her high nose bridge and tightly pursed lips. The whole picture is full of the ritual sense and resolute beauty of self-armor in the apocalyptic environment.", + "seed": 7685, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_9/index_4115_seed_7685_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07654392365366221, + "vlm_logprob": -0.07654392365366221, + "vlm_yes_no_margin": 4.074999988079071, + "vlm_token_credit": -0.07654392365366221, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, tight close-up focused on a shattered mirror with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young woman with a short orange bob haircut and emerald eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding an electronic eyeliner that emits a cold blue light and applying it to her upper eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature magenta and cyan neon lights casting hazy smears on damp brick walls with diagonal icy raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's expression appear remarkably focused and resolute while applying the eyeliner?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3506503701210022 + ], + "opposite_logprob_tokens": [ + -1.2256503105163574 + ], + "logprob_reward": -0.3506503701210022, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.3506503701210022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41470834612846375, + 0.06470406800508499, + 0.12165511399507523, + 0.06744429469108582, + 0.03900764510035515, + 0.06105241924524307, + 0.04995414987206459, + 0.08097483962774277, + 0.05139854922890663, + 0.04910055547952652 + ], + "credit_map": [ + 0.050856977701187134, + 0.06525083631277084, + 0.05320078879594803, + 0.04656665772199631, + 0.04703507199883461, + 0.045905258506536484, + 0.05042627826333046, + 0.05546645075082779, + 0.007397472858428955, + 0.007295620162039995, + 0.007793760858476162, + 0.007435584906488657, + 0.007967418059706688, + 0.006549814250320196, + 0.009525991976261139, + 0.010738404467701912, + 0.013117806054651737, + 0.014983724802732468, + 0.017311053350567818, + 0.016095338389277458, + 0.0264456607401371, + 0.00990317389369011, + 0.013893640600144863, + 0.009904720820486546, + 0.006316389888525009, + 0.015135399997234344, + 0.006861030589789152, + 0.005709294695407152, + 0.01147748064249754, + 0.0077565694227814674, + 0.007632480002939701, + 0.006555648986250162, + 0.004903615452349186, + 0.009485972113907337, + 0.0038256726693361998, + 0.004315669182687998, + 0.005298256874084473, + 0.0037933094426989555, + 0.003728133626282215, + 0.003657015273347497, + 0.007148891221731901, + 0.010549605824053288, + 0.009154058992862701, + 0.008709884248673916, + 0.006053504534065723, + 0.004377898760139942, + 0.007048527244478464, + 0.008010047487914562, + 0.005450673401355743, + 0.006837394088506699, + 0.005769083742052317, + 0.0051441490650177, + 0.007131785620003939, + 0.006003228481858969, + 0.005285062827169895, + 0.00833277776837349, + 0.011448279023170471, + 0.023008013144135475, + 0.00961734913289547, + 0.008656895719468594, + 0.010768459178507328, + 0.005589118227362633, + 0.006817744579166174, + 0.005068985745310783, + 0.008867170661687851, + 0.009208888746798038, + 0.006564992014318705, + 0.005607472267001867, + 0.007139916997402906, + 0.004261378198862076, + 0.005254795309156179, + 0.004493937827646732, + 0.00727707426995039, + 0.010775227099657059, + 0.006401080172508955, + 0.007226100191473961, + 0.005565842613577843, + 0.004020306281745434, + 0.00473279831930995, + 0.0031021242029964924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.3506503701210022 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, tight close-up focused on a shattered mirror with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young woman with a short orange bob haircut and emerald eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding an electronic eyeliner that emits a cold blue light and applying it to her upper eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature magenta and cyan neon lights casting hazy smears on damp brick walls with diagonal icy raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's expression appear remarkably focused and resolute while applying the eyeliner?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011080047115683556 + ], + "opposite_logprob_tokens": [ + -4.511080265045166 + ], + "logprob_reward": -0.011080047115683556, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080047115683556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11294782906770706, + 0.17515231668949127, + 0.20479436218738556, + 0.1720871776342392, + 0.03631126508116722, + 0.020140191540122032, + 0.0413268581032753, + 0.07758742570877075, + 0.06398997455835342, + 0.09566257894039154 + ], + "credit_map": [ + 0.012700216844677925, + 0.020630331709980965, + 0.010911430232226849, + 0.010422141291201115, + 0.012084881775081158, + 0.01358311902731657, + 0.018112553283572197, + 0.01450316235423088, + 0.020522484555840492, + 0.023333365097641945, + 0.01406017504632473, + 0.015175552107393742, + 0.03177955746650696, + 0.01793326437473297, + 0.03896962106227875, + 0.01337829977273941, + 0.02438013069331646, + 0.018263276666402817, + 0.025464948266744614, + 0.02341870777308941, + 0.032327476888895035, + 0.022482994943857193, + 0.030597209930419922, + 0.027859613299369812, + 0.0170986857265234, + 0.024144552648067474, + 0.01701376400887966, + 0.012313243933022022, + 0.032674554735422134, + 0.023171575739979744, + 0.027958907186985016, + 0.017711905762553215, + 0.0034142162185162306, + 0.00442462507635355, + 0.0022454424761235714, + 0.0034350533969700336, + 0.005823896266520023, + 0.005614622496068478, + 0.006517089903354645, + 0.004836317151784897, + 0.0022931783460080624, + 0.0032541125547140837, + 0.0025829256046563387, + 0.0019042668864130974, + 0.003061834257096052, + 0.0022941159550100565, + 0.002475994871929288, + 0.002273763995617628, + 0.004556745290756226, + 0.006990096066147089, + 0.005561097525060177, + 0.00471121771261096, + 0.006402510683983564, + 0.003249702975153923, + 0.004114642273634672, + 0.005740843713283539, + 0.011023215018212795, + 0.02189408428966999, + 0.00873363297432661, + 0.009995044209063053, + 0.009371880441904068, + 0.004773673601448536, + 0.006612864788621664, + 0.005183028988540173, + 0.008526533842086792, + 0.016262825578451157, + 0.006652058567851782, + 0.008610490709543228, + 0.008265829645097256, + 0.004627913702279329, + 0.005838068202137947, + 0.005206257104873657, + 0.012669860385358334, + 0.02293909713625908, + 0.010893830098211765, + 0.010557466186583042, + 0.015569241717457771, + 0.006018375512212515, + 0.01109094638377428, + 0.005923764780163765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.011080047115683556 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, tight close-up focused on a shattered mirror with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young woman with a short orange bob haircut and emerald eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding an electronic eyeliner that emits a cold blue light and applying it to her upper eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature magenta and cyan neon lights casting hazy smears on damp brick walls with diagonal icy raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's expression appear remarkably focused and resolute while applying the eyeliner?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009789198637008667 + ], + "opposite_logprob_tokens": [ + -4.634788990020752 + ], + "logprob_reward": -0.009789198637008667, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009789198637008667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1523321270942688, + 0.1189262643456459, + 0.22412288188934326, + 0.0737696960568428, + 0.09722990542650223, + 0.0502544641494751, + 0.08461567014455795, + 0.06691757589578629, + 0.06878548860549927, + 0.06304578483104706 + ], + "credit_map": [ + 0.01714274100959301, + 0.02679341658949852, + 0.013959131203591824, + 0.012415850535035133, + 0.01617249846458435, + 0.01733492501080036, + 0.027667583897709846, + 0.020845996215939522, + 0.012175322510302067, + 0.018286744132637978, + 0.013141385279595852, + 0.010096438229084015, + 0.015064827166497707, + 0.012948116287589073, + 0.013832339085638523, + 0.023381110280752182, + 0.025120338425040245, + 0.032786931842565536, + 0.03456695005297661, + 0.03040253557264805, + 0.02498014271259308, + 0.019083475694060326, + 0.029714588075876236, + 0.027467934414744377, + 0.008067321963608265, + 0.010527084581553936, + 0.006828714162111282, + 0.010628118179738522, + 0.01081151608377695, + 0.007983201183378696, + 0.01076418999582529, + 0.008159555494785309, + 0.01247954647988081, + 0.03099863789975643, + 0.006562764290720224, + 0.007847854867577553, + 0.014595993794500828, + 0.008930075913667679, + 0.007648013532161713, + 0.008167022839188576, + 0.006030577700585127, + 0.010345353744924068, + 0.007187318056821823, + 0.0062118638306856155, + 0.0051251668483018875, + 0.0040140459313988686, + 0.0062800622545182705, + 0.005060082767158747, + 0.009137499146163464, + 0.014149581082165241, + 0.011676646769046783, + 0.008415386080741882, + 0.021260729059576988, + 0.0051063233986496925, + 0.009062468074262142, + 0.005807050038129091, + 0.0075919134542346, + 0.015433545224368572, + 0.007215271703898907, + 0.00688438443467021, + 0.011783614754676819, + 0.003938948269933462, + 0.008008044213056564, + 0.006061868742108345, + 0.019367139786481857, + 0.007599321193993092, + 0.010184524580836296, + 0.005287660751491785, + 0.008467219769954681, + 0.0036363874096423388, + 0.006839883513748646, + 0.007403363473713398, + 0.009396865032613277, + 0.016302119940519333, + 0.008183654397726059, + 0.006747001316398382, + 0.007967483252286911, + 0.004618728067725897, + 0.005751487798988819, + 0.004078450612723827 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.009789198637008667 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, tight close-up focused on a shattered mirror with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young woman with a short orange bob haircut and emerald eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding an electronic eyeliner that emits a cold blue light and applying it to her upper eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature magenta and cyan neon lights casting hazy smears on damp brick walls with diagonal icy raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's expression appear remarkably focused and resolute while applying the eyeliner?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005946444813162088 + ], + "opposite_logprob_tokens": [ + -5.130946636199951 + ], + "logprob_reward": -0.005946444813162088, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005946444813162088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22073136270046234, + 0.03565049543976784, + 0.06824707984924316, + 0.04200423136353493, + 0.06757969409227371, + 0.13777752220630646, + 0.17596964538097382, + 0.14185389876365662, + 0.050263140350580215, + 0.059922955930233 + ], + "credit_map": [ + 0.02069016732275486, + 0.05129813402891159, + 0.02016812562942505, + 0.025342129170894623, + 0.01815963350236416, + 0.02204998768866062, + 0.030200431123375893, + 0.03282275050878525, + 0.005020354874432087, + 0.003915734123438597, + 0.003602412762120366, + 0.0036025873851031065, + 0.004829960409551859, + 0.003444867441430688, + 0.005495209246873856, + 0.005739371292293072, + 0.00748963700607419, + 0.00824250839650631, + 0.010775183327496052, + 0.006715249270200729, + 0.01607624813914299, + 0.004831366706639528, + 0.008601442910730839, + 0.00551544688642025, + 0.005377167835831642, + 0.0061905463226139545, + 0.004949553869664669, + 0.004433236550539732, + 0.006968884263187647, + 0.004625584464520216, + 0.005745592527091503, + 0.0037136643659323454, + 0.006974637974053621, + 0.0156292412430048, + 0.0061874836683273315, + 0.006139579229056835, + 0.0135303670540452, + 0.005709197837859392, + 0.007883268408477306, + 0.005525919143110514, + 0.018734922632575035, + 0.04175528138875961, + 0.010833135806024075, + 0.012511749751865864, + 0.01642313227057457, + 0.009841405786573887, + 0.012677936814725399, + 0.01499995868653059, + 0.02011650986969471, + 0.027876751497387886, + 0.0208758395165205, + 0.0205678790807724, + 0.03524985909461975, + 0.014425856061279774, + 0.01941888965666294, + 0.017438065260648727, + 0.01994827389717102, + 0.0314626544713974, + 0.015112252905964851, + 0.01172614935785532, + 0.020261619240045547, + 0.010357167571783066, + 0.015896061435341835, + 0.017089728266000748, + 0.010092943906784058, + 0.007776828017085791, + 0.006768689025193453, + 0.004547686781734228, + 0.007185073569417, + 0.004091949667781591, + 0.005164796020835638, + 0.004635173827409744, + 0.008083627559244633, + 0.015166481956839561, + 0.008708120323717594, + 0.007170805707573891, + 0.00588269205763936, + 0.004871064331382513, + 0.005618094000965357, + 0.004422068130224943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.005946444813162088 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, tight close-up focused on a shattered mirror with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young woman with a short orange bob haircut and emerald eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding an electronic eyeliner that emits a cold blue light and applying it to her upper eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature magenta and cyan neon lights casting hazy smears on damp brick walls with diagonal icy raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's expression appear remarkably focused and resolute while applying the eyeliner?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052535575814545155 + ], + "opposite_logprob_tokens": [ + -5.255253791809082 + ], + "logprob_reward": -0.0052535575814545155, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052535575814545155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26432904601097107, + 0.13170506060123444, + 0.11760266870260239, + 0.1126427948474884, + 0.04674671217799187, + 0.07654614746570587, + 0.056145258247852325, + 0.07503986358642578, + 0.04883260652422905, + 0.07040981203317642 + ], + "credit_map": [ + 0.02511369064450264, + 0.056471407413482666, + 0.02499355748295784, + 0.02749711088836193, + 0.02336208149790764, + 0.02558552660048008, + 0.056285206228494644, + 0.025020476430654526, + 0.012932516634464264, + 0.014490698464214802, + 0.013992197811603546, + 0.00907944981008768, + 0.014899486675858498, + 0.016332779079675674, + 0.022324474528431892, + 0.02765345387160778, + 0.011631270870566368, + 0.021398523822426796, + 0.011599558405578136, + 0.011376531794667244, + 0.02695063315331936, + 0.008691366761922836, + 0.013144166208803654, + 0.012810622341930866, + 0.015805073082447052, + 0.024234121665358543, + 0.010513494722545147, + 0.014197620563209057, + 0.017256105318665504, + 0.007029698230326176, + 0.012295832857489586, + 0.011310851201415062, + 0.005054858047515154, + 0.009795651771128178, + 0.0033207328524440527, + 0.006458228453993797, + 0.009831861592829227, + 0.0031816912814974785, + 0.005818413104861975, + 0.003285276936367154, + 0.008726522326469421, + 0.015831617638468742, + 0.011738890781998634, + 0.009949745610356331, + 0.007951317355036736, + 0.00484524667263031, + 0.00907007697969675, + 0.008432728238403797, + 0.0066572208888828754, + 0.010796342976391315, + 0.005128934513777494, + 0.005297407507896423, + 0.00786754209548235, + 0.008304019458591938, + 0.006415020674467087, + 0.005678768735378981, + 0.010444204322993755, + 0.01416732557117939, + 0.007799552753567696, + 0.006750729400664568, + 0.012532668188214302, + 0.006270552519708872, + 0.008543366566300392, + 0.00853146892040968, + 0.007545502856373787, + 0.00822419486939907, + 0.006141037680208683, + 0.004240621812641621, + 0.00656139338389039, + 0.004491189960390329, + 0.00561884930357337, + 0.006009814795106649, + 0.007224537432193756, + 0.015068710781633854, + 0.008410773240029812, + 0.007705926429480314, + 0.008094793185591698, + 0.006356475409120321, + 0.008472194895148277, + 0.009076403453946114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0052535575814545155 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, tight close-up locks the viewer's gaze on a shattered mirror, riddled with spiderweb-like cracks. Reflecting it is the fragmented image of a young woman. Her hair is neatly cut in a short orange bob, a few strands wet from the icy rain clinging to her smooth temples. Her emerald eyes radiate an unmistakable resolve. In the background, magenta and cyan neon lights cast hazy smears on the damp brick walls, while icy raindrops drift diagonally through the light, creating a decadent and psychedelic cyberpunk atmosphere. The girl's expression is remarkably focused. With an astonishingly steady hand, she grips an electronic eyeliner that emits a cold, blue light. Carefully, she traces the tip of the pencil along the contour of her upper lid, smoothly and decisively swiping it across, instantly creating a precise, sharp, cyborg-esque line. The cool and warm tones of neon are intertwined with the technological blue light of eyeliner, carving out distinct contours and deep shadows on her high nose bridge and tightly pursed lips. The whole picture is full of the ritual sense and resolute beauty of self-armor in the apocalyptic environment.", + "seed": 7686, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_9/index_4115_seed_7686_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03683669874444604, + "vlm_logprob": -0.03683669874444604, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.03683669874444604, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, tight close-up focused on a shattered mirror with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young woman with a short orange bob haircut and emerald eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding an electronic eyeliner that emits a cold blue light and applying it to her upper eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature magenta and cyan neon lights casting hazy smears on damp brick walls with diagonal icy raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's expression appear remarkably focused and resolute while applying the eyeliner?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14628787338733673 + ], + "opposite_logprob_tokens": [ + -2.0212879180908203 + ], + "logprob_reward": -0.14628787338733673, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14628787338733673, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19220387935638428, + 0.19787785410881042, + 0.11774362623691559, + 0.11368699371814728, + 0.1297931671142578, + 0.07300363481044769, + 0.04825008660554886, + 0.05107983946800232, + 0.05151255428791046, + 0.024848323315382004 + ], + "credit_map": [ + 0.02088523656129837, + 0.01906774938106537, + 0.02419944293797016, + 0.04011154919862747, + 0.014330531470477581, + 0.01870742440223694, + 0.030166149139404297, + 0.024735799059271812, + 0.028133193030953407, + 0.01885916106402874, + 0.031164253130555153, + 0.02225782722234726, + 0.02554047666490078, + 0.025855056941509247, + 0.024496693164110184, + 0.021571189165115356, + 0.018742062151432037, + 0.01023823767900467, + 0.014051886275410652, + 0.01030303630977869, + 0.026092788204550743, + 0.010124857537448406, + 0.019621694460511208, + 0.008569062687456608, + 0.01078511867672205, + 0.007162326946854591, + 0.019848428666591644, + 0.007249284535646439, + 0.0270532239228487, + 0.011370109394192696, + 0.017651023343205452, + 0.012567476369440556, + 0.011023348197340965, + 0.013162833638489246, + 0.01804986782371998, + 0.00926236156374216, + 0.025369662791490555, + 0.014355550520122051, + 0.020445916801691055, + 0.018123624846339226, + 0.006149564404040575, + 0.006700051948428154, + 0.009665156714618206, + 0.007001083344221115, + 0.010685558430850506, + 0.006235595792531967, + 0.015638118609786034, + 0.010928506962954998, + 0.00404059374704957, + 0.004771726671606302, + 0.005689320154488087, + 0.005214747507125139, + 0.006394615862518549, + 0.0066907680593431, + 0.007857526652514935, + 0.007590788882225752, + 0.006017478182911873, + 0.006515879184007645, + 0.007052922621369362, + 0.00624917633831501, + 0.005858247634023428, + 0.007240684702992439, + 0.006399855483323336, + 0.0057455929927527905, + 0.005079588387161493, + 0.0077193803153932095, + 0.005914817564189434, + 0.0064642722718417645, + 0.007911503314971924, + 0.005853201262652874, + 0.007311483845114708, + 0.005258309654891491, + 0.0026836430188268423, + 0.003472867188975215, + 0.002813131082803011, + 0.0034321099519729614, + 0.0030760602094233036, + 0.0034023814368993044, + 0.0025027631781995296, + 0.003465368179604411 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.14628787338733673 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, tight close-up focused on a shattered mirror with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young woman with a short orange bob haircut and emerald eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding an electronic eyeliner that emits a cold blue light and applying it to her upper eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature magenta and cyan neon lights casting hazy smears on damp brick walls with diagonal icy raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's expression appear remarkably focused and resolute while applying the eyeliner?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0110840555280447 + ], + "opposite_logprob_tokens": [ + -4.511084079742432 + ], + "logprob_reward": -0.0110840555280447, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.0110840555280447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08091547340154648, + 0.06950396299362183, + 0.07739469408988953, + 0.09830720722675323, + 0.265916109085083, + 0.14480841159820557, + 0.09401284158229828, + 0.07354643195867538, + 0.025095468387007713, + 0.07049937546253204 + ], + "credit_map": [ + 0.009227645583450794, + 0.009187624789774418, + 0.010408400557935238, + 0.013482439331710339, + 0.007481795269995928, + 0.01228225976228714, + 0.008991376496851444, + 0.009853936731815338, + 0.00779893109574914, + 0.011918500065803528, + 0.008135437034070492, + 0.013158989138901234, + 0.008039489388465881, + 0.009253100492060184, + 0.005920475348830223, + 0.005279039032757282, + 0.01135226245969534, + 0.006572804879397154, + 0.013415418565273285, + 0.008322060108184814, + 0.01715296134352684, + 0.006004177499562502, + 0.00837803352624178, + 0.006196974311023951, + 0.011995639652013779, + 0.005610934458673, + 0.028123067691922188, + 0.006089796777814627, + 0.013812522403895855, + 0.003907667938619852, + 0.024166343733668327, + 0.004601236432790756, + 0.02483549900352955, + 0.026773277670145035, + 0.04468019679188728, + 0.020368175581097603, + 0.05602497234940529, + 0.02131015621125698, + 0.050086311995983124, + 0.021837525069713593, + 0.011399817653000355, + 0.011647027917206287, + 0.02088998444378376, + 0.014210612513124943, + 0.017163293436169624, + 0.012353313155472279, + 0.02663717232644558, + 0.03050718456506729, + 0.008360123261809349, + 0.013007252477109432, + 0.010715493001043797, + 0.010987444780766964, + 0.014450469985604286, + 0.009701934643089771, + 0.017954910174012184, + 0.008835209533572197, + 0.005828583147376776, + 0.008163358084857464, + 0.00958004966378212, + 0.008923529647290707, + 0.01225606631487608, + 0.008922834880650043, + 0.011317521333694458, + 0.008554484695196152, + 0.0028252294287085533, + 0.0026574102230370045, + 0.003147928975522518, + 0.0035221492871642113, + 0.004068528302013874, + 0.0025276117958128452, + 0.003543579252436757, + 0.002803029026836157, + 0.0064768195152282715, + 0.009733039885759354, + 0.008140342310070992, + 0.013495958410203457, + 0.011921525001525879, + 0.00568376062437892, + 0.007150739897042513, + 0.00789718795567751 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.0110840555280447 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, tight close-up focused on a shattered mirror with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young woman with a short orange bob haircut and emerald eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding an electronic eyeliner that emits a cold blue light and applying it to her upper eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature magenta and cyan neon lights casting hazy smears on damp brick walls with diagonal icy raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's expression appear remarkably focused and resolute while applying the eyeliner?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009789670817553997 + ], + "opposite_logprob_tokens": [ + -4.63478946685791 + ], + "logprob_reward": -0.009789670817553997, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009789670817553997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20349319279193878, + 0.1500072479248047, + 0.08935970813035965, + 0.05733095854520798, + 0.06637110561132431, + 0.07028097659349442, + 0.04992033168673515, + 0.1890724152326584, + 0.0808512344956398, + 0.04331282153725624 + ], + "credit_map": [ + 0.01996300183236599, + 0.020746450871229172, + 0.024931669235229492, + 0.04021355137228966, + 0.017385557293891907, + 0.032117508351802826, + 0.02684267982840538, + 0.021292762830853462, + 0.02052609622478485, + 0.02178448624908924, + 0.022033045068383217, + 0.025560684502124786, + 0.018497340381145477, + 0.01640678010880947, + 0.013904072344303131, + 0.011294743977487087, + 0.01160686369985342, + 0.007267394568771124, + 0.016627445816993713, + 0.00622299499809742, + 0.021742701530456543, + 0.006451040972024202, + 0.013191958889365196, + 0.006249306723475456, + 0.006229397375136614, + 0.004311570897698402, + 0.00839246716350317, + 0.0033750797156244516, + 0.013798738829791546, + 0.005275818519294262, + 0.009108182042837143, + 0.006839701905846596, + 0.006689033005386591, + 0.0046973153948783875, + 0.010476202704012394, + 0.006075446959584951, + 0.014849498867988586, + 0.004759407136589289, + 0.010972595773637295, + 0.007851609028875828, + 0.004745394457131624, + 0.005319021642208099, + 0.009689709171652794, + 0.006517385132610798, + 0.009189484640955925, + 0.008350360207259655, + 0.011191180907189846, + 0.015278439968824387, + 0.003982716239988804, + 0.004705904051661491, + 0.00566370552405715, + 0.0070635005831718445, + 0.008973580785095692, + 0.004558490589261055, + 0.00923339556902647, + 0.005739041604101658, + 0.0147098358720541, + 0.01981896534562111, + 0.02511194907128811, + 0.02414541319012642, + 0.03111880272626877, + 0.021749846637248993, + 0.030727140605449677, + 0.02169046178460121, + 0.006598085630685091, + 0.008885565213859081, + 0.009134785272181034, + 0.013408983126282692, + 0.017267296090722084, + 0.00720333494246006, + 0.010202118195593357, + 0.00815106462687254, + 0.0045904675498604774, + 0.004686036612838507, + 0.005040075164288282, + 0.004509774502366781, + 0.008534715510904789, + 0.006031414959579706, + 0.004317395389080048, + 0.005602939985692501 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.009789670817553997 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, tight close-up focused on a shattered mirror with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young woman with a short orange bob haircut and emerald eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding an electronic eyeliner that emits a cold blue light and applying it to her upper eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature magenta and cyan neon lights casting hazy smears on damp brick walls with diagonal icy raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's expression appear remarkably focused and resolute while applying the eyeliner?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075449176132679 + ], + "opposite_logprob_tokens": [ + -4.511075496673584 + ], + "logprob_reward": -0.011075449176132679, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075449176132679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23517413437366486, + 0.14306852221488953, + 0.09231144934892654, + 0.11077816039323807, + 0.046069320291280746, + 0.054666973650455475, + 0.02746000327169895, + 0.12781094014644623, + 0.06462164223194122, + 0.09803883731365204 + ], + "credit_map": [ + 0.027556706219911575, + 0.019836852326989174, + 0.030145874246954918, + 0.032957084476947784, + 0.021796369925141335, + 0.029355840757489204, + 0.03667068853974342, + 0.03685472905635834, + 0.021986251696944237, + 0.014751989394426346, + 0.017162008211016655, + 0.021190116181969643, + 0.018004028126597404, + 0.020843366160988808, + 0.015688596293330193, + 0.013442172668874264, + 0.02054649591445923, + 0.00782701000571251, + 0.014531626366078854, + 0.009664824232459068, + 0.01189957745373249, + 0.007474425248801708, + 0.011817997321486473, + 0.008549492806196213, + 0.00897321105003357, + 0.009415555745363235, + 0.02371513843536377, + 0.006519402842968702, + 0.028297435492277145, + 0.005325342994183302, + 0.0220218226313591, + 0.006510249804705381, + 0.004799770191311836, + 0.006052603479474783, + 0.006597294006496668, + 0.003394216299057007, + 0.008982679806649685, + 0.0038117943331599236, + 0.008073359727859497, + 0.004357601050287485, + 0.0040285056456923485, + 0.0037370226345956326, + 0.007856435142457485, + 0.004832634702324867, + 0.005405631382018328, + 0.007113487459719181, + 0.010240817442536354, + 0.011452438309788704, + 0.0031890771351754665, + 0.0028016052674502134, + 0.0031975104939192533, + 0.0029365341179072857, + 0.0032537756487727165, + 0.0029384158551692963, + 0.00476114684715867, + 0.004381940234452486, + 0.01173979602754116, + 0.014866997487843037, + 0.01613074541091919, + 0.016025550663471222, + 0.021346161141991615, + 0.011144394055008888, + 0.022246118634939194, + 0.014311183243989944, + 0.005718911532312632, + 0.00815964862704277, + 0.008346524089574814, + 0.00584604823961854, + 0.015351325273513794, + 0.006941485684365034, + 0.007747940719127655, + 0.00650976225733757, + 0.00887005589902401, + 0.014785713516175747, + 0.011494455859065056, + 0.022919638082385063, + 0.00920567475259304, + 0.007988501340150833, + 0.011939109303057194, + 0.010835694149136543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.011075449176132679 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, tight close-up focused on a shattered mirror with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young woman with a short orange bob haircut and emerald eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding an electronic eyeliner that emits a cold blue light and applying it to her upper eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature magenta and cyan neon lights casting hazy smears on damp brick walls with diagonal icy raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's expression appear remarkably focused and resolute while applying the eyeliner?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005946444813162088 + ], + "opposite_logprob_tokens": [ + -5.130946636199951 + ], + "logprob_reward": -0.005946444813162088, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005946444813162088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2872830033302307, + 0.04654587432742119, + 0.10610360652208328, + 0.047182001173496246, + 0.13564512133598328, + 0.1489724963903427, + 0.10692045092582703, + 0.05597812309861183, + 0.03892737627029419, + 0.026441950350999832 + ], + "credit_map": [ + 0.030309373512864113, + 0.02826692722737789, + 0.035360101610422134, + 0.04828156530857086, + 0.027577057480812073, + 0.053423285484313965, + 0.03561967983841896, + 0.028444984927773476, + 0.006357770878821611, + 0.00602797232568264, + 0.0069943200796842575, + 0.007132757920771837, + 0.006019554100930691, + 0.006061650346964598, + 0.0040634749457240105, + 0.0038883720990270376, + 0.018060699105262756, + 0.007450185250490904, + 0.0243099182844162, + 0.011233772151172161, + 0.01688959449529648, + 0.007269458845257759, + 0.013972695916891098, + 0.006917273160070181, + 0.006065850146114826, + 0.003532704897224903, + 0.011471431702375412, + 0.004683181177824736, + 0.008471798151731491, + 0.002657457021996379, + 0.007198930252343416, + 0.0031006461940705776, + 0.012166395783424377, + 0.009831723757088184, + 0.021783307194709778, + 0.009588461369276047, + 0.027558179572224617, + 0.01293396670371294, + 0.026651903986930847, + 0.015131178312003613, + 0.009976104833185673, + 0.01190194021910429, + 0.02142983488738537, + 0.014647242613136768, + 0.01792738027870655, + 0.011576962657272816, + 0.03419902175664902, + 0.027314012870192528, + 0.011410082690417767, + 0.011037159711122513, + 0.012446999549865723, + 0.01094071939587593, + 0.01563967764377594, + 0.014289622195065022, + 0.018722431734204292, + 0.012433750554919243, + 0.005851342808455229, + 0.005734908394515514, + 0.00750669464468956, + 0.0073809935711324215, + 0.008004560135304928, + 0.005407118704169989, + 0.009512827731668949, + 0.006579673383384943, + 0.004792199935764074, + 0.003973012324422598, + 0.0050513893365859985, + 0.0037272099871188402, + 0.008742881007492542, + 0.005044757854193449, + 0.004169689025729895, + 0.0034262356348335743, + 0.0028473329730331898, + 0.004080427810549736, + 0.002804558491334319, + 0.005141195375472307, + 0.003611859865486622, + 0.0023005439434200525, + 0.0032459949143230915, + 0.0024100353475660086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.005946444813162088 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, tight close-up locks the viewer's gaze on a shattered mirror, riddled with spiderweb-like cracks. Reflecting it is the fragmented image of a young woman. Her hair is neatly cut in a short orange bob, a few strands wet from the icy rain clinging to her smooth temples. Her emerald eyes radiate an unmistakable resolve. In the background, magenta and cyan neon lights cast hazy smears on the damp brick walls, while icy raindrops drift diagonally through the light, creating a decadent and psychedelic cyberpunk atmosphere. The girl's expression is remarkably focused. With an astonishingly steady hand, she grips an electronic eyeliner that emits a cold, blue light. Carefully, she traces the tip of the pencil along the contour of her upper lid, smoothly and decisively swiping it across, instantly creating a precise, sharp, cyborg-esque line. The cool and warm tones of neon are intertwined with the technological blue light of eyeliner, carving out distinct contours and deep shadows on her high nose bridge and tightly pursed lips. The whole picture is full of the ritual sense and resolute beauty of self-armor in the apocalyptic environment.", + "seed": 7687, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_9/index_4115_seed_7687_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02246801247820258, + "vlm_logprob": -0.02246801247820258, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.02246801247820258, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, tight close-up focused on a shattered mirror with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young woman with a short orange bob haircut and emerald eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding an electronic eyeliner that emits a cold blue light and applying it to her upper eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature magenta and cyan neon lights casting hazy smears on damp brick walls with diagonal icy raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's expression appear remarkably focused and resolute while applying the eyeliner?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0629793331027031 + ], + "opposite_logprob_tokens": [ + -2.812979221343994 + ], + "logprob_reward": -0.0629793331027031, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0629793331027031, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08360925316810608, + 0.03809367120265961, + 0.6179538369178772, + 0.07456027716398239, + 0.03173638507723808, + 0.058676786720752716, + 0.015346391126513481, + 0.020214885473251343, + 0.0433947779238224, + 0.016413766890764236 + ], + "credit_map": [ + 0.017993561923503876, + 0.010123152285814285, + 0.008947565220296383, + 0.008487769402563572, + 0.011309531517326832, + 0.007620677817612886, + 0.009779170155525208, + 0.009347815997898579, + 0.004099993035197258, + 0.0023498290684074163, + 0.0031086497474461794, + 0.0036269829142838717, + 0.004695049487054348, + 0.003457703860476613, + 0.011223922483623028, + 0.005531538277864456, + 0.02253294363617897, + 0.023313328623771667, + 0.03149563819169998, + 0.04296911135315895, + 0.04810548946261406, + 0.08032651245594025, + 0.20401206612586975, + 0.16519872844219208, + 0.0074144890531897545, + 0.008164999075233936, + 0.005933484528213739, + 0.009303157217800617, + 0.014203228987753391, + 0.0067593008279800415, + 0.013233687728643417, + 0.009547927416861057, + 0.0035804184153676033, + 0.004232566338032484, + 0.00207866495475173, + 0.003427888499572873, + 0.003457200713455677, + 0.0058472673408687115, + 0.004469718784093857, + 0.0046426597982645035, + 0.0053443568758666515, + 0.0056200213730335236, + 0.0062462566420435905, + 0.00792757049202919, + 0.010930441319942474, + 0.006526726298034191, + 0.00928370002657175, + 0.0067977155558764935, + 0.0020712430123239756, + 0.002568123396486044, + 0.0011979549890384078, + 0.001949034514836967, + 0.0018483673920854926, + 0.002137228613719344, + 0.0018910514190793037, + 0.0016833872068673372, + 0.0020226729102432728, + 0.0026043446268886328, + 0.001946865115314722, + 0.002008997369557619, + 0.0037537843454629183, + 0.002335609868168831, + 0.0031697493977844715, + 0.0023728625383228064, + 0.0035039964132010937, + 0.003819909878075123, + 0.004716258030384779, + 0.005756261758506298, + 0.006141635589301586, + 0.005837662611156702, + 0.008157672360539436, + 0.005461379885673523, + 0.0020487713627517223, + 0.0020104635041207075, + 0.0020644671749323606, + 0.0024623614735901356, + 0.0016366993077099323, + 0.00187151157297194, + 0.0025714668445289135, + 0.00174802471883595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.0629793331027031 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, tight close-up focused on a shattered mirror with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young woman with a short orange bob haircut and emerald eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding an electronic eyeliner that emits a cold blue light and applying it to her upper eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature magenta and cyan neon lights casting hazy smears on damp brick walls with diagonal icy raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's expression appear remarkably focused and resolute while applying the eyeliner?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020585810765624046 + ], + "opposite_logprob_tokens": [ + -3.895585775375366 + ], + "logprob_reward": -0.020585810765624046, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020585810765624046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30412864685058594, + 0.19832584261894226, + 0.3319360315799713, + 0.04608074948191643, + 0.03604089468717575, + 0.03574829176068306, + 0.005298545118421316, + 0.014724989421665668, + 0.006284032016992569, + 0.021431967616081238 + ], + "credit_map": [ + 0.013115902431309223, + 0.018707459792494774, + 0.01598036289215088, + 0.015606887638568878, + 0.02281039021909237, + 0.042332831770181656, + 0.08987873792648315, + 0.08569607138633728, + 0.008928797207772732, + 0.014767906628549099, + 0.013423721306025982, + 0.017739148810505867, + 0.019457386806607246, + 0.016498761251568794, + 0.07377290725708008, + 0.03373721241950989, + 0.012210406363010406, + 0.012543713673949242, + 0.017340581864118576, + 0.022965895012021065, + 0.026455441489815712, + 0.042014267295598984, + 0.1104988306760788, + 0.08790690451860428, + 0.004904996138066053, + 0.0050234184600412846, + 0.0029683976899832487, + 0.003954692743718624, + 0.00974368117749691, + 0.005864044185727835, + 0.008073451928794384, + 0.005548068322241306, + 0.003122652880847454, + 0.006224953103810549, + 0.0035733291879296303, + 0.004027827642858028, + 0.004728095605969429, + 0.004129959736019373, + 0.00616973452270031, + 0.0040643420070409775, + 0.0031363703310489655, + 0.003258215030655265, + 0.0034562021028250456, + 0.003819163190200925, + 0.007396007422357798, + 0.004837604705244303, + 0.005819831043481827, + 0.004024898633360863, + 0.0005150674260221422, + 0.0005490846233442426, + 0.00042400878737680614, + 0.0005621844902634621, + 0.0008342390065081418, + 0.0008486525039188564, + 0.000855192425660789, + 0.0007101160590536892, + 0.001545193837955594, + 0.002393831964582205, + 0.001372018945403397, + 0.0012393343495205045, + 0.002309047384187579, + 0.0018889035563915968, + 0.002223620656877756, + 0.001753038028255105, + 0.000825593713670969, + 0.0008101168787106872, + 0.0008414878975600004, + 0.0008101343992166221, + 0.0007511331932619214, + 0.0007146840216591954, + 0.0008828938589431345, + 0.0006479881703853607, + 0.0016690442571416497, + 0.0026542635168880224, + 0.001757850986905396, + 0.0030607148073613644, + 0.00276833469979465, + 0.0029529414605349302, + 0.004378722980618477, + 0.00219009630382061 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.020585810765624046 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, tight close-up focused on a shattered mirror with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young woman with a short orange bob haircut and emerald eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding an electronic eyeliner that emits a cold blue light and applying it to her upper eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature magenta and cyan neon lights casting hazy smears on damp brick walls with diagonal icy raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's expression appear remarkably focused and resolute while applying the eyeliner?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018192103132605553 + ], + "opposite_logprob_tokens": [ + -4.018192291259766 + ], + "logprob_reward": -0.018192103132605553, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018192103132605553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09912377595901489, + 0.15990254282951355, + 0.4767376780509949, + 0.07501348108053207, + 0.05857498571276665, + 0.032569218426942825, + 0.025622399523854256, + 0.0225532203912735, + 0.029221614822745323, + 0.02068110927939415 + ], + "credit_map": [ + 0.012055104598402977, + 0.009449858218431473, + 0.008805443532764912, + 0.007331745233386755, + 0.008949625305831432, + 0.017785441130399704, + 0.011179262772202492, + 0.02356729283928871, + 0.0072486805729568005, + 0.007633715867996216, + 0.009561856277287006, + 0.012585204094648361, + 0.013874243013560772, + 0.014231129549443722, + 0.05886712297797203, + 0.035900600254535675, + 0.02179902046918869, + 0.02110234647989273, + 0.02467908337712288, + 0.0329999178647995, + 0.03814418986439705, + 0.06040908768773079, + 0.1538308709859848, + 0.12377315014600754, + 0.007070047315210104, + 0.009261509403586388, + 0.010921421460807323, + 0.007925805635750294, + 0.011166795156896114, + 0.011048790067434311, + 0.007888875901699066, + 0.009730231016874313, + 0.0052493782714009285, + 0.0070035564713180065, + 0.0054331375285983086, + 0.005839626770466566, + 0.00825463980436325, + 0.009004895575344563, + 0.010713820345699787, + 0.0070759267546236515, + 0.002961549209430814, + 0.0028691564220935106, + 0.00597123010084033, + 0.0046049184165894985, + 0.0048273345455527306, + 0.003429130418226123, + 0.004171928856521845, + 0.0037339720875024796, + 0.0026798718608915806, + 0.005055452696979046, + 0.0020959742832928896, + 0.0026622654404491186, + 0.0040396167896687984, + 0.0034431659150868654, + 0.003042235504835844, + 0.0026038167998194695, + 0.0019259246764704585, + 0.003214971860870719, + 0.0025236245710402727, + 0.0028707340825349092, + 0.0027103007305413485, + 0.0034703805577009916, + 0.002682735677808523, + 0.0031545476522296667, + 0.00299669592641294, + 0.004354698583483696, + 0.004991481080651283, + 0.003491477109491825, + 0.0037744324654340744, + 0.003199242055416107, + 0.0034347460605204105, + 0.0029788401443511248, + 0.004126687068492174, + 0.0030237077735364437, + 0.001669391873292625, + 0.002777418354526162, + 0.0018682095687836409, + 0.0027363302651792765, + 0.0019918186590075493, + 0.0024875467643141747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.018192103132605553 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, tight close-up focused on a shattered mirror with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young woman with a short orange bob haircut and emerald eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding an electronic eyeliner that emits a cold blue light and applying it to her upper eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature magenta and cyan neon lights casting hazy smears on damp brick walls with diagonal icy raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's expression appear remarkably focused and resolute while applying the eyeliner?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004636015277355909 + ], + "opposite_logprob_tokens": [ + -5.379635810852051 + ], + "logprob_reward": -0.004636015277355909, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004636015277355909, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1540934145450592, + 0.28130313754081726, + 0.3701665997505188, + 0.040347870439291, + 0.035785309970378876, + 0.025868019089102745, + 0.015426614321768284, + 0.02726864628493786, + 0.02017996460199356, + 0.02956041321158409 + ], + "credit_map": [ + 0.023553408682346344, + 0.011178439483046532, + 0.010576403699815273, + 0.01173440832644701, + 0.014365327544510365, + 0.025115543976426125, + 0.02102074958384037, + 0.036549124866724014, + 0.01505052950233221, + 0.014073044992983341, + 0.01837526634335518, + 0.024505237117409706, + 0.02478099800646305, + 0.021420296281576157, + 0.11069881170988083, + 0.052398961037397385, + 0.014386357739567757, + 0.014022927731275558, + 0.01964328996837139, + 0.026545187458395958, + 0.030652878805994987, + 0.04512770101428032, + 0.12423427402973175, + 0.09555399417877197, + 0.003451294032856822, + 0.004408970475196838, + 0.003960890229791403, + 0.005652574356645346, + 0.007925816811621189, + 0.0038985745050013065, + 0.006444352213293314, + 0.004605396650731564, + 0.0061684101819992065, + 0.004911396652460098, + 0.002654047217220068, + 0.004091318231076002, + 0.004644304048269987, + 0.003946349956095219, + 0.00396379129961133, + 0.005405689124017954, + 0.003696428844705224, + 0.002591065363958478, + 0.0025257100351154804, + 0.0028142649680376053, + 0.004399089142680168, + 0.0034709640312939882, + 0.003330923616886139, + 0.0030395728535950184, + 0.0019859017338603735, + 0.002582141663879156, + 0.0013517597690224648, + 0.0019505855161696672, + 0.0022026789374649525, + 0.00203252793289721, + 0.001965600298717618, + 0.0013554184697568417, + 0.0030646224040538073, + 0.002459562150761485, + 0.0030521031003445387, + 0.002774142660200596, + 0.0037468401715159416, + 0.004372561350464821, + 0.004424860700964928, + 0.0033739530481398106, + 0.0021590159740298986, + 0.002594570629298687, + 0.0024189583491533995, + 0.0017649712972342968, + 0.0034007939975708723, + 0.0022869128733873367, + 0.003059086389839649, + 0.002495654858648777, + 0.00552341528236866, + 0.004048740956932306, + 0.0022304269950836897, + 0.004072648473083973, + 0.0025239372625947, + 0.00436189491301775, + 0.0031359849963337183, + 0.0036633636336773634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.004636015277355909 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, tight close-up focused on a shattered mirror with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young woman with a short orange bob haircut and emerald eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding an electronic eyeliner that emits a cold blue light and applying it to her upper eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature magenta and cyan neon lights casting hazy smears on damp brick walls with diagonal icy raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's expression appear remarkably focused and resolute while applying the eyeliner?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005946800112724304 + ], + "opposite_logprob_tokens": [ + -5.130946636199951 + ], + "logprob_reward": -0.005946800112724304, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005946800112724304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21182358264923096, + 0.15133418142795563, + 0.2925369143486023, + 0.06974111497402191, + 0.049236174672842026, + 0.037250980734825134, + 0.033146027475595474, + 0.04435548186302185, + 0.03880636394023895, + 0.07176920771598816 + ], + "credit_map": [ + 0.021180128678679466, + 0.01069667562842369, + 0.010377681814134121, + 0.014709255658090115, + 0.01594945602118969, + 0.03184249624609947, + 0.04677693545818329, + 0.06029094383120537, + 0.01058137696236372, + 0.007934248074889183, + 0.009532344527542591, + 0.012000184506177902, + 0.0132153844460845, + 0.01368731539696455, + 0.05251016840338707, + 0.03187315911054611, + 0.014410644769668579, + 0.013621104881167412, + 0.016199316829442978, + 0.022621285170316696, + 0.02614689990878105, + 0.03163333237171173, + 0.09801815450191498, + 0.06988618522882462, + 0.010278292931616306, + 0.007193691562861204, + 0.005655098240822554, + 0.0070862555876374245, + 0.01054295338690281, + 0.009571862407028675, + 0.011524219997227192, + 0.007888744585216045, + 0.010158834978938103, + 0.006269293837249279, + 0.004490588791668415, + 0.004043371882289648, + 0.005615857895463705, + 0.0042757298797369, + 0.007293815724551678, + 0.007088678423315287, + 0.0029040214139968157, + 0.0031386883929371834, + 0.0043913149274885654, + 0.0048560830764472485, + 0.007601762656122446, + 0.004508159589022398, + 0.005465582944452763, + 0.004385368898510933, + 0.003711874131113291, + 0.00405520386993885, + 0.002438441151753068, + 0.003922699950635433, + 0.003945180214941502, + 0.006284183356910944, + 0.004707857500761747, + 0.004080584738403559, + 0.0036973219830542803, + 0.006429574452340603, + 0.004248654004186392, + 0.005136487539857626, + 0.007785695139318705, + 0.005852856673300266, + 0.006331570912152529, + 0.004873322788625956, + 0.0038320734165608883, + 0.006031413562595844, + 0.003641055664047599, + 0.004553609061986208, + 0.005072765052318573, + 0.003467586124315858, + 0.007391462102532387, + 0.004816398490220308, + 0.012729508802294731, + 0.0095339585095644, + 0.007057607173919678, + 0.007901749573647976, + 0.009102185256779194, + 0.007558738347142935, + 0.010696384124457836, + 0.007189080119132996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.005946800112724304 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file